[Intel-gfx] [PATCH] drm/i915: Treat WC a separate cache domain

[Chris Wilson]

On Fri, Apr 07, 2017 at 10:16:52AM +0300, Joonas Lahtinen wrote:
> On ke, 2017-04-05 at 22:07 +0100, Chris Wilson wrote:
> > When discussing a new WC mmap, we based the interface upon the
> > assumption that GTT was fully coherent. How naive! Commits 3b5724d702ef
> > ("drm/i915: Wait for writes through the GTT to land before reading
> > back") and ed4596ea992d ("drm/i915/guc: WA to address the Ringbuffer
> > coherency issue") demonstrate that writes through the GTT are indeed
> > delayed and may be overtaken by direct WC access. To be safe, if
> > userspace is mixing WC mmaps with other potential GTT access (pwrite,
> > GTT mmaps) it should use set_domain(WC).
> > 
> > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=96563
> > Testcase: igt/gem_pwrite/small-gtt*
> > Testcase: igt/drv_selftest/coherency
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> 
> <SNIP>
> 
> > +int
> > +i915_gem_object_set_to_wc_domain(struct drm_i915_gem_object *obj, bool write)
> > +{
> > +	int ret;
> > +
> > +	lockdep_assert_held(&obj->base.dev->struct_mutex);
> > +
> > > +	ret = i915_gem_object_wait(obj,
> > > +				   I915_WAIT_INTERRUPTIBLE |
> > > +				   I915_WAIT_LOCKED |
> > +				   (write ? I915_WAIT_ALL : 0),
> 
> Could construct into flags variable.
> 
> > +	/* Flush and acquire obj->pages so that we are coherent through
> > +	 * direct access in memory with previous cached writes through
> > +	 * shmemfs and that our cache domain tracking remains valid.
> > +	 * For example, if the obj->filp was moved to swap without us
> > +	 * being notified and releasing the pages, we would mistakenly
> > +	 * continue to assume that the obj remained out of the CPU cached
> > +	 * domain.
> > +	 */
> > +	ret = i915_gem_object_pin_pages(obj);
> > +	if (ret)
> > +		return ret;
> > +
> > +	i915_gem_object_flush_cpu_write_domain(obj);
> > +	i915_gem_object_flush_gtt_write_domain(obj);
> 
> I was thinking if WC should be a "substate" of CPU domain, because we
> don't have i915_gem_object_flush_wc_write_domain.

There isn't any to do in i915_gem_object_flush_wc_write_domain()...
Other than a wmb. I guess for completeness that should be added.
Probably via a flush_other_write_domain().
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre