[sheepdog] fix embryonic connection

[Liu Yuan]

On 09/04/2012 04:56 PM, MORITA Kazutaka wrote:
> If it works well, it's okay to me for now.  But, in future, I'd like
> to consider removing TCP keepalive.  If the target node is in the
> sheepdog node list, sheep should keep the connection even if timeout
> is fired.  If the target is not in the node list, sheep should close
> the connection ASAP.  

This sounds feasible to me. But I have no idea if this will be better
design until the code comes up. I guess this change will push complexity
to sockfd cache and the sockfd usage in gateway code.

The check of availability should be done by
> cluster drivers.  That's the reason I suggested "retry poll while
> epoch is not updated".

I am not against removing keepalive, because we have met these
undocumented problems, who ever know we'll meet yet another one. But
before removing it, I'd like see a better design: either simpler code or
efficient new mechanism.

But as for this patch set, one bug is for keepalive, another is for
connect(), which exist in sheepdog for ages.

Thanks,
Yuan