perf_counter: More aggressive frequency adjustment

Also employ the overflow handler to adjust the frequency, this results
in a stable frequency in about 40~50 samples, instead of that many ticks.

This also means we can start sampling at a sample period of 1 without
running head-first into the throttle.

It relies on sched_clock() to accurately measure the time difference
between the overflow NMIs.

Signed-off-by: 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>

3 files changed, 92 insertions, 44 deletions

diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index 49f258537cbf..240ca5630632 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -696,10 +696,11 @@ static int __hw_perf_counter_init(struct perf_counter *counter)
         if (!attr->exclude_kernel)
                 hwc->config |= ARCH_PERFMON_EVENTSEL_OS;
 
-        if (!hwc->sample_period)
+        if (!hwc->sample_period) {
                 hwc->sample_period = x86_pmu.max_period;
+                atomic64_set(&hwc->period_left, hwc->sample_period);
+        }
 
-        atomic64_set(&hwc->period_left, hwc->sample_period);
         counter->destroy = hw_perf_counter_destroy;
 
         /*
diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index 3586df840f69..282d8cc48980 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -371,6 +371,7 @@ struct hw_perf_counter {
 
         u64                             freq_count;
         u64                             freq_interrupts;
+        u64                             freq_stamp;
 #endif
 };
 
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 5eacaaf3f9cd..51c571ee4d0b 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1184,13 +1184,33 @@ static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
 static void perf_log_throttle(struct perf_counter *counter, int enable);
 static void perf_log_period(struct perf_counter *counter, u64 period);
 
-static void perf_adjust_freq(struct perf_counter_context *ctx)
+static void perf_adjust_period(struct perf_counter *counter, u64 events)
+{
+        struct hw_perf_counter *hwc = &counter->hw;
+        u64 period, sample_period;
+        s64 delta;
+
+        events *= hwc->sample_period;
+        period = div64_u64(events, counter->attr.sample_freq);
+
+        delta = (s64)(period - hwc->sample_period);
+        delta = (delta + 7) / 8; /* low pass filter */
+
+        sample_period = hwc->sample_period + delta;
+
+        if (!sample_period)
+                sample_period = 1;
+
+        perf_log_period(counter, sample_period);
+
+        hwc->sample_period = sample_period;
+}
+
+static void perf_ctx_adjust_freq(struct perf_counter_context *ctx)
 {
         struct perf_counter *counter;
         struct hw_perf_counter *hwc;
-        u64 interrupts, sample_period;
-        u64 events, period, freq;
-        s64 delta;
+        u64 interrupts, freq;
 
         spin_lock(&ctx->lock);
         list_for_each_entry(counter, &ctx->counter_list, list_entry) {
@@ -1202,6 +1222,9 @@ static void perf_adjust_freq(struct perf_counter_context *ctx)
                 interrupts = hwc->interrupts;
                 hwc->interrupts = 0;
 
+                /*
+                 * unthrottle counters on the tick
+                 */
                 if (interrupts == MAX_INTERRUPTS) {
                         perf_log_throttle(counter, 1);
                         counter->pmu->unthrottle(counter);
@@ -1211,6 +1234,9 @@ static void perf_adjust_freq(struct perf_counter_context *ctx)
                 if (!counter->attr.freq || !counter->attr.sample_freq)
                         continue;
 
+                /*
+                 * if the specified freq < HZ then we need to skip ticks
+                 */
                 if (counter->attr.sample_freq < HZ) {
                         freq = counter->attr.sample_freq;
 
@@ -1226,20 +1252,20 @@ static void perf_adjust_freq(struct perf_counter_context *ctx)
                 } else
                         freq = HZ;
 
-                events = freq * interrupts * hwc->sample_period;
-                period = div64_u64(events, counter->attr.sample_freq);
-
-                delta = (s64)(1 + period - hwc->sample_period);
-                delta >>= 1;
-
-                sample_period = hwc->sample_period + delta;
-
-                if (!sample_period)
-                        sample_period = 1;
+                perf_adjust_period(counter, freq * interrupts);
 
-                perf_log_period(counter, sample_period);
-
-                hwc->sample_period = sample_period;
+                /*
+                 * In order to avoid being stalled by an (accidental) huge
+                 * sample period, force reset the sample period if we didn't
+                 * get any events in this freq period.
+                 */
+                if (!interrupts) {
+                        perf_disable();
+                        counter->pmu->disable(counter);
+                        atomic_set(&hwc->period_left, 0);
+                        counter->pmu->enable(counter);
+                        perf_enable();
+                }
         }
         spin_unlock(&ctx->lock);
 }
@@ -1279,9 +1305,9 @@ void perf_counter_task_tick(struct task_struct *curr, int cpu)
         cpuctx = &per_cpu(perf_cpu_context, cpu);
         ctx = curr->perf_counter_ctxp;
 
-        perf_adjust_freq(&cpuctx->ctx);
+        perf_ctx_adjust_freq(&cpuctx->ctx);
         if (ctx)
-                perf_adjust_freq(ctx);
+                perf_ctx_adjust_freq(ctx);
 
         perf_counter_cpu_sched_out(cpuctx);
         if (ctx)
@@ -1647,10 +1673,10 @@ static int perf_counter_period(struct perf_counter *counter, u64 __user *arg)
 
                 counter->attr.sample_freq = value;
         } else {
+                perf_log_period(counter, value);
+
                 counter->attr.sample_period = value;
                 counter->hw.sample_period = value;
-
-                perf_log_period(counter, value);
         }
 unlock:
         spin_unlock_irq(&ctx->lock);
@@ -2853,35 +2879,41 @@ void __perf_counter_mmap(struct vm_area_struct *vma)
  * event flow.
  */
 
+struct freq_event {
+        struct perf_event_header        header;
+        u64                             time;
+        u64                             id;
+        u64                             period;
+};
+
 static void perf_log_period(struct perf_counter *counter, u64 period)
 {
         struct perf_output_handle handle;
+        struct freq_event event;
         int ret;
 
-        struct {
-                struct perf_event_header        header;
-                u64                             time;
-                u64                             id;
-                u64                             period;
-        } freq_event = {
+        if (counter->hw.sample_period == period)
+                return;
+
+        if (counter->attr.sample_type & PERF_SAMPLE_PERIOD)
+                return;
+
+        event = (struct freq_event) {
                 .header = {
                         .type = PERF_EVENT_PERIOD,
                         .misc = 0,
-                        .size = sizeof(freq_event),
+                        .size = sizeof(event),
                 },
                 .time = sched_clock(),
                 .id = counter->id,
                 .period = period,
         };
 
-        if (counter->hw.sample_period == period)
-                return;
-
-        ret = perf_output_begin(&handle, counter, sizeof(freq_event), 0, 0);
+        ret = perf_output_begin(&handle, counter, sizeof(event), 1, 0);
         if (ret)
                 return;
 
-        perf_output_put(&handle, freq_event);
+        perf_output_put(&handle, event);
         perf_output_end(&handle);
 }
 
@@ -2923,15 +2955,16 @@ int perf_counter_overflow(struct perf_counter *counter,
 {
         int events = atomic_read(&counter->event_limit);
         int throttle = counter->pmu->unthrottle != NULL;
+        struct hw_perf_counter *hwc = &counter->hw;
         int ret = 0;
 
         if (!throttle) {
-                counter->hw.interrupts++;
+                hwc->interrupts++;
         } else {
-                if (counter->hw.interrupts != MAX_INTERRUPTS) {
-                        counter->hw.interrupts++;
-                        if (HZ*counter->hw.interrupts > (u64)sysctl_perf_counter_limit) {
-                                counter->hw.interrupts = MAX_INTERRUPTS;
+                if (hwc->interrupts != MAX_INTERRUPTS) {
+                        hwc->interrupts++;
+                        if (HZ * hwc->interrupts > (u64)sysctl_perf_counter_limit) {
+                                hwc->interrupts = MAX_INTERRUPTS;
                                 perf_log_throttle(counter, 0);
                                 ret = 1;
                         }
@@ -2945,6 +2978,16 @@ int perf_counter_overflow(struct perf_counter *counter,
                 }
         }
 
+        if (counter->attr.freq) {
+                u64 now = sched_clock();
+                s64 delta = now - hwc->freq_stamp;
+
+                hwc->freq_stamp = now;
+
+                if (delta > 0 && delta < TICK_NSEC)
+                        perf_adjust_period(counter, NSEC_PER_SEC / (int)delta);
+        }
+
         /*
          * XXX event_limit might not quite work as expected on inherited
          * counters
@@ -3379,7 +3422,6 @@ static const struct pmu *tp_perf_counter_init(struct perf_counter *counter)
                 return NULL;
 
         counter->destroy = tp_perf_counter_destroy;
-        counter->hw.sample_period = counter->attr.sample_period;
 
         return &perf_ops_generic;
 }
@@ -3483,10 +3525,11 @@ perf_counter_alloc(struct perf_counter_attr *attr,
         pmu = NULL;
 
         hwc = &counter->hw;
+        hwc->sample_period = attr->sample_period;
         if (attr->freq && attr->sample_freq)
-                hwc->sample_period = div64_u64(TICK_NSEC, attr->sample_freq);
-        else
-                hwc->sample_period = attr->sample_period;
+                hwc->sample_period = 1;
+
+        atomic64_set(&hwc->period_left, hwc->sample_period);
 
         /*
          * we currently do not support PERF_SAMPLE_GROUP on inherited counters
@@ -3687,6 +3730,9 @@ inherit_counter(struct perf_counter *parent_counter,
         else
                 child_counter->state = PERF_COUNTER_STATE_OFF;
 
+        if (parent_counter->attr.freq)
+                child_counter->hw.sample_period = parent_counter->hw.sample_period;
+
         /*
          * Link it up in the child's context:
          */