Re: [PATCH v3 7/7] perf: Make perf_pmu_unregister() useable

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Apr 14, 2025 at 08:37:07AM +0800, Mi, Dapeng wrote:

> It seems this patch would break the task attached events counting like the
> below command shows.
> 

Right, found another report for that yesterday.

---
Subject: perf: Fix perf-stat / read()
From: Peter Zijlstra <peterz@infradead.org>
Date: Wed Apr 16 20:50:27 CEST 2025

In the zeal to adjust all event->state checks to include the new
REVOKED state, one adjustment was made in error. Notably it resulted
in read() on the perf filedesc to stop working for any state lower
than ERROR, specifically EXIT.

This leads to problems with (among others) perf-stat, which wants to
read the counts after a program has finished execution.

Fixes: da916e96e2de ("perf: Make perf_pmu_unregister() useable")
Reported-by: "Mi, Dapeng" <dapeng1.mi@linux.intel.com>
Reported-by: James Clark <james.clark@linaro.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/77036114-8723-4af9-a068-1d535f4e2e81@linaro.org
---
 kernel/events/core.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -6023,7 +6023,7 @@ __perf_read(struct perf_event *event, ch
 	 * error state (i.e. because it was pinned but it couldn't be
 	 * scheduled on to the CPU at some point).
 	 */
-	if (event->state <= PERF_EVENT_STATE_ERROR)
+	if (event->state == PERF_EVENT_STATE_ERROR)
 		return 0;
 
 	if (count < event->read_size)