Fix the "dump corrupts filesystems" buffer-cache bug.

Since block-dev-in-page-cache was added in 2.4.10, and the aliasing between
buffered devices and getblk() was restored in 2.4.11, getblk() and friends
can be performing IO on a page cache page without the page cache lock
being held.  The page cache IO assumes that the page cache lock is enough
to synchronise things safely, but this breaks if we get aliased IO.

In particular, block_read_full_page() assumes that it is safe to begin IO
on any non-uptodate bh, regardless of the locked status of the bh.  To be
safe, we need to test the uptodate state *after* taking the bh lock.

Already in 2.5.

--- linux-2.4.19-ext3/fs/buffer.c.=K0004=.orig	Fri Oct 11 15:48:39 2002
+++ linux-2.4.19-ext3/fs/buffer.c	Fri Oct 11 15:52:01 2002
@@ -1754,9 +1754,14 @@
 	}
 
 	/* Stage 3: start the IO */
-	for (i = 0; i < nr; i++)
-		submit_bh(READ, arr[i]);
-
+	for (i = 0; i < nr; i++) {
+		struct buffer_head * bh = arr[i];
+		if (buffer_uptodate(bh))
+			end_buffer_io_async(bh, 1);
+		else
+			submit_bh(READ, bh);
+	}
+	
 	return 0;
 }