Re: [patch V2 3/5] posix-cpu-timers: Provide mechanisms to defer timer handling to task_work

[Frederic Weisbecker <frederic@kernel.org>]

On Thu, Jul 23, 2020 at 10:32:54AM +0200, Thomas Gleixner wrote:
> Frederic Weisbecker <frederic@kernel.org> writes:
> > On Fri, Jul 17, 2020 at 12:50:34AM +0200, Peter Zijlstra wrote:
> >> On Thu, Jul 16, 2020 at 10:19:26PM +0200, Thomas Gleixner wrote:
> >> > +static void __run_posix_cpu_timers(struct task_struct *tsk)
> >> > +{
> >> > +	struct posix_cputimers *pct = &tsk->posix_cputimers;
> >> > +
> >> > +	if (!test_and_set_bit(CPUTIMERS_WORK_SCHEDULED, &pct->flags))
> >> > +		task_work_add(tsk, &pct->task_work, true);
> >> > +}
> >> > +
> >> > +static inline void posix_cpu_timers_enable_work(struct task_struct *tsk)
> >> > +{
> >> > +	clear_bit(CPUTIMERS_WORK_SCHEDULED, &tsk->posix_cputimers.flags);
> >> 	/*
> >> 	 * Ensure we observe everything before a failing test_and_set()
> >> 	 * in __run_posix_cpu_timers().
> >> 	 */
> >> 	smp_mb__after_atomic();
> >> > +}
> >> 
> >> Such that when another timer interrupt happens while we run this, we're
> >> guaranteed to either see it, or get re-queued and thus re-run the
> >> function.
> >
> > But each thread in the process enqueues its own task work and flips its
> > own flags. So if task A runs the task work and task B runs __run_posix_cpu_timers(),
> > they wouldn't be ordering against the same flags.
> 
> If two tasks queue work independent of each other then one of them will
> find it done already, which is the same as if two tasks of the same
> process execute run_posix_cpu_timers() in parallel.
> 
> I really don't want to go into the rathole of making the work or the
> synchronization process wide. That's a guarantee for disaster.
> 
> Handling task work strictly per task is straight forward and simple. The
> eventually resulting contention on sighand lock in task work is
> unavoidable, but that's a reasonable tradeoff vs. the complexity you
> need to handle task work process wide.

Definetly!

I was only commenting on the barrier suggestion. But I believe it shouldn't
be needed in the end.

If we were to have a per task work for thread timers and a per process work
for process timers, that means we would need to cut down the whole thing, and also
take care about timers firing after exit_task_work(), which isn't an issue
in the thread case as the work will simply be ignored for an exiting task but
it's a big issue in the case of process wide handling.

Anyway, the current layout is simple enough.