Re: [PATCH v8 7/7] drm/panfrost: Switch to generic memory shrinker

[Dmitry Osipenko <dmitry.osipenko@collabora.com>]

Hello Steve,

On 11/14/22 19:54, Steven Price wrote:
> On 05/11/2022 23:27, Dmitry Osipenko wrote:
>> Replace Panfrost's custom memory shrinker with a common drm-shmem
>> memory shrinker.
>>
>> Signed-off-by: Dmitry Osipenko <dmitry.osipenko@collabora.com>
> 
> Sadly this triggers GPU faults under memory pressure - it looks
> suspiciously like mappings are being freed while the jobs are still running.
> 
> I'm not sure I understand how the generic shrinker replicates the
> "gpu_usecount" atomic that Panfrost currently has, and I'm wondering if
> that's the cause?
> 
> Also just reverting this commit (so just patches 1-6) I can't actually
> get Panfrost to purge any memory. So I don't think the changes (most
> likely in patch 4) are quite right either.
> 
> At the moment I don't have the time to investigate in detail. But if
> you've any ideas for something specific I should look at I can run more
> testing.

Thank you for the testing! It just occurred to me that the shrinker
callback lost the dma_resv_test_signaled() in comparison to the previous
versions of this patchset. It appeared to me that the drm_gem_lru now
checks whether reservation is busy, but it doesn't.

I saw a similar page faults once in a while when was testing the
Panfrost driver, but then couldn't reproduce the faults after applying
the IOMMU unmap range fix that Robin made recently.

I'll re-add the dma_resv_test_signaled() in v9, it was a luck that I
didn't hit it much during my testing.

-- 
Best regards,
Dmitry