Skip to contentSkip to conversation

perf: Fix perf_event_attr::exclusive rotation

PATCH Applied Peter Zijlstra 4 participants 14 messages Oct 29, 2020 – Nov 10, 2020 perf events

lore
Share & tools
lore

Peter Zijlstra

author maintainer
#0 lore raw

v2 cover perf: Fix perf_event_attr::exclusive rotation

Hi,

Andi recently added exclusive event group support to tools/perf:

https://lkml.kernel.org/r/[email protected]

and promptly found that they didn't work as specified.

(sorry for the resend, I forgot LKML the first time)

author maintainer
#7 lore raw

PATCH v2 4/4 perf: Tweak perf_event_attr::exclusive semantics

Currently perf_event_attr::exclusive can be used to ensure an event(group) is the sole group scheduled on the PMU. One consequence is that when you have a pinned event (say the watchdog) you can no longer have regular exclusive event(group)s.

Inspired by the fact that !pinned events are considered less strict, allow !pinned,exclusive events to share the PMU with pinned,!exclusive events.

Pinned,exclusive is still fully exclusive.

---
 kernel/events/core.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/events/core.c +1 -1
@@ -2637,7 +2637,7 @@ static int group_can_go_on(struct perf_e
 	 * If this group is exclusive and there are already
 	 * events on the CPU, it can't go on.
 	 */
-	if (event->attr.exclusive && cpuctx->active_oncpu)
+	if (event->attr.exclusive && !list_empty(get_event_list(event)))
 		return 0;
 	/*
 	 * Otherwise, try to add it if all previous groups were able
author maintainer
#5 lore raw

PATCH v2 3/4 perf: Fix event multiplexing for exclusive groups

In reply to Peter Zijlstra

Commit 9e6302056f80 ("perf: Use hrtimers for event multiplexing") placed the hrtimer (re)start call in the wrong place. Instead of capturing all scheduling failures, it only considered the PMU failure.

The result is that groups using perf_event_attr::exclusive are no longer rotated.

Fixes 9e6302056f80 ("perf: Use hrtimers for event multiplexing")
---
 kernel/events/core.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/events/core.c +1 -1
@@ -2612,7 +2612,6 @@ group_sched_in(struct perf_event *group_
 
 error:
 	pmu->cancel_txn(pmu);
-	perf_mux_hrtimer_restart(cpuctx);
 	return -EAGAIN;
 }
 
@@ -3672,6 +3671,7 @@ static int merge_sched_in(struct perf_ev
 
 		*can_add_hw = 0;
 		ctx->rotate_necessary = 1;
+		perf_mux_hrtimer_restart(cpuctx);
 	}
 
 	return 0;
author maintainer
#1 lore raw

PATCH v2 1/4 perf: Simplify group_sched_out()

In reply to Peter Zijlstra

Since event_sched_out() clears cpuctx->exclusive upon removal of an exclusive event (and only group leaders can be exclusive), there is no point in group_sched_out() trying to do it too. It is impossible for cpuctx->exclusive to still be set here.

---
 kernel/events/core.c |    3 ---
 1 file changed, 3 deletions(-)
kernel/events/core.c +0 -3
@@ -2312,9 +2312,6 @@ group_sched_out(struct perf_event *group
 		event_sched_out(event, cpuctx, ctx);
 
 	perf_pmu_enable(ctx->pmu);
-
-	if (group_event->attr.exclusive)
-		cpuctx->exclusive = 0;
 }
 
 #define DETACH_GROUP	0x01UL
author maintainer
#3 lore raw

PATCH v2 2/4 perf: Simplify group_sched_in()

In reply to Peter Zijlstra

Collate the error paths. Code duplication only leads to divergence and extra bugs.

---
 kernel/events/core.c |   10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)
kernel/events/core.c +3 -7
@@ -2580,11 +2580,8 @@ group_sched_in(struct perf_event *group_
 
 	pmu->start_txn(pmu, PERF_PMU_TXN_ADD);
 
-	if (event_sched_in(group_event, cpuctx, ctx)) {
-		pmu->cancel_txn(pmu);
-		perf_mux_hrtimer_restart(cpuctx);
-		return -EAGAIN;
-	}
+	if (event_sched_in(group_event, cpuctx, ctx))
+		goto error;
 
 	/*
 	 * Schedule in siblings as one group (if any):
@@ -2613,10 +2610,9 @@ group_sched_in(struct perf_event *group_
 	}
 	event_sched_out(group_event, cpuctx, ctx);
 
+error:
 	pmu->cancel_txn(pmu);
-
 	perf_mux_hrtimer_restart(cpuctx);
-
 	return -EAGAIN;
 }
 
#9 lore raw
In reply to Peter Zijlstra
Quoted 9 lines · Peter Zijlstra

Hi,

Andi recently added exclusive event group support to tools/perf:

https://lkml.kernel.org/r/[email protected]

and promptly found that they didn't work as specified.

(sorry for the resend, I forgot LKML the first time)

hm, it's too late for me to check ;-) but should I be able to do this with exclusive event.. running both command at the same time:

	$ sudo ./perf stat -e cycles:e -I 1000
	#           time             counts unit events
	     1.002430650         33,946,849      cycles:e                                                    
	     2.004920725        502,399,986      cycles:e                                                      (67.57%)
	     3.007087631        859,745,048      cycles:e                                                      (50.00%)
	     4.009078254        845,860,723      cycles:e                                                      (50.00%)
	     5.011086104        838,457,275      cycles:e                                                      (50.00%)
	$ sudo ./perf stat -e cycles:e  -I 1000
	#           time             counts unit events
	     1.001665466        848,973,404      cycles:e                                                      (50.01%)
	     2.003658048        856,505,255      cycles:e                                                      (50.00%)
	     3.005658022        842,737,973      cycles:e                                                      (50.00%)
	     4.007657797        844,800,598      cycles:e                                                      (50.00%)

jirka

#10 lore raw
Quoted 2 lines

hm, it's too late for me to check ;-) but should I be able to do this with exclusive event.. running both command at the same time:

Yes. The exclusive part only applies during a given context, but the two commands are different contexts.

You would only see a difference when in the same context, and you have multiple groups (or events) that could in theory schedule in parallel

e.g. something like

perf stat -e '{cycles,cycles},{cycles,cycles}' ...

The main motivation is actually that the "multiple groups" algorithm in perf doesn't work all that great: it has quite a few cases where it starves groups or makes the wrong decisions. That is because it is very difficult (likely NP complete) problem and the kernel takes a lot of short cuts to avoid spending too much time on it.

With exclusive it will be possible for a tool to generate "perfect groups" in user space and assume the kernel schedules it dumbly, but at least without any starvation.

-Andi

author maintainer
#11 lore raw
Quoted 5 lines · Andi Kleen

The main motivation is actually that the "multiple groups" algorithm in perf doesn't work all that great: it has quite a few cases where it starves groups or makes the wrong decisions. That is because it is very difficult (likely NP complete) problem and the kernel takes a lot of short cuts to avoid spending too much time on it.

The event scheduling should be starvation free, except in the presence of pinned events.

If you can show starvation without pinned events, it's a bug.

It will also always do equal or better than exclusive mode wrt PMU utilization. Again, if it doesn't it's a bug.

Please provide concrete examples for these two cases, or stop spreading FUD like this.

#12 lore raw
Quoted 7 lines · Peter Zijlstra
Quoted 5 lines · Andi Kleen

The main motivation is actually that the "multiple groups" algorithm in perf doesn't work all that great: it has quite a few cases where it starves groups or makes the wrong decisions. That is because it is very difficult (likely NP complete) problem and the kernel takes a lot of short cuts to avoid spending too much time on it.

Quoted 7 lines

The event scheduling should be starvation free, except in the presence of pinned events.

If you can show starvation without pinned events, it's a bug.

It will also always do equal or better than exclusive mode wrt PMU utilization. Again, if it doesn't it's a bug.

Simple example (I think we've shown that one before):

(on skylake)

$ cat /proc/sys/kernel/nmi_watchdog
0
$ perf stat -e instructions,cycles,frontend_retired.latency_ge_2,frontend_retired.latency_ge_16 -a sleep 2

Performance counter stats for 'system wide':

       654,514,990      instructions              #    0.34  insn per cycle           (50.67%)
     1,924,297,028      cycles                                                        (74.28%)
        21,708,935      frontend_retired.latency_ge_2                                     (75.01%)
         1,769,952      frontend_retired.latency_ge_16                                     (24.99%)
       2.002426541 seconds time elapsed

The second frontend_retired should be both getting 50% and the fixed events should be getting 100%. So several events are starved.

Another similar example is trying to schedule the topdown events on Icelake in parallel to other groups. It works with one extra group, but breaks with two.

(on icelake)

$ cat /proc/sys/kernel/nmi_watchdog
0
$ perf stat -e '{slots,topdown-bad-spec,topdown-be-bound,topdown-fe-bound,topdown-retiring},{branches,branches,branches,branches,branches,branches,branches,branches},{branches,branches,branches,branches,branches,branches,branches,branches}' -a sleep 1

Performance counter stats for 'system wide':

        71,229,087      slots                                                         (60.65%)
         5,066,320      topdown-bad-spec          #      7.1% bad speculation         (60.65%)
        35,080,387      topdown-be-bound          #     49.2% backend bound           (60.65%)
        22,769,750      topdown-fe-bound          #     32.0% frontend bound          (60.65%)
         8,336,760      topdown-retiring          #     11.7% retiring                (60.65%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
       1.001312511 seconds time elapsed

A tool using exclusive hopefully will be able to do better than this.

-Andi

author maintainer
#13 lore raw
Quoted 30 lines · Andi Kleen
Quoted 7 lines · Peter Zijlstra
Quoted 5 lines · Andi Kleen

The main motivation is actually that the "multiple groups" algorithm in perf doesn't work all that great: it has quite a few cases where it starves groups or makes the wrong decisions. That is because it is very difficult (likely NP complete) problem and the kernel takes a lot of short cuts to avoid spending too much time on it.

Quoted 7 lines

The event scheduling should be starvation free, except in the presence of pinned events.

If you can show starvation without pinned events, it's a bug.

It will also always do equal or better than exclusive mode wrt PMU utilization. Again, if it doesn't it's a bug.

Simple example (I think we've shown that one before):

(on skylake)

$ cat /proc/sys/kernel/nmi_watchdog
0
$ perf stat -e instructions,cycles,frontend_retired.latency_ge_2,frontend_retired.latency_ge_16 -a sleep 2

Performance counter stats for 'system wide':

       654,514,990      instructions              #    0.34  insn per cycle           (50.67%)
     1,924,297,028      cycles                                                        (74.28%)
        21,708,935      frontend_retired.latency_ge_2                                     (75.01%)
         1,769,952      frontend_retired.latency_ge_16                                     (24.99%)
Quoted 4 lines
       2.002426541 seconds time elapsed

The second frontend_retired should be both getting 50% and the fixed events should be getting 100%. So several events are starved.

*should* how? Also, nothing is 0% so nothing is getting starved.

Quoted 32 lines

Another similar example is trying to schedule the topdown events on Icelake in parallel to other groups. It works with one extra group, but breaks with two.

(on icelake)

$ cat /proc/sys/kernel/nmi_watchdog
0
$ perf stat -e '{slots,topdown-bad-spec,topdown-be-bound,topdown-fe-bound,topdown-retiring},{branches,branches,branches,branches,branches,branches,branches,branches},{branches,branches,branches,branches,branches,branches,branches,branches}' -a sleep 1

Performance counter stats for 'system wide':

        71,229,087      slots                                                         (60.65%)
         5,066,320      topdown-bad-spec          #      7.1% bad speculation         (60.65%)
        35,080,387      topdown-be-bound          #     49.2% backend bound           (60.65%)
        22,769,750      topdown-fe-bound          #     32.0% frontend bound          (60.65%)
         8,336,760      topdown-retiring          #     11.7% retiring                (60.65%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
           424,584      branches                                                      (70.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
         3,634,075      branches                                                      (30.00%)
Quoted 3 lines
       1.001312511 seconds time elapsed

A tool using exclusive hopefully will be able to do better than this.

I don't see how, exclusive will always result in equal or worse PMU utilization, never better.

merge bot Applied
#8 lore raw

[tip: perf/urgent] perf: Tweak perf_event_attr::exclusive semantics

In reply to Peter Zijlstra
The following commit has been merged into the perf/urgent branch of tip:

The following commit has been merged into the perf/urgent branch of tip:

Commit-ID:     1908dc911792067287458fdb0800f036f4f4e0f6
Gitweb:        https://git.kernel.org/tip/1908dc911792067287458fdb0800f036f4f4e0f6
Author:        Peter Zijlstra <[email protected]>
AuthorDate:    Thu, 29 Oct 2020 16:32:22 +01:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Mon, 09 Nov 2020 18:12:36 +01:00

perf: Tweak perf_event_attr::exclusive semantics

Currently perf_event_attr::exclusive can be used to ensure an event(group) is the sole group scheduled on the PMU. One consequence is that when you have a pinned event (say the watchdog) you can no longer have regular exclusive event(group)s.

Inspired by the fact that !pinned events are considered less strict, allow !pinned,exclusive events to share the PMU with pinned,!exclusive events.

Pinned,exclusive is still fully exclusive.

---
 kernel/events/core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/events/core.c +1 -1
index 00be48a..dc568ca 100644
@@ -2637,7 +2637,7 @@ static int group_can_go_on(struct perf_event *event,
 	 * If this group is exclusive and there are already
 	 * events on the CPU, it can't go on.
 	 */
-	if (event->attr.exclusive && cpuctx->active_oncpu)
+	if (event->attr.exclusive && !list_empty(get_event_list(event)))
 		return 0;
 	/*
 	 * Otherwise, try to add it if all previous groups were able
merge bot Applied
#2 lore raw

[tip: perf/urgent] perf: Simplify group_sched_out()

In reply to Peter Zijlstra
The following commit has been merged into the perf/urgent branch of tip:

The following commit has been merged into the perf/urgent branch of tip:

Commit-ID:     8c7855d82933bab7fa5e96f0e568fc125c2e1ab4
Gitweb:        https://git.kernel.org/tip/8c7855d82933bab7fa5e96f0e568fc125c2e1ab4
Author:        Peter Zijlstra <[email protected]>
AuthorDate:    Thu, 29 Oct 2020 16:28:25 +01:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Mon, 09 Nov 2020 18:12:35 +01:00

perf: Simplify group_sched_out()

Since event_sched_out() clears cpuctx->exclusive upon removal of an exclusive event (and only group leaders can be exclusive), there is no point in group_sched_out() trying to do it too. It is impossible for cpuctx->exclusive to still be set here.

---
 kernel/events/core.c | 3 ---
 1 file changed, 3 deletions(-)
kernel/events/core.c +0 -3
index d67c9cb..9a57366 100644
@@ -2312,9 +2312,6 @@ group_sched_out(struct perf_event *group_event,
 		event_sched_out(event, cpuctx, ctx);
 
 	perf_pmu_enable(ctx->pmu);
-
-	if (group_event->attr.exclusive)
-		cpuctx->exclusive = 0;
 }
 
 #define DETACH_GROUP	0x01UL
merge bot Applied
#4 lore raw

[tip: perf/urgent] perf: Simplify group_sched_in()

In reply to Peter Zijlstra
The following commit has been merged into the perf/urgent branch of tip:

The following commit has been merged into the perf/urgent branch of tip:

Commit-ID:     251ff2d49347793d348babcff745289b11910e96
Gitweb:        https://git.kernel.org/tip/251ff2d49347793d348babcff745289b11910e96
Author:        Peter Zijlstra <[email protected]>
AuthorDate:    Thu, 29 Oct 2020 16:29:15 +01:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Mon, 09 Nov 2020 18:12:35 +01:00

perf: Simplify group_sched_in()

Collate the error paths. Code duplication only leads to divergence and extra bugs.

---
 kernel/events/core.c | 10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)
kernel/events/core.c +3 -7
index 9a57366..f0e5268 100644
@@ -2580,11 +2580,8 @@ group_sched_in(struct perf_event *group_event,
 
 	pmu->start_txn(pmu, PERF_PMU_TXN_ADD);
 
-	if (event_sched_in(group_event, cpuctx, ctx)) {
-		pmu->cancel_txn(pmu);
-		perf_mux_hrtimer_restart(cpuctx);
-		return -EAGAIN;
-	}
+	if (event_sched_in(group_event, cpuctx, ctx))
+		goto error;
 
 	/*
 	 * Schedule in siblings as one group (if any):
@@ -2613,10 +2610,9 @@ group_error:
 	}
 	event_sched_out(group_event, cpuctx, ctx);
 
+error:
 	pmu->cancel_txn(pmu);
-
 	perf_mux_hrtimer_restart(cpuctx);
-
 	return -EAGAIN;
 }
 
merge bot Applied
#6 lore raw

[tip: perf/urgent] perf: Fix event multiplexing for exclusive groups

In reply to Peter Zijlstra
The following commit has been merged into the perf/urgent branch of tip:

The following commit has been merged into the perf/urgent branch of tip:

Commit-ID:     2714c3962f304d031d5016c963c4b459337b0749
Gitweb:        https://git.kernel.org/tip/2714c3962f304d031d5016c963c4b459337b0749
Author:        Peter Zijlstra <[email protected]>
AuthorDate:    Thu, 29 Oct 2020 16:29:53 +01:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Mon, 09 Nov 2020 18:12:36 +01:00

perf: Fix event multiplexing for exclusive groups

Commit 9e6302056f80 ("perf: Use hrtimers for event multiplexing") placed the hrtimer (re)start call in the wrong place. Instead of capturing all scheduling failures, it only considered the PMU failure.

The result is that groups using perf_event_attr::exclusive are no longer rotated.

Fixes 9e6302056f80 ("perf: Use hrtimers for event multiplexing")
---
 kernel/events/core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/events/core.c +1 -1
index f0e5268..00be48a 100644
@@ -2612,7 +2612,6 @@ group_error:
 
 error:
 	pmu->cancel_txn(pmu);
-	perf_mux_hrtimer_restart(cpuctx);
 	return -EAGAIN;
 }
 
@@ -3672,6 +3671,7 @@ static int merge_sched_in(struct perf_event *event, void *data)
 
 		*can_add_hw = 0;
 		ctx->rotate_necessary = 1;
+		perf_mux_hrtimer_restart(cpuctx);
 	}
 
 	return 0;