Re: [PATCH] perf/x86/rapl: fix deadlock in rapl_pmu_event_stop

[duoming@zju.edu.cn]

Hello,

On Mon, 19 Sep 2022 13:45:38 +0200 Peter Zijlstra wrote:

> On Sat, Sep 17, 2022 at 10:47:29PM +0800, Duoming Zhou wrote:
> > There is a deadlock in rapl_pmu_event_stop(), the process is
> > shown below:
> > 
> >     (thread 1)                 |        (thread 2)
> > rapl_pmu_event_stop()          | rapl_hrtimer_handle()
> >  ...                           |  if (!pmu->n_active)
> >  raw_spin_lock_irqsave() //(1) |  ...
> >   ...                          |
> >   hrtimer_cancel()             |  raw_spin_lock_irqsave() //(2)
> >   (block forever)
> > 
> > We hold pmu->lock in position (1) and use hrtimer_cancel() to wait
> > rapl_hrtimer_handle() to stop, but rapl_hrtimer_handle() also need
> > pmu->lock in position (2). As a result, the rapl_pmu_event_stop()
> > will be blocked forever.
> > 
> > This patch extracts hrtimer_cancel() from the protection of
> > raw_spin_lock_irqsave(). As a result, the rapl_hrtimer_handle() could
> > obtain the pmu->lock. In order to prevent race conditions, we put
> > "if (!pmu->n_active)" in rapl_hrtimer_handle() under the protection
> > of raw_spin_lock_irqsave().
> > 
> > Fixes: 65661f96d3b3 ("perf/x86: Add RAPL hrtimer support")
> > Signed-off-by: Duoming Zhou <duoming@zju.edu.cn>
> > ---
> >  arch/x86/events/rapl.c | 9 ++++++---
> >  1 file changed, 6 insertions(+), 3 deletions(-)
> > 
> > diff --git a/arch/x86/events/rapl.c b/arch/x86/events/rapl.c
> > index 77e3a47af5a..97c71538d01 100644
> > --- a/arch/x86/events/rapl.c
> > +++ b/arch/x86/events/rapl.c
> > @@ -219,11 +219,11 @@ static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
> >  	struct perf_event *event;
> >  	unsigned long flags;
> >  
> > +	raw_spin_lock_irqsave(&pmu->lock, flags);
> > +
> >  	if (!pmu->n_active)
> >  		return HRTIMER_NORESTART;
> 
> Except now you return with the lock held...
> 
> >  
> > -	raw_spin_lock_irqsave(&pmu->lock, flags);
> > -
> >  	list_for_each_entry(event, &pmu->active_list, active_entry)
> >  		rapl_event_update(event);
> >  
> > @@ -281,8 +281,11 @@ static void rapl_pmu_event_stop(struct perf_event *event, int mode)
> >  	if (!(hwc->state & PERF_HES_STOPPED)) {
> >  		WARN_ON_ONCE(pmu->n_active <= 0);
> >  		pmu->n_active--;
> > -		if (pmu->n_active == 0)
> > +		if (!pmu->n_active) {
> > +			raw_spin_unlock_irqrestore(&pmu->lock, flags);
> >  			hrtimer_cancel(&pmu->hrtimer);
> > +			raw_spin_lock_irqsave(&pmu->lock, flags);
> 
> Doing a lock-break makes the nr_active and list_del thing non-atomic,
> breaking the whole purpose of the lock.

Thank you for your time and suggestions! I come up with another solution that
will not break the atomicity, the detail is shown below: 

diff --git a/arch/x86/events/rapl.c b/arch/x86/events/rapl.c
index 77e3a47af5a..7c110092c83 100644
--- a/arch/x86/events/rapl.c
+++ b/arch/x86/events/rapl.c
@@ -281,8 +281,6 @@ static void rapl_pmu_event_stop(struct perf_event *event, int mode)
        if (!(hwc->state & PERF_HES_STOPPED)) {
                WARN_ON_ONCE(pmu->n_active <= 0);
                pmu->n_active--;
-               if (pmu->n_active == 0)
-                       hrtimer_cancel(&pmu->hrtimer);
 
                list_del(&event->active_entry);
 
@@ -300,6 +298,11 @@ static void rapl_pmu_event_stop(struct perf_event *event, int mode)
                hwc->state |= PERF_HES_UPTODATE;
        }
 
+       if (!pmu->n_active) {
+               raw_spin_unlock_irqrestore(&pmu->lock, flags);
+               hrtimer_cancel(&pmu->hrtimer);
+               return;
+       }
        raw_spin_unlock_irqrestore(&pmu->lock, flags);
 }

I move the hrtimer_cancel() to the end of the rapl_pmu_event_stop() function.
As a result, the atomicity will not break and the deadlock bug could be mitigated.

> > +		}
> >  
> >  		list_del(&event->active_entry);
> 
> 
> Now; did you actually observe this deadlock or is this a code-reading
> exercise? If you saw an actual deadlock, was cpu-hotplug involved?

I found this bug through a static analysis tool written by myself.

Thanks you!

Best regards,
Duoming Zhou