[patch] Fix various bugs when using the gnutls async API

[Sjoerd Simons]

On Tue, 2010-08-17 at 17:55 +0200, Nikos Mavrogiannopoulos wrote:
> On 08/11/2010 03:49 PM, Sjoerd Simons wrote:
> > Hi,
> > 
> > Our XMPP library (wocky) uses gnutls in a quite heavily async way, which
> > unfortunately seems to have triggered some bugs in the lastest
> > versions :/. Some patches for git HEAD attached, which seems to fix
> > things nicely for me.
> > 
> > Some extra comments for some of the patches:
> > * 0003-Remember-the-amount-of-user-data-we-re-sending-out.patch
> >   I guess It could be argued that this information should be stored
> >   inside the record_send_buffer instead of in the session internal
> >   struct directly. I'm happy to change the patch around to do just
> >   that :)
> 
> Hello Sjoerd,
>  Thanks for the patches. I've applied 1 and 2. 

Thanks

> But on 3 I cannot get what was the issue that mandated the change. Why you need the user
> length instead of the byte length?

gnutls_record_send is documented to return the number of bytes send (by
the user!), so the failure case is currently like this:

gnutls_record_send () => -EAGAIN
gnutls_record_send () => byte_length

Where byte_lenght is what at that point is still left in the
record_send_buffer, it's not related to the data the user was sending.

> I am also wondering why the test program mini-egain that forces the pull
> function to return eagain now and then, it didn't trigger the errors you
> see. How do you trigger those issues?

I guess pure luck in the case of running mini-egain ? In our case every
pull and push will first return EAGAIN and only on the second call
succeed, so it will stress all the code-patch related to EAGAIN. 
-- 
Sjoerd Simons <sjoerd.simons at collabora.co.uk>
Collabora Ltd.