Skip to contentSkip to conversation

perf_counter: don't count scheduler ticks as context switches

PATCH In review Paul Mackerras 1 participant 2 messages May 11, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf_counter: don't count scheduler ticks as context switches

The context-switch software counter gives inflated values at present because each scheduler tick and each process-wide counter enable/disable prctl gets counted as a context switch. This happens because perf_counter_task_tick, perf_counter_task_disable and perf_counter_task_enable all call perf_counter_task_sched_out, which calls perf_swcounter_event to record a context switch event.

This fixes it by introducing a variant of perf_counter_task_sched_out with two underscores in front for internal use within the perf_counter code, and makes perf_counter_task_{tick,disable,enable} call it. This variant doesn't record a context switch event, and takes a struct perf_counter_context *. This adds the new variant rather than changing the behaviour or interface of perf_counter_task_sched_out because that is called from other code.

---
 kernel/perf_counter.c |   16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)
kernel/perf_counter.c +11 -5
index a5bdc93..7373b96 100644
@@ -837,6 +837,14 @@ void perf_counter_task_sched_out(struct task_struct *task, int cpu)
 	cpuctx->task_ctx = NULL;
 }
 
+static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
+{
+	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
+
+	__perf_counter_sched_out(ctx, cpuctx);
+	cpuctx->task_ctx = NULL;
+}
+
 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
 {
 	__perf_counter_sched_out(&cpuctx->ctx, cpuctx);
@@ -943,15 +951,13 @@ int perf_counter_task_disable(void)
 	struct perf_counter *counter;
 	unsigned long flags;
 	u64 perf_flags;
-	int cpu;
 
 	if (likely(!ctx->nr_counters))
 		return 0;
 
 	local_irq_save(flags);
-	cpu = smp_processor_id();
 
-	perf_counter_task_sched_out(curr, cpu);
+	__perf_counter_task_sched_out(ctx);
 
 	spin_lock(&ctx->lock);
 
@@ -989,7 +995,7 @@ int perf_counter_task_enable(void)
 	local_irq_save(flags);
 	cpu = smp_processor_id();
 
-	perf_counter_task_sched_out(curr, cpu);
+	__perf_counter_task_sched_out(ctx);
 
 	spin_lock(&ctx->lock);
 
@@ -1054,7 +1060,7 @@ void perf_counter_task_tick(struct task_struct *curr, int cpu)
 	ctx = &curr->perf_counter_ctx;
 
 	perf_counter_cpu_sched_out(cpuctx);
-	perf_counter_task_sched_out(curr, cpu);
+	__perf_counter_task_sched_out(ctx);
 
 	rotate_ctx(&cpuctx->ctx);
 	rotate_ctx(ctx);
#1 lore raw
Commit-ID:  a08b159fc243dbfe415250466d24cfc5010deee5
Gitweb:     http://git.kernel.org/tip/a08b159fc243dbfe415250466d24cfc5010deee5
Author:     Paul Mackerras <[email protected]>
AuthorDate: Mon, 11 May 2009 15:46:10 +1000
Committer:  Ingo Molnar <[email protected]>
CommitDate: Mon, 11 May 2009 12:10:53 +0200

perf_counter: don't count scheduler ticks as context switches

The context-switch software counter gives inflated values at present because each scheduler tick and each process-wide counter enable/disable prctl gets counted as a context switch.

This happens because perf_counter_task_tick, perf_counter_task_disable and perf_counter_task_enable all call perf_counter_task_sched_out, which calls perf_swcounter_event to record a context switch event.

This fixes it by introducing a variant of perf_counter_task_sched_out with two underscores in front for internal use within the perf_counter code, and makes perf_counter_task_{tick,disable,enable} call it. This variant doesn't record a context switch event, and takes a struct perf_counter_context *. This adds the new variant rather than changing the behaviour or interface of perf_counter_task_sched_out because that is called from other code.

[ Impact: fix inflated context-switch event counts ]

Cc Peter Zijlstra <[email protected]>
Cc Corey Ashford <[email protected]>

LKML-Reference: <[email protected]>

---
 kernel/perf_counter.c |   16 +++++++++++-----
 1 files changed, 11 insertions(+), 5 deletions(-)
kernel/perf_counter.c +11 -5
index a5bdc93..7373b96 100644
@@ -837,6 +837,14 @@ void perf_counter_task_sched_out(struct task_struct *task, int cpu)
 	cpuctx->task_ctx = NULL;
 }
 
+static void __perf_counter_task_sched_out(struct perf_counter_context *ctx)
+{
+	struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
+
+	__perf_counter_sched_out(ctx, cpuctx);
+	cpuctx->task_ctx = NULL;
+}
+
 static void perf_counter_cpu_sched_out(struct perf_cpu_context *cpuctx)
 {
 	__perf_counter_sched_out(&cpuctx->ctx, cpuctx);
@@ -943,15 +951,13 @@ int perf_counter_task_disable(void)
 	struct perf_counter *counter;
 	unsigned long flags;
 	u64 perf_flags;
-	int cpu;
 
 	if (likely(!ctx->nr_counters))
 		return 0;
 
 	local_irq_save(flags);
-	cpu = smp_processor_id();
 
-	perf_counter_task_sched_out(curr, cpu);
+	__perf_counter_task_sched_out(ctx);
 
 	spin_lock(&ctx->lock);
 
@@ -989,7 +995,7 @@ int perf_counter_task_enable(void)
 	local_irq_save(flags);
 	cpu = smp_processor_id();
 
-	perf_counter_task_sched_out(curr, cpu);
+	__perf_counter_task_sched_out(ctx);
 
 	spin_lock(&ctx->lock);
 
@@ -1054,7 +1060,7 @@ void perf_counter_task_tick(struct task_struct *curr, int cpu)
 	ctx = &curr->perf_counter_ctx;
 
 	perf_counter_cpu_sched_out(cpuctx);
-	perf_counter_task_sched_out(curr, cpu);
+	__perf_counter_task_sched_out(ctx);
 
 	rotate_ctx(&cpuctx->ctx);
 	rotate_ctx(ctx);