[Intel-gfx] [PATCH 2/8] drm/i915: Adds graphic address space ballooning logic

[Daniel Vetter]

On Tue, Sep 23, 2014 at 10:19:02AM +0100, Chris Wilson wrote:
> On Tue, Sep 23, 2014 at 10:26:26AM +0200, Daniel Vetter wrote:
> > On Fri, Sep 19, 2014 at 09:00:00PM +0100, Chris Wilson wrote:
> > > On Fri, Sep 19, 2014 at 06:21:46PM +0000, Tian, Kevin wrote:
> > > > > From: Chris Wilson
> > > > > The implementation also looks backwards. To work correctly with the GTT
> > > > > allocator, you need to preallocate the reserved space such that it can
> > > > > only allocate from the allowed ranges. Similarly, it should evict any
> > > > > conflicting nodes when deballooning.
> > > > 
> > > > Could you elaborate a bit for above suggestion?
> > > 
> > > My expectation was that the dev_priv->gtt.base.vm would contain exactly
> > > two holes after setup (in the mappable and non-mappable range). To do
> > > that you would explicitly reserve everything barred from this client 
> > > using a set of drm_mm_reserve_node()
> > 
> > Essentially a reserve_node implements what you open-code with
> > insert_node_range right now.
> 
> Heh, there is a big difference. One inserts exactly where you ask and
> fails if it conflicts, the other inserts where it feels like within that
> range.

Well if the the requested size matches the range exactly then it will be
the same. Which iirc is what's going on here I think.

> > One issue aside with both this and with the PDE reservations for gen7 is
> > that there are now other thins in the ggtt drm_mm allocator than just gem
> > objects. Which means our debugfs files are now less useful.
> > 
> > It might be useful to augment that dumper with one that dumps everything.
> > We could add a few bits of driver-private tags in drm_mm_node (there's
> > space) to figure out what kind of object it is. Would be a great follow-up
> > task.
> 
> I think moving the other way and making them all objects so that we can
> tie them into evection and the shrinker, use more interesting allocation
> strategies, improve integration with debugging etc.

Hm, not sure yet since it will be a lot of work at least. But I guess we
could untangle the meaning of obj->pin a bit and add an unbind vfunc which
adds some magic. But there's a lot of stuff attached to a gem bo that just
doesn't make a lot of sense really, so maybe a better option would be to
subclass a struct i915_ggtt_vma with special magic. Dunno really.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch