Nouveau - Dec 2017 - nouveau. swiotlb: coherent allocation failed for device 0000:01:00.0 size=2097152

On 2017-12-18 08:01 PM, Tobias Klausmann wrote:
> On 12/18/17 7:06 PM, Mike Galbraith wrote:
>> Greetings,
>>
>> Kernel bound workloads seem to trigger the below for whatever reason.
>>   I only see this when beating up NFS.  There was a kworker wakeup
>> latency issue, but with a bandaid applied to fix that up, I can still
>> trigger this.
> 
> 
> Hi,
> 
> i have seen this one as well with my system, but i could not find an
> easy way to trigger it for bisecting purpose. If you can trigger it
> conveniently, a bisect would be nice!
I'm seeing this (with the amdgpu and radeon drivers) when restic takes a
backup, creating memory pressure. I happen to have just finished
bisecting, the result is:

648bc3574716400acc06f99915815f80d9563783 is the first bad commit
commit 648bc3574716400acc06f99915815f80d9563783
Author: Christian König <christian.koenig at amd.com>
Date:   Thu Jul 6 09:59:43 2017 +0200

    drm/ttm: add transparent huge page support for DMA allocations v2

    Try to allocate huge pages when it makes sense.

    v2: fix comment and use ifdef


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer

On 2017-12-19 11:37 AM, Michel Dänzer wrote:
> On 2017-12-18 08:01 PM, Tobias Klausmann wrote:
>> On 12/18/17 7:06 PM, Mike Galbraith wrote:
>>> Greetings,
>>>
>>> Kernel bound workloads seem to trigger the below for whatever
reason.
>>>   I only see this when beating up NFS.  There was a kworker wakeup
>>> latency issue, but with a bandaid applied to fix that up, I can
still
>>> trigger this.
>>
>>
>> Hi,
>>
>> i have seen this one as well with my system, but i could not find an
>> easy way to trigger it for bisecting purpose. If you can trigger it
>> conveniently, a bisect would be nice!
> 
> I'm seeing this (with the amdgpu and radeon drivers) when restic takes
a
> backup, creating memory pressure. I happen to have just finished
> bisecting, the result is:
> 
> 648bc3574716400acc06f99915815f80d9563783 is the first bad commit
> commit 648bc3574716400acc06f99915815f80d9563783
> Author: Christian König <christian.koenig at amd.com>
> Date:   Thu Jul 6 09:59:43 2017 +0200
> 
>     drm/ttm: add transparent huge page support for DMA allocations v2
> 
>     Try to allocate huge pages when it makes sense.
> 
>     v2: fix comment and use ifdef
> 
> 
BTW, I haven't noticed any bad effects other than the dmesg splats, so
maybe it's just noise about transient failures for which there is a
proper fallback in place.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer

Am 19.12.2017 um 11:39 schrieb Michel Dänzer:
> On 2017-12-19 11:37 AM, Michel Dänzer wrote:
>> On 2017-12-18 08:01 PM, Tobias Klausmann wrote:
>>> On 12/18/17 7:06 PM, Mike Galbraith wrote:
>>>> Greetings,
>>>>
>>>> Kernel bound workloads seem to trigger the below for whatever
reason.
>>>>    I only see this when beating up NFS.  There was a kworker
wakeup
>>>> latency issue, but with a bandaid applied to fix that up, I can
still
>>>> trigger this.
>>>
>>> Hi,
>>>
>>> i have seen this one as well with my system, but i could not find
an
>>> easy way to trigger it for bisecting purpose. If you can trigger it
>>> conveniently, a bisect would be nice!
>> I'm seeing this (with the amdgpu and radeon drivers) when restic
takes a
>> backup, creating memory pressure. I happen to have just finished
>> bisecting, the result is:
>>
>> 648bc3574716400acc06f99915815f80d9563783 is the first bad commit
>> commit 648bc3574716400acc06f99915815f80d9563783
>> Author: Christian König <christian.koenig at amd.com>
>> Date:   Thu Jul 6 09:59:43 2017 +0200
>>
>>      drm/ttm: add transparent huge page support for DMA allocations v2
>>
>>      Try to allocate huge pages when it makes sense.
>>
>>      v2: fix comment and use ifdef
>>
>>
> BTW, I haven't noticed any bad effects other than the dmesg splats, so
> maybe it's just noise about transient failures for which there is a
> proper fallback in place.
Yeah, I think that is exactly what happens here.

We try to allocate a huge page, but fail and so fall back to using 
multiple 4k pages instead.

Going to send out a patch to suppress the warning.

Thanks for bisecting this,
Christian.