Merge tag 'xfs-5.18-merge-4' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux

Pull xfs fixes from Darrick Wong:
 "This fixes multiple problems in the reserve pool sizing functions: an
  incorrect free space calculation, a pointless infinite loop, and even
  more braindamage that could result in the pool being overfilled. The
  pile of patches from Dave fix myriad races and UAF bugs in the log
  recovery code that much to our mutual surprise nobody's tripped over.
  Dave also fixed a performance optimization that had turned into a
  regression.

  Dave Chinner is taking over as XFS maintainer starting Sunday and
  lasting until 5.19-rc1 is tagged so that I can focus on starting a
  massive design review for the (feature complete after five years)
  online repair feature. From then on, he and I will be moving XFS to a
  co-maintainership model by trading duties every other release.

  NOTE: I hope very strongly that the other pieces of the (X)FS
  ecosystem (fstests and xfsprogs) will make similar changes to spread
  their maintenance load.

  Summary:

   - Fix an incorrect free space calculation in xfs_reserve_blocks that
     could lead to a request for free blocks that will never succeed.

   - Fix a hang in xfs_reserve_blocks caused by an infinite loop and the
     incorrect free space calculation.

   - Fix yet a third problem in xfs_reserve_blocks where multiple racing
     threads can overfill the reserve pool.

   - Fix an accounting error that lead to us reporting reserved space as
     "available".

   - Fix a race condition during abnormal fs shutdown that could cause
     UAF problems when memory reclaim and log shutdown try to clean up
     inodes.

   - Fix a bug where log shutdown can race with unmount to tear down the
     log, thereby causing UAF errors.

   - Disentangle log and filesystem shutdown to reduce confusion.

   - Fix some confusion in xfs_trans_commit such that a race between
     transaction commit and filesystem shutdown can cause unlogged dirty
     inode metadata to be committed, thereby corrupting the filesystem.

   - Remove a performance optimization in the log as it was discovered
     that certain storage hardware handle async log flushes so poorly as
     to cause serious performance regressions. Recent restructuring of
     other parts of the logging code mean that no performance benefit is
     seen on hardware that handle it well"

* tag 'xfs-5.18-merge-4' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
  xfs: drop async cache flushes from CIL commits.
  xfs: shutdown during log recovery needs to mark the log shutdown
  xfs: xfs_trans_commit() path must check for log shutdown
  xfs: xfs_do_force_shutdown needs to block racing shutdowns
  xfs: log shutdown triggers should only shut down the log
  xfs: run callbacks before waking waiters in xlog_state_shutdown_callbacks
  xfs: shutdown in intent recovery has non-intent items in the AIL
  xfs: aborting inodes on shutdown may need buffer lock
  xfs: don't report reserved bnobt space as available
  xfs: fix overfilling of reserve pool
  xfs: always succeed at setting the reserve pool size
  xfs: remove infinite loop when reserving free block pool
  xfs: don't include bnobt blocks when reserving free block pool
  xfs: document the XFS_ALLOC_AGFL_RESERVE constant

18 files changed, 347 insertions, 246 deletions

diff --git a/fs/xfs/libxfs/xfs_alloc.c b/fs/xfs/libxfs/xfs_alloc.c
index 353e53b892e6..b52ed339727f 100644
--- a/fs/xfs/libxfs/xfs_alloc.c
+++ b/fs/xfs/libxfs/xfs_alloc.c
@@ -82,6 +82,24 @@ xfs_prealloc_blocks(
 }
 
 /*
+ * The number of blocks per AG that we withhold from xfs_mod_fdblocks to
+ * guarantee that we can refill the AGFL prior to allocating space in a nearly
+ * full AG.  Although the the space described by the free space btrees, the
+ * blocks used by the freesp btrees themselves, and the blocks owned by the
+ * AGFL are counted in the ondisk fdblocks, it's a mistake to let the ondisk
+ * free space in the AG drop so low that the free space btrees cannot refill an
+ * empty AGFL up to the minimum level.  Rather than grind through empty AGs
+ * until the fs goes down, we subtract this many AG blocks from the incore
+ * fdblocks to ensure user allocation does not overcommit the space the
+ * filesystem needs for the AGFLs.  The rmap btree uses a per-AG reservation to
+ * withhold space from xfs_mod_fdblocks, so we do not account for that here.
+ */
+#define XFS_ALLOCBT_AGFL_RESERVE	4
+
+/*
+ * Compute the number of blocks that we set aside to guarantee the ability to
+ * refill the AGFL and handle a full bmap btree split.
+ *
  * In order to avoid ENOSPC-related deadlock caused by out-of-order locking of
  * AGF buffer (PV 947395), we place constraints on the relationship among
  * actual allocations for data blocks, freelist blocks, and potential file data
@@ -93,14 +111,14 @@ xfs_prealloc_blocks(
  * extents need to be actually allocated. To get around this, we explicitly set
  * aside a few blocks which will not be reserved in delayed allocation.
  *
- * We need to reserve 4 fsbs _per AG_ for the freelist and 4 more to handle a
- * potential split of the file's bmap btree.
+ * For each AG, we need to reserve enough blocks to replenish a totally empty
+ * AGFL and 4 more to handle a potential split of the file's bmap btree.
  */
 unsigned int
 xfs_alloc_set_aside(
 	struct xfs_mount	*mp)
 {
-	return mp->m_sb.sb_agcount * (XFS_ALLOC_AGFL_RESERVE + 4);
+	return mp->m_sb.sb_agcount * (XFS_ALLOCBT_AGFL_RESERVE + 4);
 }
 
 /*
@@ -124,12 +142,12 @@ xfs_alloc_ag_max_usable(
 	unsigned int		blocks;
 
 	blocks = XFS_BB_TO_FSB(mp, XFS_FSS_TO_BB(mp, 4)); /* ag headers */
-	blocks += XFS_ALLOC_AGFL_RESERVE;
+	blocks += XFS_ALLOCBT_AGFL_RESERVE;
 	blocks += 3;			/* AGF, AGI btree root blocks */
 	if (xfs_has_finobt(mp))
 		blocks++;		/* finobt root block */
 	if (xfs_has_rmapbt(mp))
-		blocks++; 		/* rmap root block */
+		blocks++;		/* rmap root block */
 	if (xfs_has_reflink(mp))
 		blocks++;		/* refcount root block */
 
diff --git a/fs/xfs/libxfs/xfs_alloc.h b/fs/xfs/libxfs/xfs_alloc.h
index 1c14a0b1abea..d4c057b764f9 100644
--- a/fs/xfs/libxfs/xfs_alloc.h
+++ b/fs/xfs/libxfs/xfs_alloc.h
@@ -88,7 +88,6 @@ typedef struct xfs_alloc_arg {
 #define XFS_ALLOC_NOBUSY		(1 << 2)/* Busy extents not allowed */
 
 /* freespace limit calculations */
-#define XFS_ALLOC_AGFL_RESERVE	4
 unsigned int xfs_alloc_set_aside(struct xfs_mount *mp);
 unsigned int xfs_alloc_ag_max_usable(struct xfs_mount *mp);
 
diff --git a/fs/xfs/xfs_bio_io.c b/fs/xfs/xfs_bio_io.c
index 32fa02945f73..ae4345b37621 100644
--- a/fs/xfs/xfs_bio_io.c
+++ b/fs/xfs/xfs_bio_io.c
@@ -9,39 +9,6 @@ static inline unsigned int bio_max_vecs(unsigned int count)
 	return bio_max_segs(howmany(count, PAGE_SIZE));
 }
 
-static void
-xfs_flush_bdev_async_endio(
-	struct bio	*bio)
-{
-	complete(bio->bi_private);
-}
-
-/*
- * Submit a request for an async cache flush to run. If the request queue does
- * not require flush operations, just skip it altogether. If the caller needs
- * to wait for the flush completion at a later point in time, they must supply a
- * valid completion. This will be signalled when the flush completes.  The
- * caller never sees the bio that is issued here.
- */
-void
-xfs_flush_bdev_async(
-	struct bio		*bio,
-	struct block_device	*bdev,
-	struct completion	*done)
-{
-	struct request_queue	*q = bdev->bd_disk->queue;
-
-	if (!test_bit(QUEUE_FLAG_WC, &q->queue_flags)) {
-		complete(done);
-		return;
-	}
-
-	bio_init(bio, bdev, NULL, 0, REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC);
-	bio->bi_private = done;
-	bio->bi_end_io = xfs_flush_bdev_async_endio;
-
-	submit_bio(bio);
-}
 int
 xfs_rw_bdev(
 	struct block_device	*bdev,
diff --git a/fs/xfs/xfs_fsops.c b/fs/xfs/xfs_fsops.c
index 33e26690a8c4..68f74549fa22 100644
--- a/fs/xfs/xfs_fsops.c
+++ b/fs/xfs/xfs_fsops.c
@@ -17,6 +17,7 @@
 #include "xfs_fsops.h"
 #include "xfs_trans_space.h"
 #include "xfs_log.h"
+#include "xfs_log_priv.h"
 #include "xfs_ag.h"
 #include "xfs_ag_resv.h"
 #include "xfs_trace.h"
@@ -347,7 +348,7 @@ xfs_fs_counts(
 	cnt->allocino = percpu_counter_read_positive(&mp->m_icount);
 	cnt->freeino = percpu_counter_read_positive(&mp->m_ifree);
 	cnt->freedata = percpu_counter_read_positive(&mp->m_fdblocks) -
-						mp->m_alloc_set_aside;
+						xfs_fdblocks_unavailable(mp);
 
 	spin_lock(&mp->m_sb_lock);
 	cnt->freertx = mp->m_sb.sb_frextents;
@@ -430,46 +431,36 @@ xfs_reserve_blocks(
 	 * If the request is larger than the current reservation, reserve the
 	 * blocks before we update the reserve counters. Sample m_fdblocks and
 	 * perform a partial reservation if the request exceeds free space.
+	 *
+	 * The code below estimates how many blocks it can request from
+	 * fdblocks to stash in the reserve pool.  This is a classic TOCTOU
+	 * race since fdblocks updates are not always coordinated via
+	 * m_sb_lock.  Set the reserve size even if there's not enough free
+	 * space to fill it because mod_fdblocks will refill an undersized
+	 * reserve when it can.
 	 */
-	error = -ENOSPC;
-	do {
-		free = percpu_counter_sum(&mp->m_fdblocks) -
-						mp->m_alloc_set_aside;
-		if (free <= 0)
-			break;
-
-		delta = request - mp->m_resblks;
-		lcounter = free - delta;
-		if (lcounter < 0)
-			/* We can't satisfy the request, just get what we can */
-			fdblks_delta = free;
-		else
-			fdblks_delta = delta;
-
+	free = percpu_counter_sum(&mp->m_fdblocks) -
+						xfs_fdblocks_unavailable(mp);
+	delta = request - mp->m_resblks;
+	mp->m_resblks = request;
+	if (delta > 0 && free > 0) {
 		/*
 		 * We'll either succeed in getting space from the free block
-		 * count or we'll get an ENOSPC. If we get a ENOSPC, it means
-		 * things changed while we were calculating fdblks_delta and so
-		 * we should try again to see if there is anything left to
-		 * reserve.
+		 * count or we'll get an ENOSPC.  Don't set the reserved flag
+		 * here - we don't want to reserve the extra reserve blocks
+		 * from the reserve.
 		 *
-		 * Don't set the reserved flag here - we don't want to reserve
-		 * the extra reserve blocks from the reserve.....
+		 * The desired reserve size can change after we drop the lock.
+		 * Use mod_fdblocks to put the space into the reserve or into
+		 * fdblocks as appropriate.
 		 */
+		fdblks_delta = min(free, delta);
 		spin_unlock(&mp->m_sb_lock);
 		error = xfs_mod_fdblocks(mp, -fdblks_delta, 0);
+		if (!error)
+			xfs_mod_fdblocks(mp, fdblks_delta, 0);
 		spin_lock(&mp->m_sb_lock);
-	} while (error == -ENOSPC);
-
-	/*
-	 * Update the reserve counters if blocks have been successfully
-	 * allocated.
-	 */
-	if (!error && fdblks_delta) {
-		mp->m_resblks += fdblks_delta;
-		mp->m_resblks_avail += fdblks_delta;
 	}
-
 out:
 	if (outval) {
 		outval->resblks = mp->m_resblks;
@@ -528,8 +519,11 @@ xfs_do_force_shutdown(
 	int		tag;
 	const char	*why;
 
-	if (test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &mp->m_opstate))
+
+	if (test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &mp->m_opstate)) {
+		xlog_shutdown_wait(mp->m_log);
 		return;
+	}
 	if (mp->m_sb_bp)
 		mp->m_sb_bp->b_flags |= XBF_DONE;
 
diff --git a/fs/xfs/xfs_icache.c b/fs/xfs/xfs_icache.c
index 20186c584c7d..bffd6eb0b298 100644
--- a/fs/xfs/xfs_icache.c
+++ b/fs/xfs/xfs_icache.c
@@ -883,7 +883,7 @@ xfs_reclaim_inode(
 	 */
 	if (xlog_is_shutdown(ip->i_mount->m_log)) {
 		xfs_iunpin_wait(ip);
-		xfs_iflush_abort(ip);
+		xfs_iflush_shutdown_abort(ip);
 		goto reclaim;
 	}
 	if (xfs_ipincount(ip))
diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
index 26227d26f274..9de6205fe134 100644
--- a/fs/xfs/xfs_inode.c
+++ b/fs/xfs/xfs_inode.c
@@ -3631,7 +3631,7 @@ xfs_iflush_cluster(
 
 	/*
 	 * We must use the safe variant here as on shutdown xfs_iflush_abort()
-	 * can remove itself from the list.
+	 * will remove itself from the list.
 	 */
 	list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
 		iip = (struct xfs_inode_log_item *)lip;
diff --git a/fs/xfs/xfs_inode_item.c b/fs/xfs/xfs_inode_item.c
index 11158fa81a09..9e6ef55cf29e 100644
--- a/fs/xfs/xfs_inode_item.c
+++ b/fs/xfs/xfs_inode_item.c
@@ -544,10 +544,17 @@ xfs_inode_item_push(
 	uint			rval = XFS_ITEM_SUCCESS;
 	int			error;
 
-	ASSERT(iip->ili_item.li_buf);
+	if (!bp || (ip->i_flags & XFS_ISTALE)) {
+		/*
+		 * Inode item/buffer is being being aborted due to cluster
+		 * buffer deletion. Trigger a log force to have that operation
+		 * completed and items removed from the AIL before the next push
+		 * attempt.
+		 */
+		return XFS_ITEM_PINNED;
+	}
 
-	if (xfs_ipincount(ip) > 0 || xfs_buf_ispinned(bp) ||
-	    (ip->i_flags & XFS_ISTALE))
+	if (xfs_ipincount(ip) > 0 || xfs_buf_ispinned(bp))
 		return XFS_ITEM_PINNED;
 
 	if (xfs_iflags_test(ip, XFS_IFLUSHING))
@@ -834,46 +841,143 @@ xfs_buf_inode_io_fail(
 }
 
 /*
- * This is the inode flushing abort routine.  It is called when
- * the filesystem is shutting down to clean up the inode state.  It is
- * responsible for removing the inode item from the AIL if it has not been
- * re-logged and clearing the inode's flush state.
+ * Clear the inode logging fields so no more flushes are attempted.  If we are
+ * on a buffer list, it is now safe to remove it because the buffer is
+ * guaranteed to be locked. The caller will drop the reference to the buffer
+ * the log item held.
+ */
+static void
+xfs_iflush_abort_clean(
+	struct xfs_inode_log_item *iip)
+{
+	iip->ili_last_fields = 0;
+	iip->ili_fields = 0;
+	iip->ili_fsync_fields = 0;
+	iip->ili_flush_lsn = 0;
+	iip->ili_item.li_buf = NULL;
+	list_del_init(&iip->ili_item.li_bio_list);
+}
+
+/*
+ * Abort flushing the inode from a context holding the cluster buffer locked.
+ *
+ * This is the normal runtime method of aborting writeback of an inode that is
+ * attached to a cluster buffer. It occurs when the inode and the backing
+ * cluster buffer have been freed (i.e. inode is XFS_ISTALE), or when cluster
+ * flushing or buffer IO completion encounters a log shutdown situation.
+ *
+ * If we need to abort inode writeback and we don't already hold the buffer
+ * locked, call xfs_iflush_shutdown_abort() instead as this should only ever be
+ * necessary in a shutdown situation.
  */
 void
 xfs_iflush_abort(
 	struct xfs_inode	*ip)
 {
 	struct xfs_inode_log_item *iip = ip->i_itemp;
-	struct xfs_buf		*bp = NULL;
+	struct xfs_buf		*bp;
 
-	if (iip) {
-		/*
-		 * Clear the failed bit before removing the item from the AIL so
-		 * xfs_trans_ail_delete() doesn't try to clear and release the
-		 * buffer attached to the log item before we are done with it.
-		 */
-		clear_bit(XFS_LI_FAILED, &iip->ili_item.li_flags);
-		xfs_trans_ail_delete(&iip->ili_item, 0);
+	if (!iip) {
+		/* clean inode, nothing to do */
+		xfs_iflags_clear(ip, XFS_IFLUSHING);
+		return;
+	}
+
+	/*
+	 * Remove the inode item from the AIL before we clear its internal
+	 * state. Whilst the inode is in the AIL, it should have a valid buffer
+	 * pointer for push operations to access - it is only safe to remove the
+	 * inode from the buffer once it has been removed from the AIL.
+	 *
+	 * We also clear the failed bit before removing the item from the AIL
+	 * as xfs_trans_ail_delete()->xfs_clear_li_failed() will release buffer
+	 * references the inode item owns and needs to hold until we've fully
+	 * aborted the inode log item and detached it from the buffer.
+	 */
+	clear_bit(XFS_LI_FAILED, &iip->ili_item.li_flags);
+	xfs_trans_ail_delete(&iip->ili_item, 0);
+
+	/*
+	 * Grab the inode buffer so can we release the reference the inode log
+	 * item holds on it.
+	 */
+	spin_lock(&iip->ili_lock);
+	bp = iip->ili_item.li_buf;
+	xfs_iflush_abort_clean(iip);
+	spin_unlock(&iip->ili_lock);
 
+	xfs_iflags_clear(ip, XFS_IFLUSHING);
+	if (bp)
+		xfs_buf_rele(bp);
+}
+
+/*
+ * Abort an inode flush in the case of a shutdown filesystem. This can be called
+ * from anywhere with just an inode reference and does not require holding the
+ * inode cluster buffer locked. If the inode is attached to a cluster buffer,
+ * it will grab and lock it safely, then abort the inode flush.
+ */
+void
+xfs_iflush_shutdown_abort(
+	struct xfs_inode	*ip)
+{
+	struct xfs_inode_log_item *iip = ip->i_itemp;
+	struct xfs_buf		*bp;
+
+	if (!iip) {
+		/* clean inode, nothing to do */
+		xfs_iflags_clear(ip, XFS_IFLUSHING);
+		return;
+	}
+
+	spin_lock(&iip->ili_lock);
+	bp = iip->ili_item.li_buf;
+	if (!bp) {
+		spin_unlock(&iip->ili_lock);
+		xfs_iflush_abort(ip);
+		return;
+	}
+
+	/*
+	 * We have to take a reference to the buffer so that it doesn't get
+	 * freed when we drop the ili_lock and then wait to lock the buffer.
+	 * We'll clean up the extra reference after we pick up the ili_lock
+	 * again.
+	 */
+	xfs_buf_hold(bp);
+	spin_unlock(&iip->ili_lock);
+	xfs_buf_lock(bp);
+
+	spin_lock(&iip->ili_lock);
+	if (!iip->ili_item.li_buf) {
 		/*
-		 * Clear the inode logging fields so no more flushes are
-		 * attempted.
+		 * Raced with another removal, hold the only reference
+		 * to bp now. Inode should not be in the AIL now, so just clean
+		 * up and return;
 		 */
-		spin_lock(&iip->ili_lock);
-		iip->ili_last_fields = 0;
-		iip->ili_fields = 0;
-		iip->ili_fsync_fields = 0;
-		iip->ili_flush_lsn = 0;
-		bp = iip->ili_item.li_buf;
-		iip->ili_item.li_buf = NULL;
-		list_del_init(&iip->ili_item.li_bio_list);
+		ASSERT(list_empty(&iip->ili_item.li_bio_list));
+		ASSERT(!test_bit(XFS_LI_IN_AIL, &iip->ili_item.li_flags));
+		xfs_iflush_abort_clean(iip);
 		spin_unlock(&iip->ili_lock);
+		xfs_iflags_clear(ip, XFS_IFLUSHING);
+		xfs_buf_relse(bp);
+		return;
 	}
-	xfs_iflags_clear(ip, XFS_IFLUSHING);
-	if (bp)
-		xfs_buf_rele(bp);
+
+	/*
+	 * Got two references to bp. The first will get dropped by
+	 * xfs_iflush_abort() when the item is removed from the buffer list, but
+	 * we can't drop our reference until _abort() returns because we have to
+	 * unlock the buffer as well. Hence we abort and then unlock and release
+	 * our reference to the buffer.
+	 */
+	ASSERT(iip->ili_item.li_buf == bp);
+	spin_unlock(&iip->ili_lock);
+	xfs_iflush_abort(ip);
+	xfs_buf_relse(bp);
 }
 
+
 /*
  * convert an xfs_inode_log_format struct from the old 32 bit version
  * (which can have different field alignments) to the native 64 bit version
diff --git a/fs/xfs/xfs_inode_item.h b/fs/xfs/xfs_inode_item.h
index 1a302000d604..bbd836a44ff0 100644
--- a/fs/xfs/xfs_inode_item.h
+++ b/fs/xfs/xfs_inode_item.h
@@ -44,6 +44,7 @@ static inline int xfs_inode_clean(struct xfs_inode *ip)
 extern void xfs_inode_item_init(struct xfs_inode *, struct xfs_mount *);
 extern void xfs_inode_item_destroy(struct xfs_inode *);
 extern void xfs_iflush_abort(struct xfs_inode *);
+extern void xfs_iflush_shutdown_abort(struct xfs_inode *);
 extern int xfs_inode_item_format_convert(xfs_log_iovec_t *,
 					 struct xfs_inode_log_format *);
 
diff --git a/fs/xfs/xfs_linux.h b/fs/xfs/xfs_linux.h
index 09a8fba84ff9..cb9105d667db 100644
--- a/fs/xfs/xfs_linux.h
+++ b/fs/xfs/xfs_linux.h
@@ -197,8 +197,6 @@ static inline uint64_t howmany_64(uint64_t x, uint32_t y)
 
 int xfs_rw_bdev(struct block_device *bdev, sector_t sector, unsigned int count,
 		char *data, unsigned int op);
-void xfs_flush_bdev_async(struct bio *bio, struct block_device *bdev,
-		struct completion *done);
 
 #define ASSERT_ALWAYS(expr)	\
 	(likely(expr) ? (void)0 : assfail(NULL, #expr, __FILE__, __LINE__))
diff --git a/fs/xfs/xfs_log.c b/fs/xfs/xfs_log.c
index a8034c0afdf2..499e15b24215 100644
--- a/fs/xfs/xfs_log.c
+++ b/fs/xfs/xfs_log.c
@@ -487,7 +487,10 @@ out_error:
  * Run all the pending iclog callbacks and wake log force waiters and iclog
  * space waiters so they can process the newly set shutdown state. We really
  * don't care what order we process callbacks here because the log is shut down
- * and so state cannot change on disk anymore.
+ * and so state cannot change on disk anymore. However, we cannot wake waiters
+ * until the callbacks have been processed because we may be in unmount and
+ * we must ensure that all AIL operations the callbacks perform have completed
+ * before we tear down the AIL.
  *
  * We avoid processing actively referenced iclogs so that we don't run callbacks
  * while the iclog owner might still be preparing the iclog for IO submssion.
@@ -501,7 +504,6 @@ xlog_state_shutdown_callbacks(
 	struct xlog_in_core	*iclog;
 	LIST_HEAD(cb_list);
 
-	spin_lock(&log->l_icloglock);
 	iclog = log->l_iclog;
 	do {
 		if (atomic_read(&iclog->ic_refcnt)) {
@@ -509,26 +511,22 @@ xlog_state_shutdown_callbacks(
 			continue;
 		}
 		list_splice_init(&iclog->ic_callbacks, &cb_list);
+		spin_unlock(&log->l_icloglock);
+
+		xlog_cil_process_committed(&cb_list);
+
+		spin_lock(&log->l_icloglock);
 		wake_up_all(&iclog->ic_write_wait);
 		wake_up_all(&iclog->ic_force_wait);
 	} while ((iclog = iclog->ic_next) != log->l_iclog);
 
 	wake_up_all(&log->l_flush_wait);
-	spin_unlock(&log->l_icloglock);
-
-	xlog_cil_process_committed(&cb_list);
 }
 
 /*
  * Flush iclog to disk if this is the last reference to the given iclog and the
  * it is in the WANT_SYNC state.
  *
- * If the caller passes in a non-zero @old_tail_lsn and the current log tail
- * does not match, there may be metadata on disk that must be persisted before
- * this iclog is written.  To satisfy that requirement, set the
- * XLOG_ICL_NEED_FLUSH flag as a condition for writing this iclog with the new
- * log tail value.
- *
  * If XLOG_ICL_NEED_FUA is already set on the iclog, we need to ensure that the
  * log tail is updated correctly. NEED_FUA indicates that the iclog will be
  * written to stable storage, and implies that a commit record is contained
@@ -545,12 +543,10 @@ xlog_state_shutdown_callbacks(
  * always capture the tail lsn on the iclog on the first NEED_FUA release
  * regardless of the number of active reference counts on this iclog.
  */
-
 int
 xlog_state_release_iclog(
 	struct xlog		*log,
-	struct xlog_in_core	*iclog,
-	xfs_lsn_t		old_tail_lsn)
+	struct xlog_in_core	*iclog)
 {
 	xfs_lsn_t		tail_lsn;
 	bool			last_ref;
@@ -561,18 +557,14 @@ xlog_state_release_iclog(
 	/*
 	 * Grabbing the current log tail needs to be atomic w.r.t. the writing
 	 * of the tail LSN into the iclog so we guarantee that the log tail does
-	 * not move between deciding if a cache flush is required and writing
-	 * the LSN into the iclog below.
+	 * not move between the first time we know that the iclog needs to be
+	 * made stable and when we eventually submit it.
 	 */
-	if (old_tail_lsn || iclog->ic_state == XLOG_STATE_WANT_SYNC) {
+	if ((iclog->ic_state == XLOG_STATE_WANT_SYNC ||
+	     (iclog->ic_flags & XLOG_ICL_NEED_FUA)) &&
+	    !iclog->ic_header.h_tail_lsn) {
 		tail_lsn = xlog_assign_tail_lsn(log->l_mp);
-
-		if (old_tail_lsn && tail_lsn != old_tail_lsn)
-			iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
-
-		if ((iclog->ic_flags & XLOG_ICL_NEED_FUA) &&
-		    !iclog->ic_header.h_tail_lsn)
-			iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
+		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
 	}
 
 	last_ref = atomic_dec_and_test(&iclog->ic_refcnt);
@@ -583,11 +575,8 @@ xlog_state_release_iclog(
 		 * pending iclog callbacks that were waiting on the release of
 		 * this iclog.
 		 */
-		if (last_ref) {
-			spin_unlock(&log->l_icloglock);
+		if (last_ref)
 			xlog_state_shutdown_callbacks(log);
-			spin_lock(&log->l_icloglock);
-		}
 		return -EIO;
 	}
 
@@ -600,8 +589,6 @@ xlog_state_release_iclog(
 	}
 
 	iclog->ic_state = XLOG_STATE_SYNCING;
-	if (!iclog->ic_header.h_tail_lsn)
-		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
 	xlog_verify_tail_lsn(log, iclog);
 	trace_xlog_iclog_syncing(iclog, _RET_IP_);
 
@@ -873,7 +860,7 @@ xlog_force_iclog(
 	iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
 	if (iclog->ic_state == XLOG_STATE_ACTIVE)
 		xlog_state_switch_iclogs(iclog->ic_log, iclog, 0);
-	return xlog_state_release_iclog(iclog->ic_log, iclog, 0);
+	return xlog_state_release_iclog(iclog->ic_log, iclog);
 }
 
 /*
@@ -1373,7 +1360,7 @@ xlog_ioend_work(
 	 */
 	if (XFS_TEST_ERROR(error, log->l_mp, XFS_ERRTAG_IODONE_IOERR)) {
 		xfs_alert(log->l_mp, "log I/O error %d", error);
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 	}
 
 	xlog_state_done_syncing(iclog);
@@ -1912,7 +1899,7 @@ xlog_write_iclog(
 	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
 
 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 		return;
 	}
 	if (is_vmalloc_addr(iclog->ic_data))
@@ -2411,7 +2398,7 @@ xlog_write_copy_finish(
 		ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
 			xlog_is_shutdown(log));
 release_iclog:
-	error = xlog_state_release_iclog(log, iclog, 0);
+	error = xlog_state_release_iclog(log, iclog);
 	spin_unlock(&log->l_icloglock);
 	return error;
 }
@@ -2487,7 +2474,7 @@ xlog_write(
 		xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
 		     "ctx ticket reservation ran out. Need to up reservation");
 		xlog_print_tic_res(log->l_mp, ticket);
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 	}
 
 	len = xlog_write_calc_vec_length(ticket, log_vector, optype);
@@ -2628,7 +2615,7 @@ next_lv:
 
 	spin_lock(&log->l_icloglock);
 	xlog_state_finish_copy(log, iclog, record_cnt, data_cnt);
-	error = xlog_state_release_iclog(log, iclog, 0);
+	error = xlog_state_release_iclog(log, iclog);
 	spin_unlock(&log->l_icloglock);
 
 	return error;
@@ -3052,7 +3039,7 @@ restart:
 		 * reference to the iclog.
 		 */
 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
-			error = xlog_state_release_iclog(log, iclog, 0);
+			error = xlog_state_release_iclog(log, iclog);
 		spin_unlock(&log->l_icloglock);
 		if (error)
 			return error;
@@ -3821,9 +3808,10 @@ xlog_verify_iclog(
 #endif
 
 /*
- * Perform a forced shutdown on the log. This should be called once and once
- * only by the high level filesystem shutdown code to shut the log subsystem
- * down cleanly.
+ * Perform a forced shutdown on the log.
+ *
+ * This can be called from low level log code to trigger a shutdown, or from the
+ * high level mount shutdown code when the mount shuts down.
  *
  * Our main objectives here are to make sure that:
  *	a. if the shutdown was not due to a log IO error, flush the logs to
@@ -3832,6 +3820,8 @@ xlog_verify_iclog(
  *	   parties to find out. Nothing new gets queued after this is done.
  *	c. Tasks sleeping on log reservations, pinned objects and
  *	   other resources get woken up.
+ *	d. The mount is also marked as shut down so that log triggered shutdowns
+ *	   still behave the same as if they called xfs_forced_shutdown().
  *
  * Return true if the shutdown cause was a log IO error and we actually shut the
  * log down.
@@ -3843,25 +3833,25 @@ xlog_force_shutdown(
 {
 	bool		log_error = (shutdown_flags & SHUTDOWN_LOG_IO_ERROR);
 
-	/*
-	 * If this happens during log recovery then we aren't using the runtime
-	 * log mechanisms yet so there's nothing to shut down.
-	 */
-	if (!log || xlog_in_recovery(log))
+	if (!log)
 		return false;
 
-	ASSERT(!xlog_is_shutdown(log));
-
 	/*
 	 * Flush all the completed transactions to disk before marking the log
 	 * being shut down. We need to do this first as shutting down the log
 	 * before the force will prevent the log force from flushing the iclogs
 	 * to disk.
 	 *
-	 * Re-entry due to a log IO error shutdown during the log force is
-	 * prevented by the atomicity of higher level shutdown code.
+	 * When we are in recovery, there are no transactions to flush, and
+	 * we don't want to touch the log because we don't want to perturb the
+	 * current head/tail for future recovery attempts. Hence we need to
+	 * avoid a log force in this case.
+	 *
+	 * If we are shutting down due to a log IO error, then we must avoid
+	 * trying to write the log as that may just result in more IO errors and
+	 * an endless shutdown/force loop.
 	 */
-	if (!log_error)
+	if (!log_error && !xlog_in_recovery(log))
 		xfs_log_force(log->l_mp, XFS_LOG_SYNC);
 
 	/*
@@ -3878,12 +3868,25 @@ xlog_force_shutdown(
 	spin_lock(&log->l_icloglock);
 	if (test_and_set_bit(XLOG_IO_ERROR, &log->l_opstate)) {
 		spin_unlock(&log->l_icloglock);
-		ASSERT(0);
 		return false;
 	}
 	spin_unlock(&log->l_icloglock);
 
 	/*
+	 * If this log shutdown also sets the mount shutdown state, issue a
+	 * shutdown warning message.
+	 */
+	if (!test_and_set_bit(XFS_OPSTATE_SHUTDOWN, &log->l_mp->m_opstate)) {
+		xfs_alert_tag(log->l_mp, XFS_PTAG_SHUTDOWN_LOGERROR,
+"Filesystem has been shut down due to log error (0x%x).",
+				shutdown_flags);
+		xfs_alert(log->l_mp,
+"Please unmount the filesystem and rectify the problem(s).");
+		if (xfs_error_level >= XFS_ERRLEVEL_HIGH)
+			xfs_stack_trace();
+	}
+
+	/*
 	 * We don't want anybody waiting for log reservations after this. That
 	 * means we have to wake up everybody queued up on reserveq as well as
 	 * writeq.  In addition, we make sure in xlog_{re}grant_log_space that
@@ -3903,8 +3906,12 @@ xlog_force_shutdown(
 	wake_up_all(&log->l_cilp->xc_start_wait);
 	wake_up_all(&log->l_cilp->xc_commit_wait);
 	spin_unlock(&log->l_cilp->xc_push_lock);
+
+	spin_lock(&log->l_icloglock);
 	xlog_state_shutdown_callbacks(log);
+	spin_unlock(&log->l_icloglock);
 
+	wake_up_var(&log->l_opstate);
 	return log_error;
 }
 
diff --git a/fs/xfs/xfs_log_cil.c b/fs/xfs/xfs_log_cil.c
index 796e4464f809..ba57323bfdce 100644
--- a/fs/xfs/xfs_log_cil.c
+++ b/fs/xfs/xfs_log_cil.c
@@ -540,7 +540,7 @@ xlog_cil_insert_items(
 	spin_unlock(&cil->xc_cil_lock);
 
 	if (tp->t_ticket->t_curr_res < 0)
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 }
 
 static void
@@ -705,11 +705,21 @@ xlog_cil_set_ctx_write_state(
 		 * The LSN we need to pass to the log items on transaction
 		 * commit is the LSN reported by the first log vector write, not
 		 * the commit lsn. If we use the commit record lsn then we can
-		 * move the tail beyond the grant write head.
+		 * move the grant write head beyond the tail LSN and overwrite
+		 * it.
 		 */
 		ctx->start_lsn = lsn;
 		wake_up_all(&cil->xc_start_wait);
 		spin_unlock(&cil->xc_push_lock);
+
+		/*
+		 * Make sure the metadata we are about to overwrite in the log
+		 * has been flushed to stable storage before this iclog is
+		 * issued.
+		 */
+		spin_lock(&cil->xc_log->l_icloglock);
+		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH;
+		spin_unlock(&cil->xc_log->l_icloglock);
 		return;
 	}
 
@@ -854,7 +864,7 @@ xlog_cil_write_commit_record(
 
 	error = xlog_write(log, ctx, &vec, ctx->ticket, XLOG_COMMIT_TRANS);
 	if (error)
-		xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
+		xlog_force_shutdown(log, SHUTDOWN_LOG_IO_ERROR);
 	return error;
 }
 
@@ -888,10 +898,7 @@ xlog_cil_push_work(
 	struct xfs_trans_header thdr;
 	struct xfs_log_iovec	lhdr;
 	struct xfs_log_vec	lvhdr = { NULL };
-	xfs_lsn_t		preflush_tail_lsn;
 	xfs_csn_t		push_seq;
-	struct bio		bio;
-	DECLARE_COMPLETION_ONSTACK(bdev_flush);
 	bool			push_commit_stable;
 
 	new_ctx = xlog_cil_ctx_alloc();
@@ -962,23 +969,6 @@ xlog_cil_push_work(
 	spin_unlock(&cil->xc_push_lock);
 
 	/*
-	 * The CIL is s