[Intel-gfx] [PATCH] drm/agp: agp-intel/i915: trim stolen space to 16M

[Daniel Vetter]

On Wed, Mar 24, 2010 at 01:29:08PM -0700, Jesse Barnes wrote:
> On Wed, 24 Mar 2010 21:25:07 +0100
> Daniel Vetter <daniel at ffwll.ch> wrote:
> > Fixing up the interaction between the gtt driver in intel-agp and the i915
> > drm modules is my current pet project. We can't remove the gtt support
> > without breaking old non-gem userspace code, though. So my idea is to
> > create a private, direct interface to the gtt for gem and implement the
> > agp stuff on top of it. First step in this direction is my patch-pile to
> > seperate the gtt support code from the agp driver I've sent a few days
> > ago.
> > 
> > Does this sound like a reasonable plan to go forward? Obviously my patches
> > conflict with all recent patches against intel-agp. I've you'd ack the
> > general direction I could rebase these patches at a convenient time and
> > then start with clean-ups and killing duplicate code when you've merged
> > them.
> 
> That sounds like a good approach.  Basically the current AGP interfaces
> would be deprecated unless someone was using the old GART ioctls?

My current plan (probably needs some revising on the go) is to extract a
small interface from the current fake-agp gtt code:
- Probe function to set up/detect stolen space/gtt space/other reserved
  stuff.
- Function to insert a PTE.
- Chipset flush funtion.
Then use these functions directly from gem. As soon as gem takes (also)
care of initializing the gtt we could start dropping the old agp interface
(by simply not probing from the agp driver), at least for the chipsets
that never had userspace modesetting (ironlake and later).

Iff we ever drop support for old userspace, we could then just rip out the
fake agp driver and move intel-gtt.c from char/agp to gpu/drm/i915.
Problem is that only 2.11 doesn't use the gart ioctls anymore, if you
include xvmc (my recent cleanup killed them).

-Daniel
-- 
Daniel Vetter
Mail: daniel at ffwll.ch
Mobile: +41 (0)79 365 57 48