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>

2 files changed, 18 insertions, 6 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;
 }