On Sun, Mar 27, 2005 at 08:38:39AM -0800, Jason Thorpe wrote:
> 
> I'm not even sure my suggestion of translation ENOBUFS -> 0 is  
> actually necessary now.

One thing that bothers me is the effect of all this on the use of
watermarks on the socket buffer.  Even given that UDP is unreliable,
it does seem like we _don't_ want to do anything that will cause select
to return, or send to unblock, with a watermark set, and then have the
send syscall return ENOBUFS.  That's highly counterintuitive and, it
seems to me, will frustrate efforts to do congestion control over UDP
in any sane way (not that what Manu is trying to do seems even remotely
sane to me, at least not the way he's trying to do it).

It occurs to me that he might see significantly better throughput for
his application simply by making the interface queue longer.

-- 
 Thor Lancelot Simon	                                      tls@rek.tjls.com

"The inconsistency is startling, though admittedly, if consistency is to be
 abandoned or transcended, there is no problem."		- Noam Chomsky