general/1119: Some transfers to distant sites terminate with graphics incomplete.

[Dean Gaudet <dg...@arctic.org>]

If you use SAFE_UNSERIALIZED_ACCEPT and a single socket then there's no
accept() serialization so you wouldn't trigger a pthread mutex problem.

I wonder if 1.3 fixes PR#1119.  1.3 again has different network behaviour
over 1.2... and 1.2 certainly has different behaviour over 1.1. 

I wonder if PR#1119 is related to the bug I've fixed in 1.3 that causes
the last packet of a transfer to be delayed waiting for the hostname
lookup to complete.  This bug, which I'd been considering a performance
problem only, exists in 1.2 because of the changes to improve pipelined
HTTP/1.1 connections.  It doesn't exist in 1.1 because 1.1 bflush()s all
over the place. 

I've cc'd the PR#1119 submitter... Jon could you try disabling
HostnameLookups (or try 1.3b3) and see if the problem goes away?  If so
I'll try to backport the fix from 1.3 to 1.2. 

Dean

On Thu, 8 Jan 1998, Lars Eilebrecht wrote:

> According to Dean Gaudet:
> 
> >  It it fixed by using USE_FCNTL_SERIALIZED_ACCEPT?  'cause I suspect the
> >  pthreads stuff far more than any of the others... it's the major thing
> >  that's different between solaris and all other platforms.  And it's only
> >  used in 2.5 and later, so wouldn't affect Dirk's 2.4.
> 
> Hmm... I was running a sparc5 test box with Solaris 2.6 and 1.3b4-dev for
> some weeks without any problems. Some days ago I upgraded the server of
> my University www.uni-siegen.de to the configuration from my test box and
> haven't got any reports about broken images or other problems...
> I compiled with SAFE_UNSERIALIZED_ACCEPT.
> 
> I wonder if PR#1119 is a related problem (I posted a note about this
> to the list some time ago, see <XF...@unix-ag.org>)?
> 
> 
> ciao...
> -- 
> Lars Eilebrecht                    - Instant programmer... just add coffee!
> sfx@unix-ag.org
> http://www.si.unix-ag.org/~sfx/
> 
>