[Intel-gfx] [PATCH 5/5] drm/i915: Use partial view in mmap fault handler

[Joonas Lahtinen]

On ma, 2015-04-27 at 12:21 +0100, Chris Wilson wrote:
> On Mon, Apr 27, 2015 at 02:01:59PM +0300, Joonas Lahtinen wrote:
> > On pe, 2015-04-24 at 13:33 +0100, Chris Wilson wrote:
> > > On Fri, Apr 24, 2015 at 03:10:20PM +0300, Joonas Lahtinen wrote:
> > > > Use partial view for huge BOs (bigger than half the mappable aperture)
> > > > in fault handler so that they can be accessed withough trying to make
> > > > room for them by evicting other objects.
> > > > 
> > > > Signed-off-by: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> > > > ---
> > > >  drivers/gpu/drm/i915/i915_gem.c |   67 ++++++++++++++++++++++++++-------------
> > > >  1 file changed, 45 insertions(+), 22 deletions(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> > > > index c2c1819..eb30cee 100644
> > > > --- a/drivers/gpu/drm/i915/i915_gem.c
> > > > +++ b/drivers/gpu/drm/i915/i915_gem.c
> > > > @@ -1635,6 +1635,7 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
> > > >  	struct drm_i915_gem_object *obj = to_intel_bo(vma->vm_private_data);
> > > >  	struct drm_device *dev = obj->base.dev;
> > > >  	struct drm_i915_private *dev_priv = dev->dev_private;
> > > > +	struct i915_ggtt_view view = i915_ggtt_view_normal;
> > > >  	pgoff_t page_offset;
> > > >  	unsigned long pfn;
> > > >  	int ret = 0;
> > > > @@ -1667,8 +1668,21 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
> > > >  		goto unlock;
> > > >  	}
> > > >  
> > > > -	/* Now bind it into the GTT if needed */
> > > > -	ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_MAPPABLE);
> > > > +	/* Use a partial view if the object is bigger than half the aperture. */
> > > > +	if (obj->base.size > dev_priv->gtt.mappable_end/2) {
> > > > +		static const size_t chunk_size = 256; // 1 MiB
> > > > +		memset(&view, 0, sizeof(view));
> > > > +		view.type = I915_GGTT_VIEW_PARTIAL;
> > > > +		view.params.partial.offset = rounddown(page_offset, chunk_size);
> > > > +		view.params.partial.size =
> > > > +			min_t(size_t,
> > > > +			      chunk_size,
> > > > +			      (vma->vm_end - vma->vm_start)/PAGE_SIZE -
> > > > +			      view.params.partial.offset);
> > > 
> > > This isn't what I was imagining.
> > > 
> > > I was expecting to see error handling inside the fault path if we could
> > > not pin the object. This way we could handle fragmentation or display
> > > objects pinned outside the aperture.
> > 
> > After discussion with Daniel, the idea was dropped due to high amount of
> > trashing which would occur if each object would be attempted to fit to
> > the mappable aperture for each fault to that object.
> 
> The point is that we fail to install a partial view for pinned objects
> outside of the aperture. Or did I miss how you handle them?
> 

That is true.

By changing the comparison to be against full aperture size, then the
patch will only bring new functionality to handle the cases when the
object was actually impossible to map previously (and was early
rejected).

I'd prefer to have this version in first (with change of removing the /2
from aperture size comparison), to get some feedback from the XenGT team
about the usability of it (speed-wise).

I will extend the series with fence calculation for tield partial view,
so I would consider this in the rework. Does that sound good?

> I would suggest try first with a PIN_MAPPABLE | PIN_NOEVICT then. You
> can limit thrashing by a vma->release callback to only zap the PTE
> mapping into that vma. Also remember to destroy all of the partial vma
> when we zap the mapping for an object.
> 
> > An alternate path could be added to mark an object partially_mappable if
> > mapping it normally fails, then further attempts could be skipped to
> > avoid the trashing. But it is still questionable if an object close to
> > the whole aperture size should be attempted to map normally, which will
> > cause huge trashing if there are plenty of objects in the mappable
> > aperture (that problem existed before, if the object was just short of
> > the aperture size, it was attempted to be mapped).
> 
> Just treat the aperture as first come first served, with perhaps a
> reservation of the first 256k for transient partial vma. Fix thrashing
> when we a usecase that falls foul of the page-fault-of-doom, but first
> try to write one. I think with a mru list of individually faulted vma
> thrashing will be limited (i.e. no worse than current mru eviction and
> possibly better albeit at a higher pagefault cost).
> 

Ack.

> Also I found a usecase for temporarily mapping invidual pages:
> http://cgit.freedesktop.org/~ickle/linux-2.6/commit/?h=bug90137&id=40e8a9cddb4a1e39a2fdcff4878b61c666e1a51e

I have not peeked into execbuffers much, but looks quite
straightforward. Could use the partial views code if desired. Not sure
how effective it is for single pages.

Regards, Joonas

> -Chris
>