[PATCH v3] perf/core: Fix missing read event generation on task exit

Thaumy Cheng posted 1 patch 3 months, 2 weeks ago
There is a newer version of this series
kernel/events/core.c | 23 ++++++++++++++---------
1 file changed, 14 insertions(+), 9 deletions(-)
[PATCH v3] perf/core: Fix missing read event generation on task exit
Posted by Thaumy Cheng 3 months, 2 weeks ago
For events with inherit_stat enabled, a "read" event will be generated
to collect per task event counts on task exit.

The call chain is as follows:

do_exit
  -> perf_event_exit_task
    -> perf_event_exit_task_context
      -> perf_event_exit_event
        -> perf_remove_from_context
          -> perf_child_detach
            -> sync_child_event
              -> perf_event_read_event

However, the child event context detaches the task too early in
perf_event_exit_task_context, which causes sync_child_event to never
generate the read event in this case, since child_event->ctx->task is
always set to TASK_TOMBSTONE. Fix that by moving context lock section
backward to ensure ctx->task is not set to TASK_TOMBSTONE before
generating the read event.

Because perf_event_free_task calls perf_event_exit_task_context with
exit = false to tear down all child events from the context, and the
task never lived, accessing the task PID can lead to a use-after-free.

To fix that, let sync_child_event read task from argument and move the
call to the only place it should be triggered to avoid the effect of
setting ctx->task to TASK_TOMESTONE, and add a task parameter to
perf_event_exit_event to trigger the sync_child_event properly when
needed.

This bug can be reproduced by running "perf record -s" and attaching to
any program that generates perf events in its child tasks. If we check
the result with "perf report -T", the last line of the report will leave
an empty table like "# PID  TID", which is expected to contain the
per-task event counts by design.

Fixes: ef54c1a476ae ("perf: Rework perf_event_exit_event()")
Signed-off-by: Thaumy Cheng <thaumy.love@gmail.com>
---
Changes in v3:
- Fix the bug in a more direct way by moving the call to
  sync_child_event and bring back the task param to
  perf_event_exit_event.
  This approach avoids the event unscheduling issue in v2.

Changes in v2:
- Only trigger read event on task exit.
- Rename perf_event_exit_event to perf_event_detach_event.
- Link to v2: https://lore.kernel.org/all/20250817132742.85154-1-thaumy.love@gmail.com/

Changes in v1:
- Set TASK_TOMBSTONE after the read event is tirggered.
- Link to v1: https://lore.kernel.org/all/20250720000424.12572-1-thaumy.love@gmail.com/

 kernel/events/core.c | 23 ++++++++++++++---------
 1 file changed, 14 insertions(+), 9 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 177e57c1a362..618e7947c358 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -2316,7 +2316,8 @@ static void perf_group_detach(struct perf_event *event)
 	perf_event__header_size(leader);
 }

-static void sync_child_event(struct perf_event *child_event);
+static void sync_child_event(struct perf_event *child_event,
+			     struct task_struct *task);

 static void perf_child_detach(struct perf_event *event)
 {
@@ -2336,7 +2337,6 @@ static void perf_child_detach(struct perf_event *event)
 	lockdep_assert_held(&parent_event->child_mutex);
 	 */

-	sync_child_event(event);
 	list_del_init(&event->child_list);
 }

@@ -4587,6 +4587,7 @@ static void perf_event_enable_on_exec(struct perf_event_context *ctx)
 static void perf_remove_from_owner(struct perf_event *event);
 static void perf_event_exit_event(struct perf_event *event,
 				  struct perf_event_context *ctx,
+				  struct task_struct *task,
 				  bool revoke);

 /*
@@ -4614,7 +4615,7 @@ static void perf_event_remove_on_exec(struct perf_event_context *ctx)

 		modified = true;

-		perf_event_exit_event(event, ctx, false);
+		perf_event_exit_event(event, ctx, ctx->task, false);
 	}

 	raw_spin_lock_irqsave(&ctx->lock, flags);
@@ -12437,7 +12438,7 @@ static void __pmu_detach_event(struct pmu *pmu, struct perf_event *event,
 	/*
 	 * De-schedule the event and mark it REVOKED.
 	 */
-	perf_event_exit_event(event, ctx, true);
+	perf_event_exit_event(event, ctx, ctx->task, true);

 	/*
 	 * All _free_event() bits that rely on event->pmu:
@@ -13994,14 +13995,13 @@ void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
 }
 EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);

-static void sync_child_event(struct perf_event *child_event)
+static void sync_child_event(struct perf_event *child_event,
+			     struct task_struct *task)
 {
 	struct perf_event *parent_event = child_event->parent;
 	u64 child_val;

 	if (child_event->attr.inherit_stat) {
-		struct task_struct *task = child_event->ctx->task;
-
 		if (task && task != TASK_TOMBSTONE)
 			perf_event_read_event(child_event, task);
 	}
@@ -14020,7 +14020,9 @@ static void sync_child_event(struct perf_event *child_event)

 static void
 perf_event_exit_event(struct perf_event *event,
-		      struct perf_event_context *ctx, bool revoke)
+		      struct perf_event_context *ctx,
+		      struct task_struct *task,
+		      bool revoke)
 {
 	struct perf_event *parent_event = event->parent;
 	unsigned long detach_flags = DETACH_EXIT;
@@ -14043,6 +14045,9 @@ perf_event_exit_event(struct perf_event *event,
 		mutex_lock(&parent_event->child_mutex);
 		/* PERF_ATTACH_ITRACE might be set concurrently */
 		attach_state = READ_ONCE(event->attach_state);
+
+		if (attach_state & PERF_ATTACH_CHILD)
+			sync_child_event(event, task);
 	}

 	if (revoke)
@@ -14134,7 +14139,7 @@ static void perf_event_exit_task_context(struct task_struct *task, bool exit)
 		perf_event_task(task, ctx, 0);

 	list_for_each_entry_safe(child_event, next, &ctx->event_list, event_entry)
-		perf_event_exit_event(child_event, ctx, false);
+		perf_event_exit_event(child_event, ctx, exit ? task : NULL, false);

 	mutex_unlock(&ctx->mutex);

--
2.51.0
Re: [PATCH v3] perf/core: Fix missing read event generation on task exit
Posted by James Clark 1 day, 2 hours ago

On 24/10/2025 6:05 pm, Thaumy Cheng wrote:
> For events with inherit_stat enabled, a "read" event will be generated
> to collect per task event counts on task exit.
> 
> The call chain is as follows:
> 
> do_exit
>    -> perf_event_exit_task
>      -> perf_event_exit_task_context
>        -> perf_event_exit_event
>          -> perf_remove_from_context
>            -> perf_child_detach
>              -> sync_child_event
>                -> perf_event_read_event
> 
> However, the child event context detaches the task too early in
> perf_event_exit_task_context, which causes sync_child_event to never
> generate the read event in this case, since child_event->ctx->task is
> always set to TASK_TOMBSTONE. Fix that by moving context lock section
> backward to ensure ctx->task is not set to TASK_TOMBSTONE before
> generating the read event.
> 
> Because perf_event_free_task calls perf_event_exit_task_context with
> exit = false to tear down all child events from the context, and the
> task never lived, accessing the task PID can lead to a use-after-free.
> 
> To fix that, let sync_child_event read task from argument and move the
> call to the only place it should be triggered to avoid the effect of
> setting ctx->task to TASK_TOMESTONE, and add a task parameter to
> perf_event_exit_event to trigger the sync_child_event properly when
> needed.
> 
> This bug can be reproduced by running "perf record -s" and attaching to
> any program that generates perf events in its child tasks. If we check
> the result with "perf report -T", the last line of the report will leave
> an empty table like "# PID  TID", which is expected to contain the
> per-task event counts by design.
> 
> Fixes: ef54c1a476ae ("perf: Rework perf_event_exit_event()")
> Signed-off-by: Thaumy Cheng <thaumy.love@gmail.com>
> ---
> Changes in v3:
> - Fix the bug in a more direct way by moving the call to
>    sync_child_event and bring back the task param to
>    perf_event_exit_event.
>    This approach avoids the event unscheduling issue in v2.
> 
> Changes in v2:
> - Only trigger read event on task exit.
> - Rename perf_event_exit_event to perf_event_detach_event.
> - Link to v2: https://lore.kernel.org/all/20250817132742.85154-1-thaumy.love@gmail.com/
> 
> Changes in v1:
> - Set TASK_TOMBSTONE after the read event is tirggered.
> - Link to v1: https://lore.kernel.org/all/20250720000424.12572-1-thaumy.love@gmail.com/
> 
>   kernel/events/core.c | 23 ++++++++++++++---------
>   1 file changed, 14 insertions(+), 9 deletions(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 177e57c1a362..618e7947c358 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -2316,7 +2316,8 @@ static void perf_group_detach(struct perf_event *event)
>   	perf_event__header_size(leader);
>   }
> 
> -static void sync_child_event(struct perf_event *child_event);
> +static void sync_child_event(struct perf_event *child_event,
> +			     struct task_struct *task);
> 
>   static void perf_child_detach(struct perf_event *event)
>   {
> @@ -2336,7 +2337,6 @@ static void perf_child_detach(struct perf_event *event)
>   	lockdep_assert_held(&parent_event->child_mutex);
>   	 */
> 
> -	sync_child_event(event);
>   	list_del_init(&event->child_list);
>   }
> 
> @@ -4587,6 +4587,7 @@ static void perf_event_enable_on_exec(struct perf_event_context *ctx)
>   static void perf_remove_from_owner(struct perf_event *event);
>   static void perf_event_exit_event(struct perf_event *event,
>   				  struct perf_event_context *ctx,
> +				  struct task_struct *task,
>   				  bool revoke);
> 
>   /*
> @@ -4614,7 +4615,7 @@ static void perf_event_remove_on_exec(struct perf_event_context *ctx)
> 
>   		modified = true;
> 
> -		perf_event_exit_event(event, ctx, false);
> +		perf_event_exit_event(event, ctx, ctx->task, false);
>   	}
> 
>   	raw_spin_lock_irqsave(&ctx->lock, flags);
> @@ -12437,7 +12438,7 @@ static void __pmu_detach_event(struct pmu *pmu, struct perf_event *event,
>   	/*
>   	 * De-schedule the event and mark it REVOKED.
>   	 */
> -	perf_event_exit_event(event, ctx, true);
> +	perf_event_exit_event(event, ctx, ctx->task, true);
> 
>   	/*
>   	 * All _free_event() bits that rely on event->pmu:
> @@ -13994,14 +13995,13 @@ void perf_pmu_migrate_context(struct pmu *pmu, int src_cpu, int dst_cpu)
>   }
>   EXPORT_SYMBOL_GPL(perf_pmu_migrate_context);
> 
> -static void sync_child_event(struct perf_event *child_event)
> +static void sync_child_event(struct perf_event *child_event,
> +			     struct task_struct *task)
>   {
>   	struct perf_event *parent_event = child_event->parent;
>   	u64 child_val;
> 
>   	if (child_event->attr.inherit_stat) {
> -		struct task_struct *task = child_event->ctx->task;
> -
>   		if (task && task != TASK_TOMBSTONE)
>   			perf_event_read_event(child_event, task);
>   	}
> @@ -14020,7 +14020,9 @@ static void sync_child_event(struct perf_event *child_event)
> 
>   static void
>   perf_event_exit_event(struct perf_event *event,
> -		      struct perf_event_context *ctx, bool revoke)
> +		      struct perf_event_context *ctx,
> +		      struct task_struct *task,
> +		      bool revoke)
>   {
>   	struct perf_event *parent_event = event->parent;
>   	unsigned long detach_flags = DETACH_EXIT;
> @@ -14043,6 +14045,9 @@ perf_event_exit_event(struct perf_event *event,
>   		mutex_lock(&parent_event->child_mutex);
>   		/* PERF_ATTACH_ITRACE might be set concurrently */
>   		attach_state = READ_ONCE(event->attach_state);
> +
> +		if (attach_state & PERF_ATTACH_CHILD)
> +			sync_child_event(event, task);

Hi Thaumy and Peter,

I've been looking into a regression caused by this commit and didn't 
manage to come up with a fix. But shouldn't this be something more like:

   if (attach_state & PERF_ATTACH_CHILD && event_filter_match(event))
       sync_child_event(event, task);

As in, you only want to call sync_child_event() and write stuff to the 
ring buffer for the CPU that is currently running this exit handler? 
Although this change affects the 'total_time_enabled' tracking as well, 
but I'm not 100% sure if we're not double counting it anyway.

 From perf_event_exit_task_context(), perf_event_exit_event() is called 
on all events, which includes events on other CPUs:

   list_for_each_entry_safe(child_event, next, &ctx->event_list, ...)
     perf_event_exit_event(child_event, ctx, exit ? task : NULL, false);

Then we write into those other CPU's ring buffers, which don't support 
concurrency.

The reason I found this is because we have a tracing test that spawns 
some threads and then looks for PERF_RECORD_AUX events. When there are 
concurrent writes into the ring buffers, rb->nest tracking gets messed 
up leaving the count positive even after all nested writers have 
finished. Then all future writes don't copy the data_head pointer to the 
user page (because it thinks someone else is writing), so Perf doesn't 
copy out any data anymore leaving records missing.

An easy reproducer is to put a warning that the ring buffer being 
written to is the correct one:

   @@ -41,10 +41,11 @@ static void perf_output_get_handle(struct
   perf_output_handle *handle)
   {
  	struct perf_buffer *rb = handle->rb;

  	preempt_disable();

   +	WARN_ON(handle->event->cpu != smp_processor_id());


And then record:

   perf record -s -- stress -c 8 -t 1

Which results in:

   perf_output_begin+0x320/0x480 (P)
   perf_event_exit_event+0x178/0x2c0
   perf_event_exit_task_context+0x214/0x2f0
   perf_event_exit_task+0xb0/0x3b0
   do_exit+0x1bc/0x808
   __arm64_sys_exit+0x28/0x30
   invoke_syscall+0x4c/0xe8
   el0_svc_common+0x9c/0xf0
   do_el0_svc+0x28/0x40
   el0_svc+0x50/0x240
   el0t_64_sync_handler+0x78/0x130
   el0t_64_sync+0x198/0x1a0

I suppose there is a chance that this is only an issue when also doing 
perf_aux_output_begin()/perf_aux_output_end() from start/stop because 
that's where I saw the real race? Maybe without that, accessing the rb 
from another CPU is ok because there is some locking, but I think this 
might be a more general issue.

Thanks
James


>   	}
> 
>   	if (revoke)
> @@ -14134,7 +14139,7 @@ static void perf_event_exit_task_context(struct task_struct *task, bool exit)
>   		perf_event_task(task, ctx, 0);
> 
>   	list_for_each_entry_safe(child_event, next, &ctx->event_list, event_entry)
> -		perf_event_exit_event(child_event, ctx, false);
> +		perf_event_exit_event(child_event, ctx, exit ? task : NULL, false);
> 
>   	mutex_unlock(&ctx->mutex);
> 
> --
> 2.51.0
>
Re: [PATCH v3] perf/core: Fix missing read event generation on task exit
Posted by Peter Zijlstra 22 hours ago
On Fri, Feb 06, 2026 at 11:21:19AM +0000, James Clark wrote:

> I've been looking into a regression caused by this commit and didn't manage
> to come up with a fix. But shouldn't this be something more like:
> 
>   if (attach_state & PERF_ATTACH_CHILD && event_filter_match(event))
>       sync_child_event(event, task);
> 
> As in, you only want to call sync_child_event() and write stuff to the ring
> buffer for the CPU that is currently running this exit handler? Although
> this change affects the 'total_time_enabled' tracking as well, but I'm not
> 100% sure if we're not double counting it anyway.
> 
> From perf_event_exit_task_context(), perf_event_exit_event() is called on
> all events, which includes events on other CPUs:
> 
>   list_for_each_entry_safe(child_event, next, &ctx->event_list, ...)
>     perf_event_exit_event(child_event, ctx, exit ? task : NULL, false);
> 
> Then we write into those other CPU's ring buffers, which don't support
> concurrency.
> 
> The reason I found this is because we have a tracing test that spawns some
> threads and then looks for PERF_RECORD_AUX events. When there are concurrent
> writes into the ring buffers, rb->nest tracking gets messed up leaving the
> count positive even after all nested writers have finished. Then all future
> writes don't copy the data_head pointer to the user page (because it thinks
> someone else is writing), so Perf doesn't copy out any data anymore leaving
> records missing.
> 
> An easy reproducer is to put a warning that the ring buffer being written to
> is the correct one:
> 
>   @@ -41,10 +41,11 @@ static void perf_output_get_handle(struct
>   perf_output_handle *handle)
>   {
>  	struct perf_buffer *rb = handle->rb;
> 
>  	preempt_disable();
> 
>   +	WARN_ON(handle->event->cpu != smp_processor_id());
> 
> 
> And then record:
> 
>   perf record -s -- stress -c 8 -t 1
> 
> Which results in:
> 
>   perf_output_begin+0x320/0x480 (P)
>   perf_event_exit_event+0x178/0x2c0
>   perf_event_exit_task_context+0x214/0x2f0
>   perf_event_exit_task+0xb0/0x3b0
>   do_exit+0x1bc/0x808
>   __arm64_sys_exit+0x28/0x30
>   invoke_syscall+0x4c/0xe8
>   el0_svc_common+0x9c/0xf0
>   do_el0_svc+0x28/0x40
>   el0_svc+0x50/0x240
>   el0t_64_sync_handler+0x78/0x130
>   el0t_64_sync+0x198/0x1a0
> 
> I suppose there is a chance that this is only an issue when also doing
> perf_aux_output_begin()/perf_aux_output_end() from start/stop because that's
> where I saw the real race? Maybe without that, accessing the rb from another
> CPU is ok because there is some locking, but I think this might be a more
> general issue.

I *think* something like so.

Before the patch in question this would never happen, because of calling
things too late and always hitting that TASK_TOMBSTONE.

But irrespective of emitting that event, we do want to propagate the
count and runtime numbers.


---
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 5b5cb620499e..f566ad55b4fb 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -14086,7 +14086,7 @@ static void sync_child_event(struct perf_event *child_event,
 	u64 child_val;
 
 	if (child_event->attr.inherit_stat) {
-		if (task && task != TASK_TOMBSTONE)
+		if (task && task != TASK_TOMBSTONE && event_filter_match(child_event))
 			perf_event_read_event(child_event, task);
 	}