[PATCH] perf: update perf_cgroup time for ancestor cgroup(s)

[PATCH] perf: update perf_cgroup time for ancestor cgroup(s)

[Song Liu]

When a perf_event is attached to parent cgroup, it should count events
for all children cgroups:

 parent_group   <---- perf_event
   \
    - child_group  <---- process(es)

However, in our tests, we found this perf_event cannot report reliable
results. This is because perf_event->cgrp and cpuctx->cgrp are not
identical, thus perf_event->cgrp are not updated properly.

This patch fixes this by updating perf_cgroup properly for ancestor
cgroup(s).

Signed-off-by: Song Liu <songliubraving@fb.com>
Reported-by: Ephraim Park <ephiepark@fb.com>
---
 kernel/events/core.c | 68 +++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 67 insertions(+), 1 deletion(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 5789810..623d38f 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -722,11 +722,48 @@ static inline void __update_cgrp_time(struct perf_cgroup *cgrp)
 	info->timestamp = now;
 }
 
+static inline void
+update_ancestor_cgroup_time(struct perf_event *event,
+			    struct perf_cgroup *cgrp,
+			    struct perf_cgroup_info *cgrp_info)
+{
+	if (!is_cgroup_event(event))
+		return;
+
+	if (cgroup_is_descendant(cgrp->css.cgroup,
+				 event->cgrp->css.cgroup) &&
+	    event->cgrp != cgrp) {
+		struct perf_cgroup_info *info;
+
+		info = this_cpu_ptr(event->cgrp->info);
+		info->time += cgrp_info->timestamp - info->timestamp;
+		info->timestamp = cgrp_info->timestamp;
+	}
+}
+
 static inline void update_cgrp_time_from_cpuctx(struct perf_cpu_context *cpuctx)
 {
 	struct perf_cgroup *cgrp_out = cpuctx->cgrp;
-	if (cgrp_out)
+
+	if (cgrp_out) {
 		__update_cgrp_time(cgrp_out);
+
+		/* update time for ancestor cgroups */
+		if (cgrp_out->css.cgroup->level > 1) {
+			struct perf_cgroup_info *info
+				= this_cpu_ptr(cgrp_out->info);
+			struct perf_event *event;
+
+			list_for_each_entry(event, &cpuctx->ctx.pinned_groups,
+					    group_entry)
+				update_ancestor_cgroup_time(event, cgrp_out,
+							    info);
+			list_for_each_entry(event, &cpuctx->ctx.flexible_groups,
+					    group_entry)
+				update_ancestor_cgroup_time(event, cgrp_out,
+							    info);
+		}
+	}
 }
 
 static inline void update_cgrp_time_from_event(struct perf_event *event)
@@ -749,6 +786,24 @@ static inline void update_cgrp_time_from_event(struct perf_event *event)
 }
 
 static inline void
+perf_ancestor_cgroup_set_timestamp(struct perf_event *event,
+				   struct perf_cgroup *cgrp,
+				   struct perf_cgroup_info *cgrp_info)
+{
+	if (!is_cgroup_event(event))
+		return;
+
+	if (cgroup_is_descendant(cgrp->css.cgroup,
+				 event->cgrp->css.cgroup) &&
+	    event->cgrp != cgrp) {
+		struct perf_cgroup_info *info
+			= this_cpu_ptr(event->cgrp->info);
+
+		info->timestamp = cgrp_info->timestamp;
+	}
+}
+
+static inline void
 perf_cgroup_set_timestamp(struct task_struct *task,
 			  struct perf_event_context *ctx)
 {
@@ -766,6 +821,17 @@ perf_cgroup_set_timestamp(struct task_struct *task,
 	cgrp = perf_cgroup_from_task(task, ctx);
 	info = this_cpu_ptr(cgrp->info);
 	info->timestamp = ctx->timestamp;
+
+	/* set timestamp for ancestor cgroups */
+	if (cgrp->css.cgroup->level > 1) {
+		struct perf_cgroup_info *info = this_cpu_ptr(cgrp->info);
+		struct perf_event *event;
+
+		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
+			perf_ancestor_cgroup_set_timestamp(event, cgrp, info);
+		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
+			perf_ancestor_cgroup_set_timestamp(event, cgrp, info);
+	}
 }
 
 static DEFINE_PER_CPU(struct list_head, cgrp_cpuctx_list);
-- 
2.9.5

Re: [PATCH] perf: update perf_cgroup time for ancestor cgroup(s)

[Peter Zijlstra]

On Sun, Mar 11, 2018 at 10:32:44PM -0700, Song Liu wrote:
> When a perf_event is attached to parent cgroup, it should count events
> for all children cgroups:
> 
>  parent_group   <---- perf_event
>    \
>     - child_group  <---- process(es)
> 
> However, in our tests, we found this perf_event cannot report reliable
> results. This is because perf_event->cgrp and cpuctx->cgrp are not
> identical, thus perf_event->cgrp are not updated properly.

It might help now and in for our older selves, if you could provide a
simple reproducer for this.

> Signed-off-by: Song Liu <songliubraving@fb.com>
> Reported-by: Ephraim Park <ephiepark@fb.com>
> ---
>  kernel/events/core.c | 68 +++++++++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 67 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 5789810..623d38f 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c

> @@ -766,6 +821,17 @@ perf_cgroup_set_timestamp(struct task_struct *task,
>  	cgrp = perf_cgroup_from_task(task, ctx);
>  	info = this_cpu_ptr(cgrp->info);
>  	info->timestamp = ctx->timestamp;
> +
> +	/* set timestamp for ancestor cgroups */
> +	if (cgrp->css.cgroup->level > 1) {
> +		struct perf_cgroup_info *info = this_cpu_ptr(cgrp->info);
> +		struct perf_event *event;
> +
> +		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
> +			perf_ancestor_cgroup_set_timestamp(event, cgrp, info);
> +		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
> +			perf_ancestor_cgroup_set_timestamp(event, cgrp, info);
> +	}
>  }

That doesn't make any kind of sense... if we're interested in ancestor
groups, then WTH are you iterating _events_ ?

I'm expecting something like:

	struct cgroup_subsys_state *css;

	for (css = cgrp->css; css; css = css->parent) {
		/* fudge time */
	}

Re: [PATCH] perf: update perf_cgroup time for ancestor cgroup(s)

[Song Liu]

> On Mar 12, 2018, at 5:38 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> 
> On Sun, Mar 11, 2018 at 10:32:44PM -0700, Song Liu wrote:
>> When a perf_event is attached to parent cgroup, it should count events
>> for all children cgroups:
>> 
>> parent_group   <---- perf_event
>>   \
>>    - child_group  <---- process(es)
>> 
>> However, in our tests, we found this perf_event cannot report reliable
>> results. This is because perf_event->cgrp and cpuctx->cgrp are not
>> identical, thus perf_event->cgrp are not updated properly.
> 
> It might help now and in for our older selves, if you could provide a
> simple reproducer for this.

I will add a repro to v2. 

> 
>> Signed-off-by: Song Liu <songliubraving@fb.com>
>> Reported-by: Ephraim Park <ephiepark@fb.com>
>> ---
>> kernel/events/core.c | 68 +++++++++++++++++++++++++++++++++++++++++++++++++++-
>> 1 file changed, 67 insertions(+), 1 deletion(-)
>> 
>> diff --git a/kernel/events/core.c b/kernel/events/core.c
>> index 5789810..623d38f 100644
>> --- a/kernel/events/core.c
>> +++ b/kernel/events/core.c
> 
>> @@ -766,6 +821,17 @@ perf_cgroup_set_timestamp(struct task_struct *task,
>> 	cgrp = perf_cgroup_from_task(task, ctx);
>> 	info = this_cpu_ptr(cgrp->info);
>> 	info->timestamp = ctx->timestamp;
>> +
>> +	/* set timestamp for ancestor cgroups */
>> +	if (cgrp->css.cgroup->level > 1) {
>> +		struct perf_cgroup_info *info = this_cpu_ptr(cgrp->info);
>> +		struct perf_event *event;
>> +
>> +		list_for_each_entry(event, &ctx->pinned_groups, group_entry)
>> +			perf_ancestor_cgroup_set_timestamp(event, cgrp, info);
>> +		list_for_each_entry(event, &ctx->flexible_groups, group_entry)
>> +			perf_ancestor_cgroup_set_timestamp(event, cgrp, info);
>> +	}
>> }
> 
> That doesn't make any kind of sense... if we're interested in ancestor
> groups, then WTH are you iterating _events_ ?
> 
> I'm expecting something like:
> 
> 	struct cgroup_subsys_state *css;
> 
> 	for (css = cgrp->css; css; css = css->parent) {
> 		/* fudge time */
> 	}

I guess this is what I have been looking for! Thanks Peter! 

V2 coming soon...

Song