On Tue, Mar 24, 2009 at 03:07:21PM +0100, Jan Kara wrote: > On Tue 24-03-09 10:01:45, Chris Mason wrote: > > On Tue, 2009-03-24 at 14:26 +0100, Jan Kara wrote: > > > On Tue 24-03-09 13:55:10, Jan Kara wrote: > > > > > > And one more interesting thing I don't yet fully understand - I see pages > > > > having PageError() set when they are removed from page cache (and they have > > > > been faulted in before). It's probably some interaction with pagecache > > > > readahead... > > > Argh... So the problem seems to be that get_block() occasionally returns > > > ENOSPC and we then discard the dirty data (hmm, we could give at least a > > > warning for that). I'm not yet sure why getblock behaves like this because > > > the filesystem seems to have enough space but anyway this seems to be some > > > strange fs trouble as well. > > > > > > > Ouch. Perhaps the free space is waiting on a journal commit? > Yes, exactly. I've already found there's lot of space hold by the > committing transaction (it can easily hold a few hundred megs or a few gigs > with larger journal and my UML images aren't that big...). And writepage() > implementation in ext3 does not have a logic to retry. Also > block_write_full_page() clears buffers dirty bits so it's not easy to retry > even if we did it. I'm now looking into how to fix this... We retry block allocation in ext3_write_begin. And for mmap we should be doing something similar to ext4_page_mkwrite so that we can be sure that during writepage we don't need to do block allocation. -aneesh -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html