Re: Errors and later panics in 2.6.0-test11.

[Nathan Scott <nathans@sgi.com>]

On Thu, Dec 04, 2003 at 07:09:10AM +1100, Neil Brown wrote:
> On Wednesday December 3, torvalds@osdl.org wrote:
> > 
> > 
> > On Wed, 3 Dec 2003, Jens Axboe wrote:
> > > >
> > > > Interesting. Another RAID 0 problem report..
> > >
> > > Hmm did _all_ reports include raid-0, or just "some" raid? I'm looking
> > > at the bio_pair stuff which raid-0 is the only user of, something looks
> > > fishy there.
> > 
> > The ones I've seen seem to be raid-0, but Nathan (nathans@sgi.com)
> > reported problems in RAID-5 under load. I didn't decode the full oops on
> > that one, but it really looked like a stale "bi" bio that trapped on the
> > PAGE_ALLOC debug code.
> > 
> 
> Nathan's had a second oops that turned out to be a bi_next pointer
> being bad in a bio that raid5 had just about finished writing out.
> So there does seem to be something wrong with bio handling, quite
> possibly in raid5.
> 
> The only thing I could find was that if raid5 received two overlapping
> bios concurrently (or atleast received the second before it had
> finished with the first) it could get confused.  I've asked Nathan to
> try a patch that BUGs when that happens.

I haven't tripped the bug so far today, although have been
running with page-sized fs blocksize so far - perhaps that
is implicated, and makes it less likely to trigger (when I
say "the bug" there, I mean neither the panic, nor the new
BUG_ON(); I'll revert back to smaller block sizes next).

That error path bio_put issue you spotted in XFS, Neil, I
think is a valid problem - I'm not sure that is reachable
code in practice (possibly overly defensive XFS bio code),
I'll go investigate that some more.

cheers.

-- 
Nathan