Re: [RFC PATCH] tracing: add sched_prio_update

From: Julien Desfossez
Date: Wed Jul 06 2016 - 10:02:12 EST


On 06-Jul-2016 09:13:25 AM, Steven Rostedt wrote:
> On Tue, 5 Jul 2016 21:50:34 +0000 (UTC)
> Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
>
> > >
> > >> +
> > >> + TP_PROTO(struct task_struct *tsk),
> > >> +
> > >> + TP_ARGS(tsk),
> > >> +
> > >> + TP_STRUCT__entry(
> > >> + __array( char, comm, TASK_COMM_LEN )
> > >
> > > I could imagine this being a high frequency tracepoint, especially with
> > > a lot of boosting going on. Can we nuke the comm recording and let the
> > > userspace tools just hook to the sched_switch tracepoint for that?
> >
> > We can surely do that.
> >
> > Just to clarify: currently this tracepoint is *not* hooked on PI boosting,
> > as described in the changelog. This tracepoint is about the prio attributes
> > set by user-space. The PI boosting temporarily changes the task struct prio
> > without updating the associated policy, which seems rather
> > implementation-specific and odd to expose.
> >
> > Thoughts ?
>
> Ah, you're right, I was thinking it was at boosting. But still, it's a
> rather hefty tracepoint (lots of fields), probably want to keep from
> adding comm too.

Yes, I agree we can remove the comm field, it is easy to get from the
previous sched_switch.

> > >> + __field( pid_t, pid )
> > >> + __field( unsigned int, policy )
> > >> + __field( int, nice )
> > >> + __field( unsigned int, rt_priority )
> > >> + __field( u64, dl_runtime )
> > >> + __field( u64, dl_deadline )
> > >> + __field( u64, dl_period )
> > >> + ),
> > >> +
> > >> + TP_fast_assign(
> > >> + memcpy(__entry->comm, tsk->comm, TASK_COMM_LEN);
> > >> + __entry->pid = tsk->pid;
> > >> + __entry->policy = tsk->policy;
> > >> + __entry->nice = task_nice(tsk);
> > >> + __entry->rt_priority = tsk->rt_priority;
> > >> + __entry->dl_runtime = tsk->dl.dl_runtime;
> > >> + __entry->dl_deadline = tsk->dl.dl_deadline;
> > >> + __entry->dl_period = tsk->dl.dl_period;
> > >> + ),
> > >> +
> > >> + TP_printk("comm=%s pid=%d, policy=%s, nice=%d, rt_priority=%u, "
> > >> + "dl_runtime=%Lu, dl_deadline=%Lu, dl_period=%Lu",
> > >> + __entry->comm, __entry->pid,
> > >> + __print_symbolic(__entry->policy, SCHEDULING_POLICY),
> > >> + __entry->nice, __entry->rt_priority,
> > >> + __entry->dl_runtime, __entry->dl_deadline,
> > >> + __entry->dl_period)
> > >> +);
> > >> #endif /* _TRACE_SCHED_H */
> > >>
> > >> /* This part must be outside protection */
> > >> diff --git a/kernel/fork.c b/kernel/fork.c
> > >> index 7926993..ac4294a 100644
> > >> --- a/kernel/fork.c
> > >> +++ b/kernel/fork.c
> > >> @@ -1773,6 +1773,7 @@ long _do_fork(unsigned long clone_flags,
> > >> struct pid *pid;
> > >>
> > >> trace_sched_process_fork(current, p);
> > >> + trace_sched_prio_update(p);
>
> From the change log:
>
> "It is emitted in the code path of the sched_setscheduler,
> sched_setattr, sched_setparam, nice and the fork system calls. For fork, it is emitted
> after the sched_process_fork tracepoint for timeline consistency and
> because the PID is not yet set when sched_fork() is called."
>
> I'm not convinced this should be needed. I hate adding back to back
> tracepoints.

Indeed, having two tracepoints back to back is not pretty. We placed it
here to get the priority of the newly created threads. Maybe a more
appropriate way of doing that would be to extend the sched_process_fork
tracepoint to output the same scheduling informations. Would you prefer
that option ?

Thanks,

Julien