Re: [PATCH 2/2] selftests/seccomp: Ensure that notifications come in FIFO order

[Sargun Dhillon]

On Thu, Apr 28, 2022 at 6:15 AM Tycho Andersen <tycho@tycho.pizza> wrote:
>
> On Wed, Apr 27, 2022 at 06:54:47PM -0700, Sargun Dhillon wrote:
> > +     /* Start children, and them generate notifications */
>                            ^^ - they maybe?
>
Whoops, this was supposed to be:
 /* Start children, and generate notifications */
> > +     for (i = 0; i < ARRAY_SIZE(pids); i++) {
> > +             pid = fork();
> > +             if (pid == 0) {
> > +                     ret = syscall(__NR_getppid);
> > +                     exit(ret != USER_NOTIF_MAGIC);
> > +             }
> > +             pids[i] = pid;
> > +     }
> > +
> > +     /* This spins until all of the children are sleeping */
> > +restart_wait:
> > +     for (i = 0; i < ARRAY_SIZE(pids); i++) {
> > +             if (get_proc_stat(pids[i]) != 'S') {
> > +                     nanosleep(&delay, NULL);
> > +                     goto restart_wait;
> > +             }
> > +     }
>
> I wonder if we should/can combine this loop with the previous one, and
> wait for the child to sleep in getppid() before we fork the next one.
> Otherwise isn't racy in the case that your loop continues to the next
> iteration before the child processes are scheduled, so things might be
> out of order? Maybe I'm missing something.
>
> In any case, this change seems reasonable to me.
>
> Tycho
It's okay if the child processes are started out of order. The test just
verifies that the calls are delivered in FIFO order according to when
the syscall was called (not when the process was started), and we do
this by just looking at the notification ID. It doesn't care about which
process generated the notification.