[PATCH 2/2] drm/ttm: optimize the pool shrinker a bit v2

[Christian König]

Am 15.04.21 um 22:33 schrieb Andrew Morton:
> On Thu, 15 Apr 2021 13:56:24 +0200 "Christian König" <ckoenig.leichtzumerken at gmail.com> wrote:
>
>> @@ -530,6 +525,11 @@ void ttm_pool_fini(struct ttm_pool *pool)
>>   			for (j = 0; j < MAX_ORDER; ++j)
>>   				ttm_pool_type_fini(&pool->caching[i].orders[j]);
>>   	}
>> +
>> +	/* We removed the pool types from the LRU, but we need to also make sure
>> +	 * that no shrinker is concurrently freeing pages from the pool.
>> +	 */
>> +	sync_shrinkers();
> It isn't immediately clear to me how this works.  ttm_pool_fini() has
> already freed all the pages hasn't it?  So why would it care if some
> shrinkers are still playing with the pages?

Yes ttm_pool_fini() has freed up all pages which had been in the pool 
when the function was called.

But the problem is it is possible that a parallel running shrinker has 
taken a page from the pool and is in the process of freeing it up.

When I return here the pool structure and especially the device 
structure are freed while the parallel running shrinker is still using them.

I could go for a design where we have one shrinker per device instead, 
but that would put a bit to much pressure on the pool in my opinion.

> Or is it the case that ttm_pool_fini() is assuming that there will be
> some further action against these pages, which requires that shrinkers
> no longer be accessing the pages and which further assumes that future
> shrinker invocations will not be able to look up these pages?
>
> IOW, a bit more explanation about the dynamics here would help!

Sorry, I'm not a native speaker of English and sometimes still have a 
hard time explaining things.

Regards,
Christian.