Re: [perf] WARNING: CPU: 0 PID: 1457 at kernel/events/core.c:890 add_event_to_ctx()

[Mark Rutland]

On Sun, Jan 25, 2015 at 03:56:33PM +0000, Mark Rutland wrote:
> On Sun, Jan 25, 2015 at 04:34:28AM +0000, Fengguang Wu wrote:
> > Greetings,
> 
> Hi Fengguang,
> 
> Thanks very much for the report.
> 
> > 0day kernel testing robot got the below dmesg and the first bad commit is
> > 
> > git://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git perf/core
> > 
> > commit d26bb7f73a2881f2412c340a27438b185f0cc3dc
> > Author:     Mark Rutland <mark.rutland@xxxxxxx>
> > AuthorDate: Wed Jan 7 15:01:54 2015 +0000
> > Commit:     Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> > CommitDate: Fri Jan 23 15:17:56 2015 +0100
> > 
> >     perf: decouple unthrottling and rotating
> 
> [...]
> 
> What seems to be happening is:
> 
> * An event is created in the context of task A on CPU0. As the task's hw
>   context is empty of events, we call perf_pmu_ctx_activate. This adds
>   the cpuctx of the relevant HW PMU to the active_ctx_list. Note that we
>   checked the task's ctx for emptiness, then added the PMU's hw context.
> 
> * An event is created (as a result of a clone()) in the context of task
>   B on CPU0, and we do the same thing, finding the ctx empty of events
>   we add the HW PMU's cpuctx to the active_ctx_list. As it's already in
>   there, the WARN_ON(!list_empty(&cpuctx->active_ctx_list)) explodes.
> 
> So I guess what we actually want to do is to turn the active_ctx list
> into a list of perf_event_contexts rather than perf_event_cpu_contexts,
> and add/remove from the list when a context is scheduled in/out (or
> updated empty<->nonempty). That way we remove chances for erroneous
> add/remove, and we don't need to treat task and CPU contexts separately
> in perf_event_task_tick.
> 
> I managed to get the reproducer running on a box at home, so I'll spin a
> potential fix against that for a while, and send that out if I don't see
> explosions.

The below patch is a fix up to the broken patch above. I've given it
some testing with Fengguang's reproducer and some custom tests, and as
far as I can tell everything is working as expected.

With Fengguang's reproducer I see some failures in netlink, but I
believe that's an unrelated issue that trinity is tickling.

Peter, does the below patch look OK to you?

Thanks,
Mark.

---->8----