[Intel-gfx] [PATCH 3/4] drm/i915: use ORIGIN_CPU for frontbuffer invalidation on WC mmaps

[Daniel Vetter]

On Mon, Apr 25, 2016 at 09:20:05AM +0100, Chris Wilson wrote:
> On Mon, Apr 25, 2016 at 10:15:11AM +0200, Daniel Vetter wrote:
> > On Mon, Apr 04, 2016 at 06:17:17PM -0300, Paulo Zanoni wrote:
> > > From: Chris Wilson <chris at chris-wilson.co.uk>
> > > 
> > > ... instead of the previous ORIGIN_GTT. This should actually
> > > invalidate FBC once something is written on the frontbuffer using WC
> > > mmaps. The problem with ORIGIN_GTT is that the automatic hardware
> > > tracking is not able to detect the WC writes as it can detect the GTT
> > > writes.
> > > 
> > > This should help fix the SKL bug where nothing happens when you type
> > > your username/password on lightdm.
> > > 
> > > This patch was originally pasted on an email by Chris and converted to
> > > an actual git patch by Paulo.
> > > 
> > > Cc: <stable at vger.kernel.org> # 4.6
> > > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > 
> > This is From: Chris but lacks his sob. Can't merge as-is, pls ask him for
> > his sob on irc.
> > 
> > > Signed-off-by: Paulo Zanoni <paulo.r.zanoni at intel.com>
> > > ---
> > >  drivers/gpu/drm/i915/i915_drv.h |  1 +
> > >  drivers/gpu/drm/i915/i915_gem.c | 14 +++++++++++---
> > >  2 files changed, 12 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > > index dd18772..17b42a8 100644
> > > --- a/drivers/gpu/drm/i915/i915_drv.h
> > > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > > @@ -2157,6 +2157,7 @@ struct drm_i915_gem_object {
> > >  	unsigned int cache_dirty:1;
> > >  
> > >  	unsigned int frontbuffer_bits:INTEL_FRONTBUFFER_BITS;
> > > +	unsigned int has_wc_mmap:1;
> > >  
> > >  	unsigned int pin_display;
> > >  
> > > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> > > index 40f90c7..bfafa07 100644
> > > --- a/drivers/gpu/drm/i915/i915_gem.c
> > > +++ b/drivers/gpu/drm/i915/i915_gem.c
> > > @@ -1608,6 +1608,13 @@ static struct intel_rps_client *to_rps_client(struct drm_file *file)
> > >  	return &fpriv->rps;
> > >  }
> > >  
> > > +static enum fb_op_origin
> > > +write_origin(struct drm_i915_gem_object *obj, unsigned domain)
> > > +{
> > > +	return domain == I915_GEM_DOMAIN_GTT && !obj->has_wc_mmap ?
> > > +	       ORIGIN_GTT : ORIGIN_CPU;
> > > +}
> > > +
> > >  /**
> > >   * Called when user space prepares to use an object with the CPU, either
> > >   * through the mmap ioctl's mapping or a GTT mapping.
> > > @@ -1661,9 +1668,7 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
> > >  		ret = i915_gem_object_set_to_cpu_domain(obj, write_domain != 0);
> > >  
> > >  	if (write_domain != 0)
> > > -		intel_fb_obj_invalidate(obj,
> > > -					write_domain == I915_GEM_DOMAIN_GTT ?
> > > -					ORIGIN_GTT : ORIGIN_CPU);
> > > +		intel_fb_obj_invalidate(obj, write_origin(obj, write_domain));
> > >  
> > >  unref:
> > >  	drm_gem_object_unreference(&obj->base);
> > > @@ -1761,6 +1766,9 @@ i915_gem_mmap_ioctl(struct drm_device *dev, void *data,
> > >  		else
> > >  			addr = -ENOMEM;
> > >  		up_write(&mm->mmap_sem);
> > > +
> > > +		/* This may race, but that's ok, it only gets set */
> > 
> > This comment doesn't mesh with the
> > 
> > 	unsigned int has_wc_mmap:1;
> > 
> > above. If you depend upon atomicity of writes at the hw level, your struct
> > member _must_ be a full machine word. Which in practice means
> > 
> > 	/*
> > 	 * IMPORTANT: We have unlocked access to this, this must be a
> > 	 * stand-alone bool.
> > 	 */
> > 	bool has_wc_mmap;
> > 
> > You can only fold together different bitfields into one if they're _all_
> > protected by the same lock. gcc needs to generate read-modify-write cycles
> > to write them, which means if any of them are accessed through a separate
> > lock or lock-lessly, there's a real race.
> > 
> > With that fixed (and assuming the ABI fix has still Chris' ack/sob):
> 
> Yup, we can do better and use WRITE_ONCE() as well. That would make GCC
> complain about it being bogus.

+1 for WRITE_ONCE(). Great suggestion, totally forgot about it.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch