Re: [Xen-devel] [3.15-rc3] Bisected: xen-netback mangles packets between two guests on a bridge since merge of "TX grant mapping with SKBTX_DEV_ZEROCOPY instead of copy" series.

[David Vrabel <david.vrabel@citrix.com>]

On 06/05/14 19:07, Steven Haigh wrote:
> On 07/05/14 03:37, Sander Eikelenboom wrote:
>>
>> Tuesday, May 6, 2014, 7:13:44 PM, you wrote:
>>
>>> On 06/05/14 18:07, Steven Haigh wrote:
>>>> On a related note, I'm seeing this on an Arch Linux DomU and kernel
>>>> 3.14.2. Its running aria2c bit torrent client.
>>>>
>>>> The Dom0 is EL6 with Xen 4.2.4 / 3.14.2 kernel.
>>>>
>>>> I also lodged this on the Arch bug tracker:
>>>>       https://bugs.archlinux.org/task/40244
>>>>
>>>> In a nutshell, I see:
>>>> [ 7432.398096] xen_netfront: xennet: skb rides the rocket: 19 slots
>>>> [ 7434.270870] xen_netfront: xennet: skb rides the rocket: 19 slots
>>>> [ 7434.782199] xen_netfront: xennet: skb rides the rocket: 19 slots
>>>> [ 7435.184794] xen_netfront: xennet: skb rides the rocket: 19 slots
>>>> [ 7437.530677] xen_netfront: xennet: skb rides the rocket: 19 slots
>>>> [ 7437.973905] xen_netfront: xennet: skb rides the rocket: 19 slots
>>>> [ 7438.383947] xen_netfront: xennet: skb rides the rocket: 19 slots
>>
>>> Can you also see packet corruption? You can find a testfile generator in 
>>> Sander's mail, downloading from an Apache server with wget seems to be OK.
>>
>> The corruption shouldn't be there since he is running 3.14.2 which is fine.
>>
>> And note that the corruption with 3.15 only seems to occur in traffic between 
>> guests  and not between dom0 and guest. The bit torrent traffic is probably 
>> inet <-> dom0 <-> guest, so shouldn't be affected, even with 3.15.
> 
> Well, I'm trying to get a test together - although I don't have webdav
> anywhere... I was going to try it via FTP, however I'm hitting another
> MMU bug in Xen:
> 	https://bugs.archlinux.org/task/38596
> 
> This references:
> 	https://lkml.org/lkml/2014/1/23/596
> 
> It says that should be fixed in 3.15-rc3 (if I understand properly) -
> but I'm also hitting that on 3.14.2. Did these changes get ported back
> to the 3.14 tree?

This is fixed by 29c7787075 (mm: use paravirt friendly ops for NUMA
hinting ptes) which is tagged for stable but hasn't yet been picked up
for 3.14.y.

You can also work around it by disabling CONFIG_NUMA_BALANCING which
doesn't work with Xen PV guests anyway.

David