[Intel-gfx] [PATCH v3 3/8] drm/i915: Add a new "remapped" gtt_view

[Tvrtko Ursulin]

On 09/10/2018 12:54, Ville Syrjälä wrote:
> On Tue, Oct 09, 2018 at 09:24:33AM +0100, Tvrtko Ursulin wrote:
>>
>> On 05/10/2018 19:42, Ville Syrjälä wrote:
>>> On Mon, Oct 01, 2018 at 04:48:21PM +0100, Tvrtko Ursulin wrote:
>>>>
>>>> On 01/10/2018 16:37, Chris Wilson wrote:
>>>>> Quoting Ville Syrjälä (2018-10-01 16:27:43)
>>>>>> On Mon, Oct 01, 2018 at 04:12:09PM +0100, Chris Wilson wrote:
>>>>>>> Quoting Ville Syrjälä (2018-10-01 16:03:30)
>>>>>>>> On Wed, Sep 26, 2018 at 08:50:25AM +0100, Tvrtko Ursulin wrote:
>>>>>>>>>
>>>>>>>>> On 25/09/2018 20:37, Ville Syrjala wrote:
>>>>>>>>> One more thing, do you really need random access for this
>>>>>>>>> transformation? Or you could walk the sg list as it is? Just if you hit
>>>>>>>>> a too long chunk you need to copy a trimmed version over and know where
>>>>>>>>> to continue for the next row. If doable it would be better than having
>>>>>>>>> to kvmalloc_array.
>>>>>>>>
>>>>>>>> I think Chris suggested just using i915_gem_object_get_dma_address()
>>>>>>>> here. But I'm not sure why we're not using it for rotate_pages()
>>>>>>>> as well.
>>>>>>>
>>>>>>> Tvrtko is opposed to populating the obj->mm.pages cache with no defined
>>>>>>> release point. I say the mempressure and shrinker should to the right
>>>>>>> thing, but it's a big if.
>>>>>>
>>>>>> OK.
>>>>>>
>>>>>> Well, looks to me like i915_gem_object_get_dma_address() is the
>>>>>> only convenient looking thing for iterating the pages without
>>>>>> arowning the code in irrelevant details about sgs and whatnot.
>>>>>> I suppose it should be possible to write some helpers that avoid
>>>>>> all that and don't need the temp array, but I'm not really
>>>>>> motivated enough to do that myself.
>>>>>
>>>>> Keep it simple and use get_dma_address(). We can find ways to throw away
>>>>> the cache later if need be.
>>>>
>>>> I'd do it straight away. I think cache for a large framebuffer, the kind
>>>> which needs remapping could be quite big! Even the more fragmented
>>>> memory the bigger the cache, and so if it sticks around pointlessly for
>>>> the lifetime of the framebuffer it is a double whammy.
>>>
>>> The tree is indexed with the ggtt offset so memory fragmentation
>>> shouldn't matter I think. Or did I totally miss something?
>>
>> I think it is indexed by page index, but the number of tree entries
>> depends on the number of sg chunks, which in turn depends on the system
>> memory fragmentation. Consecutive pages are stored as "exceptional"
>> entries so that is more efficient IIRC,
> 
> Didn't look any more efficient to me. It's still one page per slot
> AFAICS. The exceptional entry just tells you where to find the start
> of the current sg chunk.
> 
>> but TBH I don't remember how
>> many of those it can store before it needs a new tree node. Anyways, if
>> I am not completely mistaken then the tree metadata size is proportional
>> to backing store fragmentation, and secondary proportional to object
>> size. (Due exceptional entry spill.)
>>
>>> The relative overhead should be highest for a single page object
>>> (576/4096 = ~15%). For big objects it should be something around
>>> .2% AFAICS.
>>
>> It is a bit annoying since radix tree does not have a helper to tell us
>> the number of allocated nodes. I don't remember how I measure this last
>> time.. Will try to recall.
> 
> I wrote myself a helper to count the nodes and that seems to
> agree with my back of the envelope calculations.

How many sg chunks and then radix nodes for some typical framebuffer did 
that show? Or maybe not typical but the one which will require 
remapping. Did you measure after a fresh boot only or also on a bit of a 
memory constrained system (some memory pressure should equal more 
fragmentation)?

If not even that shows the cache size as significant, then I guess my 
worries that large framebuffers would be sufficiently fragmented for 
this to be much more were unfounded, so feel free to go with it.

Regards,

Tvrtko