[gnutls-devel] gnutls_record_cork and maximum record size

[Nikos Mavrogiannopoulos]

On Fri, 2014-12-19 at 10:26 +0200, Jaak Ristioja wrote:

> >> As I understand the documentation, it implies that if the session is
> >> corked by gnutls_record_cork() and the size of the data queued by
> >> successive calls to gnutls_record_send() reaches the maximum record
> >> size, then gnutls_record_get_max_size() bytes of data are passed to the
> >> underlying transport. However, the source code for gnutls_record_send()
> >> implies that if the session is corked then data is queued until
> >> gnutls_record_uncork() is called (or GNUTLS_E_MEMORY_ERROR is returned).
> >> Do I misunderstand the documentation or is there a discrepancy between
> >> the source code and the documentation?
> > No you don't misunderstand the documentation, it is wrong. I'll update
> > it to match the current behavior of the code.
> Why not instead update the code to match the documentation? :D There
> might also be other projects who rely on the previously released
> documentation to be correct and might hit similar issues with a lot of
> queued records.

Well, it was never behaving like that so they couldn't have relied on
that behavior. That was the original intention of that API, but there
were practical issue with supporting every networking scenario
(blocking/non-blocking - tls/dtls), so it was simplified in the final
implementation.

> + * If called, gnutls_record_send() will no longer send partial records.
> + * All queued records will be sent when gnutls_uncork() is called.
> This is still somewhat confusing as the first sentence might imply that
> complete records will still be sent. Why not write something like this:
>   If called, gnutls_record_send() will no longer send any records until
>   gnutls_record_uncork() is called. Instead, gnutls_record_send() will
>   try to buffer all data passed to it. The data buffered by
>   gnutls_record_send() will be sent when gnutls_record_uncork() is
>   called.

I've updated it to be more precise along these lines.

> Because essentially, a gnutls_record_cork() call, followed by a number
> of gnutls_record_send() calls, followed by a gnutls_record_uncork() call
> is just a way of concatenating the data before it is finally passed to
> the underlying TLS implementation as a single chunk of memory.

Correct. The idea is to allow implementations to use record_send() for
small chunks, that will be concatenated.

> PS: Please note that the documentation uses gnutls_cork() and
> gnutls_uncork() instead of gnutls_record_cork() and gnutls_record_uncork().

Thanks, corrected.

regards,
Nikos