perf_init_event tries multiple init callbacks and does not reset the
event state between tries. When x86_pmu_event_init runs, it
unconditionally sets the destroy callback to hw_perf_event_destroy. On
the next init attempt after x86_pmu_event_init, in perf_try_init_event,
if the pmu's capabilities includes PERF_PMU_CAP_NO_EXCLUDE, the destroy
callback will be run. However, if the next init didn't set the destroy
callback, hw_perf_event_destroy will be run (since the callback wasn't
reset).
Looking at other pmu init functions, the common pattern is to only set
the destroy callback on a successful init. Resetting the callback on
failure tries to replicate that pattern.
This was discovered after commit f11dd0d80555 ("perf/x86/amd/ibs: Extend
PERF_PMU_CAP_NO_EXCLUDE to IBS Op") when the second (and only second)
run of the perf tool after a reboot results in 0 samples being
generated. The extra run of hw_perf_event_destroy results in
active_events having an extra decrement on each perf run. The second run
has active_events == 0 and every subsequent run has active_events < 0.
When active_events == 0, the NMI handler will early-out and not record
any samples.
Signed-off-by: Anand K Mistry <[email protected]>
---
arch/x86/events/core.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index e6db1a1f22d7..1f5d96ba4866 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -2284,6 +2284,7 @@ static int x86_pmu_event_init(struct perf_event *event)
if (err) {
if (event->destroy)
event->destroy(event);
+ event->destroy = NULL;
}
if (READ_ONCE(x86_pmu.attr_rdpmc) &&
--
2.33.0.800.g4c38ced690-goog
The following commit has been merged into the perf/urgent branch of tip:
Commit-ID: 02d029a41dc986e2d5a77ecca45803857b346829
Gitweb: https://git.kernel.org/tip/02d029a41dc986e2d5a77ecca45803857b346829
Author: Anand K Mistry <[email protected]>
AuthorDate: Wed, 29 Sep 2021 17:04:21 +10:00
Committer: Peter Zijlstra <[email protected]>
CommitterDate: Fri, 01 Oct 2021 13:57:54 +02:00
perf/x86: Reset destroy callback on event init failure
perf_init_event tries multiple init callbacks and does not reset the
event state between tries. When x86_pmu_event_init runs, it
unconditionally sets the destroy callback to hw_perf_event_destroy. On
the next init attempt after x86_pmu_event_init, in perf_try_init_event,
if the pmu's capabilities includes PERF_PMU_CAP_NO_EXCLUDE, the destroy
callback will be run. However, if the next init didn't set the destroy
callback, hw_perf_event_destroy will be run (since the callback wasn't
reset).
Looking at other pmu init functions, the common pattern is to only set
the destroy callback on a successful init. Resetting the callback on
failure tries to replicate that pattern.
This was discovered after commit f11dd0d80555 ("perf/x86/amd/ibs: Extend
PERF_PMU_CAP_NO_EXCLUDE to IBS Op") when the second (and only second)
run of the perf tool after a reboot results in 0 samples being
generated. The extra run of hw_perf_event_destroy results in
active_events having an extra decrement on each perf run. The second run
has active_events == 0 and every subsequent run has active_events < 0.
When active_events == 0, the NMI handler will early-out and not record
any samples.
Signed-off-by: Anand K Mistry <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Link: https://lkml.kernel.org/r/20210929170405.1.I078b98ee7727f9ae9d6df8262bad7e325e40faf0@changeid
---
arch/x86/events/core.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 2a57dbe..6dfa8dd 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -2465,6 +2465,7 @@ static int x86_pmu_event_init(struct perf_event *event)
if (err) {
if (event->destroy)
event->destroy(event);
+ event->destroy = NULL;
}
if (READ_ONCE(x86_pmu.attr_rdpmc) &&