Re: I'm back

Adrian Chadd wrote:
> On Mon, Jul 15, 2002, Henrik Nordstrom wrote:
> > 2. Finish the store_client_copy API migration started before you left..
> > This is known to break at least pipeline prefetched replies in the
> > current state (tries to restart reading at 0, which the new store_client
> > API does not like..)
>
> Oh?
>
> Tries to restart reading what at 0?
> it should be reading a new object here.
> Got a bugzilla id, and a cute way of reproducing it?

Just send a pipelined request...

The problem is in clientSendMoreData(), or actually the fact that the second
reply isn't deferred until clientSendMoreData() is invoked on it by the
store.. storeClientSendMoreData() then notices this is not the current reply
for the connection, forgets the data it got and assumes it is OK to just ask
for the data again when the reply can be forwarded by restarting the
store_client_copy() call in clientKeepaliveNextRequest()...

This is only one of the bugs in how Squid deals with pipelined requests.. that
whole section of the code is bit of mess. There is also some suspect timing
problems in clientReadRequests() and how/when/why the processing of pipelined
requests are deferred, and access.log..

Regards
Henrik
Received on Mon Jul 15 2002 - 07:01:23 MDT