Re: [PATCH] perf: Fix missing SIGTRAPs due to pending_disable abuse

[Peter Zijlstra]

On Wed, Oct 05, 2022 at 09:37:06AM +0200, Peter Zijlstra wrote:
> On Tue, Oct 04, 2022 at 07:33:55PM +0200, Marco Elver wrote:
> > It looks reasonable, but obviously needs to pass tests. :-)
> 
> Ikr :-)
> 
> > Also, see comment below (I think you're still turning signals
> > asynchronous, which we shouldn't do).
> 
> Indeed so; I tried fixing that this morning, but so far that doesn't
> seem to want to actually cure things :/ I'll need to stomp on this
> harder.
> 
> Current hackery below. The main difference is that instead of trying to
> restart the irq_work on sched_in, sched_out will now queue a task-work.
> 
> The event scheduling is done from 'regular' IRQ context and as such
> there should be a return-to-userspace for the relevant task in the
> immediate future (either directly or after scheduling).
> 
> Alas, something still isn't right...

Oh, lol, *groan*... this fixes it:

Now to find a sane way to inhibit this while a sig thing is pending :/

diff --git a/kernel/events/core.c b/kernel/events/core.c
index b981b879bcd8..92b6a2f6de1a 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3426,7 +3434,7 @@ static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
 		 */
 		raw_spin_lock(&ctx->lock);
 		raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
-		if (context_equiv(ctx, next_ctx)) {
+		if (0 && context_equiv(ctx, next_ctx)) {
 
 			WRITE_ONCE(ctx->task, next);
 			WRITE_ONCE(next_ctx->task, task);