Re: 1e918876 breaks r8169 (linux-3.18+)

["Holger Hoffstätte" <holger.hoffstaette@googlemail.com>]

On Sat, 21 Feb 2015 11:31:04 +0100, Florian Westphal wrote:

> Tomas Szepe <szepe@pinerecords.com> wrote:
>> > I tried to reproduce this without success so far on my RTL8168d/8111d device.
>> > I've been running 40 parallel netperf TCP_STREAM tests (1gbit) for the
>> > last 5 hours and so far I saw no watchdog tx timeouts.
>> > 
>> > I'll keep this running for a day or so to see if it just takes more time
>> > to trigger.
>> 
>> So, how's this coming along?  Don't you think the patch should be reverted
>> until the problem is diagnosed/understood/fixed?
> 
> Sorry.
> 
> David, please consider reverting
> 
> 1e918876853aa85435e0f17fd8b4a92dcfff53d6
> (r8169: add support for Byte Queue Limits)
> 
> 	and
> 
> 0bec3b700d106a8b0a34227b2976d1a582f1aab7
> (r8169: add support for xmit_more)
> 
> I cannot reproduce any hangs (tried for 2days with 40 parallel
> netperfs using both 100mbit and 1gbit receiver).
> 
> And I don't see anything wrong with the change either.
> Seems like some revisions of the HW are just dodgy?
> 
> I hate giving up, but I have no means to diagnose this any further.
> Even reporter says it doesn't affect all of his r8169 nics.
> 
> So I think the change is correct per se, but might be revealing some
> HW/firmware bug.

Florian,

have you experimented with offload settings? The only times r8169 seems to
hiccup is with sg/tso enabled. I've reverted my NIC settings back to mostly
defaults (which does not enable sg/tso) and had no hangs, spurious timeouts
or other problems ever since, despite BQL, xmit_more and client/server
use for 24/7.

Tomas never said whether his setup enabled any offload settings; it's not
inconceivable that a distribution might try to automatically "optimize"
things.

-h