On Wed, Jun 07, 2023 at 12:21:42AM +0300, Kirill Tkhai wrote:
> On 06.06.2023 04:24, Dave Chinner wrote:
> > On Mon, Jun 05, 2023 at 05:38:27PM -0700, Roman Gushchin wrote:
> >> On Mon, Jun 05, 2023 at 10:03:25PM +0300, Kirill Tkhai wrote:
> >>> Kernel test robot reports -88.8% regression in stress-ng.ramfs.ops_per_sec
> >>> test case caused by commit: f95bdb700bc6 ("mm: vmscan: make global slab
> >>> shrink lockless"). Qi Zheng investigated that the reason is in long SRCU's
> >>> synchronize_srcu() occuring in unregister_shrinker().
> >>>
> >>> This patch fixes the problem by using new unregistration interfaces,
> >>> which split unregister_shrinker() in two parts. First part actually only
> >>> notifies shrinker subsystem about the fact of unregistration and it prevents
> >>> future shrinker methods calls. The second part completes the unregistration
> >>> and it insures, that struct shrinker is not used during shrinker chain
> >>> iteration anymore, so shrinker memory may be freed. Since the long second
> >>> part is called from delayed work asynchronously, it hides synchronize_srcu()
> >>> delay from a user.
> >>>
> >>> Signed-off-by: Kirill Tkhai <tkhai@xxxxx>
> >>> ---
> >>> fs/super.c | 3 ++-
> >>> 1 file changed, 2 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/fs/super.c b/fs/super.c
> >>> index 8d8d68799b34..f3e4f205ec79 100644
> >>> --- a/fs/super.c
> >>> +++ b/fs/super.c
> >>> @@ -159,6 +159,7 @@ static void destroy_super_work(struct work_struct *work)
> >>> destroy_work);
> >>> int i;
> >>>
> >>> + unregister_shrinker_delayed_finalize(&s->s_shrink);
> >>> for (i = 0; i < SB_FREEZE_LEVELS; i++)
> >>> percpu_free_rwsem(&s->s_writers.rw_sem[i]);
> >>> kfree(s);
> >>> @@ -327,7 +328,7 @@ void deactivate_locked_super(struct super_block *s)
> >>> {
> >>> struct file_system_type *fs = s->s_type;
> >>> if (atomic_dec_and_test(&s->s_active)) {
> >>> - unregister_shrinker(&s->s_shrink);
> >>> + unregister_shrinker_delayed_initiate(&s->s_shrink);
> >>
> >> Hm, it makes the API more complex and easier to mess with. Like what will happen
> >> if the second part is never called? Or it's called without the first part being
> >> called first?
> >
> > Bad things.
> >
> > Also, it doesn't fix the three other unregister_shrinker() calls in
> > the XFS unmount path, nor the three in the ext4/mbcache/jbd2 unmount
> > path.
> >
> > Those are just some of the unregister_shrinker() calls that have
> > dynamic contexts that would also need this same fix; I haven't
> > audited the 3 dozen other unregister_shrinker() calls around the
> > kernel to determine if any of them need similar treatment, too.
> >
> > IOWs, this patchset is purely a band-aid to fix the reported
> > regression, not an actual fix for the underlying problems caused by
> > moving the shrinker infrastructure to SRCU protection. This is why
> > I really want the SRCU changeover reverted.
> >
> > Not only are the significant changes the API being necessary, it's
> > put the entire shrinker paths under a SRCU critical section. AIUI,
> > this means while the shrinkers are running the RCU grace period
> > cannot expire and no RCU freed memory will actually get freed until
> > the srcu read lock is dropped by the shrinker.
>
> Why so? Doesn't SRCU and RCU have different grace period and they does not prolong
> each other?
No idea - Documentation/RCU/whatisRCU.rst doesn't describe any
differences between SRCU and RCU except for "use SRCU if you need to
sleep in the read side" and there's no discussion of how they
interact, either. maybe there's some discussion in other RCU
documentation, but there's nothing in the "how to use RCU"
documentation that tells me they use different grace period
definitions...
> Also, it looks like every SRCU has it's own namespace like shrinker_srcu for shrinker.
> Don't different SRCU namespaces never prolong each other?!
RIght, SRCU vs SRCU is well defined. What is not clear from anything
I've read is SRCU vs RCU interactions, so I can only assuming from
the shared "RCU" in the name there are shared implementation
details and interactions...
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
