Ohhh, Ahhh. These write requests are made by bdflush and the
sys_bdflush syscall. There is missing the kick to the device
driver.
Here is a one line patch which solves this bug for 2.0.30. Probably
similar patch should be applied to 2.1.43 too.
This could certainly explain a lot. Are the people playing with the
swap/buffer-cache patch sets listening? 8-)