Re: [chrony-users] FreeBSD socket.test FAIL: recvmsg() error EAGAIN on accept TCP socket receive

[Matthias Andree <matthias.andree@xxxxxx>]

Am 28.08.25 um 10:07 schrieb Miroslav Lichvar:

> On Wed, Aug 27, 2025 at 04:50:36PM +0200, Matthias Andree wrote:
> > chrony 4.7 and 4.8 (and possibly others) have an issue with
> > test/util/socket.c's self test,
> > UNLESS running under "truss" (think strace in the Linux world).
> >
> > Apparently we're looking at OS-specific differences WRT non-blocking
> > accept()ed sockets for TCP between FreeBSD and, say, Linux, or with
> > something that a syscall tracer will cause to be retried.
> >       
> I have a FreeBSD 14.2 VM that I use for testing and it wasn't failing
> for me. Maybe it's too slow.
>
> > Debugging this, SCK_AcceptConnection unconditionally sets O_NONBLOCK on the
> > socket,
> >       
> It should be now fixed in git. Thanks for the report.
>

Miroslav,

thanks for the prompt reply. I  confirm the fix works for us and have pushed it to the FreeBSD ports repository.

> > Speaking of which, configure states that FreeBSD 11.0 had a broken
> > recvmmsg() - have there been bug reports at the time or fixes? Have we
> > retried this on 13.5, 14.2, 14.3? What do we need to look for?
> >       
> That bug was fixed, but it seems it's still just a loop around
> recvmsg, not a single system call, which means there is an extra
> system call made and I'm not sure if it's worth it.
>
> https://github.com/freebsd/freebsd-src/blob/main/lib/libsys/recvmmsg.c#L80
>

OK, then indeed we can just talk recvmsg directly if it's not better.

Regards,
Matthias