MFC

4 files changed, 129 insertions, 23 deletions

diff --git a/sys/ufs/ffs/ffs_alloc.c b/sys/ufs/ffs/ffs_alloc.c
index abe4073..789a7cf 100644
--- a/sys/ufs/ffs/ffs_alloc.c
+++ b/sys/ufs/ffs/ffs_alloc.c
@@ -1790,6 +1790,17 @@ fail:
 	return (0);
 }
 
+static inline struct buf *
+getinobuf(struct inode *ip, u_int cg, u_int32_t cginoblk, int gbflags)
+{
+	struct fs *fs;
+
+	fs = ip->i_fs;
+	return (getblk(ip->i_devvp, fsbtodb(fs, ino_to_fsba(fs,
+	    cg * fs->fs_ipg + cginoblk)), (int)fs->fs_bsize, 0, 0,
+	    gbflags));
+}
+
 /*
  * Determine whether an inode can be allocated.
  *
@@ -1814,9 +1825,11 @@ ffs_nodealloccg(ip, cg, ipref, mode, unused)
 	u_int8_t *inosused, *loc;
 	struct ufs2_dinode *dp2;
 	int error, start, len, i;
+	u_int32_t old_initediblk;
 
 	fs = ip->i_fs;
 	ump = ip->i_ump;
+check_nifree:
 	if (fs->fs_cs(fs, cg).cs_nifree == 0)
 		return (0);
 	UFS_UNLOCK(ump);
@@ -1828,13 +1841,13 @@ ffs_nodealloccg(ip, cg, ipref, mode, unused)
 		return (0);
 	}
 	cgp = (struct cg *)bp->b_data;
+restart:
 	if (!cg_chkmagic(cgp) || cgp->cg_cs.cs_nifree == 0) {
 		brelse(bp);
 		UFS_LOCK(ump);
 		return (0);
 	}
 	bp->b_xflags |= BX_BKGRDWRITE;
-	cgp->cg_old_time = cgp->cg_time = time_second;
 	inosused = cg_inosused(cgp);
 	if (ipref) {
 		ipref %= fs->fs_ipg;
@@ -1856,7 +1869,6 @@ ffs_nodealloccg(ip, cg, ipref, mode, unused)
 		}
 	}
 	ipref = (loc - inosused) * NBBY + ffs(~*loc) - 1;
-	cgp->cg_irotor = ipref;
 gotit:
 	/*
 	 * Check to see if we need to initialize more inodes.
@@ -1864,9 +1876,37 @@ gotit:
 	if (fs->fs_magic == FS_UFS2_MAGIC &&
 	    ipref + INOPB(fs) > cgp->cg_initediblk &&
 	    cgp->cg_initediblk < cgp->cg_niblk) {
-		ibp = getblk(ip->i_devvp, fsbtodb(fs,
-		    ino_to_fsba(fs, cg * fs->fs_ipg + cgp->cg_initediblk)),
-		    (int)fs->fs_bsize, 0, 0, 0);
+		old_initediblk = cgp->cg_initediblk;
+
+		/*
+		 * Free the cylinder group lock before writing the
+		 * initialized inode block.  Entering the
+		 * babarrierwrite() with the cylinder group lock
+		 * causes lock order violation between the lock and
+		 * snaplk.
+		 *
+		 * Another thread can decide to initialize the same
+		 * inode block, but whichever thread first gets the
+		 * cylinder group lock after writing the newly
+		 * allocated inode block will update it and the other
+		 * will realize that it has lost and leave the
+		 * cylinder group unchanged.
+		 */
+		ibp = getinobuf(ip, cg, old_initediblk, GB_LOCK_NOWAIT);
+		brelse(bp);
+		if (ibp == NULL) {
+			/*
+			 * The inode block buffer is already owned by
+			 * another thread, which must initialize it.
+			 * Wait on the buffer to allow another thread
+			 * to finish the updates, with dropped cg
+			 * buffer lock, then retry.
+			 */
+			ibp = getinobuf(ip, cg, old_initediblk, 0);
+			brelse(ibp);
+			UFS_LOCK(ump);
+			goto check_nifree;
+		}
 		bzero(ibp->b_data, (int)fs->fs_bsize);
 		dp2 = (struct ufs2_dinode *)(ibp->b_data);
 		for (i = 0; i < INOPB(fs); i++) {
@@ -1883,8 +1923,29 @@ gotit:
 		 * loading of newly created filesystems.
 		 */
 		babarrierwrite(ibp);
-		cgp->cg_initediblk += INOPB(fs);
+
+		/*
+		 * After the inode block is written, try to update the
+		 * cg initediblk pointer.  If another thread beat us
+		 * to it, then leave it unchanged as the other thread
+		 * has already set it correctly.
+		 */
+		error = bread(ip->i_devvp, fsbtodb(fs, cgtod(fs, cg)),
+		    (int)fs->fs_cgsize, NOCRED, &bp);
+		UFS_LOCK(ump);
+		ACTIVECLEAR(fs, cg);
+		UFS_UNLOCK(ump);
+		if (error != 0) {
+			brelse(bp);
+			return (error);
+		}
+		cgp = (struct cg *)bp->b_data;
+		if (cgp->cg_initediblk == old_initediblk)
+			cgp->cg_initediblk += INOPB(fs);
+		goto restart;
 	}
+	cgp->cg_old_time = cgp->cg_time = time_second;
+	cgp->cg_irotor = ipref;
 	UFS_LOCK(ump);
 	ACTIVECLEAR(fs, cg);
 	setbit(inosused, ipref);
diff --git a/sys/ufs/ffs/ffs_softdep.c b/sys/ufs/ffs/ffs_softdep.c
index 16fe134..e39fd46 100644
--- a/sys/ufs/ffs/ffs_softdep.c
+++ b/sys/ufs/ffs/ffs_softdep.c
@@ -1908,7 +1908,12 @@ softdep_flushfiles(oldmnt, flags, td)
 	int flags;
 	struct thread *td;
 {
-	int error, depcount, loopcnt, retry_flush_count, retry;
+#ifdef QUOTA
+	struct ufsmount *ump;
+	int i;
+#endif
+	int error, early, depcount, loopcnt, retry_flush_count, retry;
+	int morework;
 
 	loopcnt = 10;
 	retry_flush_count = 3;
@@ -1926,7 +1931,9 @@ retry_flush:
 		 * Do another flush in case any vnodes were brought in
 		 * as part of the cleanup operations.
 		 */
-		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
+		early = retry_flush_count == 1 || (oldmnt->mnt_kern_flag &
+		    MNTK_UNMOUNT) == 0 ? 0 : EARLYFLUSH;
+		if ((error = ffs_flushfiles(oldmnt, flags | early, td)) != 0)
 			break;
 		if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 ||
 		    depcount == 0)
@@ -1950,7 +1957,17 @@ retry_flush:
 			MNT_ILOCK(oldmnt);
 			KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0,
 			    ("softdep_flushfiles: !MNTK_NOINSMNTQ"));
-			if (oldmnt->mnt_nvnodelistsize > 0) {
+			morework = oldmnt->mnt_nvnodelistsize > 0;
+#ifdef QUOTA
+			ump = VFSTOUFS(oldmnt);
+			UFS_LOCK(ump);
+			for (i = 0; i < MAXQUOTAS; i++) {
+				if (ump->um_quotas[i] != NULLVP)
+					morework = 1;
+			}
+			UFS_UNLOCK(ump);
+#endif
+			if (morework) {
 				if (--retry_flush_count > 0) {
 					retry = 1;
 					loopcnt = 3;
diff --git a/sys/ufs/ffs/ffs_vfsops.c b/sys/ufs/ffs/ffs_vfsops.c
index 0204613..b3292d0 100644
--- a/sys/ufs/ffs/ffs_vfsops.c
+++ b/sys/ufs/ffs/ffs_vfsops.c
@@ -1351,9 +1351,10 @@ ffs_flushfiles(mp, flags, td)
 	struct thread *td;
 {
 	struct ufsmount *ump;
-	int error;
+	int qerror, error;
 
 	ump = VFSTOUFS(mp);
+	qerror = 0;
 #ifdef QUOTA
 	if (mp->mnt_flag & MNT_QUOTA) {
 		int i;
@@ -1361,11 +1362,19 @@ ffs_flushfiles(mp, flags, td)
 		if (error)
 			return (error);
 		for (i = 0; i < MAXQUOTAS; i++) {
-			quotaoff(td, mp, i);
+			error = quotaoff(td, mp, i);
+			if (error != 0) {
+				if ((flags & EARLYFLUSH) == 0)
+					return (error);
+				else
+					qerror = error;
+			}
 		}
+
 		/*
-		 * Here we fall through to vflush again to ensure
-		 * that we have gotten rid of all the system vnodes.
+		 * Here we fall through to vflush again to ensure that
+		 * we have gotten rid of all the system vnodes, unless
+		 * quotas must not be closed.
 		 */
 	}
 #endif
@@ -1380,11 +1389,21 @@ ffs_flushfiles(mp, flags, td)
 		 * that we have gotten rid of all the system vnodes.
 		 */
 	}
-        /*
-	 * Flush all the files.
+
+	/*
+	 * Do not close system files if quotas were not closed, to be
+	 * able to sync the remaining dquots.  The freeblks softupdate
+	 * workitems might hold a reference on a dquot, preventing
+	 * quotaoff() from completing.  Next round of
+	 * softdep_flushworklist() iteration should process the
+	 * blockers, allowing the next run of quotaoff() to finally
+	 * flush held dquots.
+	 *
+	 * Otherwise, flush all the files.
 	 */
-	if ((error = vflush(mp, 0, flags, td)) != 0)
+	if (qerror == 0 && (error = vflush(mp, 0, flags, td)) != 0)
 		return (error);
+
 	/*
 	 * Flush filesystem metadata.
 	 */
diff --git a/sys/ufs/ufs/ufs_quota.c b/sys/ufs/ufs/ufs_quota.c
index 87ac9a1..a949898 100644
--- a/sys/ufs/ufs/ufs_quota.c
+++ b/sys/ufs/ufs/ufs_quota.c
@@ -80,7 +80,7 @@ static int dqopen(struct vnode *, struct ufsmount *, int);
 static int dqget(struct vnode *,
 	u_long, struct ufsmount *, int, struct dquot **);
 static int dqsync(struct vnode *, struct dquot *);
-static void dqflush(struct vnode *);
+static int dqflush(struct vnode *);
 static int quotaoff1(struct thread *td, struct mount *mp, int type);
 static int quotaoff_inchange(struct thread *td, struct mount *mp, int type);
 
@@ -674,8 +674,12 @@ again:
 		vrele(vp);
 	}
 
-	dqflush(qvp);
-	/* Clear um_quotas before closing the quota vnode to prevent
+	error = dqflush(qvp);
+	if (error != 0)
+		return (error);
+
+	/*
+	 * Clear um_quotas before closing the quota vnode to prevent
 	 * access to the closed vnode from dqget/dqsync
 	 */
 	UFS_LOCK(ump);
@@ -1594,17 +1598,19 @@ out:
 /*
  * Flush all entries from the cache for a particular vnode.
  */
-static void
+static int
 dqflush(struct vnode *vp)
 {
 	struct dquot *dq, *nextdq;
 	struct dqhash *dqh;
+	int error;
 
 	/*
 	 * Move all dquot's that used to refer to this quota
 	 * file off their hash chains (they will eventually
 	 * fall off the head of the free list and be re-used).
 	 */
+	error = 0;
 	DQH_LOCK();
 	for (dqh = &dqhashtbl[dqhash]; dqh >= dqhashtbl; dqh--) {
 		for (dq = LIST_FIRST(dqh); dq; dq = nextdq) {
@@ -1612,12 +1618,15 @@ dqflush(struct vnode *vp)
 			if (dq->dq_ump->um_quotas[dq->dq_type] != vp)
 				continue;
 			if (dq->dq_cnt)
-				panic("dqflush: stray dquot");
-			LIST_REMOVE(dq, dq_hash);
-			dq->dq_ump = (struct ufsmount *)0;
+				error = EBUSY;
+			else {
+				LIST_REMOVE(dq, dq_hash);
+				dq->dq_ump = NULL;
+			}
 		}
 	}
 	DQH_UNLOCK();
+	return (error);
 }
 
 /*