> On Oct 8, 2019, at 2:35 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Wed, Oct 02, 2019 at 11:43:17PM -0700, Song Liu wrote:
>> This is a rare corner case, but it does happen:
>>
>> In perf_rotate_context(), when the first cpu flexible event fail to
>> schedule, cpu_rotate is 1, while cpu_event is NULL.
>
> You're failing to explain how we get here in the first place.
On Intel CPU, it could be trigger by something like:
perf stat -e ref-cycles:D,ref-cycles,cycles -C 5 -I 1000
I will add this to v2.
>
>> diff --git a/kernel/events/core.c b/kernel/events/core.c
>> index 4655adbbae10..50021735f367 100644
>> --- a/kernel/events/core.c
>> +++ b/kernel/events/core.c
>> @@ -3775,6 +3775,13 @@ static void rotate_ctx(struct perf_event_context *ctx, struct perf_event *event)
>> if (ctx->rotate_disable)
>> return;
>>
>> + /* if no event specified, try to rotate the first event */
>> + if (!event)
>> + event = rb_entry_safe(rb_first(&ctx->flexible_groups.tree),
>> + typeof(*event), group_node);
>> + if (!event)
>> + return;
>> +
>> perf_event_groups_delete(&ctx->flexible_groups, event);
>> perf_event_groups_insert(&ctx->flexible_groups, event);
>> }
>
> I don't think that is a very nice solution; would not something simpler
> like this be sufficient? (although possible we should rename that
> function now).
Yes, this is much cleaner. Thanks!
Let me submit v2 based on this.
Thanks,
Song
>
> And it needs more comments.
>
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 4655adbbae10..772a9854eb3a 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -3782,8 +3797,17 @@ static void rotate_ctx(struct perf_event_context *ctx, struct perf_event *event)
> static inline struct perf_event *
> ctx_first_active(struct perf_event_context *ctx)
> {
> - return list_first_entry_or_null(&ctx->flexible_active,
> - struct perf_event, active_list);
> + struct perf_event *event;
> +
> + event = list_first_entry_or_null(&ctx->flexible_active,
> + struct perf_event, active_list);
> +
> + if (!event) {
> + event = rb_entry_safe(rb_first(&ctx->flexible_groups.tree),
> + typeof(*event), group_node);
> + }
> +
> + return event;
> }
>
> static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
