Re: [RFC PATCH v2] posix-timers: Support delivery of signals to the current thread

[Dmitry Vyukov]

 On Wed, 25 Jan 2023 at 17:31, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
>
> On 01/25, Dmitry Vyukov wrote:
> >
> > > diff --git a/kernel/time/posix-timers.c b/kernel/time/posix-timers.c
> > > index 5dead89308b7..e38b53a0f814 100644
> > > --- a/kernel/time/posix-timers.c
> > > +++ b/kernel/time/posix-timers.c
> > > @@ -336,6 +336,7 @@ void posixtimer_rearm(struct kernel_siginfo *info)
> > >  int posix_timer_event(struct k_itimer *timr, int si_private)
> > >  {
> > >         enum pid_type type;
> > > +       struct pid *pid;
> > >         int ret;
> > >         /*
> > >          * FIXME: if ->sigq is queued we can race with
> > > @@ -350,8 +351,9 @@ int posix_timer_event(struct k_itimer *timr, int si_private)
> > >          */
> > >         timr->sigq->info.si_sys_private = si_private;
> > >
> > > -       type = !(timr->it_sigev_notify & SIGEV_THREAD_ID) ? PIDTYPE_TGID : PIDTYPE_PID;
> > > -       ret = send_sigqueue(timr->sigq, timr->it_pid, type);
> > > +       type = (timr->it_sigev_notify & SIGEV_THREAD_ID) ? PIDTYPE_PID : PIDTYPE_TGID;
> > > +       pid = (type == PIDTYPE_PID) ? timr->it_pid : task_pid(current);
> > > +       ret = send_sigqueue(timr->sigq, pid, type);
> > >         /* If we failed to send the signal the timer stops. */
> > >         return ret > 0;
> > >  }
> >
> > Hi Oleg,
> >
> > This is indeed much simpler!
> >
> > Do I understand correctly that:
> > 1. I would need to use SIGEV_SIGNAL (without SIGEV_THREAD_ID)
>
> Yes,
>
> > 2. The signal is still queued into process shared_pending
>
> Yes. But just in case, please note that if this signal is not realtime
> (sigev_signo < SIGRTMIN) and it is already queued, it will be dropped.
> And I do not know if this can work for you.
>
> However this is what we already have with SIGEV_SIGNAL w/o SIGEV_THREAD_ID,
> and the same is true for SIGEV_THREAD_ID if the signal is already pending in
> target_task->pending.
>
> > 3. If the current task has not blocked the signal (it shouldn't), then
> > it won't kick any other task
>
> Yes,
>
> > 4. The current task will likely deliver the signal right on the timer
> > interrupt return to userspace
> > ?
>
> Yes.
>
> But! I just noticed send_sigqueue() does pid_task(pid, type), so the patch
> above needs another change
>
>
>         --- a/kernel/signal.c
>         +++ b/kernel/signal.c
>         @@ -1970,7 +1970,8 @@ int send_sigqueue(struct sigqueue *q, struct pid *pid, enum pid_type type)
>
>                 ret = -1;
>                 rcu_read_lock();
>         -       t = pid_task(pid, type);
>         +       // comment to explain why don't we use "type"
>         +       t = pid_task(pid, PIDTYPE_PID);
>                 if (!t || !likely(lock_task_sighand(t, &flags)))
>                         goto ret;
>
>
>
> > This changes the existing behavior (the "average bear" may be surprised :))
> > https://elixir.bootlin.com/linux/v6.2-rc5/source/kernel/signal.c#L1007
>
> this comment looks a bit misleading, s/main thread/target thread/
>
> > But currnently it's also queued into shared_pending and any thread
> > could get the signal anyway. So I think this should be fine.
>
> Yes.
>
> > On the positive side: it should improve performance. Delivering to the
> > currently running task is better on all fronts (no kicking,
> > rescheduling, IPIs, better locality), right?
>
> Well, iiuc this was the goal of your patch ? ;)

No, it actually is not. The actual goal is sampling activity of
threads. For CLOCK_PROCESS_CPUTIME_ID timers you get signals
proportional to the total activity of all threads (good), but all
signals are delivered to the main thread w/o even indication of what
thread caused the signal (questionable).