Merge tag 'xfs-5.14-fixes-2' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux

Pull xfs fixes from Darrick Wong:
 "This contains a bunch of bug fixes in XFS.

  Dave and I have been busy the last couple of weeks to find and fix as
  many log recovery bugs as we can find; here are the results so far. Go
  fstests -g recoveryloop! ;)

   - Fix a number of coordination bugs relating to cache flushes for
     metadata writeback, cache flushes for multi-buffer log writes, and
     FUA writes for single-buffer log writes

   - Fix a bug with incorrect replay of attr3 blocks

   - Fix unnecessary stalls when flushing logs to disk

   - Fix spoofing problems when recovering realtime bitmap blocks"

* tag 'xfs-5.14-fixes-2' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
  xfs: prevent spoofing of rtbitmap blocks when recovering buffers
  xfs: limit iclog tail updates
  xfs: need to see iclog flags in tracing
  xfs: Enforce attr3 buffer recovery order
  xfs: logging the on disk inode LSN can make it go backwards
  xfs: avoid unnecessary waits in xfs_log_force_lsn()
  xfs: log forces imply data device cache flushes
  xfs: factor out forced iclog flushes
  xfs: fix ordering violation between cache flushes and tail updates
  xfs: fold __xlog_state_release_iclog into xlog_state_release_iclog
  xfs: external logs need to flush data device
  xfs: flush data dev on external log write

7 files changed, 244 insertions, 106 deletions

diff --git a/fs/xfs/libxfs/xfs_log_format.h b/fs/xfs/libxfs/xfs_log_format.h
index d548ea4b6aab..2c5bcbc19264 100644
--- a/fs/xfs/libxfs/xfs_log_format.h
+++ b/fs/xfs/libxfs/xfs_log_format.h
@@ -411,7 +411,16 @@ struct xfs_log_dinode {
 	/* start of the extended dinode, writable fields */
 	uint32_t	di_crc;		/* CRC of the inode */
 	uint64_t	di_changecount;	/* number of attribute changes */
-	xfs_lsn_t	di_lsn;		/* flush sequence */
+
+	/*
+	 * The LSN we write to this field during formatting is not a reflection
+	 * of the current on-disk LSN. It should never be used for recovery
+	 * sequencing, nor should it be recovered into the on-disk inode at all.
+	 * See xlog_recover_inode_commit_pass2() and xfs_log_dinode_to_disk()
+	 * for details.
+	 */
+	xfs_lsn_t	di_lsn;
+
 	uint64_t	di_flags2;	/* more random flags */
 	uint32_t	di_cowextsize;	/* basic cow extent size for file */
 	uint8_t		di_pad2[12];	/* more padding for future expansion */
diff --git a/fs/xfs/xfs_buf_item_recover.c b/fs/xfs/xfs_buf_item_recover.c
index d44e8b4a3391..4775485b4062 100644
--- a/fs/xfs/xfs_buf_item_recover.c
+++ b/fs/xfs/xfs_buf_item_recover.c
@@ -698,7 +698,8 @@ xlog_recover_do_inode_buffer(
 static xfs_lsn_t
 xlog_recover_get_buf_lsn(
 	struct xfs_mount	*mp,
-	struct xfs_buf		*bp)
+	struct xfs_buf		*bp,
+	struct xfs_buf_log_format *buf_f)
 {
 	uint32_t		magic32;
 	uint16_t		magic16;
@@ -706,11 +707,20 @@ xlog_recover_get_buf_lsn(
 	void			*blk = bp->b_addr;
 	uuid_t			*uuid;
 	xfs_lsn_t		lsn = -1;
+	uint16_t		blft;
 
 	/* v4 filesystems always recover immediately */
 	if (!xfs_sb_version_hascrc(&mp->m_sb))
 		goto recover_immediately;
 
+	/*
+	 * realtime bitmap and summary file blocks do not have magic numbers or
+	 * UUIDs, so we must recover them immediately.
+	 */
+	blft = xfs_blft_from_flags(buf_f);
+	if (blft == XFS_BLFT_RTBITMAP_BUF || blft == XFS_BLFT_RTSUMMARY_BUF)
+		goto recover_immediately;
+
 	magic32 = be32_to_cpu(*(__be32 *)blk);
 	switch (magic32) {
 	case XFS_ABTB_CRC_MAGIC:
@@ -796,6 +806,7 @@ xlog_recover_get_buf_lsn(
 	switch (magicda) {
 	case XFS_DIR3_LEAF1_MAGIC:
 	case XFS_DIR3_LEAFN_MAGIC:
+	case XFS_ATTR3_LEAF_MAGIC:
 	case XFS_DA3_NODE_MAGIC:
 		lsn = be64_to_cpu(((struct xfs_da3_blkinfo *)blk)->lsn);
 		uuid = &((struct xfs_da3_blkinfo *)blk)->uuid;
@@ -919,7 +930,7 @@ xlog_recover_buf_commit_pass2(
 	 * the verifier will be reset to match whatever recover turns that
 	 * buffer into.
 	 */
-	lsn = xlog_recover_get_buf_lsn(mp, bp);
+	lsn = xlog_recover_get_buf_lsn(mp, bp, buf_f);
 	if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
 		trace_xfs_log_recover_buf_skip(log, buf_f);
 		xlog_recover_validate_buf_type(mp, bp, buf_f, NULLCOMMITLSN);
diff --git a/fs/xfs/xfs_inode_item_recover.c b/fs/xfs/xfs_inode_item_recover.c
index 7b79518b6c20..e0072a6cd2d3 100644
--- a/fs/xfs/xfs_inode_item_recover.c
+++ b/fs/xfs/xfs_inode_item_recover.c
@@ -145,7 +145,8 @@ xfs_log_dinode_to_disk_ts(
 STATIC void
 xfs_log_dinode_to_disk(
 	struct xfs_log_dinode	*from,
-	struct xfs_dinode	*to)
+	struct xfs_dinode	*to,
+	xfs_lsn_t		lsn)
 {
 	to->di_magic = cpu_to_be16(from->di_magic);
 	to->di_mode = cpu_to_be16(from->di_mode);
@@ -182,7 +183,7 @@ xfs_log_dinode_to_disk(
 		to->di_flags2 = cpu_to_be64(from->di_flags2);
 		to->di_cowextsize = cpu_to_be32(from->di_cowextsize);
 		to->di_ino = cpu_to_be64(from->di_ino);
-		to->di_lsn = cpu_to_be64(from->di_lsn);
+		to->di_lsn = cpu_to_be64(lsn);
 		memcpy(to->di_pad2, from->di_pad2, sizeof(to->di_pad2));
 		uuid_copy(&to->di_uuid, &from->di_uuid);
 		to->di_flushiter = 0;
@@ -261,16 +262,25 @@ xlog_recover_inode_commit_pass2(
 	}
 
 	/*
-	 * If the inode has an LSN in it, recover the inode only if it's less
-	 * than the lsn of the transaction we are replaying. Note: we still
-	 * need to replay an owner change even though the inode is more recent
-	 * than the transaction as there is no guarantee that all the btree
-	 * blocks are more recent than this transaction, too.
+	 * If the inode has an LSN in it, recover the inode only if the on-disk
+	 * inode's LSN is older than the lsn of the transaction we are
+	 * replaying. We can have multiple checkpoints with the same start LSN,
+	 * so the current LSN being equal to the on-disk LSN doesn't necessarily
+	 * mean that the on-disk inode is more recent than the change being
+	 * replayed.
+	 *
+	 * We must check the current_lsn against the on-disk inode
+	 * here because the we can't trust the log dinode to contain a valid LSN
+	 * (see comment below before replaying the log dinode for details).
+	 *
+	 * Note: we still need to replay an owner change even though the inode
+	 * is more recent than the transaction as there is no guarantee that all
+	 * the btree blocks are more recent than this transaction, too.
 	 */
 	if (dip->di_version >= 3) {
 		xfs_lsn_t	lsn = be64_to_cpu(dip->di_lsn);
 
-		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) >= 0) {
+		if (lsn && lsn != -1 && XFS_LSN_CMP(lsn, current_lsn) > 0) {
 			trace_xfs_log_recover_inode_skip(log, in_f);
 			error = 0;
 			goto out_owner_change;
@@ -368,8 +378,17 @@ xlog_recover_inode_commit_pass2(
 		goto out_release;
 	}
 
-	/* recover the log dinode inode into the on disk inode */
-	xfs_log_dinode_to_disk(ldip, dip);
+	/*
+	 * Recover the log dinode inode into the on disk inode.
+	 *
+	 * The LSN in the log dinode is garbage - it can be zero or reflect
+	 * stale in-memory runtime state that isn't coherent with the changes
+	 * logged in this transaction or the changes written to the on-disk
+	 * inode.  Hence we write the current lSN into the inode because that
+	 * matches what xfs_iflush() would write inode the inode when flushing
+	 * the changes in this transaction.
+	 */
+	xfs_log_dinode_to_disk(ldip, dip, current_lsn);
 
 	fields = in_f->ilf_fields;
 	if (fields & XFS_ILOG_DEV)
diff --git a/fs/xfs/xfs_log.c b/fs/xfs/xfs_log.c
index 36fa2650b081..60ac5fd63f1e 100644
--- a/fs/xfs/xfs_log.c
+++ b/fs/xfs/xfs_log.c
@@ -78,13 +78,12 @@ xlog_verify_iclog(
 STATIC void
 xlog_verify_tail_lsn(
 	struct xlog		*log,
-	struct xlog_in_core	*iclog,
-	xfs_lsn_t		tail_lsn);
+	struct xlog_in_core	*iclog);
 #else
 #define xlog_verify_dest_ptr(a,b)
 #define xlog_verify_grant_tail(a)
 #define xlog_verify_iclog(a,b,c)
-#define xlog_verify_tail_lsn(a,b,c)
+#define xlog_verify_tail_lsn(a,b)
 #endif
 
 STATIC int
@@ -487,51 +486,80 @@ out_error:
 	return error;
 }
 
-static bool
-__xlog_state_release_iclog(
-	struct xlog		*log,
-	struct xlog_in_core	*iclog)
-{
-	lockdep_assert_held(&log->l_icloglock);
-
-	if (iclog->ic_state == XLOG_STATE_WANT_SYNC) {
-		/* update tail before writing to iclog */
-		xfs_lsn_t tail_lsn = xlog_assign_tail_lsn(log->l_mp);
-
-		iclog->ic_state = XLOG_STATE_SYNCING;
-		iclog->ic_header.h_tail_lsn = cpu_to_be64(tail_lsn);
-		xlog_verify_tail_lsn(log, iclog, tail_lsn);
-		/* cycle incremented when incrementing curr_block */
-		trace_xlog_iclog_syncing(iclog, _RET_IP_);
-		return true;
-	}
-
-	ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
-	return false;
-}
-
 /*
  * 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
+ * within the iclog. We need to ensure that the log tail does not move beyond
+ * the tail that the first commit record in the iclog ordered against, otherwise
+ * correct recovery of that checkpoint becomes dependent on future operations
+ * performed on this iclog.
+ *
+ * Hence if NEED_FUA is set and the current iclog tail lsn is empty, write the
+ * current tail into iclog. Once the iclog tail is set, future operations must
+ * not modify it, otherwise they potentially violate ordering constraints for
+ * the checkpoint commit that wrote the initial tail lsn value. The tail lsn in
+ * the iclog will get zeroed on activation of the iclog after sync, so we
+ * 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)
+	struct xlog_in_core	*iclog,
+	xfs_lsn_t		old_tail_lsn)
 {
+	xfs_lsn_t		tail_lsn;
 	lockdep_assert_held(&log->l_icloglock);
 
 	trace_xlog_iclog_release(iclog, _RET_IP_);
 	if (iclog->ic_state == XLOG_STATE_IOERROR)
 		return -EIO;
 
-	if (atomic_dec_and_test(&iclog->ic_refcnt) &&
-	    __xlog_state_release_iclog(log, iclog)) {
-		spin_unlock(&log->l_icloglock);
-		xlog_sync(log, iclog);
-		spin_lock(&log->l_icloglock);
+	/*
+	 * 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.
+	 */
+	if (old_tail_lsn || iclog->ic_state == XLOG_STATE_WANT_SYNC) {
+		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);
 	}
 
+	if (!atomic_dec_and_test(&iclog->ic_refcnt))
+		return 0;
+
+	if (iclog->ic_state != XLOG_STATE_WANT_SYNC) {
+		ASSERT(iclog->ic_state == XLOG_STATE_ACTIVE);
+		return 0;
+	}
+
+	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_);
+
+	spin_unlock(&log->l_icloglock);
+	xlog_sync(log, iclog);
+	spin_lock(&log->l_icloglock);
 	return 0;
 }
 
@@ -774,6 +802,21 @@ xfs_log_mount_cancel(
 }
 
 /*
+ * Flush out the iclog to disk ensuring that device caches are flushed and
+ * the iclog hits stable storage before any completion waiters are woken.
+ */
+static inline int
+xlog_force_iclog(
+	struct xlog_in_core	*iclog)
+{
+	atomic_inc(&iclog->ic_refcnt);
+	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);
+}
+
+/*
  * Wait for the iclog and all prior iclogs to be written disk as required by the
  * log force state machine. Waiting on ic_force_wait ensures iclog completions
  * have been ordered and callbacks run before we are woken here, hence
@@ -827,13 +870,6 @@ xlog_write_unmount_record(
 	/* account for space used by record data */
 	ticket->t_curr_res -= sizeof(ulf);
 
-	/*
-	 * For external log devices, we need to flush the data device cache
-	 * first to ensure all metadata writeback is on stable storage before we
-	 * stamp the tail LSN into the unmount record.
-	 */
-	if (log->l_targ != log->l_mp->m_ddev_targp)
-		blkdev_issue_flush(log->l_targ->bt_bdev);
 	return xlog_write(log, &vec, ticket, NULL, NULL, XLOG_UNMOUNT_TRANS);
 }
 
@@ -865,18 +901,7 @@ out_err:
 
 	spin_lock(&log->l_icloglock);
 	iclog = log->l_iclog;
-	atomic_inc(&iclog->ic_refcnt);
-	if (iclog->ic_state == XLOG_STATE_ACTIVE)
-		xlog_state_switch_iclogs(log, iclog, 0);
-	else
-		ASSERT(iclog->ic_state == XLOG_STATE_WANT_SYNC ||
-		       iclog->ic_state == XLOG_STATE_IOERROR);
-	/*
-	 * Ensure the journal is fully flushed and on stable storage once the
-	 * iclog containing the unmount record is written.
-	 */
-	iclog->ic_flags |= (XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
-	error = xlog_state_release_iclog(log, iclog);
+	error = xlog_force_iclog(iclog);
 	xlog_wait_on_iclog(iclog);
 
 	if (tic) {
@@ -1796,10 +1821,20 @@ xlog_write_iclog(
 	 * metadata writeback and causing priority inversions.
 	 */
 	iclog->ic_bio.bi_opf = REQ_OP_WRITE | REQ_META | REQ_SYNC | REQ_IDLE;
-	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH)
+	if (iclog->ic_flags & XLOG_ICL_NEED_FLUSH) {
 		iclog->ic_bio.bi_opf |= REQ_PREFLUSH;
+		/*
+		 * For external log devices, we also need to flush the data
+		 * device cache first to ensure all metadata writeback covered
+		 * by the LSN in this iclog is on stable storage. This is slow,
+		 * but it *must* complete before we issue the external log IO.
+		 */
+		if (log->l_targ != log->l_mp->m_ddev_targp)
+			blkdev_issue_flush(log->l_mp->m_ddev_targp->bt_bdev);
+	}
 	if (iclog->ic_flags & XLOG_ICL_NEED_FUA)
 		iclog->ic_bio.bi_opf |= REQ_FUA;
+
 	iclog->ic_flags &= ~(XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA);
 
 	if (xlog_map_iclog_data(&iclog->ic_bio, iclog->ic_data, count)) {
@@ -2310,7 +2345,7 @@ xlog_write_copy_finish(
 	return 0;
 
 release_iclog:
-	error = xlog_state_release_iclog(log, iclog);
+	error = xlog_state_release_iclog(log, iclog, 0);
 	spin_unlock(&log->l_icloglock);
 	return error;
 }
@@ -2529,7 +2564,7 @@ next_lv:
 		ASSERT(optype & XLOG_COMMIT_TRANS);
 		*commit_iclog = iclog;
 	} else {
-		error = xlog_state_release_iclog(log, iclog);
+		error = xlog_state_release_iclog(log, iclog, 0);
 	}
 	spin_unlock(&log->l_icloglock);
 
@@ -2567,6 +2602,7 @@ xlog_state_activate_iclog(
 	memset(iclog->ic_header.h_cycle_data, 0,
 		sizeof(iclog->ic_header.h_cycle_data));
 	iclog->ic_header.h_lsn = 0;
+	iclog->ic_header.h_tail_lsn = 0;
 }
 
 /*
@@ -2967,7 +3003,7 @@ restart:
 		 * reference to the iclog.
 		 */
 		if (!atomic_add_unless(&iclog->ic_refcnt, -1, 1))
-			error = xlog_state_release_iclog(log, iclog);
+			error = xlog_state_release_iclog(log, iclog, 0);
 		spin_unlock(&log->l_icloglock);
 		if (error)
 			return error;
@@ -3132,6 +3168,35 @@ xlog_state_switch_iclogs(
 }
 
 /*
+ * Force the iclog to disk and check if the iclog has been completed before
+ * xlog_force_iclog() returns. This can happen on synchronous (e.g.
+ * pmem) or fast async storage because we drop the icloglock to issue the IO.
+ * If completion has already occurred, tell the caller so that it can avoid an
+ * unnecessary wait on the iclog.
+ */
+static int
+xlog_force_and_check_iclog(
+	struct xlog_in_core	*iclog,
+	bool			*completed)
+{
+	xfs_lsn_t		lsn = be64_to_cpu(iclog->ic_header.h_lsn);
+	int			error;
+
+	*completed = false;
+	error = xlog_force_iclog(iclog);
+	if (error)
+		return error;
+
+	/*
+	 * If the iclog has already been completed and reused the header LSN
+	 * will have been rewritten by completion
+	 */
+	if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
+		*completed = true;
+	return 0;
+}
+
+/*
  * Write out all data in the in-core log as of this exact moment in time.
  *
  * Data may be written to the in-core log during this call.  However,
@@ -3165,7 +3230,6 @@ xfs_log_force(
 {
 	struct xlog		*log = mp->m_log;
 	struct xlog_in_core	*iclog;
-	xfs_lsn_t		lsn;
 
 	XFS_STATS_INC(mp, xs_log_force);
 	trace_xfs_log_force(mp, 0, _RET_IP_);
@@ -3193,39 +3257,33 @@ xfs_log_force(
 		iclog = iclog->ic_prev;
 	} else if (iclog->ic_state == XLOG_STATE_ACTIVE) {
 		if (atomic_read(&iclog->ic_refcnt) == 0) {
-			/*
-			 * We are the only one with access to this iclog.
-			 *
-			 * Flush it out now.  There should be a roundoff of zero
-			 * to show that someone has already taken care of the
-			 * roundoff from the previous sync.
-			 */
-			atomic_inc(&iclog->ic_refcnt);
-			lsn = be64_to_cpu(iclog->ic_header.h_lsn);
-			xlog_state_switch_iclogs(log, iclog, 0);
-			if (xlog_state_release_iclog(log, iclog))
+			/* We have exclusive access to this iclog. */
+			bool	completed;
+
+			if (xlog_force_and_check_iclog(iclog, &completed))
 				goto out_error;
 
-			if (be64_to_cpu(iclog->ic_header.h_lsn) != lsn)
+			if (completed)
 				goto out_unlock;
 		} else {
 			/*
-			 * Someone else is writing to this iclog.
-			 *
-			 * Use its call to flush out the data.  However, the
-			 * other thread may not force out this LR, so we mark
-			 * it WANT_SYNC.
+			 * Someone else is still writing to this iclog, so we
+			 * need to ensure that when they release the iclog it
+			 * gets synced immediately as we may be waiting on it.
 			 */
 			xlog_state_switch_iclogs(log, iclog, 0);
 		}
-	} else {
-		/*
-		 * If the head iclog is not active nor dirty, we just attach
-		 * ourselves to the head and go to sleep if necessary.
-		 */
-		;
 	}
 
+	/*
+	 * The iclog we are about to wait on may contain the checkpoint pushed
+	 * by the above xlog_cil_force() call, but it may not have been pushed
+	 * to disk yet. Like the ACTIVE case above, we need to make sure caches
+	 * are flushed when this iclog is written.
+	 */
+	if (iclog->ic_state == XLOG_STATE_WANT_SYNC)
+		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
+
 	if (flags & XFS_LOG_SYNC)
 		return xlog_wait_on_iclog(iclog);
 out_unlock:
@@ -3245,6 +3303,7 @@ xlog_force_lsn(
 	bool			already_slept)
 {
 	struct xlog_in_core	*iclog;
+	bool			completed;
 
 	spin_lock(&log->l_icloglock);
 	iclog = log->l_iclog;
@@ -3258,7 +3317,8 @@ xlog_force_lsn(
 			goto out_unlock;
 	}
 
-	if (iclog->ic_state == XLOG_STATE_ACTIVE) {
+	switch (iclog->ic_state) {
+	case XLOG_STATE_ACTIVE:
 		/*
 		 * We sleep here if we haven't already slept (e.g. this is the
 		 * first time we've looked at the correct iclog buf) and the
@@ -3281,12 +3341,31 @@ xlog_force_lsn(
 					&log->l_icloglock);
 			return -EAGAIN;
 		}
-		atomic_inc(&iclog->ic_refcnt);
-		xlog_state_switch_iclogs(log, iclog, 0);
-		if (xlog_state_release_iclog(log, iclog))
+		if (xlog_force_and_check_iclog(iclog, &completed))
 			goto out_error;
 		if (log_flushed)
 			*log_flushed = 1;
+		if (completed)
+			goto out_unlock;
+		break;
+	case XLOG_STATE_WANT_SYNC:
+		/*
+		 * This iclog may contain the checkpoint pushed by the
+		 * xlog_cil_force_seq() call, but there are other writers still
+		 * accessing it so it hasn't been pushed to disk yet. Like the
+		 * ACTIVE case above, we need to make sure caches are flushed
+		 * when this iclog is written.
+		 */
+		iclog->ic_flags |= XLOG_ICL_NEED_FLUSH | XLOG_ICL_NEED_FUA;
+		break;
+	default:
+		/*
+		 * The entire checkpoint was written by the CIL force and is on
+		 * its way to disk already. It will be stable when it
+		 * completes, so we don't need to manipulate caches here at all.
+		 * We just need to wait for completion if necessary.
+		 */
+		break;
 	}
 
 	if (flags & XFS_LOG_SYNC)
@@ -3559,10 +3638,10 @@ xlog_verify_grant_tail(
 STATIC void
 xlog_verify_tail_lsn(
 	struct xlog		*log,
-	struct xlog_in_core	*iclog,
-	xfs_lsn_t		tail_lsn)
+	struct xlog_in_core	*iclog)
 {
-    int blocks;
+	xfs_lsn_t	tail_lsn = be64_to_cpu(iclog->ic_header.h_tail_lsn);
+	int		blocks;
 
     if (CYCLE_LSN(tail_lsn) == log->l_prev_cycle) {
 	blocks =
diff --git a/fs/xfs/xfs_log_cil.c b/fs/xfs/xfs_log_cil.c
index b128aaa9b870..4c44bc3786c0 100644
--- a/fs/xfs/xfs_log_cil.c
+++ b/fs/xfs/xfs_log_cil.c
@@ -654,8 +654,9 @@ 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_lsn_t		commit_lsn;
-	xfs_lsn_t		push_seq;
+	xfs_csn_t		push_seq;
 	struct bio		bio;
 	DECLARE_COMPLETION_ONSTACK(bdev_flush);
 
@@ -730,7 +731,15 @@ xlog_cil_push_work(
 	 * because we hold the flush lock exclusively. Hence we can now issue
 	 * a cache flush to ensure all the completed metadata in the journal we
 	 * are about to overwrite is on stable storage.
+	 *
+	 * Because we are issuing this cache flush before we've written the
+	 * tail lsn to the iclog, we can have metadata IO completions move the
+	 * tail forwards between the completion of this flush and the iclog
+	 * being written. In this case, we need to re-issue the cache flush
+	 * before the iclog write. To detect whether the log tail moves, sample
+	 * the tail LSN *before* we issue the flush.
 	 */
+	preflush_tail_lsn = atomic64_read(&log->l_tail_lsn);
 	xfs_flush_bdev_async(&bio, log->l_mp->m_ddev_targp->bt_bdev,
 				&bdev_flush);
 
@@ -941,7 +950,7 @@ restart:
 	 * storage.
 	 */
 	commit_iclog->ic_flags |= XLOG_ICL_NEED_FUA;
-	xlog_state_release_iclog(log, commit_iclog);
+	xlog_state_release_iclog(log, commit_iclog, preflush_tail_lsn);
 	spin_unlock(&log->l_icloglock);
 	return;
 
diff --git a/fs/xfs/xfs_log_priv.h b/fs/xfs/xfs_log_priv.h
index 4c41bbfa33b0..f3e79a45d60a 100644
--- a/fs/xfs/xfs_log_priv.h
+++ b/fs/xfs/xfs_log_priv.h
@@ -59,6 +59,16 @@ enum xlog_iclog_state {
 	{ XLOG_STATE_DIRTY,	"XLOG_STATE_DIRTY" }, \
 	{ XLOG_STATE_IOERROR,	"XLOG_STATE_IOERROR" }
 
+/*
+ * In core log flags
+ */
+#define XLOG_ICL_NEED_FLUSH	(1 << 0)	/* iclog needs REQ_PREFLUSH */
+#define XLOG_ICL_NEED_FUA	(1 << 1)	/* iclog needs REQ_FUA */
+
+#define XLOG_ICL_STRINGS \
+	{ XLOG_ICL_NEED_FLUSH,	"XLOG_ICL_NEED_FLUSH" }, \
+	{ XLOG_ICL_NEED_FUA,	"XLOG_ICL_NEED_FUA" }
+
 
 /*
  * Log ticket flags
@@ -143,9 +153,6 @@ enum xlog_iclog_state {
 
 #define XLOG_COVER_OPS		5
 
-#define XLOG_ICL_NEED_FLUSH	(1 << 0)	/* iclog needs REQ_PREFLUSH */
-#define XLOG_ICL_NEED_FUA	(1 << 1)	/* iclog needs REQ_FUA */
-
 /* Ticket reservation region accounting */ 
 #define XLOG_TIC_LEN_MAX	15
 
@@ -497,7 +504,8 @@ int	xlog_commit_record(struct xlog *log, struct xlog_ticket *ticket,
 void	xfs_log_ticket_ungrant(struct xlog *log, struct xlog_ticket *ticket);
 void	xfs_log_ticket_regrant(struct xlog *log, struct xlog_ticket *ticket);
 
-int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog);
+int xlog_state_release_iclog(struct xlog *log, struct xlog_in_core *iclog,
+		xfs_lsn_t log_tail_lsn);
 
 /*
  * When we crack an atomic LSN, we sample it first so that the value will not
diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
index f9d8d605f9b1..19260291ff8b 100644
--- a/fs/xfs/xfs_trace.h
+++ b/fs/xfs/xfs_trace.h
@@ -3944,6 +3944,7 @@ DECLARE_EVENT_CLASS(xlog_iclog_class,
 		__field(uint32_t, state)
 		__field(int32_t, refcount)
 		__field(uint32_t, offset)
+		__field(uint32_t, flags)
 		__field(unsigned long long, lsn)
 		__field(unsigned long, caller_ip)
 	),
@@ -3952,15 +3953,17 @@ DECLARE_EVENT_CLASS(xlog_iclog_class,
 		__entry->state = iclog->ic_state;
 		__entry->refcount = atomic_read(&iclog->ic_refcnt);
 		__entry->offset = iclog->ic_offset;
+		__entry->flags = iclog->ic_flags;
 		__entry->lsn = be64_to_cpu(iclog->ic_header.h_lsn);
 		__entry->caller_ip = caller_ip;
 	),
-	TP_printk("dev %d:%d state %s refcnt %d offset %u lsn 0x%llx caller %pS",
+	TP_printk("dev %d:%d state %s refcnt %d offset %u lsn 0x%llx flags %s caller %pS",
 		  MAJOR(__entry->dev), MINOR(__entry->dev),
 		  __print_symbolic(__entry->state, XLOG_STATE_STRINGS),
 		  __entry->refcount,
 		  __entry->offset,
 		  __entry->lsn,
+		  __print_flags(__entry->flags, "|", XLOG_ICL_STRINGS),
 		  (char *)__entry->caller_ip)
 
 );