On Thu, May 10, 2018 at 12:18:32PM -0700, Darrick J. Wong wrote:
> From: Darrick J. Wong <darrick.wong@xxxxxxxxxx>
>
> In normal operation, the XFS convention is to take an inode's iolock
> and then allocate a transaction. However, when scrubbing parent inodes
> this is inverted -- we allocated the transaction to do the scrub, and
> now we're trying to grab the parent's iolock. This can lead to ABBA
> deadlocks: some thread grabbed the parent's iolock and is waiting for
> space for a transaction while our parent scrubber is sitting on a
> transaction trying to get the parent's iolock.
>
> Therefore, convert all iolock attempts to use trylock; if that fails,
> they can use the existing mechanisms to back off and try again.
>
> The ABBA deadlock didn't happen with a non-repair scrub because the
> transactions don't reserve any space, but repair scrubs require
> reservation in order to update metadata. However, any other concurrent
> metadata update (e.g. directory create in the parent) could also induce
> this deadlock with the parent scrubber.
>
> Signed-off-by: Darrick J. Wong <darrick.wong@xxxxxxxxxx>
> ---
> fs/xfs/scrub/common.c | 22 ++++++++++++++++++++++
> fs/xfs/scrub/common.h | 1 +
> fs/xfs/scrub/parent.c | 16 ++++++++++++++--
> 3 files changed, 37 insertions(+), 2 deletions(-)
>
>
> diff --git a/fs/xfs/scrub/common.c b/fs/xfs/scrub/common.c
> index 62b33c99efe4..518bff2be0c9 100644
> --- a/fs/xfs/scrub/common.c
> +++ b/fs/xfs/scrub/common.c
> @@ -844,3 +844,25 @@ xfs_scrub_metadata_inode_forks(
>
> return error;
> }
> +
> +/*
> + * Try to lock an inode in violation of the usual locking order rules. For
> + * example, trying to get the IOLOCK while in transaction context, or just
> + * plain breaking AG-order or inode-order inode locking rules. Either way,
> + * the only way to avoid an ABBA deadlock is to use trylock and back off if
> + * we can't.
> + */
> +int
> +xfs_scrub_ilock_inverted(
> + struct xfs_inode *ip,
> + uint lock_mode)
> +{
> + int i;
> +
> + for (i = 0; i < 20; i++) {
> + if (xfs_ilock_nowait(ip, lock_mode))
> + return 0;
> + delay(1);
> + }
> + return -EDEADLOCK;
> +}
This is definitely hacky. It would be nice if we could come up with
something cleaner to explicitly unwind or something. While this may
address the issue described in the commit log, I wonder whether these
kind of loops introduce the possibility of really long runtimes due to
racing with parent changes and whatnot (though I guess the "try_again"
thing is also capped to 20 retries, but when you factor them all
together per-inode...).
Anyways, experimental code and a better solution probably requires more
thought:
Reviewed-by: Brian Foster <bfoster@xxxxxxxxxx>
> diff --git a/fs/xfs/scrub/common.h b/fs/xfs/scrub/common.h
> index 5d78bb9602ab..119d9b6db887 100644
> --- a/fs/xfs/scrub/common.h
> +++ b/fs/xfs/scrub/common.h
> @@ -156,5 +156,6 @@ static inline bool xfs_scrub_skip_xref(struct xfs_scrub_metadata *sm)
> }
>
> int xfs_scrub_metadata_inode_forks(struct xfs_scrub_context *sc);
> +int xfs_scrub_ilock_inverted(struct xfs_inode *ip, uint lock_mode);
>
> #endif /* __XFS_SCRUB_COMMON_H__ */
> diff --git a/fs/xfs/scrub/parent.c b/fs/xfs/scrub/parent.c
> index fc336807e156..77c6b22c6bfd 100644
> --- a/fs/xfs/scrub/parent.c
> +++ b/fs/xfs/scrub/parent.c
> @@ -214,7 +214,9 @@ xfs_scrub_parent_validate(
> */
> xfs_iunlock(sc->ip, sc->ilock_flags);
> sc->ilock_flags = 0;
> - xfs_ilock(dp, XFS_IOLOCK_SHARED);
> + error = xfs_scrub_ilock_inverted(dp, XFS_IOLOCK_SHARED);
> + if (error)
> + goto out_rele;
>
> /* Go looking for our dentry. */
> error = xfs_scrub_parent_count_parent_dentries(sc, dp, &nlink);
> @@ -223,8 +225,10 @@ xfs_scrub_parent_validate(
>
> /* Drop the parent lock, relock this inode. */
> xfs_iunlock(dp, XFS_IOLOCK_SHARED);
> + error = xfs_scrub_ilock_inverted(sc->ip, XFS_IOLOCK_EXCL);
> + if (error)
> + goto out_rele;
> sc->ilock_flags = XFS_IOLOCK_EXCL;
> - xfs_ilock(sc->ip, sc->ilock_flags);
>
> /*
> * If we're an unlinked directory, the parent /won't/ have a link
> @@ -326,5 +330,13 @@ xfs_scrub_parent(
> if (try_again && tries == 20)
> xfs_scrub_set_incomplete(sc);
> out:
> + /*
> + * If we failed to lock the parent inode even after a retry, just mark
> + * this scrub incomplete and return.
> + */
> + if (sc->try_harder && error == -EDEADLOCK) {
> + error = 0;
> + xfs_scrub_set_incomplete(sc);
> + }
> return error;
> }
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
