[PATCH v2] perf/core: Fix cgroup events tracking

From: Chengming Zhou
Date: Wed Dec 07 2022 - 07:41:19 EST


We encounter perf warnings when using cgroup events like:
```
cd /sys/fs/cgroup
mkdir test
perf stat -e cycles -a -G test
```

WARNING: CPU: 0 PID: 690 at kernel/events/core.c:849 perf_cgroup_switch+0xb2/0xc0
Call Trace:
<TASK>
__schedule+0x4ae/0x9f0
? _raw_spin_unlock_irqrestore+0x23/0x40
? __cond_resched+0x18/0x20
preempt_schedule_common+0x2d/0x70
__cond_resched+0x18/0x20
wait_for_completion+0x2f/0x160
? cpu_stop_queue_work+0x9e/0x130
affine_move_task+0x18a/0x4f0

WARNING: CPU: 0 PID: 690 at kernel/events/core.c:829 ctx_sched_in+0x1cf/0x1e0
Call Trace:
<TASK>
? ctx_sched_out+0xb7/0x1b0
perf_cgroup_switch+0x88/0xc0
__schedule+0x4ae/0x9f0
? _raw_spin_unlock_irqrestore+0x23/0x40
? __cond_resched+0x18/0x20
preempt_schedule_common+0x2d/0x70
__cond_resched+0x18/0x20
wait_for_completion+0x2f/0x160
? cpu_stop_queue_work+0x9e/0x130
affine_move_task+0x18a/0x4f0

The above two warnings are not complete here since I remove other
unimportant information. The problem is caused by the perf cgroup
events tracking:

CPU0 CPU1
perf_event_open()
perf_event_alloc()
account_event()
account_event_cpu()
atomic_inc(perf_cgroup_events)
__perf_event_task_sched_out()
if (atomic_read(perf_cgroup_events))
perf_cgroup_switch()
// kernel/events/core.c:849
WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0)
if (READ_ONCE(cpuctx->cgrp) == cgrp) // false
return
perf_ctx_lock()
ctx_sched_out()
cpuctx->cgrp = cgrp
ctx_sched_in()
perf_cgroup_set_timestamp()
// kernel/events/core.c:829
WARN_ON_ONCE(!ctx->nr_cgroups)
perf_ctx_unlock()
perf_install_in_context()
cpu_function_call()
__perf_install_in_context()
add_event_to_ctx()
list_add_event()
perf_cgroup_event_enable()
ctx->nr_cgroups++
cpuctx->cgrp = X

We can see from above that we wrongly use percpu atomic perf_cgroup_events
to check if we need to perf_cgroup_switch(), which should only be used
when we know this CPU has cgroup events enabled.

The commit bd2756811766 ("perf: Rewrite core context handling") change
to have only one context per-CPU, so we can just use cpuctx->cgrp to
check if this CPU has cgroup events enabled.

So percpu atomic perf_cgroup_events is not needed.

Fixes: bd2756811766 ("perf: Rewrite core context handling")
Signed-off-by: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx>
Tested-by: Ravi Bangoria <ravi.bangoria@xxxxxxx>
---
v2:
- Remove timestamps and fix race graph in commit log per Ravi Bangoria.
- Add Fixes tag and Tested-by tag.
---
kernel/events/core.c | 42 ++++++++++--------------------------------
1 file changed, 10 insertions(+), 32 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index eacc3702654d..5d97a9f26003 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -380,7 +380,6 @@ enum event_type_t {

/*
* perf_sched_events : >0 events exist
- * perf_cgroup_events: >0 per-cpu cgroup events exist on this cpu
*/

static void perf_sched_delayed(struct work_struct *work);
@@ -389,7 +388,6 @@ static DECLARE_DELAYED_WORK(perf_sched_work, perf_sched_delayed);
static DEFINE_MUTEX(perf_sched_mutex);
static atomic_t perf_sched_count;

-static DEFINE_PER_CPU(atomic_t, perf_cgroup_events);
static DEFINE_PER_CPU(struct pmu_event_list, pmu_sb_events);

static atomic_t nr_mmap_events __read_mostly;
@@ -844,9 +842,16 @@ static void perf_cgroup_switch(struct task_struct *task)
struct perf_cpu_context *cpuctx = this_cpu_ptr(&perf_cpu_context);
struct perf_cgroup *cgrp;

- cgrp = perf_cgroup_from_task(task, NULL);
+ /*
+ * cpuctx->cgrp is set when the first cgroup event enabled,
+ * and is cleared when the last cgroup event disabled.
+ */
+ if (READ_ONCE(cpuctx->cgrp) == NULL)
+ return;

WARN_ON_ONCE(cpuctx->ctx.nr_cgroups == 0);
+
+ cgrp = perf_cgroup_from_task(task, NULL);
if (READ_ONCE(cpuctx->cgrp) == cgrp)
return;

@@ -3631,8 +3636,7 @@ void __perf_event_task_sched_out(struct task_struct *task,
* to check if we have to switch out PMU state.
* cgroup event are system-wide mode only
*/
- if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
- perf_cgroup_switch(next);
+ perf_cgroup_switch(next);
}

static bool perf_less_group_idx(const void *l, const void *r)
@@ -4974,15 +4978,6 @@ static void unaccount_pmu_sb_event(struct perf_event *event)
detach_sb_event(event);
}

-static void unaccount_event_cpu(struct perf_event *event, int cpu)
-{
- if (event->parent)
- return;
-
- if (is_cgroup_event(event))
- atomic_dec(&per_cpu(perf_cgroup_events, cpu));
-}
-
#ifdef CONFIG_NO_HZ_FULL
static DEFINE_SPINLOCK(nr_freq_lock);
#endif
@@ -5048,8 +5043,6 @@ static void unaccount_event(struct perf_event *event)
schedule_delayed_work(&perf_sched_work, HZ);
}

- unaccount_event_cpu(event, event->cpu);
-
unaccount_pmu_sb_event(event);
}

@@ -11679,15 +11672,6 @@ static void account_pmu_sb_event(struct perf_event *event)
attach_sb_event(event);
}

-static void account_event_cpu(struct perf_event *event, int cpu)
-{
- if (event->parent)
- return;
-
- if (is_cgroup_event(event))
- atomic_inc(&per_cpu(perf_cgroup_events, cpu));
-}
-
/* Freq events need the tick to stay alive (see perf_event_task_tick). */
static void account_freq_event_nohz(void)
{
@@ -11775,8 +11759,6 @@ static void account_event(struct perf_event *event)
}
enabled:

- account_event_cpu(event, event->cpu);
-
account_pmu_sb_event(event);
}

@@ -12822,13 +12804,11 @@ static void __perf_pmu_remove(struct perf_event_context *ctx,

perf_event_groups_for_cpu_pmu(event, groups, cpu, pmu) {
perf_remove_from_context(event, 0);
- unaccount_event_cpu(event, cpu);
put_pmu_ctx(event->pmu_ctx);
list_add(&event->migrate_entry, events);

for_each_sibling_event(sibling, event) {
perf_remove_from_context(sibling, 0);
- unaccount_event_cpu(sibling, cpu);
put_pmu_ctx(sibling->pmu_ctx);
list_add(&sibling->migrate_entry, events);
}
@@ -12847,7 +12827,6 @@ static void __perf_pmu_install_event(struct pmu *pmu,

if (event->state >= PERF_EVENT_STATE_OFF)
event->state = PERF_EVENT_STATE_INACTIVE;
- account_event_cpu(event, cpu);
perf_install_in_context(ctx, event, cpu);
}

@@ -13742,8 +13721,7 @@ static int __perf_cgroup_move(void *info)
struct task_struct *task = info;

preempt_disable();
- if (atomic_read(this_cpu_ptr(&perf_cgroup_events)))
- perf_cgroup_switch(task);
+ perf_cgroup_switch(task);
preempt_enable();

return 0;
--
2.37.2