diff --git a/fs/direct-io.c b/fs/direct-io.c
index 8b31b9f..1782023 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -266,8 +266,18 @@
 		dio->end_io(dio->iocb, offset, transferred, dio->private);
 
 	inode_dio_done(dio->inode);
-	if (is_async)
+	if (is_async) {
+		if (dio->rw & WRITE) {
+			int err;
+
+			err = generic_write_sync(dio->iocb->ki_filp, offset,
+						 transferred);
+			if (err < 0 && ret > 0)
+				ret = err;
+		}
+
 		aio_complete(dio->iocb, ret, 0);
+	}
 
 	kmem_cache_free(dio_cache, dio);
 	return ret;
@@ -1183,11 +1193,6 @@
 	}
 
 	/*
-	 * Will be decremented at I/O completion time.
-	 */
-	atomic_inc(&inode->i_dio_count);
-
-	/*
 	 * For file extending writes updating i_size before data
 	 * writeouts complete can expose uninitialized blocks. So
 	 * even for AIO, we need to wait for i/o to complete before
@@ -1195,11 +1200,33 @@
 	 */
 	dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
 		(end > i_size_read(inode)));
-
-	retval = 0;
-
 	dio->inode = inode;
 	dio->rw = rw;
+
+	/*
+	 * For AIO O_(D)SYNC writes we need to defer completions to a workqueue
+	 * so that we can call ->fsync.
+	 */
+	if (dio->is_async && (rw & WRITE) &&
+	    ((iocb->ki_filp->f_flags & O_DSYNC) ||
+	     IS_SYNC(iocb->ki_filp->f_mapping->host))) {
+		retval = dio_set_defer_completion(dio);
+		if (retval) {
+			/*
+			 * We grab i_mutex only for reads so we don't have
+			 * to release it here
+			 */
+			kmem_cache_free(dio_cache, dio);
+			goto out;
+		}
+	}
+
+	/*
+	 * Will be decremented at I/O completion time.
+	 */
+	atomic_inc(&inode->i_dio_count);
+
+	retval = 0;
 	sdio.blkbits = blkbits;
 	sdio.blkfactor = i_blkbits - blkbits;
 	sdio.block_in_file = offset >> blkbits;