Re: [PATCH 11/11] [v4] drm/i915/bdw: Ensure a context is loaded before RC6

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, Mar 04, 2014 at 03:30:14PM +0100, Daniel Vetter wrote:
> On Wed, Feb 19, 2014 at 10:27:20PM -0800, Ben Widawsky wrote:
> > RC6 works a lot like HW contexts in that when the GPU enters RC6 it
> > saves away the state to a context, and loads it upon wake.
> > 
> > It's to be somewhat expected that BIOS will not set up valid GPU state.
> > As a result, if loading bad state can cause the GPU to get angry, it
> > would make sense then that we need to load state first. There are two
> > ways in which we can do this:
> > 
> > 1. Create 3d state in the driver, load it up, then enable RC6.
> > 1b. Reuse a known good state, [and if needed,] just bind objects where
> > needed. Then enable RC6.
> > 2. Hold off enabling RC6 until userspace has had a chance to complete
> > batches.
> > 
> > There has been discussions in the past with #1 as it has been
> > recommended for fixes elsewhere. I'm not opposed to it, I'd just like to
> > do the easy thing now to enable the platform.
> > 
> > This patch is a hack that implement option #2. It will defer enabling
> > rc6 until the first batch from userspace has been retired. It suffers
> > two flaws. The first is, if the driver is loaded, but a batch is not
> > submitted/completed, we'll never enter rc6. The other is, it expects
> > userspace to submit a batch with 3d state first. Both of these things
> > are not actual flaws for most users because most users will boot to a
> > graphical composited desktop. Both mesa, and X will always emit the
> > necessary 3d state.
> > 
> > Once a context is loaded and we enable rc6, the default context should
> > inherit the proper state because we always inhibit the restore for the
> > default context. This assumes certain things about the workaround/issue
> > itself to which I am not privy (primarily that the indirect state
> > objects don't actually need to exist).
> > 
> > With that, there are currently 4 options for BDW:
> > 1. Don't use RC6.
> > 2. Use RC6 and expect a hang on the first batch submitted for every
> > context.
> > 3. Use RC6 and use this patch.
> > 4. Wait for another workaround implementation.
> > 
> > NOTE: This patch could be used against other platforms as well.
> > 
> > v2: Re-add accidentally dropped hunk (Ben)
> > 
> > v3: Now more compilable (Ben)
> > 
> > v4: Use the existing enable flag for rc6. This will also make the
> > suspend/resume case work properly, which is broken in v3.
> > Disable rc6 on reset, and defer re-enabling until the first batch.
> > 
> > The fact that RC6 residency continues to increment, and that this patch
> > prevents a hang on BDW silicon has been:
> > Tested-by: Kenneth Graunke <kenneth@xxxxxxxxxxxxx> (v1)
> > 
> > Cc: David E. Box <david.e.box@xxxxxxxxx>
> > Cc: Kristen Carlson Accardi <kristen@xxxxxxxxxxxxxxx>
> > Signed-off-by: Ben Widawsky <ben@xxxxxxxxxxxx>
> > 
> > squash! drm/i915/bdw: Ensure a context is loaded before RC6
> > ---
> >  drivers/gpu/drm/i915/i915_drv.c      |  4 +++-
> >  drivers/gpu/drm/i915/i915_gem.c      | 10 ++++++++++
> >  drivers/gpu/drm/i915/intel_display.c |  5 +++++
> >  3 files changed, 18 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> > index 2d05d7c..7fdfc0e 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.c
> > +++ b/drivers/gpu/drm/i915/i915_drv.c
> > @@ -679,6 +679,8 @@ int i915_reset(struct drm_device *dev)
> >  	mutex_lock(&dev->struct_mutex);
> >  
> >  	i915_gem_reset(dev);
> > +	if (IS_BROADWELL(dev))
> > +		intel_disable_gt_powersave(dev);
> >  
> >  	simulated = dev_priv->gpu_error.stop_rings != 0;
> >  
> > @@ -733,7 +735,7 @@ int i915_reset(struct drm_device *dev)
> >  		 * reset and the re-install of drm irq. Skip for ironlake per
> >  		 * previous concerns that it doesn't respond well to some forms
> >  		 * of re-init after reset. */
> > -		if (INTEL_INFO(dev)->gen > 5) {
> > +		if (INTEL_INFO(dev)->gen > 5 && !IS_BROADWELL(dev)) {
> >  			mutex_lock(&dev->struct_mutex);
> >  			intel_enable_gt_powersave(dev);
> >  			mutex_unlock(&dev->struct_mutex);
> > diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> > index 3618bb0..25a97a6 100644
> > --- a/drivers/gpu/drm/i915/i915_gem.c
> > +++ b/drivers/gpu/drm/i915/i915_gem.c
> > @@ -2420,6 +2420,7 @@ void i915_gem_reset(struct drm_device *dev)
> >  void
> >  i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
> >  {
> > +	struct drm_i915_private *dev_priv = ring->dev->dev_private;
> >  	uint32_t seqno;
> >  
> >  	if (list_empty(&ring->request_list))
> > @@ -2443,6 +2444,15 @@ i915_gem_retire_requests_ring(struct intel_ring_buffer *ring)
> >  		if (!i915_seqno_passed(seqno, obj->last_read_seqno))
> >  			break;
> >  
> > +		/* Wa: can't find the w/a name.
> > +		 * This doesn't actually implement the w/a, but it a workaround
> > +		 * for the workaround. It defers using rc6 until we know valid
> > +		 * state exists.
> > +		 */
> > +		if (IS_BROADWELL(ring->dev) && intel_enable_rc6(ring->dev) &&
> > +		    !dev_priv->rps.enabled && ring->id == RCS)
> > +			intel_enable_gt_powersave(ring->dev);
> 
> Nope, we won't rely on userspace submitting a command to enable a power
> saving feature. There's a good chance that userspace won't submitting
> anything, but still wake up from suspend, e.g. for background mail
> checking. Especially on Android where opportunistic is used to
> aggressively conserve battery power.

Your suspend example with mail is also not great. So the thing wakes up
for a second to check mail, and GPU is not in RC6, then goes back to
sleep. Big deal. If a user sees a new mail and uses the mail app, she
will get RC6. Essentially on any composited windowing system, that the
user is actively interfacing with, you get RC6 (and deeper states).

> 
> If just a MI_SET_CONTEXT isn't good enough and we need to save a valid 3d
> context then we need to (shock, horror) emit the relevant minimal set of
> 3d state commands in the kernel before enabling rc6.
> 
> Aside: Series looks good otherwise, so I'll happily merge once it's
> reviewed. No opinion on the RP* bikeshed (besides that your new names for
> dev_priv->rps.foo are definitely clearer) as long as the connection
> between the sane names and the Bspec names is documented somewhere.
> -Daniel
> 

I can't say it's completely unexpected that this would be your response,
but I do feel like you've ignored my argument that this is better than
the current situation. Not merging this patch only keeps things bad.

So I'd like you to re-consider merging this patch instead of waiting for
the perfect solution. This patch requires a lot less review than the
real fix. It has been tested by several people (I can ask them to put
their reviewed by on it). It enables rc6 for people today, and this
includes pc7, and deeper package and C states. It's very easy to revert
if/when we get a real fix. Real users benefit from this patch. Real
users are not hurt by this patch because if userspace is submitting bad
state setup, they'll have the same or worse experience than failing RC6.

As an aside, this needs to come before I enable rc6 anyway. So the order
way bad.

> > +
> >  		i915_gem_object_move_to_inactive(obj);
> >  	}
> >  
> > diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> > index f19e6ea..72c8e1d 100644
> > --- a/drivers/gpu/drm/i915/intel_display.c
> > +++ b/drivers/gpu/drm/i915/intel_display.c
> > @@ -10983,6 +10983,11 @@ void intel_modeset_init_hw(struct drm_device *dev)
> >  
> >  	intel_reset_dpio(dev);
> >  
> > +	if (IS_BROADWELL(dev)) {
> > +		DRM_DEBUG_DRIVER("Deferring RC6 enabling until first batch is complete\n");
> > +		return;
> > +	}
> > +
> >  	mutex_lock(&dev->struct_mutex);
> >  	intel_enable_gt_powersave(dev);
> >  	mutex_unlock(&dev->struct_mutex);
> > -- 
> > 1.9.0
> > 
> > _______________________________________________
> > Intel-gfx mailing list
> > Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
> > http://lists.freedesktop.org/mailman/listinfo/intel-gfx
> 
> -- 
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch

-- 
Ben Widawsky, Intel Open Source Technology Center
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx




[Index of Archives]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]
  Powered by Linux