[Intel-gfx] [PATCH 03/19] drm/i915: Micro-optimise gen8_ppgtt_insert_entries()

[Chris Wilson]

On Thu, Feb 02, 2017 at 04:39:49PM +0000, Tvrtko Ursulin wrote:
> 
> On 02/02/2017 16:10, Chris Wilson wrote:
> >On Thu, Feb 02, 2017 at 03:57:43PM +0000, Tvrtko Ursulin wrote:
> >>
> >>On 02/02/2017 15:32, Chris Wilson wrote:
> >>>On Thu, Feb 02, 2017 at 03:02:32PM +0000, Chris Wilson wrote:
> >>>>Improve the sg iteration and in hte process eliminate a bug in
> >>>>miscomputing the pml4 length as orig_nents<<PAGE_SHIFT is no longer the
> >>>>full length of the sg table.
> >>>>
> >>>
> >>>which fixes a corner case of 0c40ce130e38
> >>>Fixes: 0c40ce130e38 ("drm/i915: Trim the object sg table")
> >>
> >>What do you mean? oring_nents is definitely the full length of the
> >>sg table, especially after i915_sg_trim. Before it orig_nents was
> >>often larger than the real length of the sg table.
> >
> >The code is using orig_nents as obj->base.size/vma->size (a page count,
> >not the sg count), if I read it correctly as it is computing the address
> >range.
> 
> Oh right, I was misled by the commit message ("is not longer the
> full length of the sg table"). Nasty. That means it was broken for
> userptr objects as well.

And partial, and... pretty everything at some point.

Hmm, maybe it is the recent spat of NULL deref for gen8 ppgtt. I'd been
assuming that they had a similar unknown cause to the gen6 which have
been around for yonks.

Tart up the commit message and this could 4.10-rc material.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre