diff options
| author | Mateusz Guzik <mjguzik@gmail.com> | 2025-12-03 10:48:37 +0100 |
|---|---|---|
| committer | Christian Brauner <brauner@kernel.org> | 2025-12-15 14:33:38 +0100 |
| commit | 887e97745ec336c2f49b6c0af3c4cc00a5df3211 (patch) | |
| tree | 98ee23214c9bcfea4556a378b089a9b2bc71784a /fs/locks.c | |
| parent | 1fa4e69a54a250fa17d2afd9c5b54a59329033c1 (diff) | |
fs: track the inode having file locks with a flag in ->i_opflags
Opening and closing an inode dirties the ->i_readcount field.
Depending on the alignment of the inode, it may happen to false-share
with other fields loaded both for both operations to various extent.
This notably concerns the ->i_flctx field.
Since most inodes don't have the field populated, this bit can be managed
with a flag in ->i_opflags instead which bypasses the problem.
Here are results I obtained while opening a file read-only in a loop
with 24 cores doing the work on Sapphire Rapids. Utilizing the flag as
opposed to reading ->i_flctx field was toggled at runtime as the benchmark
was running, to make sure both results come from the same alignment.
before: 3233740
after: 3373346 (+4%)
before: 3284313
after: 3518711 (+7%)
before: 3505545
after: 4092806 (+16%)
Or to put it differently, this varies wildly depending on how (un)lucky
you get.
The primary bottleneck before and after is the avoidable lockref trip in
do_dentry_open().
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Link: https://patch.msgid.link/20251203094837.290654-2-mjguzik@gmail.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
Diffstat (limited to 'fs/locks.c')
| -rw-r--r-- | fs/locks.c | 14 |
1 files changed, 12 insertions, 2 deletions
diff --git a/fs/locks.c b/fs/locks.c index 9f565802a88c..7a63fa3ca9b4 100644 --- a/fs/locks.c +++ b/fs/locks.c @@ -178,7 +178,6 @@ locks_get_lock_context(struct inode *inode, int type) { struct file_lock_context *ctx; - /* paired with cmpxchg() below */ ctx = locks_inode_context(inode); if (likely(ctx) || type == F_UNLCK) goto out; @@ -196,7 +195,18 @@ locks_get_lock_context(struct inode *inode, int type) * Assign the pointer if it's not already assigned. If it is, then * free the context we just allocated. */ - if (cmpxchg(&inode->i_flctx, NULL, ctx)) { + spin_lock(&inode->i_lock); + if (!(inode->i_opflags & IOP_FLCTX)) { + VFS_BUG_ON_INODE(inode->i_flctx, inode); + WRITE_ONCE(inode->i_flctx, ctx); + /* + * Paired with locks_inode_context(). + */ + smp_store_release(&inode->i_opflags, inode->i_opflags | IOP_FLCTX); + spin_unlock(&inode->i_lock); + } else { + VFS_BUG_ON_INODE(!inode->i_flctx, inode); + spin_unlock(&inode->i_lock); kmem_cache_free(flctx_cache, ctx); ctx = locks_inode_context(inode); } |
