Re: [E1000-devel] "e1000_clean_tx_irq: Detected Tx Unit Hang" errors only on pppoe connect

[Kok, A. <auk...@in...>]

snowcrash+e1000 wrote:
>>  can you please eliminate Xen completely and reproduce this on a non-xen kernel?
>>  that would really help a lot...
> 
> i know it would.
> 
> unfortunately, i can't really, without basically tearing this box
> down, &/or polluting the Dom0 -- which by intent (of xen) & design i'm
> NOT monkeying with.

ok, fair enough.

Let me explain the exact issue that currently is the only known unfixed cause for
TX hangs:

The driver writes descriptors through DMA to the adapter card signifying that the
hardware should start fetching data and transmit that data. This DMA transaction
(on some non-intel chipsets) goes once in a while completely wrong and the adapter
ends up getting wrong data in it's copy of the descriptor memory.

This then means that the adapter sees wrong bits set and waits for that one
descriptor to have those proper bits set. Forever. The driver cannot recover since
it wrote the proper data to the card in the first place. The adapter cannot
recover since it received data perfectly fine from the bus/bridge and acts exactly
like it should.

So, obviously this is a very hard issue to solve and we have not yet found a
resolution since the real cause is completely outside of the driver and the
adapter, and in hardware that we do not control.

We're here trying (still) to come up with a workaround... can never really fix
this issue since it's not Intel hardware.


If your issue is caused by some new and unknown problem than it would help if you
would be willing to attempt to reproduce the issue without Xen, with an updated
driver and kernel. But that's up to you.

Cheers,

Auke