Re: [PATCH 1/1] perf/core: fix dangling cgroup pointer in cpuctx

[Yeoreum Yun <yeoreum.yun@arm.com>]

Hi David,

> > > > >
> > > > > Also, your patch couldn't solve a problem describe in
> > > > > commit a3c3c6667("perf/core: Fix child_total_time_enabled accounting bug at task exit")
> > > > > for INCATIVE event's total_enable_time.
> > > >
> > > > I do not think so.
> > > > Correct me if I am making silly  mistakes,
> > > > The patch, https://lore.kernel.org/lkml/20250603032651.3988-1-00107082@163.com/
> > > > calls perf_event_set_state() based on DETACH_EXIT flag, which cover the INACTIVE state, right?
> > > > If DETACH_EXIT is not used for this purpose? Then why should it exist at the first place?
> > > > I think I does not revert the purpose of commit a3c3c6667.....But I could be wrong
> > > > Would you show a call path where DETACH_EXIT is not set, but the changes in commit a3c3c6667 is still needed?
> > > >
> > > > Sorry for my bad explaination without detail.
> > > > Think about cpu specific event and closed by task.
> > > > If there is specific child cpu event specified in cpu 0.
> > > > 1. cpu 0 -> active
> > > > 2. scheulded to cpu1 -> inactive
> > > > 3. close the cpu event from parent -> inactive close
> > > >
> > > > Can be failed to count total_enable_time.
> > >
> > > Is this explaining the purpose of commit a3c3c6667 ?
> > > I am not arguing with it. And I also not suggest reverting it. (it is just that reverting it can fix the kernel panic.)
> >
> > In commit a3c3c6667, I explain the specific case but not with above
> > case. But the commit's purpose is "account total_enable_time" properly.
> >
> > > > And also, considering the your patch, for DETACH_EXIT case,
> > > > If it changes the state before list_del_event() that wouldn't disable
> > > > related to the cgroup. So it would make cpuctx->cgrp pointer could be dangled
> > > > as patch describe...
> > > No, I don't think so.
> > > change state before list_del_event(), this is the same behavior before commit a3c3c6667, right?
> > > And no such kernel panic happened  before commit a3c3c6667.
>
> Oh! I was wrong, before commit a3c3c6667, "change state" happened *after* list_del_event()
> >
> > That's why list_del_event() handle the perf_cgroup_disable() before the
> > commit a3c3c6667. However because of *my mistake*, I've forget to
> > perf_cgroup_disable() properly before change the event state.
> > Yes, your patch can make avoid the panic since as soon as exit,
> > the event->cgrp switched.
>
> I cannot agree with the reasoning,
> The panic dose not happened when exit, it happened when reboot/shutdown.
> (I close perf_event_open before reboot)
> >
> > However, as I said, the INACTIVE event could be failed to count
> >total_enable_time.
> >
> > So, set event should be occured before list_del_event().
> >And since it's event->state change on remove.
> >It shouldn't have any side effect the state change isn't cause of your
> > panic. But missed perf_cgroup_disable().
>
> Any procedure to bring out the impact of this missed perf_cgroup_disable()?
> My system seems all normal, where should I check it?

Here is possible senario:
  1. perf event open with cgroup.
  2. perf event open with cpu event (no cgroup).
  3. above task sets the cpuctx->cgrp the same to (1).
  3. close (1) events.
     here, perf_cgroup_event_disable() isn't called,
     cpuctx->cgrp still point the cgroup.
  4. by other task, the cgroup and is destroied.
  5. close (2) events.
     here, it is last event, in __perf_remove_from_context()
     and last event, it calls update_cgrp_time_from_cpuctx(),
     And this refers invalid pointer.

> But to fix it,  isn't following change less aggressive?
>         event_sched_out(event, ctx);
> -       perf_event_set_state(event, min(event->state, state));
>         if (flags & DETACH_GROUP)
>                 perf_group_detach(event);
>         if (flags & DETACH_CHILD)
>                perf_child_detach(event);
>         list_del_event(event, ctx);
> +       perf_event_set_state(event, min(event->state, state));

If perf_child_detach() is called first and perf_event_set_state() call,
since the parent is removed in perf_child_detatced,
It would be failed to account the total_enable_time which caculating
child_event's enable_time too.

Thanks

--
Sincerely,
Yeoreum Yun