Re: [PATCH V2] perf/x86/rapl: fix deadlock in rapl_pmu_event_stop

[duoming]

Hello,

On Wed, 21 Sep 2022 09:53:04 +0200 Ingo Molnar wrote:

> * Duoming Zhou <duoming@xxxxxxxxxx> 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.
> > 
> > Fixes: 65661f96d3b3 ("perf/x86: Add RAPL hrtimer support")
> > Signed-off-by: Duoming Zhou <duoming@xxxxxxxxxx>
> > ---
> > Changes in v2:
> >   - Move hrtimer_cancel() to the end of rapl_pmu_event_stop() function.
> > 
> >  arch/x86/events/rapl.c | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > 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;
> > +	}
> 
> Looks racy now: AFAICS now it's possible for rapl_hrtimer_handle() to 
> execute at an arbitrary moment after pmu->lock is dropped - which could be 
> use-after-free after cleanup_rapl_pmus() executes and the PMU is freed, 
> right?
> 
> There's also the quality-of-implementation issue of the hrtimer executing 
> in a delayed fashion for the *next* event that may have been added, leading 
> to possibly unexpected results.

Thank your for your suggestions! In order to solve the above problems,
I come up with the following solution.

diff --git a/arch/x86/events/rapl.c b/arch/x86/events/rapl.c
index 77e3a47af5a..a526a08ee6e 100644
--- a/arch/x86/events/rapl.c
+++ b/arch/x86/events/rapl.c
@@ -219,11 +219,13 @@ static enum hrtimer_restart rapl_hrtimer_handle(struct hrtimer *hrtimer)
        struct perf_event *event;
        unsigned long flags;

-       if (!pmu->n_active)
-               return HRTIMER_NORESTART;
-
        raw_spin_lock_irqsave(&pmu->lock, flags);

+       if (!pmu->n_active) {
+               raw_spin_unlock_irqrestore(&pmu->lock, flags);
+               return HRTIMER_NORESTART;
+       }
+
        list_for_each_entry(event, &pmu->active_list, active_entry)
                rapl_event_update(event);

@@ -282,7 +284,7 @@ static void rapl_pmu_event_stop(struct perf_event *event, int mode)
                WARN_ON_ONCE(pmu->n_active <= 0);
                pmu->n_active--;
                if (pmu->n_active == 0)
-                       hrtimer_cancel(&pmu->hrtimer);
+                       hrtimer_try_to_cancel(&pmu->hrtimer);

                list_del(&event->active_entry);

Firstly, the deadlock could be mitigated. Because if the timer callback function
is running, the hrtimer_try_to_cancel() will directly return. 

Secondly, the race could be avoided. Because we use pmu->lock to synchronize and
move the check "if (!pmu->n_active)" into the protection scope of pmu->lock.
If the rapl_pmu_event_stop() has finished, the "pmu->n_active" equals to 0 and
the rapl_hrtimer_handle() will return "HRTIMER_NORESTART".

Thirdly, this solution will not cause quality-of-implementation issue of the hrtimer.

Best regards,
Duoming Zhou