This change avoids a kernel deadlock on "snaplk" when using

snapshots on UFS filesystems running with journaled soft updates.
This is the first of several bugs that need to be fixed before
removing the restriction added in -r230250 to prevent the use
of snapshots on filesystems running with journaled soft updates.

The deadlock occurs when holding the snapshot lock (snaplk)
and then trying to flush an inode via ffs_update(). We become
blocked by another process trying to flush a different inode
contained in the same inode block that we need. It holds the
inode block for which we are waiting locked. When it tries to
write the inode block, it gets blocked waiting for the our
snaplk when it calls ffs_copyonwrite() to see if the inode
block needs to be copied in our snapshot.

The most obvious place that this deadlock arises is in the
ffs_copyonwrite() routine when it updates critical metadata
in a snapshot and tries to write it out before proceeding.
The fix here is to write the data and indirect block pointer
for the snapshot, but to skip the call to ffs_update() to
write the snapshot inode. To ensure that we will never have
to update a pointer in the inode itself, the ffs_snapshot()
routine that creates the snapshot has to ensure that all the
direct blocks are allocated as part of the creation of the
snapshot.

A less obvious place that this deadlock occurs is when we hold
the snaplk because we are deleting a snapshot. In the course of
doing the deletion, we need to allocate various soft update
dependency structures and allocate some journal space. If we
hit a resource limit while doing this we decrease the resources
in use by flushing out an existing dirty file to get it to give
up the soft dependency resources that it holds. The flush can
cause an ffs_update() to be done on the inode for the file that
we have selected to flush resulting in the same deadlock as
described above when the inode that we have chosen to flush
resides in the same inode block as the snapshot inode that we hold.
The fix is to defer cleaning up any time that the inode on which
we are operating is a snapshot.

Help and review by:    Jeff Roberson
Tested by:             Peter Holm
MFC (to 9 only) after: 2 weeks

8 files changed, 172 insertions, 99 deletions

diff --git a/sys/kern/vfs_bio.c b/sys/kern/vfs_bio.c
index c0a80f340334..4eac76d133a1 100644
--- a/sys/kern/vfs_bio.c
+++ b/sys/kern/vfs_bio.c
@@ -782,21 +782,6 @@ bremfreel(struct buf *bp)
 	}
 }
 
-
-/*
- * Get a buffer with the specified data.  Look in the cache first.  We
- * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
- * is set, the buffer is valid and we do not have to do anything ( see
- * getblk() ).  This is really just a special case of breadn().
- */
-int
-bread(struct vnode * vp, daddr_t blkno, int size, struct ucred * cred,
-    struct buf **bpp)
-{
-
-	return (breadn(vp, blkno, size, 0, 0, 0, cred, bpp));
-}
-
 /*
  * Attempt to initiate asynchronous I/O on read-ahead blocks.  We must
  * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set,
@@ -834,19 +819,28 @@ breada(struct vnode * vp, daddr_t * rablkno, int * rabsize,
 }
 
 /*
- * Operates like bread, but also starts asynchronous I/O on
- * read-ahead blocks.
+ * Entry point for bread() and breadn() via #defines in sys/buf.h.
+ *
+ * Get a buffer with the specified data.  Look in the cache first.  We
+ * must clear BIO_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
+ * is set, the buffer is valid and we do not have to do anything, see
+ * getblk(). Also starts asynchronous I/O on read-ahead blocks.
  */
 int
-breadn(struct vnode * vp, daddr_t blkno, int size,
-    daddr_t * rablkno, int *rabsize,
-    int cnt, struct ucred * cred, struct buf **bpp)
+breadn_flags(struct vnode * vp, daddr_t blkno, int size,
+    daddr_t * rablkno, int *rabsize, int cnt,
+    struct ucred * cred, int flags, struct buf **bpp)
 {
 	struct buf *bp;
 	int rv = 0, readwait = 0;
 
 	CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size);
-	*bpp = bp = getblk(vp, blkno, size, 0, 0, 0);
+	/*
+	 * Can only return NULL if GB_LOCK_NOWAIT flag is specified.
+	 */
+	*bpp = bp = getblk(vp, blkno, size, 0, 0, flags);
+	if (bp == NULL)
+		return (EBUSY);
 
 	/* if not found in cache, do some I/O */
 	if ((bp->b_flags & B_CACHE) == 0) {
diff --git a/sys/sys/buf.h b/sys/sys/buf.h
index 972b05379fa5..96021eb5ea48 100644
--- a/sys/sys/buf.h
+++ b/sys/sys/buf.h
@@ -479,10 +479,13 @@ void	bwillwrite(void);
 int	buf_dirty_count_severe(void);
 void	bremfree(struct buf *);
 void	bremfreef(struct buf *);	/* XXX Force bremfree, only for nfs. */
-int	bread(struct vnode *, daddr_t, int, struct ucred *, struct buf **);
+#define bread(vp, blkno, size, cred, bpp) \
+	    breadn_flags(vp, blkno, size, 0, 0, 0, cred, 0, bpp)
+#define breadn(vp, blkno, size, rablkno, rabsize, cnt, cred, bpp) \
+	    breadn_flags(vp, blkno, size, rablkno, rabsize, cnt, cred, 0, bpp)
+int	breadn_flags(struct vnode *, daddr_t, int, daddr_t *, int *, int,
+	    struct ucred *, int, struct buf **);
 void	breada(struct vnode *, daddr_t *, int *, int, struct ucred *);
-int	breadn(struct vnode *, daddr_t, int, daddr_t *, int *, int,
-	    struct ucred *, struct buf **);
 void	bdwrite(struct buf *);
 void	bawrite(struct buf *);
 void	bdirty(struct buf *);
diff --git a/sys/ufs/ffs/ffs_extern.h b/sys/ufs/ffs/ffs_extern.h
index bfc02564824f..258966cae2bb 100644
--- a/sys/ufs/ffs/ffs_extern.h
+++ b/sys/ufs/ffs/ffs_extern.h
@@ -167,6 +167,13 @@ void	softdep_freework(struct workhead *);
 #define FLUSH_INODES_WAIT	2
 #define FLUSH_BLOCKS		3
 #define FLUSH_BLOCKS_WAIT	4
+/*
+ * Flag to ffs_syncinode() to request flushing of data only,
+ * but skip the ffs_update() on the inode itself. Used to avoid
+ * deadlock when flushing snapshot inodes while holding snaplk.
+ * Avoid bit conflicts with MNT_WAIT values in sys/mount.h
+ */
+#define	NO_INO_UPDT		0x10
 
 int	ffs_rdonly(struct inode *);
 
diff --git a/sys/ufs/ffs/ffs_inode.c b/sys/ufs/ffs/ffs_inode.c
index 101bdb43aa97..2667e318f1ba 100644
--- a/sys/ufs/ffs/ffs_inode.c
+++ b/sys/ufs/ffs/ffs_inode.c
@@ -81,7 +81,7 @@ ffs_update(vp, waitfor)
 	struct fs *fs;
 	struct buf *bp;
 	struct inode *ip;
-	int error;
+	int flags, error;
 
 	ASSERT_VOP_ELOCKED(vp, "ffs_update");
 	ufs_itimes(vp);
@@ -92,11 +92,36 @@ ffs_update(vp, waitfor)
 	fs = ip->i_fs;
 	if (fs->fs_ronly && ip->i_ump->um_fsckpid == 0)
 		return (0);
-	error = bread(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
-		(int)fs->fs_bsize, NOCRED, &bp);
-	if (error) {
-		brelse(bp);
-		return (error);
+	/*
+	 * If we are updating a snapshot and another process is currently
+	 * writing the buffer containing the inode for this snapshot then
+	 * a deadlock can occur when it tries to check the snapshot to see
+	 * if that block needs to be copied. Thus when updating a snapshot
+	 * we check to see if the buffer is already locked, and if it is
+	 * we drop the snapshot lock until the buffer has been written
+	 * and is available to us. We have to grab a reference to the
+	 * snapshot vnode to prevent it from being removed while we are
+	 * waiting for the buffer.
+	 */
+	flags = 0;
+	if (IS_SNAPSHOT(ip))
+		flags = GB_LOCK_NOWAIT;
+	error = breadn_flags(ip->i_devvp,
+	     fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
+	     (int) fs->fs_bsize, 0, 0, 0, NOCRED, flags, &bp);
+	if (error != 0) {
+		if (error != EBUSY) {
+			brelse(bp);
+			return (error);
+		}
+		KASSERT((IS_SNAPSHOT(ip)), ("EBUSY from non-snapshot"));
+		vref(vp);	/* Protect against ffs_snapgone() */
+		VOP_UNLOCK(vp, 0);
+		(void) bread(ip->i_devvp,
+		     fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
+		     (int) fs->fs_bsize, NOCRED, &bp);
+		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
+		vrele(vp);
 	}
 	if (DOINGSOFTDEP(vp))
 		softdep_update_inodeblock(ip, bp, waitfor);
@@ -108,16 +133,16 @@ ffs_update(vp, waitfor)
 	else
 		*((struct ufs2_dinode *)bp->b_data +
 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
-	if (waitfor && !DOINGASYNC(vp)) {
-		return (bwrite(bp));
-	} else if (vm_page_count_severe() || buf_dirty_count_severe()) {
-		return (bwrite(bp));
+	if ((waitfor && !DOINGASYNC(vp)) ||
+	    (vm_page_count_severe() || buf_dirty_count_severe())) {
+		error = bwrite(bp);
 	} else {
 		if (bp->b_bufsize == fs->fs_bsize)
 			bp->b_flags |= B_CLUSTEROK;
 		bdwrite(bp);
-		return (0);
+		error = 0;
 	}
+	return (error);
 }
 
 #define	SINGLE	0	/* index of single indirect block */
@@ -253,7 +278,7 @@ ffs_truncate(vp, length, flags, cred, td)
 	}
 	if (fs->fs_ronly)
 		panic("ffs_truncate: read-only filesystem");
-	if ((ip->i_flags & SF_SNAPSHOT) != 0)
+	if (IS_SNAPSHOT(ip))
 		ffs_snapremove(vp);
 	vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
 	osize = ip->i_size;
diff --git a/sys/ufs/ffs/ffs_snapshot.c b/sys/ufs/ffs/ffs_snapshot.c
index adc1d781755e..c930f5dd71c4 100644
--- a/sys/ufs/ffs/ffs_snapshot.c
+++ b/sys/ufs/ffs/ffs_snapshot.c
@@ -203,6 +203,7 @@ ffs_snapshot(mp, snapfile)
 	ufs2_daddr_t numblks, blkno, *blkp, *snapblklist;
 	int error, cg, snaploc;
 	int i, size, len, loc;
+	ufs2_daddr_t blockno;
 	uint64_t flag;
 	struct timespec starttime = {0, 0}, endtime;
 	char saved_nice = 0;
@@ -529,7 +530,7 @@ loop:
 		    (xvp->v_usecount == 0 &&
 		     (xvp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == 0) ||
 		    xvp->v_type == VNON ||
-		    (VTOI(xvp)->i_flags & SF_SNAPSHOT)) {
+		    IS_SNAPSHOT(VTOI(xvp))) {
 			VI_UNLOCK(xvp);
 			MNT_ILOCK(mp);
 			continue;
@@ -815,21 +816,26 @@ out1:
 	if (space != NULL)
 		free(space, M_UFSMNT);
 	/*
-	 * If another process is currently writing the buffer containing
-	 * the inode for this snapshot then a deadlock can occur. Drop
-	 * the snapshot lock until the buffer has been written.
+	 * Preallocate all the direct blocks in the snapshot inode so
+	 * that we never have to write the inode itself to commit an
+	 * update to the contents of the snapshot. Note that once
+	 * created, the size of the snapshot will never change, so
+	 * there will never be a need to write the inode except to
+	 * update the non-integrity-critical time fields and
+	 * allocated-block count.
 	 */
-	VREF(vp);	/* Protect against ffs_snapgone() */
-	VOP_UNLOCK(vp, 0);
-	(void) bread(ip->i_devvp,
-		     fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
-		     (int) fs->fs_bsize, NOCRED, &nbp);
-	brelse(nbp);
-	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
-	if (ip->i_effnlink == 0)
-		error = ENOENT;		/* Snapshot file unlinked */
-	else
-		vrele(vp);		/* Drop extra reference */
+	for (blockno = 0; blockno < NDADDR; blockno++) {
+		if (DIP(ip, i_db[blockno]) != 0)
+			continue;
+		error = UFS_BALLOC(vp, lblktosize(fs, blockno),
+		    fs->fs_bsize, KERNCRED, BA_CLRBUF, &bp);
+		if (error)
+			break;
+		error = readblock(vp, bp, blockno);
+		bawrite(bp);
+		if (error != 0)
+			break;
+	}
 done:
 	free(copy_fs->fs_csp, M_UFSMNT);
 	free(copy_fs, M_UFSMNT);
@@ -1902,7 +1908,7 @@ retry:
 			bawrite(cbp);
 			if ((vtype == VDIR || dopersistence) &&
 			    ip->i_effnlink > 0)
-				(void) ffs_syncvnode(vp, MNT_WAIT);
+				(void) ffs_syncvnode(vp, MNT_WAIT|NO_INO_UPDT);
 			continue;
 		}
 		/*
@@ -1913,7 +1919,7 @@ retry:
 			bawrite(cbp);
 			if ((vtype == VDIR || dopersistence) &&
 			    ip->i_effnlink > 0)
-				(void) ffs_syncvnode(vp, MNT_WAIT);
+				(void) ffs_syncvnode(vp, MNT_WAIT|NO_INO_UPDT);
 			break;
 		}
 		savedcbp = cbp;
@@ -1931,7 +1937,7 @@ retry:
 		bawrite(savedcbp);
 		if ((vtype == VDIR || dopersistence) &&
 		    VTOI(vp)->i_effnlink > 0)
-			(void) ffs_syncvnode(vp, MNT_WAIT);
+			(void) ffs_syncvnode(vp, MNT_WAIT|NO_INO_UPDT);
 	}
 	/*
 	 * If we have been unable to allocate a block in which to do
@@ -1987,9 +1993,9 @@ ffs_snapshot_mount(mp)
 			continue;
 		}
 		ip = VTOI(vp);
-		if ((ip->i_flags & SF_SNAPSHOT) == 0 || ip->i_size ==
+		if (!IS_SNAPSHOT(ip) || ip->i_size ==
 		    lblktosize(fs, howmany(fs->fs_size, fs->fs_frag))) {
-			if ((ip->i_flags & SF_SNAPSHOT) == 0) {
+			if (!IS_SNAPSHOT(ip)) {
 				reason = "non-snapshot";
 			} else {
 				reason = "old format snapshot";
@@ -2250,7 +2256,7 @@ ffs_copyonwrite(devvp, bp)
 	int launched_async_io, prev_norunningbuf;
 	long saved_runningbufspace;
 
-	if (devvp != bp->b_vp && (VTOI(bp->b_vp)->i_flags & SF_SNAPSHOT) != 0)
+	if (devvp != bp->b_vp && IS_SNAPSHOT(VTOI(bp->b_vp)))
 		return (0);		/* Update on a snapshot file */
 	if (td->td_pflags & TDP_COWINPROGRESS)
 		panic("ffs_copyonwrite: recursive call");
@@ -2395,7 +2401,7 @@ ffs_copyonwrite(devvp, bp)
 			bawrite(cbp);
 			if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR ||
 			    dopersistence) && ip->i_effnlink > 0)
-				(void) ffs_syncvnode(vp, MNT_WAIT);
+				(void) ffs_syncvnode(vp, MNT_WAIT|NO_INO_UPDT);
 			else
 				launched_async_io = 1;
 			continue;
@@ -2408,7 +2414,7 @@ ffs_copyonwrite(devvp, bp)
 			bawrite(cbp);
 			if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR ||
 			    dopersistence) && ip->i_effnlink > 0)
-				(void) ffs_syncvnode(vp, MNT_WAIT);
+				(void) ffs_syncvnode(vp, MNT_WAIT|NO_INO_UPDT);
 			else
 				launched_async_io = 1;
 			break;
@@ -2428,7 +2434,7 @@ ffs_copyonwrite(devvp, bp)
 		bawrite(savedcbp);
 		if ((devvp == bp->b_vp || bp->b_vp->v_type == VDIR ||
 		    dopersistence) && VTOI(vp)->i_effnlink > 0)
-			(void) ffs_syncvnode(vp, MNT_WAIT);
+			(void) ffs_syncvnode(vp, MNT_WAIT|NO_INO_UPDT);
 		else
 			launched_async_io = 1;
 	}
@@ -2478,7 +2484,7 @@ ffs_sync_snap(mp, waitfor)
 	}
 	TAILQ_FOREACH(ip, &sn->sn_head, i_nextsnap) {
 		vp = ITOV(ip);
-		ffs_syncvnode(vp, waitfor);
+		ffs_syncvnode(vp, waitfor|NO_INO_UPDT);
 	}
 	lockmgr(&sn->sn_lock, LK_RELEASE, NULL);
 }
diff --git a/sys/ufs/ffs/ffs_softdep.c b/sys/ufs/ffs/ffs_softdep.c
index 9639076b04aa..debb8389e80a 100644
--- a/sys/ufs/ffs/ffs_softdep.c
+++ b/sys/ufs/ffs/ffs_softdep.c
@@ -2827,7 +2827,12 @@ softdep_prealloc(vp, waitok)
 {
 	struct ufsmount *ump;
 
-	if (DOINGSUJ(vp) == 0)
+	/*
+	 * Nothing to do if we are not running journaled soft updates.
+	 * If we currently hold the snapshot lock, we must avoid handling
+	 * other resources that could cause deadlock.
+	 */
+	if (DOINGSUJ(vp) == 0 || IS_SNAPSHOT(VTOI(vp)))
 		return (0);
 	ump = VFSTOUFS(vp->v_mount);
 	ACQUIRE_LOCK(&lk);
@@ -2873,7 +2878,12 @@ softdep_prelink(dvp, vp)
 
 	ump = VFSTOUFS(dvp->v_mount);
 	mtx_assert(&lk, MA_OWNED);
-	if (journal_space(ump, 0))
+	/*
+	 * Nothing to do if we have sufficient journal space.
+	 * If we currently hold the snapshot lock, we must avoid
+	 * handling other resources that could cause deadlock.
+	 */
+	if (journal_space(ump, 0) || (vp && IS_SNAPSHOT(VTOI(vp))))
 		return;
 	stat_journal_low++;
 	FREE_LOCK(&lk);
@@ -4304,11 +4314,15 @@ inodedep_lookup_ip(ip)
 	struct inode *ip;
 {
 	struct inodedep *inodedep;
+	int dflags;
 
 	KASSERT(ip->i_nlink >= ip->i_effnlink,
 	    ("inodedep_lookup_ip: bad delta"));
-	(void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number,
-	    DEPALLOC, &inodedep);
+	dflags = DEPALLOC;
+	if (IS_SNAPSHOT(ip))
+		dflags |= NODELAY;
+	(void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags,
+	    &inodedep);
 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
 
 	return (inodedep);
@@ -4696,7 +4710,7 @@ softdep_setup_inomapdep(bp, ip, newinum, mode)
 	 * the cylinder group map from which it was allocated.
 	 */
 	ACQUIRE_LOCK(&lk);
-	if ((inodedep_lookup(mp, newinum, DEPALLOC|NODELAY, &inodedep)))
+	if ((inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep)))
 		panic("softdep_setup_inomapdep: dependency %p for new"
 		    "inode already exists", inodedep);
 	bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum));
@@ -5437,6 +5451,7 @@ softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
 	struct allocindir *aip;
 	struct pagedep *pagedep;
 	struct mount *mp;
+	int dflags;
 
 	if (lbn != nbp->b_lblkno)
 		panic("softdep_setup_allocindir_page: lbn %jd != lblkno %jd",
@@ -5444,7 +5459,10 @@ softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
 	mp = UFSTOVFS(ip->i_ump);
 	aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn);
-	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
+	dflags = DEPALLOC;
+	if (IS_SNAPSHOT(ip))
+		dflags |= NODELAY;
+	(void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 	/*
 	 * If we are allocating a directory page, then we must
 	 * allocate an associated pagedep to track additions and
@@ -5474,11 +5492,15 @@ softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
 	struct inodedep *inodedep;
 	struct allocindir *aip;
 	ufs_lbn_t lbn;
+	int dflags;
 
 	lbn = nbp->b_lblkno;
 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
 	aip = newallocindir(ip, ptrno, newblkno, 0, lbn);
-	inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, &inodedep);
+	dflags = DEPALLOC;
+	if (IS_SNAPSHOT(ip))
+		dflags |= NODELAY;
+	inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep);
 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
 	if (setup_allocindir_phase2(bp, ip, inodedep, aip, lbn))
 		panic("softdep_setup_allocindir_meta: Block already existed");
@@ -6085,11 +6107,7 @@ softdep_journal_freeblocks(ip, cred, length, flags)
 	struct mount *mp;
 	ufs2_daddr_t extblocks, datablocks;
 	ufs_lbn_t tmpval, lbn, lastlbn;
-	int frags;
-	int lastoff, iboff;
-	int allocblock;
-	int error, i;
-	int needj;
+	int frags, lastoff, iboff, allocblock, needj, dflags, error, i;
 
 	fs = ip->i_fs;
 	mp = UFSTOVFS(ip->i_ump);
@@ -6107,7 +6125,10 @@ softdep_journal_freeblocks(ip, cred, length, flags)
 	 * we don't need to journal the block frees.  The canceled journals
 	 * for the allocations will suffice.
 	 */
-	inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
+	dflags = DEPALLOC;
+	if (IS_SNAPSHOT(ip))
+		dflags |= NODELAY;
+	inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 	if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED &&
 	    length == 0)
 		needj = 0;
@@ -6232,7 +6253,7 @@ softdep_journal_freeblocks(ip, cred, length, flags)
 		*((struct ufs2_dinode *)bp->b_data +
 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
 	ACQUIRE_LOCK(&lk);
-	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
+	(void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 	if ((inodedep->id_state & IOSTARTED) != 0)
 		panic("softdep_setup_freeblocks: inode busy");
 	/*
@@ -6310,7 +6331,7 @@ softdep_journal_freeblocks(ip, cred, length, flags)
 
 	}
 	ACQUIRE_LOCK(&lk);
-	inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
+	inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 	TAILQ_INSERT_TAIL(&inodedep->id_freeblklst, freeblks, fb_next);
 	freeblks->fb_state |= DEPCOMPLETE | ONDEPLIST;
 	/*
@@ -6398,7 +6419,7 @@ softdep_setup_freeblocks(ip, length, flags)
 	struct fs *fs;
 	ufs2_daddr_t extblocks, datablocks;
 	struct mount *mp;
-	int i, delay, error;
+	int i, delay, error, dflags;
 	ufs_lbn_t tmpval;
 	ufs_lbn_t lbn;
 
@@ -6463,7 +6484,10 @@ softdep_setup_freeblocks(ip, length, flags)
 	 * Find and eliminate any inode dependencies.
 	 */
 	ACQUIRE_LOCK(&lk);
-	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
+	dflags = DEPALLOC;
+	if (IS_SNAPSHOT(ip))
+		dflags |= NODELAY;
+	(void) inodedep_lookup(mp, ip->i_number, dflags, &inodedep);
 	if ((inodedep->id_state & IOSTARTED) != 0)
 		panic("softdep_setup_freeblocks: inode busy");
 	/*
@@ -8029,7 +8053,7 @@ softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
 	dap->da_pagedep = pagedep;
 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
 	    da_pdlist);
-	inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
+	inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep);
 	/*
 	 * If we're journaling, link the diradd into the jaddref so it
 	 * may be completed after the journal entry is written.  Otherwise,
@@ -8631,8 +8655,7 @@ newdirrem(bp, dp, ip, isrmdir, prevdirremp)
 	 * the number of freefile and freeblks structures.
 	 */
 	ACQUIRE_LOCK(&lk);
-	if (!(ip->i_flags & SF_SNAPSHOT) &&
-	    dep_current[D_DIRREM] > max_softdeps / 2)
+	if (!IS_SNAPSHOT(ip) && dep_current[D_DIRREM] > max_softdeps / 2)
 		(void) request_cleanup(ITOV(dp)->v_mount, FLUSH_BLOCKS);
 	FREE_LOCK(&lk);
 	dirrem = malloc(sizeof(struct dirrem),
@@ -8866,11 +8889,11 @@ softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
 	/*
 	 * Lookup the jaddref for this journal entry.  We must finish
 	 * initializing it and make the diradd write dependent on it.
-	 * If we're not journaling Put it on the id_bufwait list if the inode
-	 * is not yet written. If it is written, do the post-inode write
-	 * processing to put it on the id_pendinghd list.
+	 * If we're not journaling, put it on the id_bufwait list if the
+	 * inode is not yet written. If it is written, do the post-inode
+	 * write processing to put it on the id_pendinghd list.
 	 */
-	inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
+	inodedep_lookup(mp, newinum, DEPALLOC | NODELAY, &inodedep);
 	if (MOUNTEDSUJ(mp)) {
 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
 		    inoreflst);
@@ -8912,9 +8935,13 @@ softdep_change_linkcnt(ip)
 	struct inode *ip;	/* the inode with the increased link count */
 {
 	struct inodedep *inodedep;
+	int dflags;
 
 	ACQUIRE_LOCK(&lk);
-	inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, &inodedep);
+	dflags = DEPALLOC;
+	if (IS_SNAPSHOT(ip))
+		dflags |= NODELAY;
+	inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, dflags, &inodedep);
 	if (ip->i_nlink < ip->i_effnlink)
 		panic("softdep_change_linkcnt: bad delta");
 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
@@ -12544,22 +12571,25 @@ softdep_request_cleanup(fs, vp, cred, resource)
 	ufs2_daddr_t needed;
 	int error;
 
-	mp = vp->v_mount;
-	ump = VFSTOUFS(mp);
-	mtx_assert(UFS_MTX(ump), MA_OWNED);
-	if (resource == FLUSH_BLOCKS_WAIT)
-		stat_cleanup_blkrequests += 1;
-	else
-		stat_cleanup_inorequests += 1;
-
 	/*
 	 * If we are being called because of a process doing a
 	 * copy-on-write, then it is not safe to process any
 	 * worklist items as we will recurse into the copyonwrite
 	 * routine.  This will result in an incoherent snapshot.
+	 * If the vnode that we hold is a snapshot, we must avoid
+	 * handling other resources that could cause deadlock.
 	 */
-	if (curthread->td_pflags & TDP_COWINPROGRESS)
+	if ((curthread->td_pflags & TDP_COWINPROGRESS) || IS_SNAPSHOT(VTOI(vp)))
 		return (0);
+
+	if (resource == FLUSH_BLOCKS_WAIT)
+		stat_cleanup_blkrequests += 1;
+	else
+		stat_cleanup_inorequests += 1;
+
+	mp = vp->v_mount;
+	ump = VFSTOUFS(mp);
+	mtx_assert(UFS_MTX(ump), MA_OWNED);
 	UFS_UNLOCK(ump);
 	error = ffs_update(vp, 1);
 	if (error != 0) {
diff --git a/sys/ufs/ffs/ffs_vnops.c b/sys/ufs/ffs/ffs_vnops.c
index 4783f65c29f2..9e8648e85035 100644
--- a/sys/ufs/ffs/ffs_vnops.c
+++ b/sys/ufs/ffs/ffs_vnops.c
@@ -216,8 +216,10 @@ ffs_syncvnode(struct vnode *vp, int waitfor)
 	struct buf *bp;
 	struct buf *nbp;
 	ufs_lbn_t lbn;
-	int error, wait, passes;
+	int error, wait, passes, noupdate;
 
+	noupdate = waitfor & NO_INO_UPDT;
+	waitfor &= ~NO_INO_UPDT;
 	ip = VTOI(vp);
 	ip->i_flag &= ~IN_NEEDSYNC;
 	bo = &vp->v_bufobj;
@@ -300,7 +302,10 @@ next:
 	}
 	if (waitfor != MNT_WAIT) {
 		BO_UNLOCK(bo);
-		return (ffs_update(vp, waitfor));
+		if (noupdate)
+			return (0);
+		else
+			return (ffs_update(vp, waitfor));
 	}
 	/* Drain IO to see if we're done. */
 	bufobj_wwait(bo, 0, 0);
@@ -317,7 +322,7 @@ next:
 	 */
 	if (bo->bo_dirty.bv_cnt > 0) {
 		/* Write the inode after sync passes to flush deps. */
-		if (wait && DOINGSOFTDEP(vp)) {
+		if (wait && DOINGSOFTDEP(vp) && noupdate == 0) {
 			BO_UNLOCK(bo);
 			ffs_update(vp, MNT_WAIT);
 			BO_LOCK(bo);
@@ -332,7 +337,9 @@ next:
 #endif
 	}
 	BO_UNLOCK(bo);
-	error = ffs_update(vp, MNT_WAIT);
+	error = 0;
+	if (noupdate == 0)
+		error = ffs_update(vp, MNT_WAIT);
 	if (DOINGSUJ(vp))
 		softdep_journal_fsync(VTOI(vp));
 	return (error);
diff --git a/sys/ufs/ufs/inode.h b/sys/ufs/ufs/inode.h
index 3519ca24f54c..2b020007f835 100644
--- a/sys/ufs/ufs/inode.h
+++ b/sys/ufs/ufs/inode.h
@@ -158,6 +158,7 @@ struct inode {
 #define	SHORTLINK(ip) \
 	(((ip)->i_ump->um_fstype == UFS1) ? \
 	(caddr_t)(ip)->i_din1->di_db : (caddr_t)(ip)->i_din2->di_db)
+#define IS_SNAPSHOT(ip)		((ip)->i_flags & SF_SNAPSHOT)
 
 /*
  * Structure used to pass around logical block paths generated by