On Thu, Dec 05, 2013 at 05:06:52PM -0600, Mark Tinguely wrote:
> On 12/05/13 16:32, Dave Chinner wrote:
> >On Thu, Dec 05, 2013 at 03:40:07PM -0600, Mark Tinguely wrote:
> >>If freeing an extent fails during recovery, then the filesystem
> >>will be forced down with the EFI entry still on the AIL. This
> >>will result in hanging the function xfs_ail_push_all_sync().
> >>This patch is similar to the patches that removed the dquot and
> >>inode in commits 32ce90a and dea9609.
> >>Found by mounting an metadata dump that triggers a
> >>XFS_WANT_CORRUPTED_RETURN() on log recovery.
> >>Signed-off-by: Mark Tinguely<tinguely@xxxxxxx>
> >> fs/xfs/xfs_log_recover.c | 5 ++++-
> >> 1 file changed, 4 insertions(+), 1 deletion(-)
> >>Index: b/fs/xfs/xfs_log_recover.c
> >>--- a/fs/xfs/xfs_log_recover.c
> >>+++ b/fs/xfs/xfs_log_recover.c
> >>@@ -3668,7 +3668,7 @@ xlog_recover_process_efi(
> >> extp =&(efip->efi_format.efi_extents[i]);
> >> error = xfs_free_extent(tp, extp->ext_start, extp->ext_len);
> >> if (error)
> >>- goto abort_error;
> >>+ goto free_abort;
> >> xfs_trans_log_efd_extent(tp, efdp, extp->ext_start,
> >> extp->ext_len);
> >> }
> >>@@ -3677,6 +3677,9 @@ xlog_recover_process_efi(
> >> error = xfs_trans_commit(tp, 0);
> >> return error;
> >>+ set_bit(XFS_EFI_RECOVERED,&efip->efi_flags);
> >>+ xfs_efi_release(efip, efip->efi_format.efi_nextents);
> >> abort_error:
> >> xfs_trans_cancel(tp, XFS_TRANS_ABORT);
> >> return error;
> >Why does a failure of xfs_trans_reserve() in this function
> >not require the XFS_EFI_RECOVERED bit to be set?
> >If it does require setting the XFS_EFI_RECOVERED bit (I think it
> >does), then we unconditionally set that bit in
> >xlog_recover_process_efi() so it should be pulled up to be the first
> >thing the function does, not something that is handled in the error
> >paths. IOWs, we leave this function having guaranteed that we've
> >pulled the EFI from the AIL on any failure.
> or not if we move where we remove the entries from the AIL....
Sorry, I don't understand what this means. Can you explain in more
> >Bigger picture: if we fail to recover this EFI, we abort the AIL
> >walk across all the EFIs in the AIL. That means all unrecovered EFIs
> >get left in the AIL and this will result in hanging the function
> >xfs_ail_push_all_sync(). IOWs, this patch only fixes the case where
> >we get an error on the last EFI in the AIL during recovery.
> >Hence a larger fix is needed here - if we fail to recover an EFI, we
> >need to continue to walk the AIL and set the XFS_EFI_RECOVERED bit
> >on all the EFIs still in the AIL and release them.
> ...pretty sure the forced shutdown will happen later and the caller,
The forced shutdown occurs within the context of the
xlog_recover_process_efis() i.e. immediately on detection of the
corruption error in xfs_free_extent().
> is already walking the ail with a
> xfs_ail_cursor. As you mention, there can be more entries, then
> xlog_recover_process_efis() could continue to use the cursor to the
> remove the remaining entries from the AIL on error.
I'm not sure that will be sufficient, because the EFIs that are in
the AIL due to references from the EFD are relying on the EFD to
take the last reference away from the EFI and hence free it. When
the shutdown occurs, the EFDs are aborted via ->iop_committed()
after the XFS_LI_ABORTED flag is set on the EFD log items and this
struct xfs_log_item *lip,
struct xfs_efd_log_item *efdp = EFD_ITEM(lip);
* If we got a log I/O error, it's always the case that the LR with the
* EFI got unpinned and freed before the EFD got aborted.
if (!(lip->li_flags & XFS_LI_ABORTED))
i.e. the EFIs don't get released, and so continue to live in the AIL
even though they have the XFS_EFI_RECOVERED bit set on them.
Hence you have to walk the entire AIL again, releasing all EFIs that
have the XFS_EFI_RECOVERED set, and for those that don't have it set
it needs to be set and the EFI released....