[PATCH 1/3] drm: create unified vma offset manager

[Chris Wilson]

On Wed, 19 Dec 2012 11:56:18 +1000, Dave Airlie <airlied at gmail.com> wrote:
> From: Dave Airlie <airlied at redhat.com>
> 
> So we have to offset manager implementations for dealing with VMA offsets.
> GEM had one using the hash table, TTM had one using an rbtree,
> 
> I'd rather we just had one to rule them all, since ttm is using the rbtree
> variant to allow sub mappings, to keep ABI we need to use that one.
> 
> This also adds a bunch of inline helpers to avoid gem/ttm poking around
> inside the vma_offset objects. TTM needs a reset helper to remove a vma_offset
> when it copies an object to another one for buffer moves. The helpers
> also let drivers avoid the map_list.hash_key << PAGE_SHIFT nonsense.

Any clue as to the performance difference between the two
implementations? What does it add to the cost of a pagefault?

Hmm, don't have an i-g-t handy for scalability testing of the fault
handlers.

> +int drm_vma_offset_setup(struct drm_vma_offset_manager *man,
> +			 struct drm_vma_offset_node *node,
> +			 unsigned long num_pages)
> +{
> +	int ret;
> +
> +retry_pre_get:
> +	ret = drm_mm_pre_get(&man->addr_space_mm);
> +	if (unlikely(ret != 0))
> +		return ret;
> +
> +	write_lock(&man->vm_lock);
> +	node->vm_node = drm_mm_search_free(&man->addr_space_mm,
> +					   num_pages, 0, 0);
> +
> +	if (unlikely(node->vm_node == NULL)) {
> +		ret = -ENOMEM;
ret = -ENOSPC;

Depended upon by the higher layers for determining when to purge their
caches; i-g-t/gem_mmap_offset_exhaustion

> +		goto out_unlock;
> +	}
> +
> +	node->vm_node = drm_mm_get_block_atomic(node->vm_node,
> +						num_pages, 0);
> +

I'd feel happier if this tried to only take from the preallocated pool
rather than actually try a GFP_ATOMIC allocation.

> +	if (unlikely(node->vm_node == NULL)) {
> +		write_unlock(&man->vm_lock);
> +		goto retry_pre_get;
> +	}
> +
> +	node->num_pages = num_pages;
> +	drm_vma_offset_insert_rb(man, node);
> +	write_unlock(&man->vm_lock);
> +
> +	return 0;
> +out_unlock:
> +	write_unlock(&man->vm_lock);
> +	return ret;	
> +	
> +}
> +EXPORT_SYMBOL(drm_vma_offset_setup);

-- 
Chris Wilson, Intel Open Source Technology Centre