[PATCH 0/6] kernfs: stop taking kernfs_iattr_rwsem for writing in mkdir/rmdir

From: Shakeel Butt

Date: Wed Nov 18 2026 - 17:02:10 EST


In the Meta fleet, cgroup mkdir and rmdir can stall cgroup operations
on the whole host. They hold cgroup_mutex and kernfs_rwsem while they
wait for kernfs_iattr_rwsem, which every path walk takes for reading,
and they take it once or twice for every file of the cgroup. systemd,
clone3() into a cgroup, BPF queries and the OOM daemon all wait behind
them.

- Our lock monitor saw a kernfs_rwsem writer blocked on
kernfs_iattr_rwsem several thousand times on a large number of hosts
in 29 days (in mkdir and rmdir), for up to 1.8 s at a time.

- Under memory and CPU pressure the waits become host-wide stalls. In
one week, we observed several cases where the OOM daemon and the
monitoring agent were stuck for up to 22 minutes. In October, one
batch job caused these stalls on a large number of hosts. On the
sampled hosts with current kernels, the cgroup_mutex owner was always
waiting for kernfs_iattr_rwsem (in mkdir and rmdir). One specific
stall lasted 90 minutes and ended in a reboot.

The write lock is not needed for most of what it covers:

- kernfs_rwsem already serialises the parent's two counters and
KERNFS_REMOVING (patches 3 and 4).
- The parent's times get a seqlock of their own (patch 5).
- Only inode setup needs to set a directory's i_nlink, so removal can
clear it without the lock (patches 2 and 6).

The lock also kept times and link counts stable while getattr and
permission copied them into the shared inode. Patches 1 and 2 make
getattr report them from the node and stop copying them, so stat()
cannot return a stale or torn value.

After the series, mkdir and rmdir take no kernfs_iattr_rwsem for
writing. Write acquisitions, from lock_stat:

before after
5 dummy netdevs created and removed 1850 0
100 cgroup mkdirs and 100 rmdirs 7500 0

Latency in a VM with 8 vCPUs, while 32 threads walk cgroupfs paths
(worst of three 20 s runs; mkdir and rmdir by a SCHED_FIFO thread):

before after
p50 p99 max p50 p99 max
cgroup mkdir 2.8ms 9.0ms 20.1ms 25us 37us 84us
cgroup rmdir 4.2ms 12.7ms 26.6ms 28us 38us 107us

A write that only needs cgroup_mutex waited up to 23 ms before and up
to 34 us after. With 8 threads and no writer, stat() and access() on
a path whose directories have iattrs are 12-14% faster, since the
refresh copies less.

Tested with the kernfs selftests on lockdep + KASAN and KCSAN kernels;
with extra checks of link counts, times and IN_DELETE_SELF; with
concurrent mkdir/rmdir, setattr, stat, readdir and netdev churn; and
with test-only delays that force the stale stat(), lost time update,
backwards time, overwritten setattr and stat()-during-removal races.
KCSAN reports nothing new, and its set_nlink() reports from the refresh
are gone.

Based on vfs-7.5.kernfs. It also applies to v7.4-rc3.

Shakeel Butt (6):
kernfs: report times and link count from the node in getattr
kernfs: don't copy times and link count into the inode on refresh
kernfs: don't take kernfs_iattr_rwsem for the directory counters
kernfs: don't take kernfs_iattr_rwsem to publish KERNFS_REMOVING
kernfs: update a directory's times without kernfs_iattr_rwsem
kernfs: don't take kernfs_iattr_rwsem in __kernfs_remove()

fs/kernfs/dir.c | 48 +++++---------
fs/kernfs/inode.c | 125 ++++++++++++++++++++++++++++++------
fs/kernfs/kernfs-internal.h | 2 +
3 files changed, 124 insertions(+), 51 deletions(-)


base-commit: ccf2a4692fed0fa77751d32d2690585127a152c5
--
2.53.1-Meta