Re: [PATCH v5 1/2] perf/core: Prevent duplicate sigtrap task work enqueue

From: Peter Zijlstra

Date: Fri Nov 20 2026 - 12:50:06 EST


On Fri, Nov 20, 2026 at 06:28:31PM +0100, Sebastian Andrzej Siewior wrote:
> On 2026-11-20 17:26:04 [+0100], Peter Zijlstra wrote:
> > Also, sending signals on software events seems utterly insane. But I
> > don't support we can go limit that now :-(
>
> Tengda: Is this an actual valid use case that you have here or did an AI
> figured out that this broken?

I *think* the below is sufficient.

- sigtrap implies a per-task event; task_work is local to the current
task (+- the context switch thing from 2/2).

- task_work_run() is written such that the work handler can do
task_work_add(), therefore the early clear in the handler doesn't
matter.

And while recursive software events are tedious, I think the below fix
is simple enough to not worry about too much.

---
Subject: perf/core: Prevent duplicate sigtrap task work enqueue
From: Tengda Wu <wutengda@xxxxxxxxxxxxxxx>
Date: Thu, 19 Nov 2026 09:51:00 +0000

From: Tengda Wu <wutengda@xxxxxxxxxxxxxxx>

The sigtrap handling in __perf_event_overflow() checks
event->pending_work before calling task_work_add(), but only sets
pending_work after the task work has been queued.

This leaves a window where an NMI can interrupt task_work_add() and
re-enter __perf_event_overflow() for the same (software/tracepoint)
event. Since pending_work is still zero, the nested invocation can
enqueue the same pending_task again.

The re-entrant call path looks like this:

__perf_event_overflow()
task_work_add()
<NMI>
__perf_event_overflow()
task_work_add()

The nested invocation may enqueue pending_task before the outer
invocation does. When the outer invocation resumes, it inserts the
same task work entry again, potentially creating a cycle in the task
work list.

As a result, perf_pending_task() may be executed repeatedly, leading
to unbalanced event reference counting.

Fix this by atomically claiming pending_work with cmpxchg_local()
before calling task_work_add(). This ensures that only one invocation
can enqueue the task work, even when another overflow occurs
re-entrantly from an NMI.

Notably: sigtrap events must be per-task, meaning that the there is no
concurrency, only re-entrancy.

[peterz: clarify and rewrite]
Fixes: c5d93d23a260 ("perf: Enqueue SIGTRAP always via task_work.")
Signed-off-by: Tengda Wu <wutengda@xxxxxxxxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
kernel/events/core.c | 24 +++++++++++++++---------
1 file changed, 15 insertions(+), 9 deletions(-)

--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -10950,16 +10950,22 @@ static int __perf_event_overflow(struct

notify_mode = in_nmi() ? TWA_NMI_CURRENT : TWA_RESUME;

- if (!event->pending_work &&
- !task_work_add(current, &event->pending_task, notify_mode)) {
- event->pending_work = pending_id;
- local_inc(&event->ctx->nr_no_switch_fast);
- WARN_ON_ONCE(!atomic_long_inc_not_zero(&event->refcount));
-
- event->pending_addr = 0;
- if (valid_sample && (data->sample_flags & PERF_SAMPLE_ADDR))
- event->pending_addr = data->addr;
+ /*
+ * Software events can nest and attempt this recursively;
+ * claim the pending_work before task_work_add() recurses.
+ */
+ if (cmpxchg_local(&event->pending_work, 0, 1) == 0) {
+ if (!task_work_add(current, &event->pending_task, notify_mode)) {
+ event->pending_work = pending_id;
+ local_inc(&event->ctx->nr_no_switch_fast);
+ WARN_ON_ONCE(!atomic_long_inc_not_zero(&event->refcount));

+ event->pending_addr = 0;
+ if (valid_sample && (data->sample_flags & PERF_SAMPLE_ADDR))
+ event->pending_addr = data->addr;
+ } else {
+ event->pending_work = 0;
+ }
} else if (event->attr.exclude_kernel && valid_sample) {
/*
* Should not be able to return to user space without