Hello
I noticed in the 3.16-rc kernels that one of my regression tests was
failing, specifically
https://github.com/deater/perf_event_tests/blob/master/tests/bugs/read_group_attached.c
This test creates an event group with two events and PERF_FORMAT_GROUP
set.
It creates a child process, stops it, attaches the event group to it,
starts the events and the child, waits for the child to finish, stops the
events, then reads the value.
In theory a subsequent read from the group leader should return values for
*both* events (due to PERF_FORMAT_GROUP) but this stopped working, it only
returns the value for the leader.
I've bisected it down to change: 15a2d4de0eab5
"perf: Always destroy groups on exit"
So is this expected behavior due to this patch, or is it a regression?
Thanks,
Vince
On Tue, Jul 15, 2014 at 10:36:33AM -0400, Vince Weaver wrote:
> Hello
>
> I noticed in the 3.16-rc kernels that one of my regression tests was
> failing, specifically
> https://github.com/deater/perf_event_tests/blob/master/tests/bugs/read_group_attached.c
>
> This test creates an event group with two events and PERF_FORMAT_GROUP
> set.
>
> It creates a child process, stops it, attaches the event group to it,
> starts the events and the child, waits for the child to finish, stops the
> events, then reads the value.
>
> In theory a subsequent read from the group leader should return values for
> *both* events (due to PERF_FORMAT_GROUP) but this stopped working, it only
> returns the value for the leader.
>
> I've bisected it down to change: 15a2d4de0eab5
> "perf: Always destroy groups on exit"
>
> So is this expected behavior due to this patch, or is it a regression?
Regression, lemme try and wrap my head around it. Thanks!
Subject: perf: Revert ("perf: Always destroy groups on exit")
From: Peter Zijlstra <[email protected]>
Date: Tue Jul 15 17:27:27 CEST 2014
Vince reported that commit 15a2d4de0eab5 ("perf: Always destroy groups
on exit") causes a regression with grouped events. In particular his
read_group_attached.c test fails.
https://github.com/deater/perf_event_tests/blob/master/tests/bugs/read_group_attached.c
Because of the context switch optimization in
perf_event_context_sched_out() the 'original' event may end up in the
child process and when that exits the change in the patch in question
destroys the actual grouping.
Therefore revert that change and only destroy inherited groups.
Reported-by: Vince Weaver <[email protected]>
Signed-off-by: Peter Zijlstra <[email protected]>
---
kernel/events/core.c | 14 +++++++++++++-
1 file changed, 13 insertions(+), 1 deletion(-)
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -7458,7 +7458,19 @@ __perf_event_exit_task(struct perf_event
struct perf_event_context *child_ctx,
struct task_struct *child)
{
- perf_remove_from_context(child_event, true);
+ /*
+ * Do not destroy the 'original' grouping; because of the context
+ * switch optimization the original events could've ended up in a
+ * random child task.
+ *
+ * If we were to destroy the original group, all group related
+ * operations would cease to function properly after this random
+ * child dies.
+ *
+ * Do destroy all inherited groups, we don't care about those
+ * and being thorough is better.
+ */
+ perf_remove_from_context(child_event, !!child_event->parent);
/*
* It can happen that the parent exits first, and has events