Use vn_start_secondary_write() and vn_finished_secondary_write() as a

replacement for vn_write_suspend_wait() to better account for secondary write
processing.

Close race where secondary writes could be started after ffs_sync() returned
but before the file system was marked as suspended.

Detect if secondary writes or softdep processing occurred during vnode sync
loop in ffs_sync() and retry the loop if needed.

3 files changed, 210 insertions, 8 deletions

diff --git a/sys/ufs/ffs/ffs_extern.h b/sys/ufs/ffs/ffs_extern.h
index 8045704..21704bd 100644
--- a/sys/ufs/ffs/ffs_extern.h
+++ b/sys/ufs/ffs/ffs_extern.h
@@ -94,6 +94,10 @@ extern struct vop_vector ffs_fifoops2;
 /*
  * Soft update function prototypes.
  */
+
+int	softdep_check_suspend(struct mount *, struct vnode *,
+	  int, int, int, int);
+void	softdep_get_depcounts(struct mount *, int *, int *);
 void	softdep_initialize(void);
 void	softdep_uninitialize(void);
 int	softdep_mount(struct vnode *, struct mount *, struct fs *,
diff --git a/sys/ufs/ffs/ffs_softdep.c b/sys/ufs/ffs/ffs_softdep.c
index 600203d..53853cc 100644
--- a/sys/ufs/ffs/ffs_softdep.c
+++ b/sys/ufs/ffs/ffs_softdep.c
@@ -346,6 +346,68 @@ softdep_request_cleanup(fs, vp)
 	return (0);
 }
 
+int
+softdep_check_suspend(struct mount *mp,
+		      struct vnode *devvp,
+		      int softdep_deps,
+		      int softdep_accdeps,
+		      int secondary_writes,
+		      int secondary_accwrites)
+{
+	struct bufobj *bo;
+	int error;
+	
+	(void) softdep_deps,
+	(void) softdep_accdeps;
+
+	ASSERT_VI_LOCKED(devvp, "softdep_check_suspend");
+	bo = &devvp->v_bufobj;
+
+	for (;;) {
+		if (!MNT_ITRYLOCK(mp)) {
+			VI_UNLOCK(devvp);
+			MNT_ILOCK(mp);
+			MNT_IUNLOCK(mp);
+			VI_LOCK(devvp);
+			continue;
+		}
+		if (mp->mnt_secondary_writes != 0) {
+			VI_UNLOCK(devvp);
+			msleep(&mp->mnt_secondary_writes,
+			       MNT_MTX(mp),
+			       (PUSER - 1) | PDROP, "secwr", 0);
+			VI_LOCK(devvp);
+			continue;
+		}
+		break;
+	}
+
+	/*
+	 * Reasons for needing more work before suspend:
+	 * - Dirty buffers on devvp.
+	 * - Secondary writes occurred after start of vnode sync loop
+	 */
+	error = 0;
+	if (bo->bo_numoutput > 0 ||
+	    bo->bo_dirty.bv_cnt > 0 ||
+	    secondary_writes != 0 ||
+	    mp->mnt_secondary_writes != 0 ||
+	    secondary_accwrites != mp->mnt_secondary_accwrites)
+		error = EAGAIN;
+	VI_UNLOCK(devvp);
+	return (error);
+}
+
+void
+softdep_get_depcounts(struct mount *mp,
+		      int *softdepactivep,
+		      int *softdepactiveaccp)
+{
+	(void) mp;
+	*softdepactivep = 0;
+	*softdepactiveaccp = 0;
+}
+
 #else
 /*
  * These definitions need to be adapted to the system to which
@@ -500,6 +562,7 @@ static	int softdep_count_dependencies(struct buf *bp, int);
 static struct mtx lk;
 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF);
 
+#define TRY_ACQUIRE_LOCK(lk)		mtx_trylock(lk)
 #define ACQUIRE_LOCK(lk)		mtx_lock(lk)
 #define FREE_LOCK(lk)			mtx_unlock(lk)
 
@@ -588,6 +651,7 @@ workitem_alloc(item, type, mp)
 	item->wk_state = 0;
 	ACQUIRE_LOCK(&lk);
 	VFSTOUFS(mp)->softdep_deps++;
+	VFSTOUFS(mp)->softdep_accdeps++;
 	FREE_LOCK(&lk);
 }
 
@@ -873,7 +937,7 @@ process_worklist_item(mp, flags)
 	}
 	ump->softdep_on_worklist -= 1;
 	FREE_LOCK(&lk);
-	if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
+	if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
 		panic("process_worklist_item: suspended filesystem");
 	matchcnt++;
 	switch (wk->wk_type) {
@@ -903,6 +967,7 @@ process_worklist_item(mp, flags)
 		    "softdep", TYPENAME(wk->wk_type));
 		/* NOTREACHED */
 	}
+	vn_finished_secondary_write(mp);
 	ACQUIRE_LOCK(&lk);
 	return (matchcnt);
 }
@@ -6000,6 +6065,98 @@ getdirtybuf(bp, mtx, waitfor)
 	return (bp);
 }
 
+
+/*
+ * Check if it is safe to suspend the file system now.  On entry,
+ * the vnode interlock for devvp should be held.  Return 0 with
+ * the mount interlock held if the file system can be suspended now,
+ * otherwise return EAGAIN with the mount interlock held.
+ */
+int
+softdep_check_suspend(struct mount *mp,
+		      struct vnode *devvp,
+		      int softdep_deps,
+		      int softdep_accdeps,
+		      int secondary_writes,
+		      int secondary_accwrites)
+{
+	struct bufobj *bo;
+	struct ufsmount *ump;
+	int error;
+
+	ASSERT_VI_LOCKED(devvp, "softdep_check_suspend");
+	ump = VFSTOUFS(mp);
+	bo = &devvp->v_bufobj;
+
+	for (;;) {
+		if (!TRY_ACQUIRE_LOCK(&lk)) {
+			VI_UNLOCK(devvp);
+			ACQUIRE_LOCK(&lk);
+			FREE_LOCK(&lk);
+			VI_LOCK(devvp);
+			continue;
+		}
+		if (!MNT_ITRYLOCK(mp)) {
+			FREE_LOCK(&lk);
+			VI_UNLOCK(devvp);
+			MNT_ILOCK(mp);
+			MNT_IUNLOCK(mp);
+			VI_LOCK(devvp);
+			continue;
+		}
+		if (mp->mnt_secondary_writes != 0) {
+			FREE_LOCK(&lk);
+			VI_UNLOCK(devvp);
+			msleep(&mp->mnt_secondary_writes,
+			       MNT_MTX(mp),
+			       (PUSER - 1) | PDROP, "secwr", 0);
+			VI_LOCK(devvp);
+			continue;
+		}
+		break;
+	}
+
+	/*
+	 * Reasons for needing more work before suspend:
+	 * - Dirty buffers on devvp.
+	 * - Softdep activity occurred after start of vnode sync loop
+	 * - Secondary writes occurred after start of vnode sync loop
+	 */
+	error = 0;
+	if (bo->bo_numoutput > 0 ||
+	    bo->bo_dirty.bv_cnt > 0 ||
+	    softdep_deps != 0 ||
+	    ump->softdep_deps != 0 ||
+	    softdep_accdeps != ump->softdep_accdeps ||
+	    secondary_writes != 0 ||
+	    mp->mnt_secondary_writes != 0 ||
+	    secondary_accwrites != mp->mnt_secondary_accwrites)
+		error = EAGAIN;
+	FREE_LOCK(&lk);
+	VI_UNLOCK(devvp);
+	return (error);
+}
+
+
+/*
+ * Get the number of dependency structures for the file system, both
+ * the current number and the total number allocated.  These will
+ * later be used to detect that softdep processing has occurred.
+ */
+void
+softdep_get_depcounts(struct mount *mp,
+		      int *softdep_depsp,
+		      int *softdep_accdepsp)
+{
+	struct ufsmount *ump;
+
+	ump = VFSTOUFS(mp);
+	ACQUIRE_LOCK(&lk);
+	*softdep_depsp = ump->softdep_deps;
+	*softdep_accdepsp = ump->softdep_accdeps;
+	FREE_LOCK(&lk);
+}
+
 /*
  * Wait for pending output on a vnode to complete.
  * Must be called with vnode lock and interlock locked.
diff --git a/sys/ufs/ffs/ffs_vfsops.c b/sys/ufs/ffs/ffs_vfsops.c
index c6c6510..f4245db 100644
--- a/sys/ufs/ffs/ffs_vfsops.c
+++ b/sys/ufs/ffs/ffs_vfsops.c
@@ -70,7 +70,7 @@ __FBSDID("$FreeBSD$");
 
 static uma_zone_t uma_inode, uma_ufs1, uma_ufs2;
 
-static int	ffs_sbupdate(struct ufsmount *, int);
+static int	ffs_sbupdate(struct ufsmount *, int, int);
 static int	ffs_reload(struct mount *, struct thread *);
 static int	ffs_mountfs(struct vnode *, struct mount *, struct thread *);
 static void	ffs_oldfscompat_read(struct fs *, struct ufsmount *,
@@ -234,7 +234,7 @@ ffs_mount(struct mount *mp, struct thread *td)
 			}
 			if ((fs->fs_flags & (FS_UNCLEAN | FS_NEEDSFSCK)) == 0)
 				fs->fs_clean = 1;
-			if ((error = ffs_sbupdate(ump, MNT_WAIT)) != 0) {
+			if ((error = ffs_sbupdate(ump, MNT_WAIT, 0)) != 0) {
 				fs->fs_ronly = 0;
 				fs->fs_clean = 0;
 				vn_finished_write(mp);
@@ -301,7 +301,7 @@ ffs_mount(struct mount *mp, struct thread *td)
 			fs->fs_ronly = 0;
 			mp->mnt_flag &= ~MNT_RDONLY;
 			fs->fs_clean = 0;
-			if ((error = ffs_sbupdate(ump, MNT_WAIT)) != 0) {
+			if ((error = ffs_sbupdate(ump, MNT_WAIT, 0)) != 0) {
 				vn_finished_write(mp);
 				return (error);
 			}
@@ -806,7 +806,7 @@ ffs_mountfs(devvp, mp, td)
 			ffs_snapshot_mount(mp);
 		fs->fs_fmod = 1;
 		fs->fs_clean = 0;
-		(void) ffs_sbupdate(ump, MNT_WAIT);
+		(void) ffs_sbupdate(ump, MNT_WAIT, 0);
 	}
 	/*
 	 * Initialize filesystem stat information in mount struct.
@@ -988,7 +988,7 @@ ffs_unmount(mp, mntflags, td)
 	UFS_UNLOCK(ump);
 	if (fs->fs_ronly == 0) {
 		fs->fs_clean = fs->fs_flags & (FS_UNCLEAN|FS_NEEDSFSCK) ? 0 : 1;
-		error = ffs_sbupdate(ump, MNT_WAIT);
+		error = ffs_sbupdate(ump, MNT_WAIT, 0);
 		if (error) {
 			fs->fs_clean = 0;
 			return (error);
@@ -1113,6 +1113,12 @@ ffs_sync(mp, waitfor, td)
 	struct ufsmount *ump = VFSTOUFS(mp);
 	struct fs *fs;
 	int error, count, wait, lockreq, allerror = 0;
+	int suspend;
+	int suspended;
+	int secondary_writes;
+	int secondary_accwrites;
+	int softdep_deps;
+	int softdep_accdeps;
 	struct bufobj *bo;
 
 	fs = ump->um_fs;
@@ -1124,14 +1130,30 @@ ffs_sync(mp, waitfor, td)
 	 * Write back each (modified) inode.
 	 */
 	wait = 0;
+	suspend = 0;
+	suspended = 0;
 	lockreq = LK_EXCLUSIVE | LK_NOWAIT;
+	if (waitfor == MNT_SUSPEND) {
+		suspend = 1;
+		waitfor = MNT_WAIT;
+	}
 	if (waitfor == MNT_WAIT) {
 		wait = 1;
 		lockreq = LK_EXCLUSIVE;
 	}
 	lockreq |= LK_INTERLOCK | LK_SLEEPFAIL;
 	MNT_ILOCK(mp);
+	MNT_REF(mp);
 loop:
+	/* Grab snapshot of secondary write counts */
+	secondary_writes = mp->mnt_secondary_writes;
+	secondary_accwrites = mp->mnt_secondary_accwrites;
+
+	/* Grab snapshot of softdep dependency counts */
+	MNT_IUNLOCK(mp);
+	softdep_get_depcounts(mp, &softdep_deps, &softdep_accdeps);
+	MNT_ILOCK(mp);
+
 	MNT_VNODE_FOREACH(vp, mp, mvp) {
 		/*
 		 * Depend on the mntvnode_slock to keep things stable enough
@@ -1165,6 +1187,7 @@ loop:
 		vput(vp);
 		MNT_ILOCK(mp);
 	}
+	MNT_REL(mp);
 	MNT_IUNLOCK(mp);
 	/*
 	 * Force stale filesystem control information to be flushed.
@@ -1194,12 +1217,25 @@ loop:
 			MNT_ILOCK(mp);
 			goto loop;
 		}
+	} else if (suspend != 0) {
+		if (softdep_check_suspend(mp,
+					  devvp,
+					  softdep_deps,
+					  softdep_accdeps,
+					  secondary_writes,
+					  secondary_accwrites) != 0)
+			goto loop;	/* More work needed */
+		mtx_assert(MNT_MTX(mp), MA_OWNED);
+		mp->mnt_kern_flag |= MNTK_SUSPENDED;
+		MNT_IUNLOCK(mp);
+		suspended = 1;
 	} else
 		VI_UNLOCK(devvp);
 	/*
 	 * Write back modified superblock.
 	 */
-	if (fs->fs_fmod != 0 && (error = ffs_sbupdate(ump, waitfor)) != 0)
+	if (fs->fs_fmod != 0 &&
+	    (error = ffs_sbupdate(ump, waitfor, suspended)) != 0)
 		allerror = error;
 	return (allerror);
 }
@@ -1449,9 +1485,10 @@ ffs_uninit(vfsp)
  * Write a superblock and associated information back to disk.
  */
 static int
-ffs_sbupdate(mp, waitfor)
+ffs_sbupdate(mp, waitfor, suspended)
 	struct ufsmount *mp;
 	int waitfor;
+	int suspended;
 {
 	struct fs *fs = mp->um_fs;
 	struct buf *sbbp;
@@ -1482,6 +1519,8 @@ ffs_sbupdate(mp, waitfor)
 		    size, 0, 0, 0);
 		bcopy(space, bp->b_data, (u_int)size);
 		space = (char *)space + size;
+		if (suspended)
+			bp->b_flags |= B_VALIDSUSPWRT;
 		if (waitfor != MNT_WAIT)
 			bawrite(bp);
 		else if ((error = bwrite(bp)) != 0)
@@ -1513,6 +1552,8 @@ ffs_sbupdate(mp, waitfor)
 	fs->fs_time = time_second;
 	bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
 	ffs_oldfscompat_write((struct fs *)bp->b_data, mp);
+	if (suspended)
+		bp->b_flags |= B_VALIDSUSPWRT;
 	if (waitfor != MNT_WAIT)
 		bawrite(bp);
 	else if ((error = bwrite(bp)) != 0)