From: Kan Liang <[email protected]> For perf record frequency mode, the initial sample_period is 1. That's because perf doesn't know what period should be set. It uses the minimum period 1 as the first period. It will trigger an interrupt soon. Then there will be enough data to calculate the period for the given frequency. But too many very short period like 1 may cause various problems and increase the overhead. It's better to limit the 1 period to just the first several period setting.
However, for some workload, 1 period is frequently set. For example, perf record a busy loop for 10 seconds. perf record ./finity_busy_loop.sh 10 while [ "A" != "B" ] do date > /dev/null done Period was changed 150503 times in 10 seconds. 22.5% (33861 times) of the period is set to 1. That's because, in the inherit_event, the period for child event is inherit from parent's parent's event, which is usually the default sample_period 1. Each child event has to recaculate the period from 1 everytime. That brings high overhead. This patch keeps the sample period average in original parent event. Each new child event can use it as its initial sample period. Adding a ori_parent in struct perf_event to help child event access the original parent. For each new child event, the parent event refcount++. Parent will not go away until all children go away. So the stored pointer is safe to be accessed. After applying this patch, the 1 period rate reduces to 0.1%. Signed-off-by: Kan Liang <[email protected]> --- include/linux/perf_event.h | 4 ++++ kernel/events/core.c | 22 ++++++++++++++++++++-- 2 files changed, 24 insertions(+), 2 deletions(-) diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h index 486e84c..b328617 100644 --- a/include/linux/perf_event.h +++ b/include/linux/perf_event.h @@ -403,6 +403,10 @@ struct perf_event { struct list_head child_list; struct perf_event *parent; + /* Average Sample period in the original parent event */ + struct perf_event *ori_parent; + local64_t avg_sample_period; + int oncpu; int cpu; diff --git a/kernel/events/core.c b/kernel/events/core.c index af0a5ba..a8be6d3 100644 --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -2795,7 +2795,8 @@ static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bo { struct hw_perf_event *hwc = &event->hw; s64 period, sample_period; - s64 delta; + s64 delta, avg_period; + struct perf_event *head_event = event->ori_parent; period = perf_calculate_period(event, nsec, count); @@ -2809,6 +2810,9 @@ static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bo hwc->sample_period = sample_period; + avg_period = (local64_read(&head_event->avg_sample_period) + sample_period) / 2; + local64_set(&head_event->avg_sample_period, avg_period); + if (local64_read(&hwc->period_left) > 8*sample_period) { if (disable) event->pmu->stop(event, PERF_EF_UPDATE); @@ -6996,6 +7000,10 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu, event->oncpu = -1; event->parent = parent_event; + if (parent_event) + event->ori_parent = parent_event->ori_parent; + else + event->ori_parent = event; event->ns = get_pid_ns(task_active_pid_ns(current)); event->id = atomic64_inc_return(&perf_event_id); @@ -7030,8 +7038,16 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu, hwc = &event->hw; hwc->sample_period = attr->sample_period; - if (attr->freq && attr->sample_freq) + if (attr->freq && attr->sample_freq) { hwc->sample_period = 1; + if (parent_event) { + struct perf_event *head_event = event->ori_parent; + + hwc->sample_period = local64_read(&head_event->avg_sample_period); + } else { + local64_set(&event->avg_sample_period, hwc->sample_period); + } + } hwc->last_period = hwc->sample_period; local64_set(&hwc->period_left, hwc->sample_period); @@ -7904,7 +7920,9 @@ inherit_event(struct perf_event *parent_event, if (parent_event->attr.freq) { u64 sample_period = parent_event->hw.sample_period; struct hw_perf_event *hwc = &child_event->hw; + struct perf_event *head_event = child_event->ori_parent; + sample_period = local64_read(&head_event->avg_sample_period); hwc->sample_period = sample_period; hwc->last_period = sample_period; -- 1.8.3.2 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
