Re: perf_events: question about __perf_event_read()

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2011-01-21 at 19:06 +0100, Stephane Eranian wrote:
> Hi,
> 
> I think the code below still has a problem in case of a per-cpu event.
> 
> If you issue a read() on a different CPU, then you IPI to the event's cpu.
> By the time you get there, the event may be de-scheduled in which
> case you don't want to issue event->pmu_read() nor update context
> timings. The function has a test but it seems to be checking the per-cpu
> case only.
> 
> I have seen panics on P4 with this code because it goes all the way
> down to rdmsrl() with a bogus counter index (like -1).
> 
> Am I missing something here?
> 
> static void __perf_event_read(void *info)
> {
>         struct perf_event *event = info;
>         struct perf_event_context *ctx = event->ctx;
>         struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
> 
>         /*
>          * If this is a task context, we need to check whether it is
>          * the current task context of this cpu.  If not it has been
>          * scheduled out before the smp call arrived.  In that case
>          * event->count would have been updated to a recent sample
>          * when the event was scheduled out.
>          */
>         if (ctx->task && cpuctx->task_ctx != ctx)
>                 return;
> 
>         raw_spin_lock(&ctx->lock);

Shouldn't we re-check event->state once we hold this lock?

>         update_context_time(ctx);
>         update_event_times(event);
>         raw_spin_unlock(&ctx->lock);
> 
>         event->pmu->read(event);
> }