[Mesa-dev] [PATCH 25/53] i965/drm: Drop intel_chipset.h in favor of using gen_device_info.

[Chris Wilson]

On Wed, Apr 05, 2017 at 02:36:38PM -0700, Kenneth Graunke wrote:
> On Wednesday, April 5, 2017 3:33:39 AM PDT Chris Wilson wrote:
> > On Tue, Apr 04, 2017 at 05:10:15PM -0700, Kenneth Graunke wrote:
> > > This moves the PCI ID detection to intel_screen.c and makes
> > > drm_bacon_bufmgr_gem_init() take a devinfo pointer.
> > > 
> > > We also drop the HAS_LLC query stuff - devinfo has that info already,
> > > without kernel queries, and it makes no sense to have two has_llc flags
> > > set by different mechanisms.
> > 
> > Just remember the kernel query actually tells you something more than
> > what you know from just querying pci-id. As you replace the buffer cache
> > management, that will become more obvious (and directly controllable).
> 
> It does?  I must be missing something - it sure looks like the kernel
> sets dev_priv->info.has_llc based on the PCI ID, and this getparam
> just returns that that value.
> 
> I'd be happy to set our devinfo->has_llc based on the kernel query,
> if it's actually useful...but it looks kinda pointless...

What I was hinting at was that being on a llc architecture is one thing,
and knowing the actual coherency of the buffer is another. Mostly just
another warning about handling scanouts carefully, and it is something
should definitely be taken into consideration when importing a buffer as
the third party may be using the buffer incoherently even if llc. (And
for kernels older than v3.6 you would find the buffers were by default
being treated as incoherent by the kernel so just assuming they were llc
could cause excess clflushing. The same still applies, just by default
everything is now treated as coherent on llc.)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre