Btrfs: Throttle operations if the reference cache gets too large

A large reference cache is directly related to a lot of work pending
for the cleaner thread.  This throttles back new operations based on
the size of the reference cache so the cleaner thread will be able to keep
up.

Overall, this actually makes the FS faster because the cleaner thread will
be more likely to find things in cache.
Signed-off-by: Chris Mason <chris.mason@oracle.com>

fs/btrfs/ctree.h

@@ -580,6 +580,7 @@ struct btrfs_fs_info {
 	int do_barriers;
 	int closing;
 	atomic_t throttles;
+	atomic_t throttle_gen;
 
 	u64 total_pinned;
 	struct list_head dirty_cowonly_roots;

fs/btrfs/disk-io.c

@@ -1171,8 +1171,10 @@ static int transaction_kthread(void *arg)
 		vfs_check_frozen(root->fs_info->sb, SB_FREEZE_WRITE);
 		mutex_lock(&root->fs_info->transaction_kthread_mutex);
 
+		if (root->fs_info->total_ref_cache_size > 20 * 1024 * 1024) {
 			printk("btrfs: total reference cache size %Lu\n",
 				root->fs_info->total_ref_cache_size);
+		}
 
 		mutex_lock(&root->fs_info->trans_mutex);
 		cur = root->fs_info->running_transaction;
@@ -1256,6 +1258,7 @@ struct btrfs_root *open_ctree(struct super_block *sb,
 	btrfs_mapping_init(&fs_info->mapping_tree);
 	atomic_set(&fs_info->nr_async_submits, 0);
 	atomic_set(&fs_info->throttles, 0);
+	atomic_set(&fs_info->throttle_gen, 0);
 	fs_info->sb = sb;
 	fs_info->max_extent = (u64)-1;
 	fs_info->max_inline = 8192 * 1024;

fs/btrfs/extent-tree.c

@@ -2650,6 +2650,7 @@ int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
 		}
 	}
 	while(1) {
+		atomic_inc(&root->fs_info->throttle_gen);
 		wret = walk_down_tree(trans, root, path, &level);
 		if (wret > 0)
 			break;

fs/btrfs/file.c

@@ -974,6 +974,7 @@ static ssize_t btrfs_file_write(struct file *file, const char __user *buf,
 		balance_dirty_pages_ratelimited_nr(inode->i_mapping, num_pages);
 		if (num_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
 			btrfs_btree_balance_dirty(root, 1);
+		btrfs_throttle(root);
 		cond_resched();
 	}
 out:

fs/btrfs/inode.c

@@ -2482,7 +2482,7 @@ static int btrfs_create(struct inode *dir, struct dentry *dentry,
 	btrfs_update_inode_block_group(trans, dir);
 out_unlock:
 	nr = trans->blocks_used;
-	btrfs_end_transaction(trans, root);
+	btrfs_end_transaction_throttle(trans, root);
 fail:
 	if (drop_inode) {
 		inode_dec_link_count(inode);
@@ -2535,7 +2535,7 @@ static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
 		drop_inode = 1;
 
 	nr = trans->blocks_used;
-	btrfs_end_transaction(trans, root);
+	btrfs_end_transaction_throttle(trans, root);
 fail:
 	if (drop_inode) {
 		inode_dec_link_count(inode);
@@ -2609,7 +2609,7 @@ static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, int mode)
 
 out_fail:
 	nr = trans->blocks_used;
-	btrfs_end_transaction(trans, root);
+	btrfs_end_transaction_throttle(trans, root);
 
 out_unlock:
 	if (drop_on_err)
@@ -3434,7 +3434,7 @@ static int btrfs_rename(struct inode * old_dir, struct dentry *old_dentry,
 		goto out_fail;
 
 out_fail:
-	btrfs_end_transaction(trans, root);
+	btrfs_end_transaction_throttle(trans, root);
 out_unlock:
 	return ret;
 }
@@ -3548,7 +3548,7 @@ static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
 
 out_unlock:
 	nr = trans->blocks_used;
-	btrfs_end_transaction(trans, root);
+	btrfs_end_transaction_throttle(trans, root);
 out_fail:
 	if (drop_inode) {
 		inode_dec_link_count(inode);

fs/btrfs/transaction.c

@@ -202,35 +202,64 @@ static noinline int wait_for_commit(struct btrfs_root *root,
 	return 0;
 }
 
+void btrfs_throttle(struct btrfs_root *root)
+{
+	struct btrfs_fs_info *info = root->fs_info;
+
+harder:
+	if (atomic_read(&info->throttles)) {
+		DEFINE_WAIT(wait);
+		int thr;
+		int harder_count = 0;
+		thr = atomic_read(&info->throttle_gen);
+
+		do {
+			prepare_to_wait(&info->transaction_throttle,
+					&wait, TASK_UNINTERRUPTIBLE);
+			if (!atomic_read(&info->throttles)) {
+				finish_wait(&info->transaction_throttle, &wait);
+				break;
+			}
+			schedule();
+			finish_wait(&info->transaction_throttle, &wait);
+		} while (thr == atomic_read(&info->throttle_gen));
+
+		if (harder_count < 5 &&
+		    info->total_ref_cache_size > 5 * 1024 * 1024) {
+			harder_count++;
+			goto harder;
+		}
+
+		if (harder_count < 10 &&
+		    info->total_ref_cache_size > 10 * 1024 * 1024) {
+			harder_count++;
+			goto harder;
+		}
+	}
+}
+
 static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
 			  struct btrfs_root *root, int throttle)
 {
 	struct btrfs_transaction *cur_trans;
+	struct btrfs_fs_info *info = root->fs_info;
 
-	mutex_lock(&root->fs_info->trans_mutex);
-	cur_trans = root->fs_info->running_transaction;
+	mutex_lock(&info->trans_mutex);
+	cur_trans = info->running_transaction;
 	WARN_ON(cur_trans != trans->transaction);
 	WARN_ON(cur_trans->num_writers < 1);
 	cur_trans->num_writers--;
 
 	if (waitqueue_active(&cur_trans->writer_wait))
 		wake_up(&cur_trans->writer_wait);
-
-	if (throttle && atomic_read(&root->fs_info->throttles)) {
-		DEFINE_WAIT(wait);
-		mutex_unlock(&root->fs_info->trans_mutex);
-		prepare_to_wait(&root->fs_info->transaction_throttle, &wait,
-				TASK_UNINTERRUPTIBLE);
-		if (atomic_read(&root->fs_info->throttles))
-			schedule();
-		finish_wait(&root->fs_info->transaction_throttle, &wait);
-		mutex_lock(&root->fs_info->trans_mutex);
-	}
-
 	put_transaction(cur_trans);
-	mutex_unlock(&root->fs_info->trans_mutex);
+	mutex_unlock(&info->trans_mutex);
 	memset(trans, 0, sizeof(*trans));
 	kmem_cache_free(btrfs_trans_handle_cachep, trans);
+
+	if (throttle)
+		btrfs_throttle(root);
+
 	return 0;
 }
 

fs/btrfs/transaction.h

@@ -91,4 +91,5 @@ int btrfs_commit_transaction(struct btrfs_trans_handle *trans,
 			     struct btrfs_root *root);
 int btrfs_end_transaction_throttle(struct btrfs_trans_handle *trans,
 				   struct btrfs_root *root);
+void btrfs_throttle(struct btrfs_root *root);
 #endif