[PATCH] O(1) sb list traversing on syncs

This patch removes O(n^2) super block loops in sync_inodes(),
sync_filesystems() etc.  in favour of using __put_super_and_need_restart()
which I introduced earlier.  We faced a noticably long freezes on sb
syncing when there are thousands of super blocks in the system.

Signed-Off-By: Kirill Korotaev <dev@sw.ru>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

fs/fs-writeback.c

@@ -485,32 +485,6 @@ static void set_sb_syncing(int val)
 	spin_unlock(&sb_lock);
 }
 
-/*
- * Find a superblock with inodes that need to be synced
- */
-static struct super_block *get_super_to_sync(void)
-{
-	struct super_block *sb;
-restart:
-	spin_lock(&sb_lock);
-	sb = sb_entry(super_blocks.prev);
-	for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) {
-		if (sb->s_syncing)
-			continue;
-		sb->s_syncing = 1;
-		sb->s_count++;
-		spin_unlock(&sb_lock);
-		down_read(&sb->s_umount);
-		if (!sb->s_root) {
-			drop_super(sb);
-			goto restart;
-		}
-		return sb;
-	}
-	spin_unlock(&sb_lock);
-	return NULL;
-}
-
 /**
  * sync_inodes - writes all inodes to disk
  * @wait: wait for completion
@@ -530,23 +504,39 @@ restart:
  * outstanding dirty inodes, the writeback goes block-at-a-time within the
  * filesystem's write_inode().  This is extremely slow.
  */
-void sync_inodes(int wait)
+static void __sync_inodes(int wait)
 {
 	struct super_block *sb;
 
-	set_sb_syncing(0);
+	spin_lock(&sb_lock);
-	while ((sb = get_super_to_sync()) != NULL) {
+restart:
-		sync_inodes_sb(sb, 0);
+	list_for_each_entry(sb, &super_blocks, s_list) {
-		sync_blockdev(sb->s_bdev);
+		if (sb->s_syncing)
-		drop_super(sb);
+			continue;
+		sb->s_syncing = 1;
+		sb->s_count++;
+		spin_unlock(&sb_lock);
+		down_read(&sb->s_umount);
+		if (sb->s_root) {
+			sync_inodes_sb(sb, wait);
+			sync_blockdev(sb->s_bdev);
+		}
+		up_read(&sb->s_umount);
+		spin_lock(&sb_lock);
+		if (__put_super_and_need_restart(sb))
+			goto restart;
 	}
+	spin_unlock(&sb_lock);
+}
+
+void sync_inodes(int wait)
+{
+	set_sb_syncing(0);
+	__sync_inodes(0);
+
 	if (wait) {
 		set_sb_syncing(0);
-		while ((sb = get_super_to_sync()) != NULL) {
+		__sync_inodes(1);
-			sync_inodes_sb(sb, 1);
-			sync_blockdev(sb->s_bdev);
-			drop_super(sb);
-		}
 	}
 }
 

fs/quota.c

@@ -149,36 +149,6 @@ static int check_quotactl_valid(struct super_block *sb, int type, int cmd, qid_t
 	return error;
 }
 
-static struct super_block *get_super_to_sync(int type)
-{
-	struct list_head *head;
-	int cnt, dirty;
-
-restart:
-	spin_lock(&sb_lock);
-	list_for_each(head, &super_blocks) {
-		struct super_block *sb = list_entry(head, struct super_block, s_list);
-
-		/* This test just improves performance so it needn't be reliable... */
-		for (cnt = 0, dirty = 0; cnt < MAXQUOTAS; cnt++)
-			if ((type == cnt || type == -1) && sb_has_quota_enabled(sb, cnt)
-			    && info_any_dirty(&sb_dqopt(sb)->info[cnt]))
-				dirty = 1;
-		if (!dirty)
-			continue;
-		sb->s_count++;
-		spin_unlock(&sb_lock);
-		down_read(&sb->s_umount);
-		if (!sb->s_root) {
-			drop_super(sb);
-			goto restart;
-		}
-		return sb;
-	}
-	spin_unlock(&sb_lock);
-	return NULL;
-}
-
 static void quota_sync_sb(struct super_block *sb, int type)
 {
 	int cnt;
@@ -219,17 +189,35 @@ static void quota_sync_sb(struct super_block *sb, int type)
 
 void sync_dquots(struct super_block *sb, int type)
 {
+	int cnt, dirty;
+
 	if (sb) {
 		if (sb->s_qcop->quota_sync)
 			quota_sync_sb(sb, type);
+		return;
 	}
-	else {
+
-		while ((sb = get_super_to_sync(type)) != NULL) {
+	spin_lock(&sb_lock);
-			if (sb->s_qcop->quota_sync)
+restart:
-				quota_sync_sb(sb, type);
+	list_for_each_entry(sb, &super_blocks, s_list) {
-			drop_super(sb);
+		/* This test just improves performance so it needn't be reliable... */
-		}
+		for (cnt = 0, dirty = 0; cnt < MAXQUOTAS; cnt++)
+			if ((type == cnt || type == -1) && sb_has_quota_enabled(sb, cnt)
+			    && info_any_dirty(&sb_dqopt(sb)->info[cnt]))
+				dirty = 1;
+		if (!dirty)
+			continue;
+		sb->s_count++;
+		spin_unlock(&sb_lock);
+		down_read(&sb->s_umount);
+		if (sb->s_root && sb->s_qcop->quota_sync)
+			quota_sync_sb(sb, type);
+		up_read(&sb->s_umount);
+		spin_lock(&sb_lock);
+		if (__put_super_and_need_restart(sb))
+			goto restart;
 	}
+	spin_unlock(&sb_lock);
 }
 
 /* Copy parameters and call proper function */

fs/super.c

@@ -341,20 +341,22 @@ static inline void write_super(struct super_block *sb)
  */
 void sync_supers(void)
 {
-	struct super_block * sb;
+	struct super_block *sb;
-restart:
+
 	spin_lock(&sb_lock);
-	sb = sb_entry(super_blocks.next);
+restart:
-	while (sb != sb_entry(&super_blocks))
+	list_for_each_entry(sb, &super_blocks, s_list) {
 		if (sb->s_dirt) {
 			sb->s_count++;
 			spin_unlock(&sb_lock);
 			down_read(&sb->s_umount);
 			write_super(sb);
-			drop_super(sb);
+			up_read(&sb->s_umount);
-			goto restart;
+			spin_lock(&sb_lock);
-		} else
+			if (__put_super_and_need_restart(sb))
-			sb = sb_entry(sb->s_list.next);
+				goto restart;
+		}
+	}
 	spin_unlock(&sb_lock);
 }
 
@@ -381,20 +383,16 @@ void sync_filesystems(int wait)
 
 	down(&mutex);		/* Could be down_interruptible */
 	spin_lock(&sb_lock);
-	for (sb = sb_entry(super_blocks.next); sb != sb_entry(&super_blocks);
+	list_for_each_entry(sb, &super_blocks, s_list) {
-			sb = sb_entry(sb->s_list.next)) {
 		if (!sb->s_op->sync_fs)
 			continue;
 		if (sb->s_flags & MS_RDONLY)
 			continue;
 		sb->s_need_sync_fs = 1;
 	}
-	spin_unlock(&sb_lock);
 
 restart:
-	spin_lock(&sb_lock);
+	list_for_each_entry(sb, &super_blocks, s_list) {
-	for (sb = sb_entry(super_blocks.next); sb != sb_entry(&super_blocks);
-			sb = sb_entry(sb->s_list.next)) {
 		if (!sb->s_need_sync_fs)
 			continue;
 		sb->s_need_sync_fs = 0;
@@ -405,8 +403,11 @@ restart:
 		down_read(&sb->s_umount);
 		if (sb->s_root && (wait || sb->s_dirt))
 			sb->s_op->sync_fs(sb, wait);
-		drop_super(sb);
+		up_read(&sb->s_umount);
-		goto restart;
+		/* restart only when sb is no longer on the list */
+		spin_lock(&sb_lock);
+		if (__put_super_and_need_restart(sb))
+			goto restart;
 	}
 	spin_unlock(&sb_lock);
 	up(&mutex);
@@ -422,21 +423,25 @@ restart:
 
 struct super_block * get_super(struct block_device *bdev)
 {
-	struct list_head *p;
+	struct super_block *sb;
+
 	if (!bdev)
 		return NULL;
-rescan:
+
 	spin_lock(&sb_lock);
-	list_for_each(p, &super_blocks) {
+rescan:
-		struct super_block *s = sb_entry(p);
+	list_for_each_entry(sb, &super_blocks, s_list) {
-		if (s->s_bdev == bdev) {
+		if (sb->s_bdev == bdev) {
-			s->s_count++;
+			sb->s_count++;
 			spin_unlock(&sb_lock);
-			down_read(&s->s_umount);
+			down_read(&sb->s_umount);
-			if (s->s_root)
+			if (sb->s_root)
-				return s;
+				return sb;
-			drop_super(s);
+			up_read(&sb->s_umount);
-			goto rescan;
+			/* restart only when sb is no longer on the list */
+			spin_lock(&sb_lock);
+			if (__put_super_and_need_restart(sb))
+				goto rescan;
 		}
 	}
 	spin_unlock(&sb_lock);
@@ -447,20 +452,22 @@ EXPORT_SYMBOL(get_super);
  
 struct super_block * user_get_super(dev_t dev)
 {
-	struct list_head *p;
+	struct super_block *sb;
 
-rescan:
 	spin_lock(&sb_lock);
-	list_for_each(p, &super_blocks) {
+rescan:
-		struct super_block *s = sb_entry(p);
+	list_for_each_entry(sb, &super_blocks, s_list) {
-		if (s->s_dev ==  dev) {
+		if (sb->s_dev ==  dev) {
-			s->s_count++;
+			sb->s_count++;
 			spin_unlock(&sb_lock);
-			down_read(&s->s_umount);
+			down_read(&sb->s_umount);
-			if (s->s_root)
+			if (sb->s_root)
-				return s;
+				return sb;
-			drop_super(s);
+			up_read(&sb->s_umount);
-			goto rescan;
+			/* restart only when sb is no longer on the list */
+			spin_lock(&sb_lock);
+			if (__put_super_and_need_restart(sb))
+				goto rescan;
 		}
 	}
 	spin_unlock(&sb_lock);