Re: fsx-linux loosing mmap() writes under memory pressure

[Jan Kara <jack@suse.cz>]

On Thu 05-03-09 13:55:43, Nick Piggin wrote:
> On Thursday 05 March 2009 04:50:31 Jan Kara wrote:
> > On Wed 04-03-09 16:55:35, Jan Kara wrote:
> > > On Wed 04-03-09 15:51:09, Jan Kara wrote:
> > > >   first, I'd like to point out that this has happened under UML so it
> > > > can be just some obscure bug in that architecture but I belive it's
> > > > worth debugging anyway. Now to the problem:
> > > >   This has happened with today Linus's git snapshot. The filesystem is
> > > > ext3 with *1KB* blocksize. I booted UML with 64MB of memory and run
> > > > (these are test's from Andrew Morton's torture tests):
> > > >   fsx-linux -l 8000000 /mnt/testfile
> > > >   bash-shared-mapping -t 8 /mnt/bashfile 50000000
> > > > (the second test just makes the UML under memory pressure and stresses
> > > > the filesystem, otherwise it does not interact with fsx-linux in any
> > > > way). After some time (like an hour) fsx-linux reported the file is
> > > > corrupted. I tried again and it happened again so probably some
> > > > debugging should be possible.
> > > >   Both times it seems we've simply completely lost a write which
> > > > happened through mmap (2 pages in the first case, 3 pages in the second
> > > > case). Also I've checked and in the first case no blocks are allocated
> > > > for the offsets where the data should be so most probably we've lost
> > > > the write before block_write_full_page() called get_block().
> > > >   I'll debug this further but I wanted let people know there's some
> > > > problem and maybe somebody has some bright idea :).  I'm attaching the
> > > > log from fsx if someone is interested.
> > >
> > >   Testing a bit more, I managed to reproduce the problem on ext2 and
> > > what's more strange, now the lost page was written via ordinary write()
> > > (fsxlog attached). So I believe this is more likely to be UML specific...
> >
> >   And to add even more information, this also happens on ext2 with 4KB
> > blocksize (although much more rarely it seems). Again the data was written
> > by an extending write() but the block for it was not even allocated...
> 
> What block device driver are you using?
  UML was just using image file to back the filesystem I was testing on.
But I don't think that plays a big role because the blocks were not even
allocated in the fs-image so we must have lost them quite early.

> Can it be reproduced without mapped reads and writes completely? (-W -R)
  Good idea, will try.

									Honza

-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR