perf_counter: Accurate period data

We currently log hw.sample_period for PERF_SAMPLE_PERIOD, however this is
incorrect. When we adjust the period, it will only take effect the next
cycle but report it for the current cycle. So when we adjust the period
for every cycle, we're always wrong.

Solve this by keeping track of the last_period.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

4 files changed, 28 insertions, 11 deletions

diff --git a/arch/powerpc/kernel/perf_counter.c b/arch/powerpc/kernel/perf_counter.c
index 5e0bf399c433..4990ce2e5f08 100644
--- a/arch/powerpc/kernel/perf_counter.c
+++ b/arch/powerpc/kernel/perf_counter.c
@@ -767,6 +767,7 @@ static void power_pmu_unthrottle(struct perf_counter *counter)
 	perf_disable();
 	power_pmu_read(counter);
 	left = counter->hw.sample_period;
+	counter->hw.last_period = left;
 	val = 0;
 	if (left < 0x80000000L)
 		val = 0x80000000L - left;
@@ -937,7 +938,8 @@ const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
 
 	counter->hw.config = events[n];
 	counter->hw.counter_base = cflags[n];
-	atomic64_set(&counter->hw.period_left, counter->hw.sample_period);
+	counter->hw.last_period = counter->hw.sample_period;
+	atomic64_set(&counter->hw.period_left, counter->hw.last_period);
 
 	/*
 	 * See if we need to reserve the PMU.
@@ -1002,8 +1004,9 @@ static void record_and_restart(struct perf_counter *counter, long val,
 	 */
 	if (record) {
 		struct perf_sample_data data = {
-			.regs = regs,
-			.addr = 0,
+			.regs	= regs,
+			.addr	= 0,
+			.period	= counter->hw.last_period,
 		};
 
 		if (counter->attr.sample_type & PERF_SAMPLE_ADDR) {
diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index 82a23d487f92..57ae1bec81be 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -698,6 +698,7 @@ static int __hw_perf_counter_init(struct perf_counter *counter)
 
 	if (!hwc->sample_period) {
 		hwc->sample_period = x86_pmu.max_period;
+		hwc->last_period = hwc->sample_period;
 		atomic64_set(&hwc->period_left, hwc->sample_period);
 	}
 
@@ -880,12 +881,14 @@ x86_perf_counter_set_period(struct perf_counter *counter,
 	if (unlikely(left <= -period)) {
 		left = period;
 		atomic64_set(&hwc->period_left, left);
+		hwc->last_period = period;
 		ret = 1;
 	}
 
 	if (unlikely(left <= 0)) {
 		left += period;
 		atomic64_set(&hwc->period_left, left);
+		hwc->last_period = period;
 		ret = 1;
 	}
 	/*
@@ -1257,9 +1260,12 @@ static int amd_pmu_handle_irq(struct pt_regs *regs)
 		if (val & (1ULL << (x86_pmu.counter_bits - 1)))
 			continue;
 
-		/* counter overflow */
-		handled = 1;
-		inc_irq_stat(apic_perf_irqs);
+		/*
+		 * counter overflow
+		 */
+		handled		= 1;
+		data.period	= counter->hw.last_period;
+
 		if (!x86_perf_counter_set_period(counter, hwc, idx))
 			continue;
 
@@ -1267,6 +1273,9 @@ static int amd_pmu_handle_irq(struct pt_regs *regs)
 			amd_pmu_disable_counter(hwc, idx);
 	}
 
+	if (handled)
+		inc_irq_stat(apic_perf_irqs);
+
 	return handled;
 }
 
diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index d8c0eb480f9a..5b966472b458 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -366,6 +366,7 @@ struct hw_perf_counter {
 	};
 	atomic64_t			prev_count;
 	u64				sample_period;
+	u64				last_period;
 	atomic64_t			period_left;
 	u64				interrupts;
 
@@ -606,8 +607,9 @@ extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 extern void perf_counter_update_userpage(struct perf_counter *counter);
 
 struct perf_sample_data {
-	struct pt_regs	*regs;
-	u64		addr;
+	struct pt_regs		*regs;
+	u64			addr;
+	u64			period;
 };
 
 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 4fe85e804f43..8b89b40bd0f0 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -2495,7 +2495,7 @@ static void perf_counter_output(struct perf_counter *counter, int nmi,
 		perf_output_put(&handle, cpu_entry);
 
 	if (sample_type & PERF_SAMPLE_PERIOD)
-		perf_output_put(&handle, counter->hw.sample_period);
+		perf_output_put(&handle, data->period);
 
 	/*
 	 * XXX PERF_SAMPLE_GROUP vs inherited counters seems difficult.
@@ -3040,11 +3040,13 @@ static void perf_swcounter_set_period(struct perf_counter *counter)
 	if (unlikely(left <= -period)) {
 		left = period;
 		atomic64_set(&hwc->period_left, left);
+		hwc->last_period = period;
 	}
 
 	if (unlikely(left <= 0)) {
 		left += period;
 		atomic64_add(period, &hwc->period_left);
+		hwc->last_period = period;
 	}
 
 	atomic64_set(&hwc->prev_count, -left);
@@ -3086,8 +3088,9 @@ static void perf_swcounter_overflow(struct perf_counter *counter,
 				    int nmi, struct pt_regs *regs, u64 addr)
 {
 	struct perf_sample_data data = {
-		.regs = regs,
-		.addr = addr,
+		.regs	= regs,
+		.addr	= addr,
+		.period	= counter->hw.last_period,
 	};
 
 	perf_swcounter_update(counter);