Eric Sandeen wrote:
> Lachlan McIlroy wrote:
>> There's a bug in _xfs_buf_find() that will cause it to return buffers
>> that failed to be initialised.
>>
>> If a thread has a buffer locked and is waiting for I/O to initialise
>> it and another thread wants the same buffer the second thread will
>> wait on the buffer lock in _xfs_buf_find(). If the initial thread
>> gets an I/O error it marks the buffer in error and releases the
>> buffer lock. The second thread gets the buffer lock, assumes the
>> buffer has been successfully initialised, and then tries to use it.
>>
>> Some callers of xfs_buf_get_flags() will check for B_DONE, and if
>> it's not set then re-issue the I/O, bust most callers assume the
>> buffer and it's contents are good and then use the uninitialised
>> data.
>>
>> The solution I've come up with is if we lookup a buffer and find
>> it's got b_error set or has been marked stale then unhash it from
>> the buffer hashtable and retry the lookup. Also if we fail to setup
>> the buffer correctly in xfs_buf_get_flags() then mark the buffer in
>> error and unhash it. If the buffer is marked stale then in
>> xfs_buf_free() inform the page cache that the contents of the pages
>> are no longer valid.
>
> I managed to come up with a sorta-kinda testcase for this.
>
> Fragmented freespace, many files in a dir, on raid5; simply doing
> drop caches / ls in a loop triggered it.
>
> I guess raid5 is bad in this respect; in it's make_request() we have:
>
> } else {
> /* cannot get stripe for read-ahead, just give-up */
> clear_bit(BIO_UPTODATE, &bi->bi_flags);
> finish_wait(&conf->wait_for_overlap, &w);
> break;
> }
>
> and this happens fairly often. This probably explains a large
> percentage of our xfs_da_do_buf(2) errors we've seen on the list.
>
> From my testing, I think this suffices - and interestingly, Lachlan's
> original patch doesn't seem to help...
>
> Comments?
>
> Maybe could clean up the logic a bit... should this only be
> tested for XBF_READ buffers as well ... or maybe an assert that
> if !uptodate, error should be set ...
>
> diff --git a/fs/xfs/linux-2.6/xfs_buf.c b/fs/xfs/linux-2.6/xfs_buf.c
> index 965df12..cbc0541 100644
> --- a/fs/xfs/linux-2.6/xfs_buf.c
> +++ b/fs/xfs/linux-2.6/xfs_buf.c
> @@ -1142,6 +1165,8 @@ xfs_buf_bio_end_io(
> if (unlikely(bp->b_error)) {
> if (bp->b_flags & XBF_READ)
> ClearPageUptodate(page);
> + } else if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) {
> + ClearPageUptodate(bp);
> } else if (blocksize >= PAGE_CACHE_SIZE) {
> SetPageUptodate(page);
> } else if (!PagePrivate(page) &&
Ok, so that was shoot-from-the-hip, and actually it was tested on an
older kernel; upstream didn't demonstrate the problem, thanks to:
commit c2b00852fbae4f8c45c2651530ded3bd01bde814
Author: NeilBrown <neilb@xxxxxxx>
Date: Sun Dec 10 02:20:51 2006 -0800
[PATCH] md: return a non-zero error to bi_end_io as appropriate in raid5
Currently raid5 depends on clearing the BIO_UPTODATE flag to signal an
error
to higher levels. While this should be sufficient, it is safer to
explicitly
set the error code as well - less room for confusion.
Signed-off-by: Neil Brown <neilb@xxxxxxx>
Signed-off-by: Andrew Morton <akpm@xxxxxxxx>
Signed-off-by: Linus Torvalds <torvalds@xxxxxxxx>
As Neil says, xfs should probably cope too by checking uptodate itself
as well, though...
-Eric
|