sky2 rx error problems...

sky2 rx error problems...

[Daniel J Blueman]

To follow on from the previous 'sky2 rx error' problem reported, the
problem doesn't have anything to do with iptables or MTU settings, ie
can be reproduced out the box.

The race is hit after a couple of GBs of data. Initially, I thought
this may have something to do with broken MSI support on my Intel
ICH-7M chipset, but disabling MSI didn't help.

kernel: sky2 lan0: rx error, status 0x977d977d length 0

Is there any debug information I can get to help or such? Voluntary
preemption is enabled, UP, HZ=100 on 2.6.17-rc4 on i686.

Dan
-- 
Daniel J Blueman

Re: sky2 rx error problems...

[Daniel J Blueman]

Hi Stephen,

Sorry to bring you more trouble. I'm able to reproduce this 'status
0x977d977d' rx race even with a 100Mb FD connection (MTU 1500) to the
host sending the data.

Would it help if I try and capture some packet stats or a dump of
this? The packets being received at the sending host will be a steady
stream of TCP acks and SMB protocol acks (every 64KB of data), so not
too heavy. The sending host has a Pentium-M 1.7GHz with 2MB L2$ and is
quiescent apart from samba sending data via sendfile(), so is
relatively powerful.

Let me know if there is anything I can do to help.

Many thanks,
  Daniel

On 31/05/06, Stephen Hemminger <shemminger@osdl.org> wrote:
> On Tue, 30 May 2006 22:08:12 +0100
> "Daniel J Blueman" <daniel.blueman@gmail.com> wrote:
>
> > To follow on from the previous 'sky2 rx error' problem reported, the
> > problem doesn't have anything to do with iptables or MTU settings, ie
> > can be reproduced out the box.
> >
> > The race is hit after a couple of GBs of data. Initially, I thought
> > this may have something to do with broken MSI support on my Intel
> > ICH-7M chipset, but disabling MSI didn't help.
> >
> > kernel: sky2 lan0: rx error, status 0x977d977d length 0
> >
> > Is there any debug information I can get to help or such? Voluntary
> > preemption is enabled, UP, HZ=100 on 2.6.17-rc4 on i686.
> >
> > Dan
>
> No the problem is your CPU isn't keeping up with the incoming flood
> and what ever you are connected to doesn't do hardware flow control.
> The driver doesn't properly flush the FIFO (yet), and it is not clear
> even from the NDA documentation the steps required.
-- 
Daniel J Blueman