Comment 17 for bug 624877

Does the problem go away in 2.6.37? There's an ext4 commit that may be relevant, which hit mainline in 2.6.37:

5b41d92437 ext4: implement writeback livelock avoidance using page tagging

The fact that changing the IO scheduler from "deadline" to "cfq" is also interesting, given that most of the kernel developers out there use the default IO scheduler of cfq. It's not clear to me why this would be making a difference --- as is why changing the the max_maxwriteback_mb_bump would unjam things. But that suggests to me that it may be a case of livelock, which is why the above-mentioned commit that is in 2.6.37 might make a difference.

For people who want to try cherry-picking this commit back to 2.6.32, you'll also need commit f446daaea by Jan Kara.