[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <e0c43868-8201-fe46-9e8b-5e38c2611340@gmail.com>
Date: Mon, 17 Feb 2020 20:59:46 +0100
From: Heiner Kallweit <hkallweit1@...il.com>
To: Vincas Dargis <vindrg@...il.com>,
Salvatore Bonaccorso <carnil@...ian.org>
Cc: netdev@...r.kernel.org
Subject: Re: About r8169 regression 5.4
On 17.02.2020 19:08, Vincas Dargis wrote:
> 2020-02-16 01:27, Heiner Kallweit rašė:
>> One more idea:
>> Commit "r8169: enable HW csum and TSO" enables certain hardware offloading by default.
>> Maybe your chip version has a hw issue with offloading. You could try:
>>
>> 1. Disable TSO
>> ethtool -K <if> tso off
>>
>> 2. If this didn't help, disable all offloading.
>> ethtool -K <if> tx off sg off tso off
>>
>
> Unmodified 5.4 was running successfully for whole Sunday with `tx off sg off tso off`! Disabling only tso did not help, while disabling all actually avoided the timeout.
>
Great, thanks a lot for testing! Then the bisecting shouldn't be needed. Since 5.4 these features are enabled by default,
up to 5.3 they are available but have to be enabled explicitly. This should explain the observed behavior.
So it looks like this chip version has a hw issue with tx checksumming. I contacted Realtek to see whether
they are aware of any such hw issue. Depending on their feedback we may have to add a quirk for this chip version
to not enable these features by default.
> I've attached kern.log from boot 'til 5.4 got that timeout (when I did not use these off's).
>
> About bisecting - I need to figure out how to build linux-image and linux-headers package only, to reduce that almost hour build...
Powered by blists - more mailing lists