2019-05-27 08:55:01 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-02-02 19:15:33 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-03-23 03:00:32 -08:00
|
|
|
|
2007-05-10 22:23:21 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2016-12-24 11:46:01 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2011-07-26 16:09:06 -07:00
|
|
|
|
2012-12-17 16:05:02 -08:00
|
|
|
|
|
|
|
|
|
2013-02-24 14:52:17 -05:00
|
|
|
|
2013-11-12 15:10:16 -08:00
|
|
|
|
2024-02-13 06:16:45 +00:00
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: expand top-of-file overview / locking doc
The existing ~40-line "LOCKING:" banner covered the three-level lock
hierarchy (epnested_mutex > ep->mtx > ep->lock) but nothing else.
Lifetime rules, the ready-list state machine, the three removal paths,
and the POLLFREE contract are implicit in the code. The recent UAF
series (a6dc643c6931, 07712db80857, 8c2e52ebbe88, f2e467a48287) rode
on invariants that were only implicit.
Codify them at the top of the file: the subsystem overview, the lock
hierarchy and its mutex_lock_nested() subclass convention (reworded
from the old banner), a field-protection table for struct eventpoll
and struct epitem that names the two faces of the rbn/rcu union (rbn
under ep->mtx while linked into ep->rbr; rcu touched only by
kfree_rcu(epi) on the free path), the ovflist sentinel encoding and
scan-flip invariants, the three removal paths (A ep_remove, B
ep_clear_and_put, C eventpoll_release_file) and the epi_fget() pin
that orchestrates A vs C, and the POLLFREE store-release /
load-acquire handshake.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-1-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:32 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: expand top-of-file overview / locking doc
The existing ~40-line "LOCKING:" banner covered the three-level lock
hierarchy (epnested_mutex > ep->mtx > ep->lock) but nothing else.
Lifetime rules, the ready-list state machine, the three removal paths,
and the POLLFREE contract are implicit in the code. The recent UAF
series (a6dc643c6931, 07712db80857, 8c2e52ebbe88, f2e467a48287) rode
on invariants that were only implicit.
Codify them at the top of the file: the subsystem overview, the lock
hierarchy and its mutex_lock_nested() subclass convention (reworded
from the old banner), a field-protection table for struct eventpoll
and struct epitem that names the two faces of the rbn/rcu union (rbn
under ep->mtx while linked into ep->rbr; rcu touched only by
kfree_rcu(epi) on the free path), the ovflist sentinel encoding and
scan-flip invariants, the three removal paths (A ep_remove, B
ep_clear_and_put, C eventpoll_release_file) and the epi_fget() pin
that orchestrates A vs C, and the POLLFREE store-release /
load-acquire handshake.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-1-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:32 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: expand top-of-file overview / locking doc
The existing ~40-line "LOCKING:" banner covered the three-level lock
hierarchy (epnested_mutex > ep->mtx > ep->lock) but nothing else.
Lifetime rules, the ready-list state machine, the three removal paths,
and the POLLFREE contract are implicit in the code. The recent UAF
series (a6dc643c6931, 07712db80857, 8c2e52ebbe88, f2e467a48287) rode
on invariants that were only implicit.
Codify them at the top of the file: the subsystem overview, the lock
hierarchy and its mutex_lock_nested() subclass convention (reworded
from the old banner), a field-protection table for struct eventpoll
and struct epitem that names the two faces of the rbn/rcu union (rbn
under ep->mtx while linked into ep->rbr; rcu touched only by
kfree_rcu(epi) on the free path), the ovflist sentinel encoding and
scan-flip invariants, the three removal paths (A ep_remove, B
ep_clear_and_put, C eventpoll_release_file) and the epi_fget() pin
that orchestrates A vs C, and the POLLFREE store-release /
load-acquire handshake.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-1-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:32 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: expand top-of-file overview / locking doc
The existing ~40-line "LOCKING:" banner covered the three-level lock
hierarchy (epnested_mutex > ep->mtx > ep->lock) but nothing else.
Lifetime rules, the ready-list state machine, the three removal paths,
and the POLLFREE contract are implicit in the code. The recent UAF
series (a6dc643c6931, 07712db80857, 8c2e52ebbe88, f2e467a48287) rode
on invariants that were only implicit.
Codify them at the top of the file: the subsystem overview, the lock
hierarchy and its mutex_lock_nested() subclass convention (reworded
from the old banner), a field-protection table for struct eventpoll
and struct epitem that names the two faces of the rbn/rcu union (rbn
under ep->mtx while linked into ep->rbr; rcu touched only by
kfree_rcu(epi) on the free path), the ovflist sentinel encoding and
scan-flip invariants, the three removal paths (A ep_remove, B
ep_clear_and_put, C eventpoll_release_file) and the epi_fget() pin
that orchestrates A vs C, and the POLLFREE store-release /
load-acquire handshake.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-1-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:32 +02:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: expand top-of-file overview / locking doc
The existing ~40-line "LOCKING:" banner covered the three-level lock
hierarchy (epnested_mutex > ep->mtx > ep->lock) but nothing else.
Lifetime rules, the ready-list state machine, the three removal paths,
and the POLLFREE contract are implicit in the code. The recent UAF
series (a6dc643c6931, 07712db80857, 8c2e52ebbe88, f2e467a48287) rode
on invariants that were only implicit.
Codify them at the top of the file: the subsystem overview, the lock
hierarchy and its mutex_lock_nested() subclass convention (reworded
from the old banner), a field-protection table for struct eventpoll
and struct epitem that names the two faces of the rbn/rcu union (rbn
under ep->mtx while linked into ep->rbr; rcu touched only by
kfree_rcu(epi) on the free path), the ovflist sentinel encoding and
scan-flip invariants, the three removal paths (A ep_remove, B
ep_clear_and_put, C eventpoll_release_file) and the epi_fget() pin
that orchestrates A vs C, and the POLLFREE store-release /
load-acquire handshake.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-1-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:32 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-02-11 14:34:03 -08:00
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
2018-02-11 14:34:03 -08:00
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
|
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-10-11 01:21:44 -07:00
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
|
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
2020-09-02 11:45:57 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-09-02 11:45:57 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: optimizations and cleanups
Epoll is doing multiple passes over the ready set at the moment, because of
the constraints over the f_op->poll() call. Looking at the code again, I
noticed that we already hold the epoll semaphore in read, and this
(together with other locking conditions that hold while doing an
epoll_wait()) can lead to a smarter way [1] to "ship" events to userspace
(in a single pass).
This is a stress application that can be used to test the new code. It
spwans multiple thread and call epoll_wait() and epoll_ctl() from many
threads. Stress tested on my dual Opteron 254 w/out any problems.
http://www.xmailserver.org/totalmess.c
This is not a benchmark, just something that tries to stress and exploit
possible problems with the new code.
Also, I made a stupid micro-benchmark:
http://www.xmailserver.org/epwbench.c
[1] Considering that epoll must be thread-safe, there are five ways we can
be hit during an epoll_wait() transfer loop (ep_send_events()):
1) The epoll fd going away and calling ep_free
This just can't happen, since we did an fget() in sys_epoll_wait
2) An epoll_ctl(EPOLL_CTL_DEL)
This can't happen because epoll_ctl() gets ep->sem in write, and
we're holding it in read during ep_send_events()
3) An fd stored inside the epoll fd going away
This can't happen because in eventpoll_release_file() we get
ep->sem in write, and we're holding it in read during
ep_send_events()
4) Another epoll_wait() happening on another thread
They both can be inside ep_send_events() at the same time, we get
(splice) the ready-list under the spinlock, so each one will get
its own ready list. Note that an fd cannot be at the same time
inside more than one ready list, because ep_poll_callback() will
not re-queue it if it sees it already linked:
if (ep_is_linked(&epi->rdllink))
goto is_linked;
Another case that can happen, is two concurrent epoll_wait(),
coming in with a userspace event buffer of size, say, ten.
Suppose there are 50 event ready in the list. The first
epoll_wait() will "steal" the whole list, while the second, seeing
no events, will go to sleep. But at the end of ep_send_events() in
the first epoll_wait(), we will re-inject surplus ready fds, and we
will trigger the proper wake_up to the second epoll_wait().
5) ep_poll_callback() hitting us asyncronously
This is the tricky part. As I said above, the ep_is_linked() test
done inside ep_poll_callback(), will guarantee us that until the
item will result linked to a list, ep_poll_callback() will not try
to re-queue it again (read, write data on any of its members). When
we do a list_del() in ep_send_events(), the item will still satisfy
the ep_is_linked() test (whatever data is written in prev/next,
it'll never be its own pointer), so ep_poll_callback() will still
leave us alone. It's only after the eventual smp_mb()+INIT_LIST_HEAD(&epi->rdllink)
that it'll become visible to ep_poll_callback(), but at the point
we're already past it.
[akpm@osdl.org: 80 cols]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-08 00:25:41 -07:00
|
|
|
|
2013-04-30 15:27:38 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2013-11-12 15:10:16 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
2007-05-15 01:40:47 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-05-14 08:07:20 -06:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2020-09-02 11:45:57 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-27 11:18:30 -04:00
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2007-05-15 01:40:47 -07:00
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-30 23:42:29 +02:00
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
2019-03-07 16:28:53 -08:00
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:52 -07:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-06 09:43:53 +08:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
|
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-10 08:30:05 -04:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
2026-04-23 11:56:13 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
2026-03-31 15:25:32 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:42 +00:00
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:43 +00:00
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:44 +00:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
2020-04-06 20:11:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-12 17:00:01 -08:00
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
eventpoll: document loop-check / path-check globals
The globals that support EPOLL_CTL_ADD's cycle and path-length checks
are scattered: epnested_mutex, loop_check_gen, inserting_into, and
tfile_check_list sit at the top of the file; path_count[] and
path_limits[] are declared inline with the path-check code further
down. Their interaction -- the "ep->gen == loop_check_gen" trigger in
do_epoll_ctl(), the two loop_check_gen++ bumps that sandwich a check,
the EP_UNACTIVE_PTR sentinel on tfile_check_list, the -ELOOP back-edge
detection via inserting_into -- is not documented anywhere.
The area has had three recent fixes (CVE-2025-38349, the unbounded
recursion fix, and the overflow fix) whose logic depends on these
invariants. Collect the description in one block alongside the
declarations, cross-reference the path_count[] declaration that lives
with the path-check code, and name the fix commits so future readers
can find the context.
Also add a short comment on struct epitems_head describing its
dual use (wrapper for non-epoll file->f_ep versus pointing into
&ep->refs for the epoll-watches-epoll case), which the old comment
on tfile_check_list had accidentally attached to the struct.
Comment-only; no functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-2-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:33 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: document loop-check / path-check globals
The globals that support EPOLL_CTL_ADD's cycle and path-length checks
are scattered: epnested_mutex, loop_check_gen, inserting_into, and
tfile_check_list sit at the top of the file; path_count[] and
path_limits[] are declared inline with the path-check code further
down. Their interaction -- the "ep->gen == loop_check_gen" trigger in
do_epoll_ctl(), the two loop_check_gen++ bumps that sandwich a check,
the EP_UNACTIVE_PTR sentinel on tfile_check_list, the -ELOOP back-edge
detection via inserting_into -- is not documented anywhere.
The area has had three recent fixes (CVE-2025-38349, the unbounded
recursion fix, and the overflow fix) whose logic depends on these
invariants. Collect the description in one block alongside the
declarations, cross-reference the path_count[] declaration that lives
with the path-check code, and name the fix commits so future readers
can find the context.
Also add a short comment on struct epitems_head describing its
dual use (wrapper for non-epoll file->f_ep versus pointing into
&ep->refs for the epoll-watches-epoll case), which the old comment
on tfile_check_list had accidentally attached to the struct.
Comment-only; no functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-2-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:33 +02:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: document loop-check / path-check globals
The globals that support EPOLL_CTL_ADD's cycle and path-length checks
are scattered: epnested_mutex, loop_check_gen, inserting_into, and
tfile_check_list sit at the top of the file; path_count[] and
path_limits[] are declared inline with the path-check code further
down. Their interaction -- the "ep->gen == loop_check_gen" trigger in
do_epoll_ctl(), the two loop_check_gen++ bumps that sandwich a check,
the EP_UNACTIVE_PTR sentinel on tfile_check_list, the -ELOOP back-edge
detection via inserting_into -- is not documented anywhere.
The area has had three recent fixes (CVE-2025-38349, the unbounded
recursion fix, and the overflow fix) whose logic depends on these
invariants. Collect the description in one block alongside the
declarations, cross-reference the path_count[] declaration that lives
with the path-check code, and name the fix commits so future readers
can find the context.
Also add a short comment on struct epitems_head describing its
dual use (wrapper for non-epoll file->f_ep versus pointing into
&ep->refs for the epoll-watches-epoll case), which the old comment
on tfile_check_list had accidentally attached to the struct.
Comment-only; no functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-2-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:33 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: document loop-check / path-check globals
The globals that support EPOLL_CTL_ADD's cycle and path-length checks
are scattered: epnested_mutex, loop_check_gen, inserting_into, and
tfile_check_list sit at the top of the file; path_count[] and
path_limits[] are declared inline with the path-check code further
down. Their interaction -- the "ep->gen == loop_check_gen" trigger in
do_epoll_ctl(), the two loop_check_gen++ bumps that sandwich a check,
the EP_UNACTIVE_PTR sentinel on tfile_check_list, the -ELOOP back-edge
detection via inserting_into -- is not documented anywhere.
The area has had three recent fixes (CVE-2025-38349, the unbounded
recursion fix, and the overflow fix) whose logic depends on these
invariants. Collect the description in one block alongside the
declarations, cross-reference the path_count[] declaration that lives
with the path-check code, and name the fix commits so future readers
can find the context.
Also add a short comment on struct epitems_head describing its
dual use (wrapper for non-epoll file->f_ep versus pointing into
&ep->refs for the epoll-watches-epoll case), which the old comment
on tfile_check_list had accidentally attached to the struct.
Comment-only; no functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-2-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:33 +02:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: document loop-check / path-check globals
The globals that support EPOLL_CTL_ADD's cycle and path-length checks
are scattered: epnested_mutex, loop_check_gen, inserting_into, and
tfile_check_list sit at the top of the file; path_count[] and
path_limits[] are declared inline with the path-check code further
down. Their interaction -- the "ep->gen == loop_check_gen" trigger in
do_epoll_ctl(), the two loop_check_gen++ bumps that sandwich a check,
the EP_UNACTIVE_PTR sentinel on tfile_check_list, the -ELOOP back-edge
detection via inserting_into -- is not documented anywhere.
The area has had three recent fixes (CVE-2025-38349, the unbounded
recursion fix, and the overflow fix) whose logic depends on these
invariants. Collect the description in one block alongside the
declarations, cross-reference the path_count[] declaration that lives
with the path-check code, and name the fix commits so future readers
can find the context.
Also add a short comment on struct epitems_head describing its
dual use (wrapper for non-epoll file->f_ep versus pointing into
&ep->refs for the epoll-watches-epoll case), which the old comment
on tfile_check_list had accidentally attached to the struct.
Comment-only; no functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-2-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:33 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-11 16:41:59 -07:00
|
|
|
|
2020-09-10 08:30:05 -04:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: restore EP_UNACTIVE_PTR sentinel for ctx->tfile_check_list
Commit e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into
struct ep_ctl_ctx") moved tfile_check_list from a file-scope global into
the stack-allocated struct ep_ctl_ctx, and in doing so replaced the
EP_UNACTIVE_PTR sentinel with NULL on the grounds that "NULL is the
obvious 'empty' value and zero-init handles it for free", describing the
change as "No functional change". It is not.
epitems_head->next is overloaded with two roles:
1. the "next" pointer that threads a head onto ctx->tfile_check_list;
2. a membership flag: ep_remove_file() uses
!smp_load_acquire(&v->next) to mean "this head is not on any
pending ctx->tfile_check_list and is therefore safe to free".
Before that change the EP_UNACTIVE_PTR sentinel kept the two roles
disjoint: a head on the list always had a non-NULL ->next (another head,
or the sentinel at the tail), so ->next == NULL was equivalent to "never
listed". With the sentinel gone the list is NULL-terminated, so the tail
head's ->next is NULL as well. ep_remove_file()'s gate can no longer
distinguish "never listed" from "listed at the tail", and misfires on
the tail head.
The reader (reverse_path_check_proc) holds epnested_mutex +
rcu_read_lock; the freer (ep_remove_file) holds ep->mtx + file->f_lock.
The two sides share no mutex -- the sentinel was the invariant the gate
relied on to know it could skip the read side. With it gone,
ep_remove_file() frees the tail head while reverse_path_check_proc() is
still walking it, producing the slab-use-after-free read. The syzbot
reproducer hits this within seconds on a multi-CPU VM.
Restore the sentinel: initialize ctx.tfile_check_list to EP_UNACTIVE_PTR
in do_epoll_ctl_file(), and terminate the walk on "!= EP_UNACTIVE_PTR"
in reverse_path_check() and clear_tfile_check_list(). The tail head's
->next becomes the sentinel again rather than NULL, so
ep_remove_file()'s gate regains its exclusivity and stops misfiring on
the tail. ep_remove_file() itself is unchanged.
This restores the invariant the file-scope tfile_check_list relied on
before that change while preserving the ctx packaging it introduced.
Reported-by: syzbot+e70e1b6cba8714543f7c@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=e70e1b6cba8714543f7c
Fixes: e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into struct ep_ctl_ctx")
Suggested-by: Christian Brauner <brauner@kernel.org>
Link: https://lore.kernel.org/all/20260528-rotwild-summt-kuhhandel-7276ef4c33b7@brauner.io/
Signed-off-by: Zhan Wei <zhanwei919@gmail.com>
Link: https://patch.msgid.link/20260529142533.23696-1-zhanwei919@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-05-29 22:25:33 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-10-11 19:55:00 +03:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2023-10-11 19:55:00 +03:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
eventpoll: document loop-check / path-check globals
The globals that support EPOLL_CTL_ADD's cycle and path-length checks
are scattered: epnested_mutex, loop_check_gen, inserting_into, and
tfile_check_list sit at the top of the file; path_count[] and
path_limits[] are declared inline with the path-check code further
down. Their interaction -- the "ep->gen == loop_check_gen" trigger in
do_epoll_ctl(), the two loop_check_gen++ bumps that sandwich a check,
the EP_UNACTIVE_PTR sentinel on tfile_check_list, the -ELOOP back-edge
detection via inserting_into -- is not documented anywhere.
The area has had three recent fixes (CVE-2025-38349, the unbounded
recursion fix, and the overflow fix) whose logic depends on these
invariants. Collect the description in one block alongside the
declarations, cross-reference the path_count[] declaration that lives
with the path-check code, and name the fix commits so future readers
can find the context.
Also add a short comment on struct epitems_head describing its
dual use (wrapper for non-epoll file->f_ep versus pointing into
&ep->refs for the epoll-watches-epoll case), which the old comment
on tfile_check_list had accidentally attached to the struct.
Comment-only; no functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-2-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:33 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: pin files while checking reverse paths
Commit 319c15174757 ("epoll: take epitem list out of struct file")
intentionally removed temporary file references from the reverse path
check list. At the time, both epitems and their files were freed after
an RCU grace period, so unlist_file() could obtain file->f_lock through
an epitem while clear_tfile_check_list() held rcu_read_lock().
Commit 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU") made
struct file SLAB_TYPESAFE_BY_RCU and removed its RCU-delayed freeing.
RCU still protects the epitem, but no longer keeps the referenced file
from being freed and reused. A concurrent close can therefore make
unlist_file() lock or unlock f_lock in a recycled file object.
This violates the documented SLAB_TYPESAFE_BY_RCU rule requiring a
reference before acquiring an object's lock. The race was reproduced,
causing a wild unlock of f_lock in a recycled file and breaking its
mutual exclusion.
Add ->file to epitems_head to remember the pinned file independently of
->epitems. A concurrent EPOLL_CTL_DEL can empty ->epitems before the head
is unlisted, leaving no epi->ffd.file from which to drop the reference.
In list_file(), acquire the reference before adding the head to the
check list. The caller either owns a reference or holds the ep->mtx for
the epitem leading to the file. In the latter case, file_ref_get() can
fail after the last reference is dropped, but eventpoll_release_file()
must acquire the same mutex before the file can be freed. The dying leaf
can be skipped because removing links cannot increase the reverse path
count.
In unlist_file(), epnested_mutex excludes another list_file() or
unlist_file(), while head->next prevents a concurrent EPOLL_CTL_DEL from
freeing the head. Save head->file locally, clear it with head->next
under f_lock, and drop the reference after the RCU-protected operation.
Christian Brauner <brauner@kernel.org> quotes:
> SLAB_TYPESAFE_BY_RCU allows a slab slot to be reused while an RCU reader
> still holds its old address. Once that address contains a new live
> struct file, KASAN sees valid, unpoisoned memory and cannot distinguish
> the stale object identity. CONFIG_DEBUG_SPINLOCK exposes the failure
> instead.
>
> The failing interleaving is:
>
> CPU0: nested EPOLL_CTL_ADD CPU1: close/open churn
> ------------------------------------ ---------------------------------
> p = hlist_first_rcu(&head->epitems)
> epi = container_of(p, ...)
> close(victim)
> __fput()
> eventpoll_release_file()
> file_free(victim)
> // the slot is free; f_lock remains
> spin_lock(&epi->ffd.file->f_lock)
> open() reuses the slot as new_file
> spin_lock_init(&new_file->f_lock)
> spin_unlock(&epi->ffd.file->f_lock) // wild unlock of new_file's lock
>
> CONFIG_DEBUG_SPINLOCK reports:
>
> BUG: spinlock already unlocked on CPU#0, poc_unlist/150
> lock: 0xffff8880067fb200, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
> CPU: 0 UID: 1000 PID: 150 Comm: poc_unlist Not tainted 7.2.0-rc3-dirty #22 PREEMPTLAZY
> Hardware name: QEMU Ubuntu 24.04 PC v2 (i440FX + PIIX, arch_caps fix, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> Call Trace:
> <TASK>
> dump_stack_lvl+0x64/0x80
> do_raw_spin_unlock+0x75/0xb0
> _raw_spin_unlock+0xe/0x30
> clear_tfile_check_list+0x88/0xe0
> do_epoll_ctl_file+0x519/0xcf0
> ? __pfx_ep_ptable_queue_proc+0x10/0x10
> do_epoll_ctl+0x8f/0x100
> __x64_sys_epoll_ctl+0x6f/0xa0
> do_syscall_64+0xdc/0x520
> ? srso_alias_return_thunk+0x5/0xfbef5
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
> RIP: 0033:0x42034e
> Code: 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 e9 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
> RSP: 002b:00007a657ff3c198 EFLAGS: 00000202 ORIG_RAX: 00000000000000e9
> RAX: ffffffffffffffda RBX: 00007a657ff3ccdc RCX: 000000000042034e
> RDX: 0000000000000003 RSI: 0000000000000001 RDI: 0000000000000004
> RBP: 00007a657ff3c2f0 R08: 0000000000000000 R09: 00007a657ff3c6c0
> R10: 00007a657ff3c1a4 R11: 0000000000000202 R12: 00007a657ff3c6c0
> R13: ffffffffffffffb8 R14: 000000000000000d R15: 00007fffb7de0210
> </TASK>
> ------------[ cut here ]------------
>
> unlist_file() does not appear as a separate frame because it was inlined
> into clear_tfile_check_list(). This report was obtained with mdelay()
> instrumentation immediately before spin_lock() and spin_unlock() in
> unlist_file() to widen the two race windows.
>
> More importantly, this is a wild unlock. The stale unlock can target
> f_lock of a different live file and invalidate mutual exclusion for
> state protected by that lock. Turning this into a reliable exploit
> would require precise scheduling and same-slot reuse and is likely
> difficult, but the primitive is potentially exploitable.
Reported-by: Qi Tang <tpluszz77@gmail.com>
Reported-by: Junxi Qian <qjx1298677004@gmail.com>
Fixes: 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU")
Cc: stable@vger.kernel.org
Signed-off-by: Guidong Han <2045gemini@gmail.com>
Link: https://patch.msgid.link/20260718104406.27897-1-2045gemini@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-18 18:44:06 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: pin files while checking reverse paths
Commit 319c15174757 ("epoll: take epitem list out of struct file")
intentionally removed temporary file references from the reverse path
check list. At the time, both epitems and their files were freed after
an RCU grace period, so unlist_file() could obtain file->f_lock through
an epitem while clear_tfile_check_list() held rcu_read_lock().
Commit 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU") made
struct file SLAB_TYPESAFE_BY_RCU and removed its RCU-delayed freeing.
RCU still protects the epitem, but no longer keeps the referenced file
from being freed and reused. A concurrent close can therefore make
unlist_file() lock or unlock f_lock in a recycled file object.
This violates the documented SLAB_TYPESAFE_BY_RCU rule requiring a
reference before acquiring an object's lock. The race was reproduced,
causing a wild unlock of f_lock in a recycled file and breaking its
mutual exclusion.
Add ->file to epitems_head to remember the pinned file independently of
->epitems. A concurrent EPOLL_CTL_DEL can empty ->epitems before the head
is unlisted, leaving no epi->ffd.file from which to drop the reference.
In list_file(), acquire the reference before adding the head to the
check list. The caller either owns a reference or holds the ep->mtx for
the epitem leading to the file. In the latter case, file_ref_get() can
fail after the last reference is dropped, but eventpoll_release_file()
must acquire the same mutex before the file can be freed. The dying leaf
can be skipped because removing links cannot increase the reverse path
count.
In unlist_file(), epnested_mutex excludes another list_file() or
unlist_file(), while head->next prevents a concurrent EPOLL_CTL_DEL from
freeing the head. Save head->file locally, clear it with head->next
under f_lock, and drop the reference after the RCU-protected operation.
Christian Brauner <brauner@kernel.org> quotes:
> SLAB_TYPESAFE_BY_RCU allows a slab slot to be reused while an RCU reader
> still holds its old address. Once that address contains a new live
> struct file, KASAN sees valid, unpoisoned memory and cannot distinguish
> the stale object identity. CONFIG_DEBUG_SPINLOCK exposes the failure
> instead.
>
> The failing interleaving is:
>
> CPU0: nested EPOLL_CTL_ADD CPU1: close/open churn
> ------------------------------------ ---------------------------------
> p = hlist_first_rcu(&head->epitems)
> epi = container_of(p, ...)
> close(victim)
> __fput()
> eventpoll_release_file()
> file_free(victim)
> // the slot is free; f_lock remains
> spin_lock(&epi->ffd.file->f_lock)
> open() reuses the slot as new_file
> spin_lock_init(&new_file->f_lock)
> spin_unlock(&epi->ffd.file->f_lock) // wild unlock of new_file's lock
>
> CONFIG_DEBUG_SPINLOCK reports:
>
> BUG: spinlock already unlocked on CPU#0, poc_unlist/150
> lock: 0xffff8880067fb200, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
> CPU: 0 UID: 1000 PID: 150 Comm: poc_unlist Not tainted 7.2.0-rc3-dirty #22 PREEMPTLAZY
> Hardware name: QEMU Ubuntu 24.04 PC v2 (i440FX + PIIX, arch_caps fix, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> Call Trace:
> <TASK>
> dump_stack_lvl+0x64/0x80
> do_raw_spin_unlock+0x75/0xb0
> _raw_spin_unlock+0xe/0x30
> clear_tfile_check_list+0x88/0xe0
> do_epoll_ctl_file+0x519/0xcf0
> ? __pfx_ep_ptable_queue_proc+0x10/0x10
> do_epoll_ctl+0x8f/0x100
> __x64_sys_epoll_ctl+0x6f/0xa0
> do_syscall_64+0xdc/0x520
> ? srso_alias_return_thunk+0x5/0xfbef5
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
> RIP: 0033:0x42034e
> Code: 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 e9 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
> RSP: 002b:00007a657ff3c198 EFLAGS: 00000202 ORIG_RAX: 00000000000000e9
> RAX: ffffffffffffffda RBX: 00007a657ff3ccdc RCX: 000000000042034e
> RDX: 0000000000000003 RSI: 0000000000000001 RDI: 0000000000000004
> RBP: 00007a657ff3c2f0 R08: 0000000000000000 R09: 00007a657ff3c6c0
> R10: 00007a657ff3c1a4 R11: 0000000000000202 R12: 00007a657ff3c6c0
> R13: ffffffffffffffb8 R14: 000000000000000d R15: 00007fffb7de0210
> </TASK>
> ------------[ cut here ]------------
>
> unlist_file() does not appear as a separate frame because it was inlined
> into clear_tfile_check_list(). This report was obtained with mdelay()
> instrumentation immediately before spin_lock() and spin_unlock() in
> unlist_file() to widen the two race windows.
>
> More importantly, this is a wild unlock. The stale unlock can target
> f_lock of a different live file and invalidate mutual exclusion for
> state protected by that lock. Turning this into a reliable exploit
> would require precise scheduling and same-slot reuse and is likely
> difficult, but the primitive is potentially exploitable.
Reported-by: Qi Tang <tpluszz77@gmail.com>
Reported-by: Junxi Qian <qjx1298677004@gmail.com>
Fixes: 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU")
Cc: stable@vger.kernel.org
Signed-off-by: Guidong Han <2045gemini@gmail.com>
Link: https://patch.msgid.link/20260718104406.27897-1-2045gemini@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-18 18:44:06 +08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
2023-10-11 19:55:00 +03:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: pin files while checking reverse paths
Commit 319c15174757 ("epoll: take epitem list out of struct file")
intentionally removed temporary file references from the reverse path
check list. At the time, both epitems and their files were freed after
an RCU grace period, so unlist_file() could obtain file->f_lock through
an epitem while clear_tfile_check_list() held rcu_read_lock().
Commit 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU") made
struct file SLAB_TYPESAFE_BY_RCU and removed its RCU-delayed freeing.
RCU still protects the epitem, but no longer keeps the referenced file
from being freed and reused. A concurrent close can therefore make
unlist_file() lock or unlock f_lock in a recycled file object.
This violates the documented SLAB_TYPESAFE_BY_RCU rule requiring a
reference before acquiring an object's lock. The race was reproduced,
causing a wild unlock of f_lock in a recycled file and breaking its
mutual exclusion.
Add ->file to epitems_head to remember the pinned file independently of
->epitems. A concurrent EPOLL_CTL_DEL can empty ->epitems before the head
is unlisted, leaving no epi->ffd.file from which to drop the reference.
In list_file(), acquire the reference before adding the head to the
check list. The caller either owns a reference or holds the ep->mtx for
the epitem leading to the file. In the latter case, file_ref_get() can
fail after the last reference is dropped, but eventpoll_release_file()
must acquire the same mutex before the file can be freed. The dying leaf
can be skipped because removing links cannot increase the reverse path
count.
In unlist_file(), epnested_mutex excludes another list_file() or
unlist_file(), while head->next prevents a concurrent EPOLL_CTL_DEL from
freeing the head. Save head->file locally, clear it with head->next
under f_lock, and drop the reference after the RCU-protected operation.
Christian Brauner <brauner@kernel.org> quotes:
> SLAB_TYPESAFE_BY_RCU allows a slab slot to be reused while an RCU reader
> still holds its old address. Once that address contains a new live
> struct file, KASAN sees valid, unpoisoned memory and cannot distinguish
> the stale object identity. CONFIG_DEBUG_SPINLOCK exposes the failure
> instead.
>
> The failing interleaving is:
>
> CPU0: nested EPOLL_CTL_ADD CPU1: close/open churn
> ------------------------------------ ---------------------------------
> p = hlist_first_rcu(&head->epitems)
> epi = container_of(p, ...)
> close(victim)
> __fput()
> eventpoll_release_file()
> file_free(victim)
> // the slot is free; f_lock remains
> spin_lock(&epi->ffd.file->f_lock)
> open() reuses the slot as new_file
> spin_lock_init(&new_file->f_lock)
> spin_unlock(&epi->ffd.file->f_lock) // wild unlock of new_file's lock
>
> CONFIG_DEBUG_SPINLOCK reports:
>
> BUG: spinlock already unlocked on CPU#0, poc_unlist/150
> lock: 0xffff8880067fb200, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
> CPU: 0 UID: 1000 PID: 150 Comm: poc_unlist Not tainted 7.2.0-rc3-dirty #22 PREEMPTLAZY
> Hardware name: QEMU Ubuntu 24.04 PC v2 (i440FX + PIIX, arch_caps fix, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> Call Trace:
> <TASK>
> dump_stack_lvl+0x64/0x80
> do_raw_spin_unlock+0x75/0xb0
> _raw_spin_unlock+0xe/0x30
> clear_tfile_check_list+0x88/0xe0
> do_epoll_ctl_file+0x519/0xcf0
> ? __pfx_ep_ptable_queue_proc+0x10/0x10
> do_epoll_ctl+0x8f/0x100
> __x64_sys_epoll_ctl+0x6f/0xa0
> do_syscall_64+0xdc/0x520
> ? srso_alias_return_thunk+0x5/0xfbef5
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
> RIP: 0033:0x42034e
> Code: 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 e9 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
> RSP: 002b:00007a657ff3c198 EFLAGS: 00000202 ORIG_RAX: 00000000000000e9
> RAX: ffffffffffffffda RBX: 00007a657ff3ccdc RCX: 000000000042034e
> RDX: 0000000000000003 RSI: 0000000000000001 RDI: 0000000000000004
> RBP: 00007a657ff3c2f0 R08: 0000000000000000 R09: 00007a657ff3c6c0
> R10: 00007a657ff3c1a4 R11: 0000000000000202 R12: 00007a657ff3c6c0
> R13: ffffffffffffffb8 R14: 000000000000000d R15: 00007fffb7de0210
> </TASK>
> ------------[ cut here ]------------
>
> unlist_file() does not appear as a separate frame because it was inlined
> into clear_tfile_check_list(). This report was obtained with mdelay()
> instrumentation immediately before spin_lock() and spin_unlock() in
> unlist_file() to widen the two race windows.
>
> More importantly, this is a wild unlock. The stale unlock can target
> f_lock of a different live file and invalidate mutual exclusion for
> state protected by that lock. Turning this into a reliable exploit
> would require precise scheduling and same-slot reuse and is likely
> difficult, but the primitive is potentially exploitable.
Reported-by: Qi Tang <tpluszz77@gmail.com>
Reported-by: Junxi Qian <qjx1298677004@gmail.com>
Fixes: 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU")
Cc: stable@vger.kernel.org
Signed-off-by: Guidong Han <2045gemini@gmail.com>
Link: https://patch.msgid.link/20260718104406.27897-1-2045gemini@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-18 18:44:06 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: pin files while checking reverse paths
Commit 319c15174757 ("epoll: take epitem list out of struct file")
intentionally removed temporary file references from the reverse path
check list. At the time, both epitems and their files were freed after
an RCU grace period, so unlist_file() could obtain file->f_lock through
an epitem while clear_tfile_check_list() held rcu_read_lock().
Commit 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU") made
struct file SLAB_TYPESAFE_BY_RCU and removed its RCU-delayed freeing.
RCU still protects the epitem, but no longer keeps the referenced file
from being freed and reused. A concurrent close can therefore make
unlist_file() lock or unlock f_lock in a recycled file object.
This violates the documented SLAB_TYPESAFE_BY_RCU rule requiring a
reference before acquiring an object's lock. The race was reproduced,
causing a wild unlock of f_lock in a recycled file and breaking its
mutual exclusion.
Add ->file to epitems_head to remember the pinned file independently of
->epitems. A concurrent EPOLL_CTL_DEL can empty ->epitems before the head
is unlisted, leaving no epi->ffd.file from which to drop the reference.
In list_file(), acquire the reference before adding the head to the
check list. The caller either owns a reference or holds the ep->mtx for
the epitem leading to the file. In the latter case, file_ref_get() can
fail after the last reference is dropped, but eventpoll_release_file()
must acquire the same mutex before the file can be freed. The dying leaf
can be skipped because removing links cannot increase the reverse path
count.
In unlist_file(), epnested_mutex excludes another list_file() or
unlist_file(), while head->next prevents a concurrent EPOLL_CTL_DEL from
freeing the head. Save head->file locally, clear it with head->next
under f_lock, and drop the reference after the RCU-protected operation.
Christian Brauner <brauner@kernel.org> quotes:
> SLAB_TYPESAFE_BY_RCU allows a slab slot to be reused while an RCU reader
> still holds its old address. Once that address contains a new live
> struct file, KASAN sees valid, unpoisoned memory and cannot distinguish
> the stale object identity. CONFIG_DEBUG_SPINLOCK exposes the failure
> instead.
>
> The failing interleaving is:
>
> CPU0: nested EPOLL_CTL_ADD CPU1: close/open churn
> ------------------------------------ ---------------------------------
> p = hlist_first_rcu(&head->epitems)
> epi = container_of(p, ...)
> close(victim)
> __fput()
> eventpoll_release_file()
> file_free(victim)
> // the slot is free; f_lock remains
> spin_lock(&epi->ffd.file->f_lock)
> open() reuses the slot as new_file
> spin_lock_init(&new_file->f_lock)
> spin_unlock(&epi->ffd.file->f_lock) // wild unlock of new_file's lock
>
> CONFIG_DEBUG_SPINLOCK reports:
>
> BUG: spinlock already unlocked on CPU#0, poc_unlist/150
> lock: 0xffff8880067fb200, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
> CPU: 0 UID: 1000 PID: 150 Comm: poc_unlist Not tainted 7.2.0-rc3-dirty #22 PREEMPTLAZY
> Hardware name: QEMU Ubuntu 24.04 PC v2 (i440FX + PIIX, arch_caps fix, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> Call Trace:
> <TASK>
> dump_stack_lvl+0x64/0x80
> do_raw_spin_unlock+0x75/0xb0
> _raw_spin_unlock+0xe/0x30
> clear_tfile_check_list+0x88/0xe0
> do_epoll_ctl_file+0x519/0xcf0
> ? __pfx_ep_ptable_queue_proc+0x10/0x10
> do_epoll_ctl+0x8f/0x100
> __x64_sys_epoll_ctl+0x6f/0xa0
> do_syscall_64+0xdc/0x520
> ? srso_alias_return_thunk+0x5/0xfbef5
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
> RIP: 0033:0x42034e
> Code: 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 e9 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
> RSP: 002b:00007a657ff3c198 EFLAGS: 00000202 ORIG_RAX: 00000000000000e9
> RAX: ffffffffffffffda RBX: 00007a657ff3ccdc RCX: 000000000042034e
> RDX: 0000000000000003 RSI: 0000000000000001 RDI: 0000000000000004
> RBP: 00007a657ff3c2f0 R08: 0000000000000000 R09: 00007a657ff3c6c0
> R10: 00007a657ff3c1a4 R11: 0000000000000202 R12: 00007a657ff3c6c0
> R13: ffffffffffffffb8 R14: 000000000000000d R15: 00007fffb7de0210
> </TASK>
> ------------[ cut here ]------------
>
> unlist_file() does not appear as a separate frame because it was inlined
> into clear_tfile_check_list(). This report was obtained with mdelay()
> instrumentation immediately before spin_lock() and spin_unlock() in
> unlist_file() to widen the two race windows.
>
> More importantly, this is a wild unlock. The stale unlock can target
> f_lock of a different live file and invalidate mutual exclusion for
> state protected by that lock. Turning this into a reliable exploit
> would require precise scheduling and same-slot reuse and is likely
> difficult, but the primitive is potentially exploitable.
Reported-by: Qi Tang <tpluszz77@gmail.com>
Reported-by: Junxi Qian <qjx1298677004@gmail.com>
Fixes: 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU")
Cc: stable@vger.kernel.org
Signed-off-by: Guidong Han <2045gemini@gmail.com>
Link: https://patch.msgid.link/20260718104406.27897-1-2045gemini@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-18 18:44:06 +08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: pin files while checking reverse paths
Commit 319c15174757 ("epoll: take epitem list out of struct file")
intentionally removed temporary file references from the reverse path
check list. At the time, both epitems and their files were freed after
an RCU grace period, so unlist_file() could obtain file->f_lock through
an epitem while clear_tfile_check_list() held rcu_read_lock().
Commit 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU") made
struct file SLAB_TYPESAFE_BY_RCU and removed its RCU-delayed freeing.
RCU still protects the epitem, but no longer keeps the referenced file
from being freed and reused. A concurrent close can therefore make
unlist_file() lock or unlock f_lock in a recycled file object.
This violates the documented SLAB_TYPESAFE_BY_RCU rule requiring a
reference before acquiring an object's lock. The race was reproduced,
causing a wild unlock of f_lock in a recycled file and breaking its
mutual exclusion.
Add ->file to epitems_head to remember the pinned file independently of
->epitems. A concurrent EPOLL_CTL_DEL can empty ->epitems before the head
is unlisted, leaving no epi->ffd.file from which to drop the reference.
In list_file(), acquire the reference before adding the head to the
check list. The caller either owns a reference or holds the ep->mtx for
the epitem leading to the file. In the latter case, file_ref_get() can
fail after the last reference is dropped, but eventpoll_release_file()
must acquire the same mutex before the file can be freed. The dying leaf
can be skipped because removing links cannot increase the reverse path
count.
In unlist_file(), epnested_mutex excludes another list_file() or
unlist_file(), while head->next prevents a concurrent EPOLL_CTL_DEL from
freeing the head. Save head->file locally, clear it with head->next
under f_lock, and drop the reference after the RCU-protected operation.
Christian Brauner <brauner@kernel.org> quotes:
> SLAB_TYPESAFE_BY_RCU allows a slab slot to be reused while an RCU reader
> still holds its old address. Once that address contains a new live
> struct file, KASAN sees valid, unpoisoned memory and cannot distinguish
> the stale object identity. CONFIG_DEBUG_SPINLOCK exposes the failure
> instead.
>
> The failing interleaving is:
>
> CPU0: nested EPOLL_CTL_ADD CPU1: close/open churn
> ------------------------------------ ---------------------------------
> p = hlist_first_rcu(&head->epitems)
> epi = container_of(p, ...)
> close(victim)
> __fput()
> eventpoll_release_file()
> file_free(victim)
> // the slot is free; f_lock remains
> spin_lock(&epi->ffd.file->f_lock)
> open() reuses the slot as new_file
> spin_lock_init(&new_file->f_lock)
> spin_unlock(&epi->ffd.file->f_lock) // wild unlock of new_file's lock
>
> CONFIG_DEBUG_SPINLOCK reports:
>
> BUG: spinlock already unlocked on CPU#0, poc_unlist/150
> lock: 0xffff8880067fb200, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
> CPU: 0 UID: 1000 PID: 150 Comm: poc_unlist Not tainted 7.2.0-rc3-dirty #22 PREEMPTLAZY
> Hardware name: QEMU Ubuntu 24.04 PC v2 (i440FX + PIIX, arch_caps fix, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> Call Trace:
> <TASK>
> dump_stack_lvl+0x64/0x80
> do_raw_spin_unlock+0x75/0xb0
> _raw_spin_unlock+0xe/0x30
> clear_tfile_check_list+0x88/0xe0
> do_epoll_ctl_file+0x519/0xcf0
> ? __pfx_ep_ptable_queue_proc+0x10/0x10
> do_epoll_ctl+0x8f/0x100
> __x64_sys_epoll_ctl+0x6f/0xa0
> do_syscall_64+0xdc/0x520
> ? srso_alias_return_thunk+0x5/0xfbef5
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
> RIP: 0033:0x42034e
> Code: 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 e9 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
> RSP: 002b:00007a657ff3c198 EFLAGS: 00000202 ORIG_RAX: 00000000000000e9
> RAX: ffffffffffffffda RBX: 00007a657ff3ccdc RCX: 000000000042034e
> RDX: 0000000000000003 RSI: 0000000000000001 RDI: 0000000000000004
> RBP: 00007a657ff3c2f0 R08: 0000000000000000 R09: 00007a657ff3c6c0
> R10: 00007a657ff3c1a4 R11: 0000000000000202 R12: 00007a657ff3c6c0
> R13: ffffffffffffffb8 R14: 000000000000000d R15: 00007fffb7de0210
> </TASK>
> ------------[ cut here ]------------
>
> unlist_file() does not appear as a separate frame because it was inlined
> into clear_tfile_check_list(). This report was obtained with mdelay()
> instrumentation immediately before spin_lock() and spin_unlock() in
> unlist_file() to widen the two race windows.
>
> More importantly, this is a wild unlock. The stale unlock can target
> f_lock of a different live file and invalidate mutual exclusion for
> state protected by that lock. Turning this into a reliable exploit
> would require precise scheduling and same-slot reuse and is likely
> difficult, but the primitive is potentially exploitable.
Reported-by: Qi Tang <tpluszz77@gmail.com>
Reported-by: Junxi Qian <qjx1298677004@gmail.com>
Fixes: 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU")
Cc: stable@vger.kernel.org
Signed-off-by: Guidong Han <2045gemini@gmail.com>
Link: https://patch.msgid.link/20260718104406.27897-1-2045gemini@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-18 18:44:06 +08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: pin files while checking reverse paths
Commit 319c15174757 ("epoll: take epitem list out of struct file")
intentionally removed temporary file references from the reverse path
check list. At the time, both epitems and their files were freed after
an RCU grace period, so unlist_file() could obtain file->f_lock through
an epitem while clear_tfile_check_list() held rcu_read_lock().
Commit 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU") made
struct file SLAB_TYPESAFE_BY_RCU and removed its RCU-delayed freeing.
RCU still protects the epitem, but no longer keeps the referenced file
from being freed and reused. A concurrent close can therefore make
unlist_file() lock or unlock f_lock in a recycled file object.
This violates the documented SLAB_TYPESAFE_BY_RCU rule requiring a
reference before acquiring an object's lock. The race was reproduced,
causing a wild unlock of f_lock in a recycled file and breaking its
mutual exclusion.
Add ->file to epitems_head to remember the pinned file independently of
->epitems. A concurrent EPOLL_CTL_DEL can empty ->epitems before the head
is unlisted, leaving no epi->ffd.file from which to drop the reference.
In list_file(), acquire the reference before adding the head to the
check list. The caller either owns a reference or holds the ep->mtx for
the epitem leading to the file. In the latter case, file_ref_get() can
fail after the last reference is dropped, but eventpoll_release_file()
must acquire the same mutex before the file can be freed. The dying leaf
can be skipped because removing links cannot increase the reverse path
count.
In unlist_file(), epnested_mutex excludes another list_file() or
unlist_file(), while head->next prevents a concurrent EPOLL_CTL_DEL from
freeing the head. Save head->file locally, clear it with head->next
under f_lock, and drop the reference after the RCU-protected operation.
Christian Brauner <brauner@kernel.org> quotes:
> SLAB_TYPESAFE_BY_RCU allows a slab slot to be reused while an RCU reader
> still holds its old address. Once that address contains a new live
> struct file, KASAN sees valid, unpoisoned memory and cannot distinguish
> the stale object identity. CONFIG_DEBUG_SPINLOCK exposes the failure
> instead.
>
> The failing interleaving is:
>
> CPU0: nested EPOLL_CTL_ADD CPU1: close/open churn
> ------------------------------------ ---------------------------------
> p = hlist_first_rcu(&head->epitems)
> epi = container_of(p, ...)
> close(victim)
> __fput()
> eventpoll_release_file()
> file_free(victim)
> // the slot is free; f_lock remains
> spin_lock(&epi->ffd.file->f_lock)
> open() reuses the slot as new_file
> spin_lock_init(&new_file->f_lock)
> spin_unlock(&epi->ffd.file->f_lock) // wild unlock of new_file's lock
>
> CONFIG_DEBUG_SPINLOCK reports:
>
> BUG: spinlock already unlocked on CPU#0, poc_unlist/150
> lock: 0xffff8880067fb200, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1
> CPU: 0 UID: 1000 PID: 150 Comm: poc_unlist Not tainted 7.2.0-rc3-dirty #22 PREEMPTLAZY
> Hardware name: QEMU Ubuntu 24.04 PC v2 (i440FX + PIIX, arch_caps fix, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014
> Call Trace:
> <TASK>
> dump_stack_lvl+0x64/0x80
> do_raw_spin_unlock+0x75/0xb0
> _raw_spin_unlock+0xe/0x30
> clear_tfile_check_list+0x88/0xe0
> do_epoll_ctl_file+0x519/0xcf0
> ? __pfx_ep_ptable_queue_proc+0x10/0x10
> do_epoll_ctl+0x8f/0x100
> __x64_sys_epoll_ctl+0x6f/0xa0
> do_syscall_64+0xdc/0x520
> ? srso_alias_return_thunk+0x5/0xfbef5
> entry_SYSCALL_64_after_hwframe+0x76/0x7e
> RIP: 0033:0x42034e
> Code: 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 49 89 ca b8 e9 00 00 00 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
> RSP: 002b:00007a657ff3c198 EFLAGS: 00000202 ORIG_RAX: 00000000000000e9
> RAX: ffffffffffffffda RBX: 00007a657ff3ccdc RCX: 000000000042034e
> RDX: 0000000000000003 RSI: 0000000000000001 RDI: 0000000000000004
> RBP: 00007a657ff3c2f0 R08: 0000000000000000 R09: 00007a657ff3c6c0
> R10: 00007a657ff3c1a4 R11: 0000000000000202 R12: 00007a657ff3c6c0
> R13: ffffffffffffffb8 R14: 000000000000000d R15: 00007fffb7de0210
> </TASK>
> ------------[ cut here ]------------
>
> unlist_file() does not appear as a separate frame because it was inlined
> into clear_tfile_check_list(). This report was obtained with mdelay()
> instrumentation immediately before spin_lock() and spin_unlock() in
> unlist_file() to widen the two race windows.
>
> More importantly, this is a wild unlock. The stale unlock can target
> f_lock of a different live file and invalidate mutual exclusion for
> state protected by that lock. Turning this into a reliable exploit
> would require precise scheduling and same-slot reuse and is likely
> difficult, but the primitive is potentially exploitable.
Reported-by: Qi Tang <tpluszz77@gmail.com>
Reported-by: Junxi Qian <qjx1298677004@gmail.com>
Fixes: 0ede61d8589c ("file: convert to SLAB_TYPESAFE_BY_RCU")
Cc: stable@vger.kernel.org
Signed-off-by: Guidong Han <2045gemini@gmail.com>
Link: https://patch.msgid.link/20260718104406.27897-1-2045gemini@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-18 18:44:06 +08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
proc/sysctl: add shared variables for range check
In the sysctl code the proc_dointvec_minmax() function is often used to
validate the user supplied value between an allowed range. This
function uses the extra1 and extra2 members from struct ctl_table as
minimum and maximum allowed value.
On sysctl handler declaration, in every source file there are some
readonly variables containing just an integer which address is assigned
to the extra1 and extra2 members, so the sysctl range is enforced.
The special values 0, 1 and INT_MAX are very often used as range
boundary, leading duplication of variables like zero=0, one=1,
int_max=INT_MAX in different source files:
$ git grep -E '\.extra[12].*&(zero|one|int_max)' |wc -l
248
Add a const int array containing the most commonly used values, some
macros to refer more easily to the correct array member, and use them
instead of creating a local one for every object file.
This is the bloat-o-meter output comparing the old and new binary
compiled with the default Fedora config:
# scripts/bloat-o-meter -d vmlinux.o.old vmlinux.o
add/remove: 2/2 grow/shrink: 0/2 up/down: 24/-188 (-164)
Data old new delta
sysctl_vals - 12 +12
__kstrtab_sysctl_vals - 12 +12
max 14 10 -4
int_max 16 - -16
one 68 - -68
zero 128 28 -100
Total: Before=20583249, After=20583085, chg -0.00%
[mcroce@redhat.com: tipc: remove two unused variables]
Link: http://lkml.kernel.org/r/20190530091952.4108-1-mcroce@redhat.com
[akpm@linux-foundation.org: fix net/ipv6/sysctl_net_ipv6.c]
[arnd@arndb.de: proc/sysctl: make firmware loader table conditional]
Link: http://lkml.kernel.org/r/20190617130014.1713870-1-arnd@arndb.de
[akpm@linux-foundation.org: fix fs/eventpoll.c]
Link: http://lkml.kernel.org/r/20190430180111.10688-1-mcroce@redhat.com
Signed-off-by: Matteo Croce <mcroce@redhat.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-18 15:58:50 -07:00
|
|
|
|
2011-01-12 17:00:01 -08:00
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
2025-01-28 13:48:37 +01:00
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-12 17:00:01 -08:00
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
2011-01-12 17:00:01 -08:00
|
|
|
|
proc/sysctl: add shared variables for range check
In the sysctl code the proc_dointvec_minmax() function is often used to
validate the user supplied value between an allowed range. This
function uses the extra1 and extra2 members from struct ctl_table as
minimum and maximum allowed value.
On sysctl handler declaration, in every source file there are some
readonly variables containing just an integer which address is assigned
to the extra1 and extra2 members, so the sysctl range is enforced.
The special values 0, 1 and INT_MAX are very often used as range
boundary, leading duplication of variables like zero=0, one=1,
int_max=INT_MAX in different source files:
$ git grep -E '\.extra[12].*&(zero|one|int_max)' |wc -l
248
Add a const int array containing the most commonly used values, some
macros to refer more easily to the correct array member, and use them
instead of creating a local one for every object file.
This is the bloat-o-meter output comparing the old and new binary
compiled with the default Fedora config:
# scripts/bloat-o-meter -d vmlinux.o.old vmlinux.o
add/remove: 2/2 grow/shrink: 0/2 up/down: 24/-188 (-164)
Data old new delta
sysctl_vals - 12 +12
__kstrtab_sysctl_vals - 12 +12
max 14 10 -4
int_max 16 - -16
one 68 - -68
zero 128 28 -100
Total: Before=20583249, After=20583085, chg -0.00%
[mcroce@redhat.com: tipc: remove two unused variables]
Link: http://lkml.kernel.org/r/20190530091952.4108-1-mcroce@redhat.com
[akpm@linux-foundation.org: fix net/ipv6/sysctl_net_ipv6.c]
[arnd@arndb.de: proc/sysctl: make firmware loader table conditional]
Link: http://lkml.kernel.org/r/20190617130014.1713870-1-arnd@arndb.de
[akpm@linux-foundation.org: fix fs/eventpoll.c]
Link: http://lkml.kernel.org/r/20190430180111.10688-1-mcroce@redhat.com
Signed-off-by: Matteo Croce <mcroce@redhat.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Aaron Tomlin <atomlin@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-18 15:58:50 -07:00
|
|
|
|
2011-01-12 17:00:01 -08:00
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
2022-01-21 22:12:09 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-23 00:10:03 -07:00
|
|
|
|
2007-05-15 01:40:52 -07:00
|
|
|
|
2026-05-14 08:07:20 -06:00
|
|
|
|
2005-06-23 00:10:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:47 +02:00
|
|
|
|
|
|
|
|
|
2005-06-23 00:10:03 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2005-06-23 00:10:03 -07:00
|
|
|
|
|
|
|
|
|
2017-06-20 12:06:13 +02:00
|
|
|
|
2012-02-24 20:07:29 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-23 00:10:03 -07:00
|
|
|
|
2017-06-20 12:06:13 +02:00
|
|
|
|
2005-06-23 00:10:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:47 +02:00
|
|
|
|
|
|
|
|
|
2011-03-22 16:34:46 -07:00
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-03-22 16:34:46 -07:00
|
|
|
|
|
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
2024-02-13 06:16:42 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-09 05:02:33 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:42 +00:00
|
|
|
|
|
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:42 +00:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-27 11:34:12 +08:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:43 +00:00
|
|
|
|
2024-02-13 06:16:44 +00:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
2024-02-13 06:16:43 +00:00
|
|
|
|
|
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
2025-02-14 18:17:51 +00:00
|
|
|
|
2025-02-27 11:34:12 +08:00
|
|
|
|
2024-02-13 06:16:44 +00:00
|
|
|
|
2020-12-18 14:01:57 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-09 05:02:34 +00:00
|
|
|
|
|
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
2020-12-18 14:01:57 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:42 +00:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:42 +00:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
2020-12-04 12:36:04 +01:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-14 18:17:51 +00:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:45 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-09 05:02:34 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-14 18:17:51 +00:00
|
|
|
|
2024-11-09 05:02:34 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-14 18:17:51 +00:00
|
|
|
|
2024-11-09 05:02:34 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-21 21:58:19 -07:00
|
|
|
|
|
|
|
|
|
2025-02-27 11:34:12 +08:00
|
|
|
|
2018-08-21 21:58:19 -07:00
|
|
|
|
2020-12-18 14:01:57 -08:00
|
|
|
|
2018-08-21 21:58:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:45 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-09 05:02:34 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-08-21 21:58:19 -07:00
|
|
|
|
|
|
|
|
|
epoll: comment the funky #ifdef
Looking for a bug in -rt, I stumbled across this code here from: commit
2dfa4eeab0fc ("epoll keyed wakeups: teach epoll about hints coming with
the wakeup key"), specifically:
#ifdef CONFIG_DEBUG_LOCK_ALLOC
static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
unsigned long events, int subclass)
{
unsigned long flags;
spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
wake_up_locked_poll(wqueue, events);
spin_unlock_irqrestore(&wqueue->lock, flags);
}
#else
static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
unsigned long events, int subclass)
{
wake_up_poll(wqueue, events);
}
#endif
You change the function of ep_wake_up_nested() depending on whether
CONFIG_DEBUG_LOCK_ALLOC is set or not. This looks awfully suspicious,
and there's no comment to explain why. I initially thought that this
was trying to fool lockdep, and hiding a real bug.
Investigating it, I found the creation of wake_up_nested() (which no
longer exists) but was created for the sole purpose of epoll and its
strange wake ups, as explained in commit 0ccf831cbee9 ("lockdep:
annotate epoll")
Although the commit message says "annotate epoll" the change log is much
better at explaining what is happening than what is in the actual code.
Thus a comment is really necessary here. And to save the time of other
developers from having to go trudging through the git logs trying to
figure out why this code exists.
I took parts of the change log and placed it into a comment above the
affected code. This will make the description of what is happening more
visible to new developers that have to look at this code for the first
time.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-03-12 17:25:22 +08:00
|
|
|
|
|
|
|
|
|
epoll: comment the funky #ifdef
Looking for a bug in -rt, I stumbled across this code here from: commit
2dfa4eeab0fc ("epoll keyed wakeups: teach epoll about hints coming with
the wakeup key"), specifically:
#ifdef CONFIG_DEBUG_LOCK_ALLOC
static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
unsigned long events, int subclass)
{
unsigned long flags;
spin_lock_irqsave_nested(&wqueue->lock, flags, subclass);
wake_up_locked_poll(wqueue, events);
spin_unlock_irqrestore(&wqueue->lock, flags);
}
#else
static inline void ep_wake_up_nested(wait_queue_head_t *wqueue,
unsigned long events, int subclass)
{
wake_up_poll(wqueue, events);
}
#endif
You change the function of ep_wake_up_nested() depending on whether
CONFIG_DEBUG_LOCK_ALLOC is set or not. This looks awfully suspicious,
and there's no comment to explain why. I initially thought that this
was trying to fool lockdep, and hiding a real bug.
Investigating it, I found the creation of wake_up_nested() (which no
longer exists) but was created for the sole purpose of epoll and its
strange wake ups, as explained in commit 0ccf831cbee9 ("lockdep:
annotate epoll")
Although the commit message says "annotate epoll" the change log is much
better at explaining what is happening than what is in the actual code.
Thus a comment is really necessary here. And to save the time of other
developers from having to go trudging through the git logs trying to
figure out why this code exists.
I took parts of the change log and placed it into a comment above the
affected code. This will make the description of what is happening more
visible to new developers that have to look at this code for the first
time.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: David Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-03-31 15:24:22 -07:00
|
|
|
|
|
|
|
|
|
2022-11-20 10:10:53 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-04-06 20:11:23 -07:00
|
|
|
|
2019-12-04 16:52:12 -08:00
|
|
|
|
2020-04-06 20:11:23 -07:00
|
|
|
|
2009-06-17 16:25:58 -07:00
|
|
|
|
2020-04-06 20:11:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-20 10:10:53 -07:00
|
|
|
|
2020-04-06 20:11:23 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
epoll: avoid calling ep_call_nested() from ep_poll_safewake()
ep_poll_safewake() is used to wakeup potentially nested epoll file
descriptors. The function uses ep_call_nested() to prevent entering the
same wake up queue more than once, and to prevent excessively deep
wakeup paths (deeper than EP_MAX_NESTS). However, this is not necessary
since we are already preventing these conditions during EPOLL_CTL_ADD.
This saves extra function calls, and avoids taking a global lock during
the ep_call_nested() calls.
I have, however, left ep_call_nested() for the CONFIG_DEBUG_LOCK_ALLOC
case, since ep_call_nested() keeps track of the nesting level, and this
is required by the call to spin_lock_irqsave_nested(). It would be nice
to remove the ep_call_nested() calls for the CONFIG_DEBUG_LOCK_ALLOC
case as well, however its not clear how to simply pass the nesting level
through multiple wake_up() levels without more surgery. In any case, I
don't think CONFIG_DEBUG_LOCK_ALLOC is generally used for production.
This patch, also apparently fixes a workload at Google that Salman Qazi
reported by completely removing the poll_safewake_ncalls->lock from
wakeup paths.
Link: http://lkml.kernel.org/r/1507920533-8812-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:02 -08:00
|
|
|
|
|
|
|
|
|
2022-11-20 10:10:53 -07:00
|
|
|
|
2023-05-12 00:46:25 +08:00
|
|
|
|
epoll: avoid calling ep_call_nested() from ep_poll_safewake()
ep_poll_safewake() is used to wakeup potentially nested epoll file
descriptors. The function uses ep_call_nested() to prevent entering the
same wake up queue more than once, and to prevent excessively deep
wakeup paths (deeper than EP_MAX_NESTS). However, this is not necessary
since we are already preventing these conditions during EPOLL_CTL_ADD.
This saves extra function calls, and avoids taking a global lock during
the ep_call_nested() calls.
I have, however, left ep_call_nested() for the CONFIG_DEBUG_LOCK_ALLOC
case, since ep_call_nested() keeps track of the nesting level, and this
is required by the call to spin_lock_irqsave_nested(). It would be nice
to remove the ep_call_nested() calls for the CONFIG_DEBUG_LOCK_ALLOC
case as well, however its not clear how to simply pass the nesting level
through multiple wake_up() levels without more surgery. In any case, I
don't think CONFIG_DEBUG_LOCK_ALLOC is generally used for production.
This patch, also apparently fixes a workload at Google that Salman Qazi
reported by completely removing the poll_safewake_ncalls->lock from
wakeup paths.
Link: http://lkml.kernel.org/r/1507920533-8812-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:02 -08:00
|
|
|
|
2022-11-20 10:10:53 -07:00
|
|
|
|
epoll: avoid calling ep_call_nested() from ep_poll_safewake()
ep_poll_safewake() is used to wakeup potentially nested epoll file
descriptors. The function uses ep_call_nested() to prevent entering the
same wake up queue more than once, and to prevent excessively deep
wakeup paths (deeper than EP_MAX_NESTS). However, this is not necessary
since we are already preventing these conditions during EPOLL_CTL_ADD.
This saves extra function calls, and avoids taking a global lock during
the ep_call_nested() calls.
I have, however, left ep_call_nested() for the CONFIG_DEBUG_LOCK_ALLOC
case, since ep_call_nested() keeps track of the nesting level, and this
is required by the call to spin_lock_irqsave_nested(). It would be nice
to remove the ep_call_nested() calls for the CONFIG_DEBUG_LOCK_ALLOC
case as well, however its not clear how to simply pass the nesting level
through multiple wake_up() levels without more surgery. In any case, I
don't think CONFIG_DEBUG_LOCK_ALLOC is generally used for production.
This patch, also apparently fixes a workload at Google that Salman Qazi
reported by completely removing the poll_safewake_ncalls->lock from
wakeup paths.
Link: http://lkml.kernel.org/r/1507920533-8812-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Acked-by: Davidlohr Bueso <dbueso@suse.de>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-02-24 20:07:29 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-01 18:55:33 +02:00
|
|
|
|
2026-04-24 15:46:34 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-01 18:55:33 +02:00
|
|
|
|
|
|
|
|
|
2012-02-24 20:07:29 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2009-03-31 15:24:15 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-02 11:45:57 -04:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2020-09-02 11:45:57 -04:00
|
|
|
|
|
|
|
|
|
2012-02-24 20:07:29 +01:00
|
|
|
|
2009-03-31 15:24:15 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:37 +02:00
|
|
|
|
|
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2026-04-24 15:46:37 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-11-09 05:02:34 +00:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-03-31 15:25:32 +02:00
|
|
|
|
|
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:08 +02:00
|
|
|
|
2026-04-24 15:46:35 +02:00
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:08 +02:00
|
|
|
|
2026-04-24 15:46:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:08 +02:00
|
|
|
|
2026-04-24 15:46:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:08 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:05 +02:00
|
|
|
|
2026-04-23 11:56:10 +02:00
|
|
|
|
2026-04-23 11:56:05 +02:00
|
|
|
|
2026-04-24 00:23:18 +02:00
|
|
|
|
2026-04-23 11:56:06 +02:00
|
|
|
|
2026-04-23 11:56:05 +02:00
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:10 +02:00
|
|
|
|
2026-04-23 11:56:05 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:10 +02:00
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:04 +02:00
|
|
|
|
2026-04-24 15:46:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-25 11:05:16 +02:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-27 11:18:30 -04:00
|
|
|
|
2009-02-06 13:52:43 -07:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
2026-04-23 11:56:05 +02:00
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:12 +02:00
|
|
|
|
2026-04-23 11:56:05 +02:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
2013-11-12 15:10:16 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-21 14:22:05 +03:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-09-07 20:00:00 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:07 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
eventpoll: fix ep_remove struct eventpoll / struct file UAF
ep_remove() (via ep_remove_file()) cleared file->f_ep under
file->f_lock but then kept using @file inside the critical section
(is_file_epoll(), hlist_del_rcu() through the head, spin_unlock).
A concurrent __fput() taking the eventpoll_release() fastpath in
that window observed the transient NULL, skipped
eventpoll_release_file() and ran to f_op->release / file_free().
For the epoll-watches-epoll case, f_op->release is
ep_eventpoll_release() -> ep_clear_and_put() -> ep_free(), which
kfree()s the watched struct eventpoll. Its embedded ->refs
hlist_head is exactly where epi->fllink.pprev points, so the
subsequent hlist_del_rcu()'s "*pprev = next" scribbles into freed
kmalloc-192 memory.
In addition, struct file is SLAB_TYPESAFE_BY_RCU, so the slot
backing @file could be recycled by alloc_empty_file() --
reinitializing f_lock and f_ep -- while ep_remove() is still
nominally inside that lock. The upshot is an attacker-controllable
kmem_cache_free() against the wrong slab cache.
Pin @file via epi_fget() at the top of ep_remove() and gate the
critical section on the pin succeeding. With the pin held @file
cannot reach refcount zero, which holds __fput() off and
transitively keeps the watched struct eventpoll alive across the
hlist_del_rcu() and the f_lock use, closing both UAFs.
If the pin fails @file has already reached refcount zero and its
__fput() is in flight. Because we bailed before clearing f_ep,
that path takes the eventpoll_release() slow path into
eventpoll_release_file() and blocks on ep->mtx until the waiter
side's ep_clear_and_put() drops it. The bailed epi's share of
ep->refcount stays intact, so the trailing ep_refcount_dec_and_test()
in ep_clear_and_put() cannot free the eventpoll out from under
eventpoll_release_file(); the orphaned epi is then cleaned up
there.
A successful pin also proves we are not racing
eventpoll_release_file() on this epi, so drop the now-redundant
re-check of epi->dying under f_lock. The cheap lockless
READ_ONCE(epi->dying) fast-path bailout stays.
Fixes: 58c9b016e128 ("epoll: use refcount to reduce ep_mutex contention")
Reported-by: Jaeyoung Chung <jjy600901@snu.ac.kr>
Link: https://patch.msgid.link/20260423-work-epoll-uaf-v1-6-2470f9eec0f5@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-04-23 11:56:09 +02:00
|
|
|
|
2026-04-23 11:56:06 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: fix ep_remove struct eventpoll / struct file UAF
ep_remove() (via ep_remove_file()) cleared file->f_ep under
file->f_lock but then kept using @file inside the critical section
(is_file_epoll(), hlist_del_rcu() through the head, spin_unlock).
A concurrent __fput() taking the eventpoll_release() fastpath in
that window observed the transient NULL, skipped
eventpoll_release_file() and ran to f_op->release / file_free().
For the epoll-watches-epoll case, f_op->release is
ep_eventpoll_release() -> ep_clear_and_put() -> ep_free(), which
kfree()s the watched struct eventpoll. Its embedded ->refs
hlist_head is exactly where epi->fllink.pprev points, so the
subsequent hlist_del_rcu()'s "*pprev = next" scribbles into freed
kmalloc-192 memory.
In addition, struct file is SLAB_TYPESAFE_BY_RCU, so the slot
backing @file could be recycled by alloc_empty_file() --
reinitializing f_lock and f_ep -- while ep_remove() is still
nominally inside that lock. The upshot is an attacker-controllable
kmem_cache_free() against the wrong slab cache.
Pin @file via epi_fget() at the top of ep_remove() and gate the
critical section on the pin succeeding. With the pin held @file
cannot reach refcount zero, which holds __fput() off and
transitively keeps the watched struct eventpoll alive across the
hlist_del_rcu() and the f_lock use, closing both UAFs.
If the pin fails @file has already reached refcount zero and its
__fput() is in flight. Because we bailed before clearing f_ep,
that path takes the eventpoll_release() slow path into
eventpoll_release_file() and blocks on ep->mtx until the waiter
side's ep_clear_and_put() drops it. The bailed epi's share of
ep->refcount stays intact, so the trailing ep_refcount_dec_and_test()
in ep_clear_and_put() cannot free the eventpoll out from under
eventpoll_release_file(); the orphaned epi is then cleaned up
there.
A successful pin also proves we are not racing
eventpoll_release_file() on this epi, so drop the now-redundant
re-check of epi->dying under f_lock. The cheap lockless
READ_ONCE(epi->dying) fast-path bailout stays.
Fixes: 58c9b016e128 ("epoll: use refcount to reduce ep_mutex contention")
Reported-by: Jaeyoung Chung <jjy600901@snu.ac.kr>
Link: https://patch.msgid.link/20260423-work-epoll-uaf-v1-6-2470f9eec0f5@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-04-23 11:56:09 +02:00
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:13 +02:00
|
|
|
|
|
|
|
|
|
eventpoll: fix ep_remove struct eventpoll / struct file UAF
ep_remove() (via ep_remove_file()) cleared file->f_ep under
file->f_lock but then kept using @file inside the critical section
(is_file_epoll(), hlist_del_rcu() through the head, spin_unlock).
A concurrent __fput() taking the eventpoll_release() fastpath in
that window observed the transient NULL, skipped
eventpoll_release_file() and ran to f_op->release / file_free().
For the epoll-watches-epoll case, f_op->release is
ep_eventpoll_release() -> ep_clear_and_put() -> ep_free(), which
kfree()s the watched struct eventpoll. Its embedded ->refs
hlist_head is exactly where epi->fllink.pprev points, so the
subsequent hlist_del_rcu()'s "*pprev = next" scribbles into freed
kmalloc-192 memory.
In addition, struct file is SLAB_TYPESAFE_BY_RCU, so the slot
backing @file could be recycled by alloc_empty_file() --
reinitializing f_lock and f_ep -- while ep_remove() is still
nominally inside that lock. The upshot is an attacker-controllable
kmem_cache_free() against the wrong slab cache.
Pin @file via epi_fget() at the top of ep_remove() and gate the
critical section on the pin succeeding. With the pin held @file
cannot reach refcount zero, which holds __fput() off and
transitively keeps the watched struct eventpoll alive across the
hlist_del_rcu() and the f_lock use, closing both UAFs.
If the pin fails @file has already reached refcount zero and its
__fput() is in flight. Because we bailed before clearing f_ep,
that path takes the eventpoll_release() slow path into
eventpoll_release_file() and blocks on ep->mtx until the waiter
side's ep_clear_and_put() drops it. The bailed epi's share of
ep->refcount stays intact, so the trailing ep_refcount_dec_and_test()
in ep_clear_and_put() cannot free the eventpoll out from under
eventpoll_release_file(); the orphaned epi is then cleaned up
there.
A successful pin also proves we are not racing
eventpoll_release_file() on this epi, so drop the now-redundant
re-check of epi->dying under f_lock. The cheap lockless
READ_ONCE(epi->dying) fast-path bailout stays.
Fixes: 58c9b016e128 ("epoll: use refcount to reduce ep_mutex contention")
Reported-by: Jaeyoung Chung <jjy600901@snu.ac.kr>
Link: https://patch.msgid.link/20260423-work-epoll-uaf-v1-6-2470f9eec0f5@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-04-23 11:56:09 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:06 +02:00
|
|
|
|
eventpoll: fix ep_remove struct eventpoll / struct file UAF
ep_remove() (via ep_remove_file()) cleared file->f_ep under
file->f_lock but then kept using @file inside the critical section
(is_file_epoll(), hlist_del_rcu() through the head, spin_unlock).
A concurrent __fput() taking the eventpoll_release() fastpath in
that window observed the transient NULL, skipped
eventpoll_release_file() and ran to f_op->release / file_free().
For the epoll-watches-epoll case, f_op->release is
ep_eventpoll_release() -> ep_clear_and_put() -> ep_free(), which
kfree()s the watched struct eventpoll. Its embedded ->refs
hlist_head is exactly where epi->fllink.pprev points, so the
subsequent hlist_del_rcu()'s "*pprev = next" scribbles into freed
kmalloc-192 memory.
In addition, struct file is SLAB_TYPESAFE_BY_RCU, so the slot
backing @file could be recycled by alloc_empty_file() --
reinitializing f_lock and f_ep -- while ep_remove() is still
nominally inside that lock. The upshot is an attacker-controllable
kmem_cache_free() against the wrong slab cache.
Pin @file via epi_fget() at the top of ep_remove() and gate the
critical section on the pin succeeding. With the pin held @file
cannot reach refcount zero, which holds __fput() off and
transitively keeps the watched struct eventpoll alive across the
hlist_del_rcu() and the f_lock use, closing both UAFs.
If the pin fails @file has already reached refcount zero and its
__fput() is in flight. Because we bailed before clearing f_ep,
that path takes the eventpoll_release() slow path into
eventpoll_release_file() and blocks on ep->mtx until the waiter
side's ep_clear_and_put() drops it. The bailed epi's share of
ep->refcount stays intact, so the trailing ep_refcount_dec_and_test()
in ep_clear_and_put() cannot free the eventpoll out from under
eventpoll_release_file(); the orphaned epi is then cleaned up
there.
A successful pin also proves we are not racing
eventpoll_release_file() on this epi, so drop the now-redundant
re-check of epi->dying under f_lock. The cheap lockless
READ_ONCE(epi->dying) fast-path bailout stays.
Fixes: 58c9b016e128 ("epoll: use refcount to reduce ep_mutex contention")
Reported-by: Jaeyoung Chung <jjy600901@snu.ac.kr>
Link: https://patch.msgid.link/20260423-work-epoll-uaf-v1-6-2470f9eec0f5@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-04-23 11:56:09 +02:00
|
|
|
|
2026-04-24 00:23:18 +02:00
|
|
|
|
2026-04-23 11:56:12 +02:00
|
|
|
|
2026-04-24 15:46:37 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:36 +02:00
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:36 +02:00
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-11 14:24:06 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-04-23 11:56:07 +02:00
|
|
|
|
2013-09-11 14:24:06 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-04-30 15:27:40 -07:00
|
|
|
|
2026-04-24 15:46:42 +02:00
|
|
|
|
2026-04-24 15:46:37 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2024-02-13 06:16:45 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:57 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-26 18:48:57 -04:00
|
|
|
|
2013-04-30 15:27:42 -07:00
|
|
|
|
2020-09-26 18:32:48 -04:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
2020-09-26 18:32:48 -04:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
poll: add poll_requested_events() and poll_does_not_wait() functions
In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for. An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested. This is something that can happen
in the video4linux subsystem among others.
Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably. The poll_table_struct does have it: it
has a key field with the event mask. But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.
Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.
This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.
The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself. That way drivers can obtain the mask through a new
poll_requested_events inline.
The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h). In
that case poll_requested_events() returns ~0 (i.e. all events).
Very rarely drivers might want to know whether poll_wait will actually
wait. If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting. This might be useful information in order
to avoid doing expensive work.
A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation. This is now used in sock_poll_wait() in
include/net/sock.h. This was the only place in the kernel that needed
this information.
Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead. In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.
This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events. It's been replaced by a call
to poll_requested_events().
For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.
Any driver accessing the qproc or key fields directly will now fail to compile.
Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument. This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.
There are only three things a driver can do with the wait argument:
1) obtain the key field:
events = wait ? wait->key : ~0;
This will still work although it should be replaced with the new
poll_requested_events() function (which does exactly the same).
This will now even work better, since wait is no longer set to NULL
unnecessarily.
2) use the qproc callback. This could be deadly since qproc can now be
NULL. Renaming qproc should prevent this from happening. There are no
kernel drivers that actually access this callback directly, BTW.
3) test whether wait == NULL to determine whether poll would return without
waiting. This is no longer sufficient as the correct test is now
wait == NULL || wait->_qproc == NULL.
However, the worst that can happen here is a slight performance hit in
the case where wait != NULL and wait->_qproc == NULL. In that case the
driver will assume that poll_wait() will actually add the fd to the set
of waiting file descriptors. Of course, poll_wait() will not do that
since it tests for wait->_qproc. This will not break anything, though.
There is only one place in the whole kernel where this happens
(sock_poll_wait() in include/net/sock.h) and that code will be replaced
by a call to poll_does_not_wait() in the next patch.
Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
actually waiting. The next file descriptor from the set might match the
event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
2020-09-26 18:48:57 -04:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
poll: add poll_requested_events() and poll_does_not_wait() functions
In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for. An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested. This is something that can happen
in the video4linux subsystem among others.
Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably. The poll_table_struct does have it: it
has a key field with the event mask. But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.
Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.
This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.
The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself. That way drivers can obtain the mask through a new
poll_requested_events inline.
The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h). In
that case poll_requested_events() returns ~0 (i.e. all events).
Very rarely drivers might want to know whether poll_wait will actually
wait. If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting. This might be useful information in order
to avoid doing expensive work.
A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation. This is now used in sock_poll_wait() in
include/net/sock.h. This was the only place in the kernel that needed
this information.
Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead. In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.
This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events. It's been replaced by a call
to poll_requested_events().
For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.
Any driver accessing the qproc or key fields directly will now fail to compile.
Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument. This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.
There are only three things a driver can do with the wait argument:
1) obtain the key field:
events = wait ? wait->key : ~0;
This will still work although it should be replaced with the new
poll_requested_events() function (which does exactly the same).
This will now even work better, since wait is no longer set to NULL
unnecessarily.
2) use the qproc callback. This could be deadly since qproc can now be
NULL. Renaming qproc should prevent this from happening. There are no
kernel drivers that actually access this callback directly, BTW.
3) test whether wait == NULL to determine whether poll would return without
waiting. This is no longer sufficient as the correct test is now
wait == NULL || wait->_qproc == NULL.
However, the worst that can happen here is a slight performance hit in
the case where wait != NULL and wait->_qproc == NULL. In that case the
driver will assume that poll_wait() will actually add the fd to the set
of waiting file descriptors. Of course, poll_wait() will not do that
since it tests for wait->_qproc. This will not break anything, though.
There is only one place in the whole kernel where this happens
(sock_poll_wait() in include/net/sock.h) and that code will be replaced
by a call to poll_does_not_wait() in the next patch.
Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
actually waiting. The next file descriptor from the set might match the
event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
2013-04-30 15:27:42 -07:00
|
|
|
|
2020-09-26 18:32:48 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
|
|
|
|
|
2020-09-26 18:48:57 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: remove ep_call_nested() from ep_eventpoll_poll()
The use of ep_call_nested() in ep_eventpoll_poll(), which is the .poll
routine for an epoll fd, is used to prevent excessively deep epoll
nesting, and to prevent circular paths.
However, we are already preventing these conditions during
EPOLL_CTL_ADD. In terms of too deep epoll chains, we do in fact allow
deep nesting of the epoll fds themselves (deeper than EP_MAX_NESTS),
however we don't allow more than EP_MAX_NESTS when an epoll file
descriptor is actually connected to a wakeup source. Thus, we do not
require the use of ep_call_nested(), since ep_eventpoll_poll(), which is
called via ep_scan_ready_list() only continues nesting if there are
events available.
Since ep_call_nested() is implemented using a global lock, applications
that make use of nested epoll can see large performance improvements
with this change.
Davidlohr said:
: Improvements are quite obscene actually, such as for the following
: epoll_wait() benchmark with 2 level nesting on a 80 core IvyBridge:
:
: ncpus vanilla dirty delta
: 1 2447092 3028315 +23.75%
: 4 231265 2986954 +1191.57%
: 8 121631 2898796 +2283.27%
: 16 59749 2902056 +4757.07%
: 32 26837 2326314 +8568.30%
: 64 12926 1341281 +10276.61%
:
: (http://linux-scalability.org/epoll/epoll-test.c)
Link: http://lkml.kernel.org/r/1509430214-5599-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Cc: Hou Tao <houtao1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:06 -08:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2009-03-31 15:24:11 -07:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
2020-09-26 18:32:48 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
epoll: remove ep_call_nested() from ep_eventpoll_poll()
The use of ep_call_nested() in ep_eventpoll_poll(), which is the .poll
routine for an epoll fd, is used to prevent excessively deep epoll
nesting, and to prevent circular paths.
However, we are already preventing these conditions during
EPOLL_CTL_ADD. In terms of too deep epoll chains, we do in fact allow
deep nesting of the epoll fds themselves (deeper than EP_MAX_NESTS),
however we don't allow more than EP_MAX_NESTS when an epoll file
descriptor is actually connected to a wakeup source. Thus, we do not
require the use of ep_call_nested(), since ep_eventpoll_poll(), which is
called via ep_scan_ready_list() only continues nesting if there are
events available.
Since ep_call_nested() is implemented using a global lock, applications
that make use of nested epoll can see large performance improvements
with this change.
Davidlohr said:
: Improvements are quite obscene actually, such as for the following
: epoll_wait() benchmark with 2 level nesting on a 80 core IvyBridge:
:
: ncpus vanilla dirty delta
: 1 2447092 3028315 +23.75%
: 4 231265 2986954 +1191.57%
: 8 121631 2898796 +2283.27%
: 16 59749 2902056 +4757.07%
: 32 26837 2326314 +8568.30%
: 64 12926 1341281 +10276.61%
:
: (http://linux-scalability.org/epoll/epoll-test.c)
Link: http://lkml.kernel.org/r/1509430214-5599-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Cc: Hou Tao <houtao1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-02-01 15:24:58 -05:00
|
|
|
|
2017-11-28 19:43:33 -05:00
|
|
|
|
epoll: remove ep_call_nested() from ep_eventpoll_poll()
The use of ep_call_nested() in ep_eventpoll_poll(), which is the .poll
routine for an epoll fd, is used to prevent excessively deep epoll
nesting, and to prevent circular paths.
However, we are already preventing these conditions during
EPOLL_CTL_ADD. In terms of too deep epoll chains, we do in fact allow
deep nesting of the epoll fds themselves (deeper than EP_MAX_NESTS),
however we don't allow more than EP_MAX_NESTS when an epoll file
descriptor is actually connected to a wakeup source. Thus, we do not
require the use of ep_call_nested(), since ep_eventpoll_poll(), which is
called via ep_scan_ready_list() only continues nesting if there are
events available.
Since ep_call_nested() is implemented using a global lock, applications
that make use of nested epoll can see large performance improvements
with this change.
Davidlohr said:
: Improvements are quite obscene actually, such as for the following
: epoll_wait() benchmark with 2 level nesting on a 80 core IvyBridge:
:
: ncpus vanilla dirty delta
: 1 2447092 3028315 +23.75%
: 4 231265 2986954 +1191.57%
: 8 121631 2898796 +2283.27%
: 16 59749 2902056 +4757.07%
: 32 26837 2326314 +8568.30%
: 64 12926 1341281 +10276.61%
:
: (http://linux-scalability.org/epoll/epoll-test.c)
Link: http://lkml.kernel.org/r/1509430214-5599-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Cc: Hou Tao <houtao1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:06 -08:00
|
|
|
|
2024-05-03 13:36:09 -07:00
|
|
|
|
2020-08-31 13:16:39 -04:00
|
|
|
|
epoll: remove ep_call_nested() from ep_eventpoll_poll()
The use of ep_call_nested() in ep_eventpoll_poll(), which is the .poll
routine for an epoll fd, is used to prevent excessively deep epoll
nesting, and to prevent circular paths.
However, we are already preventing these conditions during
EPOLL_CTL_ADD. In terms of too deep epoll chains, we do in fact allow
deep nesting of the epoll fds themselves (deeper than EP_MAX_NESTS),
however we don't allow more than EP_MAX_NESTS when an epoll file
descriptor is actually connected to a wakeup source. Thus, we do not
require the use of ep_call_nested(), since ep_eventpoll_poll(), which is
called via ep_scan_ready_list() only continues nesting if there are
events available.
Since ep_call_nested() is implemented using a global lock, applications
that make use of nested epoll can see large performance improvements
with this change.
Davidlohr said:
: Improvements are quite obscene actually, such as for the following
: epoll_wait() benchmark with 2 level nesting on a 80 core IvyBridge:
:
: ncpus vanilla dirty delta
: 1 2447092 3028315 +23.75%
: 4 231265 2986954 +1191.57%
: 8 121631 2898796 +2283.27%
: 16 59749 2902056 +4757.07%
: 32 26837 2326314 +8568.30%
: 64 12926 1341281 +10276.61%
:
: (http://linux-scalability.org/epoll/epoll-test.c)
Link: http://lkml.kernel.org/r/1509430214-5599-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Cc: Hou Tao <houtao1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:06 -08:00
|
|
|
|
2024-05-03 13:36:09 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: remove ep_call_nested() from ep_eventpoll_poll()
The use of ep_call_nested() in ep_eventpoll_poll(), which is the .poll
routine for an epoll fd, is used to prevent excessively deep epoll
nesting, and to prevent circular paths.
However, we are already preventing these conditions during
EPOLL_CTL_ADD. In terms of too deep epoll chains, we do in fact allow
deep nesting of the epoll fds themselves (deeper than EP_MAX_NESTS),
however we don't allow more than EP_MAX_NESTS when an epoll file
descriptor is actually connected to a wakeup source. Thus, we do not
require the use of ep_call_nested(), since ep_eventpoll_poll(), which is
called via ep_scan_ready_list() only continues nesting if there are
events available.
Since ep_call_nested() is implemented using a global lock, applications
that make use of nested epoll can see large performance improvements
with this change.
Davidlohr said:
: Improvements are quite obscene actually, such as for the following
: epoll_wait() benchmark with 2 level nesting on a 80 core IvyBridge:
:
: ncpus vanilla dirty delta
: 1 2447092 3028315 +23.75%
: 4 231265 2986954 +1191.57%
: 8 121631 2898796 +2283.27%
: 16 59749 2902056 +4757.07%
: 32 26837 2326314 +8568.30%
: 64 12926 1341281 +10276.61%
:
: (http://linux-scalability.org/epoll/epoll-test.c)
Link: http://lkml.kernel.org/r/1509430214-5599-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Cc: Hou Tao <houtao1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:06 -08:00
|
|
|
|
2020-09-26 18:32:48 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-03 13:36:09 -07:00
|
|
|
|
2020-08-31 13:16:39 -04:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
|
|
|
|
|
2018-06-28 09:43:44 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2020-09-26 18:32:48 -04:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2012-12-17 16:05:02 -08:00
|
|
|
|
2014-09-29 16:08:25 -07:00
|
|
|
|
2012-12-17 16:05:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2012-12-17 16:05:02 -08:00
|
|
|
|
2017-07-12 14:34:25 -07:00
|
|
|
|
2012-12-17 16:05:02 -08:00
|
|
|
|
2017-07-12 14:34:25 -07:00
|
|
|
|
2026-03-04 10:32:42 -05:00
|
|
|
|
2014-09-29 16:08:25 -07:00
|
|
|
|
2017-07-12 14:34:25 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-09-29 16:08:25 -07:00
|
|
|
|
2012-12-17 16:05:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2012-12-17 16:05:02 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2018-06-28 09:43:44 -07:00
|
|
|
|
llseek: automatically add .llseek fop
All file_operations should get a .llseek operation so we can make
nonseekable_open the default for future file operations without a
.llseek pointer.
The three cases that we can automatically detect are no_llseek, seq_lseek
and default_llseek. For cases where we can we can automatically prove that
the file offset is always ignored, we use noop_llseek, which maintains
the current behavior of not returning an error from a seek.
New drivers should normally not use noop_llseek but instead use no_llseek
and call nonseekable_open at open time. Existing drivers can be converted
to do the same when the maintainer knows for certain that no user code
relies on calling seek on the device file.
The generated code is often incorrectly indented and right now contains
comments that clarify for each added line why a specific variant was
chosen. In the version that gets submitted upstream, the comments will
be gone and I will manually fix the indentation, because there does not
seem to be a way to do that using coccinelle.
Some amount of new code is currently sitting in linux-next that should get
the same modifications, which I will do at the end of the merge window.
Many thanks to Julia Lawall for helping me learn to write a semantic
patch that does all this.
===== begin semantic patch =====
// This adds an llseek= method to all file operations,
// as a preparation for making no_llseek the default.
//
// The rules are
// - use no_llseek explicitly if we do nonseekable_open
// - use seq_lseek for sequential files
// - use default_llseek if we know we access f_pos
// - use noop_llseek if we know we don't access f_pos,
// but we still want to allow users to call lseek
//
@ open1 exists @
identifier nested_open;
@@
nested_open(...)
{
<+...
nonseekable_open(...)
...+>
}
@ open exists@
identifier open_f;
identifier i, f;
identifier open1.nested_open;
@@
int open_f(struct inode *i, struct file *f)
{
<+...
(
nonseekable_open(...)
|
nested_open(...)
)
...+>
}
@ read disable optional_qualifier exists @
identifier read_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
expression E;
identifier func;
@@
ssize_t read_f(struct file *f, char *p, size_t s, loff_t *off)
{
<+...
(
*off = E
|
*off += E
|
func(..., off, ...)
|
E = *off
)
...+>
}
@ read_no_fpos disable optional_qualifier exists @
identifier read_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
@@
ssize_t read_f(struct file *f, char *p, size_t s, loff_t *off)
{
... when != off
}
@ write @
identifier write_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
expression E;
identifier func;
@@
ssize_t write_f(struct file *f, const char *p, size_t s, loff_t *off)
{
<+...
(
*off = E
|
*off += E
|
func(..., off, ...)
|
E = *off
)
...+>
}
@ write_no_fpos @
identifier write_f;
identifier f, p, s, off;
type ssize_t, size_t, loff_t;
@@
ssize_t write_f(struct file *f, const char *p, size_t s, loff_t *off)
{
... when != off
}
@ fops0 @
identifier fops;
@@
struct file_operations fops = {
...
};
@ has_llseek depends on fops0 @
identifier fops0.fops;
identifier llseek_f;
@@
struct file_operations fops = {
...
.llseek = llseek_f,
...
};
@ has_read depends on fops0 @
identifier fops0.fops;
identifier read_f;
@@
struct file_operations fops = {
...
.read = read_f,
...
};
@ has_write depends on fops0 @
identifier fops0.fops;
identifier write_f;
@@
struct file_operations fops = {
...
.write = write_f,
...
};
@ has_open depends on fops0 @
identifier fops0.fops;
identifier open_f;
@@
struct file_operations fops = {
...
.open = open_f,
...
};
// use no_llseek if we call nonseekable_open
////////////////////////////////////////////
@ nonseekable1 depends on !has_llseek && has_open @
identifier fops0.fops;
identifier nso ~= "nonseekable_open";
@@
struct file_operations fops = {
... .open = nso, ...
+.llseek = no_llseek, /* nonseekable */
};
@ nonseekable2 depends on !has_llseek @
identifier fops0.fops;
identifier open.open_f;
@@
struct file_operations fops = {
... .open = open_f, ...
+.llseek = no_llseek, /* open uses nonseekable */
};
// use seq_lseek for sequential files
/////////////////////////////////////
@ seq depends on !has_llseek @
identifier fops0.fops;
identifier sr ~= "seq_read";
@@
struct file_operations fops = {
... .read = sr, ...
+.llseek = seq_lseek, /* we have seq_read */
};
// use default_llseek if there is a readdir
///////////////////////////////////////////
@ fops1 depends on !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier readdir_e;
@@
// any other fop is used that changes pos
struct file_operations fops = {
... .readdir = readdir_e, ...
+.llseek = default_llseek, /* readdir is present */
};
// use default_llseek if at least one of read/write touches f_pos
/////////////////////////////////////////////////////////////////
@ fops2 depends on !fops1 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier read.read_f;
@@
// read fops use offset
struct file_operations fops = {
... .read = read_f, ...
+.llseek = default_llseek, /* read accesses f_pos */
};
@ fops3 depends on !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier write.write_f;
@@
// write fops use offset
struct file_operations fops = {
... .write = write_f, ...
+ .llseek = default_llseek, /* write accesses f_pos */
};
// Use noop_llseek if neither read nor write accesses f_pos
///////////////////////////////////////////////////////////
@ fops4 depends on !fops1 && !fops2 && !fops3 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier read_no_fpos.read_f;
identifier write_no_fpos.write_f;
@@
// write fops use offset
struct file_operations fops = {
...
.write = write_f,
.read = read_f,
...
+.llseek = noop_llseek, /* read and write both use no f_pos */
};
@ depends on has_write && !has_read && !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier write_no_fpos.write_f;
@@
struct file_operations fops = {
... .write = write_f, ...
+.llseek = noop_llseek, /* write uses no f_pos */
};
@ depends on has_read && !has_write && !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
identifier read_no_fpos.read_f;
@@
struct file_operations fops = {
... .read = read_f, ...
+.llseek = noop_llseek, /* read uses no f_pos */
};
@ depends on !has_read && !has_write && !fops1 && !fops2 && !has_llseek && !nonseekable1 && !nonseekable2 && !seq @
identifier fops0.fops;
@@
struct file_operations fops = {
...
+.llseek = noop_llseek, /* no read or write fn */
};
===== End semantic patch =====
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Julia Lawall <julia@diku.dk>
Cc: Christoph Hellwig <hch@infradead.org>
2010-08-15 18:52:59 +02:00
|
|
|
|
2024-02-13 06:16:45 +00:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-27 11:18:30 -04:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:13 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2026-04-23 11:56:06 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 00:23:18 +02:00
|
|
|
|
2026-04-23 11:56:12 +02:00
|
|
|
|
2026-04-23 11:56:06 +02:00
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2026-04-24 15:46:37 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2026-02-21 16:37:42 -08:00
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
2023-07-26 11:21:35 +08:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
eventpoll: Fix epoll_wait() report false negative
ep_events_available() checks for available events by looking at
ep->rdllist and ep_is_scanning(). However, this is done without a lock
and can report false negative if ep_start_scan() or ep_done_scan() are
executed by another task concurrently. For example:
_________________________________________________________________________
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
ep_events_available() |
!list_empty_careful(&ep->rdllist)|
|| ep_is_scanning(ep) |
| ep_enter_scan(ep)
___________________________________|_____________________________________
Another example:
_________________________________________________________________________
ep_events_available() |
|ep_start_scan()
| list_splice_init(&ep->rdllist, ...)
| ep_enter_scan(ep)
!list_empty_careful(&ep->rdllist)|
|ep_done_scan()
| ep_exit_scan(ep)
| list_splice(..., &ep->rdllist)
|| ep_is_scanning(ep) |
___________________________________|_____________________________________
In the above examples, ep_events_available() sees no event despite
events being available. In case epoll_wait() is called with timeout=0,
epoll_wait() will wrongly return "no event" to user.
Introduce a sequence lock to resolve this issue.
Measuring the time consumption of 10 million loop iterations doing
epoll_wait(), the following performance drop is observed:
timeout #event before after diff
0ms 0 3727ms 3974ms +6.6%
0ms 1 8099ms 9134ms +13%
1ms 1 13525ms 13586ms +0.45%
Considering the use case of epoll_wait() (wait for events, do something
with the events, repeat), it should only contribute to a small portion of
user's CPU consumption. Therefore this performance drop is not alarming.
Fixes: c5a282e9635e ("fs/epoll: reduce the scope of wq lock in epoll_wait()")
Suggested-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://patch.msgid.link/4363cd8e34a21d4f0d257be1b33e84dc25030fdf.1780422138.git.namcao@linutronix.de
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-06-02 19:51:46 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
2023-07-26 11:21:35 +08:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-05-14 08:07:20 -06:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-05-14 08:07:17 -06:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2011-01-30 23:42:29 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-06-20 12:06:13 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2017-07-03 20:14:56 -04:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2007-05-15 01:40:41 -07:00
|
|
|
|
2009-03-31 15:24:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-07-03 20:14:56 -04:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2009-03-31 15:24:22 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
2017-07-03 20:14:56 -04:00
|
|
|
|
|
|
|
|
|
2018-02-11 14:34:03 -08:00
|
|
|
|
|
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
|
|
|
|
|
2018-02-11 14:34:03 -08:00
|
|
|
|
|
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-04-26 16:05:48 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-20 10:10:53 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2017-09-01 18:55:33 +02:00
|
|
|
|
|
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
2017-07-03 20:14:56 -04:00
|
|
|
|
2017-09-01 18:55:33 +02:00
|
|
|
|
2026-04-24 15:46:34 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-01 18:55:33 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-02 11:55:09 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2020-09-02 11:55:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-03-31 15:24:11 -07:00
|
|
|
|
2020-09-02 11:55:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-23 00:10:03 -07:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
revert "epoll: support for disabling items, and a self-test app"
Revert commit 03a7beb55b9f ("epoll: support for disabling items, and a
self-test app") pending resolution of the issues identified by Michael
Kerrisk, copied below.
We'll revisit this for 3.8.
: I've taken a look at this patch as it currently stands in 3.7-rc1, and
: done a bit of testing. (By the way, the test program
: tools/testing/selftests/epoll/test_epoll.c does not compile...)
:
: There are one or two places where the behavior seems a little strange,
: so I have a question or two at the end of this mail. But other than
: that, I want to check my understanding so that the interface can be
: correctly documented.
:
: Just to go though my understanding, the problem is the following
: scenario in a multithreaded application:
:
: 1. Multiple threads are performing epoll_wait() operations,
: and maintaining a user-space cache that contains information
: corresponding to each file descriptor being monitored by
: epoll_wait().
:
: 2. At some point, a thread wants to delete (EPOLL_CTL_DEL)
: a file descriptor from the epoll interest list, and
: delete the corresponding record from the user-space cache.
:
: 3. The problem with (2) is that some other thread may have
: previously done an epoll_wait() that retrieved information
: about the fd in question, and may be in the middle of using
: information in the cache that relates to that fd. Thus,
: there is a potential race.
:
: 4. The race can't solved purely in user space, because doing
: so would require applying a mutex across the epoll_wait()
: call, which would of course blow thread concurrency.
:
: Right?
:
: Your solution is the EPOLL_CTL_DISABLE operation. I want to
: confirm my understanding about how to use this flag, since
: the description that has accompanied the patches so far
: has been a bit sparse
:
: 0. In the scenario you're concerned about, deleting a file
: descriptor means (safely) doing the following:
: (a) Deleting the file descriptor from the epoll interest list
: using EPOLL_CTL_DEL
: (b) Deleting the corresponding record in the user-space cache
:
: 1. It's only meaningful to use this EPOLL_CTL_DISABLE in
: conjunction with EPOLLONESHOT.
:
: 2. Using EPOLL_CTL_DISABLE without using EPOLLONESHOT in
: conjunction is a logical error.
:
: 3. The correct way to code multithreaded applications using
: EPOLL_CTL_DISABLE and EPOLLONESHOT is as follows:
:
: a. All EPOLL_CTL_ADD and EPOLL_CTL_MOD operations should
: should EPOLLONESHOT.
:
: b. When a thread wants to delete a file descriptor, it
: should do the following:
:
: [1] Call epoll_ctl(EPOLL_CTL_DISABLE)
: [2] If the return status from epoll_ctl(EPOLL_CTL_DISABLE)
: was zero, then the file descriptor can be safely
: deleted by the thread that made this call.
: [3] If the epoll_ctl(EPOLL_CTL_DISABLE) fails with EBUSY,
: then the descriptor is in use. In this case, the calling
: thread should set a flag in the user-space cache to
: indicate that the thread that is using the descriptor
: should perform the deletion operation.
:
: Is all of the above correct?
:
: The implementation depends on checking on whether
: (events & ~EP_PRIVATE_BITS) == 0
: This replies on the fact that EPOLL_CTL_AD and EPOLL_CTL_MOD always
: set EPOLLHUP and EPOLLERR in the 'events' mask, and EPOLLONESHOT
: causes those flags (as well as all others in ~EP_PRIVATE_BITS) to be
: cleared.
:
: A corollary to the previous paragraph is that using EPOLL_CTL_DISABLE
: is only useful in conjunction with EPOLLONESHOT. However, as things
: stand, one can use EPOLL_CTL_DISABLE on a file descriptor that does
: not have EPOLLONESHOT set in 'events' This results in the following
: (slightly surprising) behavior:
:
: (a) The first call to epoll_ctl(EPOLL_CTL_DISABLE) returns 0
: (the indicator that the file descriptor can be safely deleted).
: (b) The next call to epoll_ctl(EPOLL_CTL_DISABLE) fails with EBUSY.
:
: This doesn't seem particularly useful, and in fact is probably an
: indication that the user made a logic error: they should only be using
: epoll_ctl(EPOLL_CTL_DISABLE) on a file descriptor for which
: EPOLLONESHOT was set in 'events'. If that is correct, then would it
: not make sense to return an error to user space for this case?
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: "Paton J. Lewis" <palewis@adobe.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-11-08 15:53:35 -08:00
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
2012-03-16 16:34:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-25 19:48:56 -04:00
|
|
|
|
2020-08-22 23:08:37 -04:00
|
|
|
|
|
|
|
|
|
2013-11-12 15:10:16 -08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
2020-09-26 15:54:05 -04:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
2020-09-26 15:54:05 -04:00
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-05-18 21:23:14 -07:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
eventpoll: restore EP_UNACTIVE_PTR sentinel for ctx->tfile_check_list
Commit e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into
struct ep_ctl_ctx") moved tfile_check_list from a file-scope global into
the stack-allocated struct ep_ctl_ctx, and in doing so replaced the
EP_UNACTIVE_PTR sentinel with NULL on the grounds that "NULL is the
obvious 'empty' value and zero-init handles it for free", describing the
change as "No functional change". It is not.
epitems_head->next is overloaded with two roles:
1. the "next" pointer that threads a head onto ctx->tfile_check_list;
2. a membership flag: ep_remove_file() uses
!smp_load_acquire(&v->next) to mean "this head is not on any
pending ctx->tfile_check_list and is therefore safe to free".
Before that change the EP_UNACTIVE_PTR sentinel kept the two roles
disjoint: a head on the list always had a non-NULL ->next (another head,
or the sentinel at the tail), so ->next == NULL was equivalent to "never
listed". With the sentinel gone the list is NULL-terminated, so the tail
head's ->next is NULL as well. ep_remove_file()'s gate can no longer
distinguish "never listed" from "listed at the tail", and misfires on
the tail head.
The reader (reverse_path_check_proc) holds epnested_mutex +
rcu_read_lock; the freer (ep_remove_file) holds ep->mtx + file->f_lock.
The two sides share no mutex -- the sentinel was the invariant the gate
relied on to know it could skip the read side. With it gone,
ep_remove_file() frees the tail head while reverse_path_check_proc() is
still walking it, producing the slab-use-after-free read. The syzbot
reproducer hits this within seconds on a multi-CPU VM.
Restore the sentinel: initialize ctx.tfile_check_list to EP_UNACTIVE_PTR
in do_epoll_ctl_file(), and terminate the walk on "!= EP_UNACTIVE_PTR"
in reverse_path_check() and clear_tfile_check_list(). The tail head's
->next becomes the sentinel again rather than NULL, so
ep_remove_file()'s gate regains its exclusivity and stops misfiring on
the tail. ep_remove_file() itself is unchanged.
This restores the invariant the file-scope tfile_check_list relied on
before that change while preserving the ctx packaging it introduced.
Reported-by: syzbot+e70e1b6cba8714543f7c@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=e70e1b6cba8714543f7c
Fixes: e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into struct ep_ctl_ctx")
Suggested-by: Christian Brauner <brauner@kernel.org>
Link: https://lore.kernel.org/all/20260528-rotwild-summt-kuhhandel-7276ef4c33b7@brauner.io/
Signed-off-by: Zhan Wei <zhanwei919@gmail.com>
Link: https://patch.msgid.link/20260529142533.23696-1-zhanwei919@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-05-29 22:25:33 +08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
2020-10-01 14:11:00 -04:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
2020-10-01 14:11:00 -04:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
|
|
|
|
|
2020-09-24 19:41:58 -04:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
|
|
|
|
|
2019-08-06 18:48:46 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-24 19:41:58 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
|
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
|
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
|
|
|
|
|
2013-04-30 15:27:43 -07:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:39 +02:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-25 11:05:16 +02:00
|
|
|
|
|
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
|
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
eventpoll: wrap EP_UNACTIVE_PTR in typed sentinel helpers
ep->ovflist and epi->next both use EP_UNACTIVE_PTR (a cast to
(void *)-1) as a sentinel, with distinct meanings at each site:
ep->ovflist == EP_UNACTIVE_PTR no scan in progress
epi->next == EP_UNACTIVE_PTR epi not on ovflist
Call sites had to know the sentinel's value and, by convention, what
it meant in each context. Hide both behind inline helpers:
ep_is_scanning(ep) predicate for "scan in progress"
ep_enter_scan(ep) WRITE_ONCE flip to NULL (scan start)
ep_exit_scan(ep) WRITE_ONCE flip to sentinel (scan end)
epi_on_ovflist(epi) predicate for "epi is on ovflist"
epi_clear_ovflist(epi) clear epi's ovflist link slot
Convert ep_events_available(), ep_start_scan(), ep_done_scan(),
ep_poll_callback(), and ep_alloc_epitem() to use the wrappers. The
ovflist state-machine transitions are now named, not encoded in
sentinel comparisons, and the top-of-file "Ready-list state machine"
section is the single place that spells out the sentinel's meaning.
ep_alloc() keeps the raw "ep->ovflist = EP_UNACTIVE_PTR" init (no
concurrent access at that point) with an inline "not scanning"
comment, and the tfile_check_list sentinel is left alone -- it will
disappear entirely when the loop-check globals move into a
stack-allocated ep_ctl_ctx in a later commit.
Also rework ep_done_scan()'s for-loop: the combined initializer +
update clause that advanced nepi AND cleared epi->next in one step
was clever but hard to read; splitting the update into two
statements inside the body makes the epi_clear_ovflist() call
visible.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-14-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:45 +02:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-02-01 15:24:58 -05:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2020-09-26 18:15:26 -04:00
|
|
|
|
|
|
|
|
|
2026-05-14 08:07:17 -06:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
2021-09-07 20:00:00 -07:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
2026-04-23 11:56:07 +02:00
|
|
|
|
2020-09-26 18:09:29 -04:00
|
|
|
|
|
|
|
|
|
2020-09-09 22:25:06 -04:00
|
|
|
|
2020-09-27 11:03:32 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:07 +02:00
|
|
|
|
2020-09-27 11:03:32 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-09 22:25:06 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
epoll: remove ep_call_nested() from ep_eventpoll_poll()
The use of ep_call_nested() in ep_eventpoll_poll(), which is the .poll
routine for an epoll fd, is used to prevent excessively deep epoll
nesting, and to prevent circular paths.
However, we are already preventing these conditions during
EPOLL_CTL_ADD. In terms of too deep epoll chains, we do in fact allow
deep nesting of the epoll fds themselves (deeper than EP_MAX_NESTS),
however we don't allow more than EP_MAX_NESTS when an epoll file
descriptor is actually connected to a wakeup source. Thus, we do not
require the use of ep_call_nested(), since ep_eventpoll_poll(), which is
called via ep_scan_ready_list() only continues nesting if there are
events available.
Since ep_call_nested() is implemented using a global lock, applications
that make use of nested epoll can see large performance improvements
with this change.
Davidlohr said:
: Improvements are quite obscene actually, such as for the following
: epoll_wait() benchmark with 2 level nesting on a 80 core IvyBridge:
:
: ncpus vanilla dirty delta
: 1 2447092 3028315 +23.75%
: 4 231265 2986954 +1191.57%
: 8 121631 2898796 +2283.27%
: 16 59749 2902056 +4757.07%
: 32 26837 2326314 +8568.30%
: 64 12926 1341281 +10276.61%
:
: (http://linux-scalability.org/epoll/epoll-test.c)
Link: http://lkml.kernel.org/r/1509430214-5599-1-git-send-email-jbaron@akamai.com
Signed-off-by: Jason Baron <jbaron@akamai.com>
Cc: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Salman Qazi <sqazi@google.com>
Cc: Hou Tao <houtao1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-17 15:29:06 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
2020-09-26 18:09:29 -04:00
|
|
|
|
2026-04-23 11:56:07 +02:00
|
|
|
|
2020-09-26 18:09:29 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
eventpoll: split ep_insert() into alloc + register stages
ep_insert() was 130 lines and mixed four concerns in one body: user
quota charge and epitem allocation, attach-into-file-hlist plus
rbtree insert plus target-ep locking, reverse-path + EPOLLWAKEUP +
poll-queue install with rollback, and ready-list publication.
Factor the first two concerns into named helpers so the body reduces
to orchestration.
ep_alloc_epitem() charges the user's epoll_watches quota, allocates
a fresh epitem, and initializes its fields. On failure it returns
ERR_PTR(-ENOSPC) or ERR_PTR(-ENOMEM); on success the epi is not yet
linked into anything.
ep_register_epitem() installs @epi into @tfile's f_ep hlist and
@ep's rbtree, optionally chains @tfile onto tfile_check_list for the
path check, takes the tep->mtx nested lock for the epoll-watches-
epoll case, and finally takes the ep_get() reference that pairs
with ep_remove()'s ep_put() in ep_insert()'s error paths. On failure
it frees the epi and decrements epoll_watches to match
ep_alloc_epitem().
ep_insert()'s remaining body is the rollback-via-ep_remove() chain
(reverse_path_check, EPOLLWAKEUP source creation, ep_ptable_queue_proc
allocation) and the ready-list / wake publication. Remove a few
stale comments that duplicated function-level documentation or
described obvious code.
No functional change; rollback boundaries unchanged -- every error
path after ep_register_epitem() still calls ep_remove(), preserving
the ep->refcount invariant that keeps ep_remove()'s WARN_ON_ONCE safe.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-10-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:41 +02:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2017-03-24 10:08:30 -07:00
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2019-03-07 16:28:53 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2022-11-20 10:10:53 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:47 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-11-28 19:43:33 -05:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
poll: add poll_requested_events() and poll_does_not_wait() functions
In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for. An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested. This is something that can happen
in the video4linux subsystem among others.
Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably. The poll_table_struct does have it: it
has a key field with the event mask. But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.
Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.
This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.
The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself. That way drivers can obtain the mask through a new
poll_requested_events inline.
The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h). In
that case poll_requested_events() returns ~0 (i.e. all events).
Very rarely drivers might want to know whether poll_wait will actually
wait. If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting. This might be useful information in order
to avoid doing expensive work.
A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation. This is now used in sock_poll_wait() in
include/net/sock.h. This was the only place in the kernel that needed
this information.
Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead. In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.
This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events. It's been replaced by a call
to poll_requested_events().
For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.
Any driver accessing the qproc or key fields directly will now fail to compile.
Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument. This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.
There are only three things a driver can do with the wait argument:
1) obtain the key field:
events = wait ? wait->key : ~0;
This will still work although it should be replaced with the new
poll_requested_events() function (which does exactly the same).
This will now even work better, since wait is no longer set to NULL
unnecessarily.
2) use the qproc callback. This could be deadly since qproc can now be
NULL. Renaming qproc should prevent this from happening. There are no
kernel drivers that actually access this callback directly, BTW.
3) test whether wait == NULL to determine whether poll would return without
waiting. This is no longer sufficient as the correct test is now
wait == NULL || wait->_qproc == NULL.
However, the worst that can happen here is a slight performance hit in
the case where wait != NULL and wait->_qproc == NULL. In that case the
driver will assume that poll_wait() will actually add the fd to the set
of waiting file descriptors. Of course, poll_wait() will not do that
since it tests for wait->_qproc. This will not break anything, though.
There is only one place in the whole kernel where this happens
(sock_poll_wait() in include/net/sock.h) and that code will be replaced
by a call to poll_does_not_wait() in the next patch.
Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
actually waiting. The next file descriptor from the set might match the
event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
poll: add poll_requested_events() and poll_does_not_wait() functions
In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for. An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested. This is something that can happen
in the video4linux subsystem among others.
Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably. The poll_table_struct does have it: it
has a key field with the event mask. But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.
Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.
This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.
The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself. That way drivers can obtain the mask through a new
poll_requested_events inline.
The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h). In
that case poll_requested_events() returns ~0 (i.e. all events).
Very rarely drivers might want to know whether poll_wait will actually
wait. If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting. This might be useful information in order
to avoid doing expensive work.
A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation. This is now used in sock_poll_wait() in
include/net/sock.h. This was the only place in the kernel that needed
this information.
Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead. In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.
This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events. It's been replaced by a call
to poll_requested_events().
For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.
Any driver accessing the qproc or key fields directly will now fail to compile.
Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument. This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.
There are only three things a driver can do with the wait argument:
1) obtain the key field:
events = wait ? wait->key : ~0;
This will still work although it should be replaced with the new
poll_requested_events() function (which does exactly the same).
This will now even work better, since wait is no longer set to NULL
unnecessarily.
2) use the qproc callback. This could be deadly since qproc can now be
NULL. Renaming qproc should prevent this from happening. There are no
kernel drivers that actually access this callback directly, BTW.
3) test whether wait == NULL to determine whether poll would return without
waiting. This is no longer sufficient as the correct test is now
wait == NULL || wait->_qproc == NULL.
However, the worst that can happen here is a slight performance hit in
the case where wait != NULL and wait->_qproc == NULL. In that case the
driver will assume that poll_wait() will actually add the fd to the set
of waiting file descriptors. Of course, poll_wait() will not do that
since it tests for wait->_qproc. This will not break anything, though.
There is only one place in the whole kernel where this happens
(sock_poll_wait() in include/net/sock.h) and that code will be replaced
by a call to poll_does_not_wait() in the next patch.
Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
actually waiting. The next file descriptor from the set might match the
event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2009-03-31 15:24:15 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-01-01 21:20:27 +00:00
|
|
|
|
2009-03-31 15:24:15 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2013-04-30 15:27:39 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-01-01 21:20:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-07 16:28:53 -08:00
|
|
|
|
2013-01-01 21:20:27 +00:00
|
|
|
|
2019-03-07 16:28:53 -08:00
|
|
|
|
2013-01-01 21:20:27 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:47 -07:00
|
|
|
|
2007-05-15 01:40:52 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-11-28 19:56:15 -05:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-31 13:39:52 -04:00
|
|
|
|
|
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
2019-01-03 15:27:05 -08:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
poll: add poll_requested_events() and poll_does_not_wait() functions
In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for. An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested. This is something that can happen
in the video4linux subsystem among others.
Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably. The poll_table_struct does have it: it
has a key field with the event mask. But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.
Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.
This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.
The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself. That way drivers can obtain the mask through a new
poll_requested_events inline.
The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h). In
that case poll_requested_events() returns ~0 (i.e. all events).
Very rarely drivers might want to know whether poll_wait will actually
wait. If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting. This might be useful information in order
to avoid doing expensive work.
A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation. This is now used in sock_poll_wait() in
include/net/sock.h. This was the only place in the kernel that needed
this information.
Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead. In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.
This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events. It's been replaced by a call
to poll_requested_events().
For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.
Any driver accessing the qproc or key fields directly will now fail to compile.
Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument. This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.
There are only three things a driver can do with the wait argument:
1) obtain the key field:
events = wait ? wait->key : ~0;
This will still work although it should be replaced with the new
poll_requested_events() function (which does exactly the same).
This will now even work better, since wait is no longer set to NULL
unnecessarily.
2) use the qproc callback. This could be deadly since qproc can now be
NULL. Renaming qproc should prevent this from happening. There are no
kernel drivers that actually access this callback directly, BTW.
3) test whether wait == NULL to determine whether poll would return without
waiting. This is no longer sufficient as the correct test is now
wait == NULL || wait->_qproc == NULL.
However, the worst that can happen here is a slight performance hit in
the case where wait != NULL and wait->_qproc == NULL. In that case the
driver will assume that poll_wait() will actually add the fd to the set
of waiting file descriptors. Of course, poll_wait() will not do that
since it tests for wait->_qproc. This will not break anything, though.
There is only one place in the whole kernel where this happens
(sock_poll_wait() in include/net/sock.h) and that code will be replaced
by a call to poll_does_not_wait() in the next patch.
Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
actually waiting. The next file descriptor from the set might match the
event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
2020-08-31 13:39:52 -04:00
|
|
|
|
poll: add poll_requested_events() and poll_does_not_wait() functions
In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for. An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested. This is something that can happen
in the video4linux subsystem among others.
Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably. The poll_table_struct does have it: it
has a key field with the event mask. But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.
Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.
This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.
The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself. That way drivers can obtain the mask through a new
poll_requested_events inline.
The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h). In
that case poll_requested_events() returns ~0 (i.e. all events).
Very rarely drivers might want to know whether poll_wait will actually
wait. If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting. This might be useful information in order
to avoid doing expensive work.
A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation. This is now used in sock_poll_wait() in
include/net/sock.h. This was the only place in the kernel that needed
this information.
Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead. In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.
This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events. It's been replaced by a call
to poll_requested_events().
For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.
Any driver accessing the qproc or key fields directly will now fail to compile.
Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument. This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.
There are only three things a driver can do with the wait argument:
1) obtain the key field:
events = wait ? wait->key : ~0;
This will still work although it should be replaced with the new
poll_requested_events() function (which does exactly the same).
This will now even work better, since wait is no longer set to NULL
unnecessarily.
2) use the qproc callback. This could be deadly since qproc can now be
NULL. Renaming qproc should prevent this from happening. There are no
kernel drivers that actually access this callback directly, BTW.
3) test whether wait == NULL to determine whether poll would return without
waiting. This is no longer sufficient as the correct test is now
wait == NULL || wait->_qproc == NULL.
However, the worst that can happen here is a slight performance hit in
the case where wait != NULL and wait->_qproc == NULL. In that case the
driver will assume that poll_wait() will actually add the fd to the set
of waiting file descriptors. Of course, poll_wait() will not do that
since it tests for wait->_qproc. This will not break anything, though.
There is only one place in the whole kernel where this happens
(sock_poll_wait() in include/net/sock.h) and that code will be replaced
by a call to poll_does_not_wait() in the next patch.
Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
actually waiting. The next file descriptor from the set might match the
event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
2020-12-18 14:01:51 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
poll: add poll_requested_events() and poll_does_not_wait() functions
In some cases the poll() implementation in a driver has to do different
things depending on the events the caller wants to poll for. An example
is when a driver needs to start a DMA engine if the caller polls for
POLLIN, but doesn't want to do that if POLLIN is not requested but instead
only POLLOUT or POLLPRI is requested. This is something that can happen
in the video4linux subsystem among others.
Unfortunately, the current epoll/poll/select implementation doesn't
provide that information reliably. The poll_table_struct does have it: it
has a key field with the event mask. But once a poll() call matches one
or more bits of that mask any following poll() calls are passed a NULL
poll_table pointer.
Also, the eventpoll implementation always left the key field at ~0 instead
of using the requested events mask.
This was changed in eventpoll.c so the key field now contains the actual
events that should be polled for as set by the caller.
The solution to the NULL poll_table pointer is to set the qproc field to
NULL in poll_table once poll() matches the events, not the poll_table
pointer itself. That way drivers can obtain the mask through a new
poll_requested_events inline.
The poll_table_struct can still be NULL since some kernel code calls it
internally (netfs_state_poll() in ./drivers/staging/pohmelfs/netfs.h). In
that case poll_requested_events() returns ~0 (i.e. all events).
Very rarely drivers might want to know whether poll_wait will actually
wait. If another earlier file descriptor in the set already matched the
events the caller wanted to wait for, then the kernel will return from the
select() call without waiting. This might be useful information in order
to avoid doing expensive work.
A new helper function poll_does_not_wait() is added that drivers can use
to detect this situation. This is now used in sock_poll_wait() in
include/net/sock.h. This was the only place in the kernel that needed
this information.
Drivers should no longer access any of the poll_table internals, but use
the poll_requested_events() and poll_does_not_wait() access functions
instead. In order to enforce that the poll_table fields are now prepended
with an underscore and a comment was added warning against using them
directly.
This required a change in unix_dgram_poll() in unix/af_unix.c which used
the key field to get the requested events. It's been replaced by a call
to poll_requested_events().
For qproc it was especially important to change its name since the
behavior of that field changes with this patch since this function pointer
can now be NULL when that wasn't possible in the past.
Any driver accessing the qproc or key fields directly will now fail to compile.
Some notes regarding the correctness of this patch: the driver's poll()
function is called with a 'struct poll_table_struct *wait' argument. This
pointer may or may not be NULL, drivers can never rely on it being one or
the other as that depends on whether or not an earlier file descriptor in
the select()'s fdset matched the requested events.
There are only three things a driver can do with the wait argument:
1) obtain the key field:
events = wait ? wait->key : ~0;
This will still work although it should be replaced with the new
poll_requested_events() function (which does exactly the same).
This will now even work better, since wait is no longer set to NULL
unnecessarily.
2) use the qproc callback. This could be deadly since qproc can now be
NULL. Renaming qproc should prevent this from happening. There are no
kernel drivers that actually access this callback directly, BTW.
3) test whether wait == NULL to determine whether poll would return without
waiting. This is no longer sufficient as the correct test is now
wait == NULL || wait->_qproc == NULL.
However, the worst that can happen here is a slight performance hit in
the case where wait != NULL and wait->_qproc == NULL. In that case the
driver will assume that poll_wait() will actually add the fd to the set
of waiting file descriptors. Of course, poll_wait() will not do that
since it tests for wait->_qproc. This will not break anything, though.
There is only one place in the whole kernel where this happens
(sock_poll_wait() in include/net/sock.h) and that code will be replaced
by a call to poll_does_not_wait() in the next patch.
Note that even if wait->_qproc != NULL drivers cannot rely on poll_wait()
actually waiting. The next file descriptor from the set might match the
event mask and thus any possible waits will never happen.
Signed-off-by: Hans Verkuil <hans.verkuil@cisco.com>
Reviewed-by: Jonathan Corbet <corbet@lwn.net>
Reviewed-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Hans de Goede <hdegoede@redhat.com>
Cc: Mauro Carvalho Chehab <mchehab@infradead.org>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-23 15:02:27 -07:00
|
|
|
|
2020-08-31 13:39:52 -04:00
|
|
|
|
2020-08-31 13:41:30 -04:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
2019-01-03 15:27:12 -08:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2019-01-03 15:27:05 -08:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
2020-08-31 13:39:52 -04:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
2020-08-31 13:39:52 -04:00
|
|
|
|
2019-01-03 15:27:05 -08:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
2009-03-31 15:24:11 -07:00
|
|
|
|
eventpoll: extract ep_deliver_event() from ep_send_events()
ep_send_events()'s body covered two concerns: per-item work (PM
wakeup-source bookkeeping, re-poll, copy_to_user, level-trigger
re-queue, EPOLLONESHOT mask clear) and the scan-level accumulator
(maxevents cap, EFAULT preservation, txlist/rdllist splice).
Extract the per-item work as ep_deliver_event(), which returns a
tri-state int:
1 one event was delivered; caller advances the counter,
0 re-poll produced no caller-requested events (item drops
out of the ready list; a future callback will re-queue),
-EFAULT copy_to_user() faulted; item is already re-inserted at
the head of the txlist so ep_done_scan() splices it back
to rdllist.
The per-item comments (PM ordering, the "sole writer to rdllist"
invariant for the LT re-queue, the EFAULT semantics) move into
ep_deliver_event(). ep_send_events() reduces to the fatal-signal
short-circuit, scan bracket, and a short txlist walk that accumulates
the deliveries and preserves the "first error wins" EFAULT contract
(res = delivered only if no event was previously delivered; otherwise
the success count is returned and -EFAULT is reported on the next
call).
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-12-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:43 +02:00
|
|
|
|
2026-04-24 15:46:46 +02:00
|
|
|
|
2020-08-31 13:41:30 -04:00
|
|
|
|
epoll: fix epoll's own poll
Fix a bug inside the epoll's f_op->poll() code, that returns POLLIN even
though there are no actual ready monitored fds. The bug shows up if you
add an epoll fd inside another fd container (poll, select, epoll).
The problem is that callback-based wake ups used by epoll does not carry
(patches will follow, to fix this) any information about the events that
actually happened. So the callback code, since it can't call the file*
->poll() inside the callback, chains the file* into a ready-list.
So, suppose you added an fd with EPOLLOUT only, and some data shows up on
the fd, the file* mapped by the fd will be added into the ready-list (via
wakeup callback). During normal epoll_wait() use, this condition is
sorted out at the time we're actually able to call the file*'s
f_op->poll().
Inside the old epoll's f_op->poll() though, only a quick check
!list_empty(ready-list) was performed, and this could have led to
reporting POLLIN even though no ready fds would show up at a following
epoll_wait(). In order to correctly report the ready status for an epoll
fd, the ready-list must be checked to see if any really available fd+event
would be ready in a following epoll_wait().
Operation (calling f_op->poll() from inside f_op->poll()) that, like wake
ups, must be handled with care because of the fact that epoll fds can be
added to other epoll fds.
Test code:
/*
* epoll_test by Davide Libenzi (Simple code to test epoll internals)
* Copyright (C) 2008 Davide Libenzi
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
* Davide Libenzi <davidel@xmailserver.org>
*
*/
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <signal.h>
#include <limits.h>
#include <poll.h>
#include <sys/epoll.h>
#include <sys/wait.h>
#define EPWAIT_TIMEO (1 * 1000)
#ifndef POLLRDHUP
#define POLLRDHUP 0x2000
#endif
#define EPOLL_MAX_CHAIN 100L
#define EPOLL_TF_LOOP (1 << 0)
struct epoll_test_cfg {
long size;
long flags;
};
static int xepoll_create(int n) {
int epfd;
if ((epfd = epoll_create(n)) == -1) {
perror("epoll_create");
exit(2);
}
return epfd;
}
static void xepoll_ctl(int epfd, int cmd, int fd, struct epoll_event *evt) {
if (epoll_ctl(epfd, cmd, fd, evt) < 0) {
perror("epoll_ctl");
exit(3);
}
}
static void xpipe(int *fds) {
if (pipe(fds)) {
perror("pipe");
exit(4);
}
}
static pid_t xfork(void) {
pid_t pid;
if ((pid = fork()) == (pid_t) -1) {
perror("pipe");
exit(5);
}
return pid;
}
static int run_forked_proc(int (*proc)(void *), void *data) {
int status;
pid_t pid;
if ((pid = xfork()) == 0)
exit((*proc)(data));
if (waitpid(pid, &status, 0) != pid) {
perror("waitpid");
return -1;
}
return WIFEXITED(status) ? WEXITSTATUS(status): -2;
}
static int check_events(int fd, int timeo) {
struct pollfd pfd;
fprintf(stdout, "Checking events for fd %d\n", fd);
memset(&pfd, 0, sizeof(pfd));
pfd.fd = fd;
pfd.events = POLLIN | POLLOUT;
if (poll(&pfd, 1, timeo) < 0) {
perror("poll()");
return 0;
}
if (pfd.revents & POLLIN)
fprintf(stdout, "\tPOLLIN\n");
if (pfd.revents & POLLOUT)
fprintf(stdout, "\tPOLLOUT\n");
if (pfd.revents & POLLERR)
fprintf(stdout, "\tPOLLERR\n");
if (pfd.revents & POLLHUP)
fprintf(stdout, "\tPOLLHUP\n");
if (pfd.revents & POLLRDHUP)
fprintf(stdout, "\tPOLLRDHUP\n");
return pfd.revents;
}
static int epoll_test_tty(void *data) {
int epfd, ifd = fileno(stdin), res;
struct epoll_event evt;
if (check_events(ifd, 0) != POLLOUT) {
fprintf(stderr, "Something is cooking on STDIN (%d)\n", ifd);
return 1;
}
epfd = xepoll_create(1);
fprintf(stdout, "Created epoll fd (%d)\n", epfd);
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
xepoll_ctl(epfd, EPOLL_CTL_ADD, ifd, &evt);
if (check_events(epfd, 0) & POLLIN) {
res = epoll_wait(epfd, &evt, 1, 0);
if (res == 0) {
fprintf(stderr, "Epoll fd (%d) is ready when it shouldn't!\n",
epfd);
return 2;
}
}
return 0;
}
static int epoll_wakeup_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write must come after the poll(2) call inside
* check_events(). This tests the nested wakeup code in
* fs/eventpoll.c:ep_poll_safewake()
* By having the check_events() (hence poll(2)) happens first,
* we have poll wait queue filled up, and the write(2) in the
* child will trigger the wakeup chain.
*/
if ((pid = xfork()) == 0) {
sleep(1);
write(pfds[1], "w", 1);
exit(0);
}
res = check_events(epfd, 2000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
static int epoll_poll_chain(void *data) {
struct epoll_test_cfg *tcfg = data;
int i, res, epfd, bfd, nfd, pfds[2];
pid_t pid;
struct epoll_event evt;
memset(&evt, 0, sizeof(evt));
evt.events = EPOLLIN;
epfd = bfd = xepoll_create(1);
for (i = 0; i < tcfg->size; i++) {
nfd = xepoll_create(1);
xepoll_ctl(bfd, EPOLL_CTL_ADD, nfd, &evt);
bfd = nfd;
}
xpipe(pfds);
if (tcfg->flags & EPOLL_TF_LOOP)
{
xepoll_ctl(bfd, EPOLL_CTL_ADD, epfd, &evt);
/*
* If we're testing for loop, we want that the wakeup
* triggered by the write to the pipe done in the child
* process, triggers a fake event. So we add the pipe
* read size with EPOLLOUT events. This will trigger
* an addition to the ready-list, but no real events
* will be there. The the epoll kernel code will proceed
* in calling f_op->poll() of the epfd, triggering the
* loop we want to test.
*/
evt.events = EPOLLOUT;
}
xepoll_ctl(bfd, EPOLL_CTL_ADD, pfds[0], &evt);
/*
* The pipe write mush come before the poll(2) call inside
* check_events(). This tests the nested f_op->poll calls code in
* fs/eventpoll.c:ep_eventpoll_poll()
* By having the pipe write(2) happen first, we make the kernel
* epoll code to load the ready lists, and the following poll(2)
* done inside check_events() will test nested poll code in
* ep_eventpoll_poll().
*/
if ((pid = xfork()) == 0) {
write(pfds[1], "w", 1);
exit(0);
}
sleep(1);
res = check_events(epfd, 1000) & POLLIN;
if (waitpid(pid, NULL, 0) != pid) {
perror("waitpid");
return -1;
}
return res;
}
int main(int ac, char **av) {
int error;
struct epoll_test_cfg tcfg;
fprintf(stdout, "\n********** Testing TTY events\n");
error = run_forked_proc(epoll_test_tty, NULL);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short wakeup chain\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing short poll chain\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == POLLIN ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = EPOLL_MAX_CHAIN;
tcfg.flags = 0;
fprintf(stdout, "\n********** Testing long poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy wakeup chain (HOLD ON)\n");
error = run_forked_proc(epoll_wakeup_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
tcfg.size = 3;
tcfg.flags = EPOLL_TF_LOOP;
fprintf(stdout, "\n********** Testing loopy poll chain (HOLD ON)\n");
error = run_forked_proc(epoll_poll_chain, &tcfg);
fprintf(stdout, error == 0 ?
"********** OK\n": "********** FAIL (%d)\n", error);
return 0;
}
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Pavel Pisa <pisa@cmp.felk.cvut.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-31 15:24:10 -07:00
|
|
|
|
2020-08-31 13:39:52 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
2011-02-01 15:52:35 -08:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-02-01 15:52:35 -08:00
|
|
|
|
2016-05-19 17:09:05 -07:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
|
|
|
|
|
2011-02-01 15:52:35 -08:00
|
|
|
|
|
|
|
|
|
2022-06-15 14:24:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-05-30 11:32:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-15 14:24:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-19 10:22:25 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-04-16 18:58:25 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-03-22 16:34:47 -07:00
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
2011-03-22 16:34:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
|
|
|
|
|
2011-03-22 16:34:47 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
2011-03-22 16:34:47 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-04-24 15:46:47 +02:00
|
|
|
|
|
|
|
|
|
timer: convert timer_slack_ns from unsigned long to u64
This patchset introduces a /proc/<pid>/timerslack_ns interface which
would allow controlling processes to be able to set the timerslack value
on other processes in order to save power by avoiding wakeups (Something
Android currently does via out-of-tree patches).
The first patch tries to fix the internal timer_slack_ns usage which was
defined as a long, which limits the slack range to ~4 seconds on 32bit
systems. It converts it to a u64, which provides the same basically
unlimited slack (500 years) on both 32bit and 64bit machines.
The second patch introduces the /proc/<pid>/timerslack_ns interface
which allows the full 64bit slack range for a task to be read or set on
both 32bit and 64bit machines.
With these two patches, on a 32bit machine, after setting the slack on
bash to 10 seconds:
$ time sleep 1
real 0m10.747s
user 0m0.001s
sys 0m0.005s
The first patch is a little ugly, since I had to chase the slack delta
arguments through a number of functions converting them to u64s. Let me
know if it makes sense to break that up more or not.
Other than that things are fairly straightforward.
This patch (of 2):
The timer_slack_ns value in the task struct is currently a unsigned
long. This means that on 32bit applications, the maximum slack is just
over 4 seconds. However, on 64bit machines, its much much larger (~500
years).
This disparity could make application development a little (as well as
the default_slack) to a u64. This means both 32bit and 64bit systems
have the same effective internal slack range.
Now the existing ABI via PR_GET_TIMERSLACK and PR_SET_TIMERSLACK specify
the interface as a unsigned long, so we preserve that limitation on
32bit systems, where SET_TIMERSLACK can only set the slack to a unsigned
long value, and GET_TIMERSLACK will return ULONG_MAX if the slack is
actually larger then what can be stored by an unsigned long.
This patch also modifies hrtimer functions which specified the slack
delta as a unsigned long.
Signed-off-by: John Stultz <john.stultz@linaro.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Oren Laadan <orenl@cellrox.com>
Cc: Ruchi Kandoi <kandoiruchi@google.com>
Cc: Rom Lemarchand <romlem@android.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Android Kernel Team <kernel-team@android.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-17 14:20:51 -07:00
|
|
|
|
2017-06-20 12:06:13 +02:00
|
|
|
|
2010-10-27 15:34:54 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
2010-10-27 15:34:54 -07:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
|
|
|
|
|
2011-03-22 16:34:47 -07:00
|
|
|
|
|
|
|
|
|
2020-12-18 14:02:06 -08:00
|
|
|
|
2011-03-22 16:34:47 -07:00
|
|
|
|
2010-10-27 15:34:54 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2020-12-18 14:02:06 -08:00
|
|
|
|
|
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
2020-12-18 14:02:06 -08:00
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
2020-12-18 14:02:06 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-18 14:02:03 -08:00
|
|
|
|
|
|
|
|
|
2025-02-19 10:22:25 -07:00
|
|
|
|
|
|
|
|
|
2020-12-18 14:02:03 -08:00
|
|
|
|
|
|
|
|
|
2020-12-18 14:02:00 -08:00
|
|
|
|
2020-12-18 14:02:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-27 11:34:12 +08:00
|
|
|
|
2020-12-18 14:02:03 -08:00
|
|
|
|
|
|
|
|
|
2020-12-18 14:02:00 -08:00
|
|
|
|
2020-12-18 14:01:48 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: atomically remove wait entry on wake up
This patch does two things:
- fixes a lost wakeup introduced by commit 339ddb53d373 ("fs/epoll:
remove unnecessary wakeups of nested epoll")
- improves performance for events delivery.
The description of the problem is the following: if N (>1) threads are
waiting on ep->wq for new events and M (>1) events come, it is quite
likely that >1 wakeups hit the same wait queue entry, because there is
quite a big window between __add_wait_queue_exclusive() and the
following __remove_wait_queue() calls in ep_poll() function.
This can lead to lost wakeups, because thread, which was woken up, can
handle not all the events in ->rdllist. (in better words the problem is
described here: https://lkml.org/lkml/2019/10/7/905)
The idea of the current patch is to use init_wait() instead of
init_waitqueue_entry().
Internally init_wait() sets autoremove_wake_function as a callback,
which removes the wait entry atomically (under the wq locks) from the
list, thus the next coming wakeup hits the next wait entry in the wait
queue, thus preventing lost wakeups.
Problem is very well reproduced by the epoll60 test case [1].
Wait entry removal on wakeup has also performance benefits, because
there is no need to take a ep->lock and remove wait entry from the queue
after the successful wakeup. Here is the timing output of the epoll60
test case:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d373):
real 0m6.970s
user 0m49.786s
sys 0m0.113s
After this patch:
real 0m5.220s
user 0m36.879s
sys 0m0.019s
The other testcase is the stress-epoll [2], where one thread consumes
all the events and other threads produce many events:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d373):
threads events/ms run-time ms
8 5427 1474
16 6163 2596
32 6824 4689
64 7060 9064
128 6991 18309
After this patch:
threads events/ms run-time ms
8 5598 1429
16 7073 2262
32 7502 4265
64 7640 8376
128 7634 16767
(number of "events/ms" represents event bandwidth, thus higher is
better; number of "run-time ms" represents overall time spent
doing the benchmark, thus lower is better)
[1] tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c
[2] https://github.com/rouming/test-tools/blob/master/stress-epoll.c
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jason Baron <jbaron@akamai.com>
Cc: Khazhismel Kumykov <khazhy@google.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Heiher <r@hev.cc>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/20200430130326.1368509-2-rpenyaev@suse.de
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-07 18:36:16 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-06-15 14:24:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: atomically remove wait entry on wake up
This patch does two things:
- fixes a lost wakeup introduced by commit 339ddb53d373 ("fs/epoll:
remove unnecessary wakeups of nested epoll")
- improves performance for events delivery.
The description of the problem is the following: if N (>1) threads are
waiting on ep->wq for new events and M (>1) events come, it is quite
likely that >1 wakeups hit the same wait queue entry, because there is
quite a big window between __add_wait_queue_exclusive() and the
following __remove_wait_queue() calls in ep_poll() function.
This can lead to lost wakeups, because thread, which was woken up, can
handle not all the events in ->rdllist. (in better words the problem is
described here: https://lkml.org/lkml/2019/10/7/905)
The idea of the current patch is to use init_wait() instead of
init_waitqueue_entry().
Internally init_wait() sets autoremove_wake_function as a callback,
which removes the wait entry atomically (under the wq locks) from the
list, thus the next coming wakeup hits the next wait entry in the wait
queue, thus preventing lost wakeups.
Problem is very well reproduced by the epoll60 test case [1].
Wait entry removal on wakeup has also performance benefits, because
there is no need to take a ep->lock and remove wait entry from the queue
after the successful wakeup. Here is the timing output of the epoll60
test case:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d373):
real 0m6.970s
user 0m49.786s
sys 0m0.113s
After this patch:
real 0m5.220s
user 0m36.879s
sys 0m0.019s
The other testcase is the stress-epoll [2], where one thread consumes
all the events and other threads produce many events:
With explicit wakeup from ep_scan_ready_list() (the state of the
code prior 339ddb53d373):
threads events/ms run-time ms
8 5427 1474
16 6163 2596
32 6824 4689
64 7060 9064
128 6991 18309
After this patch:
threads events/ms run-time ms
8 5598 1429
16 7073 2262
32 7502 4265
64 7640 8376
128 7634 16767
(number of "events/ms" represents event bandwidth, thus higher is
better; number of "run-time ms" represents overall time spent
doing the benchmark, thus lower is better)
[1] tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c
[2] https://github.com/rouming/test-tools/blob/master/stress-epoll.c
Signed-off-by: Roman Penyaev <rpenyaev@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Jason Baron <jbaron@akamai.com>
Cc: Khazhismel Kumykov <khazhy@google.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Heiher <r@hev.cc>
Cc: <stable@vger.kernel.org>
Link: http://lkml.kernel.org/r/20200430130326.1368509-2-rpenyaev@suse.de
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-05-07 18:36:16 -07:00
|
|
|
|
|
|
|
|
|
2022-06-15 14:24:23 -07:00
|
|
|
|
2020-05-13 17:50:38 -07:00
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-05-13 17:50:38 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
eventpoll: compute timer slack lazily in ep_poll()
ep_poll() computes the timer slack via select_estimate_accuracy() up front,
before checking whether events are already available.
select_estimate_accuracy() reads the clock (ktime_get_ts64()), and the
resulting slack is only consumed by the schedule_hrtimeout_range() call on
the blocking path.
A busy poller such as an L7 proxy event loop calls epoll_wait() at a very
high rate and often finds events already pending, returning via
ep_try_send_events() without ever blocking. In that case the up-front
slack estimation - including its clock read - is pure overhead. read_tsc()
attributable to select_estimate_accuracy() sometimes shows up in perf profiles
of such a workload via the epoll_wait() path.
Move the slack estimation to the point where the thread is actually about
to sleep. The timeout passed to ep_poll() is already an absolute deadline
(ep_timeout_to_timespec()), so deferring the estimate does not change the
wakeup time; taken closer to the sleep it is, if anything, marginally more
accurate. On the common non-blocking path the clock read is skipped
entirely.
Measured on a host running a Meta production workload with the following
bpftrace script:
#!/usr/bin/bpftrace
fentry:__x64_sys_epoll_wait,
fentry:__x64_sys_epoll_pwait { @in[tid] = 1; }
fexit:__x64_sys_epoll_wait,
fexit:__x64_sys_epoll_pwait { delete(@in, tid); }
fentry:select_estimate_accuracy /@in[tid]/ { @sea++; }
fentry:schedule_hrtimeout_range /@in[tid]/ { @shr++; }
interval:s:30 {
printf("sea=%lld shr=%lld wasted=%lld (%d%%)\n",
@sea, @shr, @sea - @shr, (@sea - @shr) * 100 / @sea);
exit();
}
Over a 30s window:
sea=3,587,704 shr=3,003,920 wasted=583,784 (16%)
So ~16% of ep_poll invocations of select_estimate_accuracy have no
consumer.
Signed-off-by: Usama Arif <usama.arif@linux.dev>
Link: https://patch.msgid.link/20260707190238.3478608-1-usama.arif@linux.dev
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-07 12:02:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs/eventpoll: fix endless busy loop after timeout has expired
After commit 0a65bc27bd64 ("eventpoll: Set epoll timeout if it's in
the future"), the following program would immediately enter a busy
loop in the kernel:
```
int main() {
int e = epoll_create1(0);
struct epoll_event event = {.events = EPOLLIN};
epoll_ctl(e, EPOLL_CTL_ADD, 0, &event);
const struct timespec timeout = {.tv_nsec = 1};
epoll_pwait2(e, &event, 1, &timeout, 0);
}
```
This happens because the given (non-zero) timeout of 1 nanosecond
usually expires before ep_poll() is entered and then
ep_schedule_timeout() returns false, but `timed_out` is never set
because the code line that sets it is skipped. This quickly turns
into a soft lockup, RCU stalls and deadlocks, inflicting severe
headaches to the whole system.
When the timeout has expired, we don't need to schedule a hrtimer, but
we should set the `timed_out` variable. Therefore, I suggest moving
the ep_schedule_timeout() check into the `timed_out` expression
instead of skipping it.
brauner: Note that there was an earlier fix by Joe Damato in response to
my bug report in [1].
Fixes: 0a65bc27bd64 ("eventpoll: Set epoll timeout if it's in the future")
Cc: Joe Damato <jdamato@fastly.com>
Cc: stable@vger.kernel.org
Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
Link: https://lore.kernel.org/20250429153419.94723-1-jdamato@fastly.com [1]
Link: https://lore.kernel.org/20250429185827.3564438-1-max.kellermann@ionos.com
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-04-29 20:58:27 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: compute timer slack lazily in ep_poll()
ep_poll() computes the timer slack via select_estimate_accuracy() up front,
before checking whether events are already available.
select_estimate_accuracy() reads the clock (ktime_get_ts64()), and the
resulting slack is only consumed by the schedule_hrtimeout_range() call on
the blocking path.
A busy poller such as an L7 proxy event loop calls epoll_wait() at a very
high rate and often finds events already pending, returning via
ep_try_send_events() without ever blocking. In that case the up-front
slack estimation - including its clock read - is pure overhead. read_tsc()
attributable to select_estimate_accuracy() sometimes shows up in perf profiles
of such a workload via the epoll_wait() path.
Move the slack estimation to the point where the thread is actually about
to sleep. The timeout passed to ep_poll() is already an absolute deadline
(ep_timeout_to_timespec()), so deferring the estimate does not change the
wakeup time; taken closer to the sleep it is, if anything, marginally more
accurate. On the common non-blocking path the clock read is skipped
entirely.
Measured on a host running a Meta production workload with the following
bpftrace script:
#!/usr/bin/bpftrace
fentry:__x64_sys_epoll_wait,
fentry:__x64_sys_epoll_pwait { @in[tid] = 1; }
fexit:__x64_sys_epoll_wait,
fexit:__x64_sys_epoll_pwait { delete(@in, tid); }
fentry:select_estimate_accuracy /@in[tid]/ { @sea++; }
fentry:schedule_hrtimeout_range /@in[tid]/ { @shr++; }
interval:s:30 {
printf("sea=%lld shr=%lld wasted=%lld (%d%%)\n",
@sea, @shr, @sea - @shr, (@sea - @shr) * 100 / @sea);
exit();
}
Over a 30s window:
sea=3,587,704 shr=3,003,920 wasted=583,784 (16%)
So ~16% of ep_poll invocations of select_estimate_accuracy have no
consumer.
Signed-off-by: Usama Arif <usama.arif@linux.dev>
Link: https://patch.msgid.link/20260707190238.3478608-1-usama.arif@linux.dev
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-07 12:02:38 -07:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
fs/epoll: reduce the scope of wq lock in epoll_wait()
This patch aims at reducing ep wq.lock hold times in epoll_wait(2). For
the blocking case, there is no need to constantly take and drop the
spinlock, which is only needed to manipulate the waitqueue.
The call to ep_events_available() is now lockless, and only exposed to
benign races. Here, if false positive (returns available events and
does not see another thread deleting an epi from the list) we call into
send_events and then the list's state is correctly seen. Otoh, if a
false negative and we don't see a list_add_tail(), for example, from irq
callback, then it is rechecked again before blocking, which will see the
correct state.
In order for more accuracy to see concurrent list_del_init(), use the
list_empty_careful() variant -- of course, this won't be safe against
insertions from wakeup.
For the overflow list we obviously need to prevent load/store tearing as
we don't want to see partial values while the ready list is disabled.
[dave@stgolabs.net: forgotten fixlets]
Link: http://lkml.kernel.org/r/20181109155258.jxcr4t2pnz6zqct3@linux-r8p5
Link: http://lkml.kernel.org/r/20181108051006.18751-6-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Suggested-by: Jason Baron <jbaron@akamai.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-03 15:27:15 -08:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:47 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Replace rwlock with spinlock
The ready event list of an epoll object is protected by read-write
semaphore:
- The consumer (waiter) acquires the write lock and takes items.
- the producer (waker) takes the read lock and adds items.
The point of this design is enabling epoll to scale well with large number
of producers, as multiple producers can hold the read lock at the same
time.
Unfortunately, this implementation may cause scheduling priority inversion
problem. Suppose the consumer has higher scheduling priority than the
producer. The consumer needs to acquire the write lock, but may be blocked
by the producer holding the read lock. Since read-write semaphore does not
support priority-boosting for the readers (even with CONFIG_PREEMPT_RT=y),
we have a case of priority inversion: a higher priority consumer is blocked
by a lower priority producer. This problem was reported in [1].
Furthermore, this could also cause stall problem, as described in [2].
Fix this problem by replacing rwlock with spinlock.
This reduces the event bandwidth, as the producers now have to contend with
each other for the spinlock. According to the benchmark from
https://github.com/rouming/test-tools/blob/master/stress-epoll.c:
On 12 x86 CPUs:
Before After Diff
threads events/ms events/ms
8 7162 4956 -31%
16 8733 5383 -38%
32 7968 5572 -30%
64 10652 5739 -46%
128 11236 5931 -47%
On 4 riscv CPUs:
Before After Diff
threads events/ms events/ms
8 2958 2833 -4%
16 3323 3097 -7%
32 3451 3240 -6%
64 3554 3178 -11%
128 3601 3235 -10%
Although the numbers look bad, it should be noted that this benchmark
creates multiple threads who do nothing except constantly generating new
epoll events, thus contention on the spinlock is high. For real workload,
the event rate is likely much lower, and the performance drop is not as
bad.
Using another benchmark (perf bench epoll wait) where spinlock contention
is lower, improvement is even observed on x86:
On 12 x86 CPUs:
Before: Averaged 110279 operations/sec (+- 1.09%), total secs = 8
After: Averaged 114577 operations/sec (+- 2.25%), total secs = 8
On 4 riscv CPUs:
Before: Averaged 175767 operations/sec (+- 0.62%), total secs = 8
After: Averaged 167396 operations/sec (+- 0.23%), total secs = 8
In conclusion, no one is likely to be upset over this change. After all,
spinlock was used originally for years, and the commit which converted to
rwlock didn't mention a real workload, just that the benchmark numbers are
nice.
This patch is not exactly the revert of commit a218cc491420 ("epoll: use
rwlock in order to reduce ep_poll_callback() contention"), because git
revert conflicts in some places which are not obvious on the resolution.
This patch is intended to be backported, therefore go with the obvious
approach:
- Replace rwlock_t with spinlock_t one to one
- Delete list_add_tail_lockless() and chain_epi_lockless(). These were
introduced to allow producers to concurrently add items to the list.
But now that spinlock no longer allows producers to touch the event
list concurrently, these two functions are not necessary anymore.
Fixes: a218cc491420 ("epoll: use rwlock in order to reduce ep_poll_callback() contention")
Signed-off-by: Nam Cao <namcao@linutronix.de>
Link: https://lore.kernel.org/ec92458ea357ec503c737ead0f10b2c6e4c37d47.1752581388.git.namcao@linutronix.de
Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Cc: stable@vger.kernel.org
Reported-by: Frederic Weisbecker <frederic@kernel.org>
Closes: https://lore.kernel.org/linux-rt-users/20210825132754.GA895675@lothringen/ [1]
Reported-by: Valentin Schneider <vschneid@redhat.com>
Closes: https://lore.kernel.org/linux-rt-users/xhsmhttqvnall.mognet@vschneid.remote.csb/ [2]
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-15 14:46:34 +02:00
|
|
|
|
Revert "eventpoll: Fix priority inversion problem"
This reverts commit 8c44dac8add7503c345c0f6c7962e4863b88ba42.
I haven't figured out what the actual bug in this commit is, but I did
spend a lot of time chasing it down and eventually succeeded in
bisecting it down to this.
For some reason, this eventpoll commit ends up causing delays and stuck
user space processes, but it only happens on one of my machines, and
only during early boot or during the flurry of initial activity when
logging in.
I must be triggering some very subtle timing issue, but once I figured
out the behavior pattern that made it reasonably reliable to trigger, it
did bisect right to this, and reverting the commit fixes the problem.
Of course, that was only after I had failed at bisecting it several
times, and had flailed around blaming both the drm people and the
netlink people for the odd problems. The most obvious of which happened
at the time of the first graphical login (the most common symptom being
that some gnome app aborted due to a 30s timeout, often leading to the
whole session then failing if it was some critical component like
gnome-shell or similar).
Acked-by: Nam Cao <namcao@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Valentin Schneider <vschneid@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2025-07-11 17:10:32 -07:00
|
|
|
|
2020-12-18 14:02:03 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
2026-05-18 21:23:14 -07:00
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
2020-09-26 16:50:57 -04:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
2026-02-23 20:59:33 +01:00
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-22 23:13:27 -04:00
|
|
|
|
2020-09-10 08:30:05 -04:00
|
|
|
|
2017-09-08 16:15:18 -07:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
|
|
|
|
|
2020-09-26 16:50:57 -04:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
2026-02-23 20:59:33 +01:00
|
|
|
|
2020-09-26 16:50:57 -04:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
|
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
|
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-07-21 19:09:55 +02:00
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-09-26 16:50:57 -04:00
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
2021-03-01 15:25:51 -07:00
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: Fix semi-unbounded recursion
Ensure that epoll instances can never form a graph deeper than
EP_MAX_NESTS+1 links.
Currently, ep_loop_check_proc() ensures that the graph is loop-free and
does some recursion depth checks, but those recursion depth checks don't
limit the depth of the resulting tree for two reasons:
- They don't look upwards in the tree.
- If there are multiple downwards paths of different lengths, only one of
the paths is actually considered for the depth check since commit
28d82dc1c4ed ("epoll: limit paths").
Essentially, the current recursion depth check in ep_loop_check_proc() just
serves to prevent it from recursing too deeply while checking for loops.
A more thorough check is done in reverse_path_check() after the new graph
edge has already been created; this checks, among other things, that no
paths going upwards from any non-epoll file with a length of more than 5
edges exist. However, this check does not apply to non-epoll files.
As a result, it is possible to recurse to a depth of at least roughly 500,
tested on v6.15. (I am unsure if deeper recursion is possible; and this may
have changed with commit 8c44dac8add7 ("eventpoll: Fix priority inversion
problem").)
To fix it:
1. In ep_loop_check_proc(), note the subtree depth of each visited node,
and use subtree depths for the total depth calculation even when a subtree
has already been visited.
2. Add ep_get_upwards_depth_proc() for similarly determining the maximum
depth of an upwards walk.
3. In ep_loop_check(), use these values to limit the total path length
between epoll nodes to EP_MAX_NESTS edges.
Fixes: 22bacca48a17 ("epoll: prevent creating circular epoll structures")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/20250711-epoll-recursion-fix-v1-1-fb2457c33292@google.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-07-11 18:33:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
eventpoll: restore EP_UNACTIVE_PTR sentinel for ctx->tfile_check_list
Commit e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into
struct ep_ctl_ctx") moved tfile_check_list from a file-scope global into
the stack-allocated struct ep_ctl_ctx, and in doing so replaced the
EP_UNACTIVE_PTR sentinel with NULL on the grounds that "NULL is the
obvious 'empty' value and zero-init handles it for free", describing the
change as "No functional change". It is not.
epitems_head->next is overloaded with two roles:
1. the "next" pointer that threads a head onto ctx->tfile_check_list;
2. a membership flag: ep_remove_file() uses
!smp_load_acquire(&v->next) to mean "this head is not on any
pending ctx->tfile_check_list and is therefore safe to free".
Before that change the EP_UNACTIVE_PTR sentinel kept the two roles
disjoint: a head on the list always had a non-NULL ->next (another head,
or the sentinel at the tail), so ->next == NULL was equivalent to "never
listed". With the sentinel gone the list is NULL-terminated, so the tail
head's ->next is NULL as well. ep_remove_file()'s gate can no longer
distinguish "never listed" from "listed at the tail", and misfires on
the tail head.
The reader (reverse_path_check_proc) holds epnested_mutex +
rcu_read_lock; the freer (ep_remove_file) holds ep->mtx + file->f_lock.
The two sides share no mutex -- the sentinel was the invariant the gate
relied on to know it could skip the read side. With it gone,
ep_remove_file() frees the tail head while reverse_path_check_proc() is
still walking it, producing the slab-use-after-free read. The syzbot
reproducer hits this within seconds on a multi-CPU VM.
Restore the sentinel: initialize ctx.tfile_check_list to EP_UNACTIVE_PTR
in do_epoll_ctl_file(), and terminate the walk on "!= EP_UNACTIVE_PTR"
in reverse_path_check() and clear_tfile_check_list(). The tail head's
->next becomes the sentinel again rather than NULL, so
ep_remove_file()'s gate regains its exclusivity and stops misfiring on
the tail. ep_remove_file() itself is unchanged.
This restores the invariant the file-scope tfile_check_list relied on
before that change while preserving the ctx packaging it introduced.
Reported-by: syzbot+e70e1b6cba8714543f7c@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=e70e1b6cba8714543f7c
Fixes: e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into struct ep_ctl_ctx")
Suggested-by: Christian Brauner <brauner@kernel.org>
Link: https://lore.kernel.org/all/20260528-rotwild-summt-kuhhandel-7276ef4c33b7@brauner.io/
Signed-off-by: Zhan Wei <zhanwei919@gmail.com>
Link: https://patch.msgid.link/20260529142533.23696-1-zhanwei919@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-05-29 22:25:33 +08:00
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
epoll: prevent creating circular epoll structures
In several places, an epoll fd can call another file's ->f_op->poll()
method with ep->mtx held. This is in general unsafe, because that other
file could itself be an epoll fd that contains the original epoll fd.
The code defends against this possibility in its own ->poll() method using
ep_call_nested, but there are several other unsafe calls to ->poll
elsewhere that can be made to deadlock. For example, the following simple
program causes the call in ep_insert recursively call the original fd's
->poll, leading to deadlock:
#include <unistd.h>
#include <sys/epoll.h>
int main(void) {
int e1, e2, p[2];
struct epoll_event evt = {
.events = EPOLLIN
};
e1 = epoll_create(1);
e2 = epoll_create(2);
pipe(p);
epoll_ctl(e2, EPOLL_CTL_ADD, e1, &evt);
epoll_ctl(e1, EPOLL_CTL_ADD, p[0], &evt);
write(p[1], p, sizeof p);
epoll_ctl(e1, EPOLL_CTL_ADD, e2, &evt);
return 0;
}
On insertion, check whether the inserted file is itself a struct epoll,
and if so, do a recursive walk to detect whether inserting this file would
create a loop of epoll structures, which could lead to deadlock.
[nelhage@ksplice.com: Use epmutex to serialize concurrent inserts]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Nelson Elhage <nelhage@ksplice.com>
Reported-by: Nelson Elhage <nelhage@ksplice.com>
Tested-by: Nelson Elhage <nelhage@ksplice.com>
Cc: <stable@kernel.org> [2.6.34+, possibly earlier]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-02-25 14:44:12 -08:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2008-08-12 15:09:01 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2018-03-11 11:34:30 +01:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2025-11-23 17:33:30 +01:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2008-07-23 21:29:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-03-31 15:24:11 -07:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2009-03-31 15:24:12 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2008-07-23 21:29:43 -07:00
|
|
|
|
2009-03-31 15:24:12 -07:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2008-02-23 06:46:49 -05:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2025-11-23 17:33:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: limit paths
The current epoll code can be tickled to run basically indefinitely in
both loop detection path check (on ep_insert()), and in the wakeup paths.
The programs that tickle this behavior set up deeply linked networks of
epoll file descriptors that cause the epoll algorithms to traverse them
indefinitely. A couple of these sample programs have been previously
posted in this thread: https://lkml.org/lkml/2011/2/25/297.
To fix the loop detection path check algorithms, I simply keep track of
the epoll nodes that have been already visited. Thus, the loop detection
becomes proportional to the number of epoll file descriptor and links.
This dramatically decreases the run-time of the loop check algorithm. In
one diabolical case I tried it reduced the run-time from 15 mintues (all
in kernel time) to .3 seconds.
Fixing the wakeup paths could be done at wakeup time in a similar manner
by keeping track of nodes that have already been visited, but the
complexity is harder, since there can be multiple wakeups on different
cpus...Thus, I've opted to limit the number of possible wakeup paths when
the paths are created.
This is accomplished, by noting that the end file descriptor points that
are found during the loop detection pass (from the newly added link), are
actually the sources for wakeup events. I keep a list of these file
descriptors and limit the number and length of these paths that emanate
from these 'source file descriptors'. In the current implemetation I
allow 1000 paths of length 1, 500 of length 2, 100 of length 3, 50 of
length 4 and 10 of length 5. Note that it is sufficient to check the
'source file descriptors' reachable from the newly added link, since no
other 'source file descriptors' will have newly added links. This allows
us to check only the wakeup paths that may have gotten too long, and not
re-check all possible wakeup paths on the system.
In terms of the path limit selection, I think its first worth noting that
the most common case for epoll, is probably the model where you have 1
epoll file descriptor that is monitoring n number of 'source file
descriptors'. In this case, each 'source file descriptor' has a 1 path of
length 1. Thus, I believe that the limits I'm proposing are quite
reasonable and in fact may be too generous. Thus, I'm hoping that the
proposed limits will not prevent any workloads that currently work to
fail.
In terms of locking, I have extended the use of the 'epmutex' to all
epoll_ctl add and remove operations. Currently its only used in a subset
of the add paths. I need to hold the epmutex, so that we can correctly
traverse a coherent graph, to check the number of paths. I believe that
this additional locking is probably ok, since its in the setup/teardown
paths, and doesn't affect the running paths, but it certainly is going to
add some extra overhead. Also, worth noting is that the epmuex was
recently added to the ep_ctl add operations in the initial path loop
detection code using the argument that it was not on a critical path.
Another thing to note here, is the length of epoll chains that is allowed.
Currently, eventpoll.c defines:
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
This basically means that I am limited to a graph depth of 5 (EP_MAX_NESTS
+ 1). However, this limit is currently only enforced during the loop
check detection code, and only when the epoll file descriptors are added
in a certain order. Thus, this limit is currently easily bypassed. The
newly added check for wakeup paths, stricly limits the wakeup paths to a
length of 5, regardless of the order in which ep's are linked together.
Thus, a side-effect of the new code is a more consistent enforcement of
the graph depth.
Thus far, I've tested this, using the sample programs previously
mentioned, which now either return quickly or return -EINVAL. I've also
testing using the piptest.c epoll tester, which showed no difference in
performance. I've also created a number of different epoll networks and
tested that they behave as expectded.
I believe this solves the original diabolical test cases, while still
preserving the sane epoll nesting.
Signed-off-by: Jason Baron <jbaron@redhat.com>
Cc: Nelson Elhage <nelhage@ksplice.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-12 17:17:43 -08:00
|
|
|
|
2025-11-23 17:33:30 +01:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2018-03-11 11:34:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-01-14 14:14:25 +01:00
|
|
|
|
2008-07-23 21:29:27 -07:00
|
|
|
|
2009-05-12 13:19:44 -07:00
|
|
|
|
2008-07-23 21:29:43 -07:00
|
|
|
|
|
|
|
|
|
2018-03-11 11:34:30 +01:00
|
|
|
|
2008-07-23 21:29:27 -07:00
|
|
|
|
|
|
|
|
|
2023-03-10 17:07:54 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:38 +02:00
|
|
|
|
2020-01-08 15:05:37 -07:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:38 +02:00
|
|
|
|
2020-01-08 15:05:37 -07:00
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:38 +02:00
|
|
|
|
2020-01-08 15:05:37 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
|
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-05-14 08:07:20 -06:00
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
eventpoll: restore EP_UNACTIVE_PTR sentinel for ctx->tfile_check_list
Commit e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into
struct ep_ctl_ctx") moved tfile_check_list from a file-scope global into
the stack-allocated struct ep_ctl_ctx, and in doing so replaced the
EP_UNACTIVE_PTR sentinel with NULL on the grounds that "NULL is the
obvious 'empty' value and zero-init handles it for free", describing the
change as "No functional change". It is not.
epitems_head->next is overloaded with two roles:
1. the "next" pointer that threads a head onto ctx->tfile_check_list;
2. a membership flag: ep_remove_file() uses
!smp_load_acquire(&v->next) to mean "this head is not on any
pending ctx->tfile_check_list and is therefore safe to free".
Before that change the EP_UNACTIVE_PTR sentinel kept the two roles
disjoint: a head on the list always had a non-NULL ->next (another head,
or the sentinel at the tail), so ->next == NULL was equivalent to "never
listed". With the sentinel gone the list is NULL-terminated, so the tail
head's ->next is NULL as well. ep_remove_file()'s gate can no longer
distinguish "never listed" from "listed at the tail", and misfires on
the tail head.
The reader (reverse_path_check_proc) holds epnested_mutex +
rcu_read_lock; the freer (ep_remove_file) holds ep->mtx + file->f_lock.
The two sides share no mutex -- the sentinel was the invariant the gate
relied on to know it could skip the read side. With it gone,
ep_remove_file() frees the tail head while reverse_path_check_proc() is
still walking it, producing the slab-use-after-free read. The syzbot
reproducer hits this within seconds on a multi-CPU VM.
Restore the sentinel: initialize ctx.tfile_check_list to EP_UNACTIVE_PTR
in do_epoll_ctl_file(), and terminate the walk on "!= EP_UNACTIVE_PTR"
in reverse_path_check() and clear_tfile_check_list(). The tail head's
->next becomes the sentinel again rather than NULL, so
ep_remove_file()'s gate regains its exclusivity and stops misfiring on
the tail. ep_remove_file() itself is unchanged.
This restores the invariant the file-scope tfile_check_list relied on
before that change while preserving the ctx packaging it introduced.
Reported-by: syzbot+e70e1b6cba8714543f7c@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=e70e1b6cba8714543f7c
Fixes: e09c77d94003 ("eventpoll: hoist CTL_ADD scratch state into struct ep_ctl_ctx")
Suggested-by: Christian Brauner <brauner@kernel.org>
Link: https://lore.kernel.org/all/20260528-rotwild-summt-kuhhandel-7276ef4c33b7@brauner.io/
Signed-off-by: Zhan Wei <zhanwei919@gmail.com>
Link: https://patch.msgid.link/20260529142533.23696-1-zhanwei919@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-05-29 22:25:33 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
2024-07-19 20:17:58 -04:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2014-09-09 14:50:51 -07:00
|
|
|
|
2020-01-08 14:35:13 -07:00
|
|
|
|
2012-05-01 21:33:34 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
2020-01-08 14:35:13 -07:00
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
2020-01-08 14:35:13 -07:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
2016-01-20 14:59:24 -08:00
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2007-05-15 01:40:52 -07:00
|
|
|
|
eventpoll: extract lock dance from do_epoll_ctl() into ep_ctl_lock()
do_epoll_ctl() interleaved three concerns in one body: input
validation, the ep->mtx + epnested_mutex acquisition dance for
EPOLL_CTL_ADD on potentially-nested topologies, and the op dispatch
with final unlock. The middle concern is the error-prone one; the
error_tgt_fput label existed mainly to orchestrate it.
Extract the acquisition as ep_ctl_lock() and the release as
ep_ctl_unlock(). ep_ctl_lock() always takes ep->mtx and, for
EPOLL_CTL_ADD on a topology that can change, additionally runs the
loop / path check under epnested_mutex. The return value is a
ternary:
0 ep->mtx held.
1 ep->mtx AND epnested_mutex held (full-check mode).
-errno failure, no locks held.
The non-negative value doubles as the @full_check argument to
ep_insert() and as the argument to ep_ctl_unlock(), so the caller
neither needs an out-parameter nor a separate boolean:
full_check = ep_ctl_lock(ep, op, epfile, tfile, nonblock);
if (full_check < 0)
return full_check;
...
ep_ctl_unlock(ep, full_check);
ep_ctl_unlock() drops ep->mtx and, if full_check == 1, clears
tfile_check_list, bumps loop_check_gen, and drops epnested_mutex --
mirroring the old error_tgt_fput block.
With that in place do_epoll_ctl()'s preconditions become direct
returns (no locks held, nothing to clean up), the acquisition is a
single call, the op dispatch is unchanged, and the epilogue is a
single ep_ctl_unlock() before return. The error_tgt_fput label goes
away.
The two loop_check_gen bumps (one at the start of the full check,
one after) are preserved inside ep_ctl_lock() / ep_ctl_unlock(),
keeping the invariant that ep->gen stamps left on per-eventpoll
caches never equal loop_check_gen after the check completes.
No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-13-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:44 +02:00
|
|
|
|
|
|
|
|
|
2007-05-15 01:40:52 -07:00
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-01-08 14:35:13 -07:00
|
|
|
|
2026-05-15 17:41:05 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 11:56:07 +02:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
epoll: use refcount to reduce ep_mutex contention
We are observing huge contention on the epmutex during an http
connection/rate test:
83.17% 0.25% nginx [kernel.kallsyms] [k] entry_SYSCALL_64_after_hwframe
[...]
|--66.96%--__fput
|--60.04%--eventpoll_release_file
|--58.41%--__mutex_lock.isra.6
|--56.56%--osq_lock
The application is multi-threaded, creates a new epoll entry for
each incoming connection, and does not delete it before the
connection shutdown - that is, before the connection's fd close().
Many different threads compete frequently for the epmutex lock,
affecting the overall performance.
To reduce the contention this patch introduces explicit reference counting
for the eventpoll struct. Each registered event acquires a reference,
and references are released at ep_remove() time.
The eventpoll struct is released by whoever - among EP file close() and
and the monitored file close() drops its last reference.
Additionally, this introduces a new 'dying' flag to prevent races between
the EP file close() and the monitored file close().
ep_eventpoll_release() marks, under f_lock spinlock, each epitem as dying
before removing it, while EP file close() does not touch dying epitems.
The above is needed as both close operations could run concurrently and
drop the EP reference acquired via the epitem entry. Without the above
flag, the monitored file close() could reach the EP struct via the epitem
list while the epitem is still listed and then try to put it after its
disposal.
An alternative could be avoiding touching the references acquired via
the epitems at EP file close() time, but that could leave the EP struct
alive for potentially unlimited time after EP file close(), with nasty
side effects.
With all the above in place, we can drop the epmutex usage at disposal time.
Overall this produces a significant performance improvement in the
mentioned connection/rate scenario: the mutex operations disappear from
the topmost offenders in the perf report, and the measured connections/rate
grows by ~60%.
To make the change more readable this additionally renames ep_free() to
ep_clear_and_put(), and moves the actual memory cleanup in a separate
ep_free() helper.
Link: https://lkml.kernel.org/r/4a57788dcaf28f5eb4f8dfddcc3a8b172a7357bb.1679504153.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Co-developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Xiumei Mu <xmu@redhiat.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Davidlohr Bueso <dave@stgolabs.net>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Eric Biggers <ebiggers@kernel.org>
Cc: Jacob Keller <jacob.e.keller@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-03-22 17:57:02 +01:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
2020-01-08 14:35:13 -07:00
|
|
|
|
|
|
|
|
|
epoll: restrict EPOLLEXCLUSIVE to POLLIN and POLLOUT
In the current implementation of the EPOLLEXCLUSIVE flag (added for
4.5-rc1), if epoll waiters create different POLL* sets and register them
as exclusive against the same target fd, the current implementation will
stop waking any further waiters once it finds the first idle waiter.
This means that waiters could miss wakeups in certain cases.
For example, when we wake up a pipe for reading we do:
wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM); So if
one epoll set or epfd is added to pipe p with POLLIN and a second set
epfd2 is added to pipe p with POLLRDNORM, only epfd may receive the
wakeup since the current implementation will stop after it finds any
intersection of events with a waiter that is blocked in epoll_wait().
We could potentially address this by requiring all epoll waiters that
are added to p be required to pass the same set of POLL* events. IE the
first EPOLL_CTL_ADD that passes EPOLLEXCLUSIVE establishes the set POLL*
flags to be used by any other epfds that are added as EPOLLEXCLUSIVE.
However, I think it might be somewhat confusing interface as we would
have to reference count the number of users for that set, and so
userspace would have to keep track of that count, or we would need a
more involved interface. It also adds some shared state that we'd have
store somewhere. I don't think anybody will want to bloat
__wait_queue_head for this.
I think what we could do instead, is to simply restrict EPOLLEXCLUSIVE
such that it can only be specified with EPOLLIN and/or EPOLLOUT. So
that way if the wakeup includes 'POLLIN' and not 'POLLOUT', we can stop
once we hit the first idle waiter that specifies the EPOLLIN bit, since
any remaining waiters that only have 'POLLOUT' set wouldn't need to be
woken. Likewise, we can do the same thing if 'POLLOUT' is in the wakeup
bit set and not 'POLLIN'. If both 'POLLOUT' and 'POLLIN' are set in the
wake bit set (there is at least one example of this I saw in fs/pipe.c),
then we just wake the entire exclusive list. Having both 'POLLOUT' and
'POLLIN' both set should not be on any performance critical path, so I
think that's ok (in fs/pipe.c its in pipe_release()). We also continue
to include EPOLLERR and EPOLLHUP by default in any exclusive set. Thus,
the user can specify EPOLLERR and/or EPOLLHUP but is not required to do
so.
Since epoll waiters may be interested in other events as well besides
EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP, these can still be added by
doing a 'dup' call on the target fd and adding that as one normally
would with EPOLL_CTL_ADD. Since I think that the POLLIN and POLLOUT
events are what we are interest in balancing, I think that the 'dup'
thing could perhaps be added to only one of the waiter threads.
However, I think that EPOLLIN, EPOLLOUT, EPOLLERR and EPOLLHUP should be
sufficient for the majority of use-cases.
Since EPOLLEXCLUSIVE is intended to be used with a target fd shared
among multiple epfds, where between 1 and n of the epfds may receive an
event, it does not satisfy the semantics of EPOLLONESHOT where only 1
epfd would get an event. Thus, it is not allowed to be specified in
conjunction with EPOLLEXCLUSIVE.
EPOLL_CTL_MOD is also not allowed if the fd was previously added as
EPOLLEXCLUSIVE. It seems with the limited number of flags to not be as
interesting, but this could be relaxed at some further point.
Signed-off-by: Jason Baron <jbaron@akamai.com>
Tested-by: Madars Vitolins <m@silodev.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Al Viro <viro@ftp.linux.org.uk>
Cc: Eric Wong <normalperson@yhbt.net>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Hagen Paul Pfeifer <hagen@jauu.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-02-05 15:37:04 -08:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-24 15:46:48 +02:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-05-14 08:07:20 -06:00
|
|
|
|
2026-05-14 08:07:19 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2020-01-08 14:35:13 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-01-08 15:05:37 -07:00
|
|
|
|
2020-01-08 14:35:13 -07:00
|
|
|
|
|
|
|
|
|
2025-02-19 10:22:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-19 10:22:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-11 11:34:30 +01:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2025-02-19 10:22:24 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2024-07-19 21:19:02 -04:00
|
|
|
|
|
|
|
|
|
2012-08-28 12:52:22 -04:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2025-02-19 10:22:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-31 14:12:01 -04:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2024-07-19 21:19:02 -04:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
2018-03-11 11:34:30 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-11 11:34:30 +01:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:38 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-07-16 16:29:53 -07:00
|
|
|
|
2018-09-19 21:41:04 -07:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
2020-12-18 14:05:38 -08:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
2019-07-16 16:29:53 -07:00
|
|
|
|
2007-05-10 22:23:23 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:38 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-24 14:52:17 -05:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
2020-12-18 14:05:38 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eventpoll: relocate KCMP helpers near compat syscalls
ep_find_tfd() and get_epoll_tfile_raw_ptr() are only used when
CONFIG_KCMP=y. They implement the lookup side of the kcmp(2)
KCMP_EPOLL_TFD query. The helpers currently live between ep_find()
and ep_poll_callback(), interrupting the run of hot-path code
(callback, wait-queue setup, path check, insert, modify, send_events,
poll) with a feature-gated block.
Move the #ifdef CONFIG_KCMP block next to #ifdef CONFIG_COMPAT, which
is also a peripheral ABI extension. Hot-path code becomes a
contiguous span, and the userspace-adjacent extensions cluster at the
end of the file just before eventpoll_init().
Pure code movement; diff is 44 removed and 44 added, all within one
block. No functional change.
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Link: https://patch.msgid.link/20260424-work-epoll-rework-v1-9-249ed00a20f3@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
2026-04-24 15:46:40 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:38 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-24 14:52:17 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-07-16 16:29:53 -07:00
|
|
|
|
2018-09-19 21:41:04 -07:00
|
|
|
|
|
|
|
|
|
2013-02-24 14:52:17 -05:00
|
|
|
|
2020-12-18 14:05:38 -08:00
|
|
|
|
2020-12-18 14:05:35 -08:00
|
|
|
|
2019-07-16 16:29:53 -07:00
|
|
|
|
2013-02-24 14:52:17 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-12-18 14:05:38 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-24 14:52:17 -05:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-01-29 14:25:26 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
epoll: introduce resource usage limits
It has been thought that the per-user file descriptors limit would also
limit the resources that a normal user can request via the epoll
interface. Vegard Nossum reported a very simple program (a modified
version attached) that can make a normal user to request a pretty large
amount of kernel memory, well within the its maximum number of fds. To
solve such problem, default limits are now imposed, and /proc based
configuration has been introduced. A new directory has been created,
named /proc/sys/fs/epoll/ and inside there, there are two configuration
points:
max_user_instances = Maximum number of devices - per user
max_user_watches = Maximum number of "watched" fds - per user
The current default for "max_user_watches" limits the memory used by epoll
to store "watches", to 1/32 of the amount of the low RAM. As example, a
256MB 32bit machine, will have "max_user_watches" set to roughly 90000.
That should be enough to not break existing heavy epoll users. The
default value for "max_user_instances" is set to 128, that should be
enough too.
This also changes the userspace, because a new error code can now come out
from EPOLL_CTL_ADD (-ENOSPC). The EMFILE from epoll_create() was already
listed, so that should be ok.
[akpm@linux-foundation.org: use get_current_user()]
Signed-off-by: Davide Libenzi <davidel@xmailserver.org>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: <stable@kernel.org>
Cc: Cyrill Gorcunov <gorcunov@gmail.com>
Reported-by: Vegard Nossum <vegardno@ifi.uio.no>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-12-01 13:13:55 -08:00
|
|
|
|
2011-01-12 17:00:01 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-04-30 15:27:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2017-11-17 15:28:59 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-17 15:28:59 -08:00
|
|
|
|
2022-01-21 22:12:09 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
epoll: take epitem list out of struct file
Move the head of epitem list out of struct file; for epoll ones it's
moved into struct eventpoll (->refs there), for non-epoll - into
the new object (struct epitem_head). In place of ->f_ep_links we
leave a pointer to the list head (->f_ep).
->f_ep is protected by ->f_lock and it's zeroed as soon as the list
of epitems becomes empty (that can happen only in ep_remove() by
now).
The list of files for reverse path check is *not* going through
struct file now - it's a single-linked list going through epitem_head
instances. It's terminated by ERR_PTR(-1) (== EP_UNACTIVE_POINTER),
so the elements of list can be distinguished by head->next != NULL.
epitem_head instances are allocated at ep_insert() time (by
attach_epitem()) and freed either by ep_remove() (if it empties
the set of epitems *and* epitem_head does not belong to the
reverse path check list) or by clear_tfile_check_list() when
the list is emptied (if the set of epitems is empty by that
point). Allocations are done from a separate slab - minimal kmalloc()
size is too large on some architectures.
As the result, we trim struct file _and_ get rid of the games with
temporary file references.
Locking and barriers are interesting (aren't they always); see unlist_file()
and ep_remove() for details. The non-obvious part is that ep_remove() needs
to decide if it will be the one to free the damn thing *before* actually
storing NULL to head->epitems.first - that's what smp_load_acquire is for
in there. unlist_file() lockless path is safe, since we hit it only if
we observe NULL in head->epitems.first and whoever had done that store is
guaranteed to have observed non-NULL in head->next. IOW, their last access
had been the store of NULL into ->epitems.first and we can safely free
the sucker. OTOH, we are under rcu_read_lock() and both epitem and
epitem->file have their freeing RCU-delayed. So if we see non-NULL
->epitems.first, we can grab ->f_lock (all epitems in there share the
same struct file) and safely recheck the emptiness of ->epitems; again,
->next is still non-NULL, so ep_remove() couldn't have freed head yet.
->f_lock serializes us wrt ep_remove(); the rest is trivial.
Note that once head->epitems becomes NULL, nothing can get inserted into
it - the only remaining reference to head after that point is from the
reverse path check list.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-10-01 20:45:51 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2007-05-10 22:23:22 -07:00
|
|
|
|
