On Sun, Mar 24, 2013 at 05:42:41AM +0000, Huang, Xiong wrote:

quoted

quoted

quoted

The caputered packets show that a 16 byte segment overwrites later
data in the same packet, sometimes even multiple contiguous 16 byte

segments.

quoted

quoted

The corrupted packets look like this:

...ABC....DBBEF..

Each character represents a 16 byte region, the dots are dont-care
placeholders. The BB-region is overwritten by the data of the first B-

region.

quoted

quoted

It looks like a DMA issue. Not checksum related. :(

Yes, I think so, too. Until this is resolved I would propose to not set
CHECKSUM_UNNECESSARY because it could corrupt data (like for the original
poster).

Do you have any idea where how these errors could be resolved? I e.g. could
not reproduce it with FreeBSD and the driver seems to be related to the linux
driver.

1.  pages cross 4GB bouandary ?  or overlap with each other ?

This should not be the case.

2. try to set hw->dmaw_block = atl1e_dma_req_128 

Sorry, we can still reproduce corrupted frames with this change.

3. can you print reg15C0 content when the NIC is in normal tx/rx status ?

Yes, dma register is reg15C0 == 0x247c24 all the time (vanilla kernel without
any of your proposed changes).

Couldn't this also be because of a timing issue? It does take much more time
on slower machines(eeepc) to reproduce this issue. On a normal desktop machine it
does nearly occur instantaneous.

Thanks,

  Hannes