[Mesa-dev] [PATCH] i965: Embrace "unlimited" GTT mmap support

[Chris Wilson]

On Thu, Aug 25, 2016 at 01:43:53PM -0700, Ian Romanick wrote:
> On 08/24/2016 12:42 PM, Chris Wilson wrote:
> > +#define I915_PARAM_MMAP_GTT_VERSION 40 /* XXX delete me with new libdrm */
> > +   if (intel_get_integer(intelScreen, I915_PARAM_MMAP_GTT_VERSION) >= 1) {
> > +      /* Theorectically unlimited! */
> > +      intelScreen->max_gtt_map_object_size = UINT64_MAX;
> 
> Well... not quite unlimited, right?  Isn't the actual VMA space less
> than 64-bits?  I thought it was more like 48 bits.

The object which can be mapped is unconstrained by the GPU's address
space. We are only about 2MiB of the mappable aperture to map a portion
of the object at a time. The limits upon the object size are that we
have a global 64bit space to allocate GTT offsets from shared by all
objects across all clients. And of course the CPU only has a 40bit
address space, and far less physical memory.

Wrt the interface we can claim to be able to support any object up to
64bits in size. You can apply the same argument as for the mappable
aperture and halve/quarter it so that you can fit more than one huge
object at a time - but if you are planning to allocate lots of 60+bit
objects, we will run out of virtual offsets for them rapidly.

A safe value then would be 1<<64 / <typical number of objects>, so say
1<<48. By the time we can allocate more than one 1<<48 object, I
anticipate we will have moved onto 128bit address spaces.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre