On 4/29/2019 11:04 AM, Mark Rutland wrote:
On Mon, Apr 29, 2019 at 07:44:02AM -0700, [email protected] wrote:From: Kan Liang <[email protected]> When counting system-wide events and cgroup events simultaneously, the value of system-wide events are miscounting. For example, perf stat -e cycles,instructions -e cycles,instructions -G cgroup1,cgroup1,cgroup2,cgroup2 -a -e cycles,instructions -I 1000 1.096265502 12,375,398,872 cycles cgroup1 1.096265502 8,396,184,503 instructions cgroup1 # 0.10 insn per cycle 1.096265502 109,609,027,112 cycles cgroup2 1.096265502 11,533,690,148 instructions cgroup2 # 0.14 insn per cycle 1.096265502 121,672,937,058 cycles 1.096265502 19,331,496,727 instructions # 0.24 insn per cycle The events are identical events for system-wide and cgroup. The value of system-wide events is less than the sum of cgroup events, which is wrong. Both system-wide and cgroup are per-cpu. They share the same cpuctx groups, cpuctx->flexible_groups/pinned_groups. In context switch, cgroup switch tries to schedule all the events in the cpuctx groups. The unmatched cgroup events can be filtered by its event->cgrp. However, system-wide events, which event->cgrp is NULL, are unconditionally switched, which causes miscounting.Why exactly does that cause mis-counting? Are the system-wide events not switched back in? Or filtered erroneously?
The small period between the prev cgroup sched_out and the new cgroup sched_in is missed for system-wide events.
Current code mistakenly sched_out of system-wide events during that period.
Introduce cgrp_switch in cpuctx to indicate the cgroup context switch. If it's system-wide event in context switch, don't try to switch it. Fixes: e5d1367f17ba ("perf: Add cgroup support") Signed-off-by: Kan Liang <[email protected]> --- include/linux/perf_event.h | 1 + kernel/events/core.c | 30 ++++++++++++++++++++++++++++-- 2 files changed, 29 insertions(+), 2 deletions(-) diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h index e47ef76..039e2f2 100644 --- a/include/linux/perf_event.h +++ b/include/linux/perf_event.h @@ -795,6 +795,7 @@ struct perf_cpu_context { #ifdef CONFIG_CGROUP_PERF struct perf_cgroup *cgrp; struct list_head cgrp_cpuctx_entry; + unsigned int cgrp_switch :1; #endifstruct list_head sched_cb_entry;diff --git a/kernel/events/core.c b/kernel/events/core.c index dc7dead..388dd42 100644 --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -809,6 +809,7 @@ static void perf_cgroup_switch(struct task_struct *task, int mode)perf_ctx_lock(cpuctx, cpuctx->task_ctx);perf_pmu_disable(cpuctx->ctx.pmu); + cpuctx->cgrp_switch = true;if (mode & PERF_CGROUP_SWOUT) {cpu_ctx_sched_out(cpuctx, EVENT_ALL); @@ -832,6 +833,7 @@ static void perf_cgroup_switch(struct task_struct *task, int mode) &cpuctx->ctx); cpu_ctx_sched_in(cpuctx, EVENT_ALL, task); } + cpuctx->cgrp_switch = false; perf_pmu_enable(cpuctx->ctx.pmu); perf_ctx_unlock(cpuctx, cpuctx->task_ctx); } @@ -2944,13 +2946,25 @@ static void ctx_sched_out(struct perf_event_context *ctx,perf_pmu_disable(ctx->pmu);if (is_active & EVENT_PINNED) { - list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list) + list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list) { +#ifdef CONFIG_CGROUP_PERF + /* Don't sched system-wide event when cgroup context switch */ + if (cpuctx->cgrp_switch && !event->cgrp) + continue; +#endifThis pattern is duplicated several times, and should probably be a helper like: static bool skip_cgroup_switch(struct perf_cpu_context *cpuctx, struct perf_event *event); { return IS_ENABLED(CONFIG_CGROUP_PERF) && cpuctx->cgrp_switch && !event->cgrp; } ... allowing the above to be an unconditional: if (skip_cgroup_switch(cpuctx, event)) continue; ... and likewise for the other cases.
I will change it in V2. Thanks, Kan
Thanks, Mark.group_sched_out(event, cpuctx, ctx); + } }if (is_active & EVENT_FLEXIBLE) {- list_for_each_entry_safe(event, tmp, &ctx->flexible_active, active_list) + list_for_each_entry_safe(event, tmp, &ctx->flexible_active, active_list) { +#ifdef CONFIG_CGROUP_PERF + /* Don't sched system-wide event when cgroup context switch */ + if (cpuctx->cgrp_switch && !event->cgrp) + continue; +#endif group_sched_out(event, cpuctx, ctx); + } } perf_pmu_enable(ctx->pmu); } @@ -3280,6 +3294,12 @@ static int pinned_sched_in(struct perf_event *event, void *data) if (event->state <= PERF_EVENT_STATE_OFF) return 0;+#ifdef CONFIG_CGROUP_PERF+ /* Don't sched system-wide event when cgroup context switch */ + if (sid->cpuctx->cgrp_switch && !event->cgrp) + return 0; +#endif + if (!event_filter_match(event)) return 0;@@ -3305,6 +3325,12 @@ static int flexible_sched_in(struct perf_event *event, void *data)if (event->state <= PERF_EVENT_STATE_OFF) return 0;+#ifdef CONFIG_CGROUP_PERF+ /* Don't sched system-wide event when cgroup context switch */ + if (sid->cpuctx->cgrp_switch && !event->cgrp) + return 0; +#endif + if (!event_filter_match(event)) return 0;--2.7.4
