Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210824112946.9324-1-laoar.shao@gmail.com> <YS3/jhuRNS2GGVrp@hirez.programming.kicks-ass.net>
In-Reply-To: <YS3/jhuRNS2GGVrp@hirez.programming.kicks-ass.net>
From:   Yafang Shao <laoar.shao@gmail.com>
Date:   Tue, 31 Aug 2021 20:57:43 +0800
Message-ID: <CALOAHbBve8h7r5YQfkp-652RY8oq3N8bwXDLVT7wVhnPJpZZdg@mail.gmail.com>
Subject: Re: [PATCH v3 0/7] sched: support schedstats for RT sched class
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     Ingo Molnar <mingo@redhat.com>, Mel Gorman <mgorman@suse.de>,
        Juri Lelli <juri.lelli@redhat.com>,
        Vincent Guittot <vincent.guittot@linaro.org>,
        Dietmar Eggemann <dietmar.eggemann@arm.com>,
        Steven Rostedt <rostedt@goodmis.org>,
        Benjamin Segall <bsegall@google.com>,
        Daniel Bristot de Oliveira <bristot@redhat.com>,
        Alison Chaiken <achaiken@aurora.tech>,
        kbuild test robot <lkp@intel.com>,
        LKML <linux-kernel@vger.kernel.org>,
        linux-rt-users <linux-rt-users@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Aug 31, 2021 at 6:08 PM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Tue, Aug 24, 2021 at 11:29:39AM +0000, Yafang Shao wrote:
> > Hi Ingo, Peter,
> >
> > This feature is useful to trace the sched details of RT tasks. Hopefully
> > you can give some feedback on it.
> >
> > We want to measure the latency of RT tasks in our production
> > environment with schedstats facility, but currently schedstats is only
> > supported for fair sched class. In order to support if for other sched
> > classes, we should make it independent of fair sched class. The struct
> > sched_statistics is the schedular statistics of a task_struct or a
> > task_group, both of which are independent of sched class. So we can move
> > struct sched_statistics into struct task_struct and struct task_group to
> > achieve the goal.
>
> Do you really want schedstats or do you want the tracepoints?

I really want the schedstats, which is very helpful to help us profile
thread-level latency.
The tracepoints is a bonus.

> In general
> I really want to cut back on the built-in statistics crud we carry,

Pls. don't.
There are really use cases of statistics.
Our use case as follows,

Userspace Code Scope         Profiler

{
    user_func_abc();   <----      uprobe_begin() get the start statistics
    ...
    user_func_xyz();   <----       uprobe_end()  get the end statistics
}

Then with this profiler we can easily get what happened in this scope
and why its latency was great:
    scope_latency = Wait + Sleep + Blocked [1]  + Run (stime + utime)

If there is no schedstats, we have to trace the heavy sched::sched_switch.

[1]. With patch #5 and don't include sum_block_runtime in sum_sleep_runtime

> there's too much and it seems to keep growing forever :-(
>
> (as is the case here, you're extending it as well)
>
> That said; making schedstats cover the other classes can be seen as
> fixing an inconsistency, but then you forgot deadline.
>

There's no deadline task on our server, so I didn't support it for deadline.
But with this patchset, it is very easy to extend it to deadline and
any other sched classes.


> > After the patchset, schestats are orgnized as follows,
> > struct task_struct {
> >     ...
> >     struct sched_statistics statistics;
> >     ...
> >     struct sched_entity *se;
> >     struct sched_rt_entity *rt;
> >     ...
> > };
> >
> > struct task_group {                    |---> stats[0] : of CPU0
> >     ...                                |
> >     struct sched_statistics **stats; --|---> stats[1] : of CPU1
> >     ...                                |
> >                                        |---> stats[n] : of CPUn
> >  #ifdef CONFIG_FAIR_GROUP_SCHED
> >     struct sched_entity **se;
> >  #endif
> >  #ifdef CONFIG_RT_GROUP_SCHED
> >     struct sched_rt_entity  **rt_se;
> >  #endif
> >     ...
> > };
>
> Yeah, this seems to give a terrible mess, let me see if I can come up
> with anything less horrible.


-- 
Thanks
Yafang