On Tue, Oct 27, 2015 at 10:03 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Tue, Oct 27, 2015 at 04:57:05PM -0700, Paul Turner wrote:
> > +static void rseq_sched_out(struct preempt_notifier *pn,
> > + struct task_struct *next)
> > +{
> > + set_thread_flag(TIF_NOTIFY_RESUME);
> > +}
> >
> > static __read_mostly struct preempt_ops rseq_preempt_ops = {
> > .sched_in = rseq_sched_in_nop,
> > - .sched_out = rseq_sched_out_nop,
> > + .sched_out = rseq_sched_out,
> > };
>
> Since we're unconditionally setting this TIF flag for these tasks, can't
> we introduce something similar to the (contested) TIF_NOHZ_FULL thing
> which is kept on the task indefinitely.
>
So Andy and I talked about this also, I'm in favor, in particular this
has two nice effects:
a) In exit_to_usermode_loop() we can ensure that this is evaluated
prior to _TIF_SIGPENDING. This removes the current requirement that
we also validate this state in setup_rt_frame() [which can perturb
this state prior to our existing notifier].
b) We avoid spurious interactions with other things that use notify resume.
> That avoids having the preempt notifiers and this atomic op in the
> schedule path.
So we still want something there (although it can be definitely be
inlined as opposed to a preempt_notifier) since this allows us to only
evaluate this check on returns to user-space that might matter as
opposed to every syscall.
--
To unsubscribe from this list: send the line "unsubscribe linux-api" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
