Hello,

On Tue, Apr 30, 2019 at 10:12:27AM +0200, Uwe Kleine-König wrote:

On Thu, Apr 25, 2019 at 09:17:32PM +0200, Aurelien Jarno wrote:

quoted

On 2019-04-25 14:50, Aurelien Jarno wrote:

quoted

On 2019-04-23 22:16, Aurelien Jarno wrote:

quoted

Source: linux
Version: 4.19.28-2
Severity: important

After upgrading hartmann.debian.org (an armhf buildd using an Armada XP
GP board) from buster to stretch, the ethernet device is not working

"upgrading from buster to stretch" doesn't make sense. I think you meant
from stretch to buster.

quoted

quoted

More precisely the board is a "Marvell Armada XP Development Board
DB-MV784MP-GP"

quoted

anymore. Using tcpdump on both the buildd and a remote host, it appears
that the packets correctly leave the board and that the reception side
fails.

If you can send to a remote host at least ARP (or ND) must be working,
so some reception still works, right?

quoted

quoted

quoted

The module used for the ethernet device is mvneta. The corresponding DT
compatible entry is "marvell,armada-xp-neta".

I have started a "bisection" with the kernels from snapshot. This is
what I have found so far:

This one works:
- linux-image-4.19.0-rc6-armmp-lpae_4.19~rc6-1~exp1_armhf.deb 

The following ones don't:
- linux-image-4.19.0-rc7-armmp-lpae_4.19~rc7-1~exp1_armhf.deb
- linux-image-5.0.0-trunk-armmp_5.0.2-1~exp1_armhf.deb

My guess (I don't have time to try more now) is that the issue is caused
by the following change:

|  [ Uwe Kleine-König ]
|  * [armhf] enable MVNETA_BM_ENABLE and CAN_FLEXCAN as a module

I confirm this is the issue. Disabling MVNETA_BM_ENABLE on kernel 
4.19.28-2 fixes the issue. Note that it breaks the ABI.

A colleague happens to work with an XP based machine with a (nearly)
vanilla kernel based on 5.1.0-rc6 and there enabling MVNETA_BM_ENABLE
doesn't render networking nonfunctional.

Looking through the changes to drivers/net/ethernet/marvell/mvneta*
between 5.0 and 5.1-rc6 there isn't something that would explain a fix
though. There doesn't seem to be a good explanation in the debian
specific patches either.

So this problem is either machine specific or it works with the mvneta
driver builtin. (I didn't double check, but guess that my colleague uses
=y and the Debian kernel =m). Well, or I missed something.

Is it possible to test a few things on hartmann? I'd suggest:

 - try (vanilla) 5.1-rc6 with MVNETA=y
 - try an older kernel (maybe 4.6 as the buffer manager stuff was
   introduced in dc35a10f68d3 ("net: mvneta: bm: add support for
   hardware buffer management") which made it into 4.6-rc1) with
   MVNETA_BM_ENABLE=[ym].

Thanks to Steve McIntyre I got access to a DB-MV784MP-GP and did the
second test. I used mvebu_v7_defconfig and enabled CGROUPS and AUTOFS4
(to please systemd) and MVNETA_BM_ENABLE=y on 4.6. The latter breaks
networking similar to newer kernel versions.

So I guess the buffer management never worked on that board.

I don't have the time to debug this issue further, but will disable
buffer management for the Debian kernel again.

Best regards
Uwe