On 07/14/2010 05:34 PM, Amos Jeffries wrote:
> Re: squid-3-10601.patch
> 
> Alex,
>  would you mind explaining why this error is not a problem and what its
> meant to solve? I think I get it and can relate to some issues people have
> reported but want to be sure if and before it gets ported to 3.1.


I believe we get COMM_NOMESSAGE when we thought there was a connection
ready to be accepted but we were wrong. I do not know exactly when such
things are happening without SMP. With SMP, COMM_NOMESSAGE is probably
the result of a benign race between two or more workers trying to accept
a connection from a shared listening socket.

The code I fixed in r10601 thought that COMM_NOMESSAGE is some kind of a
fatal error and we should stop listening after it (despite code comments
saying the opposite). My understanding is that COMM_NOMESSAGE is like
EAGAIN: not a true error but more like a "try again later" notification.

HTH,

Alex.

Reply via email to