[PATCH 4/4] drm/ttm: optimize ttm pool shrinker a bit

[Daniel Vetter]

On Thu, Jan 07, 2021 at 01:49:45PM +0100, Christian König wrote:
> Am 22.12.20 um 14:51 schrieb Daniel Vetter:
> > On Fri, Dec 18, 2020 at 06:55:38PM +0100, Christian König wrote:
> > > Only initialize the DMA coherent pools if they are used.
> > > 
> > > Signed-off-by: Christian König <christian.koenig at amd.com>
> > Ah, just realized the answer to my question on patch 2: The pools are
> > per-device, due to dma_alloc_coherent being per-device (but really mostly
> > it isn't, but that's what we have to deal with fighting the dma-api
> > abstraction).
> > 
> > I think this would make a lot more sense if the shrinkers are per-pool
> > (and also most of the debugfs files), since as-is in a multi-gpu system
> > the first gpu's pool gets preferrentially thrashed. Which isn't a nice
> > design. Splitting that into per gpu shrinkers means we get equal shrinking
> > without having to maintain a global lru. This is how xfs seems to set up
> > their shrinkers, and in general xfs people have a solid understanding of
> > this stuff.
> 
> Well fairness and not trashing the first GPUs pool is the reason why I
> implemented just one shrinker plus a global LRU.

That's kinda defeating the point of how the core mm works. At least of how
I'm understanding how it works. Imo we shouldn't try to re-implement this
kind of balancing across different pools in our callback, since core mm
tries pretty hard to equally shrink already (it only shrinks each shrinker
a little bit each round, but does a lot of rounds under memory pressure).

Also maintaining your own global lru means global locking for the usual
case of none-to-little memory contention, unecessarily wasting the fast
path.

> In other words shrink_slab() just uses list_for_each_entry() on all
> shrinkers.
> 
> In the pool shrinker callback shrink one pool and move it to the end of the
> shrinker list.
> 
> > 
> > Aside: I think it also would make tons of sense to split up your new ttm
> > bo shrinker up into a per-device lru, and throw the global system memory
> > lru out the window completely :-) Assuming we can indeed get rid of it,
> > and vmwgfx doesn't need it somewhere still.
> 
> Yeah, I already have that as a patch set here, but I have this dependent on
> a larger rename of the device structures.

Hm maybe include that in the next round, just for the bigger picture?
Don't have to merge it all in one go, just want to make sure we agree on
where we're going.

> > Aside from this lgtm, but I guess will change a bit with that shuffling.
> 
> Thanks for the review, going to send out a new version with the
> fs_reclaim_acquire/release added in a minute.

Cool.

Cheers, Daniel

> 
> Christian.
> 
> > -Daniel
> > 
> > > ---
> > >   drivers/gpu/drm/ttm/ttm_pool.c | 23 ++++++++++++++++-------
> > >   1 file changed, 16 insertions(+), 7 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/ttm/ttm_pool.c b/drivers/gpu/drm/ttm/ttm_pool.c
> > > index 1cdacd58753a..f09e34614226 100644
> > > --- a/drivers/gpu/drm/ttm/ttm_pool.c
> > > +++ b/drivers/gpu/drm/ttm/ttm_pool.c
> > > @@ -504,10 +504,12 @@ void ttm_pool_init(struct ttm_pool *pool, struct device *dev,
> > >   	pool->use_dma_alloc = use_dma_alloc;
> > >   	pool->use_dma32 = use_dma32;
> > > -	for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
> > > -		for (j = 0; j < MAX_ORDER; ++j)
> > > -			ttm_pool_type_init(&pool->caching[i].orders[j],
> > > -					   pool, i, j);
> > > +	if (use_dma_alloc) {
> > > +		for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
> > > +			for (j = 0; j < MAX_ORDER; ++j)
> > > +				ttm_pool_type_init(&pool->caching[i].orders[j],
> > > +						   pool, i, j);
> > > +	}
> > >   }
> > >   EXPORT_SYMBOL(ttm_pool_init);
> > > @@ -523,9 +525,11 @@ void ttm_pool_fini(struct ttm_pool *pool)
> > >   {
> > >   	unsigned int i, j;
> > > -	for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
> > > -		for (j = 0; j < MAX_ORDER; ++j)
> > > -			ttm_pool_type_fini(&pool->caching[i].orders[j]);
> > > +	if (pool->use_dma_alloc) {
> > > +		for (i = 0; i < TTM_NUM_CACHING_TYPES; ++i)
> > > +			for (j = 0; j < MAX_ORDER; ++j)
> > > +				ttm_pool_type_fini(&pool->caching[i].orders[j]);
> > > +	}
> > >   }
> > >   EXPORT_SYMBOL(ttm_pool_fini);
> > > @@ -630,6 +634,11 @@ int ttm_pool_debugfs(struct ttm_pool *pool, struct seq_file *m)
> > >   {
> > >   	unsigned int i;
> > > +	if (!pool->use_dma_alloc) {
> > > +		seq_puts(m, "unused\n");
> > > +		return 0;
> > > +	}
> > > +
> > >   	ttm_pool_debugfs_header(m);
> > >   	spin_lock(&shrinker_lock);
> > > -- 
> > > 2.25.1
> > > 
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch