perf: Fix missing SIGTRAPs

Marco reported:

Due to the implementation of how SIGTRAP are delivered if
perf_event_attr::sigtrap is set, we've noticed 3 issues:

  1. Missing SIGTRAP due to a race with event_sched_out() (more
     details below).

  2. Hardware PMU events being disabled due to returning 1 from
     perf_event_overflow(). The only way to re-enable the event is
     for user space to first "properly" disable the event and then
     re-enable it.

  3. The inability to automatically disable an event after a
     specified number of overflows via PERF_EVENT_IOC_REFRESH.

The worst of the 3 issues is problem (1), which occurs when a
pending_disable is "consumed" by a racing event_sched_out(), observed
as follows:

		CPU0			|	CPU1
	--------------------------------+---------------------------
	__perf_event_overflow()		|
	 perf_event_disable_inatomic()	|
	  pending_disable = CPU0	| ...
					| _perf_event_enable()
					|  event_function_call()
					|   task_function_call()
					|    /* sends IPI to CPU0 */
	<IPI>				| ...
	 __perf_event_enable()		+---------------------------
	  ctx_resched()
	   task_ctx_sched_out()
	    ctx_sched_out()
	     group_sched_out()
	      event_sched_out()
	       pending_disable = -1
	</IPI>
	<IRQ-work>
	 perf_pending_event()
	  perf_pending_event_disable()
	   /* Fails to send SIGTRAP because no pending_disable! */
	</IRQ-work>

In the above case, not only is that particular SIGTRAP missed, but also
all future SIGTRAPs because 'event_limit' is not reset back to 1.

To fix, rework pending delivery of SIGTRAP via IRQ-work by introduction
of a separate 'pending_sigtrap', no longer using 'event_limit' and
'pending_disable' for its delivery.

Additionally; and different to Marco's proposed patch:

 - recognise that pending_disable effectively duplicates oncpu for
   the case where it is set. As such, change the irq_work handler to
   use ->oncpu to target the event and use pending_* as boolean toggles.

 - observe that SIGTRAP targets the ctx->task, so the context switch
   optimization that carries contexts between tasks is invalid. If
   the irq_work were delayed enough to hit after a context switch the
   SIGTRAP would be delivered to the wrong task.

 - observe that if the event gets scheduled out
   (rotation/migration/context-switch/...) the irq-work would be
   insufficient to deliver the SIGTRAP when the event gets scheduled
   back in (the irq-work might still be pending on the old CPU).

   Therefore have event_sched_out() convert the pending sigtrap into a
   task_work which will deliver the signal at return_to_user.

Fixes: 97ba62b2

 ("perf: Add support for SIGTRAP on perf events")
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Debugged-by: Dmitry Vyukov <dvyukov@google.com>
Reported-by: Marco Elver <elver@google.com>
Debugged-by: Marco Elver <elver@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Marco Elver <elver@google.com>
Tested-by: Marco Elver <elver@google.com>