Merge tag 'vfs-7.4-rc1.kernfs' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs

0ddc8871 · Linus Torvalds <[email protected]> · 2026-10-19 18:52:22 -10:00

Pull kernfs updates from Christian Brauner:
 "This contains kernfs scalability work from Shakeel driven by what Meta
  sees in their fleet. kernfs_rwsem is per kernfs root, so sysfs and
  cgroupfs each have one for the whole machine, and every create, remove
  and rename takes it for writing. rwsems don't let readers steal the
  lock once a writer is queued, so one cgroup being created or a device
  being renamed parks the entire incoming reader stream in
  uninterruptible sleep. Daemons that poll cgroup files in a loop turn
  that into a convoy and the two that hurt most, oomd and below, are the
  ones that are supposed to relieve memory pressure and record what
  happened.

  kernfs_dop_revalidate() took kernfs_rwsem for read once per path
  component of every walk into a kernfs mount. Nothing it reads needs
  the semaphore. kn->active is an atomic_t, kn->__parent and kn->name
  are rcu pointers, kn->ns can be compared instead of dereferenced and
  the directory revision is a plain counter. The callback runs under rcu
  now. Readers walking cgroupfs and sysfs while another thread churns
  cgroups, renames netdevs and plugs devices went from 48,593,360 read
  acquisitions and 1,744,517 contentions to 1,280,280 and 429,846 on an
  8-CPU VM and path walk throughput improved 35-53%. LOOKUP_RCU still
  returns ECHILD because kernfs_iop_permission() forces every walk out
  of rcu-walk anyway. That is left for later.

  kernfs_fop_readdir() held the semaphore across dir_emit(). dir_emit()
  copies to userspace and under memory pressure that fault goes to
  reclaim with the lock every create, remove and rename needs held for
  reading. Meta sees 120 second allocation stalls in filldir64() from
  below and systemd-oomd stuck behind it. The lock is dropped around
  dir_emit() now. Since the entry the listing stopped on can disappear
  in that window and two entries can share a hash, readdir resumes on
  the full (hash, ns, name) key and the fallback search lands after the
  missing entry instead of repeating or skipping one. The repeat was a
  pre-existing bug between two getdents(2) calls.

  Three more lock holds are shortened. kernfs_notify_workfn() no longer
  holds kernfs_rwsem across fsnotify() which allocates against the memcg
  of the watching cgroup that is already under pressure. The open node
  is allocated outside the hashed open file mutex, the most contended
  lock in kernfs on their fleet with 2.19M waiters over 29 days and a
  77ms median hold. A new node is activated without dropping and
  retaking the write lock, lock_stat says 1200 -> 830 write acquisitions
  for five dummy netdevs, and kernfs_rename_ns() allocates the new name
  before and frees the old one after the lock so a rename can't enter
  reclaim or wait out a grace period with it held.

  Three standalone fixes come with selftests. Same-parent renames take
  kernfs_rename_lock too so kernfs_path_from_node() can't build a path
  that never existed. Decoding a file handle takes the semaphore so a
  decode racing rmdir() can't hash an inode after the removal's
  ilookup() pass and lose IN_DELETE_SELF. And kernfs_create_link() reads
  the target's uid and gid under kernfs_iattr_rwsem so a chown in
  between can't give the link an owner the target never had. The kernfs
  selftests grew from two xattr checks to coverage of dentry
  revalidation, rename, namespace tagging, file handles and readdir
  under churn"

* tag 'vfs-7.4-rc1.kernfs' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs:
  kernfs: free the old name outside kernfs_rwsem
  kernfs: allocate the new name outside kernfs_rwsem
  kernfs: activate a new node without dropping kernfs_rwsem
  selftests: cover readdir resuming at a removed entry
  kernfs: don't hold kernfs_rwsem across dir_emit()
  kernfs: resume readdir on the full entry key
  kernfs: don't repeat or skip an entry when readdir resumes
  kernfs: don't hold kernfs_rwsem across notification delivery
  kernfs: allocate the open node outside the open file mutex
  kernfs: fix up the unlocked attribute reads on the creation paths
  kernfs: don't lose IN_DELETE_SELF when decoding a file handle
  kernfs: take kernfs_rename_lock for same-parent renames too
  selftests: cover kernfs file handles and same-parent rename
  selftests: cover kernfs dentry revalidation
  kernfs: Remove kernfs_rwsem from dentry revalidation
  kernfs: Avoid namespace dereference in d_revalidate()
  kernfs: Prepare directory revisions for lockless reads
  kernfs: Use VFS lookup context in d_revalidate()

Commit

0ddc887102d8a13ce85041a41c45f2053b483119

Parents

27dd2772 · 1c45110a

Merge commit. Changes below are compared with the first parent: 27dd2772.

Changes

8 files changed · +1414 −121