Re: [PATCH v4] fs/namespace: defer RCU sync for MNT_DETACH umount

[Eric Chanudet]

On Wed, Apr 09, 2025 at 03:04:43PM +0200, Mateusz Guzik wrote:
> On Tue, Apr 08, 2025 at 04:58:34PM -0400, Eric Chanudet wrote:
> > Defer releasing the detached file-system when calling namespace_unlock()
> > during a lazy umount to return faster.
> > 
> > When requesting MNT_DETACH, the caller does not expect the file-system
> > to be shut down upon returning from the syscall. Calling
> > synchronize_rcu_expedited() has a significant cost on RT kernel that
> > defaults to rcupdate.rcu_normal_after_boot=1. Queue the detached struct
> > mount in a separate list and put it on a workqueue to run post RCU
> > grace-period.
> > 
> > w/o patch, 6.15-rc1 PREEMPT_RT:
> > perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount mnt
> >     0.02455 +- 0.00107 seconds time elapsed  ( +-  4.36% )
> > perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount -l mnt
> >     0.02555 +- 0.00114 seconds time elapsed  ( +-  4.46% )
> > 
> > w/ patch, 6.15-rc1 PREEMPT_RT:
> > perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount mnt
> >     0.026311 +- 0.000869 seconds time elapsed  ( +-  3.30% )
> > perf stat -r 10 --null --pre 'mount -t tmpfs tmpfs mnt' -- umount -l mnt
> >     0.003194 +- 0.000160 seconds time elapsed  ( +-  5.01% )
> > 
> 
> Christian wants the patch done differently and posted his diff, so I'm
> not going to comment on it.
> 
> I do have some feedback about the commit message though.
> 
> In v1 it points out a real user which runs into it, while this one does
> not. So I would rewrite this and put in bench results from the actual
> consumer -- as it is one is left to wonder why patching up lazy unmount
> is of any significance.

Certainly. Doing the test mentioned in v1 again with v4+Christian's
suggested changes:
- QEMU x86_64, 8cpus, PREEMPT_RT, w/o patch:
# perf stat -r 10 --table --null -- crun run test
    0.07584 +- 0.00440 seconds time elapsed  ( +-  5.80% )
- QEMU x86_64, 8cpus, PREEMPT_RT, w/ patch:
# perf stat -r 10 --table --null -- crun run test
    0.01421 +- 0.00387 seconds time elapsed  ( +- 27.26% )

I will add that to the commit message.

> I had to look up what rcupdate.rcu_normal_after_boot=1 is. Docs claim it
> makes everyone use normal grace-periods, which explains the difference.
> But without that one is left to wonder if perhaps there is a perf bug in
> RCU instead where this is taking longer than it should despite the
> option. Thus I would also denote how the delay shows up.

I tried the test above while trying to force expedited RCU on the
cmdline with:
    rcupdate.rcu_normal_after_boot=0 rcupdate.rcu_expedited=1

Unfortunately, rcupdate.rcu_normal_after_boot=0 has no effect and
rcupdate_announce_bootup_oddness() reports:
[    0.015251] 	No expedited grace period (rcu_normal_after_boot).

Which yielded similar results:
- QEMU x86_64, 8cpus, PREEMPT_RT, w/o patch:
# perf stat -r 10 --table --null -- crun run test
    0.07838 +- 0.00322 seconds time elapsed  ( +-  4.11% )
- QEMU x86_64, 8cpus, PREEMPT_RT, w/ patch:
# perf stat -r 10 --table --null -- crun run test
    0.01582 +- 0.00353 seconds time elapsed  ( +- 22.30% )

I don't think rcupdate.rcu_expedited=1 had an effect, but I have not
confirmed that yet.

> v1 for reference:
> > v1: https://lore.kernel.org/all/20230119205521.497401-1-echanude@xxxxxxxxxx/

-- 
Eric Chanudet