Re: Race condition in storeClose()

The patch seems to work, squid hasn't crashed in almost an hour and the
cache_log now contains entries that say "this be aioCancel"

Only one problem: It seems that my cache is running out of filedescriptors.
My cachemanager shows an abnormal number of files queued to be open, so what
I think is happening is that if the pending aio operation is a open, it
get's cancelled, however, since the callback doesn't ever get executed,
Opening_FD never gets decremented.

I could, however, be way off here.

cheers

-bmd

On Wed, Nov 22, 2000 at 10:16:21AM -0800, Brian Degenhardt wrote:
> This would make sense why my tests of 2.4 were so disastrous. Because I'm
> serving such large files, the chances of dropping the connection before it's
> done transfering is much more likely.
>
> Also, crashes seemed to cascade across peers which would make sense if the
> peers were satisfying sibling hits and then suddenly went down, dropping a
> ton of connections.
>
> I'll test out this patch and see if helps.
>
> cheers
>
> -bmd
>
> On Wed, Nov 22, 2000 at 11:36:14PM +0800, Adrian Chadd wrote:
> >
> > .. to replicate this on my linux box I would run an aufs squid at a
> > highish throughput (90-100 req is that on my box), and after a minute
> > or two hit ctrl-C on polyclt. This generates lots of dead clients,
> > which triggers this bug pretty reliably.
> >
> >
> > adrian
Received on Wed Nov 22 2000 - 13:11:45 MST