[Intel-gfx] [PATCH] drm/i915/ttm: consider all placements for the page alignment

[Matthew Auld]

On 22/06/2021 13:29, Thomas Hellström wrote:
> 
> On 6/22/21 2:15 PM, Matthew Auld wrote:
>> On Tue, 22 Jun 2021 at 11:11, Thomas Hellström
>> <thomas.hellstrom at linux.intel.com> wrote:
>>>
>>> On 6/22/21 11:58 AM, Matthew Auld wrote:
>>>> Just checking the current region is not enough, if we later migrate the
>>>> object somewhere else. For example if the placements are {SMEM, LMEM},
>>>> then we might get this wrong. Another idea might be to make the
>>>> page_alignment part of the ttm_place, instead of the BO.
>>>>
>>>> Signed-off-by: Matthew Auld <matthew.auld at intel.com>
>>>> Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
>>>> ---
>>>>    drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 21 ++++++++++++++++++++-
>>>>    1 file changed, 20 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c 
>>>> b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
>>>> index c5deb8b7227c..5d894bba6430 100644
>>>> --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
>>>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
>>>> @@ -753,6 +753,25 @@ void i915_ttm_bo_destroy(struct 
>>>> ttm_buffer_object *bo)
>>>>                call_rcu(&obj->rcu, __i915_gem_free_object_rcu);
>>>>    }
>>>>
>>>> +static u64 i915_gem_object_page_size(struct drm_i915_gem_object *obj)
>>>> +{
>>>> +     u64 page_size;
>>>> +     int i;
>>>> +
>>>> +     if (!obj->mm.n_placements)
>>>> +             return obj->mm.region->min_page_size;
>>>> +
>>>> +     page_size = 0;
>>>> +     for (i = 0; i < obj->mm.n_placements; i++) {
>>>> +             struct intel_memory_region *mr = obj->mm.placements[i];
>>>> +
>>>> +             page_size = max_t(u64, mr->min_page_size, page_size);
>>>> +     }
>>>> +
>>>> +     GEM_BUG_ON(!page_size);
>>>> +     return page_size;
>>>> +}
>>>> +
>>>>    /**
>>>>     * __i915_gem_ttm_object_init - Initialize a ttm-backed i915 gem 
>>>> object
>>>>     * @mem: The initial memory region for the object.
>>>> @@ -793,7 +812,7 @@ int __i915_gem_ttm_object_init(struct 
>>>> intel_memory_region *mem,
>>>>        obj->base.vma_node.driver_private = i915_gem_to_ttm(obj);
>>>>        ret = ttm_bo_init(&i915->bdev, i915_gem_to_ttm(obj), size,
>>>>                          bo_type, &i915_sys_placement,
>>>> -                       mem->min_page_size >> PAGE_SHIFT,
>>>> +                       i915_gem_object_page_size(obj) >> PAGE_SHIFT,
>>> Hmm, can't we just have the buddy manager silently enforce its
>>> min_page_size?
>> Maybe, but we need some way of overriding it for all of our page-table
>> allocations(and some other stuff also), so being able to control the
>> page_alignment at the object level here seems reasonable? Could maybe
>> pass it through with create_lmem_with_page_size(..., page_size)? Ok,
>> it might be best to first type it and then see how it will all fit
>> together.
>>
> Hmm, OK, I'm not 100% sure what the various requirements are here on the 
> object level. But for region requirements, I think we've historically 
> enforced that through the manager, taking also the bo->page_alignment 
> into account and applying the larger of the two,
> 
> There is an example in vmw_thp_insert_aligned().

Yeah, so for our use case we need to support page_alignment < 
min_page_size, for page-tables(4K). I guess pushing the min_page_size 
into buddy_man, and then letting page_alignment override that, with the 
added caveat that it can be smaller could work. Otherwise just using 
mm.chunk_size would already fit the bill quite nicely.

> 
> /Thomas
> 
>