lkml.org 
[lkml]   [2011]   [Nov]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRE: [PATCH v2 5/5] perf: Enable applicable siblings when group leader is enable-on-exec
Date
> ________________________________________
> From: Peter Zijlstra [a.p.zijlstra@chello.nl]
> Sent: Tuesday, November 22, 2011 9:45 PM
> To: Zhu, DengCheng
> Cc: Barzilay, Eyal; Fortuna, Zenon; Paul Mackerras; Ingo Molnar; Arnaldo Carvalho de Melo; ralf@linux-mips.org; LKML
> Subject: RE: [PATCH v2 5/5] perf: Enable applicable siblings when group leader is enable-on-exec
>
> On Tue, 2011-11-22 at 13:24 +0000, Zhu, DengCheng wrote:
>> > ________________________________________
>> > From: Peter Zijlstra [a.p.zijlstra@chello.nl]
>> > Sent: Tuesday, November 22, 2011 6:51 PM
>> > To: Zhu, DengCheng
>> > Cc: Barzilay, Eyal; Fortuna, Zenon; Paul Mackerras; Ingo Molnar; Arnaldo Carvalho de Melo; ralf@linux-mips.org; LKML
>> > Subject: Re: [PATCH v2 5/5] perf: Enable applicable siblings when group leader is enable-on-exec
>> >
>> > On Tue, 2011-11-22 at 11:30 +0800, Deng-Cheng Zhu wrote:
>> >> Currently, when grouped events are created disabled and enable-on-exec, the
>> >> siblings won't be enabled on exec in fact. The problem looks like this:
>> >
>> > Arguably that's a daft thing to do, since if the leader is disabled the
>> > group won't get scheduled anyway. But I guess we should at least try to
>> > deal with it when people do do it.
>>
>> Well, by "grouped events" I mean "all of the grouped events", not only the
>> group leader. In fact the leader (and only the leader) will be enabled by
>> going through ctx->flexible_groups in perf_event_enable_on_exec().
>
> Right, I understood that. What I said was daft was to tag the
> non-leaders as enabled_on_exec,disabled. They wouldn't get scheduled
> anyway for as long as the leader is off.
>
>> > Seems perf-stat is a bit daft this way.
>> >
>> >> This patch fixes it.
>> >
>> > I guess it does, but its not pretty, event_enable_on_exec() already
>> > calls __perf_event_mark_enable(), now this recursion is limited because
>> > siblings can't have a sibling list of their own, but still.
>>
>> I did it like this just by reading the code comment of
>> __perf_event_mark_enabled(): "Enabling the leader of a group effectively
>> enables all the group members that aren't explicitly disabled ... Note:
>> this works for group members as well as group leaders since the non-leader
>> members' sibling_lists will be empty."
>>
>> So I suppose dealing with siblings' state in this traversal is the right
>> thing to do and introduces minimal code turmoil, although the latter is by
>> no means critical.
>
> Yeah, I just don't really like the recursion thing... Also, there's more
> ways to get to __perf_event_mark_enabled() and not all those want to
> actually do enable_on_exec().

Yep, two other functions call it. And whether doing enable_on_exec() in
__perf_event_mark_enabled() depends on how we interpret the meaning of the
latter. And if we do enable_on_exec() in it, uninterested events will be filtered
out in enable_on_exec().

One thing in your patch is uncertain to me:

> @@ -2463,11 +2461,25 @@ static int event_enable_on_exec(struct p
> if (event->state >= PERF_EVENT_STATE_INACTIVE)
> return 0;
>
> - __perf_event_mark_enabled(event, ctx);
> + event->state = PERF_EVENT_STATE_INACTIVE;
>
> return 1;
> }

By simply setting the event state in here, we bypass time stamp stuff as a result.
This might lead to inaccuracies...


Deng-Cheng


\
 
 \ /
  Last update: 2011-11-22 15:23    [W:0.127 / U:0.040 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site