Nouveau - Feb 2021 - drm/nouneau: 5.11 cycle regression bisected to 461619f5c324 "drm/nouveau: switch to new allocator"

On Wed, 2021-02-10 at 14:26 +0100, Christian K?nig
wrote:
>
> Am 10.02.21 um 13:22 schrieb Mike Galbraith:
> > On Wed, 2021-02-10 at 12:44 +0100, Christian K?nig wrote:
> >> Please try to add a "return NULL" at the beginning of
ttm_pool_type_take().
> >>
> >> That should effectively disable using the pool.
> > That did away with the yield looping, but it doesn't take long for
the
> > display to freeze.  I ssh'd in from lappy, but there was nada in
dmesg.
>
> Yeah, that is expected. Without taking pages from the pool we leak
> memory like sieve.
>
> At least we could narrow down the problem quite a bit with that.
>
> Can you test the attached patch and see if it helps?
Yup, that seems to have fixed it all up.  Another one bites the dust ;)

	-Mike

Am 10.02.21 um 16:30 schrieb Mike Galbraith:
> On Wed, 2021-02-10 at 14:26 +0100, Christian K?nig wrote:
>> Am 10.02.21 um 13:22 schrieb Mike Galbraith:
>>> On Wed, 2021-02-10 at 12:44 +0100, Christian K?nig wrote:
>>>> Please try to add a "return NULL" at the beginning of
ttm_pool_type_take().
>>>>
>>>> That should effectively disable using the pool.
>>> That did away with the yield looping, but it doesn't take long
for the
>>> display to freeze.  I ssh'd in from lappy, but there was nada
in dmesg.
>> Yeah, that is expected. Without taking pages from the pool we leak
>> memory like sieve.
>>
>> At least we could narrow down the problem quite a bit with that.
>>
>> Can you test the attached patch and see if it helps?
> Yup, that seems to have fixed it all up.  Another one bites the dust ;)
Ah! So basically nouveau relies on that the pages are cleared.

Yeah, that wasn't consitently implemented before. Some code path was 
clearing the page, some wasn't.

Any objections that I add a Reported-and-tested-by: Mike Galbraith 
<efault at gmx.de> ?

Regards,
Christian.
>
> 	-Mike
>