Messages in this thread | ![/](/images/icornerl.gif) | | Date | Wed, 10 Apr 2024 12:38:50 +0200 | From | Frederic Weisbecker <> | Subject | Re: [PATCH v3 3/4] perf: Remove perf_swevent_get_recursion_context() from perf_pending_task(). |
| |
Le Tue, Apr 09, 2024 at 03:33:36PM +0200, Sebastian Andrzej Siewior a écrit : > On 2024-04-09 14:00:49 [+0200], Frederic Weisbecker wrote: > > Le Tue, Apr 09, 2024 at 12:54:05PM +0200, Sebastian Andrzej Siewior a écrit : > > > On 2024-04-09 12:35:46 [+0200], Frederic Weisbecker wrote: > > > > > > > @@ -6800,10 +6792,6 @@ static void perf_pending_task(struct callback_head *head) > > > > > > > local_dec(&event->ctx->nr_pending); > > > > > > > } > > > > > > > > > > > > > > - if (rctx >= 0) > > > > > > > - perf_swevent_put_recursion_context(rctx); > > > > > > > - preempt_enable_notrace(); > > > > > > > > > > > > Well, if a software event happens during perf_sigtrap(), the task work > > > > > > may be requeued endlessly and the task may get stuck in task_work_run()... > > > > > > > > > > The last time I checked it had no users in the task context. How would > > > > > that happen? > > > > > > > > I guess many tracepoint events would do the trick. Such as trace_lock_acquire() > > > > for example. > > > > > > So the perf_trace_buf_alloc() is invoked from that trace point and > > > avoids the recursion. And any trace event from within perf_sigtrap() > > > would trigger the endless loop? > > > > No sure I'm following: > > > > 1) event->perf_event_overflow() -> task_work_add() > > //return to userspace > > 2) task_work_run() -> perf_pending_task() -> perf_sigtrap() -> tracepoint event > > -> perf_event_overflow() -> task_work_add() > > 3) task_work_run() -> perf_pending_task() -> etc... > > > > What am I missing? > > Yes, that is what I tried to say.
Oh ok.. :o)
> Anyway, I misunderstood the concept > before. That means we need to keep that counter here and a > migrate_disable() is needed to avoid CPU migration which is sad.
I fear that won't work either. The task is then pinned but another task can come up on that CPU and its software events will be ignored...
Some alternatives:
_ Clear event->pending_work = 0 after perf_sigtrap(), preventing an event in there from adding a new task work. We may miss a signal though...
_ Make the recursion context per task on -RT...
> > Thanks. > > Sebastian
| ![\](/images/icornerr.gif) |