2019-05-19 13:08:55 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-05 19:15:50 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-20 11:39:11 -07:00
|
|
|
|
2005-08-08 13:52:16 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-11-16 23:57:37 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2018-10-30 15:09:49 -07:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
|
|
|
|
|
2013-08-28 10:18:00 +10:00
|
|
|
|
2006-09-30 20:52:18 +02:00
|
|
|
|
2011-11-23 19:26:23 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-06-04 12:30:02 -07:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
|
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
|
|
|
|
|
2013-08-28 10:17:55 +10:00
|
|
|
|
2011-01-07 17:49:31 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
2011-01-07 17:49:33 +11:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
|
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2011-01-07 17:49:38 +11:00
|
|
|
|
2013-08-28 10:17:55 +10:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
VFS: don't keep disconnected dentries on d_anon
The original purpose of the per-superblock d_anon list was to
keep disconnected dentries in the cache between consecutive
requests to the NFS server. Dentries can be disconnected if
a client holds a file open and repeatedly performs IO on it,
and if the server drops the dentry, whether due to memory
pressure, server restart, or "echo 3 > /proc/sys/vm/drop_caches".
This purpose was thwarted by commit 75a6f82a0d10 ("freeing unlinked
file indefinitely delayed") which caused disconnected dentries
to be freed as soon as their refcount reached zero.
This means that, when a dentry being used by nfsd gets disconnected, a
new one needs to be allocated for every request (unless requests
overlap). As the dentry has no name, no parent, and no children,
there is little of value to cache. As small memory allocations are
typically fast (from per-cpu free lists) this likely has little cost.
This means that the original purpose of s_anon is no longer relevant:
there is no longer any need to keep disconnected dentries on a list so
they appear to be hashed.
However, s_anon now has a new use. When you mount an NFS filesystem,
the dentry stored in s_root is just a placebo. The "real" root dentry
is allocated using d_obtain_root() and so it kept on the s_anon list.
I don't know the reason for this, but suspect it related to NFSv4
where a mount of "server:/some/path" require NFS to look up the root
filehandle on the server, then walk down "/some" and "/path" to get
the filehandle to mount.
Whatever the reason, NFS depends on the s_anon list and on
shrink_dcache_for_umount() pruning all dentries on this list. So we
cannot simply remove s_anon.
We could just leave the code unchanged, but apart from that being
potentially confusing, the (unfair) bit-spin-lock which protects
s_anon can become a bottle neck when lots of disconnected dentries are
being created.
So this patch renames s_anon to s_roots, and stops storing
disconnected dentries on the list. Only dentries obtained with
d_obtain_root() are now stored on this list. There are many fewer of
these (only NFS and NILFS2 use the call, and only during filesystem
mount) so contention on the bit-lock will not be a problem.
Possibly an alternate solution should be found for NFS and NILFS2, but
that would require understanding their needs first.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2017-12-21 09:45:40 +11:00
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
2011-01-07 17:49:33 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
2025-01-11 15:07:47 +08:00
|
|
|
|
2025-05-11 16:36:24 +08:00
|
|
|
|
2025-01-11 15:07:47 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-11 16:36:24 +08:00
|
|
|
|
2025-01-11 15:07:47 +08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
[PATCH] audit: watching subtrees
New kind of audit rule predicates: "object is visible in given subtree".
The part that can be sanely implemented, that is. Limitations:
* if you have hardlink from outside of tree, you'd better watch
it too (or just watch the object itself, obviously)
* if you mount something under a watched tree, tell audit
that new chunk should be added to watched subtrees
* if you umount something in a watched tree and it's still mounted
elsewhere, you will get matches on events happening there. New command
tells audit to recalculate the trees, trimming such sources of false
positives.
Note that it's _not_ about path - if something mounted in several places
(multiple mount, bindings, different namespaces, etc.), the match does
_not_ depend on which one we are using for access.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2007-07-22 08:04:18 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2011-01-07 17:49:37 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-11-05 16:36:22 +01:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-07-04 17:25:22 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-04-15 19:46:50 -04:00
|
|
|
|
|
|
|
|
|
2017-07-04 17:25:22 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-29 11:20:49 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2024-08-29 11:20:49 -07:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
2024-08-29 11:20:49 -07:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
vfs: dcache: move hashlen_hash() from callers into d_hash()
Both __d_lookup_rcu() and __d_lookup_rcu_op_compare() have the full
'name_hash' value of the qstr that they want to look up, and mask it off
to just the low 32-bit hash before calling down to d_hash().
Other callers just load the 32-bit hash and pass it as the argument.
If we move the masking into d_hash() itself, it simplifies the two
callers that currently do the masking, and is a no-op for the other
cases. It doesn't actually change the generated code since the compiler
will inline d_hash() and see that the end result is the same.
[ Technically, since the parse tree changes, the code generation may not
be 100% the same, and for me on x86-64, this does result in gcc
switching the operands around for one 'cmpl' instruction. So not
necessarily the exact same code generation, but equivalent ]
However, this does encapsulate the 'd_hash()' operation more, and makes
the shift operation in particular be a "shift 32 bits right, return full
word". Which matches the instruction semantics on both x86-64 and arm64
better, since a 32-bit shift will clear the upper bits.
That makes the next step of introducing a "shift by runtime constant"
more obvious and generates the shift with no extraneous type masking.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2024-06-10 11:47:06 -07:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
2024-06-04 12:30:02 -07:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
fs: bump inode and dentry counters to long
This series reworks our current object cache shrinking infrastructure in
two main ways:
* Noticing that a lot of users copy and paste their own version of LRU
lists for objects, we put some effort in providing a generic version.
It is modeled after the filesystem users: dentries, inodes, and xfs
(for various tasks), but we expect that other users could benefit in
the near future with little or no modification. Let us know if you
have any issues.
* The underlying list_lru being proposed automatically and
transparently keeps the elements in per-node lists, and is able to
manipulate the node lists individually. Given this infrastructure, we
are able to modify the up-to-now hammer called shrink_slab to proceed
with node-reclaim instead of always searching memory from all over like
it has been doing.
Per-node lru lists are also expected to lead to less contention in the lru
locks on multi-node scans, since we are now no longer fighting for a
global lock. The locks usually disappear from the profilers with this
change.
Although we have no official benchmarks for this version - be our guest to
independently evaluate this - earlier versions of this series were
performance tested (details at
http://permalink.gmane.org/gmane.linux.kernel.mm/100537) yielding no
visible performance regressions while yielding a better qualitative
behavior in NUMA machines.
With this infrastructure in place, we can use the list_lru entry point to
provide memcg isolation and per-memcg targeted reclaim. Historically,
those two pieces of work have been posted together. This version presents
only the infrastructure work, deferring the memcg work for a later time,
so we can focus on getting this part tested. You can see more about the
history of such work at http://lwn.net/Articles/552769/
Dave Chinner (18):
dcache: convert dentry_stat.nr_unused to per-cpu counters
dentry: move to per-sb LRU locks
dcache: remove dentries from LRU before putting on dispose list
mm: new shrinker API
shrinker: convert superblock shrinkers to new API
list: add a new LRU list type
inode: convert inode lru list to generic lru list code.
dcache: convert to use new lru list infrastructure
list_lru: per-node list infrastructure
shrinker: add node awareness
fs: convert inode and dentry shrinking to be node aware
xfs: convert buftarg LRU to generic code
xfs: rework buffer dispose list tracking
xfs: convert dquot cache lru to list_lru
fs: convert fs shrinkers to new scan/count API
drivers: convert shrinkers to new count/scan API
shrinker: convert remaining shrinkers to count/scan API
shrinker: Kill old ->shrink API.
Glauber Costa (7):
fs: bump inode and dentry counters to long
super: fix calculation of shrinkable objects for small numbers
list_lru: per-node API
vmscan: per-node deferred work
i915: bail out earlier when shrinker cannot acquire mutex
hugepage: convert huge zero page shrinker to new shrinker API
list_lru: dynamically adjust node arrays
This patch:
There are situations in very large machines in which we can have a large
quantity of dirty inodes, unused dentries, etc. This is particularly true
when umounting a filesystem, where eventually since every live object will
eventually be discarded.
Dave Chinner reported a problem with this while experimenting with the
shrinker revamp patchset. So we believe it is time for a change. This
patch just moves int to longs. Machines where it matters should have a
big long anyway.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:17:53 +10:00
|
|
|
|
2013-08-28 10:17:54 +10:00
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
2024-09-29 20:28:31 +08:00
|
|
|
|
2010-10-10 05:36:23 -04:00
|
|
|
|
|
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-08-28 10:17:54 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: bump inode and dentry counters to long
This series reworks our current object cache shrinking infrastructure in
two main ways:
* Noticing that a lot of users copy and paste their own version of LRU
lists for objects, we put some effort in providing a generic version.
It is modeled after the filesystem users: dentries, inodes, and xfs
(for various tasks), but we expect that other users could benefit in
the near future with little or no modification. Let us know if you
have any issues.
* The underlying list_lru being proposed automatically and
transparently keeps the elements in per-node lists, and is able to
manipulate the node lists individually. Given this infrastructure, we
are able to modify the up-to-now hammer called shrink_slab to proceed
with node-reclaim instead of always searching memory from all over like
it has been doing.
Per-node lru lists are also expected to lead to less contention in the lru
locks on multi-node scans, since we are now no longer fighting for a
global lock. The locks usually disappear from the profilers with this
change.
Although we have no official benchmarks for this version - be our guest to
independently evaluate this - earlier versions of this series were
performance tested (details at
http://permalink.gmane.org/gmane.linux.kernel.mm/100537) yielding no
visible performance regressions while yielding a better qualitative
behavior in NUMA machines.
With this infrastructure in place, we can use the list_lru entry point to
provide memcg isolation and per-memcg targeted reclaim. Historically,
those two pieces of work have been posted together. This version presents
only the infrastructure work, deferring the memcg work for a later time,
so we can focus on getting this part tested. You can see more about the
history of such work at http://lwn.net/Articles/552769/
Dave Chinner (18):
dcache: convert dentry_stat.nr_unused to per-cpu counters
dentry: move to per-sb LRU locks
dcache: remove dentries from LRU before putting on dispose list
mm: new shrinker API
shrinker: convert superblock shrinkers to new API
list: add a new LRU list type
inode: convert inode lru list to generic lru list code.
dcache: convert to use new lru list infrastructure
list_lru: per-node list infrastructure
shrinker: add node awareness
fs: convert inode and dentry shrinking to be node aware
xfs: convert buftarg LRU to generic code
xfs: rework buffer dispose list tracking
xfs: convert dquot cache lru to list_lru
fs: convert fs shrinkers to new scan/count API
drivers: convert shrinkers to new count/scan API
shrinker: convert remaining shrinkers to count/scan API
shrinker: Kill old ->shrink API.
Glauber Costa (7):
fs: bump inode and dentry counters to long
super: fix calculation of shrinkable objects for small numbers
list_lru: per-node API
vmscan: per-node deferred work
i915: bail out earlier when shrinker cannot acquire mutex
hugepage: convert huge zero page shrinker to new shrinker API
list_lru: dynamically adjust node arrays
This patch:
There are situations in very large machines in which we can have a large
quantity of dirty inodes, unused dentries, etc. This is particularly true
when umounting a filesystem, where eventually since every live object will
eventually be discarded.
Dave Chinner reported a problem with this while experimenting with the
shrinker revamp patchset. So we believe it is time for a change. This
patch just moves int to longs. Machines where it matters should have a
big long anyway.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:17:53 +10:00
|
|
|
|
fs: use fast counters for vfs caches
percpu_counter library generates quite nasty code, so unless you need
to dynamically allocate counters or take fast approximate value, a
simple per cpu set of counters is much better.
The percpu_counter can never be made to work as well, because it has an
indirection from pointer to percpu memory, and it can't use direct
this_cpu_inc interfaces because it doesn't use static PER_CPU data, so
code will always be worse.
In the fastpath, it is the difference between this:
incl %gs:nr_dentry # nr_dentry
and this:
movl percpu_counter_batch(%rip), %edx # percpu_counter_batch,
movl $1, %esi #,
movq $nr_dentry, %rdi #,
call __percpu_counter_add # (plus I clobber registers)
__percpu_counter_add:
pushq %rbp #
movq %rsp, %rbp #,
subq $32, %rsp #,
movq %rbx, -24(%rbp) #,
movq %r12, -16(%rbp) #,
movq %r13, -8(%rbp) #,
movq %rdi, %rbx # fbc, fbc
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
incl -8124(%rax) # <variable>.preempt_count
movq 32(%rdi), %r12 # <variable>.counters, tcp_ptr__
#APP
# 78 "lib/percpu_counter.c" 1
add %gs:this_cpu_off, %r12 # this_cpu_off, tcp_ptr__
# 0 "" 2
#NO_APP
movslq (%r12),%r13 #* tcp_ptr__, tmp73
movslq %edx,%rax # batch, batch
addq %rsi, %r13 # amount, count
cmpq %rax, %r13 # batch, count
jge .L27 #,
negl %edx # tmp76
movslq %edx,%rdx # tmp76, tmp77
cmpq %rdx, %r13 # tmp77, count
jg .L28 #,
.L27:
movq %rbx, %rdi # fbc,
call _raw_spin_lock #
addq %r13, 8(%rbx) # count, <variable>.count
movq %rbx, %rdi # fbc,
movl $0, (%r12) #,* tcp_ptr__
call _raw_spin_unlock #
.L29:
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
decl -8124(%rax) # <variable>.preempt_count
movq -8136(%rax), %rax #, D.14625
testb $8, %al #, D.14625
jne .L32 #,
.L31:
movq -24(%rbp), %rbx #,
movq -16(%rbp), %r12 #,
movq -8(%rbp), %r13 #,
leave
ret
.p2align 4,,10
.p2align 3
.L28:
movl %r13d, (%r12) # count,*
jmp .L29 #
.L32:
call preempt_schedule #
.p2align 4,,6
jmp .L31 #
.size __percpu_counter_add, .-__percpu_counter_add
.p2align 4,,15
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:19 +11:00
|
|
|
|
|
|
|
|
|
fs: bump inode and dentry counters to long
This series reworks our current object cache shrinking infrastructure in
two main ways:
* Noticing that a lot of users copy and paste their own version of LRU
lists for objects, we put some effort in providing a generic version.
It is modeled after the filesystem users: dentries, inodes, and xfs
(for various tasks), but we expect that other users could benefit in
the near future with little or no modification. Let us know if you
have any issues.
* The underlying list_lru being proposed automatically and
transparently keeps the elements in per-node lists, and is able to
manipulate the node lists individually. Given this infrastructure, we
are able to modify the up-to-now hammer called shrink_slab to proceed
with node-reclaim instead of always searching memory from all over like
it has been doing.
Per-node lru lists are also expected to lead to less contention in the lru
locks on multi-node scans, since we are now no longer fighting for a
global lock. The locks usually disappear from the profilers with this
change.
Although we have no official benchmarks for this version - be our guest to
independently evaluate this - earlier versions of this series were
performance tested (details at
http://permalink.gmane.org/gmane.linux.kernel.mm/100537) yielding no
visible performance regressions while yielding a better qualitative
behavior in NUMA machines.
With this infrastructure in place, we can use the list_lru entry point to
provide memcg isolation and per-memcg targeted reclaim. Historically,
those two pieces of work have been posted together. This version presents
only the infrastructure work, deferring the memcg work for a later time,
so we can focus on getting this part tested. You can see more about the
history of such work at http://lwn.net/Articles/552769/
Dave Chinner (18):
dcache: convert dentry_stat.nr_unused to per-cpu counters
dentry: move to per-sb LRU locks
dcache: remove dentries from LRU before putting on dispose list
mm: new shrinker API
shrinker: convert superblock shrinkers to new API
list: add a new LRU list type
inode: convert inode lru list to generic lru list code.
dcache: convert to use new lru list infrastructure
list_lru: per-node list infrastructure
shrinker: add node awareness
fs: convert inode and dentry shrinking to be node aware
xfs: convert buftarg LRU to generic code
xfs: rework buffer dispose list tracking
xfs: convert dquot cache lru to list_lru
fs: convert fs shrinkers to new scan/count API
drivers: convert shrinkers to new count/scan API
shrinker: convert remaining shrinkers to count/scan API
shrinker: Kill old ->shrink API.
Glauber Costa (7):
fs: bump inode and dentry counters to long
super: fix calculation of shrinkable objects for small numbers
list_lru: per-node API
vmscan: per-node deferred work
i915: bail out earlier when shrinker cannot acquire mutex
hugepage: convert huge zero page shrinker to new shrinker API
list_lru: dynamically adjust node arrays
This patch:
There are situations in very large machines in which we can have a large
quantity of dirty inodes, unused dentries, etc. This is particularly true
when umounting a filesystem, where eventually since every live object will
eventually be discarded.
Dave Chinner reported a problem with this while experimenting with the
shrinker revamp patchset. So we believe it is time for a change. This
patch just moves int to longs. Machines where it matters should have a
big long anyway.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:17:53 +10:00
|
|
|
|
fs: use fast counters for vfs caches
percpu_counter library generates quite nasty code, so unless you need
to dynamically allocate counters or take fast approximate value, a
simple per cpu set of counters is much better.
The percpu_counter can never be made to work as well, because it has an
indirection from pointer to percpu memory, and it can't use direct
this_cpu_inc interfaces because it doesn't use static PER_CPU data, so
code will always be worse.
In the fastpath, it is the difference between this:
incl %gs:nr_dentry # nr_dentry
and this:
movl percpu_counter_batch(%rip), %edx # percpu_counter_batch,
movl $1, %esi #,
movq $nr_dentry, %rdi #,
call __percpu_counter_add # (plus I clobber registers)
__percpu_counter_add:
pushq %rbp #
movq %rsp, %rbp #,
subq $32, %rsp #,
movq %rbx, -24(%rbp) #,
movq %r12, -16(%rbp) #,
movq %r13, -8(%rbp) #,
movq %rdi, %rbx # fbc, fbc
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
incl -8124(%rax) # <variable>.preempt_count
movq 32(%rdi), %r12 # <variable>.counters, tcp_ptr__
#APP
# 78 "lib/percpu_counter.c" 1
add %gs:this_cpu_off, %r12 # this_cpu_off, tcp_ptr__
# 0 "" 2
#NO_APP
movslq (%r12),%r13 #* tcp_ptr__, tmp73
movslq %edx,%rax # batch, batch
addq %rsi, %r13 # amount, count
cmpq %rax, %r13 # batch, count
jge .L27 #,
negl %edx # tmp76
movslq %edx,%rdx # tmp76, tmp77
cmpq %rdx, %r13 # tmp77, count
jg .L28 #,
.L27:
movq %rbx, %rdi # fbc,
call _raw_spin_lock #
addq %r13, 8(%rbx) # count, <variable>.count
movq %rbx, %rdi # fbc,
movl $0, (%r12) #,* tcp_ptr__
call _raw_spin_unlock #
.L29:
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
decl -8124(%rax) # <variable>.preempt_count
movq -8136(%rax), %rax #, D.14625
testb $8, %al #, D.14625
jne .L32 #,
.L31:
movq -24(%rbp), %rbx #,
movq -16(%rbp), %r12 #,
movq -8(%rbp), %r13 #,
leave
ret
.p2align 4,,10
.p2align 3
.L28:
movl %r13d, (%r12) # count,*
jmp .L29 #
.L32:
call preempt_schedule #
.p2align 4,,6
jmp .L31 #
.size __percpu_counter_add, .-__percpu_counter_add
.p2align 4,,15
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:19 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-08-28 10:17:54 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
sysctl: treewide: constify the ctl_table argument of proc_handlers
const qualify the struct ctl_table argument in the proc_handler function
signatures. This is a prerequisite to moving the static ctl_table
structs into .rodata data which will ensure that proc_handler function
pointers cannot be modified.
This patch has been generated by the following coccinelle script:
```
virtual patch
@r1@
identifier ctl, write, buffer, lenp, ppos;
identifier func !~ "appldata_(timer|interval)_handler|sched_(rt|rr)_handler|rds_tcp_skbuf_handler|proc_sctp_do_(hmac_alg|rto_min|rto_max|udp_port|alpha_beta|auth|probe_interval)";
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos);
@r2@
identifier func, ctl, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int write, void *buffer, size_t *lenp, loff_t *ppos)
{ ... }
@r3@
identifier func;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int , void *, size_t *, loff_t *);
@r4@
identifier func, ctl;
@@
int func(
- struct ctl_table *ctl
+ const struct ctl_table *ctl
,int , void *, size_t *, loff_t *);
@r5@
identifier func, write, buffer, lenp, ppos;
@@
int func(
- struct ctl_table *
+ const struct ctl_table *
,int write, void *buffer, size_t *lenp, loff_t *ppos);
```
* Code formatting was adjusted in xfs_sysctl.c to comply with code
conventions. The xfs_stats_clear_proc_handler,
xfs_panic_mask_proc_handler and xfs_deprecated_dointvec_minmax where
adjusted.
* The ctl_table argument in proc_watchdog_common was const qualified.
This is called from a proc_handler itself and is calling back into
another proc_handler, making it necessary to change it as part of the
proc_handler migration.
Co-developed-by: Thomas Weißschuh <linux@weissschuh.net>
Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
Co-developed-by: Joel Granados <j.granados@samsung.com>
Signed-off-by: Joel Granados <j.granados@samsung.com>
2024-07-24 20:59:29 +02:00
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
2010-10-10 05:36:23 -04:00
|
|
|
|
fs: use fast counters for vfs caches
percpu_counter library generates quite nasty code, so unless you need
to dynamically allocate counters or take fast approximate value, a
simple per cpu set of counters is much better.
The percpu_counter can never be made to work as well, because it has an
indirection from pointer to percpu memory, and it can't use direct
this_cpu_inc interfaces because it doesn't use static PER_CPU data, so
code will always be worse.
In the fastpath, it is the difference between this:
incl %gs:nr_dentry # nr_dentry
and this:
movl percpu_counter_batch(%rip), %edx # percpu_counter_batch,
movl $1, %esi #,
movq $nr_dentry, %rdi #,
call __percpu_counter_add # (plus I clobber registers)
__percpu_counter_add:
pushq %rbp #
movq %rsp, %rbp #,
subq $32, %rsp #,
movq %rbx, -24(%rbp) #,
movq %r12, -16(%rbp) #,
movq %r13, -8(%rbp) #,
movq %rdi, %rbx # fbc, fbc
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
incl -8124(%rax) # <variable>.preempt_count
movq 32(%rdi), %r12 # <variable>.counters, tcp_ptr__
#APP
# 78 "lib/percpu_counter.c" 1
add %gs:this_cpu_off, %r12 # this_cpu_off, tcp_ptr__
# 0 "" 2
#NO_APP
movslq (%r12),%r13 #* tcp_ptr__, tmp73
movslq %edx,%rax # batch, batch
addq %rsi, %r13 # amount, count
cmpq %rax, %r13 # batch, count
jge .L27 #,
negl %edx # tmp76
movslq %edx,%rdx # tmp76, tmp77
cmpq %rdx, %r13 # tmp77, count
jg .L28 #,
.L27:
movq %rbx, %rdi # fbc,
call _raw_spin_lock #
addq %r13, 8(%rbx) # count, <variable>.count
movq %rbx, %rdi # fbc,
movl $0, (%r12) #,* tcp_ptr__
call _raw_spin_unlock #
.L29:
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
decl -8124(%rax) # <variable>.preempt_count
movq -8136(%rax), %rax #, D.14625
testb $8, %al #, D.14625
jne .L32 #,
.L31:
movq -24(%rbp), %rbx #,
movq -16(%rbp), %r12 #,
movq -8(%rbp), %r13 #,
leave
ret
.p2align 4,,10
.p2align 3
.L28:
movl %r13d, (%r12) # count,*
jmp .L29 #
.L32:
call preempt_schedule #
.p2align 4,,6
jmp .L31 #
.size __percpu_counter_add, .-__percpu_counter_add
.p2align 4,,15
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:19 +11:00
|
|
|
|
2013-08-28 10:17:54 +10:00
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
fs: bump inode and dentry counters to long
This series reworks our current object cache shrinking infrastructure in
two main ways:
* Noticing that a lot of users copy and paste their own version of LRU
lists for objects, we put some effort in providing a generic version.
It is modeled after the filesystem users: dentries, inodes, and xfs
(for various tasks), but we expect that other users could benefit in
the near future with little or no modification. Let us know if you
have any issues.
* The underlying list_lru being proposed automatically and
transparently keeps the elements in per-node lists, and is able to
manipulate the node lists individually. Given this infrastructure, we
are able to modify the up-to-now hammer called shrink_slab to proceed
with node-reclaim instead of always searching memory from all over like
it has been doing.
Per-node lru lists are also expected to lead to less contention in the lru
locks on multi-node scans, since we are now no longer fighting for a
global lock. The locks usually disappear from the profilers with this
change.
Although we have no official benchmarks for this version - be our guest to
independently evaluate this - earlier versions of this series were
performance tested (details at
http://permalink.gmane.org/gmane.linux.kernel.mm/100537) yielding no
visible performance regressions while yielding a better qualitative
behavior in NUMA machines.
With this infrastructure in place, we can use the list_lru entry point to
provide memcg isolation and per-memcg targeted reclaim. Historically,
those two pieces of work have been posted together. This version presents
only the infrastructure work, deferring the memcg work for a later time,
so we can focus on getting this part tested. You can see more about the
history of such work at http://lwn.net/Articles/552769/
Dave Chinner (18):
dcache: convert dentry_stat.nr_unused to per-cpu counters
dentry: move to per-sb LRU locks
dcache: remove dentries from LRU before putting on dispose list
mm: new shrinker API
shrinker: convert superblock shrinkers to new API
list: add a new LRU list type
inode: convert inode lru list to generic lru list code.
dcache: convert to use new lru list infrastructure
list_lru: per-node list infrastructure
shrinker: add node awareness
fs: convert inode and dentry shrinking to be node aware
xfs: convert buftarg LRU to generic code
xfs: rework buffer dispose list tracking
xfs: convert dquot cache lru to list_lru
fs: convert fs shrinkers to new scan/count API
drivers: convert shrinkers to new count/scan API
shrinker: convert remaining shrinkers to count/scan API
shrinker: Kill old ->shrink API.
Glauber Costa (7):
fs: bump inode and dentry counters to long
super: fix calculation of shrinkable objects for small numbers
list_lru: per-node API
vmscan: per-node deferred work
i915: bail out earlier when shrinker cannot acquire mutex
hugepage: convert huge zero page shrinker to new shrinker API
list_lru: dynamically adjust node arrays
This patch:
There are situations in very large machines in which we can have a large
quantity of dirty inodes, unused dentries, etc. This is particularly true
when umounting a filesystem, where eventually since every live object will
eventually be discarded.
Dave Chinner reported a problem with this while experimenting with the
shrinker revamp patchset. So we believe it is time for a change. This
patch just moves int to longs. Machines where it matters should have a
big long anyway.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Dave Chinner <dchinner@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:17:53 +10:00
|
|
|
|
2010-10-10 05:36:23 -04:00
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
2025-01-28 13:48:37 +01:00
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-29 20:28:31 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
|
|
|
|
|
2025-01-11 15:07:47 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-11 16:36:24 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-01-11 15:07:47 +08:00
|
|
|
|
|
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
|
|
|
|
|
2025-01-11 15:07:47 +08:00
|
|
|
|
2022-01-21 22:12:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-10-10 05:36:23 -04:00
|
|
|
|
|
|
|
|
|
2012-03-04 15:51:42 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-03 10:16:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-10 12:19:19 -07:00
|
|
|
|
2012-03-04 15:51:42 -08:00
|
|
|
|
2012-03-06 11:16:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-02-01 21:00:51 +03:00
|
|
|
|
2012-05-03 10:16:43 -07:00
|
|
|
|
2012-03-06 11:16:17 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-12-12 17:40:21 +00:00
|
|
|
|
2012-03-06 11:16:17 -08:00
|
|
|
|
2012-05-03 10:16:43 -07:00
|
|
|
|
|
|
|
|
|
2012-03-06 11:16:17 -08:00
|
|
|
|
2012-05-03 10:16:43 -07:00
|
|
|
|
2012-05-10 12:19:19 -07:00
|
|
|
|
2012-05-03 10:16:43 -07:00
|
|
|
|
2012-03-04 15:51:42 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-03 10:16:43 -07:00
|
|
|
|
|
|
|
|
|
2012-05-10 12:19:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-10-24 11:22:48 +01:00
|
|
|
|
2012-05-10 12:19:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-10-24 11:22:48 +01:00
|
|
|
|
2016-03-26 09:12:10 +00:00
|
|
|
|
2012-05-21 16:14:04 -07:00
|
|
|
|
2012-05-10 12:19:19 -07:00
|
|
|
|
|
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-10-10 05:36:22 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2010-10-10 05:36:22 -04:00
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2018-10-26 15:05:41 -07:00
|
|
|
|
dcache: account external names as indirectly reclaimable memory
I received a report about suspicious growth of unreclaimable slabs on
some machines. I've found that it happens on machines with low memory
pressure, and these unreclaimable slabs are external names attached to
dentries.
External names are allocated using generic kmalloc() function, so they
are accounted as unreclaimable. But they are held by dentries, which
are reclaimable, and they will be reclaimed under the memory pressure.
In particular, this breaks MemAvailable calculation, as it doesn't take
unreclaimable slabs into account. This leads to a silly situation, when
a machine is almost idle, has no memory pressure and therefore has a big
dentry cache. And the resulting MemAvailable is too low to start a new
workload.
To address the issue, the NR_INDIRECTLY_RECLAIMABLE_BYTES counter is
used to track the amount of memory, consumed by external names. The
counter is increased in the dentry allocation path, if an external name
structure is allocated; and it's decreased in the dentry freeing path.
To reproduce the problem I've used the following Python script:
import os
for iter in range (0, 10000000):
try:
name = ("/some_long_name_%d" % iter) + "_" * 220
os.stat(name)
except Exception:
pass
Without this patch:
$ cat /proc/meminfo | grep MemAvailable
MemAvailable: 7811688 kB
$ python indirect.py
$ cat /proc/meminfo | grep MemAvailable
MemAvailable: 2753052 kB
With the patch:
$ cat /proc/meminfo | grep MemAvailable
MemAvailable: 7809516 kB
$ python indirect.py
$ cat /proc/meminfo | grep MemAvailable
MemAvailable: 7749144 kB
[guro@fb.com: fix indirectly reclaimable memory accounting for CONFIG_SLOB]
Link: http://lkml.kernel.org/r/20180312194140.19517-1-guro@fb.com
[guro@fb.com: fix indirectly reclaimable memory accounting]
Link: http://lkml.kernel.org/r/20180313125701.7955-1-guro@fb.com
Link: http://lkml.kernel.org/r/20180305133743.12746-5-guro@fb.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-04-10 16:27:44 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2014-10-12 12:45:37 -04:00
|
|
|
|
|
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2014-10-12 12:45:37 -04:00
|
|
|
|
|
|
|
|
|
2017-07-07 14:51:19 -04:00
|
|
|
|
|
|
|
|
|
2024-12-08 22:27:51 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2017-07-07 14:51:19 -04:00
|
|
|
|
2024-12-08 22:27:51 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
2024-12-08 22:27:51 -05:00
|
|
|
|
|
|
|
|
|
2017-07-07 14:51:19 -04:00
|
|
|
|
2024-12-08 22:27:51 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-07-07 14:51:19 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2017-07-07 14:51:19 -04:00
|
|
|
|
2019-04-26 13:07:27 -04:00
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
|
|
|
|
|
2017-07-07 14:51:19 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-03-05 14:09:22 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-10 14:32:05 -05:00
|
|
|
|
2015-03-05 14:09:22 +00:00
|
|
|
|
fs/namei.c: fix missing barriers when checking positivity
Pinned negative dentries can, generally, be made positive
by another thread. Conditions that prevent that are
* ->d_lock on dentry in question
* parent directory held at least shared
* nobody else could have observed the address of dentry
Most of the places working with those fall into one of those
categories; however, d_lookup() and friends need to be used
with some care. Fortunately, there's not a lot of call sites,
and with few exceptions all of those fall under one of the
cases above.
Exceptions are all in fs/namei.c - in lookup_fast(), lookup_dcache()
and mountpoint_last(). Another one is lookup_slow() - there
dcache lookup is done with parent held shared, but the result
is used after we'd drop the lock. The same happens in do_last() -
the lookup (in lookup_one()) is done with parent locked, but
result is used after unlocking.
lookup_fast(), do_last() and mountpoint_last() flat-out reject
negatives.
Most of lookup_dcache() calls are made with parent locked at least
shared; the only exception is lookup_one_len_unlocked(). It might
return pinned negative, needs serious care from callers. Fortunately,
almost nobody calls it directly anymore; all but two callers have
converted to lookup_positive_unlocked(), which rejects negatives.
lookup_slow() is called by the same lookup_one_len_unlocked() (see
above), mountpoint_last() and walk_component(). In those two negatives
are rejected.
In other words, there is a small set of places where we need to
check carefully if a pinned potentially negative dentry is, in
fact, positive. After that check we want to be sure that both
->d_inode and type bits in ->d_flags are stable and observed.
The set consists of follow_managed() (where the rejection happens
for lookup_fast(), walk_component() and do_last()), last_mountpoint()
and lookup_positive_unlocked().
Solution:
1) transition from negative to positive (in __d_set_inode_and_type())
stores ->d_inode, then uses smp_store_release() to set ->d_flags type bits.
2) aforementioned 3 places in fs/namei.c fetch ->d_flags with
smp_load_acquire() and bugger off if it type bits say "negative".
That way anyone downstream of those checks has dentry know positive pinned,
with ->d_inode and type bits of ->d_flags stable and observed.
I considered splitting off d_lookup_positive(), so that the checks could
be done right there, under ->d_lock. However, that leads to massive
duplication of rather subtle code in fs/namei.c and fs/dcache.c. It's
worse than it might seem, thanks to autofs ->d_manage() getting involved ;-/
No matter what, autofs_d_manage()/autofs_d_automount() must live with
the possibility of pinned negative dentry passed their way, becoming
positive under them - that's the intended behaviour when lookup comes
in the middle of automount in progress, so we can't keep them out of
the area that has to deal with those, more's the pity...
Reported-by: Ritesh Harjani <riteshh@linux.ibm.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-11-12 16:13:06 -05:00
|
|
|
|
2015-03-05 14:09:22 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-10 14:32:05 -05:00
|
|
|
|
2015-03-05 14:09:22 +00:00
|
|
|
|
|
|
|
|
|
vfs: don't mod negative dentry count when on shrinker list
The nr_dentry_negative counter is intended to only account negative
dentries that are present on the superblock LRU. Therefore, the LRU
add, remove and isolate helpers modify the counter based on whether
the dentry is negative, but the shrinker list related helpers do not
modify the counter, and the paths that change a dentry between
positive and negative only do so if DCACHE_LRU_LIST is set.
The problem with this is that a dentry on a shrinker list still has
DCACHE_LRU_LIST set to indicate ->d_lru is in use. The additional
DCACHE_SHRINK_LIST flag denotes whether the dentry is on LRU or a
shrink related list. Therefore if a relevant operation (i.e. unlink)
occurs while a dentry is present on a shrinker list, and the
associated codepath only checks for DCACHE_LRU_LIST, then it is
technically possible to modify the negative dentry count for a
dentry that is off the LRU. Since the shrinker list related helpers
do not modify the negative dentry count (because non-LRU dentries
should not be included in the count) when the dentry is ultimately
removed from the shrinker list, this can cause the negative dentry
count to become permanently inaccurate.
This problem can be reproduced via a heavy file create/unlink vs.
drop_caches workload. On an 80xcpu system, I start 80 tasks each
running a 1k file create/delete loop, and one task spinning on
drop_caches. After 10 minutes or so of runtime, the idle/clean cache
negative dentry count increases from somewhere in the range of 5-10
entries to several hundred (and increasingly grows beyond
nr_dentry_unused).
Tweak the logic in the paths that turn a dentry negative or positive
to filter out the case where the dentry is present on a shrink
related list. This allows the above workload to maintain an accurate
negative dentry count.
Fixes: af0c9af1b3f6 ("fs/dcache: Track & report number of negative dentries")
Signed-off-by: Brian Foster <bfoster@redhat.com>
Link: https://lore.kernel.org/r/20240703121301.247680-1-bfoster@redhat.com
Acked-by: Ian Kent <ikent@redhat.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Waiman Long <longman@redhat.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-07-03 08:13:01 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
2015-03-05 14:09:22 +00:00
|
|
|
|
|
|
|
|
|
2026-04-22 07:29:48 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-04-29 23:40:14 -04:00
|
|
|
|
|
|
|
|
|
2026-04-22 07:29:48 -04:00
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
2026-04-22 07:29:48 -04:00
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
|
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-04-29 23:40:14 -04:00
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2014-04-29 23:40:14 -04:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2014-04-29 23:40:14 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2016-05-29 20:13:30 -04:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
2016-02-29 12:12:46 -05:00
|
|
|
|
2018-08-09 10:15:54 -04:00
|
|
|
|
2015-03-05 14:09:22 +00:00
|
|
|
|
2026-04-15 19:29:53 -04:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
|
|
|
|
|
2026-04-15 19:29:53 -04:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
|
|
|
|
|
2018-08-09 10:15:54 -04:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
|
|
|
|
|
list_lru: allow explicit memcg and NUMA node selection
Patch series "workload-specific and memory pressure-driven zswap
writeback", v8.
There are currently several issues with zswap writeback:
1. There is only a single global LRU for zswap, making it impossible to
perform worload-specific shrinking - an memcg under memory pressure
cannot determine which pages in the pool it owns, and often ends up
writing pages from other memcgs. This issue has been previously
observed in practice and mitigated by simply disabling
memcg-initiated shrinking:
https://lore.kernel.org/all/20230530232435.3097106-1-nphamcs@gmail.com/T/#u
But this solution leaves a lot to be desired, as we still do not
have an avenue for an memcg to free up its own memory locked up in
the zswap pool.
2. We only shrink the zswap pool when the user-defined limit is hit.
This means that if we set the limit too high, cold data that are
unlikely to be used again will reside in the pool, wasting precious
memory. It is hard to predict how much zswap space will be needed
ahead of time, as this depends on the workload (specifically, on
factors such as memory access patterns and compressibility of the
memory pages).
This patch series solves these issues by separating the global zswap LRU
into per-memcg and per-NUMA LRUs, and performs workload-specific (i.e
memcg- and NUMA-aware) zswap writeback under memory pressure. The new
shrinker does not have any parameter that must be tuned by the user, and
can be opted in or out on a per-memcg basis.
As a proof of concept, we ran the following synthetic benchmark: build the
linux kernel in a memory-limited cgroup, and allocate some cold data in
tmpfs to see if the shrinker could write them out and improved the overall
performance. Depending on the amount of cold data generated, we observe
from 14% to 35% reduction in kernel CPU time used in the kernel builds.
This patch (of 6):
The interface of list_lru is based on the assumption that the list node
and the data it represents belong to the same allocated on the correct
node/memcg. While this assumption is valid for existing slab objects LRU
such as dentries and inodes, it is undocumented, and rather inflexible for
certain potential list_lru users (such as the upcoming zswap shrinker and
the THP shrinker). It has caused us a lot of issues during our
development.
This patch changes list_lru interface so that the caller must explicitly
specify numa node and memcg when adding and removing objects. The old
list_lru_add() and list_lru_del() are renamed to list_lru_add_obj() and
list_lru_del_obj(), respectively.
It also extends the list_lru API with a new function, list_lru_putback,
which undoes a previous list_lru_isolate call. Unlike list_lru_add, it
does not increment the LRU node count (as list_lru_isolate does not
decrement the node count). list_lru_putback also allows for explicit
memcg and NUMA node selection.
Link: https://lkml.kernel.org/r/20231130194023.4102148-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231130194023.4102148-2-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Bagas Sanjaya <bagasdotme@gmail.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-30 11:40:18 -08:00
|
|
|
|
|
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
|
|
|
|
|
list_lru: allow explicit memcg and NUMA node selection
Patch series "workload-specific and memory pressure-driven zswap
writeback", v8.
There are currently several issues with zswap writeback:
1. There is only a single global LRU for zswap, making it impossible to
perform worload-specific shrinking - an memcg under memory pressure
cannot determine which pages in the pool it owns, and often ends up
writing pages from other memcgs. This issue has been previously
observed in practice and mitigated by simply disabling
memcg-initiated shrinking:
https://lore.kernel.org/all/20230530232435.3097106-1-nphamcs@gmail.com/T/#u
But this solution leaves a lot to be desired, as we still do not
have an avenue for an memcg to free up its own memory locked up in
the zswap pool.
2. We only shrink the zswap pool when the user-defined limit is hit.
This means that if we set the limit too high, cold data that are
unlikely to be used again will reside in the pool, wasting precious
memory. It is hard to predict how much zswap space will be needed
ahead of time, as this depends on the workload (specifically, on
factors such as memory access patterns and compressibility of the
memory pages).
This patch series solves these issues by separating the global zswap LRU
into per-memcg and per-NUMA LRUs, and performs workload-specific (i.e
memcg- and NUMA-aware) zswap writeback under memory pressure. The new
shrinker does not have any parameter that must be tuned by the user, and
can be opted in or out on a per-memcg basis.
As a proof of concept, we ran the following synthetic benchmark: build the
linux kernel in a memory-limited cgroup, and allocate some cold data in
tmpfs to see if the shrinker could write them out and improved the overall
performance. Depending on the amount of cold data generated, we observe
from 14% to 35% reduction in kernel CPU time used in the kernel builds.
This patch (of 6):
The interface of list_lru is based on the assumption that the list node
and the data it represents belong to the same allocated on the correct
node/memcg. While this assumption is valid for existing slab objects LRU
such as dentries and inodes, it is undocumented, and rather inflexible for
certain potential list_lru users (such as the upcoming zswap shrinker and
the THP shrinker). It has caused us a lot of issues during our
development.
This patch changes list_lru interface so that the caller must explicitly
specify numa node and memcg when adding and removing objects. The old
list_lru_add() and list_lru_del() are renamed to list_lru_add_obj() and
list_lru_del_obj(), respectively.
It also extends the list_lru API with a new function, list_lru_putback,
which undoes a previous list_lru_isolate call. Unlike list_lru_add, it
does not increment the LRU node count (as list_lru_isolate does not
decrement the node count). list_lru_putback also allows for explicit
memcg and NUMA node selection.
Link: https://lkml.kernel.org/r/20231130194023.4102148-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231130194023.4102148-2-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Bagas Sanjaya <bagasdotme@gmail.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-30 11:40:18 -08:00
|
|
|
|
|
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
|
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
|
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
|
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
2013-09-13 22:55:10 -04:00
|
|
|
|
|
|
|
|
|
2017-11-10 15:45:41 +11:00
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
2013-10-04 11:09:01 -04:00
|
|
|
|
2018-03-07 00:49:10 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
2017-11-10 15:45:41 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-07 00:49:10 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-10 15:45:41 +11:00
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
|
|
|
|
|
2021-01-14 09:04:39 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:30 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-16 11:50:41 -04:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
2026-04-16 11:50:41 -04:00
|
|
|
|
|
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
2026-04-16 11:50:41 -04:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
2026-04-16 11:50:41 -04:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
|
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 04:17:02 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 16:17:19 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
2026-04-11 16:17:19 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 04:17:02 -04:00
|
|
|
|
2011-01-07 17:49:48 +11:00
|
|
|
|
dentry_kill(): don't try to remove from shrink list
If the victim in on the shrink list, don't remove it from there.
If shrink_dentry_list() manages to remove it from the list before
we are done - fine, we'll just free it as usual. If not - mark
it with new flag (DCACHE_MAY_FREE) and leave it there.
Eventually, shrink_dentry_list() will get to it, remove the sucker
from shrink list and call dentry_kill(dentry, 0). Which is where
we'll deal with freeing.
Since now dentry_kill(dentry, 0) may happen after or during
dentry_kill(dentry, 1), we need to recognize that (by seeing
DCACHE_DENTRY_KILLED already set), unlock everything
and either free the sucker (in case DCACHE_MAY_FREE has been
set) or leave it for ongoing dentry_kill(dentry, 1) to deal with.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-05-01 10:30:00 -04:00
|
|
|
|
|
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2026-04-11 04:17:02 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vfs: use lockred "dead" flag to mark unrecoverably dead dentries
This simplifies the RCU to refcounting code in particular.
I was originally intending to leave this for later, but walking through
all the dput() logic (see previous commit), I realized that the dput()
"might_sleep()" check was misleadingly weak. And I removed it as
misleading, both for performance profiling and for debugging.
However, the might_sleep() debugging case is actually true: the final
dput() can indeed sleep, if the inode of the dentry that you are
releasing ends up sleeping at iput time (see dentry_iput()). So the
problem with the might_sleep() in dput() wasn't that it wasn't true, it
was that it wasn't actually testing and triggering on the interesting
case.
In particular, just about *any* dput() can indeed sleep, if you happen
to race with another thread deleting the file in question, and you then
lose the race to the be the last dput() for that file. But because it's
a very rare race, the debugging code would never trigger it in practice.
Why is this problematic? The new d_rcu_to_refcount() (see commit
15570086b590: "vfs: reimplement d_rcu_to_refcount() using
lockref_get_or_lock()") does a dput() for the failure case, and it does
it under the RCU lock. So potentially sleeping really is a bug.
But there's no way I'm going to fix this with the previous complicated
"lockref_get_or_lock()" interface. And rather than revert to the old
and crufty nested dentry locking code (which did get this right by
delaying the reference count updates until they were verified to be
safe), let's make forward progress.
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-08 13:46:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-10-28 10:02:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-05-30 11:39:02 -04:00
|
|
|
|
2013-04-15 14:13:21 +08:00
|
|
|
|
|
|
|
|
|
2014-04-29 23:42:52 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:48 +11:00
|
|
|
|
|
|
|
|
|
2016-05-29 20:13:30 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-04-29 15:45:28 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
dentry_kill(): don't try to remove from shrink list
If the victim in on the shrink list, don't remove it from there.
If shrink_dentry_list() manages to remove it from the list before
we are done - fine, we'll just free it as usual. If not - mark
it with new flag (DCACHE_MAY_FREE) and leave it there.
Eventually, shrink_dentry_list() will get to it, remove the sucker
from shrink list and call dentry_kill(dentry, 0). Which is where
we'll deal with freeing.
Since now dentry_kill(dentry, 0) may happen after or during
dentry_kill(dentry, 1), we need to recognize that (by seeing
DCACHE_DENTRY_KILLED already set), unlock everything
and either free the sucker (in case DCACHE_MAY_FREE has been
set) or leave it for ongoing dentry_kill(dentry, 1) to deal with.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-05-01 10:30:00 -04:00
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
|
|
|
|
|
2023-11-10 14:07:43 -05:00
|
|
|
|
dentry_kill(): don't try to remove from shrink list
If the victim in on the shrink list, don't remove it from there.
If shrink_dentry_list() manages to remove it from the list before
we are done - fine, we'll just free it as usual. If not - mark
it with new flag (DCACHE_MAY_FREE) and leave it there.
Eventually, shrink_dentry_list() will get to it, remove the sucker
from shrink list and call dentry_kill(dentry, 0). Which is where
we'll deal with freeing.
Since now dentry_kill(dentry, 0) may happen after or during
dentry_kill(dentry, 1), we need to recognize that (by seeing
DCACHE_DENTRY_KILLED already set), unlock everything
and either free the sucker (in case DCACHE_MAY_FREE has been
set) or leave it for ongoing dentry_kill(dentry, 1) to deal with.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-05-01 10:30:00 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
shrink_dentry_list(): take parent's ->d_lock earlier
The cause of livelocks there is that we are taking ->d_lock on
dentry and its parent in the wrong order, forcing us to use
trylock on the parent's one. d_walk() takes them in the right
order, and unfortunately it's not hard to create a situation
when shrink_dentry_list() can't make progress since trylock
keeps failing, and shrink_dcache_parent() or check_submounts_and_drop()
keeps calling d_walk() disrupting the very shrink_dentry_list() it's
waiting for.
Solution is straightforward - if that trylock fails, let's unlock
the dentry itself and take locks in the right order. We need to
stabilize ->d_parent without holding ->d_lock, but that's doable
using RCU. And we'd better do that in the very beginning of the
loop in shrink_dentry_list(), since the checks on refcount, etc.
would need to be redone anyway.
That deals with a half of the problem - killing dentries on the
shrink list itself. Another one (dropping their parents) is
in the next commit.
locking parent is interesting - it would be easy to do rcu_read_lock(),
lock whatever we think is a parent, lock dentry itself and check
if the parent is still the right one. Except that we need to check
that *before* locking the dentry, or we are risking taking ->d_lock
out of order. Fortunately, once the D1 is locked, we can check if
D2->d_parent is equal to D1 without the need to lock D2; D2->d_parent
can start or stop pointing to D1 only under D1->d_lock, so taking
D1->d_lock is enough. In other words, the right solution is
rcu_read_lock/lock what looks like parent right now/check if it's
still our parent/rcu_read_unlock/lock the child.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-05-29 08:54:52 -04:00
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
shrink_dentry_list(): take parent's ->d_lock earlier
The cause of livelocks there is that we are taking ->d_lock on
dentry and its parent in the wrong order, forcing us to use
trylock on the parent's one. d_walk() takes them in the right
order, and unfortunately it's not hard to create a situation
when shrink_dentry_list() can't make progress since trylock
keeps failing, and shrink_dcache_parent() or check_submounts_and_drop()
keeps calling d_walk() disrupting the very shrink_dentry_list() it's
waiting for.
Solution is straightforward - if that trylock fails, let's unlock
the dentry itself and take locks in the right order. We need to
stabilize ->d_parent without holding ->d_lock, but that's doable
using RCU. And we'd better do that in the very beginning of the
loop in shrink_dentry_list(), since the checks on refcount, etc.
would need to be redone anyway.
That deals with a half of the problem - killing dentries on the
shrink list itself. Another one (dropping their parents) is
in the next commit.
locking parent is interesting - it would be easy to do rcu_read_lock(),
lock whatever we think is a parent, lock dentry itself and check
if the parent is still the right one. Except that we need to check
that *before* locking the dentry, or we are risking taking ->d_lock
out of order. Fortunately, once the D1 is locked, we can check if
D2->d_parent is equal to D1 without the need to lock D2; D2->d_parent
can start or stop pointing to D1 only under D1->d_lock, so taking
D1->d_lock is enough. In other words, the right solution is
rcu_read_lock/lock what looks like parent right now/check if it's
still our parent/rcu_read_unlock/lock the child.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-05-29 08:54:52 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-02-23 21:07:24 -05:00
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
2018-02-23 21:07:24 -05:00
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-02-23 21:07:24 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
2018-02-23 21:07:24 -05:00
|
|
|
|
|
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-02-23 21:07:24 -05:00
|
|
|
|
|
|
|
|
|
2020-04-30 07:41:37 -07:00
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
2020-04-30 07:41:37 -07:00
|
|
|
|
|
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-06 21:37:31 -05:00
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-03-06 21:37:31 -05:00
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
2018-02-23 21:07:24 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-30 07:41:37 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-24 17:58:48 -05:00
|
|
|
|
2020-04-30 07:41:37 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Coccinelle-based conversion to use ->i_state accessors
All places were patched by coccinelle with the default expecting that
->i_lock is held, afterwards entries got fixed up by hand to use
unlocked variants as needed.
The script:
@@
expression inode, flags;
@@
- inode->i_state & flags
+ inode_state_read(inode) & flags
@@
expression inode, flags;
@@
- inode->i_state &= ~flags
+ inode_state_clear(inode, flags)
@@
expression inode, flag1, flag2;
@@
- inode->i_state &= ~flag1 & ~flag2
+ inode_state_clear(inode, flag1 | flag2)
@@
expression inode, flags;
@@
- inode->i_state |= flags
+ inode_state_set(inode, flags)
@@
expression inode, flags;
@@
- inode->i_state = flags
+ inode_state_assign(inode, flags)
@@
expression inode, flags;
@@
- flags = inode->i_state
+ flags = inode_state_read(inode)
@@
expression inode, flags;
@@
- READ_ONCE(inode->i_state) & flags
+ inode_state_read(inode) & flags
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-10-09 09:59:18 +02:00
|
|
|
|
2020-04-30 07:41:37 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-31 00:45:40 -04:00
|
|
|
|
|
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
2018-08-01 19:39:05 -05:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
fast_dput(): handle underflows gracefully
If refcount is less than 1, we should just warn, unlock dentry and
return true, so that the caller doesn't try to do anything else.
Taking care of that leaves the rest of "lockref_put_return() has
failed" case equivalent to "decrement refcount and rejoin the
normal slow path after the point where we grab ->d_lock".
NOTE: lockref_put_return() is strictly a fastpath thing - unlike
the rest of lockref primitives, it does not contain a fallback.
Caller (and it looks like fast_dput() is the only legitimate one
in the entire kernel) has to do that itself. Reasons for
lockref_put_return() failures:
* ->d_lock held by somebody
* refcount <= 0
* ... or an architecture not supporting lockref use of
cmpxchg - sparc, anything non-SMP, config with spinlock debugging...
We could add a fallback, but it would be a clumsy API - we'd have
to distinguish between:
(1) refcount > 1 - decremented, lock not held on return
(2) refcount < 1 - left alone, probably no sense to hold the lock
(3) refcount is 1, no cmphxcg - decremented, lock held on return
(4) refcount is 1, cmphxcg supported - decremented, lock *NOT* held
on return.
We want to return with no lock held in case (4); that's the whole point of that
thing. We very much do not want to have the fallback in case (3) return without
a lock, since the caller might have to retake it in that case.
So it wouldn't be more convenient than doing the fallback in the caller and
it would be very easy to screw up, especially since the test coverage would
suck - no way to test (3) and (4) on the same kernel build.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-01 01:08:54 -04:00
|
|
|
|
|
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
|
|
|
|
|
2018-08-01 19:39:05 -05:00
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
|
|
|
|
|
2018-08-01 19:39:05 -05:00
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fast_dput(): handle underflows gracefully
If refcount is less than 1, we should just warn, unlock dentry and
return true, so that the caller doesn't try to do anything else.
Taking care of that leaves the rest of "lockref_put_return() has
failed" case equivalent to "decrement refcount and rejoin the
normal slow path after the point where we grab ->d_lock".
NOTE: lockref_put_return() is strictly a fastpath thing - unlike
the rest of lockref primitives, it does not contain a fallback.
Caller (and it looks like fast_dput() is the only legitimate one
in the entire kernel) has to do that itself. Reasons for
lockref_put_return() failures:
* ->d_lock held by somebody
* refcount <= 0
* ... or an architecture not supporting lockref use of
cmpxchg - sparc, anything non-SMP, config with spinlock debugging...
We could add a fallback, but it would be a clumsy API - we'd have
to distinguish between:
(1) refcount > 1 - decremented, lock not held on return
(2) refcount < 1 - left alone, probably no sense to hold the lock
(3) refcount is 1, no cmphxcg - decremented, lock held on return
(4) refcount is 1, cmphxcg supported - decremented, lock *NOT* held
on return.
We want to return with no lock held in case (4); that's the whole point of that
thing. We very much do not want to have the fallback in case (3) return without
a lock, since the caller might have to retake it in that case.
So it wouldn't be more convenient than doing the fallback in the caller and
it would be very easy to screw up, especially since the test coverage would
suck - no way to test (3) and (4) on the same kernel build.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-01 01:08:54 -04:00
|
|
|
|
2023-11-23 17:33:21 -05:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
2018-08-01 19:39:05 -05:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
2018-08-01 19:39:05 -05:00
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
|
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 03:14:19 -04:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-01-09 15:19:03 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-10-27 18:32:21 -04:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-12 23:39:16 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
2026-04-12 14:35:38 -04:00
|
|
|
|
|
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
2023-10-30 01:06:06 -04:00
|
|
|
|
2026-04-12 14:35:38 -04:00
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
2026-04-12 14:35:38 -04:00
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
2026-04-12 23:39:16 -04:00
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-11 03:56:42 -04:00
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
2026-04-12 23:39:16 -04:00
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
|
|
|
|
|
2013-09-02 11:29:22 -07:00
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
2019-10-31 01:43:31 -04:00
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
2013-09-02 11:29:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-10-31 01:43:31 -04:00
|
|
|
|
locking/atomics, fs/dcache: Convert ACCESS_ONCE() to READ_ONCE()/WRITE_ONCE()
For several reasons, it is desirable to use {READ,WRITE}_ONCE() in
preference to ACCESS_ONCE(), and new code is expected to use one of the
former. So far, there's been no reason to change most existing uses of
ACCESS_ONCE(), as these aren't currently harmful.
However, for some features it is necessary to instrument reads and
writes separately, which is not possible with ACCESS_ONCE(). This
distinction is critical to correct operation.
It's possible to transform the bulk of kernel code using the Coccinelle
script below. However, this doesn't handle comments, leaving references
to ACCESS_ONCE() instances which have been removed. As a preparatory
step, this patch converts the dcache code and comments to use
{READ,WRITE}_ONCE() consistently.
----
virtual patch
@ depends on patch @
expression E1, E2;
@@
- ACCESS_ONCE(E1) = E2
+ WRITE_ONCE(E1, E2)
@ depends on patch @
expression E;
@@
- ACCESS_ONCE(E)
+ READ_ONCE(E)
----
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: davem@davemloft.net
Cc: linux-arch@vger.kernel.org
Cc: mpe@ellerman.id.au
Cc: shuah@kernel.org
Cc: snitzer@redhat.com
Cc: thor.thayer@linux.intel.com
Cc: tj@kernel.org
Cc: will.deacon@arm.com
Link: http://lkml.kernel.org/r/1508792849-3115-4-git-send-email-paulmck@linux.vnet.ibm.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 14:07:14 -07:00
|
|
|
|
2013-09-02 11:29:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-10-31 01:43:31 -04:00
|
|
|
|
2013-09-02 11:29:22 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
2011-01-07 17:49:44 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
2011-01-07 17:49:44 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:44 +11:00
|
|
|
|
2013-08-28 18:24:59 -07:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:32 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fix a race between d_find_any_alias() and final dput() of NORCU dentries
Refcount of a NORCU dentry must not be incremented after having dropped
to zero. Otherwise we might end up with the following race:
CPU1: in fast_dput(d), rcu_read_lock();
CPU1: decrements refcount of d to 0
CPU1: notice that it's unhashed
CPU2: grab a reference to d
CPU2: dput(d), freeing d
CPU1: ... looks like we need to evict d, let's grab ->d_lock, recheck
the refcount, etc.
and that spin_lock(&d->d_lock) ends up a UAF, despite still being in
an RCU read-side critical area started back when the refcount had been
positive. If not for DCACHE_NORCU in d->d_flags freeing would've been
RCU-delayed, so we'd have grabbed ->d_lock, noticed the negative value
stored into refcount by __dentry_kill(), dropped the locks and that would
be it. For NORCU dentries freeing is _not_ delayed, though.
Most of the non-counting references are excluded for NORCU dentries -
they are not allowed to be hashed, they never get placed on LRU, they
never get placed into anyone's list of children and while dput_to_list()
might put them into a shrink list, nobody bumps refcount of something
that had been reached that way.
However, inode's list of aliases can be a problem - it does not contribute
to dentry refcount (for obvious reasons) and we *do* have places that
grab references to something found on that list - that's precisely what
d_find_alias() is. In case of d_find_alias() we are safe - it skips
unhashed aliases, so all NORCU ones are ignored there. d_find_any_alias()
is *not* limited to hashed ones, though, and while it's usually called
for directories (which never get NORCU dentries), there are callers that
use it to get something for non-directories with no hashed aliases.
Having d_find_any_alias() hit a NORCU dentry is not impossible - it can
be easily arranged if you have CAP_DAC_READ_SEARCH (memfd_create() + mmap()
+ name_to_handle_at() for /proc/self/map_files/<...> + munmap() +
open_by_handle_at() will do that, and adding a second memfd_create() for
mount_fd makes it possible to do that without having memfd pinned).
The race window is narrow, and it's probably not feasible on bare hardware,
but...
It's not hard to fix, fortunately:
* separate __d_find_dir_alias() (== current __d_find_any_alias()) to
be used for directory inodes.
* provide dget_alias_ilocked() that would return false for NORCU
dentries with zero refcount and return true incrementing refcount otherwise
* make __d_find_any_alias() go over the list of aliases, using
dget_alias_ilocked() and returning the alias it succeeds on (normally the
first one). Any NORCU alias with zero refcount is going to be evicted by
the thread that had dropped the final reference; this makes __d_find_any_alias()
pretend it had lost the race with eviction.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-05-03 23:00:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-25 10:52:25 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2023-10-29 13:41:06 -04:00
|
|
|
|
2018-04-25 10:52:25 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fix a race between d_find_any_alias() and final dput() of NORCU dentries
Refcount of a NORCU dentry must not be incremented after having dropped
to zero. Otherwise we might end up with the following race:
CPU1: in fast_dput(d), rcu_read_lock();
CPU1: decrements refcount of d to 0
CPU1: notice that it's unhashed
CPU2: grab a reference to d
CPU2: dput(d), freeing d
CPU1: ... looks like we need to evict d, let's grab ->d_lock, recheck
the refcount, etc.
and that spin_lock(&d->d_lock) ends up a UAF, despite still being in
an RCU read-side critical area started back when the refcount had been
positive. If not for DCACHE_NORCU in d->d_flags freeing would've been
RCU-delayed, so we'd have grabbed ->d_lock, noticed the negative value
stored into refcount by __dentry_kill(), dropped the locks and that would
be it. For NORCU dentries freeing is _not_ delayed, though.
Most of the non-counting references are excluded for NORCU dentries -
they are not allowed to be hashed, they never get placed on LRU, they
never get placed into anyone's list of children and while dput_to_list()
might put them into a shrink list, nobody bumps refcount of something
that had been reached that way.
However, inode's list of aliases can be a problem - it does not contribute
to dentry refcount (for obvious reasons) and we *do* have places that
grab references to something found on that list - that's precisely what
d_find_alias() is. In case of d_find_alias() we are safe - it skips
unhashed aliases, so all NORCU ones are ignored there. d_find_any_alias()
is *not* limited to hashed ones, though, and while it's usually called
for directories (which never get NORCU dentries), there are callers that
use it to get something for non-directories with no hashed aliases.
Having d_find_any_alias() hit a NORCU dentry is not impossible - it can
be easily arranged if you have CAP_DAC_READ_SEARCH (memfd_create() + mmap()
+ name_to_handle_at() for /proc/self/map_files/<...> + munmap() +
open_by_handle_at() will do that, and adding a second memfd_create() for
mount_fd makes it possible to do that without having memfd pinned).
The race window is narrow, and it's probably not feasible on bare hardware,
but...
It's not hard to fix, fortunately:
* separate __d_find_dir_alias() (== current __d_find_any_alias()) to
be used for directory inodes.
* provide dget_alias_ilocked() that would return false for NORCU
dentries with zero refcount and return true incrementing refcount otherwise
* make __d_find_any_alias() go over the list of aliases, using
dget_alias_ilocked() and returning the alias it succeeds on (normally the
first one). Any NORCU alias with zero refcount is going to be evicted by
the thread that had dropped the final reference; this makes __d_find_any_alias()
pretend it had lost the race with eviction.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-05-03 23:00:09 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-25 10:52:25 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-16 11:15:51 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-04-25 10:52:25 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fix a race between d_find_any_alias() and final dput() of NORCU dentries
Refcount of a NORCU dentry must not be incremented after having dropped
to zero. Otherwise we might end up with the following race:
CPU1: in fast_dput(d), rcu_read_lock();
CPU1: decrements refcount of d to 0
CPU1: notice that it's unhashed
CPU2: grab a reference to d
CPU2: dput(d), freeing d
CPU1: ... looks like we need to evict d, let's grab ->d_lock, recheck
the refcount, etc.
and that spin_lock(&d->d_lock) ends up a UAF, despite still being in
an RCU read-side critical area started back when the refcount had been
positive. If not for DCACHE_NORCU in d->d_flags freeing would've been
RCU-delayed, so we'd have grabbed ->d_lock, noticed the negative value
stored into refcount by __dentry_kill(), dropped the locks and that would
be it. For NORCU dentries freeing is _not_ delayed, though.
Most of the non-counting references are excluded for NORCU dentries -
they are not allowed to be hashed, they never get placed on LRU, they
never get placed into anyone's list of children and while dput_to_list()
might put them into a shrink list, nobody bumps refcount of something
that had been reached that way.
However, inode's list of aliases can be a problem - it does not contribute
to dentry refcount (for obvious reasons) and we *do* have places that
grab references to something found on that list - that's precisely what
d_find_alias() is. In case of d_find_alias() we are safe - it skips
unhashed aliases, so all NORCU ones are ignored there. d_find_any_alias()
is *not* limited to hashed ones, though, and while it's usually called
for directories (which never get NORCU dentries), there are callers that
use it to get something for non-directories with no hashed aliases.
Having d_find_any_alias() hit a NORCU dentry is not impossible - it can
be easily arranged if you have CAP_DAC_READ_SEARCH (memfd_create() + mmap()
+ name_to_handle_at() for /proc/self/map_files/<...> + munmap() +
open_by_handle_at() will do that, and adding a second memfd_create() for
mount_fd makes it possible to do that without having memfd pinned).
The race window is narrow, and it's probably not feasible on bare hardware,
but...
It's not hard to fix, fortunately:
* separate __d_find_dir_alias() (== current __d_find_any_alias()) to
be used for directory inodes.
* provide dget_alias_ilocked() that would return false for NORCU
dentries with zero refcount and return true incrementing refcount otherwise
* make __d_find_any_alias() go over the list of aliases, using
dget_alias_ilocked() and returning the alias it succeeds on (normally the
first one). Any NORCU alias with zero refcount is going to be evicted by
the thread that had dropped the final reference; this makes __d_find_any_alias()
pretend it had lost the race with eviction.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-05-03 23:00:09 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-01-24 17:58:48 -05:00
|
|
|
|
2011-01-07 17:49:33 +11:00
|
|
|
|
2018-04-25 10:52:25 -04:00
|
|
|
|
2023-11-06 14:44:17 +01:00
|
|
|
|
2014-01-16 17:17:31 -05:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:33 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2021-01-14 09:04:39 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:33 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-03-25 03:06:36 -08:00
|
|
|
|
|
|
|
|
|
2012-06-09 13:51:19 -04:00
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2014-01-16 11:15:51 -05:00
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2006-03-25 03:06:36 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2021-01-05 14:13:52 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Coccinelle-based conversion to use ->i_state accessors
All places were patched by coccinelle with the default expecting that
->i_lock is held, afterwards entries got fixed up by hand to use
unlocked variants as needed.
The script:
@@
expression inode, flags;
@@
- inode->i_state & flags
+ inode_state_read(inode) & flags
@@
expression inode, flags;
@@
- inode->i_state &= ~flags
+ inode_state_clear(inode, flags)
@@
expression inode, flag1, flag2;
@@
- inode->i_state &= ~flag1 & ~flag2
+ inode_state_clear(inode, flag1 | flag2)
@@
expression inode, flags;
@@
- inode->i_state |= flags
+ inode_state_set(inode, flags)
@@
expression inode, flags;
@@
- inode->i_state = flags
+ inode_state_assign(inode, flags)
@@
expression inode, flags;
@@
- flags = inode->i_state
+ flags = inode_state_read(inode)
@@
expression inode, flags;
@@
- READ_ONCE(inode->i_state) & flags
+ inode_state_read(inode) & flags
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-10-09 09:59:18 +02:00
|
|
|
|
2021-01-05 14:13:52 -05:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2021-01-05 14:13:52 -05:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2021-01-05 14:13:52 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-03 14:46:14 -04:00
|
|
|
|
2005-09-10 00:27:07 -07:00
|
|
|
|
2023-11-03 14:46:14 -04:00
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2026-04-12 23:39:16 -04:00
|
|
|
|
|
|
|
|
|
2026-05-04 02:49:20 -04:00
|
|
|
|
|
|
|
|
|
2026-04-12 23:39:16 -04:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2023-11-03 14:46:14 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
2018-02-23 21:54:18 -05:00
|
|
|
|
2026-04-11 03:14:19 -04:00
|
|
|
|
simplify safety for lock_for_kill() slowpath
rcu_read_lock() scopes in dentry eviction machinery are too wide
and badly structured; we end up with too many of those, quite
a few essentially identical. Worse, quite a few of the function
involved are not neutral wrt that, making them harder to reason about.
rcu_read_lock() scope is not the only thing establishing an
RCU read-side critical area - spin_lock scope does the same and
they can be mixed - the sequence
rcu_read_lock()
...
spin_lock()
...
rcu_read_unlock()
...
rcu_read_lock()
...
spun_unlock()
...
rcu_read_unlock()
is an unbroken RCU read-side critical area.
Use of that observation allows to simplify things. First of all,
lock_for_kill() relies upon being in an unbroken RCU read-side
critical area. It's always called with ->d_lock held, and normally
returns without having ever dropped that spinlock. We would not
need rcu_read_lock() at all, if not for the slow path - if trylock
of inode->i_lock fails, we need to drop and retake ->d_lock.
Having all calls of lock_for_kill() inside an rcu_read_lock() scope
takes care of that, but to show that lock_for_kill() slow path is safe,
we need to demonstrate such rcu_read_lock() scope for any call chain
leading to lock_for_kill(). Which is not fun, seeing that there are
10 such scopes, with 5 distinct beginnings between them.
Case 1: opens in dput() proceeds through fast_dput() grabbing ->d_lock,
returning false into dput() and there a call of finish_dput() which calls
dentry_kill(), which calls lock_for_kill(); ends in dentry_kill(), either
right after lock_for_kill() success or right after dropping ->d_lock
on lock_for_kill() failure. ->d_lock is held continuously all the way
into lock_for_kill().
Case 2: opens in dentry_kill(), where we proceed to the same call of
dentry_kill() as in case 1. ->d_lock is held since before the
beginning of the scope and all the way into lock_for_kill().
Case 3: opens in select_collect2(), proceeds through the return to
d_walk() and to shrink_dcache_tree() where we grab ->d_lock and
proceed to call shrink_kill(), which calls dentry_kill(), then as
in the previous scopes.
Case 4: opens in shrink_dentry_list(), followed by call of shrink_kill(),
then same as in case 3. ->d_lock is held since before the beginning
of the scope and all the way into lock_for_kill().
Case 5: opens in shrink_kill(), where it's immediately followed by
call of dentry_kill(), then same as in the previous scopes. ->d_lock
is held since before the beginning of the scope all the way into
lock_for_kill().
Note that in cases 2, 4 and 5 the slow path of lock_for_kill() is the
only part of rcu_read_lock() scope that is not covered by spinlock
scopes. In case 1 we have the area in fast_dput() as well and in
case 3 - the return path from select_collect2() and chunk in shrink_dcache_tree()
up to grabbing ->d_lock.
Seeing that the reasons we need rcu_read_lock() in these additional
areas are completely unrelated to lock_for_kill() slow path, the things
get much more straightforward with
* explicit rcu_read_lock() scope surrounding the area in slow path
of lock_for_kill() where ->d_lock is not held
* shrink_dentry_list() dropping rcu_read_lock() as soon as it has
grabbed ->d_lock.
* dput() dropping rcu_read_lock() just before calling finish_dput().
* rcu_read_lock() calls in finish_dput(), shrink_kill() and
shrink_dentry_list() are removed, along with rcu_read_unlock() calls in
dentry_kill().
RCU read-side critical areas are unchanged by that, safety of lock_for_kill()
slow path is trivial to verify and a bunch of rcu_read_lock() scopes either
gone or become easier to describe.
Update the comments on locking conventions and memory safety considerations,
including the NORCU case.
Incidentally, all calls of fast_dput() are immediately preceded by rcu_read_lock()
and followed by rcu_read_unlock() now, which will allow to simplify those on
the next step...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-11 03:24:28 -04:00
|
|
|
|
2018-02-23 21:54:18 -05:00
|
|
|
|
|
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-05-02 15:38:39 -04:00
|
|
|
|
2023-11-07 15:21:33 -05:00
|
|
|
|
2018-02-23 21:54:18 -05:00
|
|
|
|
2014-05-02 15:38:39 -04:00
|
|
|
|
2011-01-07 17:49:47 +11:00
|
|
|
|
shrink_dentry_list(): start with removing from shrink list
Currently we leave dentry on the list until we are done with
lock_for_kill(). That guarantees that it won't have been
even scheduled for removal until we remove it from the list
and drop ->d_lock. We grab ->d_lock and rcu_read_lock()
and call lock_for_kill(). There are four possible cases:
1) lock_for_kill() has succeeded; dentry and its inode
(if any) are locked, dentry refcount is zero and we can
remove it from shrink list and feed it to shrink_kill().
2) lock_for_kill() fails since dentry has become busy.
Nothing to do, rcu_read_unlock(), remove from shrink list,
drop ->d_lock and move on.
3) lock_for_kill() fails since dentry is currently
being killed - already entered __dentry_kill(), but hasn't
reached dentry_unlist() yet. Nothing to do, we should just
do rcu_read_unlock(), remove from shrink list so that
whoever's executing __dentry_kill() would free it once they
are done, drop ->d_lock and move on - same actions as in
case (2).
4) lock_for_kill() fails since dentry has been killed
(reached dentry_unlist(), DCACHE_DENTRY_KILLED set in ->d_flags).
In that case whoever had been killing it had already seen it
on our shrink list and skipped freeing it. At that point it's
just a passive chunk of memory; rcu_read_unlock(), remove from
the list, drop ->d_lock and use dentry_free() to schedule
freeing.
While that works, there's a simpler way to do it:
* grab ->d_lock
* remove dentry from our shrink list
* if DCACHE_DENTRY_KILLED is already set, drop ->d_lock,
call dentry_free() and move on.
* otherwise grab rcu_read_lock() and call lock_for_free()
* if lock_for_kill() succeeds, feed dentry
to shrink_kill(), otherwise drop the locks and move on.
The end result is equivalent to the old variant. The only difference
arises if at the time we grab ->d_lock dentry had refcount 0 and
lock_for_kill() had failed spin_trylock() and had to drop and regain
->d_lock. Otherwise nobody can observe at which point within the
unbroken ->d_lock scope dentry had been removed from the shrink list -
all accesses to ->d_lru are under ->d_lock.
If ->d_lock had been dropped and regained, it is possible for another
thread to feed that dentry to __dentry_kill(); if it doesn't get to
dentry_unlist() before we regain ->d_lock, behaviour is still identical -
it's case (3) and by the time __dentry_kill() would've gotten around
to checking if the victim is on shrink list, it would've been already
removed from ours.
If __dentry_kill() from another thread *does* get to dentry_unlist(),
in the old variant we would have __dentry_kill() leave calling
dentry_free() to us and in the new one __dentry_kill() would've called
dentry_free() itself. Since we are under rcu_read_lock(), we are
guaranteed that actual freeing won't happen until we get around to
rcu_read_unlock(). IOW, the new variant is still safe wrt UAF, if
not for the same reason as the old one, and overall result is the same;
the only difference is which threads ends up scheduling the actual
freeing of dentry.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-11 01:52:53 -04:00
|
|
|
|
|
|
|
|
|
fix soft lock up at NFS mount via per-SB LRU-list of unused dentries
[Summary]
Split LRU-list of unused dentries to one per superblock to avoid soft
lock up during NFS mounts and remounting of any filesystem.
Previously I posted here:
http://lkml.org/lkml/2008/3/5/590
[Descriptions]
- background
dentry_unused is a list of dentries which are not referenced.
dentry_unused grows up when references on directories or files are
released. This list can be very long if there is huge free memory.
- the problem
When shrink_dcache_sb() is called, it scans all dentry_unused linearly
under spin_lock(), and if dentry->d_sb is differnt from given
superblock, scan next dentry. This scan costs very much if there are
many entries, and very ineffective if there are many superblocks.
IOW, When we need to shrink unused dentries on one dentry, but scans
unused dentries on all superblocks in the system. For example, we scan
500 dentries to unmount a filesystem, but scans 1,000,000 or more unused
dentries on other superblocks.
In our case , At mounting NFS*, shrink_dcache_sb() is called to shrink
unused dentries on NFS, but scans 100,000,000 unused dentries on
superblocks in the system such as local ext3 filesystems. I hear NFS
mounting took 1 min on some system in use.
* : NFS uses virtual filesystem in rpc layer, so NFS is affected by
this problem.
100,000,000 is possible number on large systems.
Per-superblock LRU of unused dentried can reduce the cost in
reasonable manner.
- How to fix
I found this problem is solved by David Chinner's "Per-superblock
unused dentry LRU lists V3"(1), so I rebase it and add some fix to
reclaim with fairness, which is in Andrew Morton's comments(2).
1) http://lkml.org/lkml/2006/5/25/318
2) http://lkml.org/lkml/2006/5/25/320
Split LRU-list of unused dentries to each superblocks. Then, NFS
mounting will check dentries under a superblock instead of all. But
this spliting will break LRU of dentry-unused. So, I've attempted to
make reclaim unused dentrins with fairness by calculate number of
dentries to scan on this sb based on following way
number of dentries to scan on this sb =
count * (number of dentries on this sb / number of dentries in the machine)
- ToDo
- I have to measuring performance number and do stress tests.
- When unmount occurs during prune_dcache(), scanning on same
superblock, It is unable to reach next superblock because it is gone
away. We restart scannig superblock from first one, it causes
unfairness of reclaim unused dentries on first superblock. But I think
this happens very rarely.
- Test Results
Result on 6GB boxes with excessive unused dentries.
Without patch:
$ cat /proc/sys/fs/dentry-state
10181835 10180203 45 0 0 0
# mount -t nfs 10.124.60.70:/work/kernel-src nfs
real 0m1.830s
user 0m0.001s
sys 0m1.653s
With this patch:
$ cat /proc/sys/fs/dentry-state
10236610 10234751 45 0 0 0
# mount -t nfs 10.124.60.70:/work/kernel-src nfs
real 0m0.106s
user 0m0.002s
sys 0m0.032s
[akpm@linux-foundation.org: fix comments]
Signed-off-by: Kentaro Makita <k-makita@np.css.fujitsu.com>
Cc: Neil Brown <neilb@suse.de>
Cc: Trond Myklebust <trond.myklebust@fys.uio.no>
Cc: David Chinner <dgc@sgi.com>
Cc: "J. Bruce Fields" <bfields@fieldses.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-23 21:27:13 -07:00
|
|
|
|
shrink_dentry_list(): start with removing from shrink list
Currently we leave dentry on the list until we are done with
lock_for_kill(). That guarantees that it won't have been
even scheduled for removal until we remove it from the list
and drop ->d_lock. We grab ->d_lock and rcu_read_lock()
and call lock_for_kill(). There are four possible cases:
1) lock_for_kill() has succeeded; dentry and its inode
(if any) are locked, dentry refcount is zero and we can
remove it from shrink list and feed it to shrink_kill().
2) lock_for_kill() fails since dentry has become busy.
Nothing to do, rcu_read_unlock(), remove from shrink list,
drop ->d_lock and move on.
3) lock_for_kill() fails since dentry is currently
being killed - already entered __dentry_kill(), but hasn't
reached dentry_unlist() yet. Nothing to do, we should just
do rcu_read_unlock(), remove from shrink list so that
whoever's executing __dentry_kill() would free it once they
are done, drop ->d_lock and move on - same actions as in
case (2).
4) lock_for_kill() fails since dentry has been killed
(reached dentry_unlist(), DCACHE_DENTRY_KILLED set in ->d_flags).
In that case whoever had been killing it had already seen it
on our shrink list and skipped freeing it. At that point it's
just a passive chunk of memory; rcu_read_unlock(), remove from
the list, drop ->d_lock and use dentry_free() to schedule
freeing.
While that works, there's a simpler way to do it:
* grab ->d_lock
* remove dentry from our shrink list
* if DCACHE_DENTRY_KILLED is already set, drop ->d_lock,
call dentry_free() and move on.
* otherwise grab rcu_read_lock() and call lock_for_free()
* if lock_for_kill() succeeds, feed dentry
to shrink_kill(), otherwise drop the locks and move on.
The end result is equivalent to the old variant. The only difference
arises if at the time we grab ->d_lock dentry had refcount 0 and
lock_for_kill() had failed spin_trylock() and had to drop and regain
->d_lock. Otherwise nobody can observe at which point within the
unbroken ->d_lock scope dentry had been removed from the shrink list -
all accesses to ->d_lru are under ->d_lock.
If ->d_lock had been dropped and regained, it is possible for another
thread to feed that dentry to __dentry_kill(); if it doesn't get to
dentry_unlist() before we regain ->d_lock, behaviour is still identical -
it's case (3) and by the time __dentry_kill() would've gotten around
to checking if the victim is on shrink list, it would've been already
removed from ours.
If __dentry_kill() from another thread *does* get to dentry_unlist(),
in the old variant we would have __dentry_kill() leave calling
dentry_free() to us and in the new one __dentry_kill() would've called
dentry_free() itself. Since we are under rcu_read_lock(), we are
guaranteed that actual freeing won't happen until we get around to
rcu_read_unlock(). IOW, the new variant is still safe wrt UAF, if
not for the same reason as the old one, and overall result is the same;
the only difference is which threads ends up scheduling the actual
freeing of dentry.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-11 01:52:53 -04:00
|
|
|
|
2014-05-28 09:48:44 -04:00
|
|
|
|
|
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
2025-09-16 14:53:07 +01:00
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
2024-11-05 01:52:57 +08:00
|
|
|
|
2013-08-28 10:18:00 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
2013-08-28 10:18:00 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
list_lru: allow explicit memcg and NUMA node selection
Patch series "workload-specific and memory pressure-driven zswap
writeback", v8.
There are currently several issues with zswap writeback:
1. There is only a single global LRU for zswap, making it impossible to
perform worload-specific shrinking - an memcg under memory pressure
cannot determine which pages in the pool it owns, and often ends up
writing pages from other memcgs. This issue has been previously
observed in practice and mitigated by simply disabling
memcg-initiated shrinking:
https://lore.kernel.org/all/20230530232435.3097106-1-nphamcs@gmail.com/T/#u
But this solution leaves a lot to be desired, as we still do not
have an avenue for an memcg to free up its own memory locked up in
the zswap pool.
2. We only shrink the zswap pool when the user-defined limit is hit.
This means that if we set the limit too high, cold data that are
unlikely to be used again will reside in the pool, wasting precious
memory. It is hard to predict how much zswap space will be needed
ahead of time, as this depends on the workload (specifically, on
factors such as memory access patterns and compressibility of the
memory pages).
This patch series solves these issues by separating the global zswap LRU
into per-memcg and per-NUMA LRUs, and performs workload-specific (i.e
memcg- and NUMA-aware) zswap writeback under memory pressure. The new
shrinker does not have any parameter that must be tuned by the user, and
can be opted in or out on a per-memcg basis.
As a proof of concept, we ran the following synthetic benchmark: build the
linux kernel in a memory-limited cgroup, and allocate some cold data in
tmpfs to see if the shrinker could write them out and improved the overall
performance. Depending on the amount of cold data generated, we observe
from 14% to 35% reduction in kernel CPU time used in the kernel builds.
This patch (of 6):
The interface of list_lru is based on the assumption that the list node
and the data it represents belong to the same allocated on the correct
node/memcg. While this assumption is valid for existing slab objects LRU
such as dentries and inodes, it is undocumented, and rather inflexible for
certain potential list_lru users (such as the upcoming zswap shrinker and
the THP shrinker). It has caused us a lot of issues during our
development.
This patch changes list_lru interface so that the caller must explicitly
specify numa node and memcg when adding and removing objects. The old
list_lru_add() and list_lru_del() are renamed to list_lru_add_obj() and
list_lru_del_obj(), respectively.
It also extends the list_lru API with a new function, list_lru_putback,
which undoes a previous list_lru_isolate call. Unlike list_lru_add, it
does not increment the LRU node count (as list_lru_isolate does not
decrement the node count). list_lru_putback also allows for explicit
memcg and NUMA node selection.
Link: https://lkml.kernel.org/r/20231130194023.4102148-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231130194023.4102148-2-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Bagas Sanjaya <bagasdotme@gmail.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-30 11:40:18 -08:00
|
|
|
|
2013-08-28 10:18:00 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
2013-08-28 10:18:00 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
2011-08-23 18:56:24 +10:00
|
|
|
|
|
|
|
|
|
list_lru: introduce list_lru_shrink_{count,walk}
Kmem accounting of memcg is unusable now, because it lacks slab shrinker
support. That means when we hit the limit we will get ENOMEM w/o any
chance to recover. What we should do then is to call shrink_slab, which
would reclaim old inode/dentry caches from this cgroup. This is what
this patch set is intended to do.
Basically, it does two things. First, it introduces the notion of
per-memcg slab shrinker. A shrinker that wants to reclaim objects per
cgroup should mark itself as SHRINKER_MEMCG_AWARE. Then it will be
passed the memory cgroup to scan from in shrink_control->memcg. For
such shrinkers shrink_slab iterates over the whole cgroup subtree under
the target cgroup and calls the shrinker for each kmem-active memory
cgroup.
Secondly, this patch set makes the list_lru structure per-memcg. It's
done transparently to list_lru users - everything they have to do is to
tell list_lru_init that they want memcg-aware list_lru. Then the
list_lru will automatically distribute objects among per-memcg lists
basing on which cgroup the object is accounted to. This way to make FS
shrinkers (icache, dcache) memcg-aware we only need to make them use
memcg-aware list_lru, and this is what this patch set does.
As before, this patch set only enables per-memcg kmem reclaim when the
pressure goes from memory.limit, not from memory.kmem.limit. Handling
memory.kmem.limit is going to be tricky due to GFP_NOFS allocations, and
it is still unclear whether we will have this knob in the unified
hierarchy.
This patch (of 9):
NUMA aware slab shrinkers use the list_lru structure to distribute
objects coming from different NUMA nodes to different lists. Whenever
such a shrinker needs to count or scan objects from a particular node,
it issues commands like this:
count = list_lru_count_node(lru, sc->nid);
freed = list_lru_walk_node(lru, sc->nid, isolate_func,
isolate_arg, &sc->nr_to_scan);
where sc is an instance of the shrink_control structure passed to it
from vmscan.
To simplify this, let's add special list_lru functions to be used by
shrinkers, list_lru_shrink_count() and list_lru_shrink_walk(), which
consolidate the nid and nr_to_scan arguments in the shrink_control
structure.
This will also allow us to avoid patching shrinkers that use list_lru
when we make shrink_slab() per-memcg - all we will have to do is extend
the shrink_control structure to include the target memcg and make
list_lru_shrink_{count,walk} handle this appropriately.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Suggested-by: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Glauber Costa <glommer@gmail.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-12 14:58:47 -08:00
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
list_lru: introduce list_lru_shrink_{count,walk}
Kmem accounting of memcg is unusable now, because it lacks slab shrinker
support. That means when we hit the limit we will get ENOMEM w/o any
chance to recover. What we should do then is to call shrink_slab, which
would reclaim old inode/dentry caches from this cgroup. This is what
this patch set is intended to do.
Basically, it does two things. First, it introduces the notion of
per-memcg slab shrinker. A shrinker that wants to reclaim objects per
cgroup should mark itself as SHRINKER_MEMCG_AWARE. Then it will be
passed the memory cgroup to scan from in shrink_control->memcg. For
such shrinkers shrink_slab iterates over the whole cgroup subtree under
the target cgroup and calls the shrinker for each kmem-active memory
cgroup.
Secondly, this patch set makes the list_lru structure per-memcg. It's
done transparently to list_lru users - everything they have to do is to
tell list_lru_init that they want memcg-aware list_lru. Then the
list_lru will automatically distribute objects among per-memcg lists
basing on which cgroup the object is accounted to. This way to make FS
shrinkers (icache, dcache) memcg-aware we only need to make them use
memcg-aware list_lru, and this is what this patch set does.
As before, this patch set only enables per-memcg kmem reclaim when the
pressure goes from memory.limit, not from memory.kmem.limit. Handling
memory.kmem.limit is going to be tricky due to GFP_NOFS allocations, and
it is still unclear whether we will have this knob in the unified
hierarchy.
This patch (of 9):
NUMA aware slab shrinkers use the list_lru structure to distribute
objects coming from different NUMA nodes to different lists. Whenever
such a shrinker needs to count or scan objects from a particular node,
it issues commands like this:
count = list_lru_count_node(lru, sc->nid);
freed = list_lru_walk_node(lru, sc->nid, isolate_func,
isolate_arg, &sc->nr_to_scan);
where sc is an instance of the shrink_control structure passed to it
from vmscan.
To simplify this, let's add special list_lru functions to be used by
shrinkers, list_lru_shrink_count() and list_lru_shrink_walk(), which
consolidate the nid and nr_to_scan arguments in the shrink_control
structure.
This will also allow us to avoid patching shrinkers that use list_lru
when we make shrink_slab() per-memcg - all we will have to do is extend
the shrink_control structure to include the target memcg and make
list_lru_shrink_{count,walk} handle this appropriately.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Suggested-by: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Glauber Costa <glommer@gmail.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-12 14:58:47 -08:00
|
|
|
|
|
|
|
|
|
2011-08-23 18:56:24 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
list_lru: introduce list_lru_shrink_{count,walk}
Kmem accounting of memcg is unusable now, because it lacks slab shrinker
support. That means when we hit the limit we will get ENOMEM w/o any
chance to recover. What we should do then is to call shrink_slab, which
would reclaim old inode/dentry caches from this cgroup. This is what
this patch set is intended to do.
Basically, it does two things. First, it introduces the notion of
per-memcg slab shrinker. A shrinker that wants to reclaim objects per
cgroup should mark itself as SHRINKER_MEMCG_AWARE. Then it will be
passed the memory cgroup to scan from in shrink_control->memcg. For
such shrinkers shrink_slab iterates over the whole cgroup subtree under
the target cgroup and calls the shrinker for each kmem-active memory
cgroup.
Secondly, this patch set makes the list_lru structure per-memcg. It's
done transparently to list_lru users - everything they have to do is to
tell list_lru_init that they want memcg-aware list_lru. Then the
list_lru will automatically distribute objects among per-memcg lists
basing on which cgroup the object is accounted to. This way to make FS
shrinkers (icache, dcache) memcg-aware we only need to make them use
memcg-aware list_lru, and this is what this patch set does.
As before, this patch set only enables per-memcg kmem reclaim when the
pressure goes from memory.limit, not from memory.kmem.limit. Handling
memory.kmem.limit is going to be tricky due to GFP_NOFS allocations, and
it is still unclear whether we will have this knob in the unified
hierarchy.
This patch (of 9):
NUMA aware slab shrinkers use the list_lru structure to distribute
objects coming from different NUMA nodes to different lists. Whenever
such a shrinker needs to count or scan objects from a particular node,
it issues commands like this:
count = list_lru_count_node(lru, sc->nid);
freed = list_lru_walk_node(lru, sc->nid, isolate_func,
isolate_arg, &sc->nr_to_scan);
where sc is an instance of the shrink_control structure passed to it
from vmscan.
To simplify this, let's add special list_lru functions to be used by
shrinkers, list_lru_shrink_count() and list_lru_shrink_walk(), which
consolidate the nid and nr_to_scan arguments in the shrink_control
structure.
This will also allow us to avoid patching shrinkers that use list_lru
when we make shrink_slab() per-memcg - all we will have to do is extend
the shrink_control structure to include the target memcg and make
list_lru_shrink_{count,walk} handle this appropriately.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Suggested-by: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Glauber Costa <glommer@gmail.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-12 14:58:47 -08:00
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
2013-08-28 10:18:00 +10:00
|
|
|
|
|
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
list_lru: introduce list_lru_shrink_{count,walk}
Kmem accounting of memcg is unusable now, because it lacks slab shrinker
support. That means when we hit the limit we will get ENOMEM w/o any
chance to recover. What we should do then is to call shrink_slab, which
would reclaim old inode/dentry caches from this cgroup. This is what
this patch set is intended to do.
Basically, it does two things. First, it introduces the notion of
per-memcg slab shrinker. A shrinker that wants to reclaim objects per
cgroup should mark itself as SHRINKER_MEMCG_AWARE. Then it will be
passed the memory cgroup to scan from in shrink_control->memcg. For
such shrinkers shrink_slab iterates over the whole cgroup subtree under
the target cgroup and calls the shrinker for each kmem-active memory
cgroup.
Secondly, this patch set makes the list_lru structure per-memcg. It's
done transparently to list_lru users - everything they have to do is to
tell list_lru_init that they want memcg-aware list_lru. Then the
list_lru will automatically distribute objects among per-memcg lists
basing on which cgroup the object is accounted to. This way to make FS
shrinkers (icache, dcache) memcg-aware we only need to make them use
memcg-aware list_lru, and this is what this patch set does.
As before, this patch set only enables per-memcg kmem reclaim when the
pressure goes from memory.limit, not from memory.kmem.limit. Handling
memory.kmem.limit is going to be tricky due to GFP_NOFS allocations, and
it is still unclear whether we will have this knob in the unified
hierarchy.
This patch (of 9):
NUMA aware slab shrinkers use the list_lru structure to distribute
objects coming from different NUMA nodes to different lists. Whenever
such a shrinker needs to count or scan objects from a particular node,
it issues commands like this:
count = list_lru_count_node(lru, sc->nid);
freed = list_lru_walk_node(lru, sc->nid, isolate_func,
isolate_arg, &sc->nr_to_scan);
where sc is an instance of the shrink_control structure passed to it
from vmscan.
To simplify this, let's add special list_lru functions to be used by
shrinkers, list_lru_shrink_count() and list_lru_shrink_walk(), which
consolidate the nid and nr_to_scan arguments in the shrink_control
structure.
This will also allow us to avoid patching shrinkers that use list_lru
when we make shrink_slab() per-memcg - all we will have to do is extend
the shrink_control structure to include the target memcg and make
list_lru_shrink_{count,walk} handle this appropriately.
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Suggested-by: Dave Chinner <david@fromorbit.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Glauber Costa <glommer@gmail.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Lameter <cl@linux.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-12 14:58:47 -08:00
|
|
|
|
|
|
|
|
|
2013-08-28 10:18:00 +10:00
|
|
|
|
2013-08-28 10:17:57 +10:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
list_lru: remove special case function list_lru_dispose_all.
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code). At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway. However, it's
implementation is buggy in the following way:
When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST. However, this is done without the
nlru->lock taken. The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time. When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is. Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.
The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock). To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.
Such mechanism already exists in the form of the list_lru_walk_cb
callback. So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.
With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:18:03 +10:00
|
|
|
|
2024-11-05 01:52:57 +08:00
|
|
|
|
2013-08-28 10:17:55 +10:00
|
|
|
|
list_lru: remove special case function list_lru_dispose_all.
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code). At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway. However, it's
implementation is buggy in the following way:
When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST. However, this is done without the
nlru->lock taken. The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time. When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is. Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.
The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock). To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.
Such mechanism already exists in the form of the list_lru_walk_cb
callback. So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.
With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:18:03 +10:00
|
|
|
|
|
|
|
|
|
2013-08-28 10:17:55 +10:00
|
|
|
|
list_lru: remove special case function list_lru_dispose_all.
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code). At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway. However, it's
implementation is buggy in the following way:
When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST. However, this is done without the
nlru->lock taken. The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time. When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is. Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.
The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock). To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.
Such mechanism already exists in the form of the list_lru_walk_cb
callback. So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.
With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:18:03 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-02-12 14:59:35 -08:00
|
|
|
|
list_lru: remove special case function list_lru_dispose_all.
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code). At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway. However, it's
implementation is buggy in the following way:
When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST. However, this is done without the
nlru->lock taken. The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time. When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is. Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.
The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock). To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.
Such mechanism already exists in the form of the list_lru_walk_cb
callback. So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.
With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:18:03 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-08-28 10:17:55 +10:00
|
|
|
|
|
|
|
|
|
list_lru: remove special case function list_lru_dispose_all.
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code). At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway. However, it's
implementation is buggy in the following way:
When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST. However, this is done without the
nlru->lock taken. The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time. When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is. Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.
The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock). To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.
Such mechanism already exists in the form of the list_lru_walk_cb
callback. So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.
With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:18:03 +10:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2010-10-10 05:36:25 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
list_lru: remove special case function list_lru_dispose_all.
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code). At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway. However, it's
implementation is buggy in the following way:
When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST. However, this is done without the
nlru->lock taken. The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time. When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is. Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.
The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock). To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.
Such mechanism already exists in the form of the list_lru_walk_cb
callback. So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.
With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:18:03 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:36 -05:00
|
|
|
|
2017-07-10 15:50:00 -07:00
|
|
|
|
list_lru: remove special case function list_lru_dispose_all.
The list_lru implementation has one function, list_lru_dispose_all, with
only one user (the dentry code). At first, such function appears to make
sense because we are really not interested in the result of isolating each
dentry separately - all of them are going away anyway. However, it's
implementation is buggy in the following way:
When we call list_lru_dispose_all in fs/dcache.c, we scan all dentries
marking them with DCACHE_SHRINK_LIST. However, this is done without the
nlru->lock taken. The imediate result of that is that someone else may
add or remove the dentry from the LRU at the same time. When list_lru_del
happens in that scenario we will see an element that is not yet marked
with DCACHE_SHRINK_LIST (even though it will be in the future) and
obviously remove it from an lru where the element no longer is. Since
list_lru_dispose_all will in effect count down nlru's nr_items and
list_lru_del will do the same, this will lead to an imbalance.
The solution for this would not be so simple: we can obviously just keep
the lru_lock taken, but then we have no guarantees that we will be able to
acquire the dentry lock (dentry->d_lock). To properly solve this, we need
a communication mechanism between the lru and dentry code, so they can
coordinate this with each other.
Such mechanism already exists in the form of the list_lru_walk_cb
callback. So it is possible to construct a dcache-side prune function
that does the right thing only by calling list_lru_walk in a loop until no
more dentries are available.
With only one user, plus the fact that a sane solution for the problem
would involve boucing between dcache and list_lru anyway, I see little
justification to keep the special case list_lru_dispose_all in tree.
Signed-off-by: Glauber Costa <glommer@openvz.org>
Cc: Michal Hocko <mhocko@suse.cz>
Acked-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-08-28 10:18:03 +10:00
|
|
|
|
2017-07-10 15:50:00 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2026-01-28 06:31:50 -08:00
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-15 18:27:23 -04:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
2018-04-15 18:27:23 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2013-09-09 15:22:25 -04:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:37 +11:00
|
|
|
|
2011-01-07 17:49:39 +11:00
|
|
|
|
2013-09-09 15:22:25 -04:00
|
|
|
|
2011-01-07 17:49:39 +11:00
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2026-05-05 00:20:19 -04:00
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2019-09-19 12:09:40 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-26 19:31:10 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
|
|
|
|
|
2013-10-25 17:04:27 -04:00
|
|
|
|
2026-04-21 15:52:13 -04:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2013-10-25 17:04:27 -04:00
|
|
|
|
2026-04-21 15:52:13 -04:00
|
|
|
|
2013-10-25 17:04:27 -04:00
|
|
|
|
2014-10-26 19:31:10 -04:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:37 +11:00
|
|
|
|
2015-05-28 23:09:19 -04:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-10-26 19:31:10 -04:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2013-09-09 15:22:25 -04:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
2011-01-07 17:49:39 +11:00
|
|
|
|
|
|
|
|
|
2014-10-26 19:31:10 -04:00
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
2013-09-09 15:22:25 -04:00
|
|
|
|
2011-01-07 17:49:39 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
2016-11-24 08:03:41 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-02 19:55:41 -04:00
|
|
|
|
2016-11-24 08:03:41 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-02 19:55:41 -04:00
|
|
|
|
2018-04-15 18:27:23 -04:00
|
|
|
|
2016-11-24 08:03:41 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-05 14:39:11 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-02-13 09:39:37 -08:00
|
|
|
|
2013-09-05 14:39:11 +02:00
|
|
|
|
2026-09-30 15:32:07 +02:00
|
|
|
|
|
|
|
|
|
2013-09-05 14:39:11 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-25 13:21:14 -05:00
|
|
|
|
2013-09-05 14:39:11 +02:00
|
|
|
|
2014-02-13 09:39:37 -08:00
|
|
|
|
2013-09-05 14:39:11 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-09-30 15:32:07 +02:00
|
|
|
|
2017-01-03 14:18:43 +13:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-05 14:39:11 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-25 13:21:14 -05:00
|
|
|
|
2013-09-05 14:39:11 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2012-09-18 16:35:51 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
2026-04-12 23:39:16 -04:00
|
|
|
|
__dentry_kill(): new locking scheme
Currently we enter __dentry_kill() with parent (along with the victim
dentry and victim's inode) held locked. Then we
mark dentry refcount as dead
call ->d_prune()
remove dentry from hash
remove it from the parent's list of children
unlock the parent, don't need it from that point on
detach dentry from inode, unlock dentry and drop the inode
(via ->d_iput())
call ->d_release()
regain the lock on dentry
check if it's on a shrink list (in which case freeing its empty husk
has to be left to shrink_dentry_list()) or not (in which case we can free it
ourselves). In the former case, mark it as an empty husk, so that
shrink_dentry_list() would know it can free the sucker.
drop the lock on dentry
... and usually the caller proceeds to drop a reference on the parent,
possibly retaking the lock on it.
That is painful for a bunch of reasons, starting with the need to take locks
out of order, but not limited to that - the parent of positive dentry can
change if we drop its ->d_lock, so getting these locks has to be done with
care. Moreover, as soon as dentry is out of the parent's list of children,
shrink_dcache_for_umount() won't see it anymore, making it appear as if
the parent is inexplicably busy. We do work around that by having
shrink_dentry_list() decrement the parent's refcount first and put it on
shrink list to be evicted once we are done with __dentry_kill() of child,
but that may in some cases lead to ->d_iput() on child called after the
parent got killed. That doesn't happen in cases where in-tree ->d_iput()
instances might want to look at the parent, but that's brittle as hell.
Solution: do removal from the parent's list of children in the very
end of __dentry_kill(). As the result, the callers do not need to
lock the parent and by the time we really need the parent locked,
dentry is negative and is guaranteed not to be moved around.
It does mean that ->d_prune() will be called with parent not locked.
It also means that we might see dentries in process of being torn
down while going through the parent's list of children; those dentries
will be unhashed, negative and with refcount marked dead. In practice,
that's enough for in-tree code that looks through the list of children
to do the right thing as-is. Out-of-tree code might need to be adjusted.
Calling conventions: __dentry_kill(dentry) is called with dentry->d_lock
held, along with ->i_lock of its inode (if any). It either returns
the parent (locked, with refcount decremented to 0) or NULL (if there'd
been no parent or if refcount decrement for parent hadn't reached 0).
lock_for_kill() is adjusted for new requirements - it doesn't touch
the parent's ->d_lock at all.
Callers adjusted. Note that for dput() we don't need to bother with
fast_dput() for the parent - we just need to check retain_dentry()
for it, since its ->d_lock is still held since the moment when
__dentry_kill() had taken it to remove the victim from the list of
children.
The kludge with early decrement of parent's refcount in
shrink_dentry_list() is no longer needed - shrink_dcache_for_umount()
sees the half-killed dentries in the list of children for as long
as they are pinning the parent. They are easily recognized and
accounted for by select_collect(), so we know we are not done yet.
As the result, we always have the expected ordering for ->d_iput()/->d_release()
vs. __dentry_kill() of the parent, no exceptions. Moreover, the current
rules for shrink lists (one must make sure that shrink_dcache_for_umount()
won't happen while any dentries from the superblock in question are on
any shrink lists) are gone - shrink_dcache_for_umount() will do the
right thing in all cases, taking such dentries out. Their empty
husks (memory occupied by struct dentry itself + its external name,
if any) will remain on the shrink lists, but they are no obstacles
to filesystem shutdown. And such husks will get freed as soon as
shrink_dentry_list() of the list they are on gets to them.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-07 16:14:08 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-05-03 00:02:25 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
introduce a flag for explicitly marking persistently pinned dentries
Some filesystems use a kinda-sorta controlled dentry refcount leak to pin
dentries of created objects in dcache (and undo it when removing those).
Reference is grabbed and not released, but it's not actually _stored_
anywhere. That works, but it's hard to follow and verify; among other
things, we have no way to tell _which_ of the increments is intended
to be an unpaired one. Worse, on removal we need to decide whether
the reference had already been dropped, which can be non-trivial if
that removal is on umount and we need to figure out if this dentry is
pinned due to e.g. unlink() not done. Usually that is handled by using
kill_litter_super() as ->kill_sb(), but there are open-coded special
cases of the same (consider e.g. /proc/self).
Things get simpler if we introduce a new dentry flag (DCACHE_PERSISTENT)
marking those "leaked" dentries. Having it set claims responsibility
for +1 in refcount.
The end result this series is aiming for:
* get these unbalanced dget() and dput() replaced with new primitives that
would, in addition to adjusting refcount, set and clear persistency flag.
* instead of having kill_litter_super() mess with removing the remaining
"leaked" references (e.g. for all tmpfs files that hadn't been removed
prior to umount), have the regular shrink_dcache_for_umount() strip
DCACHE_PERSISTENT of all dentries, dropping the corresponding
reference if it had been set. After that kill_litter_super() becomes
an equivalent of kill_anon_super().
Doing that in a single step is not feasible - it would affect too many places
in too many filesystems. It has to be split into a series.
Here we
* introduce the new flag
* teach shrink_dcache_for_umount() to handle it (i.e. remove
and drop refcount on anything that survives to umount with that flag
still set)
* teach kill_litter_super() that anything with that flag does
*not* need to be unpinned.
Next commits will add primitives for maintaing that flag and convert the
common helpers to those. After that - a long series of per-filesystem
patches converting to those primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:15:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
2026-04-12 23:39:16 -04:00
|
|
|
|
2026-04-11 04:01:28 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
introduce a flag for explicitly marking persistently pinned dentries
Some filesystems use a kinda-sorta controlled dentry refcount leak to pin
dentries of created objects in dcache (and undo it when removing those).
Reference is grabbed and not released, but it's not actually _stored_
anywhere. That works, but it's hard to follow and verify; among other
things, we have no way to tell _which_ of the increments is intended
to be an unpaired one. Worse, on removal we need to decide whether
the reference had already been dropped, which can be non-trivial if
that removal is on umount and we need to figure out if this dentry is
pinned due to e.g. unlink() not done. Usually that is handled by using
kill_litter_super() as ->kill_sb(), but there are open-coded special
cases of the same (consider e.g. /proc/self).
Things get simpler if we introduce a new dentry flag (DCACHE_PERSISTENT)
marking those "leaked" dentries. Having it set claims responsibility
for +1 in refcount.
The end result this series is aiming for:
* get these unbalanced dget() and dput() replaced with new primitives that
would, in addition to adjusting refcount, set and clear persistency flag.
* instead of having kill_litter_super() mess with removing the remaining
"leaked" references (e.g. for all tmpfs files that hadn't been removed
prior to umount), have the regular shrink_dcache_for_umount() strip
DCACHE_PERSISTENT of all dentries, dropping the corresponding
reference if it had been set. After that kill_litter_super() becomes
an equivalent of kill_anon_super().
Doing that in a single step is not feasible - it would affect too many places
in too many filesystems. It has to be split into a series.
Here we
* introduce the new flag
* teach shrink_dcache_for_umount() to handle it (i.e. remove
and drop refcount on anything that survives to umount with that flag
still set)
* teach kill_litter_super() that anything with that flag does
*not* need to be unpinned.
Next commits will add primitives for maintaing that flag and convert the
common helpers to those. After that - a long series of per-filesystem
patches converting to those primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:15:35 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
introduce a flag for explicitly marking persistently pinned dentries
Some filesystems use a kinda-sorta controlled dentry refcount leak to pin
dentries of created objects in dcache (and undo it when removing those).
Reference is grabbed and not released, but it's not actually _stored_
anywhere. That works, but it's hard to follow and verify; among other
things, we have no way to tell _which_ of the increments is intended
to be an unpaired one. Worse, on removal we need to decide whether
the reference had already been dropped, which can be non-trivial if
that removal is on umount and we need to figure out if this dentry is
pinned due to e.g. unlink() not done. Usually that is handled by using
kill_litter_super() as ->kill_sb(), but there are open-coded special
cases of the same (consider e.g. /proc/self).
Things get simpler if we introduce a new dentry flag (DCACHE_PERSISTENT)
marking those "leaked" dentries. Having it set claims responsibility
for +1 in refcount.
The end result this series is aiming for:
* get these unbalanced dget() and dput() replaced with new primitives that
would, in addition to adjusting refcount, set and clear persistency flag.
* instead of having kill_litter_super() mess with removing the remaining
"leaked" references (e.g. for all tmpfs files that hadn't been removed
prior to umount), have the regular shrink_dcache_for_umount() strip
DCACHE_PERSISTENT of all dentries, dropping the corresponding
reference if it had been set. After that kill_litter_super() becomes
an equivalent of kill_anon_super().
Doing that in a single step is not feasible - it would affect too many places
in too many filesystems. It has to be split into a series.
Here we
* introduce the new flag
* teach shrink_dcache_for_umount() to handle it (i.e. remove
and drop refcount on anything that survives to umount with that flag
still set)
* teach kill_litter_super() that anything with that flag does
*not* need to be unpinned.
Next commits will add primitives for maintaing that flag and convert the
common helpers to those. After that - a long series of per-filesystem
patches converting to those primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:15:35 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
introduce a flag for explicitly marking persistently pinned dentries
Some filesystems use a kinda-sorta controlled dentry refcount leak to pin
dentries of created objects in dcache (and undo it when removing those).
Reference is grabbed and not released, but it's not actually _stored_
anywhere. That works, but it's hard to follow and verify; among other
things, we have no way to tell _which_ of the increments is intended
to be an unpaired one. Worse, on removal we need to decide whether
the reference had already been dropped, which can be non-trivial if
that removal is on umount and we need to figure out if this dentry is
pinned due to e.g. unlink() not done. Usually that is handled by using
kill_litter_super() as ->kill_sb(), but there are open-coded special
cases of the same (consider e.g. /proc/self).
Things get simpler if we introduce a new dentry flag (DCACHE_PERSISTENT)
marking those "leaked" dentries. Having it set claims responsibility
for +1 in refcount.
The end result this series is aiming for:
* get these unbalanced dget() and dput() replaced with new primitives that
would, in addition to adjusting refcount, set and clear persistency flag.
* instead of having kill_litter_super() mess with removing the remaining
"leaked" references (e.g. for all tmpfs files that hadn't been removed
prior to umount), have the regular shrink_dcache_for_umount() strip
DCACHE_PERSISTENT of all dentries, dropping the corresponding
reference if it had been set. After that kill_litter_super() becomes
an equivalent of kill_anon_super().
Doing that in a single step is not feasible - it would affect too many places
in too many filesystems. It has to be split into a series.
Here we
* introduce the new flag
* teach shrink_dcache_for_umount() to handle it (i.e. remove
and drop refcount on anything that survives to umount with that flag
still set)
* teach kill_litter_super() that anything with that flag does
*not* need to be unpinned.
Next commits will add primitives for maintaing that flag and convert the
common helpers to those. After that - a long series of per-filesystem
patches converting to those primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:15:35 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
2026-04-16 11:50:41 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
introduce a flag for explicitly marking persistently pinned dentries
Some filesystems use a kinda-sorta controlled dentry refcount leak to pin
dentries of created objects in dcache (and undo it when removing those).
Reference is grabbed and not released, but it's not actually _stored_
anywhere. That works, but it's hard to follow and verify; among other
things, we have no way to tell _which_ of the increments is intended
to be an unpaired one. Worse, on removal we need to decide whether
the reference had already been dropped, which can be non-trivial if
that removal is on umount and we need to figure out if this dentry is
pinned due to e.g. unlink() not done. Usually that is handled by using
kill_litter_super() as ->kill_sb(), but there are open-coded special
cases of the same (consider e.g. /proc/self).
Things get simpler if we introduce a new dentry flag (DCACHE_PERSISTENT)
marking those "leaked" dentries. Having it set claims responsibility
for +1 in refcount.
The end result this series is aiming for:
* get these unbalanced dget() and dput() replaced with new primitives that
would, in addition to adjusting refcount, set and clear persistency flag.
* instead of having kill_litter_super() mess with removing the remaining
"leaked" references (e.g. for all tmpfs files that hadn't been removed
prior to umount), have the regular shrink_dcache_for_umount() strip
DCACHE_PERSISTENT of all dentries, dropping the corresponding
reference if it had been set. After that kill_litter_super() becomes
an equivalent of kill_anon_super().
Doing that in a single step is not feasible - it would affect too many places
in too many filesystems. It has to be split into a series.
Here we
* introduce the new flag
* teach shrink_dcache_for_umount() to handle it (i.e. remove
and drop refcount on anything that survives to umount with that flag
still set)
* teach kill_litter_super() that anything with that flag does
*not* need to be unpinned.
Next commits will add primitives for maintaing that flag and convert the
common helpers to those. After that - a long series of per-filesystem
patches converting to those primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:15:35 -05:00
|
|
|
|
|
|
|
|
|
2018-04-19 23:58:48 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:35 +02:00
|
|
|
|
|
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
2026-04-11 04:01:28 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
simplify safety for lock_for_kill() slowpath
rcu_read_lock() scopes in dentry eviction machinery are too wide
and badly structured; we end up with too many of those, quite
a few essentially identical. Worse, quite a few of the function
involved are not neutral wrt that, making them harder to reason about.
rcu_read_lock() scope is not the only thing establishing an
RCU read-side critical area - spin_lock scope does the same and
they can be mixed - the sequence
rcu_read_lock()
...
spin_lock()
...
rcu_read_unlock()
...
rcu_read_lock()
...
spun_unlock()
...
rcu_read_unlock()
is an unbroken RCU read-side critical area.
Use of that observation allows to simplify things. First of all,
lock_for_kill() relies upon being in an unbroken RCU read-side
critical area. It's always called with ->d_lock held, and normally
returns without having ever dropped that spinlock. We would not
need rcu_read_lock() at all, if not for the slow path - if trylock
of inode->i_lock fails, we need to drop and retake ->d_lock.
Having all calls of lock_for_kill() inside an rcu_read_lock() scope
takes care of that, but to show that lock_for_kill() slow path is safe,
we need to demonstrate such rcu_read_lock() scope for any call chain
leading to lock_for_kill(). Which is not fun, seeing that there are
10 such scopes, with 5 distinct beginnings between them.
Case 1: opens in dput() proceeds through fast_dput() grabbing ->d_lock,
returning false into dput() and there a call of finish_dput() which calls
dentry_kill(), which calls lock_for_kill(); ends in dentry_kill(), either
right after lock_for_kill() success or right after dropping ->d_lock
on lock_for_kill() failure. ->d_lock is held continuously all the way
into lock_for_kill().
Case 2: opens in dentry_kill(), where we proceed to the same call of
dentry_kill() as in case 1. ->d_lock is held since before the
beginning of the scope and all the way into lock_for_kill().
Case 3: opens in select_collect2(), proceeds through the return to
d_walk() and to shrink_dcache_tree() where we grab ->d_lock and
proceed to call shrink_kill(), which calls dentry_kill(), then as
in the previous scopes.
Case 4: opens in shrink_dentry_list(), followed by call of shrink_kill(),
then same as in case 3. ->d_lock is held since before the beginning
of the scope and all the way into lock_for_kill().
Case 5: opens in shrink_kill(), where it's immediately followed by
call of dentry_kill(), then same as in the previous scopes. ->d_lock
is held since before the beginning of the scope all the way into
lock_for_kill().
Note that in cases 2, 4 and 5 the slow path of lock_for_kill() is the
only part of rcu_read_lock() scope that is not covered by spinlock
scopes. In case 1 we have the area in fast_dput() as well and in
case 3 - the return path from select_collect2() and chunk in shrink_dcache_tree()
up to grabbing ->d_lock.
Seeing that the reasons we need rcu_read_lock() in these additional
areas are completely unrelated to lock_for_kill() slow path, the things
get much more straightforward with
* explicit rcu_read_lock() scope surrounding the area in slow path
of lock_for_kill() where ->d_lock is not held
* shrink_dentry_list() dropping rcu_read_lock() as soon as it has
grabbed ->d_lock.
* dput() dropping rcu_read_lock() just before calling finish_dput().
* rcu_read_lock() calls in finish_dput(), shrink_kill() and
shrink_dentry_list() are removed, along with rcu_read_unlock() calls in
dentry_kill().
RCU read-side critical areas are unchanged by that, safety of lock_for_kill()
slow path is trivial to verify and a bunch of rcu_read_lock() scopes either
gone or become easier to describe.
Update the comments on locking conventions and memory safety considerations,
including the NORCU case.
Incidentally, all calls of fast_dput() are immediately preceded by rcu_read_lock()
and followed by rcu_read_unlock() now, which will allow to simplify those on
the next step...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-11 03:24:28 -04:00
|
|
|
|
2026-04-11 04:01:28 -04:00
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
2026-04-16 11:50:41 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
Teach shrink_dcache_parent() to cope with mixed-filesystem shrink lists
Currently, running into a shrink list that contains dentries from different
filesystems can cause several unpleasant things for shrink_dcache_parent()
and for umount(2).
The first problem is that there's a window during shrink_dentry_list() between
__dentry_kill() takes a victim out and dropping reference to its parent. During
that window the parent looks like a genuine busy dentry. shrink_dcache_parent()
(or, worse yet, shrink_dcache_for_umount()) coming at that time will see no
eviction candidates and no indication that it needs to wait for some
shrink_dentry_list() to proceed further.
That applies for any shrink list that might intersect with the subtree we are
trying to shrink; the only reason it does not blow on umount(2) in the mainline
is that we unregister the memory shrinker before hitting shrink_dcache_for_umount().
Another problem happens if something in a mixed-filesystem shrink list gets
be stuck in e.g. iput(), getting umount of unrelated fs to spin waiting for
the stuck shrinker to get around to our dentries.
Solution:
1) have shrink_dentry_list() decrement the parent's refcount and
make sure it's on a shrink list (ours unless it already had been on some
other) before calling __dentry_kill(). That eliminates the window when
shrink_dcache_parent() would've blown past the entire subtree without
noticing anything with zero refcount not on shrink lists.
2) when shrink_dcache_parent() has found no eviction candidates,
but some dentries are still sitting on shrink lists, rather than
repeating the scan in hope that shrinkers have progressed, scan looking
for something on shrink lists with zero refcount. If such a thing is
found, grab rcu_read_lock() and stop the scan, with caller locking
it for eviction, dropping out of RCU and doing __dentry_kill(), with
the same treatment for parent as shrink_dentry_list() would do.
Note that right now mixed-filesystem shrink lists do not occur, so this
is not a mainline bug. Howevere, there's a bunch of uses for such
beasts (e.g. the "try and evict everything we can out of given page"
patches; there are potential uses in mount-related code, considerably
simplifying the life in fs/namespace.c, etc.)
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-06-29 18:31:24 -04:00
|
|
|
|
|
|
|
|
|
2026-04-16 11:50:41 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs/dcache.c: add cond_resched() to shrink_dcache_parent()
Call cond_resched() in shrink_dcache_parent() to maintain interactivity.
Before this patch:
void shrink_dcache_parent(struct dentry * parent)
{
while ((found = select_parent(parent, &dispose)) != 0)
shrink_dentry_list(&dispose);
}
select_parent() populates the dispose list with dentries which
shrink_dentry_list() then deletes. select_parent() carefully uses
need_resched() to avoid doing too much work at once. But neither
shrink_dcache_parent() nor its called functions call cond_resched(). So
once need_resched() is set select_parent() will return single dentry
dispose list which is then deleted by shrink_dentry_list(). This is
inefficient when there are a lot of dentry to process. This can cause
softlockup and hurts interactivity on non preemptable kernels.
This change adds cond_resched() in shrink_dcache_parent(). The benefit
of this is that need_resched() is quickly cleared so that future calls
to select_parent() are able to efficiently return a big batch of dentry.
These additional cond_resched() do not seem to impact performance, at
least for the workload below.
Here is a program which can cause soft lockup if other system activity
sets need_resched().
int main()
{
struct rlimit rlim;
int i;
int f[100000];
char buf[20];
struct timeval t1, t2;
double diff;
/* cleanup past run */
system("rm -rf x");
/* boost nfile rlimit */
rlim.rlim_cur = 200000;
rlim.rlim_max = 200000;
if (setrlimit(RLIMIT_NOFILE, &rlim))
err(1, "setrlimit");
/* make directory for files */
if (mkdir("x", 0700))
err(1, "mkdir");
if (gettimeofday(&t1, NULL))
err(1, "gettimeofday");
/* populate directory with open files */
for (i = 0; i < 100000; i++) {
snprintf(buf, sizeof(buf), "x/%d", i);
f[i] = open(buf, O_CREAT);
if (f[i] == -1)
err(1, "open");
}
/* close some of the files */
for (i = 0; i < 85000; i++)
close(f[i]);
/* unlink all files, even open ones */
system("rm -rf x");
if (gettimeofday(&t2, NULL))
err(1, "gettimeofday");
diff = (((double)t2.tv_sec * 1000000 + t2.tv_usec) -
((double)t1.tv_sec * 1000000 + t1.tv_usec));
printf("done: %g elapsed\n", diff/1e6);
return 0;
}
Signed-off-by: Greg Thelen <gthelen@google.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-04-30 15:26:48 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
introduce a flag for explicitly marking persistently pinned dentries
Some filesystems use a kinda-sorta controlled dentry refcount leak to pin
dentries of created objects in dcache (and undo it when removing those).
Reference is grabbed and not released, but it's not actually _stored_
anywhere. That works, but it's hard to follow and verify; among other
things, we have no way to tell _which_ of the increments is intended
to be an unpaired one. Worse, on removal we need to decide whether
the reference had already been dropped, which can be non-trivial if
that removal is on umount and we need to figure out if this dentry is
pinned due to e.g. unlink() not done. Usually that is handled by using
kill_litter_super() as ->kill_sb(), but there are open-coded special
cases of the same (consider e.g. /proc/self).
Things get simpler if we introduce a new dentry flag (DCACHE_PERSISTENT)
marking those "leaked" dentries. Having it set claims responsibility
for +1 in refcount.
The end result this series is aiming for:
* get these unbalanced dget() and dput() replaced with new primitives that
would, in addition to adjusting refcount, set and clear persistency flag.
* instead of having kill_litter_super() mess with removing the remaining
"leaked" references (e.g. for all tmpfs files that hadn't been removed
prior to umount), have the regular shrink_dcache_for_umount() strip
DCACHE_PERSISTENT of all dentries, dropping the corresponding
reference if it had been set. After that kill_litter_super() becomes
an equivalent of kill_anon_super().
Doing that in a single step is not feasible - it would affect too many places
in too many filesystems. It has to be split into a series.
Here we
* introduce the new flag
* teach shrink_dcache_for_umount() to handle it (i.e. remove
and drop refcount on anything that survives to umount with that flag
still set)
* teach kill_litter_super() that anything with that flag does
*not* need to be unpinned.
Next commits will add primitives for maintaing that flag and convert the
common helpers to those. After that - a long series of per-filesystem
patches converting to those primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:15:35 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-03-04 10:32:42 -05:00
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
|
|
|
|
|
2026-03-04 10:32:42 -05:00
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
introduce a flag for explicitly marking persistently pinned dentries
Some filesystems use a kinda-sorta controlled dentry refcount leak to pin
dentries of created objects in dcache (and undo it when removing those).
Reference is grabbed and not released, but it's not actually _stored_
anywhere. That works, but it's hard to follow and verify; among other
things, we have no way to tell _which_ of the increments is intended
to be an unpaired one. Worse, on removal we need to decide whether
the reference had already been dropped, which can be non-trivial if
that removal is on umount and we need to figure out if this dentry is
pinned due to e.g. unlink() not done. Usually that is handled by using
kill_litter_super() as ->kill_sb(), but there are open-coded special
cases of the same (consider e.g. /proc/self).
Things get simpler if we introduce a new dentry flag (DCACHE_PERSISTENT)
marking those "leaked" dentries. Having it set claims responsibility
for +1 in refcount.
The end result this series is aiming for:
* get these unbalanced dget() and dput() replaced with new primitives that
would, in addition to adjusting refcount, set and clear persistency flag.
* instead of having kill_litter_super() mess with removing the remaining
"leaked" references (e.g. for all tmpfs files that hadn't been removed
prior to umount), have the regular shrink_dcache_for_umount() strip
DCACHE_PERSISTENT of all dentries, dropping the corresponding
reference if it had been set. After that kill_litter_super() becomes
an equivalent of kill_anon_super().
Doing that in a single step is not feasible - it would affect too many places
in too many filesystems. It has to be split into a series.
Here we
* introduce the new flag
* teach shrink_dcache_for_umount() to handle it (i.e. remove
and drop refcount on anything that survives to umount with that flag
still set)
* teach kill_litter_super() that anything with that flag does
*not* need to be unpinned.
Next commits will add primitives for maintaing that flag and convert the
common helpers to those. After that - a long series of per-filesystem
patches converting to those primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:15:35 -05:00
|
|
|
|
2018-04-15 18:27:23 -04:00
|
|
|
|
dcache: keep shrink_dcache_for_umount() making progress on busy roots
Commit e9895609cb7f ("wind ->s_roots via ->d_sib instead of ->d_hash")
moved secondary roots from ->d_hash to ->d_sib. Secondary roots are now
d_unhashed(), so __d_drop() returns without removing them from
->s_roots. Consequently, d_drop() in do_one_tree() no longer
guarantees progress through the list.
If a secondary root is still busy once do_one_tree() is done with it,
its final dput() cannot evict it. The root remains ->s_roots.first and
the loop selects it forever, holding ->s_umount for write and repeatedly
reporting the same dentry.
The root does not need a leaked reference of its own for that. Every
child pins its parent (d_alloc() takes a reference on it) and
umount_check() deliberately reports a busy descendant instead of
complaining about its ancestors, so a single leaked dentry reference
anywhere below a secondary root is enough. For filesystems that build
->s_root with d_obtain_root() - nfs, ceph, nilfs2 snapshot mounts -
that is the entire tree.
Before e9895609cb7f, ___d_drop() special-cased IS_ROOT dentries and
removed them from ->s_roots regardless of their refcount, so the
d_drop() in do_one_tree() detached the root from the superblock no
matter what. Commit 9c8c10e262e0 ("more graceful recovery in
umount_collect()") deliberately made busy dentries nonfatal: report
them and finish the unmount rather than BUG() while holding
->s_umount.
Restore that by detaching the root in do_one_tree() itself, next to
the d_drop() that used to do it. That covers both callers - the
->s_roots loop and ->s_root, which for the filesystems above is a
secondary root as well. In the normal case dentry_unlist() finds
->d_sib already unhashed when eviction occurs.
A permanently leaked reference remains leaked after unmount, as it did
before e9895609cb7f; if the extra reference is merely delayed, its
final dput() may run after teardown has advanced. Leaving the root on
->s_roots is not an alternative: the superblock would then be freed
with a live dentry still linked into it, and that dentry's
dentry_unlist() would take ->s_roots_lock on freed memory.
Christian Brauner <brauner@kernel.org> says:
Moved the ->s_roots removal from the shrink_dcache_for_umount() loop
into do_one_tree(), so a busy ->s_root obtained from d_obtain_root() is
detached on the first pass instead of being reported a second time when
the loop picks it off ->s_roots. Extended the commit message with the
pinned-ancestor case.
Fixes: e9895609cb7f ("wind ->s_roots via ->d_sib instead of ->d_hash")
Signed-off-by: Karl Mehltretter <kmehltretter@gmail.com>
Link: https://patch.msgid.link/20260729005933.15858-1-kmehltretter@gmail.com
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-07-29 02:59:33 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-02 14:37:19 +02:00
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-05-02 20:36:10 -04:00
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-11-08 12:31:16 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2023-11-06 14:44:17 +01:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
|
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-02-13 09:39:37 -08:00
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2014-02-13 09:46:25 -08:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2014-02-13 09:39:37 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-02-13 09:46:25 -08:00
|
|
|
|
2014-02-13 09:39:37 -08:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2014-02-13 09:39:37 -08:00
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2014-02-13 09:46:25 -08:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
2018-04-15 18:27:23 -04:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-06-03 07:20:09 +01:00
|
|
|
|
vfs: Lazily remove mounts on unlinked files and directories.
With the introduction of mount namespaces and bind mounts it became
possible to access files and directories that on some paths are mount
points but are not mount points on other paths. It is very confusing
when rm -rf somedir returns -EBUSY simply because somedir is mounted
somewhere else. With the addition of user namespaces allowing
unprivileged mounts this condition has gone from annoying to allowing
a DOS attack on other users in the system.
The possibility for mischief is removed by updating the vfs to support
rename, unlink and rmdir on a dentry that is a mountpoint and by
lazily unmounting mountpoints on deleted dentries.
In particular this change allows rename, unlink and rmdir system calls
on a dentry without a mountpoint in the current mount namespace to
succeed, and it allows rename, unlink, and rmdir performed on a
distributed filesystem to update the vfs cache even if when there is a
mount in some namespace on the original dentry.
There are two common patterns of maintaining mounts: Mounts on trusted
paths with the parent directory of the mount point and all ancestory
directories up to / owned by root and modifiable only by root
(i.e. /media/xxx, /dev, /dev/pts, /proc, /sys, /sys/fs/cgroup/{cpu,
cpuacct, ...}, /usr, /usr/local). Mounts on unprivileged directories
maintained by fusermount.
In the case of mounts in trusted directories owned by root and
modifiable only by root the current parent directory permissions are
sufficient to ensure a mount point on a trusted path is not removed
or renamed by anyone other than root, even if there is a context
where the there are no mount points to prevent this.
In the case of mounts in directories owned by less privileged users
races with users modifying the path of a mount point are already a
danger. fusermount already uses a combination of chdir,
/proc/<pid>/fd/NNN, and UMOUNT_NOFOLLOW to prevent these races. The
removable of global rename, unlink, and rmdir protection really adds
nothing new to consider only a widening of the attack window, and
fusermount is already safe against unprivileged users modifying the
directory simultaneously.
In principle for perfect userspace programs returning -EBUSY for
unlink, rmdir, and rename of dentires that have mounts in the local
namespace is actually unnecessary. Unfortunately not all userspace
programs are perfect so retaining -EBUSY for unlink, rmdir and rename
of dentries that have mounts in the current mount namespace plays an
important role of maintaining consistency with historical behavior and
making imperfect userspace applications hard to exploit.
v2: Remove spurious old_dentry.
v3: Optimized shrink_submounts_and_drop
Removed unsued afs label
v4: Simplified the changes to check_submounts_and_drop
Do not rename check_submounts_and_drop shrink_submounts_and_drop
Document what why we need atomicity in check_submounts_and_drop
Rely on the parent inode mutex to make d_revalidate and d_invalidate
an atomic unit.
v5: Refcount the mountpoint to detach in case of simultaneous
renames.
Reviewed-by: Miklos Szeredi <miklos@szeredi.hu>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-10-01 18:33:48 -07:00
|
|
|
|
2018-04-15 18:21:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
|
|
|
|
|
2014-02-13 09:39:37 -08:00
|
|
|
|
2013-09-05 11:44:36 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2026-01-28 06:31:50 -08:00
|
|
|
|
2011-07-07 15:03:58 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-10-25 14:08:24 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-28 11:47:32 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-03-22 14:41:09 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-21 16:14:04 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
dcache: unpoison the inline name buffer in __d_alloc()
syzbot reported:
BUG: KMSAN: uninit-value in dentry_string_cmp fs/dcache.c:291 [inline]
BUG: KMSAN: uninit-value in dentry_cmp fs/dcache.c:322 [inline]
BUG: KMSAN: uninit-value in __d_lookup_rcu+0x37d/0x5e0 fs/dcache.c:2522
dentry_string_cmp fs/dcache.c:291 [inline]
dentry_cmp fs/dcache.c:322 [inline]
__d_lookup_rcu+0x37d/0x5e0 fs/dcache.c:2522
lookup_fast+0x194/0xa40 fs/namei.c:1854
lookup_fast_for_open fs/namei.c:4545 [inline]
open_last_lookups fs/namei.c:4579 [inline]
path_openat+0x9ef/0x6540 fs/namei.c:4856
do_file_open+0x2aa/0x680 fs/namei.c:4888
do_sys_openat2+0x17c/0x390 fs/open.c:1395
do_sys_open fs/open.c:1401 [inline]
__do_sys_openat fs/open.c:1417 [inline]
__se_sys_openat fs/open.c:1412 [inline]
__x64_sys_openat+0x240/0x300 fs/open.c:1412
x64_sys_call+0x2445/0x3ea0 arch/x86/include/generated/asm/syscalls_64.h:258
do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
do_syscall_64+0x15d/0x3c0 arch/x86/entry/syscall_64.c:94
entry_SYSCALL_64_after_hwframe+0x77/0x7f
Uninit was stored to memory at:
copy_name fs/dcache.c:3031 [inline]
__d_move+0xd29/0x21f0 fs/dcache.c:3099
d_move+0x71/0xf0 fs/dcache.c:3147
vfs_rename+0x2619/0x2770 fs/namei.c:6085
filename_renameat2+0xa59/0x1230 fs/namei.c:6188
__do_sys_rename fs/namei.c:6232 [inline]
__se_sys_rename+0xc5/0x5c0 fs/namei.c:6228
__x64_sys_rename+0x78/0xb0 fs/namei.c:6228
x64_sys_call+0x329/0x3ea0 arch/x86/include/generated/asm/syscalls_64.h:83
do_syscall_x64 arch/x86/entry/syscall_64.c:63
do_syscall_64+0x15d/0x3c0 arch/x86/entry/syscall_64.c:94
entry_SYSCALL_64_after_hwframe+0x77/0x7f
Uninit was created at:
slab_post_alloc_hook mm/slub.c:4617 [inline]
slab_alloc_node mm/slub.c:4939 [inline]
kmem_cache_alloc_lru_noprof+0x376/0x1230 mm/s
__d_alloc+0x52/0x9f0 fs/dcache.c:1902
d_alloc+0x57/0x300 fs/dcache.c:1981
lookup_one_qstr_excl+0x19d/0x7a0 fs/namei.c:1806
__start_renaming+0x341/0x850 fs/namei.c:3888
filename_renameat2+0x625/0x1230 fs/namei.c:6163
__do_sys_rename fs/namei.c:6232 [inline]
__se_sys_rename+0xc5/0x5c0 fs/namei.c:6228
__x64_sys_rename+0x78/0xb0 fs/namei.c:6228
x64_sys_call+0x329/0x3ea0 arch/x86/include/generated/asm/syscalls_64.h:83
do_syscall_x64 arch/x86/entry/syscall_64.c:63
do_syscall_64+0x15d/0x3c0 arch/x86/entry/syscall_64.c:94
entry_SYSCALL_64_after_hwframe+0x77/0x7f
The race is between a concurrent open() and rename() of the same path.
__d_alloc() only stores the name itself and its terminating NUL, so the
rest of the inline buffer (d_shortname, DNAME_INLINE_LEN bytes) is left
uninitialized. copy_name(), called from rename(), copies that buffer as a
whole, so the uninitialized tail is propagated into the dentry that is
being moved. Meanwhile __d_lookup_rcu(), called from open(), is an
optimistic lockless lookup: it checks d_name.hash_len first and leaves the
seqcount retry to its caller, so it can end up comparing against a dentry
whose name a rename is rewriting in place, using a stale (longer) length.
The comparison then runs past the terminating NUL and reads bytes of the
uninitialized tail, which KMSAN reports.
The read is harmless by design: it stays inside the buffer, the name is
still NUL-terminated, and the result is thrown away by the seqcount retry.
It is not specific to KMSAN either - with CONFIG_DCACHE_WORD_ACCESS
enabled the very same bytes are read by read_word_at_a_time(), which is
__no_sanitize_or_inline and therefore invisible to KMSAN. KMSAN builds
only see the instrumented byte-at-a-time dentry_string_cmp() because
CONFIG_DCACHE_WORD_ACCESS is disabled when KMSAN is enabled on x86:
commit 7cf8f44a5a1c ("x86: fs: kmsan: disable CONFIG_DCACHE_WORD_ACCESS")
Zeroing the inline buffer would hide the report, but it would add a
memset() to a hot allocation path just to initialize bytes that are never
used as part of a name. Instead, tell KMSAN the inline buffer is
initialized: kmsan_unpoison_memory() compiles to nothing unless
CONFIG_KMSAN is set, and doing it at allocation time is enough for every
dentry, because copy_name() and swap_names() copy the whole buffer and
thus propagate its shadow.
Reported-by: syzbot+7ff3adde89dd795ad4c4@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=7ff3adde89dd795ad4c4
Signed-off-by: Drif Abdelmalek Mohamed Said <drifabdelmalekmohamedsaid@gmail.com>
Changes in v3:
- Annotate for KMSAN instead of zeroing, as suggested in review: the read
is harmless, so unpoison the inline buffer in __d_alloc() with
kmsan_unpoison_memory() (a no-op unless CONFIG_KMSAN) rather than adding
a memset() to the dentry allocation path.
- Document why only KMSAN builds report this at all: with
CONFIG_DCACHE_WORD_ACCESS the same read goes through
read_word_at_a_time(), which KMSAN does not instrument.
- Rewrite the commit message; the previous one had several truncated
lines.
Link: https://patch.msgid.link/20260918224204.3056-1-drifabdelmalekmohamedsaid@gmail.com
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Christian Brauner (Amutable) <brauner@kernel.org>
2026-09-18 23:42:04 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-24 20:38:43 -04:00
|
|
|
|
2017-07-04 17:25:22 +01:00
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2016-03-24 20:38:43 -04:00
|
|
|
|
2026-04-17 11:42:40 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-10-26 15:05:41 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
2018-10-26 15:05:41 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-05-21 16:14:04 -07:00
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2012-05-21 16:14:04 -07:00
|
|
|
|
vfs: get rid of insane dentry hashing rules
The dentry hashing rules have been really quite complicated for a long
while, in odd ways. That made functions like __d_drop() very fragile
and non-obvious.
In particular, whether a dentry was hashed or not was indicated with an
explicit DCACHE_UNHASHED bit. That's despite the fact that the hash
abstraction that the dentries use actually have a 'is this entry hashed
or not' model (which is a simple test of the 'pprev' pointer).
The reason that was done is because we used the normal 'is this entry
unhashed' model to mark whether the dentry had _ever_ been hashed in the
dentry hash tables, and that logic goes back many years (commit
b3423415fbc2: "dcache: avoid RCU for never-hashed dentries").
That, in turn, meant that __d_drop had totally different unhashing logic
for the dentry hash table case and for the anonymous dcache case,
because in order to use the "is this dentry hashed" logic as a flag for
whether it had ever been on the RCU hash table, we had to unhash such a
dentry differently so that we'd never think that it wasn't 'unhashed'
and wouldn't be free'd correctly.
That's just insane. It made the logic really hard to follow, when there
were two different kinds of "unhashed" states, and one of them (the one
that used "list_bl_unhashed()") really had nothing at all to do with
being unhashed per se, but with a very subtle lifetime rule instead.
So turn all of it around, and make it logical.
Instead of having a DENTRY_UNHASHED bit in d_flags to indicate whether
the dentry is on the hash chains or not, use the hash chain unhashed
logic for that. Suddenly "d_unhashed()" just uses "list_bl_unhashed()",
and everything makes sense.
And for the lifetime rule, just use an explicit DENTRY_RCUACCEES bit.
If we ever insert the dentry into the dentry hash table so that it is
visible to RCU lookup, we mark it DENTRY_RCUACCESS to show that it now
needs the RCU lifetime rules. Now suddently that test at dentry free
time makes sense too.
And because unhashing now is sane and doesn't depend on where the dentry
got unhashed from (because the dentry hash chain details doesn't have
some subtle side effects), we can re-unify the __d_drop() logic and use
common code for the unhashing.
Also fix one more open-coded hash chain bit_spin_lock() that I missed in
the previous chain locking cleanup commit.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-04-24 07:58:46 -07:00
|
|
|
|
2025-01-30 14:56:23 +01:00
|
|
|
|
2020-07-20 17:55:24 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2011-07-07 15:03:58 -04:00
|
|
|
|
|
|
|
|
|
2025-02-23 19:53:00 -05:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2016-06-28 11:47:32 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: use fast counters for vfs caches
percpu_counter library generates quite nasty code, so unless you need
to dynamically allocate counters or take fast approximate value, a
simple per cpu set of counters is much better.
The percpu_counter can never be made to work as well, because it has an
indirection from pointer to percpu memory, and it can't use direct
this_cpu_inc interfaces because it doesn't use static PER_CPU data, so
code will always be worse.
In the fastpath, it is the difference between this:
incl %gs:nr_dentry # nr_dentry
and this:
movl percpu_counter_batch(%rip), %edx # percpu_counter_batch,
movl $1, %esi #,
movq $nr_dentry, %rdi #,
call __percpu_counter_add # (plus I clobber registers)
__percpu_counter_add:
pushq %rbp #
movq %rsp, %rbp #,
subq $32, %rsp #,
movq %rbx, -24(%rbp) #,
movq %r12, -16(%rbp) #,
movq %r13, -8(%rbp) #,
movq %rdi, %rbx # fbc, fbc
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
incl -8124(%rax) # <variable>.preempt_count
movq 32(%rdi), %r12 # <variable>.counters, tcp_ptr__
#APP
# 78 "lib/percpu_counter.c" 1
add %gs:this_cpu_off, %r12 # this_cpu_off, tcp_ptr__
# 0 "" 2
#NO_APP
movslq (%r12),%r13 #* tcp_ptr__, tmp73
movslq %edx,%rax # batch, batch
addq %rsi, %r13 # amount, count
cmpq %rax, %r13 # batch, count
jge .L27 #,
negl %edx # tmp76
movslq %edx,%rdx # tmp76, tmp77
cmpq %rdx, %r13 # tmp77, count
jg .L28 #,
.L27:
movq %rbx, %rdi # fbc,
call _raw_spin_lock #
addq %r13, 8(%rbx) # count, <variable>.count
movq %rbx, %rdi # fbc,
movl $0, (%r12) #,* tcp_ptr__
call _raw_spin_unlock #
.L29:
#APP
# 216 "/home/npiggin/usr/src/linux-2.6/arch/x86/include/asm/thread_info.h" 1
movq %gs:kernel_stack,%rax #, pfo_ret__
# 0 "" 2
#NO_APP
decl -8124(%rax) # <variable>.preempt_count
movq -8136(%rax), %rax #, D.14625
testb $8, %al #, D.14625
jne .L32 #,
.L31:
movq -24(%rbp), %rbx #,
movq -16(%rbp), %r12 #,
movq -8(%rbp), %r13 #,
leave
ret
.p2align 4,,10
.p2align 3
.L28:
movl %r13d, (%r12) # count,*
jmp .L29 #
.L32:
call preempt_schedule #
.p2align 4,,6
jmp .L31 #
.size __percpu_counter_add, .-__percpu_counter_add
.p2align 4,,15
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:19 +11:00
|
|
|
|
2010-10-10 05:36:23 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2011-07-07 15:03:58 -04:00
|
|
|
|
|
|
|
|
|
2026-01-28 06:31:50 -08:00
|
|
|
|
2011-07-07 15:03:58 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-06 14:44:17 +01:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
2011-07-07 15:03:58 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-01-19 11:39:52 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
2018-01-19 11:39:52 +01:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
2026-05-05 00:20:19 -04:00
|
|
|
|
2016-06-10 11:32:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-06-29 16:20:47 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-03-15 22:58:11 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-06-29 16:20:47 -04:00
|
|
|
|
2011-01-07 17:50:07 +11:00
|
|
|
|
|
|
|
|
|
2023-11-12 00:04:46 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
2023-11-12 00:04:46 -05:00
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
2025-02-23 20:18:15 -05:00
|
|
|
|
2025-02-23 19:39:47 -05:00
|
|
|
|
2025-02-23 20:18:15 -05:00
|
|
|
|
2023-11-12 00:04:46 -05:00
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
2011-01-07 17:50:07 +11:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-10 07:51:30 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2009-09-29 20:09:42 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-02-23 15:39:26 -05:00
|
|
|
|
|
|
|
|
|
2025-02-23 15:39:26 -05:00
|
|
|
|
|
|
|
|
|
2025-02-23 15:39:26 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-23 20:18:15 -05:00
|
|
|
|
2011-01-07 17:49:55 +11:00
|
|
|
|
2025-02-23 15:39:26 -05:00
|
|
|
|
2011-01-14 13:26:18 -08:00
|
|
|
|
2025-02-23 15:39:26 -05:00
|
|
|
|
2011-01-07 17:49:55 +11:00
|
|
|
|
2025-02-23 15:39:26 -05:00
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:55 +11:00
|
|
|
|
|
|
|
|
|
2025-02-23 19:39:47 -05:00
|
|
|
|
|
|
|
|
|
2025-02-23 19:53:00 -05:00
|
|
|
|
2025-02-23 19:39:47 -05:00
|
|
|
|
2025-02-23 19:53:00 -05:00
|
|
|
|
2025-02-23 19:39:47 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-12 19:22:53 +01:00
|
|
|
|
|
|
|
|
|
2015-01-29 12:02:29 +00:00
|
|
|
|
2013-09-12 19:22:53 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-01-29 12:02:29 +00:00
|
|
|
|
2013-09-12 19:22:53 +01:00
|
|
|
|
|
|
|
|
|
2015-01-29 12:02:29 +00:00
|
|
|
|
2015-11-17 10:20:54 -05:00
|
|
|
|
2015-01-29 12:02:29 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-09-12 19:22:53 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
|
|
|
|
|
2013-09-12 19:22:53 +01:00
|
|
|
|
2016-04-14 19:52:13 -04:00
|
|
|
|
2013-09-12 19:22:53 +01:00
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
vfs: don't mod negative dentry count when on shrinker list
The nr_dentry_negative counter is intended to only account negative
dentries that are present on the superblock LRU. Therefore, the LRU
add, remove and isolate helpers modify the counter based on whether
the dentry is negative, but the shrinker list related helpers do not
modify the counter, and the paths that change a dentry between
positive and negative only do so if DCACHE_LRU_LIST is set.
The problem with this is that a dentry on a shrinker list still has
DCACHE_LRU_LIST set to indicate ->d_lru is in use. The additional
DCACHE_SHRINK_LIST flag denotes whether the dentry is on LRU or a
shrink related list. Therefore if a relevant operation (i.e. unlink)
occurs while a dentry is present on a shrinker list, and the
associated codepath only checks for DCACHE_LRU_LIST, then it is
technically possible to modify the negative dentry count for a
dentry that is off the LRU. Since the shrinker list related helpers
do not modify the negative dentry count (because non-LRU dentries
should not be included in the count) when the dentry is ultimately
removed from the shrinker list, this can cause the negative dentry
count to become permanently inaccurate.
This problem can be reproduced via a heavy file create/unlink vs.
drop_caches workload. On an 80xcpu system, I start 80 tasks each
running a 1k file create/delete loop, and one task spinning on
drop_caches. After 10 minutes or so of runtime, the idle/clean cache
negative dentry count increases from somewhere in the range of 5-10
entries to several hundred (and increasingly grows beyond
nr_dentry_unused).
Tweak the logic in the paths that turn a dentry negative or positive
to filter out the case where the dentry is present on a shrink
related list. This allows the above workload to maintain an accurate
negative dentry count.
Fixes: af0c9af1b3f6 ("fs/dcache: Track & report number of negative dentries")
Signed-off-by: Brian Foster <bfoster@redhat.com>
Link: https://lore.kernel.org/r/20240703121301.247680-1-bfoster@redhat.com
Acked-by: Ian Kent <ikent@redhat.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Waiman Long <longman@redhat.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-07-03 08:13:01 -04:00
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
vfs: don't mod negative dentry count when on shrinker list
The nr_dentry_negative counter is intended to only account negative
dentries that are present on the superblock LRU. Therefore, the LRU
add, remove and isolate helpers modify the counter based on whether
the dentry is negative, but the shrinker list related helpers do not
modify the counter, and the paths that change a dentry between
positive and negative only do so if DCACHE_LRU_LIST is set.
The problem with this is that a dentry on a shrinker list still has
DCACHE_LRU_LIST set to indicate ->d_lru is in use. The additional
DCACHE_SHRINK_LIST flag denotes whether the dentry is on LRU or a
shrink related list. Therefore if a relevant operation (i.e. unlink)
occurs while a dentry is present on a shrinker list, and the
associated codepath only checks for DCACHE_LRU_LIST, then it is
technically possible to modify the negative dentry count for a
dentry that is off the LRU. Since the shrinker list related helpers
do not modify the negative dentry count (because non-LRU dentries
should not be included in the count) when the dentry is ultimately
removed from the shrinker list, this can cause the negative dentry
count to become permanently inaccurate.
This problem can be reproduced via a heavy file create/unlink vs.
drop_caches workload. On an 80xcpu system, I start 80 tasks each
running a 1k file create/delete loop, and one task spinning on
drop_caches. After 10 minutes or so of runtime, the idle/clean cache
negative dentry count increases from somewhere in the range of 5-10
entries to several hundred (and increasingly grows beyond
nr_dentry_unused).
Tweak the logic in the paths that turn a dentry negative or positive
to filter out the case where the dentry is present on a shrink
related list. This allows the above workload to maintain an accurate
negative dentry count.
Fixes: af0c9af1b3f6 ("fs/dcache: Track & report number of negative dentries")
Signed-off-by: Brian Foster <bfoster@redhat.com>
Link: https://lore.kernel.org/r/20240703121301.247680-1-bfoster@redhat.com
Acked-by: Ian Kent <ikent@redhat.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Waiman Long <longman@redhat.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-07-03 08:13:01 -04:00
|
|
|
|
|
|
|
|
|
2019-01-30 13:52:38 -05:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2016-02-29 12:12:46 -05:00
|
|
|
|
2015-03-05 14:09:22 +00:00
|
|
|
|
2016-02-29 12:12:46 -05:00
|
|
|
|
2016-05-29 18:35:12 -04:00
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-24 01:32:07 -05:00
|
|
|
|
2016-03-09 18:05:42 -05:00
|
|
|
|
2016-04-11 00:53:26 -04:00
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
2016-03-09 18:05:42 -05:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2016-03-09 18:05:42 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-05-04 08:23:01 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-24 01:32:07 -05:00
|
|
|
|
2018-05-04 08:23:01 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
2018-05-04 08:23:01 -04:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
Coccinelle-based conversion to use ->i_state accessors
All places were patched by coccinelle with the default expecting that
->i_lock is held, afterwards entries got fixed up by hand to use
unlocked variants as needed.
The script:
@@
expression inode, flags;
@@
- inode->i_state & flags
+ inode_state_read(inode) & flags
@@
expression inode, flags;
@@
- inode->i_state &= ~flags
+ inode_state_clear(inode, flags)
@@
expression inode, flag1, flag2;
@@
- inode->i_state &= ~flag1 & ~flag2
+ inode_state_clear(inode, flag1 | flag2)
@@
expression inode, flags;
@@
- inode->i_state |= flags
+ inode_state_set(inode, flags)
@@
expression inode, flags;
@@
- inode->i_state = flags
+ inode_state_assign(inode, flags)
@@
expression inode, flags;
@@
- flags = inode->i_state
+ flags = inode_state_read(inode)
@@
expression inode, flags;
@@
- READ_ONCE(inode->i_state) & flags
+ inode_state_read(inode) & flags
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-10-09 09:59:18 +02:00
|
|
|
|
2026-04-21 20:25:38 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Coccinelle-based conversion to use ->i_state accessors
All places were patched by coccinelle with the default expecting that
->i_lock is held, afterwards entries got fixed up by hand to use
unlocked variants as needed.
The script:
@@
expression inode, flags;
@@
- inode->i_state & flags
+ inode_state_read(inode) & flags
@@
expression inode, flags;
@@
- inode->i_state &= ~flags
+ inode_state_clear(inode, flags)
@@
expression inode, flag1, flag2;
@@
- inode->i_state &= ~flag1 & ~flag2
+ inode_state_clear(inode, flag1 | flag2)
@@
expression inode, flags;
@@
- inode->i_state |= flags
+ inode_state_set(inode, flags)
@@
expression inode, flags;
@@
- inode->i_state = flags
+ inode_state_assign(inode, flags)
@@
expression inode, flags;
@@
- flags = inode->i_state
+ flags = inode_state_read(inode)
@@
expression inode, flags;
@@
- READ_ONCE(inode->i_state) & flags
+ inode_state_read(inode) & flags
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
2025-10-09 09:59:18 +02:00
|
|
|
|
2024-08-23 14:47:38 +02:00
|
|
|
|
2018-05-04 08:23:01 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-01-08 16:49:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-01-19 11:39:52 +01:00
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
2012-01-08 16:49:21 -05:00
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
2012-01-08 16:49:21 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-01-19 11:39:52 +01:00
|
|
|
|
2008-08-11 15:48:57 +02:00
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
|
|
|
|
|
2008-08-11 15:48:57 +02:00
|
|
|
|
|
|
|
|
|
2008-08-11 15:49:04 +02:00
|
|
|
|
2008-08-11 15:48:57 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
2008-08-11 15:49:12 +02:00
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-08-11 15:49:12 +02:00
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
2008-08-11 15:48:57 +02:00
|
|
|
|
2011-01-07 17:49:38 +11:00
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-08-11 15:49:12 +02:00
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-18 16:42:43 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-08-11 15:49:12 +02:00
|
|
|
|
|
|
|
|
|
2008-08-11 15:48:57 +02:00
|
|
|
|
2014-02-14 17:35:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-01-19 11:39:52 +01:00
|
|
|
|
2014-02-14 17:35:37 -05:00
|
|
|
|
2009-02-27 14:02:59 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-02-14 17:35:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-01-19 11:39:52 +01:00
|
|
|
|
2014-02-14 17:35:37 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-30 13:20:54 +02:00
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-02-24 12:00:48 -08:00
|
|
|
|
|
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-08-07 23:49:07 +02:00
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
|
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
2009-01-05 19:10:37 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-07-30 00:29:05 -04:00
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
2013-02-11 23:20:37 -05:00
|
|
|
|
2008-05-21 16:50:46 +10:00
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-05-16 11:23:19 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-30 13:20:54 +02:00
|
|
|
|
2022-05-16 11:23:19 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
2016-06-25 23:33:49 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
2016-07-31 16:37:25 -04:00
|
|
|
|
2016-06-25 23:33:49 -04:00
|
|
|
|
|
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
2022-05-16 11:23:19 +08:00
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
2022-08-11 15:29:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vfs: dcache: move hashlen_hash() from callers into d_hash()
Both __d_lookup_rcu() and __d_lookup_rcu_op_compare() have the full
'name_hash' value of the qstr that they want to look up, and mask it off
to just the low 32-bit hash before calling down to d_hash().
Other callers just load the 32-bit hash and pass it as the argument.
If we move the masking into d_hash() itself, it simplifies the two
callers that currently do the masking, and is a no-op for the other
cases. It doesn't actually change the generated code since the compiler
will inline d_hash() and see that the end result is the same.
[ Technically, since the parse tree changes, the code generation may not
be 100% the same, and for me on x86-64, this does result in gcc
switching the operands around for one 'cmpl' instruction. So not
necessarily the exact same code generation, but equivalent ]
However, this does encapsulate the 'd_hash()' operation more, and makes
the shift operation in particular be a "shift 32 bits right, return full
word". Which matches the instruction semantics on both x86-64 and arm64
better, since a 32-bit shift will clear the upper bits.
That makes the next step of introducing a "shift by runtime constant"
more obvious and generates the shift with no extraneous type masking.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2024-06-10 11:47:06 -07:00
|
|
|
|
2022-08-11 15:29:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-03-18 21:23:05 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2012-03-02 14:23:30 -08:00
|
|
|
|
|
|
|
|
|
2013-05-21 15:22:44 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2012-05-10 13:14:12 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
vfs: dcache: move hashlen_hash() from callers into d_hash()
Both __d_lookup_rcu() and __d_lookup_rcu_op_compare() have the full
'name_hash' value of the qstr that they want to look up, and mask it off
to just the low 32-bit hash before calling down to d_hash().
Other callers just load the 32-bit hash and pass it as the argument.
If we move the masking into d_hash() itself, it simplifies the two
callers that currently do the masking, and is a no-op for the other
cases. It doesn't actually change the generated code since the compiler
will inline d_hash() and see that the end result is the same.
[ Technically, since the parse tree changes, the code generation may not
be 100% the same, and for me on x86-64, this does result in gcc
switching the operands around for one 'cmpl' instruction. So not
necessarily the exact same code generation, but equivalent ]
However, this does encapsulate the 'd_hash()' operation more, and makes
the shift operation in particular be a "shift 32 bits right, return full
word". Which matches the instruction semantics on both x86-64 and arm64
better, since a 32-bit shift will clear the upper bits.
That makes the next step of introducing a "shift by runtime constant"
more obvious and generates the shift with no extraneous type masking.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2024-06-10 11:47:06 -07:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-08-11 15:29:46 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-22 15:31:32 +09:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2011-04-23 22:32:03 -07:00
|
|
|
|
2012-03-02 14:23:30 -08:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
|
|
|
|
|
2013-05-21 15:22:44 -07:00
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
|
|
|
|
|
2013-05-21 15:22:44 -07:00
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-25 23:33:49 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces
The calling conventions for __d_lookup_rcu() and dentry_cmp() are
annoying in different ways, and there is actually one single underlying
reason for both of the annoyances.
The fundamental reason is that we do the returned dentry sequence number
check inside __d_lookup_rcu() instead of doing it in the caller. This
results in two annoyances:
- __d_lookup_rcu() now not only needs to return the dentry and the
sequence number that goes along with the lookup, it also needs to
return the inode pointer that was validated by that sequence number
check.
- and because we did the sequence number check early (to validate the
name pointer and length) we also couldn't just pass the dentry itself
to dentry_cmp(), we had to pass the counted string that contained the
name.
So that sequence number decision caused two separate ugly calling
conventions.
Both of these problems would be solved if we just did the sequence
number check in the caller instead. There's only one caller, and that
caller already has to do the sequence number check for the parent
anyway, so just do that.
That allows us to stop returning the dentry->d_inode in that in-out
argument (pointer-to-pointer-to-inode), so we can make the inode
argument just a regular input inode pointer. The caller can just load
the inode from dentry->d_inode, and then do the sequence number check
after that to make sure that it's synchronized with the name we looked
up.
And it allows us to just pass in the dentry to dentry_cmp(), which is
what all the callers really wanted. Sure, dentry_cmp() has to be a bit
careful about the dentry (which is not stable during RCU lookup), but
that's actually very simple.
And now that dentry_cmp() can clearly see that the first string argument
is a dentry, we can use the direct word access for that, instead of the
careful unaligned zero-padding. The dentry name is always properly
aligned, since it is a single path component that is either embedded
into the dentry itself, or was allocated with kmalloc() (see __d_alloc).
Finally, this also uninlines the nasty slow-case for dentry comparisons:
that one *does* need to do a sequence number check, since it will call
in to the low-level filesystems, and we want to give those a stable
inode pointer and path component length/start arguments. Doing an extra
sequence check for that slow case is not a problem, though.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-04 14:59:14 -07:00
|
|
|
|
|
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
2022-08-11 15:29:46 -07:00
|
|
|
|
|
|
|
|
|
2025-11-27 14:15:26 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-08-11 15:29:46 -07:00
|
|
|
|
2013-05-21 15:22:44 -07:00
|
|
|
|
2016-06-25 23:33:49 -04:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: remove extra lookup in __lookup_hash
fs: remove extra lookup in __lookup_hash
Optimize lookup for create operations, where no dentry should often be
common-case. In cases where it is not, such as unlink, the added overhead
is much smaller than the removed.
Also, move comments about __d_lookup racyness to the __d_lookup call site.
d_lookup is intuitive; __d_lookup is what needs commenting. So in that same
vein, add kerneldoc comments to __d_lookup and clean up some of the comments:
- We are interested in how the RCU lookup works here, particularly with
renames. Make that explicit, and point to the document where it is explained
in more detail.
- RCU is pretty standard now, and macros make implementations pretty mindless.
If we want to know about RCU barrier details, we look in RCU code.
- Delete some boring legacy comments because we don't care much about how the
code used to work, more about the interesting parts of how it works now. So
comments about lazy LRU may be interesting, but would better be done in the
LRU or refcount management code.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2010-08-18 04:37:34 +10:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
fs: remove extra lookup in __lookup_hash
fs: remove extra lookup in __lookup_hash
Optimize lookup for create operations, where no dentry should often be
common-case. In cases where it is not, such as unlink, the added overhead
is much smaller than the removed.
Also, move comments about __d_lookup racyness to the __d_lookup call site.
d_lookup is intuitive; __d_lookup is what needs commenting. So in that same
vein, add kerneldoc comments to __d_lookup and clean up some of the comments:
- We are interested in how the RCU lookup works here, particularly with
renames. Make that explicit, and point to the document where it is explained
in more detail.
- RCU is pretty standard now, and macros make implementations pretty mindless.
If we want to know about RCU barrier details, we look in RCU code.
- Delete some boring legacy comments because we don't care much about how the
code used to work, more about the interesting parts of how it works now. So
comments about lazy LRU may be interesting, but would better be done in the
LRU or refcount management code.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2010-08-18 04:37:34 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-01-24 18:29:34 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2011-01-07 17:49:37 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-08-11 11:46:53 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
fs: remove extra lookup in __lookup_hash
fs: remove extra lookup in __lookup_hash
Optimize lookup for create operations, where no dentry should often be
common-case. In cases where it is not, such as unlink, the added overhead
is much smaller than the removed.
Also, move comments about __d_lookup racyness to the __d_lookup call site.
d_lookup is intuitive; __d_lookup is what needs commenting. So in that same
vein, add kerneldoc comments to __d_lookup and clean up some of the comments:
- We are interested in how the RCU lookup works here, particularly with
renames. Make that explicit, and point to the document where it is explained
in more detail.
- RCU is pretty standard now, and macros make implementations pretty mindless.
If we want to know about RCU barrier details, we look in RCU code.
- Delete some boring legacy comments because we don't care much about how the
code used to work, more about the interesting parts of how it works now. So
comments about lazy LRU may be interesting, but would better be done in the
LRU or refcount management code.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2010-08-18 04:37:34 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-01-24 18:27:00 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2016-06-10 07:51:30 -07:00
|
|
|
|
2011-01-07 17:50:05 +11:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2005-11-07 00:59:17 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs: remove extra lookup in __lookup_hash
fs: remove extra lookup in __lookup_hash
Optimize lookup for create operations, where no dentry should often be
common-case. In cases where it is not, such as unlink, the added overhead
is much smaller than the removed.
Also, move comments about __d_lookup racyness to the __d_lookup call site.
d_lookup is intuitive; __d_lookup is what needs commenting. So in that same
vein, add kerneldoc comments to __d_lookup and clean up some of the comments:
- We are interested in how the RCU lookup works here, particularly with
renames. Make that explicit, and point to the document where it is explained
in more detail.
- RCU is pretty standard now, and macros make implementations pretty mindless.
If we want to know about RCU barrier details, we look in RCU code.
- Delete some boring legacy comments because we don't care much about how the
code used to work, more about the interesting parts of how it works now. So
comments about lazy LRU may be interesting, but would better be done in the
LRU or refcount management code.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2010-08-18 04:37:34 +10:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-22 15:31:32 +09:00
|
|
|
|
fs: remove extra lookup in __lookup_hash
fs: remove extra lookup in __lookup_hash
Optimize lookup for create operations, where no dentry should often be
common-case. In cases where it is not, such as unlink, the added overhead
is much smaller than the removed.
Also, move comments about __d_lookup racyness to the __d_lookup call site.
d_lookup is intuitive; __d_lookup is what needs commenting. So in that same
vein, add kerneldoc comments to __d_lookup and clean up some of the comments:
- We are interested in how the RCU lookup works here, particularly with
renames. Make that explicit, and point to the document where it is explained
in more detail.
- RCU is pretty standard now, and macros make implementations pretty mindless.
If we want to know about RCU barrier details, we look in RCU code.
- Delete some boring legacy comments because we don't care much about how the
code used to work, more about the interesting parts of how it works now. So
comments about lazy LRU may be interesting, but would better be done in the
LRU or refcount management code.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2010-08-18 04:37:34 +10:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2011-04-23 22:32:03 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Fix NULL pointer dereference in proc_sys_compare
The VFS interface for the 'd_compare()' is a bit special (read: 'odd'),
because it really just essentially replaces a memcmp(). The filesystem
is supposed to just compare the two names with whatever case-independent
or other function.
And when I say 'is supposed to', I obviously mean that 'procfs does odd
things, and actually looks at the dentry that we don't even pass down,
rather than just the name'. Which results in problems, because we
actually call d_compare before we have even verified that the dentry is
still hashed at all.
And that causes a problm since the inode that procfs looks at may have
been free'd and the d_inode pointer is NULL. procfs just assumes that
all dentries are positive, since procfs itself never generates a
negative one. But memory pressure will still result in the dentry
getting torn down, and as it is removed by RCU, it still remains visible
on some lists - and to d_compare.
If the filesystem just did a name comparison, we wouldn't care. And we
could just fix procfs to know about negative dentries too. But rather
than have the low-level filesystems know about internal VFS details,
just move the check for a unhashed dentry up a bit, so that we will only
call d_compare on dentries that are still active.
The actual oops this caused didn't look like a NULL pointer dereference
because procfs did a 'container_of(inode, struct proc_inode, vfs_inode)'
to get at its internal proc_inode information from the inode pointer,
and accessed a field below the inode. So the oops would look something
like
BUG: unable to handle kernel paging request at fffffffffffffff0
IP: [<ffffffff802bc6c6>] proc_sys_compare+0x36/0x50
and was seen on both x86-64 (Alexey Dobriyan and Hugh Dickins) and
ppc64 (Hugh Dickins).
Reported-by: Alexey Dobriyan <adobriyan@gmail.com>
Acked-by: Hugh Dickins <hugh@veritas.com>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Reviewed-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-of-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-29 07:42:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-25 23:33:49 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-08-28 18:24:59 -07:00
|
|
|
|
Fix NULL pointer dereference in proc_sys_compare
The VFS interface for the 'd_compare()' is a bit special (read: 'odd'),
because it really just essentially replaces a memcmp(). The filesystem
is supposed to just compare the two names with whatever case-independent
or other function.
And when I say 'is supposed to', I obviously mean that 'procfs does odd
things, and actually looks at the dentry that we don't even pass down,
rather than just the name'. Which results in problems, because we
actually call d_compare before we have even verified that the dentry is
still hashed at all.
And that causes a problm since the inode that procfs looks at may have
been free'd and the d_inode pointer is NULL. procfs just assumes that
all dentries are positive, since procfs itself never generates a
negative one. But memory pressure will still result in the dentry
getting torn down, and as it is removed by RCU, it still remains visible
on some lists - and to d_compare.
If the filesystem just did a name comparison, we wouldn't care. And we
could just fix procfs to know about negative dentries too. But rather
than have the low-level filesystems know about internal VFS details,
just move the check for a unhashed dentry up a bit, so that we will only
call d_compare on dentries that are still active.
The actual oops this caused didn't look like a NULL pointer dereference
because procfs did a 'container_of(inode, struct proc_inode, vfs_inode)'
to get at its internal proc_inode information from the inode pointer,
and accessed a field below the inode. So the oops would look something
like
BUG: unable to handle kernel paging request at fffffffffffffff0
IP: [<ffffffff802bc6c6>] proc_sys_compare+0x36/0x50
and was seen on both x86-64 (Alexey Dobriyan and Hugh Dickins) and
ppc64 (Hugh Dickins).
Reported-by: Alexey Dobriyan <adobriyan@gmail.com>
Acked-by: Hugh Dickins <hugh@veritas.com>
Cc: Al Viro <viro@ZenIV.linux.org.uk>
Reviewed-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-of-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-29 07:42:57 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-03-31 02:31:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-02-11 23:20:37 -05:00
|
|
|
|
2006-03-31 02:31:43 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-10 07:51:30 -07:00
|
|
|
|
2011-01-07 17:49:55 +11:00
|
|
|
|
2013-05-21 15:22:44 -07:00
|
|
|
|
2013-02-11 23:20:37 -05:00
|
|
|
|
|
|
|
|
|
2006-03-31 02:31:43 -08:00
|
|
|
|
2013-02-11 23:20:37 -05:00
|
|
|
|
2006-03-31 02:31:43 -08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-29 09:39:34 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-29 09:39:34 -07:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-02-23 21:02:31 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-08-28 18:24:59 -07:00
|
|
|
|
2024-09-29 20:28:31 +08:00
|
|
|
|
|
|
|
|
|
2010-05-21 16:11:04 -04:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2018-02-23 21:02:31 -05:00
|
|
|
|
2024-05-29 09:39:34 -07:00
|
|
|
|
2018-02-23 21:02:31 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2016-07-29 17:45:21 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2016-07-29 17:45:21 -04:00
|
|
|
|
2017-11-10 15:45:41 +11:00
|
|
|
|
2011-04-25 14:01:36 -04:00
|
|
|
|
2011-04-23 22:32:03 -07:00
|
|
|
|
2011-04-25 14:01:36 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-28 06:31:50 -08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-07-29 17:45:21 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
2022-08-25 18:41:25 +02:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
2025-08-11 14:52:38 +02:00
|
|
|
|
|
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
2022-08-25 18:41:25 +02:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-24 10:36:46 -05:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-24 10:36:46 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-10-02 17:55:07 +02:00
|
|
|
|
|
|
|
|
|
2025-02-24 10:36:46 -05:00
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
2018-02-19 14:55:54 +00:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
2018-02-19 14:55:54 +00:00
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
2018-02-19 14:55:54 +00:00
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
2018-02-19 14:55:55 +00:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-25 23:33:49 -04:00
|
|
|
|
|
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
|
|
|
|
|
2016-06-20 01:35:59 -04:00
|
|
|
|
2026-07-24 19:14:22 +02:00
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
2016-06-20 01:35:59 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
2016-06-20 01:35:59 -04:00
|
|
|
|
2026-04-23 19:29:18 +01:00
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-06-25 23:33:49 -04:00
|
|
|
|
|
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
2016-04-15 03:33:13 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs/dcache: Move the wakeup from __d_lookup_done() to the caller.
__d_lookup_done() wakes waiters on dentry->d_wait. On PREEMPT_RT we are
not allowed to do that with preemption disabled, since the wakeup
acquired wait_queue_head::lock, which is a "sleeping" spinlock on RT.
Calling it under dentry->d_lock is not a problem, since that is also a
"sleeping" spinlock on the same configs. Unfortunately, two of its
callers (__d_add() and __d_move()) are holding more than just ->d_lock
and that needs to be dealt with.
The key observation is that wakeup can be moved to any point before
dropping ->d_lock.
As a first step to solve this, move the wake up outside of the
hlist_bl_lock() held section.
This is safe because:
Waiters get inserted into ->d_wait only after they'd taken ->d_lock
and observed DCACHE_PAR_LOOKUP in flags. As long as they are
woken up (and evicted from the queue) between the moment __d_lookup_done()
has removed DCACHE_PAR_LOOKUP and dropping ->d_lock, we are safe,
since the waitqueue ->d_wait points to won't get destroyed without
having __d_lookup_done(dentry) called (under ->d_lock).
->d_wait is set only by d_alloc_parallel() and only in case when
it returns a freshly allocated in-lookup dentry. Whenever that happens,
we are guaranteed that __d_lookup_done() will be called for resulting
dentry (under ->d_lock) before the wq in question gets destroyed.
With two exceptions wq lives in call frame of the caller of
d_alloc_parallel() and we have an explicit d_lookup_done() on the
resulting in-lookup dentry before we leave that frame.
One of those exceptions is nfs_call_unlink(), where wq is embedded into
(dynamically allocated) struct nfs_unlinkdata. It is destroyed in
nfs_async_unlink_release() after an explicit d_lookup_done() on the
dentry wq went into.
Remaining exception is d_add_ci(). There wq is what we'd found in
->d_wait of d_add_ci() argument. Callers of d_add_ci() are two
instances of ->d_lookup() and they must have been given an in-lookup
dentry. Which means that they'd been called by __lookup_slow() or
lookup_open(), with wq in the call frame of one of those.
Result of d_alloc_parallel() in d_add_ci() is fed to
d_splice_alias(), which either returns non-NULL (and d_add_ci() does
d_lookup_done()) or feeds dentry to __d_add() that will do
__d_lookup_done() under ->d_lock. That concludes the analysis.
Let __d_lookup_unhash():
1) Lock the lookup hash and clear DCACHE_PAR_LOOKUP
2) Unhash the dentry
3) Retrieve and clear dentry::d_wait
4) Unlock the hash and return the retrieved waitqueue head pointer
5) Let the caller handle the wake up.
6) Rename __d_lookup_done() to __d_lookup_unhash_wake() to enforce
build failures for OOT code that used __d_lookup_done() and is not
aware of the new return value.
This does not yet solve the PREEMPT_RT problem completely because
preemption is still disabled due to i_dir_seq being held for write. This
will be addressed in subsequent steps.
An alternative solution would be to switch the waitqueue to a simple
waitqueue, but aside of Linus not being a fan of them, moving the wake up
closer to the place where dentry::lock is unlocked reduces lock contention
time for the woken up waiter.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20220613140712.77932-3-bigeasy@linutronix.de
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-27 13:49:03 +02:00
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
fs/dcache: Move the wakeup from __d_lookup_done() to the caller.
__d_lookup_done() wakes waiters on dentry->d_wait. On PREEMPT_RT we are
not allowed to do that with preemption disabled, since the wakeup
acquired wait_queue_head::lock, which is a "sleeping" spinlock on RT.
Calling it under dentry->d_lock is not a problem, since that is also a
"sleeping" spinlock on the same configs. Unfortunately, two of its
callers (__d_add() and __d_move()) are holding more than just ->d_lock
and that needs to be dealt with.
The key observation is that wakeup can be moved to any point before
dropping ->d_lock.
As a first step to solve this, move the wake up outside of the
hlist_bl_lock() held section.
This is safe because:
Waiters get inserted into ->d_wait only after they'd taken ->d_lock
and observed DCACHE_PAR_LOOKUP in flags. As long as they are
woken up (and evicted from the queue) between the moment __d_lookup_done()
has removed DCACHE_PAR_LOOKUP and dropping ->d_lock, we are safe,
since the waitqueue ->d_wait points to won't get destroyed without
having __d_lookup_done(dentry) called (under ->d_lock).
->d_wait is set only by d_alloc_parallel() and only in case when
it returns a freshly allocated in-lookup dentry. Whenever that happens,
we are guaranteed that __d_lookup_done() will be called for resulting
dentry (under ->d_lock) before the wq in question gets destroyed.
With two exceptions wq lives in call frame of the caller of
d_alloc_parallel() and we have an explicit d_lookup_done() on the
resulting in-lookup dentry before we leave that frame.
One of those exceptions is nfs_call_unlink(), where wq is embedded into
(dynamically allocated) struct nfs_unlinkdata. It is destroyed in
nfs_async_unlink_release() after an explicit d_lookup_done() on the
dentry wq went into.
Remaining exception is d_add_ci(). There wq is what we'd found in
->d_wait of d_add_ci() argument. Callers of d_add_ci() are two
instances of ->d_lookup() and they must have been given an in-lookup
dentry. Which means that they'd been called by __lookup_slow() or
lookup_open(), with wq in the call frame of one of those.
Result of d_alloc_parallel() in d_add_ci() is fed to
d_splice_alias(), which either returns non-NULL (and d_add_ci() does
d_lookup_done()) or feeds dentry to __d_add() that will do
__d_lookup_done() under ->d_lock. That concludes the analysis.
Let __d_lookup_unhash():
1) Lock the lookup hash and clear DCACHE_PAR_LOOKUP
2) Unhash the dentry
3) Retrieve and clear dentry::d_wait
4) Unlock the hash and return the retrieved waitqueue head pointer
5) Let the caller handle the wake up.
6) Rename __d_lookup_done() to __d_lookup_unhash_wake() to enforce
build failures for OOT code that used __d_lookup_done() and is not
aware of the new return value.
This does not yet solve the PREEMPT_RT problem completely because
preemption is still disabled due to i_dir_seq being held for write. This
will be addressed in subsequent steps.
An alternative solution would be to switch the waitqueue to a simple
waitqueue, but aside of Linus not being a fan of them, moving the wake up
closer to the place where dentry::lock is unlocked reduces lock contention
time for the woken up waiter.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20220613140712.77932-3-bigeasy@linutronix.de
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-27 13:49:03 +02:00
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
2016-04-14 19:52:13 -04:00
|
|
|
|
fs/dcache: Move the wakeup from __d_lookup_done() to the caller.
__d_lookup_done() wakes waiters on dentry->d_wait. On PREEMPT_RT we are
not allowed to do that with preemption disabled, since the wakeup
acquired wait_queue_head::lock, which is a "sleeping" spinlock on RT.
Calling it under dentry->d_lock is not a problem, since that is also a
"sleeping" spinlock on the same configs. Unfortunately, two of its
callers (__d_add() and __d_move()) are holding more than just ->d_lock
and that needs to be dealt with.
The key observation is that wakeup can be moved to any point before
dropping ->d_lock.
As a first step to solve this, move the wake up outside of the
hlist_bl_lock() held section.
This is safe because:
Waiters get inserted into ->d_wait only after they'd taken ->d_lock
and observed DCACHE_PAR_LOOKUP in flags. As long as they are
woken up (and evicted from the queue) between the moment __d_lookup_done()
has removed DCACHE_PAR_LOOKUP and dropping ->d_lock, we are safe,
since the waitqueue ->d_wait points to won't get destroyed without
having __d_lookup_done(dentry) called (under ->d_lock).
->d_wait is set only by d_alloc_parallel() and only in case when
it returns a freshly allocated in-lookup dentry. Whenever that happens,
we are guaranteed that __d_lookup_done() will be called for resulting
dentry (under ->d_lock) before the wq in question gets destroyed.
With two exceptions wq lives in call frame of the caller of
d_alloc_parallel() and we have an explicit d_lookup_done() on the
resulting in-lookup dentry before we leave that frame.
One of those exceptions is nfs_call_unlink(), where wq is embedded into
(dynamically allocated) struct nfs_unlinkdata. It is destroyed in
nfs_async_unlink_release() after an explicit d_lookup_done() on the
dentry wq went into.
Remaining exception is d_add_ci(). There wq is what we'd found in
->d_wait of d_add_ci() argument. Callers of d_add_ci() are two
instances of ->d_lookup() and they must have been given an in-lookup
dentry. Which means that they'd been called by __lookup_slow() or
lookup_open(), with wq in the call frame of one of those.
Result of d_alloc_parallel() in d_add_ci() is fed to
d_splice_alias(), which either returns non-NULL (and d_add_ci() does
d_lookup_done()) or feeds dentry to __d_add() that will do
__d_lookup_done() under ->d_lock. That concludes the analysis.
Let __d_lookup_unhash():
1) Lock the lookup hash and clear DCACHE_PAR_LOOKUP
2) Unhash the dentry
3) Retrieve and clear dentry::d_wait
4) Unlock the hash and return the retrieved waitqueue head pointer
5) Let the caller handle the wake up.
6) Rename __d_lookup_done() to __d_lookup_unhash_wake() to enforce
build failures for OOT code that used __d_lookup_done() and is not
aware of the new return value.
This does not yet solve the PREEMPT_RT problem completely because
preemption is still disabled due to i_dir_seq being held for write. This
will be addressed in subsequent steps.
An alternative solution would be to switch the waitqueue to a simple
waitqueue, but aside of Linus not being a fan of them, moving the wake up
closer to the place where dentry::lock is unlocked reduces lock contention
time for the woken up waiter.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20220613140712.77932-3-bigeasy@linutronix.de
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-27 13:49:03 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
2016-04-14 19:52:13 -04:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2016-04-15 02:42:04 -04:00
|
|
|
|
get rid of busy-waiting in shrink_dcache_tree()
If shrink_dcache_tree() runs into a potential victim that is already
dying, it must wait for that dentry to go away. To avoid busy-waiting
we need some object to wait on and a way for dentry_unlist() to see that
we need to be notified.
The obvious place for the object to wait on would be on our stack frame.
We will store a pointer to that object (struct completion_list) in victim
dentry; if there's more than one thread wanting to wait for the same
dentry to finish dying, we'll have their instances linked into a list,
with reference in dentry pointing to the head of that list.
* new object - struct completion_list. A pair of struct completion and
pointer to the next instance. That's what shrink_dcache_tree() will wait
on if needed.
* add a new member (->waiters, opaque pointer to struct completion_list)
to struct dentry. It is defined for negative live dentries that are
not in-lookup ones and it will remain NULL for almost all of them.
It does not conflict with ->d_rcu (defined for killed dentries), ->d_alias
(defined for positive dentries, all live) or ->d_in_lookup_hash (defined
for in-lookup dentries, all live negative). That allows to colocate
all four members.
* make sure that all places where dentry enters the state where ->waiters
is defined (live, negative, not-in-lookup) initialize ->waiters to NULL.
* if select_collect2() runs into a dentry that is already dying, have
its caller insert a local instance of struct completion_list into the
head of the list hanging off dentry->waiters and wait for completion.
* if dentry_unlist() sees non-NULL ->waiters, have it carefully walk
through the completion_list instances in that list, calling complete()
for each.
For now struct completion_list is local to fs/dcache.c; it's obviously
dentry-agnostic, and it can be trivially lifted into linux/completion.h
if somebody finds a reason to do so...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-01-21 18:17:12 -05:00
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-14 19:52:13 -04:00
|
|
|
|
fs/dcache: Move the wakeup from __d_lookup_done() to the caller.
__d_lookup_done() wakes waiters on dentry->d_wait. On PREEMPT_RT we are
not allowed to do that with preemption disabled, since the wakeup
acquired wait_queue_head::lock, which is a "sleeping" spinlock on RT.
Calling it under dentry->d_lock is not a problem, since that is also a
"sleeping" spinlock on the same configs. Unfortunately, two of its
callers (__d_add() and __d_move()) are holding more than just ->d_lock
and that needs to be dealt with.
The key observation is that wakeup can be moved to any point before
dropping ->d_lock.
As a first step to solve this, move the wake up outside of the
hlist_bl_lock() held section.
This is safe because:
Waiters get inserted into ->d_wait only after they'd taken ->d_lock
and observed DCACHE_PAR_LOOKUP in flags. As long as they are
woken up (and evicted from the queue) between the moment __d_lookup_done()
has removed DCACHE_PAR_LOOKUP and dropping ->d_lock, we are safe,
since the waitqueue ->d_wait points to won't get destroyed without
having __d_lookup_done(dentry) called (under ->d_lock).
->d_wait is set only by d_alloc_parallel() and only in case when
it returns a freshly allocated in-lookup dentry. Whenever that happens,
we are guaranteed that __d_lookup_done() will be called for resulting
dentry (under ->d_lock) before the wq in question gets destroyed.
With two exceptions wq lives in call frame of the caller of
d_alloc_parallel() and we have an explicit d_lookup_done() on the
resulting in-lookup dentry before we leave that frame.
One of those exceptions is nfs_call_unlink(), where wq is embedded into
(dynamically allocated) struct nfs_unlinkdata. It is destroyed in
nfs_async_unlink_release() after an explicit d_lookup_done() on the
dentry wq went into.
Remaining exception is d_add_ci(). There wq is what we'd found in
->d_wait of d_add_ci() argument. Callers of d_add_ci() are two
instances of ->d_lookup() and they must have been given an in-lookup
dentry. Which means that they'd been called by __lookup_slow() or
lookup_open(), with wq in the call frame of one of those.
Result of d_alloc_parallel() in d_add_ci() is fed to
d_splice_alias(), which either returns non-NULL (and d_add_ci() does
d_lookup_done()) or feeds dentry to __d_add() that will do
__d_lookup_done() under ->d_lock. That concludes the analysis.
Let __d_lookup_unhash():
1) Lock the lookup hash and clear DCACHE_PAR_LOOKUP
2) Unhash the dentry
3) Retrieve and clear dentry::d_wait
4) Unlock the hash and return the retrieved waitqueue head pointer
5) Let the caller handle the wake up.
6) Rename __d_lookup_done() to __d_lookup_unhash_wake() to enforce
build failures for OOT code that used __d_lookup_done() and is not
aware of the new return value.
This does not yet solve the PREEMPT_RT problem completely because
preemption is still disabled due to i_dir_seq being held for write. This
will be addressed in subsequent steps.
An alternative solution would be to switch the waitqueue to a simple
waitqueue, but aside of Linus not being a fan of them, moving the wake up
closer to the place where dentry::lock is unlocked reduces lock contention
time for the woken up waiter.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20220613140712.77932-3-bigeasy@linutronix.de
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-27 13:49:03 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
|
|
|
|
|
fs/dcache: Move the wakeup from __d_lookup_done() to the caller.
__d_lookup_done() wakes waiters on dentry->d_wait. On PREEMPT_RT we are
not allowed to do that with preemption disabled, since the wakeup
acquired wait_queue_head::lock, which is a "sleeping" spinlock on RT.
Calling it under dentry->d_lock is not a problem, since that is also a
"sleeping" spinlock on the same configs. Unfortunately, two of its
callers (__d_add() and __d_move()) are holding more than just ->d_lock
and that needs to be dealt with.
The key observation is that wakeup can be moved to any point before
dropping ->d_lock.
As a first step to solve this, move the wake up outside of the
hlist_bl_lock() held section.
This is safe because:
Waiters get inserted into ->d_wait only after they'd taken ->d_lock
and observed DCACHE_PAR_LOOKUP in flags. As long as they are
woken up (and evicted from the queue) between the moment __d_lookup_done()
has removed DCACHE_PAR_LOOKUP and dropping ->d_lock, we are safe,
since the waitqueue ->d_wait points to won't get destroyed without
having __d_lookup_done(dentry) called (under ->d_lock).
->d_wait is set only by d_alloc_parallel() and only in case when
it returns a freshly allocated in-lookup dentry. Whenever that happens,
we are guaranteed that __d_lookup_done() will be called for resulting
dentry (under ->d_lock) before the wq in question gets destroyed.
With two exceptions wq lives in call frame of the caller of
d_alloc_parallel() and we have an explicit d_lookup_done() on the
resulting in-lookup dentry before we leave that frame.
One of those exceptions is nfs_call_unlink(), where wq is embedded into
(dynamically allocated) struct nfs_unlinkdata. It is destroyed in
nfs_async_unlink_release() after an explicit d_lookup_done() on the
dentry wq went into.
Remaining exception is d_add_ci(). There wq is what we'd found in
->d_wait of d_add_ci() argument. Callers of d_add_ci() are two
instances of ->d_lookup() and they must have been given an in-lookup
dentry. Which means that they'd been called by __lookup_slow() or
lookup_open(), with wq in the call frame of one of those.
Result of d_alloc_parallel() in d_add_ci() is fed to
d_splice_alias(), which either returns non-NULL (and d_add_ci() does
d_lookup_done()) or feeds dentry to __d_add() that will do
__d_lookup_done() under ->d_lock. That concludes the analysis.
Let __d_lookup_unhash():
1) Lock the lookup hash and clear DCACHE_PAR_LOOKUP
2) Unhash the dentry
3) Retrieve and clear dentry::d_wait
4) Unlock the hash and return the retrieved waitqueue head pointer
5) Let the caller handle the wake up.
6) Rename __d_lookup_done() to __d_lookup_unhash_wake() to enforce
build failures for OOT code that used __d_lookup_done() and is not
aware of the new return value.
This does not yet solve the PREEMPT_RT problem completely because
preemption is still disabled due to i_dir_seq being held for write. This
will be addressed in subsequent steps.
An alternative solution would be to switch the waitqueue to a simple
waitqueue, but aside of Linus not being a fan of them, moving the wake up
closer to the place where dentry::lock is unlocked reduces lock contention
time for the woken up waiter.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/20220613140712.77932-3-bigeasy@linutronix.de
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-27 13:49:03 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-09 19:52:39 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-24 12:46:49 -05:00
|
|
|
|
|
|
|
|
|
2016-03-09 19:52:39 -05:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
2016-04-14 19:40:56 -04:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
2025-02-24 12:46:49 -05:00
|
|
|
|
|
|
|
|
|
2016-03-09 19:52:39 -05:00
|
|
|
|
2016-04-14 19:40:56 -04:00
|
|
|
|
2026-01-27 22:51:37 -05:00
|
|
|
|
2016-04-14 19:40:56 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-29 18:35:12 -04:00
|
|
|
|
2016-03-09 19:52:39 -05:00
|
|
|
|
2016-07-29 17:45:21 -04:00
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-14 19:40:56 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-03-09 19:52:39 -05:00
|
|
|
|
|
|
|
|
|
2016-03-08 21:01:03 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-11 00:53:26 -04:00
|
|
|
|
|
|
|
|
|
2016-03-09 19:52:39 -05:00
|
|
|
|
2016-04-11 00:53:26 -04:00
|
|
|
|
2025-02-24 12:46:49 -05:00
|
|
|
|
2016-03-08 21:01:03 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
|
|
|
|
|
2026-01-24 01:32:07 -05:00
|
|
|
|
primitives for maintaining persisitency
* d_make_persistent(dentry, inode) - bump refcount, mark persistent and
make hashed positive. Return value is a borrowed reference to dentry;
it can be used until something removes persistency (at the very least,
until the parent gets unlocked, but some filesystems may have stronger
exclusion).
* d_make_discardable() - remove persistency mark and drop reference.
d_make_persistent() is similar to combination of d_instantiate(), dget()
and setting flag. The only difference is that unlike d_instantiate()
it accepts hashed and unhashed negatives alike. It is always called in
strong locking environment (parent held exclusive, or, in some cases,
dentry coming from d_alloc_name()); if we ever start using it with parent
held only shared and dentry coming from d_alloc_parallel(), we'll need
to copy the in-lookup logics from __d_add().
d_make_discardable() is eqiuvalent to combination of removing flag and
dput(); since flag removal requires ->d_lock, there's no point trying
to avoid taking that for refcount decrement as fast_dput() does.
The slow path of dput() has been taken into a helper and reused in
d_make_discardable() instead.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2025-03-03 19:38:04 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2014-04-01 17:08:43 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
|
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
2024-12-09 20:03:33 -05:00
|
|
|
|
|
|
|
|
|
Allow sharing external names after __d_move()
* external dentry names get a small structure prepended to them
(struct external_name).
* it contains an atomic refcount, matching the number of struct dentry
instances that have ->d_name.name pointing to that external name. The
first thing free_dentry() does is decrementing refcount of external name,
so the instances that are between the call of free_dentry() and
RCU-delayed actual freeing do not contribute.
* __d_move(x, y, false) makes the name of x equal to the name of y,
external or not. If y has an external name, extra reference is grabbed
and put into x->d_name.name. If x used to have an external name, the
reference to the old name is dropped and, should it reach zero, freeing
is scheduled via kfree_rcu().
* free_dentry() in dentry with external name decrements the refcount of
that name and, should it reach zero, does RCU-delayed call that will
free both the dentry and external name. Otherwise it does what it
used to do, except that __d_free() doesn't even look at ->d_name.name;
it simply frees the dentry.
All non-RCU accesses to dentry external name are safe wrt freeing since they
all should happen before free_dentry() is called. RCU accesses might run
into a dentry seen by free_dentry() or into an old name that got already
dropped by __d_move(); however, in both cases dentry must have been
alive and refer to that name at some point after we'd done rcu_read_lock(),
which means that any freeing must be still pending.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-09-29 14:54:27 -04:00
|
|
|
|
|
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2011-07-12 21:42:24 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2014-04-01 17:08:43 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-06-22 23:01:32 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-04-01 17:08:43 +02:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-03-11 15:15:46 -04:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-03-11 15:15:46 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
2018-03-11 15:15:46 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-07 17:49:34 +11:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
parallel lookups machinery, part 2
We'll need to verify that there's neither a hashed nor in-lookup
dentry with desired parent/name before adding to in-lookup set.
One possible solution would be to hold the parent's ->d_lock through
both checks, but while the in-lookup set is relatively small at any
time, dcache is not. And holding the parent's ->d_lock through
something like __d_lookup_rcu() would suck too badly.
So we leave the parent's ->d_lock alone, which means that we watch
out for the following scenario:
* we verify that there's no hashed match
* existing in-lookup match gets hashed by another process
* we verify that there's no in-lookup matches and decide
that everything's fine.
Solution: per-directory kinda-sorta seqlock, bumped around the times
we hash something that used to be in-lookup or move (and hash)
something in place of in-lookup. Then the above would turn into
* read the counter
* do dcache lookup
* if no matches found, check for in-lookup matches
* if there had been none of those either, check if the
counter has changed; repeat if it has.
The "kinda-sorta" part is due to the fact that we don't have much spare
space in inode. There is a spare word (shared with i_bdev/i_cdev/i_pipe),
so the counter part is not a problem, but spinlock is a different story.
We could use the parent's ->d_lock, and it would be less painful in
terms of contention, for __d_add() it would be rather inconvenient to
grab; we could do that (using lock_parent()), but...
Fortunately, we can get serialization on the counter itself, and it
might be a good idea in general; we can use cmpxchg() in a loop to
get from even to odd and smp_store_release() from odd to even.
This commit adds the counter and updating logics; the readers will be
added in the next commit.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 00:58:55 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2013-10-07 15:51:59 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
2016-07-29 17:45:21 -04:00
|
|
|
|
2018-03-07 00:49:10 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-09-26 23:06:14 -04:00
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
dcache: sort the freeing-without-RCU-delay mess for good.
For lockless accesses to dentries we don't have pinned we rely
(among other things) upon having an RCU delay between dropping
the last reference and actually freeing the memory.
On the other hand, for things like pipes and sockets we neither
do that kind of lockless access, nor want to deal with the
overhead of an RCU delay every time a socket gets closed.
So delay was made optional - setting DCACHE_RCUACCESS in ->d_flags
made sure it would happen. We tried to avoid setting it unless
we knew we need it. Unfortunately, that had led to recurring
class of bugs, in which we missed the need to set it.
We only really need it for dentries that are created by
d_alloc_pseudo(), so let's not bother with trying to be smart -
just make having an RCU delay the default. The ones that do
*not* get it set the replacement flag (DCACHE_NORCU) and we'd
better use that sparingly. d_alloc_pseudo() is the only
such user right now.
FWIW, the race that finally prompted that switch had been
between __lock_parent() of immediate subdirectory of what's
currently the root of a disconnected tree (e.g. from
open-by-handle in progress) racing with d_splice_alias()
elsewhere picking another alias for the same inode, either
on outright corrupted fs image, or (in case of open-by-handle
on NFS) that subdirectory having been just moved on server.
It's not easy to hit, so the sky is not falling, but that's
not the first race on similar missed cases and the logics
for settinf DCACHE_RCUACCESS has gotten ridiculously
convoluted.
Cc: stable@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-03-15 22:23:19 -04:00
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
|
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2023-11-07 02:00:39 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
|
|
|
|
|
2019-03-20 11:39:11 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
fs: rcu-walk for path lookup
Perform common cases of path lookups without any stores or locking in the
ancestor dentry elements. This is called rcu-walk, as opposed to the current
algorithm which is a refcount based walk, or ref-walk.
This results in far fewer atomic operations on every path element,
significantly improving path lookup performance. It also avoids cacheline
bouncing on common dentries, significantly improving scalability.
The overall design is like this:
* LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk.
* Take the RCU lock for the entire path walk, starting with the acquiring
of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are
not required for dentry persistence.
* synchronize_rcu is called when unregistering a filesystem, so we can
access d_ops and i_ops during rcu-walk.
* Similarly take the vfsmount lock for the entire path walk. So now mnt
refcounts are not required for persistence. Also we are free to perform mount
lookups, and to assume dentry mount points and mount roots are stable up and
down the path.
* Have a per-dentry seqlock to protect the dentry name, parent, and inode,
so we can load this tuple atomically, and also check whether any of its
members have changed.
* Dentry lookups (based on parent, candidate string tuple) recheck the parent
sequence after the child is found in case anything changed in the parent
during the path walk.
* inode is also RCU protected so we can load d_inode and use the inode for
limited things.
* i_mode, i_uid, i_gid can be tested for exec permissions during path walk.
* i_op can be loaded.
When we reach the destination dentry, we lock it, recheck lookup sequence,
and increment its refcount and mountpoint refcount. RCU and vfsmount locks
are dropped. This is termed "dropping rcu-walk". If the dentry refcount does
not match, we can not drop rcu-walk gracefully at the current point in the
lokup, so instead return -ECHILD (for want of a better errno). This signals the
path walking code to re-do the entire lookup with a ref-walk.
Aside from the final dentry, there are other situations that may be encounted
where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take
a reference on the last good dentry) and continue with a ref-walk. Again, if
we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup
using ref-walk. But it is very important that we can continue with ref-walk
for most cases, particularly to avoid the overhead of double lookups, and to
gain the scalability advantages on common path elements (like cwd and root).
The cases where rcu-walk cannot continue are:
* NULL dentry (ie. any uncached path element)
* parent with d_inode->i_op->permission or ACLs
* dentries with d_revalidate
* Following links
In future patches, permission checks and d_revalidate become rcu-walk aware. It
may be possible eventually to make following links rcu-walk aware.
Uncached path elements will always require dropping to ref-walk mode, at the
very least because i_mutex needs to be grabbed, and objects allocated.
Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 17:49:52 +11:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
VFS: use wait_var_event for waiting in d_alloc_parallel()
Parallel lookup starts with a call of d_alloc_parallel(). That primitive
either returns a matching hashed dentry or allocates a new one in the
in-lookup state and returns it to the caller. Once the caller is done
with lookup, it indicates so either by call of d_{splice_alias,add}()
or by call of d_done_lookup(); at that point dentry leaves the in-lookup
state.
If d_alloc_parallel() finds a matching in-lookup dentry, it must wait for
that dentry to leave the in-lookup state, one way or another. Currently
by supplying wait_queue_head to d_alloc_parallel(). If d_alloc_parallel()
creates a new in-lookup dentry, the address of that wait_queue_head is stored
in ->d_wait of new dentry and stays there while it's in the in-lookup;
subsequent d_alloc_parallel() will wait on the queue found in the matching
in-lookup dentry. Transition out of in-lookup state wakes waiters on that
queue (if any).
That works, but the calling conventions are inconvenient - the caller must
supply wait_queue_head and make sure that it survives at least until the new
in-lookup dentry leaves the in-lookup state. That amounts to boilerplate
in the d_alloc_parallel() callers that are followed by a call of d_lookup_done()
in the same function; in cases like nfs asynchronous unlink it gets worse than
that.
This patch changes d_alloc_parallel() to use wake_up_var_locked() to
wake up waiters, and wait_var_event_spinlock() to wait. dentry->d_lock
is used for synchronisation as it is already held and the relevant
times.
That eliminates the need of caller-supplied wait_queue_head, simplifying
the calling conventions. Better yet, we only need one bit of information
stored in dentry itself: whether there are any waiters to be woken up,
and that can be easily stored in ->d_flags; ->d_wait goes away.
The reason we need that bit (DCACHE_LOOKUP_WAITERS) is that with wait_var
machinery the queues are shared with all kinds of stuff and there's
no way tell if any of the waiters have anything to do with our dentry;
most of the time none of them will be relevant, so we need to avoid the
pointless wakeups.
Another benefit of the new scheme comes from the fact that wakeups
have to be done outside of write-side critical areas of ->i_dir_seq;
with the old scheme we need to carry the value picked from ->d_wait from
__d_lookup_unhash() to the place where we actually wake the waiters up.
Now we can just leave DCACHE_LOOKUP_WAITERS in ->d_flags until we get
to doing wakeups - that's done within the same ->d_lock scope, so we
are fine; new bit is accessed only under ->d_lock and it's seen only
on dentries with DCACHE_PAR_LOOKUP in ->d_flags.
__d_lookup_unhash() no longer needs to re-init ->d_lru. That was
previously shared (in a union) with ->d_wait but ->d_wait is now gone
so it no longer corrupts ->d_lru.
Co-developed-by: Al Viro <viro@zeniv.linux.org.uk> # saner handling of flags
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-30 15:42:43 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-07-12 21:42:24 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-07-26 13:33:16 -04:00
|
|
|
|
|
|
|
|
|
2011-07-12 21:42:24 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-04-01 17:08:43 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2010-01-05 13:45:18 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2014-04-01 17:08:43 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-09-24 13:49:49 +01:00
|
|
|
|
2014-04-01 17:08:43 +02:00
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-10-16 07:50:27 +09:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-22 23:01:32 -05:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-23 18:11:00 -05:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2016-04-15 15:08:36 -04:00
|
|
|
|
|
|
|
|
|
2015-02-10 10:55:53 -05:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-04-15 15:08:36 -04:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2016-04-15 15:08:36 -04:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2025-02-03 16:16:09 -05:00
|
|
|
|
2025-01-05 21:33:17 -05:00
|
|
|
|
|
|
|
|
|
vfs: Lazily remove mounts on unlinked files and directories.
With the introduction of mount namespaces and bind mounts it became
possible to access files and directories that on some paths are mount
points but are not mount points on other paths. It is very confusing
when rm -rf somedir returns -EBUSY simply because somedir is mounted
somewhere else. With the addition of user namespaces allowing
unprivileged mounts this condition has gone from annoying to allowing
a DOS attack on other users in the system.
The possibility for mischief is removed by updating the vfs to support
rename, unlink and rmdir on a dentry that is a mountpoint and by
lazily unmounting mountpoints on deleted dentries.
In particular this change allows rename, unlink and rmdir system calls
on a dentry without a mountpoint in the current mount namespace to
succeed, and it allows rename, unlink, and rmdir performed on a
distributed filesystem to update the vfs cache even if when there is a
mount in some namespace on the original dentry.
There are two common patterns of maintaining mounts: Mounts on trusted
paths with the parent directory of the mount point and all ancestory
directories up to / owned by root and modifiable only by root
(i.e. /media/xxx, /dev, /dev/pts, /proc, /sys, /sys/fs/cgroup/{cpu,
cpuacct, ...}, /usr, /usr/local). Mounts on unprivileged directories
maintained by fusermount.
In the case of mounts in trusted directories owned by root and
modifiable only by root the current parent directory permissions are
sufficient to ensure a mount point on a trusted path is not removed
or renamed by anyone other than root, even if there is a context
where the there are no mount points to prevent this.
In the case of mounts in directories owned by less privileged users
races with users modifying the path of a mount point are already a
danger. fusermount already uses a combination of chdir,
/proc/<pid>/fd/NNN, and UMOUNT_NOFOLLOW to prevent these races. The
removable of global rename, unlink, and rmdir protection really adds
nothing new to consider only a widening of the attack window, and
fusermount is already safe against unprivileged users modifying the
directory simultaneously.
In principle for perfect userspace programs returning -EBUSY for
unlink, rmdir, and rename of dentires that have mounts in the local
namespace is actually unnecessary. Unfortunately not all userspace
programs are perfect so retaining -EBUSY for unlink, rmdir and rename
of dentries that have mounts in the current mount namespace plays an
important role of maintaining consistency with historical behavior and
making imperfect userspace applications hard to exploit.
v2: Remove spurious old_dentry.
v3: Optimized shrink_submounts_and_drop
Removed unsued afs label
v4: Simplified the changes to check_submounts_and_drop
Do not rename check_submounts_and_drop shrink_submounts_and_drop
Document what why we need atomicity in check_submounts_and_drop
Rely on the parent inode mutex to make d_revalidate and d_invalidate
an atomic unit.
v5: Refcount the mountpoint to detach in case of simultaneous
renames.
Reviewed-by: Miklos Szeredi <miklos@szeredi.hu>
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-10-01 18:33:48 -07:00
|
|
|
|
2025-02-03 16:16:09 -05:00
|
|
|
|
2025-01-05 21:33:17 -05:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
|
|
|
|
|
2016-04-15 15:08:36 -04:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-02-24 12:46:49 -05:00
|
|
|
|
|
|
|
|
|
2014-02-18 14:11:26 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-08-22 20:06:07 -04:00
|
|
|
|
|
|
|
|
|
2016-03-09 18:05:42 -05:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
2016-03-09 18:05:42 -05:00
|
|
|
|
2016-04-11 00:53:26 -04:00
|
|
|
|
2011-01-07 17:50:06 +11:00
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
fix a race between d_find_any_alias() and final dput() of NORCU dentries
Refcount of a NORCU dentry must not be incremented after having dropped
to zero. Otherwise we might end up with the following race:
CPU1: in fast_dput(d), rcu_read_lock();
CPU1: decrements refcount of d to 0
CPU1: notice that it's unhashed
CPU2: grab a reference to d
CPU2: dput(d), freeing d
CPU1: ... looks like we need to evict d, let's grab ->d_lock, recheck
the refcount, etc.
and that spin_lock(&d->d_lock) ends up a UAF, despite still being in
an RCU read-side critical area started back when the refcount had been
positive. If not for DCACHE_NORCU in d->d_flags freeing would've been
RCU-delayed, so we'd have grabbed ->d_lock, noticed the negative value
stored into refcount by __dentry_kill(), dropped the locks and that would
be it. For NORCU dentries freeing is _not_ delayed, though.
Most of the non-counting references are excluded for NORCU dentries -
they are not allowed to be hashed, they never get placed on LRU, they
never get placed into anyone's list of children and while dput_to_list()
might put them into a shrink list, nobody bumps refcount of something
that had been reached that way.
However, inode's list of aliases can be a problem - it does not contribute
to dentry refcount (for obvious reasons) and we *do* have places that
grab references to something found on that list - that's precisely what
d_find_alias() is. In case of d_find_alias() we are safe - it skips
unhashed aliases, so all NORCU ones are ignored there. d_find_any_alias()
is *not* limited to hashed ones, though, and while it's usually called
for directories (which never get NORCU dentries), there are callers that
use it to get something for non-directories with no hashed aliases.
Having d_find_any_alias() hit a NORCU dentry is not impossible - it can
be easily arranged if you have CAP_DAC_READ_SEARCH (memfd_create() + mmap()
+ name_to_handle_at() for /proc/self/map_files/<...> + munmap() +
open_by_handle_at() will do that, and adding a second memfd_create() for
mount_fd makes it possible to do that without having memfd pinned).
The race window is narrow, and it's probably not feasible on bare hardware,
but...
It's not hard to fix, fortunately:
* separate __d_find_dir_alias() (== current __d_find_any_alias()) to
be used for directory inodes.
* provide dget_alias_ilocked() that would return false for NORCU
dentries with zero refcount and return true incrementing refcount otherwise
* make __d_find_any_alias() go over the list of aliases, using
dget_alias_ilocked() and returning the alias it succeeds on (normally the
first one). Any NORCU alias with zero refcount is going to be evicted by
the thread that had dropped the final reference; this makes __d_find_any_alias()
pretend it had lost the race with eviction.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-05-03 23:00:09 -04:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
2015-08-15 13:36:41 -05:00
|
|
|
|
|
|
|
|
|
2011-07-12 21:42:24 -04:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
2011-07-12 21:42:24 -04:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
2023-11-23 18:11:00 -05:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
make non-exchanging __d_move() copy ->d_parent rather than swap them
Currently d_move(from, to) does the following:
* name/parent of from <- old name/parent of to, from hashed there
* to is unhashed
* name of to is preserved
* if from used to be detached, to gets detached
* if from used to be attached, parent of to <- old parent of from.
That's both user-visibly bogus and complicates reasoning a lot.
Much saner semantics would be
* name/parent of from <- name/parent of to, from hashed there.
* to is unhashed
* name/parent of to is unchanged.
The price, of course, is that old parent of from might lose a reference.
However,
* all potentially cross-directory callers of d_move() have both
parents pinned directly; typically, dentries themselves are grabbed
only after we have grabbed and locked both parents. IOW, the decrement
of old parent's refcount in case of d_move() won't reach zero.
* __d_move() from d_splice_alias() is done to detached alias.
No refcount decrements in that case
* __d_move() from __d_unalias() *can* get the refcount to zero.
So let's grab a reference to alias' old parent before calling __d_unalias()
and dput() it after we'd dropped rename_lock.
That does make d_splice_alias() potentially blocking. However, it has
no callers in non-sleepable contexts (and the case where we'd grown
that dget/dput pair is _very_ rare, so performance is not an issue).
Another thing that needs adjustment is unlocking in the end of __d_move();
folded it in. And cleaned the remnants of bogus ordering from the
"lock them in the beginning" counterpart - it's never been right and
now (well, for 7 years now) we have that thing always serialized on
rename_lock anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-10 23:15:52 -05:00
|
|
|
|
2011-07-12 21:42:24 -04:00
|
|
|
|
wind ->s_roots via ->d_sib instead of ->d_hash
shrink_dcache_for_umount() is supposed to handle the possibility of
some of the dentries to be evicted being in other threads shrink
lists; it either kills them, leaving an empty husk to be freed by
the owner of shrink list whenever it gets around to that, or it
waits for the eviction in progress to get completed.
That relies upon dentry remaining attached to the tree until the
eviction reaches dentry_unlist() and its ->d_sib gets removed
from the list. Unfortunately, the secondary roots are linked
via ->d_hash, rather than ->d_sib and they become removed from
that list before their inode references are dropped.
If shrink_dentry_list() from another thread ends up evicting
one of the secondary roots and gets to that point in dentry_kill()
when shrink_dcache_for_umount() is looking for secondary roots,
the latter will *not* notice anything, possibly leading to
warnings about busy inodes at umount time and all kinds of breakage
after that.
Moreover, shrink_dcache_for_umount() walks the list of secondary
roots with no protection whatsoever, so it might end up calling
dget() on a dentry that already passed through
lockref_mark_dead(&dentry->d_lockref);
ending up with corrupted refcount and possible UAF.
AFAICS, the most straightforward way to deal with that would be
to have secondary roots linked via ->d_sib rather than ->d_hash;
then they would remain on the list until killed, and we could
use d_add_waiter() machinery to wait for eviction in progress.
Changes:
* secondary roots look the same as ->s_root from d_unhashed()
and d_unlinked() POV now.
* secondary roots are represented as "no parent, but on ->d_sib"
instead of "no parent, but on ->d_hash".
* since ->d_sib is a plain hlist, we protect it with per-superblock
spinlock (sb->s_roots_lock) instead of the LSB of the head pointer (for
non-root dentries it would be protected by ->d_lock of parent).
* __d_obtain_alias() uses ->d_sib for linkage when allocating
a secondary root.
* d_splice_alias_ops() detects splicing of a secondary root and
removes it from the list before calling __d_move().
* dentry_unlist() detects eviction of a secondary root and
removes it from the list; no need to play the games for d_walk() sake,
since the latter is not going to look for the next sibling of those
anyway.
* ___d_drop() doesn't care about ->s_roots anymore.
* shrink_dcache_for_umount() uses proper locking for access to
the list of secondary roots and if it runs into one that is in the middle
of eviction waits for that to finish.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2026-04-18 18:39:03 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
|
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
|
|
|
|
|
2006-10-21 10:24:20 -07:00
|
|
|
|
2006-08-22 20:06:07 -04:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
2025-02-24 12:46:49 -05:00
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
2006-08-22 20:06:07 -04:00
|
|
|
|
2025-02-24 12:46:49 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-12 22:16:02 -04:00
|
|
|
|
2006-08-22 20:06:07 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-17 14:40:11 +08:00
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-11-17 14:40:11 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-07-02 21:03:26 +02:00
|
|
|
|
2011-01-07 17:49:37 +11:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
2015-11-17 14:40:11 +08:00
|
|
|
|
2008-10-16 07:50:28 +09:00
|
|
|
|
2024-07-02 21:03:26 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-01-11 11:33:24 +02:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2018-07-09 23:27:33 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2022-09-24 07:00:00 +02:00
|
|
|
|
|
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2026-01-24 01:32:07 -05:00
|
|
|
|
2013-06-07 01:20:27 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-09 14:57:43 -05:00
|
|
|
|
2013-06-07 01:20:27 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-07-09 23:27:33 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-14 11:37:21 -03:00
|
|
|
|
2026-04-07 16:58:09 -03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-14 11:37:21 -03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-07 16:58:09 -03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-14 11:37:21 -03:00
|
|
|
|
2026-04-07 16:58:09 -03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-07-09 23:27:33 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-06-07 01:20:27 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2013-06-07 01:20:27 -04:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2024-06-27 18:11:52 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-12-16 15:52:37 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-04-23 22:32:03 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:39:11 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-11-20 18:05:52 +03:00
|
|
|
|
2026-01-30 11:48:53 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-11-20 18:05:07 +03:00
|
|
|
|
2024-06-04 12:30:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 23:26:30 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-07-06 15:39:11 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:39:11 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-11-05 16:36:22 +01:00
|
|
|
|
2024-02-24 13:53:15 +00:00
|
|
|
|
2024-12-09 19:35:36 -05:00
|
|
|
|
2025-11-05 16:36:22 +01:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-04-23 22:32:03 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
2017-07-06 15:39:11 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-11-20 18:05:52 +03:00
|
|
|
|
2026-01-30 11:48:53 +08:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2017-11-20 18:05:07 +03:00
|
|
|
|
2024-06-04 12:30:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-06-27 18:19:11 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-08-06 15:46:20 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
2025-11-19 19:19:24 -05:00
|
|
|
|
2007-10-16 23:26:30 -07:00
|
|
|
|
|
|
|
|
|
2015-08-06 15:46:20 -07:00
|
|
|
|
|
|
|
|
|
2007-10-16 23:26:30 -07:00
|
|
|
|
2005-04-16 15:20:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
