[Intel-gfx] [PATCH 10/10] drm/i915: Use MI_BATCH_BUFFER_START on 830/845

[Ville Syrjälä]

On Tue, Dec 15, 2015 at 10:24:13AM +0000, Chris Wilson wrote:
> On Tue, Dec 15, 2015 at 10:09:30AM +0000, Chris Wilson wrote:
> > On Mon, Dec 14, 2015 at 07:25:13PM +0200, Ville Syrjälä wrote:
> > > On Mon, Dec 14, 2015 at 04:58:54PM +0000, Chris Wilson wrote:
> > > > On Mon, Dec 14, 2015 at 06:23:49PM +0200, ville.syrjala at linux.intel.com wrote:
> > > > > From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > > > 
> > > > > MI_BATCH_BUFFER is nasty since it requires that userspace pass in the
> > > > > correct batch length.
> > > > > 
> > > > > Let's switch to using MI_BATCH_BUFFER_START instead (like we do on
> > > > > other platforms). Then we don't have to specify the batch length
> > > > > at all, and the CS will instead execute until it sees the
> > > > > MI_BATCH_BUFFER_END.
> > > > 
> > > > Oh. My. Gosh. There's a BB_START?!!!
> > > 
> > > Looks like ;) At least my 830 seems perfectly happy with it. Well,
> > > as happy as it's ever been. Though I still couldn't get it complete
> > > a piglit run without hanging last I tried.
> > > 
> > > > 
> > > > > We still need the batch length since we do the CS TLB workaround
> > > > > and copy the batch into the permanently pinned scratch object
> > > > > and execute it from there. But for this we can simply use the
> > > > > batch object length when the user hasn't specified the actual
> > > > > batch length. So specifying the batch length becomes just a
> > > > > way to optimize the batch copy a little bit.
> > > > > 
> > > > > We lost batch_len from a bunch of igts (including the quiesce batch)
> > > > > so without this igt is utterly broken on 830/845. Also some igts such
> > > > > as gem_cpu_reloc never specified the batch_len and so didn't work.
> > > > > With MI_BATCH_BUFFER_START we don't have to fix up igt every time
> > > > > someone forgets that 830/845 exist.
> > > > > 
> > > > > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > > > > Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > > 
> > > > Looks sane.
> > 
> > Checked against bspec, and it does indeed list BB_START for 830/845 and
> > we already comply with the errata.

Yeah, the coloring stuff should be enough.

> 
> The other question, does this obsolete our work around? Can I be that
> optimisitic?

The CS TLB one? I think I tried it at some point, and things till
failed. But stuff fails even with the w/a (like I said piglit will
hang the GPU eventually), so I can't be sure that I actually tested
the CS TLB fail. I think I need to retest at some point.

As far as the docs go, I only remember it mentioning some TLB fail
affecting the blitter. I guess the CS TLB fail isn't actually
documented anywhere?

-- 
Ville Syrjälä
Intel OTC