Skip to contentSkip to conversation

perf_events: fix event scheduling issues introduced by transactional API

PATCH In review Stephane Eranian 2 participants 2 messages May 24, 2010 – May 25, 2010

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf_events: fix event scheduling issues introduced by transactional API

The transactional API patch between the generic and model-specific code introduced several important bugs with event scheduling, at least on X86. If you had pinned events, e.g., watchdog, and were over-committing the PMU, you would get bogus counts. The bug was showing up on Intel CPU because events would move around more often that on AMD. But the problem also existed on AMD, though harder to expose.

The issues were:

- group_sched_in() was missing a cancel_txn() in the error path

- cpuc->n_added was not properly maintained, leading to missing actions in hw_perf_enable(), i.e., n_running being 0. You cannot update n_added until you know the transaction has succeeded. In case of failed transaction n_added was not adjusted back.

- in case of failed transactions, event_sched_out() was called and eventually invoked x86_disable_event() to touch the HW reg. But with transactions, on X86, event_sched_in() does not touch HW registers, it simply collects events into a list. Thus, you could end up calling x86_disable_event() on a counter which did not correspond to the current event when idx != -1.

The patch modifies the generic and X86 code to avoid all those problems.

First, n_added is now ONLY updated once we know the transaction has succeeded.

Second, we encapsulate the event_sched_in() and event_sched_out() in group_sched_in() inside the transaction. That way, in the X6 code, we can detect that we are being asked to stop an event which was not yet started by checking cpuc->group_flag.

With this patch, you can now overcommit the PMU even with pinned system-wide events present and still get valid counts.

arch/x86/kernel/cpu/perf_event.c +27 -8
index fd4db0d..25dfd30 100644
@@ -105,6 +105,7 @@ struct cpu_hw_events {
 	int			enabled;
 
 	int			n_events;
+	int			n_events_trans;
 	int			n_added;
 	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 	u64			tags[X86_PMC_IDX_MAX];
@@ -591,7 +592,7 @@ void hw_perf_disable(void)
 	if (!cpuc->enabled)
 		return;
 
-	cpuc->n_added = 0;
+	cpuc->n_added = cpuc->n_events_trans = 0;
 	cpuc->enabled = 0;
 	barrier();
 
@@ -820,7 +821,7 @@ void hw_perf_enable(void)
 		 * apply assignment obtained either from
 		 * hw_perf_group_sched_in() or x86_pmu_enable()
 		 *
-		 * step1: save events moving to new counters
+		 * step1: stop-save old events moving to new counters
 		 * step2: reprogram moved events into new counters
 		 */
 		for (i = 0; i < n_running; i++) {
@@ -982,7 +983,7 @@ static int x86_pmu_enable(struct perf_event *event)
 
 out:
 	cpuc->n_events = n;
-	cpuc->n_added += n - n0;
+	cpuc->n_events_trans += n - n0;
 
 	return 0;
 }
@@ -1070,7 +1071,7 @@ static void x86_pmu_stop(struct perf_event *event)
 	struct hw_perf_event *hwc = &event->hw;
 	int idx = hwc->idx;
 
-	if (!__test_and_clear_bit(idx, cpuc->active_mask))
+	if (idx == -1 || !__test_and_clear_bit(idx, cpuc->active_mask))
 		return;
 
 	x86_pmu.disable(event);
@@ -1087,14 +1088,30 @@ static void x86_pmu_stop(struct perf_event *event)
 static void x86_pmu_disable(struct perf_event *event)
 {
 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	bool no_trans;
 	int i;
 
-	x86_pmu_stop(event);
+	/*
+	 * when coming from a transaction, then the event has not
+	 * actually been scheduled in yet. It has only been collected
+	 * (collect_events), thus we cannot apply a full disable:
+	 * - put_constraints() assumes went thru schedule_events()
+	 * - x86_pmu_stop() assumes went thru x86_pmu_start() because
+	 *   maybe idx != -1 and thus we may overwrite another the wrong
+	 *   counter (e.g, pinned event).
+	 *
+	 * When called from a transaction, we need to un-collect the
+	 * event, i..e, remove the event from event_list[]
+	 */
+	no_trans = !(cpuc->group_flag & PERF_EVENT_TXN_STARTED);
+
+	if (no_trans)
+		x86_pmu_stop(event);
 
 	for (i = 0; i < cpuc->n_events; i++) {
 		if (event == cpuc->event_list[i]) {
 
-			if (x86_pmu.put_event_constraints)
+			if (no_trans && x86_pmu.put_event_constraints)
 				x86_pmu.put_event_constraints(cpuc, event);
 
 			while (++i < cpuc->n_events)
@@ -1104,7 +1121,8 @@ static void x86_pmu_disable(struct perf_event *event)
 			break;
 		}
 	}
-	perf_event_update_userpage(event);
+	if (no_trans)
+		perf_event_update_userpage(event);
 }
 
 static int x86_pmu_handle_irq(struct pt_regs *regs)
@@ -1418,7 +1436,8 @@ static int x86_pmu_commit_txn(const struct pmu *pmu)
 	 * will be used by hw_perf_enable()
 	 */
 	memcpy(cpuc->assign, assign, n*sizeof(int));
-
+	cpuc->n_added += cpuc->n_events_trans;
+	cpuc->n_events_trans = 0;
 	return 0;
 }
 
kernel/perf_event.c +7 -3
index 7e3bcf1..b0ccf4b 100644
@@ -652,8 +652,11 @@ group_sched_in(struct perf_event *group_event,
 	if (txn)
 		pmu->start_txn(pmu);
 
-	if (event_sched_in(group_event, cpuctx, ctx))
+	if (event_sched_in(group_event, cpuctx, ctx)) {
+		if (txn)
+			pmu->cancel_txn(pmu);
 		return -EAGAIN;
+	}
 
 	/*
 	 * Schedule in siblings as one group (if any):
@@ -675,8 +678,6 @@ group_sched_in(struct perf_event *group_event,
 	}
 
 group_error:
-	if (txn)
-		pmu->cancel_txn(pmu);
 
 	/*
 	 * Groups can be scheduled in as one unit only, so undo any
@@ -689,6 +690,9 @@ group_error:
 	}
 	event_sched_out(group_event, cpuctx, ctx);
 
+	if (txn)
+		pmu->cancel_txn(pmu);
+
 	return -EAGAIN;
 }
 
#1 lore raw
Quoted 17 lines · Stephane Eranian

The transactional API patch between the generic and model-specific code introduced several important bugs with event scheduling, at least on X86. If you had pinned events, e.g., watchdog, and were over-committing the PMU, you would get bogus counts. The bug was showing up on Intel CPU because events would move around more often that on AMD. But the problem also existed on AMD, though harder to expose.

The issues were:

- group_sched_in() was missing a cancel_txn() in the error path

- cpuc->n_added was not properly maintained, leading to missing actions in hw_perf_enable(), i.e., n_running being 0. You cannot update n_added until you know the transaction has succeeded. In case of failed transaction n_added was not adjusted back.

Quoted 4 lines

- in case of failed transactions, event_sched_out() was called and eventually invoked x86_disable_event() to touch the HW reg. But with transactions, on X86, event_sched_in() does not touch HW registers, it simply collects events into a list. Thus, you

event_sched_in always does not touch HW registers, both with and without transactions.

Quoted 125 lines

could end up calling x86_disable_event() on a counter which did not correspond to the current event when idx != -1.

The patch modifies the generic and X86 code to avoid all those problems.

First, n_added is now ONLY updated once we know the transaction has succeeded.

Second, we encapsulate the event_sched_in() and event_sched_out() in group_sched_in() inside the transaction. That way, in the X6 code, we can detect that we are being asked to stop an event which was not yet started by checking cpuc->group_flag.

With this patch, you can now overcommit the PMU even with pinned system-wide events present and still get valid counts.

arch/x86/kernel/cpu/perf_event.c +27 -8
index fd4db0d..25dfd30 100644
@@ -105,6 +105,7 @@ struct cpu_hw_events {
 	int			enabled;
 
 	int			n_events;
+	int			n_events_trans;
 	int			n_added;
 	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 	u64			tags[X86_PMC_IDX_MAX];
@@ -591,7 +592,7 @@ void hw_perf_disable(void)
 	if (!cpuc->enabled)
 		return;
 
-	cpuc->n_added = 0;
+	cpuc->n_added = cpuc->n_events_trans = 0;
 	cpuc->enabled = 0;
 	barrier();
 
@@ -820,7 +821,7 @@ void hw_perf_enable(void)
 		 * apply assignment obtained either from
 		 * hw_perf_group_sched_in() or x86_pmu_enable()
 		 *
-		 * step1: save events moving to new counters
+		 * step1: stop-save old events moving to new counters
 		 * step2: reprogram moved events into new counters
 		 */
 		for (i = 0; i < n_running; i++) {
@@ -982,7 +983,7 @@ static int x86_pmu_enable(struct perf_event *event)
 
 out:
 	cpuc->n_events = n;
-	cpuc->n_added += n - n0;
+	cpuc->n_events_trans += n - n0;
 
 	return 0;
 }
@@ -1070,7 +1071,7 @@ static void x86_pmu_stop(struct perf_event *event)
 	struct hw_perf_event *hwc = &event->hw;
 	int idx = hwc->idx;
 
-	if (!__test_and_clear_bit(idx, cpuc->active_mask))
+	if (idx == -1 || !__test_and_clear_bit(idx, cpuc->active_mask))
 		return;
 
 	x86_pmu.disable(event);
@@ -1087,14 +1088,30 @@ static void x86_pmu_stop(struct perf_event *event)
 static void x86_pmu_disable(struct perf_event *event)
 {
 	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	bool no_trans;
 	int i;
 
-	x86_pmu_stop(event);
+	/*
+	 * when coming from a transaction, then the event has not
+	 * actually been scheduled in yet. It has only been collected
+	 * (collect_events), thus we cannot apply a full disable:
+	 * - put_constraints() assumes went thru schedule_events()
+	 * - x86_pmu_stop() assumes went thru x86_pmu_start() because
+	 *   maybe idx != -1 and thus we may overwrite another the wrong
+	 *   counter (e.g, pinned event).
+	 *
+	 * When called from a transaction, we need to un-collect the
+	 * event, i..e, remove the event from event_list[]
+	 */
+	no_trans = !(cpuc->group_flag & PERF_EVENT_TXN_STARTED);
+
+	if (no_trans)
+		x86_pmu_stop(event);
 
 	for (i = 0; i < cpuc->n_events; i++) {
 		if (event == cpuc->event_list[i]) {
 
-			if (x86_pmu.put_event_constraints)
+			if (no_trans && x86_pmu.put_event_constraints)
 				x86_pmu.put_event_constraints(cpuc, event);
 
 			while (++i < cpuc->n_events)
@@ -1104,7 +1121,8 @@ static void x86_pmu_disable(struct perf_event *event)
 			break;
 		}
 	}
-	perf_event_update_userpage(event);
+	if (no_trans)
+		perf_event_update_userpage(event);
 }
 
 static int x86_pmu_handle_irq(struct pt_regs *regs)
@@ -1418,7 +1436,8 @@ static int x86_pmu_commit_txn(const struct pmu *pmu)
 	 * will be used by hw_perf_enable()
 	 */
 	memcpy(cpuc->assign, assign, n*sizeof(int));
-
+	cpuc->n_added += cpuc->n_events_trans;
+	cpuc->n_events_trans = 0;
 	return 0;
 }
 
kernel/perf_event.c +0 -0
index 7e3bcf1..b0ccf4b 100644

kernel/perf_event.c group_sched_in(struct perf_event *group_event, · line 659

Quoted 9 lines
kernel/perf_event.c +4 -1
@@ -652,8 +652,11 @@ group_sched_in(struct perf_event *group_event,
 	if (txn)
 		pmu->start_txn(pmu);
 
-	if (event_sched_in(group_event, cpuctx, ctx))
+	if (event_sched_in(group_event, cpuctx, ctx)) {
+		if (txn)
+			pmu->cancel_txn(pmu);
 		return -EAGAIN;
+	}

Can't imagine I forgot to call cancel_txn here.
Thanks very much.

Quoted 13 lines
kernel/perf_event.c +0 -2
 	/*
 	 * Schedule in siblings as one group (if any):
@@ -675,8 +678,6 @@ group_sched_in(struct perf_event *group_event,
 	}
 
 group_error:
-	if (txn)
-		pmu->cancel_txn(pmu);
 
 	/*
 	 * Groups can be scheduled in as one unit only, so undo any

kernel/perf_event.c group_error: · line 698

Quoted 9 lines
kernel/perf_event.c +3 -0
@@ -689,6 +690,9 @@ group_error:
 	}
 	event_sched_out(group_event, cpuctx, ctx);
 
+	if (txn)
+		pmu->cancel_txn(pmu);
+
 	return -EAGAIN;
 }
 

Looks good to me.

Thanks,
Lin Ming