Re: [BUG] perf/core: Task stuck on global_ctx_data_rwsem

[Namhyung Kim <namhyung@kernel.org>]

On Wed, Jan 07, 2026 at 11:32:56PM +0100, Peter Zijlstra wrote:
> On Wed, Jan 07, 2026 at 11:28:24PM +0100, Peter Zijlstra wrote:
> > On Wed, Jan 07, 2026 at 11:01:53AM -0800, Namhyung Kim wrote:
> > 
> > > > But yes, I suppose this can do. The question is however, how do you get
> > > > into this predicament to begin with? Are you creating and destroying a
> > > > lot of global LBR events or something?
> > > 
> > > I think it's just because there are too many tasks in the system like
> > > O(100K).  And any thread going to exit needs to wait for
> > > attach_global_ctx_data() to finish the iteration over every task.
> > 
> > OMG, so many tasks ...
> > 
> > > > Would it make sense to delay detach_global_ctx_data() for a second or
> > > > so? That is, what is your event creation pattern?
> > > 
> > > I don't think it has a special pattern, but I'm curious how we can
> > > handle a race like below.
> > > 
> > >   attach_global_ctx_data
> > >     check p->flags & PF_EXITING
> > >                                               do_exit
> > >     (preemption)                                set PF_EXITING
> > >                                                 detach_task_ctx_data()
> > >     check p->perf_ctx_data
> > >     attach_task_ctx_data()   ---> memory leak
> > 
> > Oh right. Something like so perhaps?
> > 
> > ---
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 3c2a491200c6..e5e716420eb3 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -5421,9 +5421,19 @@ attach_task_ctx_data(struct task_struct *task, struct kmem_cache *ctx_cache,
> >  		return -ENOMEM;
> >  
> >  	for (;;) {
> > -		if (try_cmpxchg((struct perf_ctx_data **)&task->perf_ctx_data, &old, cd)) {
> > +		if (try_cmpxchg(&task->perf_ctx_data, &old, cd)) {
> >  			if (old)
> >  				perf_free_ctx_data_rcu(old);
> > +			/*
> > +			 * try_cmpxchg() pairs with try_cmpxchg() from
> > +			 * detach_task_ctx_data() such that
> > +			 * if we race with perf_event_exit_task(), we must
> > +			 * observe PF_EXITING.
> > +			 */
> > +			if (task->flags & PF_EXITING) {
> > +				task->perf_ctx_data = NULL;
> > +				perf_free_ctx_data_rcu(cd);
> 
> Ugh and now it can race and do a double free, another try_cmpxchg() is
> needed here.

Thanks!  Something like this?

Namhyung


diff --git a/kernel/events/core.c b/kernel/events/core.c
index 376fb07d869b8b50..cf252d8f49b2b259 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -5421,9 +5421,20 @@ attach_task_ctx_data(struct task_struct *task, struct kmem_cache *ctx_cache,
 		return -ENOMEM;
 
 	for (;;) {
-		if (try_cmpxchg((struct perf_ctx_data **)&task->perf_ctx_data, &old, cd)) {
+		if (try_cmpxchg(&task->perf_ctx_data, &old, cd)) {
 			if (old)
 				perf_free_ctx_data_rcu(old);
+			/*
+			 * try_cmpxchg() pairs with try_cmpxchg() from
+			 * detach_task_ctx_data() such that
+			 * if we race with perf_event_exit_task(), we must
+			 * observe PF_EXITING.
+			 */
+			if (task->flags & PF_EXITING) {
+				/* detach_task_ctx_data() may free it already */
+				if (try_cmpxchg(&task->perf_ctx_data, &cd, NULL))
+					perf_free_ctx_data_rcu(cd);
+			}
 			return 0;
 		}
 
@@ -5469,6 +5480,8 @@ attach_global_ctx_data(struct kmem_cache *ctx_cache)
 	/* Allocate everything */
 	scoped_guard (rcu) {
 		for_each_process_thread(g, p) {
+			if (p->flags & PF_EXITING)
+				continue;
 			cd = rcu_dereference(p->perf_ctx_data);
 			if (cd && !cd->global) {
 				cd->global = 1;
@@ -14562,8 +14575,11 @@ void perf_event_exit_task(struct task_struct *task)
 
 	/*
 	 * Detach the perf_ctx_data for the system-wide event.
+	 *
+	 * Done without holding global_ctx_data_rwsem; typically
+	 * attach_global_ctx_data() will skip over this task, but otherwise
+	 * attach_task_ctx_data() will observe PF_EXITING.
 	 */
-	guard(percpu_read)(&global_ctx_data_rwsem);
 	detach_task_ctx_data(task);
 }