On Thu, Dec 06, 2012 at 05:39:17PM -0500, Jeff Mahoney wrote:
> When using lvconvert to convert a linear mapping to a dm-raid1 mirror,
> we encountered issues where the log would be flooded with messages like:
>
> metadata I/O error: block 0xee7060 ("xfs_trans_read_buf") error 11 numblks 8
>
> The cause is that dm-mirror (and striping, and others) will return
> -EWOULDBLOCK for readahead requests while the mirror is rebuilding.
That's nasty - since when has DM been doing this? I doubt anything
handles a EAGAIN error from the storage layer properly - it's not
an error the filesystem expects from the lower layers at all.
> XFS's
> end_io routine caches the errno and then xfs_buf_iowait bails out early
> when it encounters it after issuing the i/o request.
That doesn't sound right. when XFS issues buffer readahead, it does
not wait for it to complete. i.e. we never get to xfs_buf_iowait()
on readahead buffers.
If something then issues a read on the buffer that failed the
readahead, then we enter xfs_buf_iowait() after reissuing the IO.
If it's aborting because of a stale EWOULDBLOCK as a result of
readahead, then the problem is either:
- failed readahead should not be leaving an error in
b_error; or
- the read IO did not zero b_error before starting the IO
> The I/O eventually
> succeeds and the endio routine resets bp->b_error,
AFAICT, it's a different IO that succeeds (i.e. the resubmitted one
that is being waited for), not the same one.
> but the original read
> request has already returned -EWOULDBLOCK to the user and added the log
> message above to the kernel log, freaking everyone out.
>
> This patch ignores EWOULDBLOCK when deciding whether to wait for the I/O
> to complete and tries again, allowing the read to succeed as expected.
Which does not appear to be the correct fix - preventing failed
readahead from leaving a stale error on the buffer seems like the
right thing to do here...
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|