Re: Race condition in storeClose()

Adrian Chadd wrote:
>
> (1) - if you took reiserfs_raw as an example, you COULD have two squid
> processes running and BOTH accessing the same object store ..
> you could do some funny socket/tcp tricks at the upper levels and
> utilise a fine-grained-locking FS to get concurrancy without ever
> putting a mutex inside squid itself.

Funny you should mention this. For experimentation sake, Moez has
written a module to his Abyss web server to serve cache hits from the
same object store Squid is using...and pass misses through to a running
Squid. Call it a prototype for some future 'fast path' dual server
approach (the fast path being the hits only server which answers hits
quickly).

So...proof of concept has been achieved. When he and I get back from
our respective vacations we're going to see if we can debuggify it, and
do some experimenting.

Though it may be best to simply implement faster http client side code
in Squid. We'll have to see, but obviously this objective took less
time than rewriting the http layer and the hits-path from scratch.

-- 
Joe Cooper <joe@swelltech.com>
http://www.swelltech.com