Re: [PATCH 2/3] cgroup: add lockless fast-path checks to cgroup_file_notify()

[Shakeel Butt]

Hi Chen, thanks for taking a look.

On Mon, Mar 02, 2026 at 09:50:53AM +0800, Chen Ridong wrote:
> 
> Hi Shakeel,
> 
> Good series to move away from the global lock.
> 
> On 2026/2/28 22:20, Shakeel Butt wrote:
> > Add two lockless checks before acquiring the lock:
> > 
> > 1. READ_ONCE(cfile->kn) NULL check to skip torn-down files.
> > 2. READ_ONCE(cfile->notified_at) check to skip when within the
> >    rate-limit window (~10ms).
> > 
> > Both checks have safe error directions -- a stale read can only cause
> > unnecessary lock acquisition, never a missed notification.  Annotate
> > all write sites with WRITE_ONCE() to pair with the lockless readers.
> > 
> > The trade-off is that trailing timer_reduce() calls during bursts are
> > skipped, so the deferred notification that delivers the final state
> > may be lost.  This is acceptable for the primary callers like
> > __memcg_memory_event() where events keep arriving.
> > 
> > Signed-off-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
> > Reported-by: Jakub Kicinski <kuba@xxxxxxxxxx>
> > ---
> >  kernel/cgroup/cgroup.c | 21 ++++++++++++++-------
> >  1 file changed, 14 insertions(+), 7 deletions(-)
> > 
> > diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
> > index 33282c7d71e4..5473ebd0f6c1 100644
> > --- a/kernel/cgroup/cgroup.c
> > +++ b/kernel/cgroup/cgroup.c
> > @@ -1749,7 +1749,7 @@ static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
> >  		struct cgroup_file *cfile = (void *)css + cft->file_offset;
> >  
> >  		spin_lock_irq(&cgroup_file_kn_lock);
> > -		cfile->kn = NULL;
> > +		WRITE_ONCE(cfile->kn, NULL);
> >  		spin_unlock_irq(&cgroup_file_kn_lock);
> >  
> >  		timer_delete_sync(&cfile->notify_timer);
> > @@ -4430,7 +4430,7 @@ static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
> >  		timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
> >  
> >  		spin_lock_irq(&cgroup_file_kn_lock);
> > -		cfile->kn = kn;
> > +		WRITE_ONCE(cfile->kn, kn);
> >  		spin_unlock_irq(&cgroup_file_kn_lock);
> >  	}
> >  
> > @@ -4686,20 +4686,27 @@ int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
> >   */
> >  void cgroup_file_notify(struct cgroup_file *cfile)
> >  {
> > -	unsigned long flags;
> > +	unsigned long flags, last, next;
> >  	struct kernfs_node *kn = NULL;
> >  
> > +	if (!READ_ONCE(cfile->kn))
> > +		return;
> > +
> > +	last = READ_ONCE(cfile->notified_at);
> > +	if (time_before_eq(jiffies, last + CGROUP_FILE_NOTIFY_MIN_INTV))
> > +		return;
> > +
> 
> Previously, if a notification arrived within the rate-limit window, we would
> still call timer_reduce(&cfile->notify_timer, next) to schedule a deferred
> notification.
> 
> With this change, returning early here bypasses that timer scheduling entirely.
> Does this risk missing notifications that would have been delivered by the timer?
> 

You are indeed right that this can cause missed notifications. After giving some
thought I think the lockless check-and-return can be pretty much simplified to
timer_pending() check. If timer is active, just do nothing and the notification
will be delivered eventually.

I will send the updated version soon. Any comments on the other two patches?