--- /opt/kernel/linux-2.4.0-loop/drivers/block/loop.c	Mon Jan  8 01:31:27 2001
+++ linux/drivers/block/loop.c	Mon Jan  8 02:40:37 2001
@@ -292,9 +292,11 @@
 {
 	struct loop_device *lo = &loop_dev[MINOR(bh->b_dev)];
 	struct buffer_head *rbh = bh->b_private;
-	int rw = test_and_clear_bit(BH_Dirty, &bh->b_state);
 
-	if (rw == READ && uptodate)
+	/*
+	 * only need to do transfer on (uptodate) READ
+	 */
+	if (!test_bit(BH_Dirty, &bh->b_state) && uptodate)
 		uptodate = !lo_do_transfer(lo, READ, bh->b_data, rbh->b_data,
 					   bh->b_size, bh->b_blocknr);
 
@@ -309,8 +311,8 @@
 
 	/*
 	 * any better ideas? one could argue that since
-	 * (BUF_BUFFER & __GFP_WAIT) kmem_cache_alloc should
-	 * always return an object
+	 *	(GFP_BUFFER & __GFP_WAIT)
+	 * kmem_cache_alloc should always return an object
 	 */
 	do {
 		if ((bh = kmem_cache_alloc(loop_bhp, SLAB_BUFFER)))
@@ -321,6 +323,7 @@
 	memset(bh, 0, sizeof(*bh));
 
 	bh->b_size = rbh->b_size;
+	bh->b_dev = rbh->b_dev;
 	bh->b_rdev = lo->lo_device;
 	bh->b_state = (1 << BH_Req) | (1 << BH_Mapped) | (1 << BH_Lock);