RE: hash_string broken

> > I've logged a bugzilla entry on this, but we seem to have
> found a fault
> > with hash_string:
> >
> > The string
> > "cif_bm\\raffagli" in a hash of size 7921, sometimes gets
> inserted at
> > 7482, and sometimes at 6475.
>
> This sounds similar to a problem that I saw some time ago. I would
> get a coredump in one of the hash functions. The coredump showed
> that two buckets had been "joined." The beginning entries in
> the list were all in the right bucket, but at some point the hash
> values didn't match.
>
> If I recall correctly, it turned out to be a bug with the filemap
> routines stomping on memory out of bounds. I found it finally by
> noticing that one of the hash_link pointers was wrong by just one
> bit.

Just when you want something to crash, it doesn't.
Today it was up at 2.4 million hits, without a crash.

For some reason it was spinning in a loop without servicing
filedescriptors (it accepted connections, but didn't serve
any data).
[looks at the cache.log] Hm... maybe the NTLM helpers were all busy, that
might be
a good reason for the slowness.

On an unrelated note, I spent the last two days at Linux Expo Milan.
There I've spent some time with Andrea Arcangeli, Linux kernel hacker,
and we've discussed some things which are getting into the Linux kernel
which could IMO be VERY useful to us (such as ways to DMA stuff
directly from the HDDs to the network card buffer).

Expect a report from me as soon as my hangover wears off ;]

-- 
	/kinkie