[PATCH v2 0/3] xfs: rework quotaoff to avoid log deadlock

[PATCH v2 0/3] xfs: rework quotaoff to avoid log deadlock

[Brian Foster]

This series reworks the quotaoff algorithm to eliminate a log
reservation deadlock vector. v2 is primarily a rebase/repost of v1.

Note that this conflicts with the recently posted append ioend cleanup,
but this series is still based on for-next. I can post another rebased
series if either of these progress.

Brian

v2:
- Rebased to for-next.
v1: https://lore.kernel.org/linux-xfs/20201001150310.141467-1-bfoster@redhat.com/

Brian Foster (3):
  xfs: skip dquot reservations if quota is inactive
  xfs: transaction subsystem quiesce mechanism
  xfs: rework quotaoff logging to avoid log deadlock on active fs

 fs/xfs/xfs_aops.c        |   2 +
 fs/xfs/xfs_mount.h       |   3 +
 fs/xfs/xfs_qm_syscalls.c | 133 +++++++++++++++++++--------------------
 fs/xfs/xfs_super.c       |   8 +++
 fs/xfs/xfs_trans.c       |   4 +-
 fs/xfs/xfs_trans.h       |  20 ++++++
 fs/xfs/xfs_trans_dquot.c |  22 +++----
 7 files changed, 111 insertions(+), 81 deletions(-)

-- 
2.26.3

[PATCH v2 1/3] xfs: skip dquot reservations if quota is inactive

[Brian Foster]

The dquot reservation helper currently performs the associated
reservation for any provided dquots. The dquots could have been
acquired from inode references or explicit dquot allocation
requests. Some reservation callers may have already checked that the
associated quota subsystem is active (xfs_qm_dqget() returns an
error otherwise), while others might not have checked at all
(xfs_trans_reserve_quota_nblks() passes the inode references).
Further, subsequent dquot modifications do actually check that the
associated quota is active before making transactional changes
(xfs_trans_mod_dquot_byino()).

Given all of that, the behavior to unconditionally perform
reservation on any provided dquots is somewhat ad hoc. While it is
currently harmless, it is not without side effect. If the quota is
inactive by the time a transaction attempts a quota reservation, the
dquot will be attached to the transaction and subsequently logged,
even though no dquot modifications are ultimately made.

This is a problem for upcoming quotaoff changes that intend to
implement a strict transactional barrier for logging dquots during a
quotaoff operation. If a dquot is logged after the subsystem
deactivated and the barrier released, a subsequent log recovery can
incorrectly replay dquot changes into the filesystem.

Therefore, update the dquot reservation path to also check that a
particular quota mode is active before associating a dquot with a
transaction. This should have no noticeable impact on the current
code that already accommodates checking active quota state at points
before and after quota reservations are made.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
---
 fs/xfs/xfs_trans_dquot.c | 20 +++++++++-----------
 1 file changed, 9 insertions(+), 11 deletions(-)

diff --git a/fs/xfs/xfs_trans_dquot.c b/fs/xfs/xfs_trans_dquot.c
index 48e09ea30ee5..eec640999148 100644
--- a/fs/xfs/xfs_trans_dquot.c
+++ b/fs/xfs/xfs_trans_dquot.c
@@ -40,14 +40,12 @@ xfs_trans_dqjoin(
 }
 
 /*
- * This is called to mark the dquot as needing
- * to be logged when the transaction is committed.  The dquot must
- * already be associated with the given transaction.
- * Note that it marks the entire transaction as dirty. In the ordinary
- * case, this gets called via xfs_trans_commit, after the transaction
- * is already dirty. However, there's nothing stop this from getting
- * called directly, as done by xfs_qm_scall_setqlim. Hence, the TRANS_DIRTY
- * flag.
+ * This is called to mark the dquot as needing to be logged when the transaction
+ * is committed. The dquot must already be associated with the given
+ * transaction. Note that it marks the entire transaction as dirty. In the
+ * ordinary case, this gets called via xfs_trans_commit, after the transaction
+ * is already dirty. However, there's nothing stop this from getting called
+ * directly, as done by xfs_qm_scall_setqlim. Hence, the TRANS_DIRTY flag.
  */
 void
 xfs_trans_log_dquot(
@@ -743,19 +741,19 @@ xfs_trans_reserve_quota_bydquots(
 
 	ASSERT(flags & XFS_QMOPT_RESBLK_MASK);
 
-	if (udqp) {
+	if (XFS_IS_UQUOTA_ON(mp) && udqp) {
 		error = xfs_trans_dqresv(tp, mp, udqp, nblks, ninos, flags);
 		if (error)
 			return error;
 	}
 
-	if (gdqp) {
+	if (XFS_IS_GQUOTA_ON(mp) && gdqp) {
 		error = xfs_trans_dqresv(tp, mp, gdqp, nblks, ninos, flags);
 		if (error)
 			goto unwind_usr;
 	}
 
-	if (pdqp) {
+	if (XFS_IS_PQUOTA_ON(mp) && pdqp) {
 		error = xfs_trans_dqresv(tp, mp, pdqp, nblks, ninos, flags);
 		if (error)
 			goto unwind_grp;
-- 
2.26.3

Re: [PATCH v2 1/3] xfs: skip dquot reservations if quota is inactive

[Christoph Hellwig]

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH v2 1/3] xfs: skip dquot reservations if quota is inactive

[Darrick J. Wong]

On Tue, Apr 06, 2021 at 10:42:36AM -0400, Brian Foster wrote:
> The dquot reservation helper currently performs the associated
> reservation for any provided dquots. The dquots could have been
> acquired from inode references or explicit dquot allocation
> requests. Some reservation callers may have already checked that the
> associated quota subsystem is active (xfs_qm_dqget() returns an
> error otherwise), while others might not have checked at all
> (xfs_trans_reserve_quota_nblks() passes the inode references).
> Further, subsequent dquot modifications do actually check that the
> associated quota is active before making transactional changes
> (xfs_trans_mod_dquot_byino()).
> 
> Given all of that, the behavior to unconditionally perform
> reservation on any provided dquots is somewhat ad hoc. While it is
> currently harmless, it is not without side effect. If the quota is
> inactive by the time a transaction attempts a quota reservation, the
> dquot will be attached to the transaction and subsequently logged,
> even though no dquot modifications are ultimately made.
> 
> This is a problem for upcoming quotaoff changes that intend to
> implement a strict transactional barrier for logging dquots during a
> quotaoff operation. If a dquot is logged after the subsystem
> deactivated and the barrier released, a subsequent log recovery can
> incorrectly replay dquot changes into the filesystem.
> 
> Therefore, update the dquot reservation path to also check that a
> particular quota mode is active before associating a dquot with a
> transaction. This should have no noticeable impact on the current
> code that already accommodates checking active quota state at points
> before and after quota reservations are made.
> 
> Signed-off-by: Brian Foster <bfoster@redhat.com>
> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>

If you end up reposting this, please change to:

Reviewed-by: Darrick J. Wong <djwong@kernel.org>

--D

> ---
>  fs/xfs/xfs_trans_dquot.c | 20 +++++++++-----------
>  1 file changed, 9 insertions(+), 11 deletions(-)
> 
> diff --git a/fs/xfs/xfs_trans_dquot.c b/fs/xfs/xfs_trans_dquot.c
> index 48e09ea30ee5..eec640999148 100644
> --- a/fs/xfs/xfs_trans_dquot.c
> +++ b/fs/xfs/xfs_trans_dquot.c
> @@ -40,14 +40,12 @@ xfs_trans_dqjoin(
>  }
>  
>  /*
> - * This is called to mark the dquot as needing
> - * to be logged when the transaction is committed.  The dquot must
> - * already be associated with the given transaction.
> - * Note that it marks the entire transaction as dirty. In the ordinary
> - * case, this gets called via xfs_trans_commit, after the transaction
> - * is already dirty. However, there's nothing stop this from getting
> - * called directly, as done by xfs_qm_scall_setqlim. Hence, the TRANS_DIRTY
> - * flag.
> + * This is called to mark the dquot as needing to be logged when the transaction
> + * is committed. The dquot must already be associated with the given
> + * transaction. Note that it marks the entire transaction as dirty. In the
> + * ordinary case, this gets called via xfs_trans_commit, after the transaction
> + * is already dirty. However, there's nothing stop this from getting called
> + * directly, as done by xfs_qm_scall_setqlim. Hence, the TRANS_DIRTY flag.
>   */
>  void
>  xfs_trans_log_dquot(
> @@ -743,19 +741,19 @@ xfs_trans_reserve_quota_bydquots(
>  
>  	ASSERT(flags & XFS_QMOPT_RESBLK_MASK);
>  
> -	if (udqp) {
> +	if (XFS_IS_UQUOTA_ON(mp) && udqp) {
>  		error = xfs_trans_dqresv(tp, mp, udqp, nblks, ninos, flags);
>  		if (error)
>  			return error;
>  	}
>  
> -	if (gdqp) {
> +	if (XFS_IS_GQUOTA_ON(mp) && gdqp) {
>  		error = xfs_trans_dqresv(tp, mp, gdqp, nblks, ninos, flags);
>  		if (error)
>  			goto unwind_usr;
>  	}
>  
> -	if (pdqp) {
> +	if (XFS_IS_PQUOTA_ON(mp) && pdqp) {
>  		error = xfs_trans_dqresv(tp, mp, pdqp, nblks, ninos, flags);
>  		if (error)
>  			goto unwind_grp;
> -- 
> 2.26.3
> 

[PATCH v2 2/3] xfs: transaction subsystem quiesce mechanism

[Brian Foster]

The updated quotaoff logging algorithm depends on a runtime quiesce
of the transaction subsystem to guarantee all transactions after a
certain point detect quota subsystem changes. Implement this
mechanism using an internal lock, similar to the external filesystem
freeze mechanism. This is also somewhat analogous to the old percpu
transaction counter mechanism, but we don't actually need a counter.

Signed-off-by: Brian Foster <bfoster@redhat.com>
---
 fs/xfs/xfs_aops.c  |  2 ++
 fs/xfs/xfs_mount.h |  3 +++
 fs/xfs/xfs_super.c |  8 ++++++++
 fs/xfs/xfs_trans.c |  4 ++--
 fs/xfs/xfs_trans.h | 20 ++++++++++++++++++++
 5 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 1cc7c36d98e9..dce52943e5a7 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -58,6 +58,7 @@ xfs_setfilesize_trans_alloc(
 	 * we released it.
 	 */
 	__sb_writers_release(ioend->io_inode->i_sb, SB_FREEZE_FS);
+	percpu_rwsem_release(&mp->m_trans_rwsem, true, _THIS_IP_);
 	/*
 	 * We hand off the transaction to the completion thread now, so
 	 * clear the flag here.
@@ -127,6 +128,7 @@ xfs_setfilesize_ioend(
 	 */
 	xfs_trans_set_context(tp);
 	__sb_writers_acquired(VFS_I(ip)->i_sb, SB_FREEZE_FS);
+	percpu_rwsem_acquire(&ip->i_mount->m_trans_rwsem, true, _THIS_IP_);
 
 	/* we abort the update if there was an IO error */
 	if (error) {
diff --git a/fs/xfs/xfs_mount.h b/fs/xfs/xfs_mount.h
index 81829d19596e..27a2a53abb4f 100644
--- a/fs/xfs/xfs_mount.h
+++ b/fs/xfs/xfs_mount.h
@@ -171,6 +171,9 @@ typedef struct xfs_mount {
 	 */
 	struct percpu_counter	m_delalloc_blks;
 
+	/* lock for transaction quiesce (used by quotaoff) */
+	struct percpu_rw_semaphore	m_trans_rwsem;
+
 	struct radix_tree_root	m_perag_tree;	/* per-ag accounting info */
 	spinlock_t		m_perag_lock;	/* lock for m_perag_tree */
 	uint64_t		m_resblks;	/* total reserved blocks */
diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c
index 8d079c5e7099..64feab042dea 100644
--- a/fs/xfs/xfs_super.c
+++ b/fs/xfs/xfs_super.c
@@ -1001,8 +1001,15 @@ xfs_init_percpu_counters(
 	if (error)
 		goto free_fdblocks;
 
+	/* not a counter, but close enough... */
+	error = percpu_init_rwsem(&mp->m_trans_rwsem);
+	if (error)
+		goto free_delalloc;
+
 	return 0;
 
+free_delalloc:
+	percpu_counter_destroy(&mp->m_delalloc_blks);
 free_fdblocks:
 	percpu_counter_destroy(&mp->m_fdblocks);
 free_ifree:
@@ -1025,6 +1032,7 @@ static void
 xfs_destroy_percpu_counters(
 	struct xfs_mount	*mp)
 {
+	percpu_free_rwsem(&mp->m_trans_rwsem);
 	percpu_counter_destroy(&mp->m_icount);
 	percpu_counter_destroy(&mp->m_ifree);
 	percpu_counter_destroy(&mp->m_fdblocks);
diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c
index bc25afc10245..c46943f0fc77 100644
--- a/fs/xfs/xfs_trans.c
+++ b/fs/xfs/xfs_trans.c
@@ -74,7 +74,7 @@ xfs_trans_free(
 	trace_xfs_trans_free(tp, _RET_IP_);
 	xfs_trans_clear_context(tp);
 	if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
-		sb_end_intwrite(tp->t_mountp->m_super);
+		xfs_trans_end(tp->t_mountp);
 	xfs_trans_free_dqinfo(tp);
 	kmem_cache_free(xfs_trans_zone, tp);
 }
@@ -265,7 +265,7 @@ xfs_trans_alloc(
 retry:
 	tp = kmem_cache_zalloc(xfs_trans_zone, GFP_KERNEL | __GFP_NOFAIL);
 	if (!(flags & XFS_TRANS_NO_WRITECOUNT))
-		sb_start_intwrite(mp->m_super);
+		xfs_trans_start(mp);
 	xfs_trans_set_context(tp);
 
 	/*
diff --git a/fs/xfs/xfs_trans.h b/fs/xfs/xfs_trans.h
index 9dd745cf77c9..95da3e179150 100644
--- a/fs/xfs/xfs_trans.h
+++ b/fs/xfs/xfs_trans.h
@@ -226,6 +226,26 @@ xfs_trans_read_buf(
 				      flags, bpp, ops);
 }
 
+/*
+ * Context tracking helpers for external (i.e. fs freeze) and internal
+ * transaction quiesce.
+ */
+static inline void
+xfs_trans_start(
+	struct xfs_mount	*mp)
+{
+	sb_start_intwrite(mp->m_super);
+	percpu_down_read(&mp->m_trans_rwsem);
+}
+
+static inline void
+xfs_trans_end(
+	struct xfs_mount	*mp)
+{
+	percpu_up_read(&mp->m_trans_rwsem);
+	sb_end_intwrite(mp->m_super);
+}
+
 struct xfs_buf	*xfs_trans_getsb(struct xfs_trans *);
 
 void		xfs_trans_brelse(xfs_trans_t *, struct xfs_buf *);
-- 
2.26.3

Re: [PATCH v2 2/3] xfs: transaction subsystem quiesce mechanism

[Christoph Hellwig]

On Tue, Apr 06, 2021 at 10:42:37AM -0400, Brian Foster wrote:
> The updated quotaoff logging algorithm depends on a runtime quiesce
> of the transaction subsystem to guarantee all transactions after a
> certain point detect quota subsystem changes. Implement this
> mechanism using an internal lock, similar to the external filesystem
> freeze mechanism. This is also somewhat analogous to the old percpu
> transaction counter mechanism, but we don't actually need a counter.

Stupid question that already came up when seeing the replies to my
s_inodes patch:  Why do we even care about quotaoff?  Is there any
real life use case for quotaoff, at least the kind that disables
accounting (vs enforcement)?  IMHO we spend a lot of effort on this
corner case that has no practical value, and just removing support
for quotaoff might serve us much better in the long run.

Re: [PATCH v2 2/3] xfs: transaction subsystem quiesce mechanism

[Brian Foster]

On Wed, Apr 07, 2021 at 09:00:41AM +0100, Christoph Hellwig wrote:
> On Tue, Apr 06, 2021 at 10:42:37AM -0400, Brian Foster wrote:
> > The updated quotaoff logging algorithm depends on a runtime quiesce
> > of the transaction subsystem to guarantee all transactions after a
> > certain point detect quota subsystem changes. Implement this
> > mechanism using an internal lock, similar to the external filesystem
> > freeze mechanism. This is also somewhat analogous to the old percpu
> > transaction counter mechanism, but we don't actually need a counter.
> 
> Stupid question that already came up when seeing the replies to my
> s_inodes patch:  Why do we even care about quotaoff?  Is there any
> real life use case for quotaoff, at least the kind that disables
> accounting (vs enforcement)?  IMHO we spend a lot of effort on this
> corner case that has no practical value, and just removing support
> for quotaoff might serve us much better in the long run.
> 

Hm, fair point. I think the historical fragility and complexity makes it
reasonable to question whether it's worth continued support. Looking
back through my notes, ISTM that the original report of the log
reservation deadlock came from fstests, so not necessarily an end user
report. I'm not aware of any real user reports around quotaoff, but then
again it's fairly boring functionality that probably just works most of
the time. It's kind of hard to surmise external dependencies from that
alone.

Personally, I'd probably have to think about it some more, but initially
I don't have any strong objection to removing quotaoff support. More
practically, I suspect we'd have to deprecate it for some period of time
given that it's a generic interface, has userspace tools, regression
tests, etc., and may or may not have real users who might want the
opportunity to object (or adjust).

Though perhaps potentially avoiding that mess is what you mean by "...
disables accounting vs.  enforcement." I.e., retain the interface and
general ability to turn off enforcement, but require a mount cycle in
the future to disable accounting..? Hmm... that seems like a potentially
nicer/easier path forward and a less disruptive change. I wonder even if
we could just (eventually) ignore the accounting disablement flags from
userspace and if any users would have reason to care about that change
in behavior.

Brian

Re: [PATCH v2 2/3] xfs: transaction subsystem quiesce mechanism

[Christoph Hellwig]

On Wed, Apr 07, 2021 at 07:36:37AM -0400, Brian Foster wrote:
> Personally, I'd probably have to think about it some more, but initially
> I don't have any strong objection to removing quotaoff support. More
> practically, I suspect we'd have to deprecate it for some period of time
> given that it's a generic interface, has userspace tools, regression
> tests, etc., and may or may not have real users who might want the
> opportunity to object (or adjust).
> 
> Though perhaps potentially avoiding that mess is what you mean by "...
> disables accounting vs.  enforcement." I.e., retain the interface and
> general ability to turn off enforcement, but require a mount cycle in
> the future to disable accounting..? Hmm... that seems like a potentially
> nicer/easier path forward and a less disruptive change. I wonder even if
> we could just (eventually) ignore the accounting disablement flags from
> userspace and if any users would have reason to care about that change
> in behavior.

I'm currently testing a series that just ignores disabling of accounting
and logs a message and that seems to do ok so far.  I'll check if
clearing the on-disk flags as well could work out even better.

Re: [PATCH v2 2/3] xfs: transaction subsystem quiesce mechanism

[Darrick J. Wong]

On Wed, Apr 07, 2021 at 02:24:55PM +0100, Christoph Hellwig wrote:
> On Wed, Apr 07, 2021 at 07:36:37AM -0400, Brian Foster wrote:
> > Personally, I'd probably have to think about it some more, but initially
> > I don't have any strong objection to removing quotaoff support. More
> > practically, I suspect we'd have to deprecate it for some period of time
> > given that it's a generic interface, has userspace tools, regression
> > tests, etc., and may or may not have real users who might want the
> > opportunity to object (or adjust).
> > 
> > Though perhaps potentially avoiding that mess is what you mean by "...
> > disables accounting vs.  enforcement." I.e., retain the interface and
> > general ability to turn off enforcement, but require a mount cycle in
> > the future to disable accounting..? Hmm... that seems like a potentially
> > nicer/easier path forward and a less disruptive change. I wonder even if
> > we could just (eventually) ignore the accounting disablement flags from
> > userspace and if any users would have reason to care about that change
> > in behavior.
> 
> I'm currently testing a series that just ignores disabling of accounting
> and logs a message and that seems to do ok so far.  I'll check if
> clearing the on-disk flags as well could work out even better.

While I was rejiggering the inode walk parts of quotaoff I did wonder
why it even mattered to dqpurge the affected dquots **now**.  With patch
1 applied, we could just turn off the _ACTIVE flag and let reclaim
erase them slowly.

--D

[PATCH v2 3/3] xfs: rework quotaoff logging to avoid log deadlock on active fs

[Brian Foster]

The quotaoff operation logs two log items. The start item is
committed first, followed by the bulk of the in-core quotaoff
processing, and then the quotaoff end item is committed to release
the start item from the log. The problem with this mechanism is that
quite a bit of processing can be required to release dquots from all
in-core inodes and subsequently flush/purge all dquots in the
system. This processing work doesn't generally generate much log
traffic itself, but the start item pins the tail of the log. If an
external workload consumes the remaining log space before the
transaction for the end item is allocated, a log deadlock can occur.

The purpose of the separate start and end log items is primarily to
ensure that log recovery does not incorrectly recover dquot data
after an fs crash where a quotaoff was in progress. If we only
logged a single quotaoff item, for example, it could fall behind the
tail of the log before the last dquot modification was made and
incorrectly replay dquot changes that might have occurred after the
start item committed but before quotaoff deactivated the quota.

With that context, we can make some small changes to the quotaoff
algorithm to provide the same general log ordering guarantee without
such a large window to create a log deadlock vector. Rather than
place a start item in the log for the duration of quotaoff
processing, we can quiesce the transaction subsystem up front to
guarantee that no further dquots are logged from that point forward.
IOW, we pause the transaction subsystem, commit the quotaoff start
and end items, deactivate the associated quota such that subsequent
transactions no longer modify associated dquots, and resume the
transaction subsystem. The transaction pause is somewhat of a heavy
weight operation, but quotaoff is already a rare, slow and
performance disruptive operation and the quiesce is only required
for two small transactions.

Altogether, this means that the dquot rele/purge sequence occurs
after the quotaoff end item has committed and thus can technically
fall off the end of the log. This is safe because the remaining
processing is in-core work that doesn't involve logging dquots and
we've guaranteed that no further dquots are modified by external
transactions. This allows quotaoff to complete without risking log
deadlock regardless of how much dquot processing is required.

Suggested-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
---
 fs/xfs/xfs_qm_syscalls.c | 133 +++++++++++++++++++--------------------
 fs/xfs/xfs_trans_dquot.c |   2 +
 2 files changed, 67 insertions(+), 68 deletions(-)

diff --git a/fs/xfs/xfs_qm_syscalls.c b/fs/xfs/xfs_qm_syscalls.c
index ca1b57d291dc..b8e55f4947bd 100644
--- a/fs/xfs/xfs_qm_syscalls.c
+++ b/fs/xfs/xfs_qm_syscalls.c
@@ -29,7 +29,8 @@ xfs_qm_log_quotaoff(
 	int			error;
 	struct xfs_qoff_logitem	*qoffi;
 
-	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_quotaoff, 0, 0, 0, &tp);
+	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_quotaoff, 0, 0,
+				XFS_TRANS_NO_WRITECOUNT, &tp);
 	if (error)
 		goto out;
 
@@ -67,7 +68,8 @@ xfs_qm_log_quotaoff_end(
 	int			error;
 	struct xfs_qoff_logitem	*qoffi;
 
-	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_equotaoff, 0, 0, 0, &tp);
+	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_qm_equotaoff, 0, 0,
+				XFS_TRANS_NO_WRITECOUNT, &tp);
 	if (error)
 		return error;
 
@@ -106,8 +108,8 @@ xfs_qm_scall_quotaoff(
 
 	/*
 	 * No file system can have quotas enabled on disk but not in core.
-	 * Note that quota utilities (like quotaoff) _expect_
-	 * errno == -EEXIST here.
+	 * Note that quota utilities (like quotaoff) _expect_ errno == -EEXIST
+	 * here.
 	 */
 	if ((mp->m_qflags & flags) == 0)
 		return -EEXIST;
@@ -116,17 +118,14 @@ xfs_qm_scall_quotaoff(
 	flags &= (XFS_ALL_QUOTA_ACCT | XFS_ALL_QUOTA_ENFD);
 
 	/*
-	 * We don't want to deal with two quotaoffs messing up each other,
-	 * so we're going to serialize it. quotaoff isn't exactly a performance
+	 * We don't want to deal with two quotaoffs messing up each other, so
+	 * we're going to serialize it. quotaoff isn't exactly a performance
 	 * critical thing.
-	 * If quotaoff, then we must be dealing with the root filesystem.
 	 */
 	ASSERT(q);
 	mutex_lock(&q->qi_quotaofflock);
 
-	/*
-	 * If we're just turning off quota enforcement, change mp and go.
-	 */
+	/* if we're just turning off quota enforcement, change mp and go */
 	if ((flags & XFS_ALL_QUOTA_ACCT) == 0) {
 		mp->m_qflags &= ~(flags);
 
@@ -142,9 +141,9 @@ xfs_qm_scall_quotaoff(
 	dqtype = 0;
 	inactivate_flags = 0;
 	/*
-	 * If accounting is off, we must turn enforcement off, clear the
-	 * quota 'CHKD' certificate to make it known that we have to
-	 * do a quotacheck the next time this quota is turned on.
+	 * If accounting is off, we must turn enforcement off, clear the quota
+	 * 'CHKD' certificate to make it known that we have to do a quotacheck
+	 * the next time this quota is turned on.
 	 */
 	if (flags & XFS_UQUOTA_ACCT) {
 		dqtype |= XFS_QMOPT_UQUOTA;
@@ -163,89 +162,87 @@ xfs_qm_scall_quotaoff(
 	}
 
 	/*
-	 * Nothing to do?  Don't complain. This happens when we're just
-	 * turning off quota enforcement.
+	 * Nothing to do? Don't complain. This happens when we're just turning
+	 * off quota enforcement.
 	 */
 	if ((mp->m_qflags & flags) == 0)
 		goto out_unlock;
 
 	/*
-	 * Write the LI_QUOTAOFF log record, and do SB changes atomically,
-	 * and synchronously. If we fail to write, we should abort the
-	 * operation as it cannot be recovered safely if we crash.
+	 * Quotaoff must deactivate the associated quota mode(s), release dquots
+	 * from inodes and purge them from the system all while the filesystem
+	 * remains active. We have two quotaoff log records that traditionally
+	 * bound the start and end of this sequence. This guarantees that no
+	 * dquots are modified after the end item hits the log, but quotaoff can
+	 * be time consuming and thus prone to deadlock because the start item
+	 * pins the tail the of log in the meantime (and we can't hold the end
+	 * transaction open across the dqrele scan).
+	 *
+	 * The critical aspect of correctly logging quotaoff is that no dquots
+	 * are modified after the quotaoff end item hits the on-disk log.
+	 * Otherwise the quotaoff can fall off the tail and log recovery can
+	 * replay incorrect data. Instead of letting the start item sit in the
+	 * log while quotaoff completes, we can provide the same guarantee via a
+	 * runtime barrier for dquot modifications. Specifically, we pause all
+	 * transactions on the system via the transaction subsystem lock, log
+	 * both start and end items (via sync transactions, which drains the
+	 * CIL), deactivate the quota, and then resume the transaction subsystem
+	 * while quotaoff completes.
+	 *
+	 * This is safe because the remaining quotaoff work is in-core cleanup
+	 * and all subsequent transactions should see the updated quota state
+	 * due to memory ordering provided by the lock. We also avoid deadlock
+	 * by committing both items sequentially with near exclusive access to
+	 * the transaction subsystem.
 	 */
+	percpu_down_write(&mp->m_trans_rwsem);
+
 	error = xfs_qm_log_quotaoff(mp, &qoffstart, flags);
-	if (error)
+	if (error) {
+		percpu_up_write(&mp->m_trans_rwsem);
 		goto out_unlock;
+	}
 
-	/*
-	 * Next we clear the XFS_MOUNT_*DQ_ACTIVE bit(s) in the mount struct
-	 * to take care of the race between dqget and quotaoff. We don't take
-	 * any special locks to reset these bits. All processes need to check
-	 * these bits *after* taking inode lock(s) to see if the particular
-	 * quota type is in the process of being turned off. If *ACTIVE, it is
-	 * guaranteed that all dquot structures and all quotainode ptrs will all
-	 * stay valid as long as that inode is kept locked.
-	 *
-	 * There is no turning back after this.
-	 */
 	mp->m_qflags &= ~inactivate_flags;
 
+	error = xfs_qm_log_quotaoff_end(mp, &qoffstart, flags);
+	if (error) {
+		percpu_up_write(&mp->m_trans_rwsem);
+		/* We're screwed now. Shutdown is the only option. */
+		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
+		goto out_unlock;
+	}
+
+	percpu_up_write(&mp->m_trans_rwsem);
+
 	/*
-	 * Give back all the dquot reference(s) held by inodes.
-	 * Here we go thru every single incore inode in this file system, and
-	 * do a dqrele on the i_udquot/i_gdquot that it may have.
-	 * Essentially, as long as somebody has an inode locked, this guarantees
-	 * that quotas will not be turned off. This is handy because in a
-	 * transaction once we lock the inode(s) and check for quotaon, we can
-	 * depend on the quota inodes (and other things) being valid as long as
-	 * we keep the lock(s).
+	 * Release dquot references held by inodes. Technically some contexts
+	 * might not pick up the quota state change until the inode lock is
+	 * cycled if there is no transaction. We don't care about that above
+	 * because a dquot can't be logged without a transaction and we can't
+	 * release/purge a dquot here until we've cycled the locks of all inodes
+	 * that reference it.
 	 */
 	xfs_qm_dqrele_all_inodes(mp, flags);
 
 	/*
 	 * Next we make the changes in the quota flag in the mount struct.
-	 * This isn't protected by a particular lock directly, because we
-	 * don't want to take a mrlock every time we depend on quotas being on.
+	 * This isn't protected by a particular lock directly, because we don't
+	 * want to take a mrlock every time we depend on quotas being on.
 	 */
 	mp->m_qflags &= ~flags;
 
-	/*
-	 * Go through all the dquots of this file system and purge them,
-	 * according to what was turned off.
-	 */
+	/* purge all deactivated dquots from the filesystem */
 	xfs_qm_dqpurge_all(mp, dqtype);
 
-	/*
-	 * Transactions that had started before ACTIVE state bit was cleared
-	 * could have logged many dquots, so they'd have higher LSNs than
-	 * the first QUOTAOFF log record does. If we happen to crash when
-	 * the tail of the log has gone past the QUOTAOFF record, but
-	 * before the last dquot modification, those dquots __will__
-	 * recover, and that's not good.
-	 *
-	 * So, we have QUOTAOFF start and end logitems; the start
-	 * logitem won't get overwritten until the end logitem appears...
-	 */
-	error = xfs_qm_log_quotaoff_end(mp, &qoffstart, flags);
-	if (error) {
-		/* We're screwed now. Shutdown is the only option. */
-		xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
-		goto out_unlock;
-	}
-
-	/*
-	 * If all quotas are completely turned off, close shop.
-	 */
+	/* if all quotas are completely turned off, close shop */
 	if (mp->m_qflags == 0) {
 		mutex_unlock(&q->qi_quotaofflock);
 		xfs_qm_destroy_quotainfo(mp);
 		return 0;
 	}
 
-	/*
-	 * Release our quotainode references if we don't need them anymore.
-	 */
+	/* release our quotainode references if we don't need them anymore */
 	if ((dqtype & XFS_QMOPT_UQUOTA) && q->qi_uquotaip) {
 		xfs_irele(q->qi_uquotaip);
 		q->qi_uquotaip = NULL;
diff --git a/fs/xfs/xfs_trans_dquot.c b/fs/xfs/xfs_trans_dquot.c
index eec640999148..e484747d83ba 100644
--- a/fs/xfs/xfs_trans_dquot.c
+++ b/fs/xfs/xfs_trans_dquot.c
@@ -53,6 +53,8 @@ xfs_trans_log_dquot(
 	struct xfs_dquot	*dqp)
 {
 	ASSERT(XFS_DQ_IS_LOCKED(dqp));
+	/* quotaoff expects no dquots logged after deactivation */
+	ASSERT(xfs_this_quota_on(tp->t_mountp, xfs_dquot_type(dqp)));
 
 	/* Upgrade the dquot to bigtime format if possible. */
 	if (dqp->q_id != 0 &&
-- 
2.26.3