diff --git a/fs/btrfs/btrfs_inode.h b/fs/btrfs/btrfs_inode.h
index 2411baf35220..2a8c242bc4f5 100644
--- a/fs/btrfs/btrfs_inode.h
+++ b/fs/btrfs/btrfs_inode.h
@@ -91,6 +91,9 @@ struct btrfs_inode {
 
 	unsigned long runtime_flags;
 
+	/* Keep track of who's O_SYNC/fsycing currently */
+	atomic_t sync_writers;
+
 	/* full 64 bit generation number, struct vfs_inode doesn't have a big
 	 * enough field for this.
 	 */
diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
index 71c2dc1ea15a..7f4654a15207 100644
--- a/fs/btrfs/file.c
+++ b/fs/btrfs/file.c
@@ -1472,6 +1472,7 @@ static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
 	ssize_t num_written = 0;
 	ssize_t err = 0;
 	size_t count, ocount;
+	bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
 
 	sb_start_write(inode->i_sb);
 
@@ -1529,6 +1530,9 @@ static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
 		}
 	}
 
+	if (sync)
+		atomic_inc(&BTRFS_I(inode)->sync_writers);
+
 	if (unlikely(file->f_flags & O_DIRECT)) {
 		num_written = __btrfs_direct_write(iocb, iov, nr_segs,
 						   pos, ppos, count, ocount);
@@ -1563,6 +1567,8 @@ static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
 			num_written = err;
 	}
 out:
+	if (sync)
+		atomic_dec(&BTRFS_I(inode)->sync_writers);
 	sb_end_write(inode->i_sb);
 	current->backing_dev_info = NULL;
 	return num_written ? num_written : err;
@@ -1613,7 +1619,9 @@ int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
 	 * out of the ->i_mutex. If so, we can flush the dirty pages by
 	 * multi-task, and make the performance up.
 	 */
+	atomic_inc(&BTRFS_I(inode)->sync_writers);
 	ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
+	atomic_dec(&BTRFS_I(inode)->sync_writers);
 	if (ret)
 		return ret;
 
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 123815f3b454..7855aac36706 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -1622,6 +1622,7 @@ static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
 	int ret = 0;
 	int skip_sum;
 	int metadata = 0;
+	int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
 
 	skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
 
@@ -1644,7 +1645,7 @@ static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
 				goto out;
 		}
 		goto mapit;
-	} else if (!skip_sum) {
+	} else if (async && !skip_sum) {
 		/* csum items have already been cloned */
 		if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
 			goto mapit;
@@ -1655,6 +1656,10 @@ static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
 				   __btrfs_submit_bio_start,
 				   __btrfs_submit_bio_done);
 		goto out;
+	} else if (!skip_sum) {
+		ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
+		if (ret)
+			goto out;
 	}
 
 mapit:
@@ -6333,6 +6338,9 @@ static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
 	struct btrfs_root *root = BTRFS_I(inode)->root;
 	int ret;
 
+	if (async_submit)
+		async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
+
 	bio_get(bio);
 
 	if (!write) {
@@ -7113,6 +7121,7 @@ struct inode *btrfs_alloc_inode(struct super_block *sb)
 	extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
 	ei->io_tree.track_uptodate = 1;
 	ei->io_failure_tree.track_uptodate = 1;
+	atomic_set(&ei->sync_writers, 0);
 	mutex_init(&ei->log_mutex);
 	mutex_init(&ei->delalloc_mutex);
 	btrfs_ordered_inode_tree_init(&ei->ordered_tree);