IDE fs corruption with 2.3.33 to 2.3.35-pre6

[Jamie Lokier]

David Ford wrote:
> Disable DMA for your harddrive, don't even compile it into the kernel. 
> Something in the recent kernels is busted wrt IDE DMA.

There is a data corruption problem too, for reading and writing.

I am getting data corruption some time after the DMA error happens if I
have to reboot the machine (for other reasons like X crashing and sysrq
not working).

It's not the normal "I rebooted and have a few inconsistencies".  fsck
complains about runs of contiguous inodes having mode 0177777, which
sometimes are actual files whose contents are lost.  Sometimes fsck core
dumps (oh well).

I imagine all-ones blocks are being written.

Also, immediately after the DMA irq timeout, even though a message
appears saying DMA has been turned off, it seems some data that's been
read is corrupted.  Because an executable loading at the time crashes,
and continues to crash each time it is started until I clear the cache.

I haven't seen it with 2.3.35 proper yet, but I've only run that for a
few minutes so far.

-- Jamie