[PATCH] perf/core: Fix ITRACE start suppression for inherited events

[PATCH] perf/core: Fix ITRACE start suppression for inherited events

[Leo Yan]

PMU drivers call perf_event_itrace_started() for the event that has
started tracing. This sets PERF_ATTACH_ITRACE in that event's
attach_state.

For inherited events, however, perf_log_itrace_start() replaces the
child event with its parent before checking PERF_ATTACH_ITRACE. The
setter and checker therefore operate on different events. If the
parent's flag is clear, the child continues to emit ITRACE_START
records on subsequent schedule-ins. If the parent has already started,
its flag can instead suppress the child's initial record.

Remove the parent substitution so that perf_log_itrace_start() checks
the same event that the PMU driver marks as started.

This is safe for tool consumers. Intel PT uses the ITRACE_START record
to set the current thread context. CoreSight ETM uses the record only
to find or create the corresponding thread. Neither decoder depends on
the parent event.

Reported-by: Sashiko AI <sashiko-bot@kernel.org>
Closes: https://lore.kernel.org/linux-perf-users/20260901164705.042781F000E9@smtp.kernel.org/
Fixes: 9a6694cfa239 ("perf/x86/intel/pt: Do not force sync packets on every schedule-in")
Signed-off-by: Leo Yan <leo.yan@arm.com>
---
 kernel/events/core.c | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index a6c8e38a311042afab6b65814a84c67b87ba929b..991ae214d46ebe8b0d9de497255f25a4ad12397f 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -10572,9 +10572,6 @@ static void perf_log_itrace_start(struct perf_event *event)
 	} rec;
 	int ret;
 
-	if (event->parent)
-		event = event->parent;
-
 	if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
 	    event->attach_state & PERF_ATTACH_ITRACE)
 		return;

---
base-commit: 940de590b839f71d6dc846160534bf202401b8b7
change-id: 20260903-perf_core_itrace_start_fix_inherit_event-a8930d792899

Best regards,
-- 
Leo Yan <leo.yan@arm.com>

Re: [PATCH] perf/core: Fix ITRACE start suppression for inherited events

[Leo Yan]

On Thu, Sep 03, 2026 at 10:35:12AM +0100, Leo Yan wrote:
> PMU drivers call perf_event_itrace_started() for the event that has
> started tracing. This sets PERF_ATTACH_ITRACE in that event's
> attach_state.
> 
> For inherited events, however, perf_log_itrace_start() replaces the
> child event with its parent before checking PERF_ATTACH_ITRACE. The
> setter and checker therefore operate on different events. If the
> parent's flag is clear, the child continues to emit ITRACE_START
> records on subsequent schedule-ins. If the parent has already started,
> its flag can instead suppress the child's initial record.
> 
> Remove the parent substitution so that perf_log_itrace_start() checks
> the same event that the PMU driver marks as started.
> 
> This is safe for tool consumers. Intel PT uses the ITRACE_START record
> to set the current thread context. CoreSight ETM uses the record only
> to find or create the corresponding thread. Neither decoder depends on
> the parent event.
> 
> Reported-by: Sashiko AI <sashiko-bot@kernel.org>
> Closes: https://lore.kernel.org/linux-perf-users/20260901164705.042781F000E9@smtp.kernel.org/
> Fixes: 9a6694cfa239 ("perf/x86/intel/pt: Do not force sync packets on every schedule-in")
> Signed-off-by: Leo Yan <leo.yan@arm.com>

Gentle ping ...

> ---
>  kernel/events/core.c | 3 ---
>  1 file changed, 3 deletions(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index a6c8e38a311042afab6b65814a84c67b87ba929b..991ae214d46ebe8b0d9de497255f25a4ad12397f 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -10572,9 +10572,6 @@ static void perf_log_itrace_start(struct perf_event *event)
>  	} rec;
>  	int ret;
>  
> -	if (event->parent)
> -		event = event->parent;
> -
>  	if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
>  	    event->attach_state & PERF_ATTACH_ITRACE)
>  		return;
> 
> ---
> base-commit: 940de590b839f71d6dc846160534bf202401b8b7
> change-id: 20260903-perf_core_itrace_start_fix_inherit_event-a8930d792899
> 
> Best regards,
> -- 
> Leo Yan <leo.yan@arm.com>
> 

Re: [PATCH] perf/core: Fix ITRACE start suppression for inherited events

[James Clark]

On 03/09/2026 10:35, Leo Yan wrote:
> PMU drivers call perf_event_itrace_started() for the event that has
> started tracing. This sets PERF_ATTACH_ITRACE in that event's
> attach_state.
> 
> For inherited events, however, perf_log_itrace_start() replaces the
> child event with its parent before checking PERF_ATTACH_ITRACE. The
> setter and checker therefore operate on different events. If the
> parent's flag is clear, the child continues to emit ITRACE_START
> records on subsequent schedule-ins. If the parent has already started,
> its flag can instead suppress the child's initial record.

This last part would need an earlier fixes: commit. The problem of 
suppressing child ITRACE_START records existed since the beginning on 
ec0d772 ("perf: Add ITRACE_START record to indicate that tracing has 
started").

Although the fixes: commit would be correct if the only problem was that 
setting and getting are on different events.

> 
> Remove the parent substitution so that perf_log_itrace_start() checks
> the same event that the PMU driver marks as started.
> 
> This is safe for tool consumers. Intel PT uses the ITRACE_START record
> to set the current thread context. CoreSight ETM uses the record only
> to find or create the corresponding thread. Neither decoder depends on
> the parent event.
> 

It's probably harmless to emit more ITRACE_STARTs, but it doesn't fit 
the original purpose of why it was added. It seems to be for when 
tracing first starts, which would be _after_ the corresponding sched 
event. Once tracing has started you can follow the subsequent sched 
events, so you don't need more ITRACE_START records for each child. 
Could we not fix it by changing the setter to follow the parent event to 
match, which would respect the original meaning?


> Reported-by: Sashiko AI <sashiko-bot@kernel.org>
> Closes: https://lore.kernel.org/linux-perf-users/20260901164705.042781F000E9@smtp.kernel.org/
> Fixes: 9a6694cfa239 ("perf/x86/intel/pt: Do not force sync packets on every schedule-in")
> Signed-off-by: Leo Yan <leo.yan@arm.com>
> ---
>   kernel/events/core.c | 3 ---
>   1 file changed, 3 deletions(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index a6c8e38a311042afab6b65814a84c67b87ba929b..991ae214d46ebe8b0d9de497255f25a4ad12397f 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -10572,9 +10572,6 @@ static void perf_log_itrace_start(struct perf_event *event)
>   	} rec;
>   	int ret;
>   
> -	if (event->parent)
> -		event = event->parent;
> -
>   	if (!(event->pmu->capabilities & PERF_PMU_CAP_ITRACE) ||
>   	    event->attach_state & PERF_ATTACH_ITRACE)
>   		return;
> 
> ---
> base-commit: 940de590b839f71d6dc846160534bf202401b8b7
> change-id: 20260903-perf_core_itrace_start_fix_inherit_event-a8930d792899
> 
> Best regards,