[Intel-gfx] [PATCH 23/24] drm/i915: Keep a recent cache of freed contexts objects for reuse

[Chris Wilson]

On Mon, May 22, 2017 at 11:51:27AM +0100, Tvrtko Ursulin wrote:
> 
> On 18/05/2017 10:46, Chris Wilson wrote:
> >Keep the recently freed context objects for reuse. This allows us to use
> >the current GGTT bindings and dma bound pages, avoiding any clflushes as
> >required. We mark the objects as purgeable under memory pressure, and
> >reap the list of freed objects as soon as the device is idle.
> 
> One additional thought - how about making the batch pool more
> generic, or more precisely extracting a layer from it to be called
> object pool, and then having batch pool and context pool on top of
> that?
> 
> There would be some details to flesh out, like do we want to
> strictly split internal from shmemfs backed objects, or can allow
> batch pool to get either, the API and similar.

That was the big difference; different object types that shouldn't be
mixed. So you going to have different lists for the the different
classes and behaviours.

On top of this, there's also the question of whether to keep old objects
around for userspace. Userspace caching is still more efficient for
userspace, but there are classes of objects that are cycled between
processes and discarded instead of being recycled in userspace. The
tricky part is making sure that the object state, after alloc from
cache, doesn't surprise the user.

> But if it could be done with not a lot of code it may be
> preferential to have one implementation of the similar basic idea.

No denying that, just not enough in common yet and still mostly
searching for the problem to be solved. (Outside of unrealistic
benchmarks and test cases.)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre