UMOUNT_CONNECTED as implemented allows for the creation of reference
count cycles. Here's a simple example
mkdir /x; mkfifo /ready /go
unshare -m sh -c 'mount -t tmpfs tmpfs /x
truncate -s 8M /x/img; mkfs.ext4 -q /x/img
dev=$(losetup -f --show /x/img)
mkdir /x/mp; mount $dev /x/mp
echo $dev > /ready; read r < /go' &
read dev < /ready
rmdir /x
echo > /go
wait
losetup -d $dev
losetup -a
Take a directory /x on the host, create a new mount namespace, mount a
tmpfs on /x, use a file on that tmpfs as the backing file for a loop
device, mount that loop device on that tmpfs. Now rmdir /x on the host.
This will lazily unmount the mount on top of /x in the container with
UMOUNT_CONNECTED. Once the namespace exits nothing references the mount
anymore. Now the tmpfs is pinned by the backing file of the loop device
and the loop mount is owned by the tmpfs superblock.
Fun fact, such cycles can be formed by at least the following
subsystems and I have added reproducers for all of them:
(1) a loop mount P from an image on a tmpfs next to it, so that P's
death shows as the loop device giving up its backing file
(2) autofs with a FIFO on P as its pipe, zram with a device node on P as
its writeback device, both on a minix image since vfat has neither
(3) ecryptfs with its lower directory on P, under a passphrase token
added to the session keyring
(4) binfmt_misc in a new user namespace with an 'F' interpreter on P
(5) a fuse server that answers FUSE_INIT with passthrough on and
registers a file on P as a backing file
(6) zloop with its zone files in a directory on P
(7) a mass storage gadget on the dummy UDC with its LUN file on P,
mounted from the SCSI disk the gadget shows up as
(8) md with a RAID1 of one loop device and its bitmap file on P, which
skips while SET_BITMAP_FILE has no way to succeed
(9) rmdir of P's mountpoint from the parent, then the child exits, then
the device must be free and LOOP_CLR_FD must release the file
The underlying mechanism is UMOUNT_CONNECTED (MNT_LOCKED falls into the
same class). With UMOUNT_CONNECTED an unmounted mount stays attached to
its parent. This is used to protect revealing the underlying mount and
is a non-negotiable security mechanism. So now the parent owns that
mount and is put on the parent's final mntput(). That moves it to
mnt_stuck_children and ultimately it's cleaned up by cleanup_mnt().
The fact that ownership of the child mount gets transferred to the
parent turns every reference from a child's superblock back to one of
its ancestors into a cycle.
Don't keep the child attached at all. What the parent needs is that a
lookup at the child's mountpoint keeps finding some mount, not the child
itself and it's not a guarantee we have given really.
When an unmounted mount would have stayed attached to its unmounted
parent disconnect it like every other unmounted mount and leave a
marker behind.
A lookup on the parent that misses the mount hash and hits a marker
finds knullfs. Either a file or a directory. Nothing leads from a marker
to any other mount.
The marker is owned by the parent and dropped by the parent's final
mntput() or by __detach_mounts() when the mountpoint is deleted from
under it. It is allocated together with the mount.
With that every unmounted mount is a root and holds only its own
reference which namespace_unlock() drops. No unmounted mount owns
another one. A superblock that pins an ancestor can't form a cycle.
It has a visible change. Mounts left connected (rmdir etc.) used to stay
traversable through the parent for as long as something held the parent.
Now it is detached with the umount. It lives as long as something
references it but it isn't reachable through the parent anymore and ".."
inside it leads nowhere which is the same as for every other lazily
unmounted mount.
Link: https://gist.github.com/mvo5/63ef46482349f3b1c3957d463a0c9c6f
Link: https://patch.msgid.link/20261002-work-mount-cover-v1-2-232a8f52b43c@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Add nullfs_new_file() to allocate an empty immutable regular file on a
nullfs instance as a dentry of its own. It is never hashed under the
root and so can't be found by lookup. Reads return nothing, changes are
refused, file locks, leases and delegations are refused as.
Link: https://patch.msgid.link/20261002-work-mount-cover-v1-1-232a8f52b43c@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
The private nullfs instance that kernel threads are confined to is only
ever reachable through init_task's root and pwd. The following patches
point mounts at it as well, so keep it in a global.
No functional changes.
Link: https://patch.msgid.link/20261002-work-mount-fixes-4-v1-14-dd44b89d44ce@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
mnt_notify_add() puts a mount on notify_list. It doesn't check whether
the mount is on the list already. That's fine as long as every mount is
queued at most once per namespace_sem hold. It isn't.
propagate_umount() moves a surviving overmount off a stack of mounts
that are going away and queues it as moved. shrink_submounts() and
mark_mounts_for_expiry() call umount_tree() for several mounts under a
single namespace_sem hold. So the second umount_tree() can take down
exactly the mount the first one reparented, reparent it once more and so
end up queueing it a second time.
The second list_add_tail() cuts the mounts queued in between out of the
list while the head still points at the last of them. notify_mnt_list()
then only visits that one mount and the others are freed after the grace
period with notify_list still pointing at them. Every later mount
operation on the host walks freed memory:
BUG: KASAN: slab-use-after-free in __list_add_valid_or_report
Read of size 8 at addr ffff8881003d57e0 by task notify_dq/162
__list_add_valid_or_report+0x15c/0x1a0
mnt_add_to_ns+0x322/0x890
attach_recursive_mnt.isra.0+0xf2a/0x1a00
path_mount+0x139c/0x1d40
This needs a mount with MNT_SHRINKABLE to bind from. That's what
finish_automount() creates (submounts of NFS, AFS and CIFS and the
tracefs mount below debugfs). Bind mounts inherit the flag. A mount
namespace with a FAN_MARK_MNTNS mark does the rest and root in a user
namespace can have both.
Initialize to_notify in alloc_vfsmnt() and leave a mount alone that is
queued already. mnt_notify() looks at the state the mount has when the
list is drained so a single entry per mount is enough. The move event
for a mount that is unmounted under the same hold is lost. That's fine.
The detach is what matters.
Fixes: bf630c4016 ("vfs: add notifications for mount attach and detach")
Cc: stable@vger.kernel.org # v6.15+
Link: https://patch.msgid.link/20260930-work-mount-fixes-3-v1-1-be34c83956ae@kernel.org
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
propagate_mount_busy() uses mnt_get_count() via do_refcount_check() to
figure out whether a synchronous umount may proceed or not by summing
the per-CPU mnt_count counter under mount_lock.
The fastpath for mntget() and mntput() doesn't require that lock. A task
that already holds a reference and takes another one on a cpu the loop
has summed can get migrated and drops that reference again on another
cpu. The loop might not have reached that cpu yet. That ends up
undercounting. The original reference of the task is missed and umount()
ends up succeeding with the file still open. The mount is destroyed
later when the task mntput()s so it degrades to a lazy umount.
That's basically the umount(2) variant of commit 9ea0a46ca2 ("fix
mntput/mntput race") closed for the final mntput(). The same
cpu-migration caused a UAF.
Luckily this isn't a UAF. It's just a hidden reference pinning the
mount but it's wrong nonetheless.
Like the other race it's a narrow window but I can reproduce it reliably
with some patching. The debug patch below prints the sum the loop
computed next to an immediate re-sum of the same counters. It reports
who drops the last reference to that mount later:
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -34,6 +34,8 @@
#include <linux/mnt_idmapping.h>
#include <linux/pidfs.h>
#include <linux/nstree.h>
+#include <linux/debugfs.h>
+#include <linux/delay.h>
#include "pnode.h"
#include "internal.h"
@@ -263,17 +265,48 @@
#endif
}
+static u32 mnt_get_count_delay_ms;
+static int mnt_get_count_debug_id;
+
+static int __init mnt_get_count_debug_init(void)
+{
+ debugfs_create_u32("mnt_get_count_delay_ms", 0600, NULL,
+ &mnt_get_count_delay_ms);
+ return 0;
+}
+late_initcall(mnt_get_count_debug_init);
+
/*
* vfsmount lock must be held for write
*/
int mnt_get_count(struct mount *mnt)
{
#ifdef CONFIG_SMP
- int count = 0;
+ static int seen[NR_CPUS];
+ bool slow = mnt_get_count_delay_ms &&
+ !(mnt->mnt.mnt_flags & MNT_UMOUNT);
+ int count = 0, again = 0;
int cpu;
for_each_possible_cpu(cpu) {
- count += per_cpu_ptr(mnt->mnt_pcp, cpu)->mnt_count;
+ seen[cpu] = per_cpu_ptr(mnt->mnt_pcp, cpu)->mnt_count;
+ count += seen[cpu];
+ if (slow)
+ mdelay(mnt_get_count_delay_ms);
+ }
+ if (slow) {
+ for_each_possible_cpu(cpu) {
+ int now;
+
+ now = per_cpu_ptr(mnt->mnt_pcp, cpu)->mnt_count;
+ again += now;
+ pr_err("mount %d: cpu%d read %d, now %d\n",
+ mnt->mnt_id, cpu, seen[cpu], now);
+ }
+ mnt_get_count_debug_id = mnt->mnt_id;
+ pr_err("mount %d: sum %d, re-sum %d, %ps\n",
+ mnt->mnt_id, count, again,
+ __builtin_return_address(0));
}
return count;
@@ -1350,6 +1383,9 @@
smp_mb();
mnt_add_count(mnt, -1);
count = mnt_get_count(mnt);
+ if (count == 0 && mnt->mnt_id == mnt_get_count_debug_id)
+ pr_err("mount %d: last put by %s (%d)\n", mnt->mnt_id,
+ current->comm, task_pid_nr(current));
if (count != 0) {
WARN_ON(count < 0);
rcu_read_unlock();
A task holds open a directory fd on a tmpfs. A synchronous rumount() of
the tmpfs sums up. Concurrently a task calls fchdir() on that fd on
CPU 0 and fchdir()s back to / on CPU 3. This will cause a path_get() and
a path_put() with a migration in between without mount_lock held.
Adding a delay per CPU makes the count loop miss the get in a cpu it
already visited but count the put in one it hasn't.
The reproducer, run as root with debugfs mounted:
/*
* gcc -static -o umount-torn umount-torn.c
*/
#define _GNU_SOURCE
#include <errno.h>
#include <fcntl.h>
#include <sched.h>
#include <stdio.h>
#include <string.h>
#include <unistd.h>
#include <sys/mount.h>
#include <sys/stat.h>
#include <sys/wait.h>
#define KNOB "/sys/kernel/debug/mnt_get_count_delay_ms"
#define MNT "/mnt/x"
static int kmsg;
static void pin(int cpu)
{
cpu_set_t s;
CPU_ZERO(&s);
CPU_SET(cpu, &s);
sched_setaffinity(0, sizeof(s), &s);
}
static void knob(const char *v)
{
int fd = open(KNOB, O_WRONLY);
write(fd, v, strlen(v));
close(fd);
}
int main(void)
{
int last = sysconf(_SC_NPROCESSORS_ONLN) - 1;
int p[2], fd, ret;
pid_t pid;
char c;
kmsg = open("/dev/kmsg", O_WRONLY);
mkdir(MNT, 0755);
mount("none", MNT, "tmpfs", 0, NULL);
fd = open(MNT "/f", O_WRONLY | O_CREAT, 0644);
write(fd, "still here\n", 11);
close(fd);
pipe(p);
pid = fork();
if (pid == 0) {
int dirfd = open(MNT, O_RDONLY | O_DIRECTORY);
int rootfd = open("/", O_RDONLY | O_DIRECTORY);
struct stat st;
char buf[16];
/* one get/put pair while the umount is summing */
usleep(150000);
pin(0);
fchdir(dirfd); /* mntget() */
dprintf(kmsg, "holder %d: get on cpu %d\n", getpid(),
sched_getcpu());
pin(last);
fchdir(rootfd); /* mntput() */
dprintf(kmsg, "holder %d: put on cpu %d\n", getpid(),
sched_getcpu());
/* the umount has returned, is the fd still good? */
read(p[0], &c, 1);
fd = openat(dirfd, "f", O_RDONLY);
ret = read(fd, buf, sizeof(buf));
dprintf(kmsg, "holder %d: fstat %d openat %d read %d\n",
getpid(), fstat(dirfd, &st), fd, ret);
_exit(0);
}
pin(1);
knob("100");
ret = umount2(MNT, 0);
knob("0");
if (ret)
dprintf(kmsg, "umount: %s\n", strerror(errno));
else
dprintf(kmsg, "umount: 0\n");
write(p[1], "x", 1);
waitpid(pid, NULL, 0);
return 0;
}
The log shows the busy check summing 2 and when read right after it
reads 3. The loop read CPU 0 before the get landed there and CPU 3
after the put did. The umount succeeds while there's still someone
with an fd. The mount is then dropped when the fd goes:
[ 2.511944] holder 90: get on cpu 0
[ 2.512023] holder 90: put on cpu 3
[ 2.671388] mount 28: cpu0 read 1, now 2
[ 2.671398] mount 28: cpu1 read 2, now 2
[ 2.671401] mount 28: cpu2 read 0, now 0
[ 2.671404] mount 28: cpu3 read -1, now -1
[ 2.671407] mount 28: sum 2, re-sum 3, propagate_mount_busy
[ 2.671773] umount: 0
[ 2.672963] holder 90: fstat 0 openat 8 read 11
[ 2.673663] mount 28: last put by umount-torn (90)
Easy fix would be to order mntget() against the loop that sums up. But
that's a full barrier and a flag test in mntget() afaict. No bueno...
That would cost every path_get().
So learn from SRCU. Keep gets and puts in separate per-cpu counters and
make mnt_get_count() sum all puts first and all gets second with a full
barrier in between.
A put is always after its own get (hopefully...) independent of the cpu
the task is on. So a put the first pass counted must have its get
visible to the second pass. Ergo, we can't undercount.
This is the two pass sum srcu_readers_active_idx_check() does for
srcu_locks and srcu_unlocks. mntget() keeps a single
this_cpu_inc(). mntput() gains an smp_wmb() before its increment, a
compiler barrier on x86 and a store fence on the weakly ordered
architectures, since the two counters are two words and a put stored
after a get could otherwise become visible before it. The counters
are unsigned so the difference stays right after they wrap.
The count can still come out high when a get is counted whose put
the first pass did not see. We don't care. A task can only take a new
reference if:
(1) it holds one already
(2) through __legitimize_mnt() which the seqcount orders
against the write side
So a get in this window means someone has a reference and -EBUSY is
correct. The final mntput() is unchanged. Once ->mnt_ns is NULL every
put takes the slow path under mount_lock and so no put is in flight
while it counts. Here's an attempted proof for that crap:
(i) A put never becomes visible before its get. Within one task the
smp_wmb() in mntput_no_expire() orders every earlier store before the
increment of the puts.
Any task that migrated in between had its stores flushed by the
scheduler. When that reference is handed to another task that's
trivially true.
(ii) The double count in mnt_get_count() is ordered via smp_mb(). That
keeps every read of the puts counter before every read of the gets
counter.
(iii) We never undercount by (ii).
(iv) There are no spurious counts/differences. A get is counted only if
it happened and a put is left out only if it hasn't happened when
the counter was read. With (iii) this means it's equivalent to the
old count whenever nothing is taken or dropped during the loop.
(v) do_umount() holds a reference and mount_lock hence. New references
come from: (v.a) __legitimize_mnt() managed by the seqcount. Either
it's smp_mb() made the get visible to the loop or it takes the lock
and waits for the loop. By (iii) a mount with a additional
references is refused. By (iv) refusing the umount just means that
any additional reference existed, same as the old counter.
(vi) mntput_no_expire_slowpath() runs with ->mnt_ns already NULL. Hence,
every other put must take the slowpath as well and hangs on mount_lock.
The grace period in namespace_unlock() will make every fast path do
a put that still saw ->mnt_ns.
No put is in flight during the loop.
A get that is in flight comes from a task holding another reference.
That must be included by the count or from __legitimize_mnt() as in
(v).
Ergo, zero means the last reference and non-zero means a reference
is live.
(vii) may_umount() uses propagate_mount_busy() as well and
may_umount_tree() sums mnt_get_count() itself.
With this patch the race is dead:
[ 3.066786] mount 28: sum 3, re-sum 3, propagate_mount_busy
[ 3.105867] umount: Device or resource busy
[ 3.106047] holder 90: fstat 0 openat 8 read 11
Link: https://patch.msgid.link/20260923-work-mount-fixes-v1-7-f424cf8d3242@kernel.org
Fixes: b3e19d924b ("fs: scale mntget/mntput")
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Please consider pulling these changes from the signed vfs-7.2-rc1.procfs tag.
Thanks!
Christian
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQRAhzRXHqcMeLMyaSiRxhvAZXjcogUCaiwLKgAKCRCRxhvAZXjc
otp1AQD6ONFiqQZJhnHbVg1MJlA8iP80x0qlil0OKGUMOby4NAD+MZymcLjtP+4O
T1b9Uca2/qz+L1F5o7vJAD3OAd4BzQA=
=/KGF
-----END PGP SIGNATURE-----
Merge tag 'vfs-7.2-rc1.procfs' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs
Pull procfs updates from Christian Brauner:
- Revamp fs/filesystems.c
The file was a mess with a hand-rolled linked list in desperate need
of a cleanup. The filesystems list is now RCU-ified, /proc files can
be marked permanent from outside fs/proc/, and the string emitted
when reading /proc/filesystems is pre-generated and cached instead of
pointer-chasing and printfing entry by entry on every read.
The file is read frequently because libselinux reads it and is linked
into numerous frequently used programs (even ones you would not
suspect, like sed!). Scalability also improves since reference
maintenance on open/close is bypassed.
open+read+close cycle single-threaded (ops/s):
before: 442732
after: 1063462 (+140%)
open+read+close cycle with 20 processes (ops/s):
before: 606177
after: 3300576 (+444%)
A follow-up patch adds missing unlocks in some corner cases and
tidies things up.
- Relax the mount visibility check for subset=pid mounts
When procfs is mounted with subset=pid, all static files become
unavailable and only the dynamic pid information is accessible. In
that case there is no point in imposing the full mount visibility
restrictions on the mounter - everything that can be hidden in procfs
is already inaccessible. These restrictions prevented procfs from
being mounted inside rootless containers since almost all container
implementations overmount parts of procfs to hide certain
directories.
As part of this /proc/self/net is only shown in subset=pid mounts for
CAP_NET_ADMIN, reconfiguring subset=pid is rejected, the
SB_I_USERNS_VISIBLE superblock flag is replaced with an
FS_USERNS_MOUNT_RESTRICTED filesystem flag, fully visible mounts are
recorded in a list, and the mount restrictions are finally
documented.
- Protect ptrace_may_access() with exec_update_lock in procfs
Most uses of ptrace_may_access() in procfs should hold
exec_update_lock to avoid TOCTOU issues with concurrent privileged
execve() (like setuid binary execution).
This fixes the easy cases - the owner and visibility checks and the
FD link permission checks - with the gnarlier ones to follow later.
* tag 'vfs-7.2-rc1.procfs' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs:
fs: fix ups and tidy ups to /proc/filesystems caching
proc: protect ptrace_may_access() with exec_update_lock (FD links)
proc: protect ptrace_may_access() with exec_update_lock (part 1)
docs: proc: add documentation about mount restrictions
proc: handle subset=pid separately in userns visibility checks
proc: prevent reconfiguring subset=pid
proc: subset=pid: Show /proc/self/net only for CAP_NET_ADMIN
fs: cache the string generated by reading /proc/filesystems
sysfs: remove trivial sysfs_get_tree() wrapper
fs: RCU-ify filesystems list
fs: move SB_I_USERNS_VISIBLE to FS_USERNS_MOUNT_RESTRICTED
proc: allow to mark /proc files permanent outside of fs/proc/
namespace: record fully visible mounts in list
may_decode_fh() accesses mount::mnt_ns without holding any locks; that
means the mount can concurrently be unmounted, and the mnt_namespace can
concurrently be freed after an RCU grace period.
This race can happens as follows, assuming that the mount point was
created by open_tree(..., OPEN_TREE_CLONE):
thread 1 thread 2 RCU
__do_sys_open_by_handle_at
do_handle_open
handle_to_path
may_decode_fh
is_mounted
[mount::mnt_ns access]
[mount::mnt_ns access]
__do_sys_close
fput_close_sync
__fput
dissolve_on_fput
umount_tree
class_namespace_excl_destructor
namespace_unlock
free_mnt_ns
mnt_ns_tree_remove
call_rcu(mnt_ns_release_rcu)
mnt_ns_release_rcu
mnt_ns_release
kfree
[mnt_namespace::user_ns access] **UAF**
Fix it by taking rcu_read_lock() around the mount::mnt_ns access, like
in __prepend_path().
Additionally, document the semantics of mount::mnt_ns, and use WRITE_ONCE()
for writers that can race with lockless readers.
This bug is unreachable unless one of the following is set:
- CONFIG_PREEMPTION
- CONFIG_RCU_STRICT_GRACE_PERIOD
because it requires an RCU grace period to happen during a syscall without
an explicit preemption.
This doesn't seem to have interesting security impact; worst-case, it could
leak the result of an integer comparison to userspace (from the level
check in cap_capable()), cause an endless loop, or crash the kernel by
dereferencing an invalid address.
Fixes: 620c266f39 ("fhandle: relax open_by_handle_at() permission checks")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Link: https://patch.msgid.link/20260603-vfs-fhandle-uaf-fix-v2-1-d05db76a5084@google.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
Instead of wading through all the mounts in the mount namespace rbtree
to find fully visible procfs and sysfs mounts, be honest about them
being special cruft and record them in a separate per-mount namespace
list.
Link: https://patch.msgid.link/684859a8e0ac929cb89c1fbe16ce15b30c70eb1f.1777278334.git.legion@kernel.org
Reviewed-by: Aleksa Sarai <aleksa@amutable.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
Currently pivot_root() doesn't work on the real rootfs because it
cannot be unmounted. Userspace has to do a recursive removal of the
initramfs contents manually before continuing the boot.
Really all we want from the real rootfs is to serve as the parent mount
for anything that is actually useful such as the tmpfs or ramfs for
initramfs unpacking or the rootfs itself. There's no need for the real
rootfs to actually be anything meaningful or useful. Add a immutable
rootfs called "nullfs" that can be selected via the "nullfs_rootfs"
kernel command line option.
The kernel will mount a tmpfs/ramfs on top of it, unpack the initramfs
and fire up userspace which mounts the rootfs and can then just do:
chdir(rootfs);
pivot_root(".", ".");
umount2(".", MNT_DETACH);
and be done with it. (Ofc, userspace can also choose to retain the
initramfs contents by using something like pivot_root(".", "/initramfs")
without unmounting it.)
Technically this also means that the rootfs mount in unprivileged
namespaces doesn't need to become MNT_LOCKED anymore as it's guaranteed
that the immutable rootfs remains permanently empty so there cannot be
anything revealed by unmounting the covering mount.
In the future this will also allow us to create completely empty mount
namespaces without risking to leak anything.
systemd already handles this all correctly as it tries to pivot_root()
first and falls back to MS_MOVE only when that fails.
This goes back to various discussion in previous years and a LPC 2024
presentation about this very topic.
Link: https://patch.msgid.link/20260112-work-immutable-rootfs-v2-3-88dd1c34a204@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
Stop playing games with the namespace id and use a boolean instead:
* This will remove the special-casing we need to do everywhere for mount
namespaces.
* It will allow us to use asserts on the namespace id for initial
namespaces everywhere.
* It will allow us to put anonymous mount namespaces on the namespaces
trees in the future and thus make them available to statmount() and
listmount().
Link: https://patch.msgid.link/20251110-work-namespace-nstree-fixes-v1-10-e8a9264e0fb9@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
* saner handling of guards in fs/namespace.c, getting
rid of needlessly strong locking in some of the users.
* lock_mount() calling conventions change - have it set
the environment for attaching to given location, storing the
results in caller-supplied object, without altering the passed
struct path. Make unlock_mount() called as __cleanup for those
objects. It's not exactly guard(), but similar to it.
* MNT_WRITE_HOLD done right - mnt_hold_writers() does *not*
mess with ->mnt_flags anymore, so insertion of a new mount into
->s_mounts of underlying superblock does not, in itself, expose
->mnt_flags of that mount to concurrent modifications.
* getting rid of pathological cases when umount() spends
quadratic time removing the victims from propagation graph -
part of that had been dealt with last cycle, this should finish
it.
* a bunch of stuff constified.
* assorted cleanups.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQQqUNBr3gm4hGXdBJlZ7Krx/gZQ6wUCaNhzLAAKCRBZ7Krx/gZQ
63/IAP4yxJ6e3Pt66Uw0MeuSNmeLsQwb7mYo72lsYHpxjYANZAEAspMaLDU9NHxM
Dy6WDVoJnf7+aDlD6E443YMfPX8XRQM=
=5T+t
-----END PGP SIGNATURE-----
Merge tag 'pull-mount' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
Pull vfs mount updates from Al Viro:
"Several piles this cycle, this mount-related one being the largest and
trickiest:
- saner handling of guards in fs/namespace.c, getting rid of
needlessly strong locking in some of the users
- lock_mount() calling conventions change - have it set the
environment for attaching to given location, storing the results in
caller-supplied object, without altering the passed struct path.
Make unlock_mount() called as __cleanup for those objects. It's not
exactly guard(), but similar to it
- MNT_WRITE_HOLD done right.
mnt_hold_writers() does *not* mess with ->mnt_flags anymore, so
insertion of a new mount into ->s_mounts of underlying superblock
does not, in itself, expose ->mnt_flags of that mount to concurrent
modifications
- getting rid of pathological cases when umount() spends quadratic
time removing the victims from propagation graph - part of that had
been dealt with last cycle, this should finish it
- a bunch of stuff constified
- assorted cleanups
* tag 'pull-mount' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (64 commits)
constify {__,}mnt_is_readonly()
WRITE_HOLD machinery: no need for to bump mount_lock seqcount
struct mount: relocate MNT_WRITE_HOLD bit
preparations to taking MNT_WRITE_HOLD out of ->mnt_flags
setup_mnt(): primitive for connecting a mount to filesystem
simplify the callers of mnt_unhold_writers()
copy_mnt_ns(): use guards
copy_mnt_ns(): use the regular mechanism for freeing empty mnt_ns on failure
open_detached_copy(): separate creation of namespace into helper
open_detached_copy(): don't bother with mount_lock_hash()
path_has_submounts(): use guard(mount_locked_reader)
fs/namespace.c: sanitize descriptions for {__,}lookup_mnt()
ecryptfs: get rid of pointless mount references in ecryptfs dentries
umount_tree(): take all victims out of propagation graph at once
do_mount(): use __free(path_put)
do_move_mount_old(): use __free(path_put)
constify can_move_mount_beneath() arguments
path_umount(): constify struct path argument
may_copy_tree(), __do_loopback(): constify struct path argument
path_mount(): constify struct path argument
...
Move the mount namespace to the generic ns lookup infrastructure.
This allows us to drop a bunch of members from struct mnt_namespace.
Signed-off-by: Christian Brauner <brauner@kernel.org>
... from ->mnt_flags to LSB of ->mnt_pprev_for_sb.
This is safe - we always set and clear it within the same mount_lock
scope, so we won't interfere with list operations - traversals are
always forward, so they don't even look at ->mnt_prev_for_sb and
both insertions and removals are in mount_lock scopes of their own,
so that bit will be clear in *all* mount instances during those.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
We have an unpleasant wart in accessibility rules for struct mount. There
are per-superblock lists of mounts, used by sb_prepare_remount_readonly()
to check if any of those is currently claimed for write access and to
block further attempts to get write access on those until we are done.
As soon as it is attached to a filesystem, mount becomes reachable
via that list. Only sb_prepare_remount_readonly() traverses it and
it only accesses a few members of struct mount. Unfortunately,
->mnt_flags is one of those and it is modified - MNT_WRITE_HOLD set
and then cleared. It is done under mount_lock, so from the locking
rules POV everything's fine.
However, it has easily overlooked implications - once mount has been
attached to a filesystem, it has to be treated as globally visible.
In particular, initializing ->mnt_flags *must* be done either prior
to that point or under mount_lock. All other members are still
private at that point.
Life gets simpler if we move that bit (and that's *all* that can get
touched by access via this list) out of ->mnt_flags. It's not even
hard to do - currently the list is implemented as list_head one,
anchored in super_block->s_mounts and linked via mount->mnt_instance.
As the first step, switch it to hlist-like open-coded structure -
address of the first mount in the set is stored in ->s_mounts
and ->mnt_instance replaced with ->mnt_next_for_sb and ->mnt_pprev_for_sb -
the former either NULL or pointing to the next mount in set, the
latter - address of either ->s_mounts or ->mnt_next_for_sb in the
previous element of the set.
In the next commit we'll steal the LSB of ->mnt_pprev_for_sb as
replacement for MNT_WRITE_HOLD.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Returns the final (topmost) mount in the chain of overmounts
starting at given mount. Same locking rules as for any mount
tree traversal - either the spinlock side of mount_lock, or
rcu + sample the seqcount side of mount_lock before the call
and recheck afterwards.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
mount_writer: write_seqlock; that's an equivalent of {un,}lock_mount_hash()
mount_locked_reader: read_seqlock_excl; these tend to be open-coded.
No bulk conversions, please - if nothing else, quite a few places take
use mount_writer form when mount_locked_reader is sufficent. It needs
to be dealt with carefully.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The only place that really needs to be adjusted is commit_tree() -
there we need to iterate through the copy and we might as well
use next_mnt() for that. However, in case when our tree has been
slid under something already mounted (propagation to a mountpoint
that already has something mounted on it or a 'beneath' move_mount)
we need to take care not to walk into the overmounting tree.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Several flags are updated and checked only under namespace_sem; we are
already making use of that when we are checking them without mount_lock,
but we have to hold mount_lock for all updates, which makes things
clumsier than they have to be.
Take MNT_SHARED, MNT_UNBINDABLE, MNT_MARKED and MNT_UMOUNT_CANDIDATE
into a separate field (->mnt_t_flags), renaming them to T_SHARED,
etc. to avoid confusion. All accesses must be under namespace_sem.
That changes locking requirements for mnt_change_propagation() and
set_mnt_shared() - only namespace_sem is needed now. The same goes
for SET_MNT_MARKED et.al.
There might be more flags moved from ->mnt_flags to that field;
this is just the initial set.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
struct mountpoint has an odd kinda-sorta refcount in it. It's always
either equal to or one above the number of mounts attached to that
mountpoint.
"One above" happens when a function takes a temporary reference to
mountpoint. Things get simpler if we express that as inserting
a local object into ->m_list and removing it to drop the reference.
New calling conventions:
1) lock_mount(), do_lock_mount(), get_mountpoint() and lookup_mountpoint()
take an extra struct pinned_mountpoint * argument and returns 0/-E...
(or true/false in case of lookup_mountpoint()) instead of returning
struct mountpoint pointers. In case of success, the struct mountpoint *
we used to get can be found as pinned_mountpoint.mp
2) unlock_mount() (always paired with lock_mount()/do_lock_mount()) takes
an address of struct pinned_mountpoint - the same that had been passed to
lock_mount()/do_lock_mount().
3) put_mountpoint() for a temporary reference (paired with get_mountpoint()
or lookup_mountpoint()) is replaced with unpin_mountpoint(), which takes
the address of pinned_mountpoint we passed to matching {get,lookup}_mountpoint().
4) all instances of pinned_mountpoint are local variables; they always live on
stack. {} is used for initializer, after successful {get,lookup}_mountpoint()
we must make sure to call unpin_mountpoint() before leaving the scope and
after successful {do_,}lock_mount() we must make sure to call unlock_mount()
before leaving the scope.
5) all manipulations of ->m_count are gone, along with ->m_count itself.
struct mountpoint lives while its ->m_list is non-empty.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The variant currently in the tree has problems; trying to prove
correctness has caught at least one class of bugs (reparenting
that ends up moving the visible location of reparented mount, due
to not excluding some of the counterparts on propagation that
should've been included).
I tried to prove that it's the only bug there; I'm still not sure
whether it is. If anyone can reconstruct and write down an analysis
of the mainline implementation, I'll gladly review it; as it is,
I ended up doing a different implementation. Candidate collection
phase is similar, but trimming the set down until it satisfies the
constraints turned out pretty different.
I hoped to do transformation as a massage series, but that turns out
to be too convoluted. So it's a single patch replacing propagate_umount()
and friends in one go, with notes and analysis in D/f/propagate_umount.txt
(in addition to inline comments).
As far I can tell, it is provably correct and provably linear by the number
of mounts we need to look at in order to decide what should be unmounted.
It even builds and seems to survive testing...
Another nice thing that fell out of that is that ->mnt_umounting is no longer
needed.
Compared to the first version:
* explicit MNT_UMOUNT_CANDIDATE flag for is_candidate()
* trim_ancestors() only clears that flag, leaving the suckers on list
* trim_one() and handle_locked() take the stuff with flag cleared off
the list. That allows to iterate with list_for_each_entry_safe() when calling
trim_one() - it removes at most one element from the list now.
* no globals - I didn't bother with any kind of context, not worth it.
* Notes updated accordingly; I have not touch the terms yet.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
checks if mount is the root of an anonymouns namespace.
Switch open-coded equivalents to using it.
For mounts that belong to anon namespace !mnt_has_parent(mount)
is the same as mount == ns->root, and intent is more obvious in
the latter form.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Currently it's still possible to run into a pathological situation when
two hashed mounts share both parent and mountpoint. That does not work
well, for obvious reasons.
We are not far from getting rid of that; the only remaining gap is
attach_recursive_mnt() not being careful enough when sliding a tree
under existing mount (for propagated copies or in 'beneath' case for
the original one).
To deal with that cleanly we need to be able to find overmounts
(i.e. mounts on top of parent's root); we could do hash lookups or scan
the list of children but either would be costly. Since one of the results
we get from that will be prevention of multiple parallel overmounts, let's
just bite the bullet and store a (non-counting) reference to overmount
in struct mount.
With that done, closing the hole in attach_recursive_mnt() becomes easy
- we just need to follow the chain of overmounts before we change the
mountpoint of the mount we are sliding things under.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
All versions up to 6.14 did not propagate mount events into detached
tree. Shortly after 6.14 a merge of vfs-6.15-rc1.mount.namespace
(130e696aa6) has changed that.
Unfortunately, that has caused userland regressions (reported in
https://lore.kernel.org/all/CAOYeF9WQhFDe+BGW=Dp5fK8oRy5AgZ6zokVyTj1Wp4EUiYgt4w@mail.gmail.com/)
Straight revert wouldn't be an option - in particular, the variant in 6.14
had a bug that got fixed in d1ddc6f1d9 ("fix IS_MNT_PROPAGATING uses")
and we don't want to bring the bug back.
This is a modification of manual revert posted by Christian, with changes
needed to avoid reintroducing the breakage in scenario described in
d1ddc6f1d9.
Cc: stable@vger.kernel.org
Reported-by: Allison Karlitskaya <lis@redhat.com>
Tested-by: Allison Karlitskaya <lis@redhat.com>
Acked-by: Christian Brauner <brauner@kernel.org>
Co-developed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQRAhzRXHqcMeLMyaSiRxhvAZXjcogUCZ90r2wAKCRCRxhvAZXjc
ouC6AQCk3MoqskN0WeNcaZT23dB7dHbEhf/7YXOFC9MFRMKXqQD9Fbn95+GuIe3U
nBVPbVyQfDtfXE08ml6gbDJrCsbkkQI=
=Xm1C
-----END PGP SIGNATURE-----
Merge tag 'vfs-6.15-rc1.mount.namespace' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs
Pull vfs mount namespace updates from Christian Brauner:
"This expands the ability of anonymous mount namespaces:
- Creating detached mounts from detached mounts
Currently, detached mounts can only be created from attached
mounts. This limitaton prevents various use-cases. For example, the
ability to mount a subdirectory without ever having to make the
whole filesystem visible first.
The current permission modelis:
(1) Check that the caller is privileged over the owning user
namespace of it's current mount namespace.
(2) Check that the caller is located in the mount namespace of the
mount it wants to create a detached copy of.
While it is not strictly necessary to do it this way it is
consistently applied in the new mount api. This model will also be
used when allowing the creation of detached mount from another
detached mount.
The (1) requirement can simply be met by performing the same check
as for the non-detached case, i.e., verify that the caller is
privileged over its current mount namespace.
To meet the (2) requirement it must be possible to infer the origin
mount namespace that the anonymous mount namespace of the detached
mount was created from.
The origin mount namespace of an anonymous mount is the mount
namespace that the mounts that were copied into the anonymous mount
namespace originate from.
In order to check the origin mount namespace of an anonymous mount
namespace the sequence number of the original mount namespace is
recorded in the anonymous mount namespace.
With this in place it is possible to perform an equivalent check
(2') to (2). The origin mount namespace of the anonymous mount
namespace must be the same as the caller's mount namespace. To
establish this the sequence number of the caller's mount namespace
and the origin sequence number of the anonymous mount namespace are
compared.
The caller is always located in a non-anonymous mount namespace
since anonymous mount namespaces cannot be setns()ed into. The
caller's mount namespace will thus always have a valid sequence
number.
The owning namespace of any mount namespace, anonymous or
non-anonymous, can never change. A mount attached to a
non-anonymous mount namespace can never change mount namespace.
If the sequence number of the non-anonymous mount namespace and the
origin sequence number of the anonymous mount namespace match, the
owning namespaces must match as well.
Hence, the capability check on the owning namespace of the caller's
mount namespace ensures that the caller has the ability to copy the
mount tree.
- Allow mount detached mounts on detached mounts
Currently, detached mounts can only be mounted onto attached
mounts. This limitation makes it impossible to assemble a new
private rootfs and move it into place. Instead, a detached tree
must be created, attached, then mounted open and then either moved
or detached again. Lift this restriction.
In order to allow mounting detached mounts onto other detached
mounts the same permission model used for creating detached mounts
from detached mounts can be used (cf. above).
Allowing to mount detached mounts onto detached mounts leaves three
cases to consider:
(1) The source mount is an attached mount and the target mount is
a detached mount. This would be equivalent to moving a mount
between different mount namespaces. A caller could move an
attached mount to a detached mount. The detached mount can now
be freely attached to any mount namespace. This changes the
current delegatioh model significantly for no good reason. So
this will fail.
(2) Anonymous mount namespaces are always attached fully, i.e., it
is not possible to only attach a subtree of an anoymous mount
namespace. This simplifies the implementation and reasoning.
Consequently, if the anonymous mount namespace of the source
detached mount and the target detached mount are the identical
the mount request will fail.
(3) The source mount's anonymous mount namespace is different from
the target mount's anonymous mount namespace.
In this case the source anonymous mount namespace of the
source mount tree must be freed after its mounts have been
moved to the target anonymous mount namespace. The source
anonymous mount namespace must be empty afterwards.
By allowing to mount detached mounts onto detached mounts a caller
may do the following:
fd_tree1 = open_tree(-EBADF, "/mnt", OPEN_TREE_CLONE)
fd_tree2 = open_tree(-EBADF, "/tmp", OPEN_TREE_CLONE)
fd_tree1 and fd_tree2 refer to two different detached mount trees
that belong to two different anonymous mount namespace.
It is important to note that fd_tree1 and fd_tree2 both refer to
the root of their respective anonymous mount namespaces.
By allowing to mount detached mounts onto detached mounts the
caller may now do:
move_mount(fd_tree1, "", fd_tree2, "",
MOVE_MOUNT_F_EMPTY_PATH | MOVE_MOUNT_T_EMPTY_PATH)
This will cause the detached mount referred to by fd_tree1 to be
mounted on top of the detached mount referred to by fd_tree2.
Thus, the detached mount fd_tree1 is moved from its separate
anonymous mount namespace into fd_tree2's anonymous mount
namespace.
It also means that while fd_tree2 continues to refer to the root of
its respective anonymous mount namespace fd_tree1 doesn't anymore.
This has the consequence that only fd_tree2 can be moved to another
anonymous or non-anonymous mount namespace. Moving fd_tree1 will
now fail as fd_tree1 doesn't refer to the root of an anoymous mount
namespace anymore.
Now fd_tree1 and fd_tree2 refer to separate detached mount trees
referring to the same anonymous mount namespace.
This is conceptually fine. The new mount api does allow for this to
happen already via:
mount -t tmpfs tmpfs /mnt
mkdir -p /mnt/A
mount -t tmpfs tmpfs /mnt/A
fd_tree3 = open_tree(-EBADF, "/mnt", OPEN_TREE_CLONE | AT_RECURSIVE)
fd_tree4 = open_tree(-EBADF, "/mnt/A", 0)
Both fd_tree3 and fd_tree4 refer to two different detached mount
trees but both detached mount trees refer to the same anonymous
mount namespace. An as with fd_tree1 and fd_tree2, only fd_tree3
may be moved another mount namespace as fd_tree3 refers to the root
of the anonymous mount namespace just while fd_tree4 doesn't.
However, there's an important difference between the
fd_tree3/fd_tree4 and the fd_tree1/fd_tree2 example.
Closing fd_tree4 and releasing the respective struct file will have
no further effect on fd_tree3's detached mount tree.
However, closing fd_tree3 will cause the mount tree and the
respective anonymous mount namespace to be destroyed causing the
detached mount tree of fd_tree4 to be invalid for further mounting.
By allowing to mount detached mounts on detached mounts as in the
fd_tree1/fd_tree2 example both struct files will affect each other.
Both fd_tree1 and fd_tree2 refer to struct files that have
FMODE_NEED_UNMOUNT set.
To handle this we use the fact that @fd_tree1 will have a parent
mount once it has been attached to @fd_tree2.
When dissolve_on_fput() is called the mount that has been passed in
will refer to the root of the anonymous mount namespace. If it
doesn't it would mean that mounts are leaked. So before allowing to
mount detached mounts onto detached mounts this would be a bug.
Now that detached mounts can be mounted onto detached mounts it
just means that the mount has been attached to another anonymous
mount namespace and thus dissolve_on_fput() must not unmount the
mount tree or free the anonymous mount namespace as the file
referring to the root of the namespace hasn't been closed yet.
If it had been closed yet it would be obvious because the mount
namespace would be NULL, i.e., the @fd_tree1 would have already
been unmounted. If @fd_tree1 hasn't been unmounted yet and has a
parent mount it is safe to skip any cleanup as closing @fd_tree2
will take care of all cleanup operations.
- Allow mount propagation for detached mount trees
In commit ee2e3f5062 ("mount: fix mounting of detached mounts
onto targets that reside on shared mounts") I fixed a bug where
propagating the source mount tree of an anonymous mount namespace
into a target mount tree of a non-anonymous mount namespace could
be used to trigger an integer overflow in the non-anonymous mount
namespace causing any new mounts to fail.
The cause of this was that the propagation algorithm was unable to
recognize mounts from the source mount tree that were already
propagated into the target mount tree and then reappeared as
propagation targets when walking the destination propagation mount
tree.
When fixing this I disabled mount propagation into anonymous mount
namespaces. Make it possible for anonymous mount namespace to
receive mount propagation events correctly. This is now also a
correctness issue now that we allow mounting detached mount trees
onto detached mount trees.
Mark the source anonymous mount namespace with MNTNS_PROPAGATING
indicating that all mounts belonging to this mount namespace are
currently in the process of being propagated and make the
propagation algorithm discard those if they appear as propagation
targets"
* tag 'vfs-6.15-rc1.mount.namespace' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs: (21 commits)
selftests: test subdirectory mounting
selftests: add test for detached mount tree propagation
fs: namespace: fix uninitialized variable use
mount: handle mount propagation for detached mount trees
fs: allow creating detached mounts from fsmount() file descriptors
selftests: seventh test for mounting detached mounts onto detached mounts
selftests: sixth test for mounting detached mounts onto detached mounts
selftests: fifth test for mounting detached mounts onto detached mounts
selftests: fourth test for mounting detached mounts onto detached mounts
selftests: third test for mounting detached mounts onto detached mounts
selftests: second test for mounting detached mounts onto detached mounts
selftests: first test for mounting detached mounts onto detached mounts
fs: mount detached mounts onto detached mounts
fs: support getname_maybe_null() in move_mount()
selftests: create detached mounts from detached mounts
fs: create detached mounts from detached mounts
fs: add may_copy_tree()
fs: add fastpath for dissolve_on_fput()
fs: add assert for move_mount()
fs: add mnt_ns_empty() helper
...
In commit ee2e3f5062 ("mount: fix mounting of detached mounts onto
targets that reside on shared mounts") I fixed a bug where propagating
the source mount tree of an anonymous mount namespace into a target
mount tree of a non-anonymous mount namespace could be used to trigger
an integer overflow in the non-anonymous mount namespace causing any new
mounts to fail.
The cause of this was that the propagation algorithm was unable to
recognize mounts from the source mount tree that were already propagated
into the target mount tree and then reappeared as propagation targets
when walking the destination propagation mount tree.
When fixing this I disabled mount propagation into anonymous mount
namespaces. Make it possible for anonymous mount namespace to receive
mount propagation events correctly. This is no also a correctness issue
now that we allow mounting detached mount trees onto detached mount
trees.
Mark the source anonymous mount namespace with MNTNS_PROPAGATING
indicating that all mounts belonging to this mount namespace are
currently in the process of being propagated and make the propagation
algorithm discard those if they appear as propagation targets.
Link: https://lore.kernel.org/r/20250225-work-mount-propagation-v1-1-e6e3724500eb@kernel.org
Signed-off-by: Christian Brauner <brauner@kernel.org>
Add notifications for attaching and detaching mounts. The following new
event masks are added:
FAN_MNT_ATTACH - Mount was attached
FAN_MNT_DETACH - Mount was detached
If a mount is moved, then the event is reported with (FAN_MNT_ATTACH |
FAN_MNT_DETACH).
These events add an info record of type FAN_EVENT_INFO_TYPE_MNT containing
these fields identifying the affected mounts:
__u64 mnt_id - the ID of the mount (see statmount(2))
FAN_REPORT_MNT must be supplied to fanotify_init() to receive these events
and no other type of event can be received with this report type.
Marks are added with FAN_MARK_MNTNS, which records the mount namespace from
an nsfs file (e.g. /proc/self/ns/mnt).
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Link: https://lore.kernel.org/r/20250129165803.72138-3-mszeredi@redhat.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
This is just the plumbing between the event source (fs/namespace.c) and the
event consumer (fanotify). In itself it does nothing.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Link: https://lore.kernel.org/r/20250129165803.72138-2-mszeredi@redhat.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
We already made the rbtree lookup lockless for the simple lookup case.
However, walking the list of mount namespaces via nsfs still happens
with taking the read lock blocking concurrent additions of new mount
namespaces pointlessly. Plus, such additions are rare anyway so allow
lockless lookup of the previous and next mount namespace by keeping a
separate list. This also allows to make some things simpler in the code.
Link: https://lore.kernel.org/r/20241213-work-mount-rbtree-lockless-v3-5-6e3cdaf9b280@kernel.org
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
Currently we use a read-write lock but for the simple search case we can
make this lockless. Creating a new mount namespace is a rather rare
event compared with querying mounts in a foreign mount namespace. Once
this is picked up by e.g., systemd to list mounts in another mount in
it's isolated services or in containers this will be used a lot so this
seems worthwhile doing.
Link: https://lore.kernel.org/r/20241213-work-mount-rbtree-lockless-v3-3-6e3cdaf9b280@kernel.org
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
Move mnt->mnt_node into the union with mnt->mnt_rcu and mnt->mnt_llist
instead of keeping it with mnt->mnt_list. This allows us to use
RB_CLEAR_NODE(&mnt->mnt_node) in umount_tree() as well as
list_empty(&mnt->mnt_node). That in turn allows us to remove MNT_ONRB.
This also fixes the bug reported in [1] where seemingly MNT_ONRB wasn't
set in @mnt->mnt_flags even though the mount was present in the mount
rbtree of the mount namespace.
The root cause is the following race. When a btrfs subvolume is mounted
a temporary mount is created:
btrfs_get_tree_subvol()
{
mnt = fc_mount()
// Register the newly allocated mount with sb->mounts:
lock_mount_hash();
list_add_tail(&mnt->mnt_instance, &mnt->mnt.mnt_sb->s_mounts);
unlock_mount_hash();
}
and registered on sb->s_mounts. Later it is added to an anonymous mount
namespace via mount_subvol():
-> mount_subvol()
-> mount_subtree()
-> alloc_mnt_ns()
mnt_add_to_ns()
vfs_path_lookup()
put_mnt_ns()
The mnt_add_to_ns() call raises MNT_ONRB in @mnt->mnt_flags. If someone
concurrently does a ro remount:
reconfigure_super()
-> sb_prepare_remount_readonly()
{
list_for_each_entry(mnt, &sb->s_mounts, mnt_instance) {
}
all mounts registered in sb->s_mounts are visited and first
MNT_WRITE_HOLD is raised, then MNT_READONLY is raised, and finally
MNT_WRITE_HOLD is removed again.
The flag modification for MNT_WRITE_HOLD/MNT_READONLY and MNT_ONRB race
so MNT_ONRB might be lost.
Fixes: 2eea9ce431 ("mounts: keep list of mounts in an rbtree")
Cc: <stable@kernel.org> # v6.8+
Link: https://lore.kernel.org/r/20241215-vfs-6-14-mount-work-v1-1-fd55922c4af8@kernel.org
Link: https://lore.kernel.org/r/ec6784ed-8722-4695-980a-4400d4e7bd1a@gmx.com [1]
Signed-off-by: Christian Brauner <brauner@kernel.org>
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQRAhzRXHqcMeLMyaSiRxhvAZXjcogUCZuQEmwAKCRCRxhvAZXjc
otRsAQCUdlBS/ky2JiYn3ePURKYVBgRq/+PnmhRrBNDuv+ToZwD+NRLNlOM8FzQy
c8BMSq0rkwO2C5Aax3kGxgTPMEuuCwc=
=QLvm
-----END PGP SIGNATURE-----
Merge tag 'vfs-6.12.mount' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs
Pull vfs mount updates from Christian Brauner:
"Recently, we added the ability to list mounts in other mount
namespaces and the ability to retrieve namespace file descriptors
without having to go through procfs by deriving them from pidfds.
This extends nsfs in two ways:
(1) Add the ability to retrieve information about a mount namespace
via NS_MNT_GET_INFO.
This will return the mount namespace id and the number of mounts
currently in the mount namespace. The number of mounts can be
used to size the buffer that needs to be used for listmount() and
is in general useful without having to actually iterate through
all the mounts.
The structure is extensible.
(2) Add the ability to iterate through all mount namespaces over
which the caller holds privilege returning the file descriptor
for the next or previous mount namespace.
To retrieve a mount namespace the caller must be privileged wrt
to it's owning user namespace. This means that PID 1 on the host
can list all mounts in all mount namespaces or that a container
can list all mounts of its nested containers.
Optionally pass a structure for NS_MNT_GET_INFO with
NS_MNT_GET_{PREV,NEXT} to retrieve information about the mount
namespace in one go.
(1) and (2) can be implemented for other namespace types easily.
Together with recent api additions this means one can iterate through
all mounts in all mount namespaces without ever touching procfs.
The commit message in 49224a345c ('Merge patch series "nsfs: iterate
through mount namespaces"') contains example code how to do this"
* tag 'vfs-6.12.mount' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs:
nsfs: iterate through mount namespaces
file: add fput() cleanup helper
fs: add put_mnt_ns() cleanup helper
fs: allow mount namespace fd
It is already possible to list mounts in other mount namespaces and to
retrieve namespace file descriptors without having to go through procfs
by deriving them from pidfds.
Augment these abilities by adding the ability to retrieve information
about a mount namespace via NS_MNT_GET_INFO. This will return the mount
namespace id and the number of mounts currently in the mount namespace.
The number of mounts can be used to size the buffer that needs to be
used for listmount() and is in general useful without having to actually
iterate through all the mounts. The structure is extensible.
And add the ability to iterate through all mount namespaces over which
the caller holds privilege returning the file descriptor for the next or
previous mount namespace.
To retrieve a mount namespace the caller must be privileged wrt to it's
owning user namespace. This means that PID 1 on the host can list all
mounts in all mount namespaces or that a container can list all mounts
of its nested containers.
Optionally pass a structure for NS_MNT_GET_INFO with
NS_MNT_GET_{PREV,NEXT} to retrieve information about the mount namespace
in one go. Both ioctls can be implemented for other namespace types
easily.
Together with recent api additions this means one can iterate through
all mounts in all mount namespaces without ever touching procfs.
Link: https://lore.kernel.org/r/20240719-work-mount-namespace-v1-5-834113cab0d2@kernel.org
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
-----BEGIN PGP SIGNATURE-----
iHUEABYKAB0WIQRAhzRXHqcMeLMyaSiRxhvAZXjcogUCZpEHCgAKCRCRxhvAZXjc
om+gAQCC4nJqJqs9bJZIItRtZ71GnxZQO3HVohhIlNM2KKh0VgEA47JhD0O0Srfb
CleII4cQWqB32BfB/vMeff6hpNa7SA4=
=7ltk
-----END PGP SIGNATURE-----
Merge tag 'vfs-6.11.mount' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs
Pull vfs mount query updates from Christian Brauner:
"This contains work to extend the abilities of listmount() and
statmount() and various fixes and cleanups.
Features:
- Allow iterating through mounts via listmount() from newest to
oldest. This makes it possible for mount(8) to keep iterating the
mount table in reverse order so it gets newest mounts first.
- Relax permissions on listmount() and statmount().
It's not necessary to have capabilities in the initial namespace:
it is sufficient to have capabilities in the owning namespace of
the mount namespace we're located in to list unreachable mounts in
that namespace.
- Extend both listmount() and statmount() to list and stat mounts in
foreign mount namespaces.
Currently the only way to iterate over mount entries in mount
namespaces that aren't in the caller's mount namespace is by
crawling through /proc in order to find /proc/<pid>/mountinfo for
the relevant mount namespace.
This is both very clumsy and hugely inefficient. So extend struct
mnt_id_req with a new member that allows to specify the mount
namespace id of the mount namespace we want to look at.
Luckily internally we already have most of the infrastructure for
this so we just need to expose it to userspace. Give userspace a
way to retrieve the id of a mount namespace via statmount() and
through a new nsfs ioctl() on mount namespace file descriptor.
This comes with appropriate selftests.
- Expose mount options through statmount().
Currently if userspace wants to get mount options for a mount and
with statmount(), they still have to open /proc/<pid>/mountinfo to
parse mount options. Simply the information through statmount()
directly.
Afterwards it's possible to only rely on statmount() and
listmount() to retrieve all and more information than
/proc/<pid>/mountinfo provides.
This comes with appropriate selftests.
Fixes:
- Avoid copying to userspace under the namespace semaphore in
listmount.
Cleanups:
- Simplify the error handling in listmount by relying on our newly
added cleanup infrastructure.
- Refuse invalid mount ids early for both listmount and statmount"
* tag 'vfs-6.11.mount' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs:
fs: reject invalid last mount id early
fs: refuse mnt id requests with invalid ids early
fs: find rootfs mount of the mount namespace
fs: only copy to userspace on success in listmount()
sefltests: extend the statmount test for mount options
fs: use guard for namespace_sem in statmount()
fs: export mount options via statmount()
fs: rename show_mnt_opts -> show_vfsmnt_opts
selftests: add a test for the foreign mnt ns extensions
fs: add an ioctl to get the mnt ns id from nsfs
fs: Allow statmount() in foreign mount namespace
fs: Allow listmount() in foreign mount namespace
fs: export the mount ns id via statmount
fs: keep an index of current mount namespaces
fs: relax permissions for statmount()
listmount: allow listing in reverse order
fs: relax permissions for listmount()
fs: simplify error handling
fs: don't copy to userspace under namespace semaphore
path: add cleanup helper
In order to allow for listmount() to be used on different namespaces we
need a way to lookup a mount ns by its id. Keep a rbtree of the current
!anonymous mount name spaces indexed by ID that we can use to look up
the namespace.
Co-developed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Link: https://lore.kernel.org/r/e5fdd78a90f5b00a75bd893962a70f52a2c015cd.1719243756.git.josef@toxicpanda.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
A current limitation of open_by_handle_at() is that it's currently not possible
to use it from within containers at all because we require CAP_DAC_READ_SEARCH
in the initial namespace. That's unfortunate because there are scenarios where
using open_by_handle_at() from within containers.
Two examples:
(1) cgroupfs allows to encode cgroups to file handles and reopen them with
open_by_handle_at().
(2) Fanotify allows placing filesystem watches they currently aren't usable in
containers because the returned file handles cannot be used.
Here's a proposal for relaxing the permission check for open_by_handle_at().
(1) Opening file handles when the caller has privileges over the filesystem
(1.1) The caller has an unobstructed view of the filesystem.
(1.2) The caller has permissions to follow a path to the file handle.
This doesn't address the problem of opening a file handle when only a portion
of a filesystem is exposed as is common in containers by e.g., bind-mounting a
subtree. The proposal to solve this use-case is:
(2) Opening file handles when the caller has privileges over a subtree
(2.1) The caller is able to reach the file from the provided mount fd.
(2.2) The caller has permissions to construct an unobstructed path to the
file handle.
(2.3) The caller has permissions to follow a path to the file handle.
The relaxed permission checks are currently restricted to directory file
handles which are what both cgroupfs and fanotify need. Handling disconnected
non-directory file handles would lead to a potentially non-deterministic api.
If a disconnected non-directory file handle is provided we may fail to decode
a valid path that we could use for permission checking. That in itself isn't a
problem as we would just return EACCES in that case. However, confusion may
arise if a non-disconnected dentry ends up in the cache later and those opening
the file handle would suddenly succeed.
* It's potentially possible to use timing information (side-channel) to infer
whether a given inode exists. I don't think that's particularly
problematic. Thanks to Jann for bringing this to my attention.
* An unrelated note (IOW, these are thoughts that apply to
open_by_handle_at() generically and are unrelated to the changes here):
Jann pointed out that we should verify whether deleted files could
potentially be reopened through open_by_handle_at(). I don't think that's
possible though.
Another potential thing to check is whether open_by_handle_at() could be
abused to open internal stuff like memfds or gpu stuff. I don't think so
but I haven't had the time to completely verify this.
This dates back to discussions Amir and I had quite some time ago and thanks to
him for providing a lot of details around the export code and related patches!
Link: https://lore.kernel.org/r/20240524-vfs-open_by_handle_at-v1-1-3d4b7d22736b@kernel.org
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
When adding a mount to a namespace insert it into an rbtree rooted in the
mnt_namespace instead of a linear list.
The mnt.mnt_list is still used to set up the mount tree and for
propagation, but not after the mount has been added to a namespace. Hence
mnt_list can live in union with rb_node. Use MNT_ONRB mount flag to
validate that the mount is on the correct list.
This allows removing the cursor used for reading /proc/$PID/mountinfo. The
mnt_id_unique of the next mount can be used as an index into the seq file.
Tested by inserting 100k bind mounts, unsharing the mount namespace, and
unmounting. No performance regressions have been observed.
For the last mount in the 100k list the statmount() call was more than 100x
faster due to the mount ID lookup not having to do a linear search. This
patch makes the overhead of mount ID lookup non-observable in this range.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Link: https://lore.kernel.org/r/20231025140205.3586473-3-mszeredi@redhat.com
Reviewed-by: Ian Kent <raven@themaw.net>
Signed-off-by: Christian Brauner <brauner@kernel.org>
If a mount is released then its mnt_id can immediately be reused. This is
bad news for user interfaces that want to uniquely identify a mount.
Implementing a unique mount ID is trivial (use a 64bit counter).
Unfortunately userspace assumes 32bit size and would overflow after the
counter reaches 2^32.
Introduce a new 64bit ID alongside the old one. Initialize the counter to
2^32, this guarantees that the old and new IDs are never mixed up.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Link: https://lore.kernel.org/r/20231025140205.3586473-2-mszeredi@redhat.com
Reviewed-by: Ian Kent <raven@themaw.net>
Signed-off-by: Christian Brauner <brauner@kernel.org>
The tricky case (__legitimize_mnt() failing after having grabbed
a reference) can be trivially dealt with by leaving nd->path.mnt
non-NULL, for terminate_walk() to drop it.
legitimize_mnt() becomes static after that.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The lock_mount_hash() and unlock_mount_hash() helpers are never called
outside a single file. Remove them from the header and make them static
to reflect this fact. There's no need to have them callable from other
places right now, as Christoph observed.
Link: https://lore.kernel.org/r/20210121131959.646623-31-christian.brauner@ubuntu.com
Cc: David Howells <dhowells@redhat.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: linux-fsdevel@vger.kernel.org
Suggested-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
Switch over mount namespaces to use the newly introduced common lifetime
counter.
Currently every namespace type has its own lifetime counter which is stored
in the specific namespace struct. The lifetime counters are used
identically for all namespaces types. Namespaces may of course have
additional unrelated counters and these are not altered.
This introduces a common lifetime counter into struct ns_common. The
ns_common struct encompasses information that all namespaces share. That
should include the lifetime counter since its common for all of them.
It also allows us to unify the type of the counters across all namespaces.
Most of them use refcount_t but one uses atomic_t and at least one uses
kref. Especially the last one doesn't make much sense since it's just a
wrapper around refcount_t since 2016 and actually complicates cleanup
operations by having to use container_of() to cast the correct namespace
struct out of struct ns_common.
Having the lifetime counter for the namespaces in one place reduces
maintenance cost. Not just because after switching all namespaces over we
will have removed more code than we added but also because the logic is
more easily understandable and we indicate to the user that the basic
lifetime requirements for all namespaces are currently identical.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Acked-by: Christian Brauner <christian.brauner@ubuntu.com>
Link: https://lore.kernel.org/r/159644980287.604812.761686947449081169.stgit@localhost.localdomain
Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
If mounts are deleted after a read(2) call on /proc/self/mounts (or its
kin), the subsequent read(2) could miss a mount that comes after the
deleted one in the list. This is because the file position is interpreted
as the number mount entries from the start of the list.
E.g. first read gets entries #0 to #9; the seq file index will be 10. Then
entry #5 is deleted, resulting in #10 becoming #9 and #11 becoming #10,
etc... The next read will continue from entry #10, and #9 is missed.
Solve this by adding a cursor entry for each open instance. Taking the
global namespace_sem for write seems excessive, since we are only dealing
with a per-namespace list. Instead add a per-namespace spinlock and use
that together with namespace_sem taken for read to protect against
concurrent modification of the mount list. This may reduce parallelism of
is_local_mountpoint(), but it's hardly a big contention point. We could
also use RCU freeing of cursors to make traversal not need additional
locks, if that turns out to be neceesary.
Only move the cursor once for each read (cursor is not added on open) to
minimize cacheline invalidation. When EOF is reached, the cursor is taken
off the list, in order to prevent an excessive number of cursors due to
inactive open file descriptors.
Reported-by: Karel Zak <kzak@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
