Link bouncing with multiple drivers.

Link bouncing with multiple drivers.

[Ben Greear]

I have a strange issue:

In several different scenarios, when we run pktgen at line speed
on 1G through our proprietary bridge-ish module, we see link go
up and down every minute or so.  This happens with e1000, e1000e,
and have also seen it on ixgbe (though not line speed here...but
we were driving it as hard as the systems could handle).

We cannot reproduce this when a normal bridge is substituted for
our proprietary module, so it must either be a bug in our code somewhere,
or something to do with the fact that our module causes more work to
be done than a bridge (backed up driver queues on rx/tx, time-stamps, etc).

Since it happens across multiple drivers and hardware (and operating systems:  F5, F8, F11,
but all with kernels based on 2.6.31), it must be some general issue.

If anyone has any ideas where I should start poking, I would
be grateful.  For now, I'm off to dig in the e1000e code...

Thanks,
Ben


-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com

Re: Link bouncing with multiple drivers.

[Ben Greear]

On 10/12/2009 10:55 AM, Ben Greear wrote:
> I have a strange issue:
>
> In several different scenarios, when we run pktgen at line speed
> on 1G through our proprietary bridge-ish module, we see link go
> up and down every minute or so. This happens with e1000, e1000e,
> and have also seen it on ixgbe (though not line speed here...but
> we were driving it as hard as the systems could handle).
>
> We cannot reproduce this when a normal bridge is substituted for
> our proprietary module, so it must either be a bug in our code somewhere,
> or something to do with the fact that our module causes more work to
> be done than a bridge (backed up driver queues on rx/tx, time-stamps, etc).
>
> Since it happens across multiple drivers and hardware (and operating
> systems: F5, F8, F11,
> but all with kernels based on 2.6.31), it must be some general issue.
>
> If anyone has any ideas where I should start poking, I would
> be grateful. For now, I'm off to dig in the e1000e code...

Ahh, it was indeed my bug.  I forgot to update the txqueue trans_start and that
caused the watchdog to fire way too often.

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com