diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c
index 92e14dd9bddbe673d582b90202e2135f0fc02e66..bbba14b629d217d2f96e19707bb6e9f3eb82c6f1 100644
--- a/fs/btrfs/disk-io.c
+++ b/fs/btrfs/disk-io.c
@@ -429,7 +429,7 @@ int btrfs_bio_wq_end_io(struct btrfs_fs_info *info, struct bio *bio,
 	return 0;
 }
 
-static unsigned long async_submit_limit(struct btrfs_fs_info *info)
+unsigned long btrfs_async_submit_limit(struct btrfs_fs_info *info)
 {
 	unsigned long limit = min_t(unsigned long,
 				    info->workers.max_workers,
@@ -439,7 +439,8 @@ static unsigned long async_submit_limit(struct btrfs_fs_info *info)
 
 int btrfs_congested_async(struct btrfs_fs_info *info, int iodone)
 {
-	return atomic_read(&info->nr_async_bios) > async_submit_limit(info);
+	return atomic_read(&info->nr_async_bios) >
+		btrfs_async_submit_limit(info);
 }
 
 static void run_one_async_submit(struct btrfs_work *work)
@@ -451,12 +452,13 @@ static void run_one_async_submit(struct btrfs_work *work)
 	async = container_of(work, struct  async_submit_bio, work);
 	fs_info = BTRFS_I(async->inode)->root->fs_info;
 
-	limit = async_submit_limit(fs_info);
+	limit = btrfs_async_submit_limit(fs_info);
 	limit = limit * 2 / 3;
 
 	atomic_dec(&fs_info->nr_async_submits);
 
-	if (atomic_read(&fs_info->nr_async_submits) < limit)
+	if (atomic_read(&fs_info->nr_async_submits) < limit &&
+	    waitqueue_active(&fs_info->async_submit_wait))
 		wake_up(&fs_info->async_submit_wait);
 
 	async->submit_bio_hook(async->inode, async->rw, async->bio,
@@ -469,7 +471,7 @@ int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode,
 			extent_submit_bio_hook_t *submit_bio_hook)
 {
 	struct async_submit_bio *async;
-	int limit = async_submit_limit(fs_info);
+	int limit = btrfs_async_submit_limit(fs_info);
 
 	async = kmalloc(sizeof(*async), GFP_NOFS);
 	if (!async)
@@ -1863,10 +1865,10 @@ void btrfs_btree_balance_dirty(struct btrfs_root *root, unsigned long nr)
 	struct extent_io_tree *tree;
 	u64 num_dirty;
 	u64 start = 0;
-	unsigned long thresh = 12 * 1024 * 1024;
+	unsigned long thresh = 96 * 1024 * 1024;
 	tree = &BTRFS_I(root->fs_info->btree_inode)->io_tree;
 
-	if (current_is_pdflush())
+	if (current_is_pdflush() || current->flags & PF_MEMALLOC)
 		return;
 
 	num_dirty = count_range_bits(tree, &start, (u64)-1,
diff --git a/fs/btrfs/disk-io.h b/fs/btrfs/disk-io.h
index e904a69347a41e5b22cf1b308014d10a1299cd42..2562a273ae18c9c9298a59c3e31c987d764c0eb2 100644
--- a/fs/btrfs/disk-io.h
+++ b/fs/btrfs/disk-io.h
@@ -73,4 +73,5 @@ int btrfs_wq_submit_bio(struct btrfs_fs_info *fs_info, struct inode *inode,
 			int rw, struct bio *bio, int mirror_num,
 			extent_submit_bio_hook_t *submit_bio_hook);
 int btrfs_congested_async(struct btrfs_fs_info *info, int iodone);
+unsigned long btrfs_async_submit_limit(struct btrfs_fs_info *info);
 #endif
diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index 6bcb0876f9bbca78a0c009a81c803ce9f2b27401..eff3ad72991b264f4844946f50f8dba77ae0d117 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c
@@ -322,8 +322,6 @@ int btrfs_write_and_wait_transaction(struct btrfs_trans_handle *trans,
 		if (ret)
 			break;
 		while(start <= end) {
-			if (btrfs_congested_async(root->fs_info, 0))
-				congestion_wait(WRITE, HZ/10);
 			cond_resched();
 
 			index = start >> PAGE_CACHE_SHIFT;
diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
index 23a5b0aba00a51c728133d4344bb52db3a5f636d..2652660e6079be6b8da1424989636d13518a27c7 100644
--- a/fs/btrfs/volumes.c
+++ b/fs/btrfs/volumes.c
@@ -138,12 +138,18 @@ int run_scheduled_bios(struct btrfs_device *device)
 {
 	struct bio *pending;
 	struct backing_dev_info *bdi;
+	struct btrfs_fs_info *fs_info;
 	struct bio *tail;
 	struct bio *cur;
 	int again = 0;
 	unsigned long num_run = 0;
+	unsigned long limit;
 
 	bdi = device->bdev->bd_inode->i_mapping->backing_dev_info;
+	fs_info = device->dev_root->fs_info;
+	limit = btrfs_async_submit_limit(fs_info);
+	limit = limit * 2 / 3;
+
 loop:
 	spin_lock(&device->io_lock);
 
@@ -179,7 +185,11 @@ int run_scheduled_bios(struct btrfs_device *device)
 		cur = pending;
 		pending = pending->bi_next;
 		cur->bi_next = NULL;
-		atomic_dec(&device->dev_root->fs_info->nr_async_bios);
+		atomic_dec(&fs_info->nr_async_bios);
+
+		if (atomic_read(&fs_info->nr_async_bios) < limit &&
+		    waitqueue_active(&fs_info->async_submit_wait))
+			wake_up(&fs_info->async_submit_wait);
 
 		BUG_ON(atomic_read(&cur->bi_cnt) == 0);
 		bio_get(cur);
@@ -2135,6 +2145,7 @@ int schedule_bio(struct btrfs_root *root, struct btrfs_device *device,
 		 int rw, struct bio *bio)
 {
 	int should_queue = 1;
+	unsigned long limit;
 
 	/* don't bother with additional async steps for reads, right now */
 	if (!(rw & (1 << BIO_RW))) {
@@ -2171,6 +2182,11 @@ int schedule_bio(struct btrfs_root *root, struct btrfs_device *device,
 	if (should_queue)
 		btrfs_queue_worker(&root->fs_info->submit_workers,
 				   &device->work);
+
+	limit = btrfs_async_submit_limit(root->fs_info);
+	wait_event_timeout(root->fs_info->async_submit_wait,
+			   (atomic_read(&root->fs_info->nr_async_bios) < limit),
+			   HZ/10);
 	return 0;
 }