On Sat, Oct 23, 2021 at 10:40:56AM +0100, Martin Pieuchot wrote:
> Diff below switches both poll(2) and select(2) to the kqueue-based
> implementation.
Passed regress on i386.
> In addition it switches libevent(3) to use poll(2) by default for
> testing purposes.
It created some test failures in syslogd regress as it could not
see kqueue in ktrace. This is an expected failure.
> I don't have any open bug left with this diff and I'm happily running
> GNOME with it. So I'd be happy if you could try to break it and report
> back.
I think we should give it a try.
> +#ifdef DIAGNOSTIC
> + /*
> + * Lazily delete spurious events.
> + *
> + * This should not happen as long as kqueue_purge() is called
> + * at the end of every syscall. It migh be interesting to do
> + * like DragonFlyBSD and not always allocated a new knote in
> + * kqueue_register() with that lazy removal makes sense.
> + */
> + if (i >= nfds) {
> + DPRINTFN(0, "poll get out of range udata %lu vs serial %lu\n",
> + (unsigned long)kevp->udata, p->p_kq_serial);
> + kevp->flags = EV_DISABLE|EV_DELETE;
> + kqueue_register(p->p_kq, kevp, p);
> + return (0);
> + }
> + if ((int)kevp->ident != pl[i].fd) {
> + DPRINTFN(0, "poll get %lu/%d mismatch fd %u!=%d serial %lu\n",
> + i+1, nfds, (int)kevp->ident, pl[i].fd, p->p_kq_serial);
> + return (0);
> + }
> +#endif
DIAGNOSTIC should not change behavior by returning 0.
> + /* NOTE: POLLHUP and POLLOUT/POLLWRNORM are mutually exclusive*/
Remove "NOTE:" and the comment will fit into the line.
> +#if 1
> + if (pl[i].events != 0 && pl[i].events != POLLOUT)
> + DPRINTFN(0, "weird events %x\n", pl[i].events);
> +#endif
Why #if 1, should it be DIAGNOSTIC ?
> + /* Tell uper layer to ask for POLLUP only */
upper
