Re: [Qemu-devel] [PATCH] qemu-char: another io_add_watch_poll fix

[Hans de Goede <hdegoede@redhat.com>]

Ji,

On 04/16/2013 11:34 AM, Paolo Bonzini wrote:
> Il 16/04/2013 11:15, Gerd Hoffmann ha scritto:
>> On 04/10/13 15:23, Paolo Bonzini wrote:
>>> After attaching the source, we have to remove the reference we hold
>>> to it, because we do not hold anymore a pointer to the source.
>>>
>>> If we do not do this, removing the source will not finalize it and
>>> will not drop the "real" I/O watch source.
>>>
>>> This showed up when backporting the new flow control patches to older
>>> versions of QEMU that still used select.  The whole select then failed
>>> with EBADF (poll instead will reporting POLLNVAL on a single pollfd)
>>> and QEMU froze.
>>
>> I get freezes now in master, bisecting points to this patch.
>>
>> Reproducer: "qemu -serial pty".
>>
>> qemu is pretty much unusable with libvirt now as libvirt uses pty
>> chardevs by default for serial & monitor ...
>
> I'm not sure why all users of qemu_chr_fe_add_watch believe that the
> watch will be one-shot.  This is definitely not what g_io_create_watch
> does...

It is supposed to depend on the return value of the callback, if you
return False, then the source should be removed, in essence making
the watch one-shot, see:
http://www.gtk.org/api/2.6/glib/glib-The-Main-Event-Loop.html#GSourceFunc

Which specifically says:

Returns:	it should return FALSE if the source should be removed.

And using sources in a one-shot mode by making the callback return false
is quite normal in the glib world.

Regards,

Hans