[Intel-gfx] [RFC] Introduce dynamic PPAT allocation for GVT-g

[Chris Wilson]

Quoting Zhi Wang (2017-08-14 13:04:34)
> [ BACKGROUND ]
> 
> Due to different guests will have different PPAT configuration under 
> virtualization environment, for example, Windows guest vs Linux guest, a 
> possible idea to handle this open is
> 
> 1)Letting host have the PPAT entries needed by guest

More or less requires static entries forevermore, locking the guest and
host into not changing in future. That doesn't seem a viable assumption.
 
> 2)Mapping PPAT indexes in guest page table to the corresponding host 
> PPAT entries during building the shadow page table.
> 
> To achieve 1), one approach which has been used before is statically 
> adding the collection of guest PPAT configuration into host i915. So 
> when performing 2) GVT-g will always find an available PPAT entry to map.
> 
> Another possible approach is GVT-g could dynamically find/add/delete the 
> PPAT configuration for different guests. To prevent regression of host 
> i915 due to this changes, it’s better to have two types of PPAT entry: 
> one is “static PPAT entry”, which is to keep current host i915 PPAT 
> configuration unchanged, e.g. the mapping between i915 cache level and 
> PPAT index, etc… just keep everything used by host i915 unchanged and a 
> *static* PPAT entry will not be modified.

Seems reasonable, just something like:

int intel_ppat_find(struct drm_i915_private *i915, struct ppat_config *opts);

That can return the index of an existing ppat entry or assign new one
(or error code, or a struct + ERR_PTR). And with a use-counter
(intel_ppat_get/intel_ppat_put, where intel_ppat_find always returns with
the use-counter preincremented) allows for reconfiguration.

The downside is the remapping required, which I expect will not be too
tricky? As you copy the PTE from the guest into the host you just run it
through a translation table, right?
-Chris