lkml.org 
[lkml]   [2020]   [Mar]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] perf/core: Fix endless multiplex timer
On Tue, Mar 03, 2020 at 12:28:19PM -0800, kan.liang@linux.intel.com wrote:
> From: Kan Liang <kan.liang@linux.intel.com>
>
> A lot of time are spent in writing uncore MSRs even though no perf is
> running.
>
> 4.66% swapper [kernel.kallsyms] [k] native_write_msr
> |
> --4.56%--native_write_msr
> |
> |--1.68%--snbep_uncore_msr_enable_box
> | perf_mux_hrtimer_handler
> | __hrtimer_run_queues
> | hrtimer_interrupt
> | smp_apic_timer_interrupt
> | apic_timer_interrupt
> | cpuidle_enter_state
> | cpuidle_enter
> | do_idle
> | cpu_startup_entry
> | start_kernel
> | secondary_startup_64
>
> The root cause is that multiplex timer was not stopped when perf stat
> finished.
> Current perf relies on rotate_necessary to determine whether the
> multiplex timer should be stopped. The variable only be reset in
> ctx_sched_out(), which is not enough for system-wide event.
> Perf stat invokes PERF_EVENT_IOC_DISABLE to stop system-wide event
> before closing it.
> perf_ioctl()
> perf_event_disable()
> event_sched_out()
> The rotate_necessary will never be reset.
>
> The issue is a generic issue, not just impact the uncore.
>
> Check whether we had been multiplexing. If yes, reset rotate_necessary
> for the last active event in __perf_event_disable().
>
> Fixes: fd7d55172d1e ("perf/cgroups: Don't rotate events for cgroups unnecessarily")
> Reported-by: Andi Kleen <ak@linux.intel.com>
> Reviewed-by: Andi Kleen <ak@linux.intel.com>
> Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
> ---
> kernel/events/core.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 3f1f77de7247..50688de56181 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -2242,6 +2242,16 @@ static void __perf_event_disable(struct perf_event *event,
> update_cgrp_time_from_event(event);
> }
>
> + /*
> + * If we had been multiplexing,
> + * stop the rotations for the last active event.
> + * Only need to check system wide events.
> + * For task events, it will be checked in ctx_sched_out().
> + */
> + if ((cpuctx->ctx.nr_events != cpuctx->ctx.nr_active) &&
> + (cpuctx->ctx.nr_active == 1))
> + cpuctx->ctx.rotate_necessary = 0;
> +
> if (event == event->group_leader)
> group_sched_out(event, cpuctx, ctx);
> else


I'm thinking this is wrong.

That is, yes, this fixes the observed problem, but it also misses at
least one other site. Which seems to suggest we ought to take a
different approach.

But even with that; I wonder if the actual condition isn't wrong.
Suppose the event was exclusive, and other events weren't scheduled
because of that. Then you disable the one exclusive event _and_ kill
rotation, so then nothing else will ever get on.

So what I think was supposed to happen is rotation killing itself;
rotation will schedule out the context -- which will clear the flag, and
then schedule the thing back in -- which will set the flag again when
needed.

Now, that isn't happening, and I think I see why, because when we drop
to !nr_active, we terminate ctx_sched_out() before we get to clearing
the flag, oops!

So how about something like this?

---
diff --git a/kernel/events/core.c b/kernel/events/core.c
index e453589da97c..7947bd3271a9 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -2182,6 +2182,7 @@ __perf_remove_from_context(struct perf_event *event,

if (!ctx->nr_events && ctx->is_active) {
ctx->is_active = 0;
+ ctx->rotate_necessary = 0;
if (ctx->task) {
WARN_ON_ONCE(cpuctx->task_ctx != ctx);
cpuctx->task_ctx = NULL;
@@ -3074,15 +3075,15 @@ static void ctx_sched_out(struct perf_event_context *ctx,

is_active ^= ctx->is_active; /* changed bits */

- if (!ctx->nr_active || !(is_active & EVENT_ALL))
- return;
-
/*
* If we had been multiplexing, no rotations are necessary, now no events
* are active.
*/
ctx->rotate_necessary = 0;

+ if (!ctx->nr_active || !(is_active & EVENT_ALL))
+ return;
+
perf_pmu_disable(ctx->pmu);
if (is_active & EVENT_PINNED) {
list_for_each_entry_safe(event, tmp, &ctx->pinned_active, active_list)
\
 
 \ /
  Last update: 2020-03-03 22:09    [W:0.675 / U:0.008 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site