Re: [libvirt PATCH 2/2] tests: don't set G_DEBUG=fatal-warnings on macOS
On Fri, Apr 29, 2022 at 01:05:42AM -0700, Andrea Bolognani wrote:
On Thu, Apr 28, 2022 at 05:55:41PM +0100, Daniel P. Berrangé wrote:
> On Thu, Apr 28, 2022 at 08:33:46AM -0700, Andrea Bolognani wrote:
> > In other words, the current implementation of g_poll() on macOS
> > doesn't follow the contract defined by GLib itself. It seems to me
> > that this is a (fairly serious) bug in the library, no?
>
> It is significant, but long standing. GLib actually had this behaviour
> forever on macOS, but it regressed when Meson was introduced, until
> the recent fix.
>
> The question is whether efficiency trumps API semantics. Normally I'm
> heavily biased towards API semantics, but poll is a performance
> critical API, so it isn't so easy to declare we must workaround all
> the quirks.
>
> I filed a bug to raise the subject for discussion though
>
> https://gitlab.gnome.org/GNOME/glib/-/issues/2644
That's an excellent bug report! Thanks for spending time on it, and
let's see where the upstream discussion leads :)
Just so I don't sway the potential discussion in the issue to the wrong
way, I'll ask here. Since BROKEN_POLL is used due to poll() on OS/X not
behaving correctly with fds=NULL, nfds=0 [0], wouldn't it be possible to
"fix" at least the internal calls to g_poll() which are done when
running the main loop since they always have at least one FD (the
eventfd used to signal the main loop [1])? I know it would not fix the
problem, but it would at least workaround the part that is used the
most, I presume.
Martin
[0] https://gitlab.gnome.org/GNOME/glib/-/blob/main/glib/gpoll.c#L59
[1] this is a presumption I did not find in the code, but I've seen in
gdb and strace
Attachments:
- signature.asc (application/pgp-signature — 833 bytes)