On Fri, Jul 06, 2012 at 04:42:50AM -0600, Jan Schmidt wrote:

... down here in the stack. The warning is printed from two levels above,
__readahead_hook.

Either I'm absolutely blind and there's code along the (rather short) road
between those two that might do this I haven't seen. Or someone else messes
with our extent buffers or the underlying pages. What really confuses me is
that it happens so reproducibly.

I've no good idea at the moment how to go on. It might help to get a feeling if
it's shifting around at least a little bit or really constant in the timing of
occurrence. So can you please apply the next patch on top of the other two and
give it some more failure tries? The "checksum mismatch [1234]" line will be of
most interest. I'm also curious what the additional debug variables will say in
the extended version of the very first printk. You can leave out the stack
traces if you like, they won't matter much anyway.

I would suggest turning on slab debug and CONFIG_DEBUG_PAGEALLOC.
Something really strange is happening here.

-chris