[Spice-devel] [PATCH spice-gtk] spice-channel: return if has_error is TRUE in spice_channel_write_msg

[Victor Toso]

Hi,

On Thu, Jun 27, 2019 at 02:27:57PM +0200, Jakub Janku wrote:
> Hi,
> 
> On Thu, Jun 27, 2019 at 10:10 AM Frediano Ziglio <fziglio at redhat.com> wrote:
> >
> > >
> > > Avoid linearizing if the message isn't written out anyway
> >
> > "linearizing" ? What do you mean about that?
> > Looking at definitions it seems not correct to me.
> 
> I was simply referring to the spice_marshaller_linearize() call.
> >
> > > (spice_channel_flush_wire checks() this condition as well).
> > >
> > > This also silences the following error:
> > >
> > >     (spicy:32087): GSpice-CRITICAL **: 16:22:03.147:
> > >     spice_session_get_read_only: assertion 'SPICE_IS_SESSION(self)' failed
> > >
> > > that can be seen if the channel gets disconnected
> > > by the session while having non-empty write queue.
> > >
> > > spice_session_channel_destroy() sets channel->priv->session to NULL,
> > > but spice_channel_write_msg() subsequently attempts to call
> > > spice_session_get_read_only() with NULL pointer.

Right, after spice_channel_disconnect() we shouldn't try to
read/write anymore indeed. As spice_session_channel_destroy()
sets c->has_error = TRUE, I think this match is almost fine.

> > Minor: this is the explanation why the error on the previous
> > paragraph should not be treated like an error, I think it should
> > be the same paragraphs.
> 
> Makes sense.
> >
> > OT: maybe channel session should never be NULL?
> 
> It indeed does seem weird that
> g_clear_object(&channel->priv->session); is called when the
> "spice-session" property of the channel is G_PARAM_CONSTRUCT_ONLY --
> with this flag, I would expect the property to not change after the
> construction.

If I'm not reading it wrong, this only happens on
spice_channel_dispose() which means we shouldn't be relying on
this object at this point.
 
> Spice session waits for the channel destruction anyway
> (channel_finally_destroyed callback), so it should be imho fine that
> the channel holds a reference to the session while it is being
> destroyed. So I think we could remove that
> g_clear_object(&channel->priv->session); call in
> spice_session_channel_destroy().

Maybe moving to spice_channel_finalize() would already be enough
if you really need to postpone this but I don't see why allow
using priv->session while we are disposing the resources of the
channel already.

> >
> > > Signed-off-by: Jakub Janků <jjanku at redhat.com>
> > > ---
> > >  src/spice-channel.c | 5 +++++
> > >  1 file changed, 5 insertions(+)
> > >
> > > diff --git a/src/spice-channel.c b/src/spice-channel.c
> > > index 61de177..aa80edf 100644
> > > --- a/src/spice-channel.c
> > > +++ b/src/spice-channel.c
> > > @@ -897,6 +897,11 @@ static void spice_channel_write_msg(SpiceChannel
> > > *channel, SpiceMsgOut *out)
> > >      g_return_if_fail(out != NULL);
> > >      g_return_if_fail(channel == out->channel);
> > >
> > > +    if (channel->priv->has_error) {
> > > +        spice_msg_out_unref(out);
> > > +        return;
> > > +    }

I'm thinking that it might actually be a programming error that
we are calling _write_msg() after channel_disconnect is called,
so a warning might be welcomed? 

Not sure of a different fix; how easy is to reproduce this?

Cheers,
Victor

> > > +
> > >      if (out->ro_check &&
> > >          spice_channel_get_read_only(channel)) {
> > >          g_warning("Try to send message while read-only. Please report a
> > >          bug.");
> >
> > Frediano
> _______________________________________________
> Spice-devel mailing list
> Spice-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/spice-devel
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/spice-devel/attachments/20190711/8719306e/attachment.sig>