Re: [PATCH v2 6/8] drm/shmem-helper: Add generic memory shrinker

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

On 3/17/22 19:13, Rob Clark wrote:
...
>>>> +               /* prevent racing with job submission code paths */
>>>> +               if (!dma_resv_trylock(obj->resv))
>>>> +                       goto shrinker_lock;
>>>
>>> jfwiw, the trylock here is in the msm code isn't so much for madvise
>>> (it is an error to submit jobs that reference DONTNEED objects), but
>>> instead for the case of evicting WILLNEED but inactive objects to
>>> swap.  Ie. in the case that we need to move bo's back in to memory, we
>>> don't want to unpin/evict a buffer that is later on the list for the
>>> same job.. msm shrinker re-uses the same scan loop for both
>>> inactive_dontneed (purge) and inactive_willneed (evict)
>>
>> I don't see connection between the objects on the shrinker's list and
>> the job's BOs. Jobs indeed must not have any objects marked as DONTNEED,
>> this case should never happen in practice, but we still need to protect
>> from it.
> 
> Hmm, let me try to explain with a simple example.. hopefully this makes sense.
> 
> Say you have a job with two bo's, A and B..  bo A is not backed with
> memory (either hasn't been used before or was evicted.  Allocating
> pages for A triggers shrinker.  But B is still on the
> inactive_willneed list, however it is already locked (because we don't
> want to evict B to obtain backing pages for A).

I see now what you're talking about, thanks. My intention of locking the
reservations is different since eviction isn't supported by this v2. But
we probably will be able to re-use this try_lock for protecting from
swapping out job's BOs as well.

>>> I suppose using trylock is not technically wrong, and it would be a
>>> good idea if the shmem helpers supported eviction as well.  But I
>>> think in the madvise/purge case if you lose the trylock then there is
>>> something else bad going on.
>>
>> This trylock is intended for protecting job's submission path from
>> racing with madvise ioctl invocation followed by immediate purging of
>> BOs while job is in a process of submission, i.e. it protects from a
>> use-after-free.
> 
> ahh, ok
> 
>> If you'll lose this trylock, then shrinker can't use
>> dma_resv_test_signaled() reliably anymore and shrinker may purge BO
>> before job had a chance to add fence to the BO's reservation.
>>
>>> Anyways, from the PoV of minimizing lock contention when under memory
>>> pressure, this all looks good to me.
>>
>> Thank you. I may try to add generic eviction support to the v3.
> 
> eviction is a trickier thing to get right, I wouldn't blame you for
> splitting that out into it's own patchset ;-)
> 
> You probably also would want to make it a thing that is opt-in for
> drivers using the shmem helpers

I had the same thoughts, will see.