We were mixing end_request with end_that_request_chunk, which
is apparantly bad news.  Also, the handycapped pxamci driver
was telling us that it had transferred all data successfully
on error, which is obviously wrong.