Patch: aliasing bug in blockdev-in-pagecache?

Hi all,

I think there's a data-corruption possible in both ext2 and ext3 (and in fact any filesystem which uses bread) if user space is reading from the filesystem's buffered block device during live fs activity.

I've recently been seeing an ext3 assert failure where the filesystem comes across a buffer which is unexpectedly locked, while running a dump(8) on the live filesystem. Now, I do _not_ want to start another debate about whether it's sane to do that or not! But the mounted kernel filesystem should still survive if the user is only reading from the buffered device.

The problem turns out to be in block_read_full_page(). The page cache IO code assumes that IO is synchronised at the page level, but any kernel code using getblk() (or sb_read() or related functions) to access the buffers at the same time will bypass the page locking.

So if block_read_full_page() encounters a page which bread() is already reading into cache, it will see the page unlocked but the buffer_head locked and !uptodate. However, it ignores the bh lock, seeing only that the buffer is !uptodate, and so it then locks the bh itself and submits a read IO.

Unfortunately, if the bread() wins the race for the wait_on_buffer after the initial IO, we can already have started to modify the buffer contents by the time that block_read_full_page() starts the new IO. So we read stale contents from disk on top of the modified contents in cache.

To solve this, we really do need to have block_read_full_page() test the uptodate state under protection of the buffer_head lock. We already go through 3 stages in block_read_full_page(): gather the buffers needing IO, then lock them, then submit the IO. To be safe, we need a final test for buffer_uptodate() *after* we have locked the required buffers.

I've verified that the scenario above is definitely happening, and I'm currently testing the patch below. I'm using 2.4 for the testing right now, but 2.5 seems to have the same problem at first glance.

Comments?

--Stephen

@@ -1902,9 +1902,14 @@
 	}
 
 	/* 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;
 }
 

Ouch.

I suspect that correct fix is to do that test in submit_bh() itself (and remove it from ll_rw_block()). IMO it's cleaner than messing with all callers out there... Linus?

Hi,

Actually, if we move the test to submit_bh(), we _do_ need to mess with all callers.

submit_bh() is currently an unconditional demand to perform a given IO regardless of buffer state, so lots of callers (ext3 journal writes, soft raid mirror writes etc) call it with private buffer_heads which just don't have any persistent dirty / uptodate state. Changing submit_bh() still means we need to audit all callers to make sure that they set up those state flags correctly even for private bh'es.

Cheers,
Stephen