Re: [PATCH] kthread: Add kthread_work tracepoints

[Rob Clark]

On Mon, Oct 12, 2020 at 6:59 AM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
> On Sat, 10 Oct 2020 11:03:22 -0700
> Rob Clark <robdclark@xxxxxxxxx> wrote:
>
> > /**
> > + * sched_kthread_work_execute_start - called immediately before the work callback
> > + * @work:    pointer to struct kthread_work
> > + *
> > + * Allows to track kthread work execution.
> > + */
> > +TRACE_EVENT(sched_kthread_work_execute_start,
> > +
> > +     TP_PROTO(struct kthread_work *work),
> > +
> > +     TP_ARGS(work),
> > +
> > +     TP_STRUCT__entry(
> > +             __field( void *,        work    )
> > +             __field( void *,        function)
> > +     ),
> > +
> > +     TP_fast_assign(
> > +             __entry->work           = work;
> > +             __entry->function       = work->func;
> > +     ),
> > +
> > +     TP_printk("work struct %p: function %ps", __entry->work, __entry->function)
> > +);
> > +
> > +/**
> > + * sched_kthread_work_execute_end - called immediately after the work callback
> > + * @work:    pointer to struct work_struct
> > + * @function:   pointer to worker function
> > + *
> > + * Allows to track workqueue execution.
> > + */
> > +TRACE_EVENT(sched_kthread_work_execute_end,
> > +
> > +     TP_PROTO(struct kthread_work *work, kthread_work_func_t function),
> > +
> > +     TP_ARGS(work, function),
> > +
> > +     TP_STRUCT__entry(
> > +             __field( void *,        work    )
> > +             __field( void *,        function)
> > +     ),
> > +
> > +     TP_fast_assign(
> > +             __entry->work           = work;
> > +             __entry->function       = function;
> > +     ),
> > +
> > +     TP_printk("work struct %p: function %ps", __entry->work, __entry->function)
> > +);
> > +
>
>
> Please combine the above into:
>
> DECLARE_EVENT_CLASS(sched_kthread_work_execute_template,
>
>         TP_PROTO(struct kthread_work *work),
>
>         TP_ARGS(work),
>
>         TP_STRUCT__entry(
>                 __field( void *,        work    )
>                 __field( void *,        function)
>         ),
>
>         TP_fast_assign(
>                 __entry->work           = work;
>                 __entry->function       = work->func;
>         ),
>
>         TP_printk("work struct %p: function %ps", __entry->work, __entry->function)
> );
>
> DEFINE_EVENT(sched_kthread_work_execute_template, sched_kthread_work_execute_start,
>         TP_PROTO(struct kthread_work *work),
>         TP_ARGS(work));
>
> DEFINE_EVENT(sched_kthread_work_execute_template, sched_kthread_work_execute_end,
>         TP_PROTO(struct kthread_work *work),
>         TP_ARGS(work));
>
> As events are cheap, classes are expensive (size wise), and a TRACE_EVENT()
> is really just a CLASS and EVENT for a single instance.

I think we cannot quite do this, because we should not rely on being
able to dereference work after it finishes.  Although I suppose I
could just define it to explicitly pass the fxn ptr to both
tracepoints..

BR,
-R