perf/core: Fix installing cgroup event into cpu

PATCH RESEND Superseded [email protected] 3 participants 5 messages Feb 8, 2018 – Feb 10, 2018 perf events

lore

Share & tools

lore

[email protected]

author

2018-02-08 03:37 UTC #0 lore raw

PATCH RESEND v4 1/1 perf/core: Fix installing cgroup event into cpu

From: "leilei.lin" <[email protected]>

Do not install cgroup event into the CPU context and schedule it if the cgroup is not running on this CPU

While there is no task of cgroup running specified CPU, current kernel still install cgroup event into CPU context that causes another cgroup event can't be installed into this CPU.

This patch prevent scheduling events at __perf_install_in_context() and installing events at list_update_cgroup_event() if cgroup isn't running on specified CPU.

Signed-off-by leilei.lin <[email protected]>

---
v2: Set cpuctx->cgrp only if the same cgroup is running on this
CPU otherwise following events couldn't be activated immediately
v3: Enhance the comments and commit message
v4: Adjust to config

 kernel/events/core.c | 50 +++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 37 insertions(+), 13 deletions(-)

kernel/events/core.c +37 ~~-13~~

index 4df5b69..fd28d61 100644
@@ -933,31 +933,41 @@ list_update_cgroup_event(struct perf_event *event,
 {
 	struct perf_cpu_context *cpuctx;
 	struct list_head *cpuctx_entry;
+	struct perf_cgroup *cgrp;
 
 	if (!is_cgroup_event(event))
 		return;
 
-	if (add && ctx->nr_cgroups++)
-		return;
-	else if (!add && --ctx->nr_cgroups)
-		return;
 	/*
 	 * Because cgroup events are always per-cpu events,
 	 * this will always be called from the right CPU.
 	 */
 	cpuctx = __get_cpu_context(ctx);
-	cpuctx_entry = &cpuctx->cgrp_cpuctx_entry;
-	/* cpuctx->cgrp is NULL unless a cgroup event is active in this CPU .*/
-	if (add) {
-		struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
+	cgrp = perf_cgroup_from_task(current, ctx);
 
-		list_add(cpuctx_entry, this_cpu_ptr(&cgrp_cpuctx_list));
-		if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
+	/*
+	 * if only the cgroup is running on this cpu,
+	 * we put/remove this cgroup into cpu context.
+	 * Or it would case mismatch in following cgroup
+	 * events at event_filter_match()
+	 */
+	if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) {
+		if (add)
 			cpuctx->cgrp = cgrp;
-	} else {
-		list_del(cpuctx_entry);
-		cpuctx->cgrp = NULL;
+		else
+			cpuctx->cgrp = NULL;
 	}
+
+	if (add && ctx->nr_cgroups++)
+		return;
+	else if (!add && --ctx->nr_cgroups)
+		return;
+
+	cpuctx_entry = &cpuctx->cgrp_cpuctx_entry;
+	if (add)
+		list_add(cpuctx_entry, this_cpu_ptr(&cgrp_cpuctx_list));
+	else
+		list_del(cpuctx_entry);
 }
 
 #else /* !CONFIG_CGROUP_PERF */

@@ -2311,6 +2321,20 @@ static int  __perf_install_in_context(void *info)
 		raw_spin_lock(&task_ctx->lock);
 	}
 
+#ifdef CONFIG_CGROUP_PERF
+	if (is_cgroup_event(event)) {
+		/*
+		 * Only care about cgroup events.
+		 *
+		 * If only the task belongs to cgroup of this event,
+		 * we will continue the installment
+		 */
+		struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
+		reprogram = cgroup_is_descendant(cgrp->css.cgroup,
+					event->cgrp->css.cgroup);
+	}
+#endif
+
 	if (reprogram) {
 		ctx_sched_out(ctx, cpuctx, EVENT_TIME);
 		add_event_to_ctx(event, ctx);

Signature · 1 line

2.8.4.31.g9ed660f

Lin Xiulei

author

2018-02-09 01:52 UTC #2 lore raw

2018-02-08 23:36 GMT+08:00 Jiri Olsa <[email protected]>:

Quoted 38 lines · On Thu, Feb 08, 2018 at 11:33:44AM +0800, [email protected] wrote:

Quoted 36 lines
From: "leilei.lin" <[email protected]>
Do not install cgroup event into the CPU context and schedule it if the cgroup is not running on this CPU
While there is no task of cgroup running specified CPU, current kernel still install cgroup event into CPU context that causes another cgroup event can't be installed into this CPU.
This patch prevent scheduling events at __perf_install_in_context() and installing events at list_update_cgroup_event() if cgroup isn't running on specified CPU.
Signed-off-by leilei.lin <[email protected]>
---
v2: Set cpuctx->cgrp only if the same cgroup is running on this
CPU otherwise following events couldn't be activated immediately
v3: Enhance the comments and commit message
v4: Adjust to config
 kernel/events/core.c | 50 +++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 37 insertions(+), 13 deletions(-)
kernel/events/core.c +1 -0
index 4df5b69..fd28d61 100644
@@ -933,31 +933,41 @@ list_update_cgroup_event(struct perf_event *event,
 {
      struct perf_cpu_context *cpuctx;
      struct list_head *cpuctx_entry;
+     struct perf_cgroup *cgrp;
 
      if (!is_cgroup_event(event))
              return;
 

Quoted 8 lines

kernel/events/core.c list_update_cgroup_event(struct perf_event *event, · line 943
Quoted 4 lines
kernel/events/core.c +0 -4
-     if (add && ctx->nr_cgroups++)
-             return;
-     else if (!add && --ctx->nr_cgroups)
-             return;
I might be missing something, but should this check stay on the top regardles of the cgroup_is_descendant check below?

I don't think so, if event A on cgroup A is opened and immediately followed by a event B opened on cgroup B, then
"if (add && ctx->nr_cgroups++)" would __return__ with
cpuctx->cgrp = cgroup A, that is incorrect.

And previous thread is here https://lkml.org/lkml/2018/1/24/79

Quoted 4 lines

you could put NULL into cpuctx->cgrp on context with cgroup event in the list

what's the harm? It's invoked by perf_remove_from_context() when an event is ready to be released. And whenever process/cgroup is scheduled, cpuctx->cgrp will be set

In case that the patch was not sorted well, I put the patched code

```
static inline void
list_update_cgroup_event(struct perf_event *event,
             struct perf_event_context *ctx, bool add)
{
    struct perf_cpu_context *cpuctx;
    struct list_head *cpuctx_entry;
    struct perf_cgroup *cgrp;

    if (!is_cgroup_event(event))
        return;

    /*
     * Because cgroup events are always per-cpu events,
     * this will always be called from the right CPU.
     */
    cpuctx = __get_cpu_context(ctx);
    cgrp = perf_cgroup_from_task(current, ctx);

    /*
     * if only the cgroup is running on this cpu,
     * we put/remove this cgroup into cpu context.
     * Or it would case mismatch in following cgroup
     * events at event_filter_match()
     */
    if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) {
        if (add)
            cpuctx->cgrp = cgrp;
        else
            cpuctx->cgrp = NULL;
    }

    if (add && ctx->nr_cgroups++)
        return;
    else if (!add && --ctx->nr_cgroups)
        return;

    cpuctx_entry = &cpuctx->cgrp_cpuctx_entry;
    if (add)
        list_add(cpuctx_entry, this_cpu_ptr(&cgrp_cpuctx_list));
    else
        list_del(cpuctx_entry);
}
```

thanks

Quoted 34 lines

thanks,
jirka

Quoted 27 lines

kernel/events/core.c +11 -9

      /*
       * Because cgroup events are always per-cpu events,
       * this will always be called from the right CPU.
       */
      cpuctx = __get_cpu_context(ctx);
-     cpuctx_entry = &cpuctx->cgrp_cpuctx_entry;
-     /* cpuctx->cgrp is NULL unless a cgroup event is active in this CPU .*/
-     if (add) {
-             struct perf_cgroup *cgrp = perf_cgroup_from_task(current, ctx);
+     cgrp = perf_cgroup_from_task(current, ctx);
 
-             list_add(cpuctx_entry, this_cpu_ptr(&cgrp_cpuctx_list));
-             if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
+     /*
+      * if only the cgroup is running on this cpu,
+      * we put/remove this cgroup into cpu context.
+      * Or it would case mismatch in following cgroup
+      * events at event_filter_match()
+      */
+     if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) {
+             if (add)
                      cpuctx->cgrp = cgrp;
-     } else {
-             list_del(cpuctx_entry);
-             cpuctx->cgrp = NULL;
+             else
+                     cpuctx->cgrp = NULL;

SNIP

Peter Zijlstra

maintainer

2018-02-09 10:11 UTC #3 lore raw

In reply to [email protected]

Quoted 45 lines · On Thu, Feb 08, 2018 at 11:33:44AM +0800, [email protected] wrote:

From: "leilei.lin" <[email protected]>
Do not install cgroup event into the CPU context and schedule it if the cgroup is not running on this CPU
While there is no task of cgroup running specified CPU, current kernel still install cgroup event into CPU context that causes another cgroup event can't be installed into this CPU.
This patch prevent scheduling events at __perf_install_in_context() and installing events at list_update_cgroup_event() if cgroup isn't running on specified CPU.
Signed-off-by leilei.lin <[email protected]>
---
v2: Set cpuctx->cgrp only if the same cgroup is running on this
CPU otherwise following events couldn't be activated immediately
v3: Enhance the comments and commit message
v4: Adjust to config
 kernel/events/core.c | 50 +++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 37 insertions(+), 13 deletions(-)
kernel/events/core.c +4 -0
index 4df5b69..fd28d61 100644
@@ -933,31 +933,41 @@ list_update_cgroup_event(struct perf_event *event,
 {
 	struct perf_cpu_context *cpuctx;
 	struct list_head *cpuctx_entry;
+	struct perf_cgroup *cgrp;
 
 	if (!is_cgroup_event(event))
 		return;
 
 	/*
 	 * Because cgroup events are always per-cpu events,
 	 * this will always be called from the right CPU.
 	 */
 	cpuctx = __get_cpu_context(ctx);
+	cgrp = perf_cgroup_from_task(current, ctx);
 
+	/*
+	 * if only the cgroup is running on this cpu,

kernel/events/core.c list_update_cgroup_event(struct perf_event *event, · line 959

Quoted 10 lines

kernel/events/core.c +8 -0

+	 * we put/remove this cgroup into cpu context.
+	 * Or it would case mismatch in following cgroup
+	 * events at event_filter_match()
+	 */
+	if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) {
+		if (add)
 			cpuctx->cgrp = cgrp;
+		else
+			cpuctx->cgrp = NULL;
 	}

I am still not convinced this is correct.

Suppose we have

R
/ \
A B

/ \

And our current task is of B, and B has an event.

We then install an event in C, if we then destroy our event in C, it would clear cpuctx->cgrp, which is wrong, since there is still an event in B.

Simpler still, if B were to have 2 events, and we'd remove one, that would still clear cpuctx->cgrp, even though there is an event left.

This is the exact issue I pointed out last time, and I still don't see how it would now be correct.

Northing explains why its ok to have NULL cpuctx->cgrp when there are in fact still cgroup events on the CPU.

Lin Xiulei

author

2018-02-10 12:28 UTC #4 lore raw

2018-02-09 18:11 GMT+08:00 Peter Zijlstra <[email protected]>:

Quoted 73 lines · On Thu, Feb 08, 2018 at 11:33:44AM +0800, [email protected] wrote:

Quoted 45 lines
From: "leilei.lin" <[email protected]>
Do not install cgroup event into the CPU context and schedule it if the cgroup is not running on this CPU
While there is no task of cgroup running specified CPU, current kernel still install cgroup event into CPU context that causes another cgroup event can't be installed into this CPU.
This patch prevent scheduling events at __perf_install_in_context() and installing events at list_update_cgroup_event() if cgroup isn't running on specified CPU.
Signed-off-by leilei.lin <[email protected]>
---
v2: Set cpuctx->cgrp only if the same cgroup is running on this
CPU otherwise following events couldn't be activated immediately
v3: Enhance the comments and commit message
v4: Adjust to config
 kernel/events/core.c | 50 +++++++++++++++++++++++++++++++++++++-------------
 1 file changed, 37 insertions(+), 13 deletions(-)
kernel/events/core.c +4 -0
index 4df5b69..fd28d61 100644
@@ -933,31 +933,41 @@ list_update_cgroup_event(struct perf_event *event,
 {
      struct perf_cpu_context *cpuctx;
      struct list_head *cpuctx_entry;
+     struct perf_cgroup *cgrp;
 
      if (!is_cgroup_event(event))
              return;
 
      /*
       * Because cgroup events are always per-cpu events,
       * this will always be called from the right CPU.
       */
      cpuctx = __get_cpu_context(ctx);
+     cgrp = perf_cgroup_from_task(current, ctx);
 
+     /*
+      * if only the cgroup is running on this cpu,
kernel/events/core.c list_update_cgroup_event(struct perf_event *event, · line 959
Quoted 10 lines
kernel/events/core.c +8 -0
+      * we put/remove this cgroup into cpu context.
+      * Or it would case mismatch in following cgroup
+      * events at event_filter_match()
+      */
+     if (cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup)) {
+             if (add)
                      cpuctx->cgrp = cgrp;
+             else
+                     cpuctx->cgrp = NULL;
      }
I am still not convinced this is correct.
Suppose we have
R
/ \
A B
    / \
C
And our current task is of B, and B has an event.
We then install an event in C, if we then destroy our event in C, it would clear cpuctx->cgrp, which is wrong, since there is still an event in B.

Quoted 8 lines

Simpler still, if B were to have 2 events, and we'd remove one, that would still clear cpuctx->cgrp, even though there is an event left.
This is the exact issue I pointed out last time, and I still don't see how it would now be correct.
Northing explains why its ok to have NULL cpuctx->cgrp when there are in fact still cgroup events on the CPU.

I got your point now, sorry for misunderstanding it last time. I wanna confirm it
that logic in __add__ is correct and I'd like to make a slight improvement

```
/* We only have to care about the first time of initiating cpuctx->cgrp,
 * which is when cpuctx->cgrp == NULL, otherwise cpuctx->cgrp was set
 * in perf_cgroup_switch() correctly.
 */
if (add && !cpuctx->cgrp &&
     cgroup_is_descendant(cgrp->css.cgroup, event->cgrp->css.cgroup))
    cpuctx->cgrp = cgrp;
```

And logic in __del__ should be rolled back to previous code that once ctx->nr_cgroups == 0, set cpuctx->cgrp to NULL.

thanks