Re: [PATCH 1/3] drm/i915: Check for a stalled page flip after each vblank

["Ville Syrjälä" <ville.syrjala@linux.intel.com>]

On Fri, May 30, 2014 at 05:16:34PM +0100, Chris Wilson wrote:
> Long ago, back in the racy haydays of 915gm interrupt handling, page
> flips would occasionally go astray and leave the hardware stuck, and the
> display not updating. This annoyed people who relied on their systems
> being able to display continuously updating information 24/7, and so
> some code to detect when the driver missed the page flip completion
> signal was added. Until recently, it was presumed that the interrupt
> handling was now flawless, but once again Simon Farnsworth has found a
> system whose display will stall. Reinstate the pageflip stall detection,
> which works by checking to see if the hardware has been updated to the
> new framebuffer address following each vblank. If the hardware is
> scanning out from the new framebuffer, but we still think the flip is
> pending, then we kick our driver into submision.
> 
> This is a continuation of the effort started with
> commit 4e5359cd053bfb7d8dabe4a63624a5726848ffbc
> Author: Simon Farnsworth <simon.farnsworth@onelan.co.uk>
> Date:   Wed Sep 1 17:47:52 2010 +0100
> 
>     drm/i915: Avoid pageflipping freeze when we miss the flip prepare interrupt
> 
> This now includes a belt-and-braces approach to make sure the driver
> (or the hardware) doesn't miss an interrupt and cause us to stop
> updating the display should the unthinkable happen and the pageflip fail - i.e.
> that the user is able to continue submitting flips.
> 
> Reported-by: Simon Farnsworth <simon@farnz.org.uk>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=75502
> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/i915_debugfs.c  |  33 ++++++++---
>  drivers/gpu/drm/i915/i915_irq.c      |  85 ++++++++-------------------
>  drivers/gpu/drm/i915/intel_display.c | 109 ++++++++++++++++++++++++++++++-----
>  drivers/gpu/drm/i915/intel_drv.h     |   2 +
>  4 files changed, 144 insertions(+), 85 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
> index 0b063c03d188..a05a33ab4b33 100644
> --- a/drivers/gpu/drm/i915/i915_debugfs.c
> +++ b/drivers/gpu/drm/i915/i915_debugfs.c
> @@ -581,6 +581,7 @@ static int i915_gem_pageflip_info(struct seq_file *m, void *data)
>  {
>  	struct drm_info_node *node = m->private;
>  	struct drm_device *dev = node->minor->dev;
> +	struct drm_i915_private *dev_priv = dev->dev_private;
>  	unsigned long flags;
>  	struct intel_crtc *crtc;
>  
> @@ -602,23 +603,37 @@ static int i915_gem_pageflip_info(struct seq_file *m, void *data)
>  				seq_printf(m, "Flip pending (waiting for vsync) on pipe %c (plane %c)\n",
>  					   pipe, plane);
>  			}
> +			seq_printf(m, "Flip queued on frame %d, now %d\n",
> +				   work->sbc, atomic_read(&dev->vblank[crtc->pipe].count));
>  			if (work->enable_stall_check)
>  				seq_puts(m, "Stall check enabled, ");
>  			else
>  				seq_puts(m, "Stall check waiting for page flip ioctl, ");
>  			seq_printf(m, "%d prepares\n", atomic_read(&work->pending));
>  
> -			if (work->old_fb_obj) {
> -				struct drm_i915_gem_object *obj = work->old_fb_obj;
> -				if (obj)
> -					seq_printf(m, "Old framebuffer gtt_offset 0x%08lx\n",
> -						   i915_gem_obj_ggtt_offset(obj));
> +			{
> +				u32 addr;
> +
> +				if (INTEL_INFO(dev)->gen >= 4)
> +					addr = DSPSURF(crtc->plane);
> +				else
> +					addr = DSPADDR(crtc->plane);
> +
> +				seq_printf(m, "Current scanout address 0x%08x\n", 
> +					   I915_READ(addr));

"current" makes me think of the live address. But right now I can't think of a
better term to use there.

>  			}

I don't like the extra {} block. Such things always give me an
impression that this is a debug hack someone forgot to remove.

> +
>  			if (work->pending_flip_obj) {
> -				struct drm_i915_gem_object *obj = work->pending_flip_obj;
> -				if (obj)
> -					seq_printf(m, "New framebuffer gtt_offset 0x%08lx\n",
> -						   i915_gem_obj_ggtt_offset(obj));
> +				bool complete;
> +
> +				seq_printf(m, "New framebuffer address 0x%08lx\n", (long)work->gtt_offset);
> +
> +				if (INTEL_INFO(dev)->gen >= 4)
> +					complete = I915_HI_DISPBASE(I915_READ(DSPSURF(crtc->plane))) == work->gtt_offset;
> +				else
> +					complete = I915_READ(DSPADDR(crtc->plane)) == work->gtt_offset;
> +
> +				seq_printf(m, "MMIO update completed? %d\n", complete);
>  			}
>  		}
>  		spin_unlock_irqrestore(&dev->event_lock, flags);
> diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> index bce9afbc34bb..3be22794b53c 100644
> --- a/drivers/gpu/drm/i915/i915_irq.c
> +++ b/drivers/gpu/drm/i915/i915_irq.c
> @@ -1802,8 +1802,9 @@ static void valleyview_pipestat_irq_handler(struct drm_device *dev, u32 iir)
>  	spin_unlock(&dev_priv->irq_lock);
>  
>  	for_each_pipe(pipe) {
> -		if (pipe_stats[pipe] & PIPE_START_VBLANK_INTERRUPT_STATUS)
> -			intel_pipe_handle_vblank(dev, pipe);
> +		if (pipe_stats[pipe] & PIPE_START_VBLANK_INTERRUPT_STATUS &&
> +		    intel_pipe_handle_vblank(dev, pipe))
> +			intel_check_page_flip(dev, pipe);

Can't we stick this call into intel_pipe_handle_vblank()?

>  
>  		if (pipe_stats[pipe] & PLANE_FLIP_DONE_INT_STATUS_VLV) {
>  			intel_prepare_page_flip(dev, pipe);
> @@ -2079,8 +2080,9 @@ static void ilk_display_irq_handler(struct drm_device *dev, u32 de_iir)
>  		DRM_ERROR("Poison interrupt\n");
>  
>  	for_each_pipe(pipe) {
> -		if (de_iir & DE_PIPE_VBLANK(pipe))
> -			intel_pipe_handle_vblank(dev, pipe);
> +		if (de_iir & DE_PIPE_VBLANK(pipe) &&
> +		    intel_pipe_handle_vblank(dev, pipe))
> +			intel_check_page_flip(dev, pipe);
>  
>  		if (de_iir & DE_PIPE_FIFO_UNDERRUN(pipe))
>  			if (intel_set_cpu_fifo_underrun_reporting(dev, pipe, false))
> @@ -2129,8 +2131,9 @@ static void ivb_display_irq_handler(struct drm_device *dev, u32 de_iir)
>  		intel_opregion_asle_intr(dev);
>  
>  	for_each_pipe(pipe) {
> -		if (de_iir & (DE_PIPE_VBLANK_IVB(pipe)))
> -			intel_pipe_handle_vblank(dev, pipe);
> +		if (de_iir & (DE_PIPE_VBLANK_IVB(pipe)) &&
> +		    intel_pipe_handle_vblank(dev, pipe))
> +			intel_check_page_flip(dev, pipe);
>  
>  		/* plane/pipes map 1:1 on ilk+ */
>  		if (de_iir & DE_PLANE_FLIP_DONE_IVB(pipe)) {
> @@ -2265,8 +2268,10 @@ static irqreturn_t gen8_irq_handler(int irq, void *arg)
>  			continue;
>  
>  		pipe_iir = I915_READ(GEN8_DE_PIPE_IIR(pipe));
> -		if (pipe_iir & GEN8_PIPE_VBLANK)
> -			intel_pipe_handle_vblank(dev, pipe);
> +
> +		if (pipe_iir & GEN8_PIPE_VBLANK &&
> +		    intel_pipe_handle_vblank(dev, pipe))
> +			intel_check_page_flip(dev, pipe);
>  
>  		if (pipe_iir & GEN8_PIPE_PRIMARY_FLIP_DONE) {
>  			intel_prepare_page_flip(dev, pipe);
> @@ -2575,52 +2580,6 @@ void i915_handle_error(struct drm_device *dev, bool wedged,
>  	schedule_work(&dev_priv->gpu_error.work);
>  }
>  
> -static void __always_unused i915_pageflip_stall_check(struct drm_device *dev, int pipe)
> -{
> -	struct drm_i915_private *dev_priv = dev->dev_private;
> -	struct drm_crtc *crtc = dev_priv->pipe_to_crtc_mapping[pipe];
> -	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
> -	struct drm_i915_gem_object *obj;
> -	struct intel_unpin_work *work;
> -	unsigned long flags;
> -	bool stall_detected;
> -
> -	/* Ignore early vblank irqs */
> -	if (intel_crtc == NULL)
> -		return;
> -
> -	spin_lock_irqsave(&dev->event_lock, flags);
> -	work = intel_crtc->unpin_work;
> -
> -	if (work == NULL ||
> -	    atomic_read(&work->pending) >= INTEL_FLIP_COMPLETE ||
> -	    !work->enable_stall_check) {
> -		/* Either the pending flip IRQ arrived, or we're too early. Don't check */
> -		spin_unlock_irqrestore(&dev->event_lock, flags);
> -		return;
> -	}
> -
> -	/* Potential stall - if we see that the flip has happened, assume a missed interrupt */
> -	obj = work->pending_flip_obj;
> -	if (INTEL_INFO(dev)->gen >= 4) {
> -		int dspsurf = DSPSURF(intel_crtc->plane);
> -		stall_detected = I915_HI_DISPBASE(I915_READ(dspsurf)) ==
> -					i915_gem_obj_ggtt_offset(obj);
> -	} else {
> -		int dspaddr = DSPADDR(intel_crtc->plane);
> -		stall_detected = I915_READ(dspaddr) == (i915_gem_obj_ggtt_offset(obj) +
> -							crtc->y * crtc->primary->fb->pitches[0] +
> -							crtc->x * crtc->primary->fb->bits_per_pixel/8);
> -	}
> -
> -	spin_unlock_irqrestore(&dev->event_lock, flags);
> -
> -	if (stall_detected) {
> -		DRM_DEBUG_DRIVER("Pageflip stall detected\n");
> -		intel_prepare_page_flip(dev, intel_crtc->plane);
> -	}
> -}
> -
>  /* Called from drm generic code, passed 'crtc' which
>   * we use as a pipe index
>   */
> @@ -3726,7 +3685,7 @@ static bool i8xx_handle_vblank(struct drm_device *dev,
>  		return false;
>  
>  	if ((iir & flip_pending) == 0)
> -		return false;
> +		goto check_page_flip;
>  
>  	intel_prepare_page_flip(dev, plane);
>  
> @@ -3737,11 +3696,14 @@ static bool i8xx_handle_vblank(struct drm_device *dev,
>  	 * an interrupt per se, we watch for the change at vblank.
>  	 */
>  	if (I915_READ16(ISR) & flip_pending)
> -		return false;
> +		goto check_page_flip;
>  
>  	intel_finish_page_flip(dev, pipe);
> -
>  	return true;
> +
> +check_page_flip:
> +	intel_check_page_flip(dev, pipe);
> +	return false;

Ah right this guy might not appreciate the check in
intel_pipe_handle_vblank(). Oh well.

>  }
>  
>  static irqreturn_t i8xx_irq_handler(int irq, void *arg)
> @@ -3911,7 +3873,7 @@ static bool i915_handle_vblank(struct drm_device *dev,
>  		return false;
>  
>  	if ((iir & flip_pending) == 0)
> -		return false;
> +		goto check_page_flip;
>  
>  	intel_prepare_page_flip(dev, plane);
>  
> @@ -3922,11 +3884,14 @@ static bool i915_handle_vblank(struct drm_device *dev,
>  	 * an interrupt per se, we watch for the change at vblank.
>  	 */
>  	if (I915_READ(ISR) & flip_pending)
> -		return false;
> +		goto check_page_flip;
>  
>  	intel_finish_page_flip(dev, pipe);
> -
>  	return true;
> +
> +check_page_flip:
> +	intel_check_page_flip(dev, pipe);
> +	return false;
>  }
>  
>  static irqreturn_t i915_irq_handler(int irq, void *arg)
> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> index 318747b444c6..764b277e5937 100644
> --- a/drivers/gpu/drm/i915/intel_display.c
> +++ b/drivers/gpu/drm/i915/intel_display.c
> @@ -3303,6 +3303,22 @@ bool intel_has_pending_fb_unpin(struct drm_device *dev)
>  	return false;
>  }
>  
> +static void page_flip_completed(struct drm_i915_private *dev_priv,
> +				struct intel_crtc *intel_crtc,
> +				struct intel_unpin_work *work)
> +{

assert_spin_locked(event_lock) ?

> +	if (work->event)
> +		drm_send_vblank_event(dev_priv->dev,
> +				      intel_crtc->pipe,
> +				      work->event);
> +
> +	wake_up_all(&dev_priv->pending_flip_queue);
> +	queue_work(dev_priv->wq, &work->work);
> +
> +	trace_i915_flip_complete(intel_crtc->plane,
> +				 work->pending_flip_obj);
> +}
> +
>  void intel_crtc_wait_for_pending_flips(struct drm_crtc *crtc)
>  {
>  	struct drm_device *dev = crtc->dev;
> @@ -8848,21 +8864,12 @@ static void do_intel_finish_page_flip(struct drm_device *dev,
>  
>  	/* and that the unpin work is consistent wrt ->pending. */
>  	smp_rmb();
> -
>  	intel_crtc->unpin_work = NULL;
>  
> -	if (work->event)
> -		drm_send_vblank_event(dev, intel_crtc->pipe, work->event);
> -
> +	page_flip_completed(dev_priv, intel_crtc, work);
>  	drm_crtc_vblank_put(crtc);
>  
>  	spin_unlock_irqrestore(&dev->event_lock, flags);
> -
> -	wake_up_all(&dev_priv->pending_flip_queue);
> -
> -	queue_work(dev_priv->wq, &work->work);
> -
> -	trace_i915_flip_complete(intel_crtc->plane, work->pending_flip_obj);
>  }
>  
>  void intel_finish_page_flip(struct drm_device *dev, int pipe)
> @@ -8940,11 +8947,17 @@ void intel_prepare_page_flip(struct drm_device *dev, int plane)
>  	spin_unlock_irqrestore(&dev->event_lock, flags);
>  }
>  
> +static inline int crtc_sbc(struct intel_crtc *crtc)
> +{
> +	return atomic_read(&crtc->base.dev->vblank[crtc->pipe].count);
> +}

msc?

> +
>  static inline void intel_mark_page_flip_active(struct intel_crtc *intel_crtc)
>  {
>  	/* Ensure that the work item is consistent when activating it ... */
>  	smp_wmb();
>  	atomic_set(&intel_crtc->unpin_work->pending, INTEL_FLIP_PENDING);
> +	intel_crtc->unpin_work->sbc = crtc_sbc(intel_crtc);
>  	/* and that it is marked active as soon as the irq could fire. */
>  	smp_wmb();
>  }
> @@ -9196,6 +9209,58 @@ static int intel_default_queue_flip(struct drm_device *dev,
>  	return -ENODEV;
>  }
>  
> +static bool __intel_pageflip_stall_check(struct drm_device *dev,
> +					 struct drm_crtc *crtc,
> +					 struct intel_unpin_work *work)
> +{
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
> +	struct drm_i915_gem_object *obj;
> +	bool stall_detected;
> +
> +	if (atomic_read(&work->pending) >= INTEL_FLIP_COMPLETE)
> +		return true;
> +
> +	if (!work->enable_stall_check)
> +		return false;
> +
> +	if ((crtc_sbc(intel_crtc) - intel_crtc->unpin_work->sbc) < 3)
> +		return false;
> +
> +	/* Potential stall - if we see that the flip has happened, assume a missed interrupt */
> +	obj = work->pending_flip_obj;
> +	if (INTEL_INFO(dev)->gen >= 4) {
> +		int dspsurf = DSPSURF(intel_crtc->plane);
> +		stall_detected = I915_HI_DISPBASE(I915_READ(dspsurf)) == work->gtt_offset;
> +	} else {
> +		int dspaddr = DSPADDR(intel_crtc->plane);
> +		stall_detected = I915_READ(dspaddr) == work->gtt_offset;
> +	}

Hmm. There's a slight danger of premature flip completion if the user
issues a flip to the same address, and it really takes longer than three
frames to complete. The hardware flip counter could be used to avoid that.
I guess for for pre-ctg there isn't much we can do except hope that flips
never take more than three frames.

> +
> +	return stall_detected;
> +}
> +
> +void intel_check_page_flip(struct drm_device *dev, int pipe)
> +{
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct drm_crtc *crtc = dev_priv->pipe_to_crtc_mapping[pipe];
> +	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
> +	unsigned long flags;
> +
> +	if (crtc == NULL)
> +		return;
> +
> +	spin_lock_irqsave(&dev->event_lock, flags);
> +	if (intel_crtc->unpin_work &&
> +	    __intel_pageflip_stall_check(dev, crtc, intel_crtc->unpin_work)) {
> +		WARN_ONCE(1, "Kicking stuck page flip\n");
> +		drm_vblank_put(dev, intel_crtc->pipe);
> +		page_flip_completed(dev_priv, intel_crtc, intel_crtc->unpin_work);
> +		intel_crtc->unpin_work = NULL;
> +	}
> +	spin_unlock_irqrestore(&dev->event_lock, flags);
> +}
> +
>  static int intel_crtc_page_flip(struct drm_crtc *crtc,
>  				struct drm_framebuffer *fb,
>  				struct drm_pending_vblank_event *event,
> @@ -9243,12 +9308,25 @@ static int intel_crtc_page_flip(struct drm_crtc *crtc,
>  	/* We borrow the event spin lock for protecting unpin_work */
>  	spin_lock_irqsave(&dev->event_lock, flags);
>  	if (intel_crtc->unpin_work) {
> -		spin_unlock_irqrestore(&dev->event_lock, flags);
> -		kfree(work);
> +		bool stall;
> +
> +		/* Before declaring the flip queue wedged, check if
> +		 * the hardware completed the operation behind our backs.
> +		 */
> +		stall = __intel_pageflip_stall_check(dev, crtc,
> +						     intel_crtc->unpin_work);
>  		drm_crtc_vblank_put(crtc);
>  
> -		DRM_DEBUG_DRIVER("flip queue: crtc already busy\n");
> -		return -EBUSY;
> +		if (stall) {
> +			DRM_DEBUG_DRIVER("flip queue: previous flip completed, continuing\n");
> +			page_flip_completed(dev_priv, intel_crtc, intel_crtc->unpin_work);

I was going to say you're leaking the vblank ref here until I saw that
it was dropped earlier. Feels a bit strange to basically drop the ref
you just got for this flip, and take over the ref from the earlier flip.

I'd just move the drm_vblank_put() call into page_flip_completed().

> +		} else {
> +			DRM_DEBUG_DRIVER("flip queue: crtc already busy\n");
> +			spin_unlock_irqrestore(&dev->event_lock, flags);
> +
> +			kfree(work);
> +			return -EBUSY;
> +		}
>  	}
>  	intel_crtc->unpin_work = work;
>  	spin_unlock_irqrestore(&dev->event_lock, flags);
> @@ -9268,8 +9346,6 @@ static int intel_crtc_page_flip(struct drm_crtc *crtc,
>  
>  	work->pending_flip_obj = obj;
>  
> -	work->enable_stall_check = true;
> -
>  	atomic_inc(&intel_crtc->unpin_work_count);
>  	intel_crtc->reset_counter = atomic_read(&dev_priv->gpu_error.reset_counter);
>  
> @@ -9292,6 +9368,7 @@ static int intel_crtc_page_flip(struct drm_crtc *crtc,
>  
>  	work->gtt_offset =
>  		i915_gem_obj_ggtt_offset(obj) + intel_crtc->dspaddr_offset;
> +	work->enable_stall_check = true;
>  
>  	ret = dev_priv->display.queue_flip(dev, crtc, fb, obj, ring, page_flip_flags);
>  	if (ret)
> diff --git a/drivers/gpu/drm/i915/intel_drv.h b/drivers/gpu/drm/i915/intel_drv.h
> index c597b0da93b6..b17c295fe967 100644
> --- a/drivers/gpu/drm/i915/intel_drv.h
> +++ b/drivers/gpu/drm/i915/intel_drv.h
> @@ -619,6 +619,7 @@ struct intel_unpin_work {
>  	u32 flip_count;
>  	u32 gtt_offset;
>  	bool enable_stall_check;
> +	int sbc;
>  };
>  
>  struct intel_set_config {
> @@ -761,6 +762,7 @@ __intel_framebuffer_create(struct drm_device *dev,
>  void intel_prepare_page_flip(struct drm_device *dev, int plane);
>  void intel_finish_page_flip(struct drm_device *dev, int pipe);
>  void intel_finish_page_flip_plane(struct drm_device *dev, int plane);
> +void intel_check_page_flip(struct drm_device *dev, int pipe);
>  struct intel_shared_dpll *intel_crtc_to_shared_dpll(struct intel_crtc *crtc);
>  void assert_shared_dpll(struct drm_i915_private *dev_priv,
>  			struct intel_shared_dpll *pll,
> -- 
> 2.0.0.rc4
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Ville Syrjälä
Intel OTC