diff mbox

[1/2] drm/i915: Replaced Blitter ring based flips with MMIO flips

Message ID 1401261122-23593-2-git-send-email-sourab.gupta@intel.com (mailing list archive)
State New, archived
Headers show

Commit Message

sourab.gupta@intel.com May 28, 2014, 7:12 a.m. UTC 
From: Sourab Gupta <sourab.gupta@intel.com>

Using MMIO based flips on Gen5+. The MMIO flips are useful for the Media power
well residency optimization. These maybe enabled on architectures where
Render and Blitter engines reside in different power wells.
The blitter ring is currently being used just for command streamer based
flip calls. For pure 3D workloads in such cases, with MMIO flips, there will
be no use of blitter ring and this will ensure the 100% residency for Media well.

v2: The MMIO flips now use the interrupt driven mechanism for issuing the
flips when target seqno is reached. (Incorporating Ville's idea)

v3: Rebasing on latest code. Code restructuring after incorporating
Damien's comments

v4: Addressing Ville's review comments
    -general cleanup
    -updating only base addr instead of calling update_primary_plane
    -extending patch for gen5+ platforms

v5: Addressed Ville's review comments
    -Making mmio flip vs cs flip selection based on module parameter
    -Adding check for DRIVER_MODESET feature in notify_ring before calling
     notify mmio flip.
    -Other changes mostly in function arguments

v6: -Having a seperate function to check condition for using mmio flips (Ville)
    -propogating error code from i915_gem_check_olr (Ville)

v7: -Adding __must_check with i915_gem_check_olr (Chris)
    -Renaming mmio_flip_data to mmio_flip (Chris)
    -Rebasing on latest nightly

v8: -Rebasing on latest code
    -squash 3rd patch in series(mmio setbase vs page flip race) with this patch
    -Added new tiling mode update in intel_do_mmio_flip (Chris)

Signed-off-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Akash Goel <akash.goel@intel.com>
---
 drivers/gpu/drm/i915/i915_dma.c      |   1 +
 drivers/gpu/drm/i915/i915_drv.h      |   8 ++
 drivers/gpu/drm/i915/i915_gem.c      |   2 +-
 drivers/gpu/drm/i915/i915_irq.c      |   3 +
 drivers/gpu/drm/i915/i915_params.c   |   4 +
 drivers/gpu/drm/i915/intel_display.c | 140 +++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/i915/intel_drv.h     |   6 ++
 7 files changed, 163 insertions(+), 1 deletion(-)

Comments

Chris Wilson May 28, 2014, 7:30 a.m. UTC | #1 
On Wed, May 28, 2014 at 12:42:01PM +0530, sourab.gupta@intel.com wrote:
> +static int intel_postpone_flip(struct drm_i915_gem_object *obj)
> +{
> +	int ret;
> +
> +	if (!obj->ring)
> +		return 0;
> +
> +	if (i915_seqno_passed(obj->ring->get_seqno(obj->ring, true),
> +				obj->last_write_seqno))
> +		return 0;

obj->last_write_seqno could be 0 here. To be correct, test against
obj->last_write_seqno == 0 instead of obj->ring == NULL first.

> +	ret = i915_gem_check_olr(obj->ring, obj->last_write_seqno);
> +	if (ret)
> +		return ret;
> +
> +	if (WARN_ON(!obj->ring->irq_get(obj->ring)))
> +		return 0;
> +
> +	return 1;
> +}
> +

>  static int intel_default_queue_flip(struct drm_device *dev,
>  				    struct drm_crtc *crtc,
>  				    struct drm_framebuffer *fb,
> @@ -11515,6 +11652,9 @@ static void intel_init_display(struct drm_device *dev)
>  		break;
>  	}
>  
> +	if (intel_use_mmio_flip(dev))
> +		dev_priv->display.queue_flip = intel_queue_mmio_flip;

I'd still like to see this as a tristate, i.e. use_mmio_flip(dev) > 1
here. E.g. we will want to use mmio flips for !BCS, but GPU flips when
BCS is active.
-Chris
Chris Wilson May 28, 2014, 7:31 a.m. UTC | #2 
On Wed, May 28, 2014 at 12:42:01PM +0530, sourab.gupta@intel.com wrote:
> +static void intel_do_mmio_flip(struct intel_crtc *intel_crtc)
> +{
> +	struct drm_device *dev = intel_crtc->base.dev;
> +	struct drm_i915_private *dev_priv = dev->dev_private;
> +	struct intel_framebuffer *intel_fb =
> +		to_intel_framebuffer(intel_crtc->base.primary->fb);
> +	struct drm_i915_gem_object *obj = intel_fb->obj;
> +	u32 dspcntr;
> +	u32 reg;
> +
> +	intel_mark_page_flip_active(intel_crtc);
> +
> +	reg = DSPCNTR(intel_crtc->plane);
> +	dspcntr = I915_READ(reg);
> +
> +	if (INTEL_INFO(dev)->gen >= 4) {
> +		if (obj->tiling_mode != I915_TILING_NONE)
> +			dspcntr |= DISPPLANE_TILED;
> +		else
> +			dspcntr &= ~DISPPLANE_TILED;
> +	}
> +	I915_WRITE(reg, dspcntr);
> +
> +	I915_WRITE(DSPSURF(intel_crtc->plane),
> +			intel_crtc->unpin_work->gtt_offset);
> +	POSTING_READ(DSPSURF(intel_crtc->plane));
> +}

So other than byt, why would we not use LRI here and avoid waking the
CPU up?
-Chris
Ville Syrjala May 28, 2014, 8:12 a.m. UTC | #3 
On Wed, May 28, 2014 at 08:31:52AM +0100, Chris Wilson wrote:
> On Wed, May 28, 2014 at 12:42:01PM +0530, sourab.gupta@intel.com wrote:
> > +static void intel_do_mmio_flip(struct intel_crtc *intel_crtc)
> > +{
> > +	struct drm_device *dev = intel_crtc->base.dev;
> > +	struct drm_i915_private *dev_priv = dev->dev_private;
> > +	struct intel_framebuffer *intel_fb =
> > +		to_intel_framebuffer(intel_crtc->base.primary->fb);
> > +	struct drm_i915_gem_object *obj = intel_fb->obj;
> > +	u32 dspcntr;
> > +	u32 reg;
> > +
> > +	intel_mark_page_flip_active(intel_crtc);
> > +
> > +	reg = DSPCNTR(intel_crtc->plane);
> > +	dspcntr = I915_READ(reg);
> > +
> > +	if (INTEL_INFO(dev)->gen >= 4) {
> > +		if (obj->tiling_mode != I915_TILING_NONE)
> > +			dspcntr |= DISPPLANE_TILED;
> > +		else
> > +			dspcntr &= ~DISPPLANE_TILED;
> > +	}
> > +	I915_WRITE(reg, dspcntr);
> > +
> > +	I915_WRITE(DSPSURF(intel_crtc->plane),
> > +			intel_crtc->unpin_work->gtt_offset);
> > +	POSTING_READ(DSPSURF(intel_crtc->plane));
> > +}
> 
> So other than byt, why would we not use LRI here and avoid waking the
> CPU up?

The plan is to eventually expand this thing to handle the nuclear flip
and we're going use mmio for that. So going for mmio from the start
seems fine to me. Especially since we need the mmio path for byt anyway.
sourab.gupta@intel.com May 28, 2014, 9:42 a.m. UTC | #4 
On Wed, 2014-05-28 at 07:30 +0000, Chris Wilson wrote:
> On Wed, May 28, 2014 at 12:42:01PM +0530, sourab.gupta@intel.com wrote:
> > +static int intel_postpone_flip(struct drm_i915_gem_object *obj)
> > +{
> > +	int ret;
> > +
> > +	if (!obj->ring)
> > +		return 0;
> > +
> > +	if (i915_seqno_passed(obj->ring->get_seqno(obj->ring, true),
> > +				obj->last_write_seqno))
> > +		return 0;
> 
> obj->last_write_seqno could be 0 here. To be correct, test against
> obj->last_write_seqno == 0 instead of obj->ring == NULL first.
> 
> > +	ret = i915_gem_check_olr(obj->ring, obj->last_write_seqno);
> > +	if (ret)
> > +		return ret;
> > +
> > +	if (WARN_ON(!obj->ring->irq_get(obj->ring)))
> > +		return 0;
> > +
> > +	return 1;
> > +}
> > +
> 
> >  static int intel_default_queue_flip(struct drm_device *dev,
> >  				    struct drm_crtc *crtc,
> >  				    struct drm_framebuffer *fb,
> > @@ -11515,6 +11652,9 @@ static void intel_init_display(struct drm_device *dev)
> >  		break;
> >  	}
> >  
> > +	if (intel_use_mmio_flip(dev))
> > +		dev_priv->display.queue_flip = intel_queue_mmio_flip;
> 
> I'd still like to see this as a tristate, i.e. use_mmio_flip(dev) > 1
> here. E.g. we will want to use mmio flips for !BCS, but GPU flips when
> BCS is active.
> -Chris
> 
Hi Chris,
We can extend the module param to have tristate here, but since the
current design assigns the .queue_flip parameter at the init time, not
at the page flip time, this kind of usecase of having driver discretion
in terms of assigning CS vs MMIO flip can't be applied in this patch.

If required, we can have a subsequent patch to make a flip time decision
of MMIO vs CS flip based on module parameter.

Regards,
Sourab
diff mbox

Patch

diff --git a/drivers/gpu/drm/i915/i915_dma.c b/drivers/gpu/drm/i915/i915_dma.c
index b9159ad..532733a 100644
--- a/drivers/gpu/drm/i915/i915_dma.c
+++ b/drivers/gpu/drm/i915/i915_dma.c
@@ -1572,6 +1572,7 @@  int i915_driver_load(struct drm_device *dev, unsigned long flags)
 	spin_lock_init(&dev_priv->backlight_lock);
 	spin_lock_init(&dev_priv->uncore.lock);
 	spin_lock_init(&dev_priv->mm.object_stat_lock);
+	spin_lock_init(&dev_priv->mmio_flip_lock);
 	mutex_init(&dev_priv->dpio_lock);
 	mutex_init(&dev_priv->modeset_restore_lock);
 
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index bea9ab40..36e1b2a 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1367,6 +1367,9 @@  struct drm_i915_private {
 	/* protects the irq masks */
 	spinlock_t irq_lock;
 
+	/* protects the mmio flip data */
+	spinlock_t mmio_flip_lock;
+
 	bool display_irqs_enabled;
 
 	/* To control wakeup latency, e.g. for irq-driven dp aux transfers. */
@@ -2036,6 +2039,7 @@  struct i915_params {
 	bool reset;
 	bool disable_display;
 	bool disable_vtd_wa;
+	bool use_mmio_flip;
 };
 extern struct i915_params i915 __read_mostly;
 
@@ -2231,6 +2235,8 @@  bool i915_gem_retire_requests(struct drm_device *dev);
 void i915_gem_retire_requests_ring(struct intel_engine_cs *ring);
 int __must_check i915_gem_check_wedge(struct i915_gpu_error *error,
 				      bool interruptible);
+int __must_check i915_gem_check_olr(struct intel_engine_cs *ring, u32 seqno);
+
 static inline bool i915_reset_in_progress(struct i915_gpu_error *error)
 {
 	return unlikely(atomic_read(&error->reset_counter)
@@ -2601,6 +2607,8 @@  int i915_reg_read_ioctl(struct drm_device *dev, void *data,
 int i915_get_reset_stats_ioctl(struct drm_device *dev, void *data,
 			       struct drm_file *file);
 
+void intel_notify_mmio_flip(struct intel_engine_cs *ring);
+
 /* overlay */
 extern struct intel_overlay_error_state *intel_overlay_capture_error_state(struct drm_device *dev);
 extern void intel_overlay_print_error_state(struct drm_i915_error_state_buf *e,
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 70b4f41..ab663ca 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1096,7 +1096,7 @@  i915_gem_check_wedge(struct i915_gpu_error *error,
  * Compare seqno against outstanding lazy request. Emit a request if they are
  * equal.
  */
-static int
+__must_check int
 i915_gem_check_olr(struct intel_engine_cs *ring, u32 seqno)
 {
 	int ret;
diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index 4ef6423..e0edb1f 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -1218,6 +1218,9 @@  static void notify_ring(struct drm_device *dev,
 
 	trace_i915_gem_request_complete(ring);
 
+	if (drm_core_check_feature(dev, DRIVER_MODESET))
+		intel_notify_mmio_flip(ring);
+
 	wake_up_all(&ring->irq_queue);
 	i915_queue_hangcheck(dev);
 }
diff --git a/drivers/gpu/drm/i915/i915_params.c b/drivers/gpu/drm/i915/i915_params.c
index d05a2af..e0d44df 100644
--- a/drivers/gpu/drm/i915/i915_params.c
+++ b/drivers/gpu/drm/i915/i915_params.c
@@ -48,6 +48,7 @@  struct i915_params i915 __read_mostly = {
 	.disable_display = 0,
 	.enable_cmd_parser = 1,
 	.disable_vtd_wa = 0,
+	.use_mmio_flip = 0,
 };
 
 module_param_named(modeset, i915.modeset, int, 0400);
@@ -156,3 +157,6 @@  MODULE_PARM_DESC(disable_vtd_wa, "Disable all VT-d workarounds (default: false)"
 module_param_named(enable_cmd_parser, i915.enable_cmd_parser, int, 0600);
 MODULE_PARM_DESC(enable_cmd_parser,
 		 "Enable command parsing (1=enabled [default], 0=disabled)");
+
+module_param_named(use_mmio_flip, i915.use_mmio_flip, bool, 0600);
+MODULE_PARM_DESC(use_mmio_flip, "use MMIO flips (default: false)");
diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
index 731cd01..f11abfb 100644
--- a/drivers/gpu/drm/i915/intel_display.c
+++ b/drivers/gpu/drm/i915/intel_display.c
@@ -9183,6 +9183,143 @@  static int intel_gen7_queue_flip(struct drm_device *dev,
 	return 0;
 }
 
+static bool intel_use_mmio_flip(struct drm_device *dev)
+{
+	/* If module parameter is disabled, use CS flips.
+	 * Otherwise, use MMIO flips starting from Gen5.
+	 * This is not being used for older platforms, because
+	 * non-availability of flip done interrupt forces us to use
+	 * CS flips. Older platforms derive flip done using some clever
+	 * tricks involving the flip_pending status bits and vblank irqs.
+	 * So using MMIO flips there would disrupt this mechanism.
+	 */
+
+	if (i915.use_mmio_flip == 0)
+		return false;
+
+	if (INTEL_INFO(dev)->gen >= 5)
+		return true;
+	else
+		return false;
+}
+
+static void intel_do_mmio_flip(struct intel_crtc *intel_crtc)
+{
+	struct drm_device *dev = intel_crtc->base.dev;
+	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_framebuffer *intel_fb =
+		to_intel_framebuffer(intel_crtc->base.primary->fb);
+	struct drm_i915_gem_object *obj = intel_fb->obj;
+	u32 dspcntr;
+	u32 reg;
+
+	intel_mark_page_flip_active(intel_crtc);
+
+	reg = DSPCNTR(intel_crtc->plane);
+	dspcntr = I915_READ(reg);
+
+	if (INTEL_INFO(dev)->gen >= 4) {
+		if (obj->tiling_mode != I915_TILING_NONE)
+			dspcntr |= DISPPLANE_TILED;
+		else
+			dspcntr &= ~DISPPLANE_TILED;
+	}
+	I915_WRITE(reg, dspcntr);
+
+	I915_WRITE(DSPSURF(intel_crtc->plane),
+			intel_crtc->unpin_work->gtt_offset);
+	POSTING_READ(DSPSURF(intel_crtc->plane));
+}
+
+static int intel_postpone_flip(struct drm_i915_gem_object *obj)
+{
+	int ret;
+
+	if (!obj->ring)
+		return 0;
+
+	if (i915_seqno_passed(obj->ring->get_seqno(obj->ring, true),
+				obj->last_write_seqno))
+		return 0;
+
+	ret = i915_gem_check_olr(obj->ring, obj->last_write_seqno);
+	if (ret)
+		return ret;
+
+	if (WARN_ON(!obj->ring->irq_get(obj->ring)))
+		return 0;
+
+	return 1;
+}
+
+void intel_notify_mmio_flip(struct intel_engine_cs *ring)
+{
+	struct drm_i915_private *dev_priv = ring->dev->dev_private;
+	struct intel_crtc *intel_crtc;
+	unsigned long irq_flags;
+	u32 seqno;
+
+	seqno = ring->get_seqno(ring, false);
+
+	spin_lock_irqsave(&dev_priv->mmio_flip_lock, irq_flags);
+	for_each_intel_crtc(ring->dev, intel_crtc) {
+		struct intel_mmio_flip *mmio_flip;
+
+		mmio_flip = &intel_crtc->mmio_flip;
+
+		if (mmio_flip->seqno == 0)
+			continue;
+		if (ring->id != mmio_flip->ring_id)
+			continue;
+
+		if (i915_seqno_passed(seqno, mmio_flip->seqno)) {
+			intel_do_mmio_flip(intel_crtc);
+			mmio_flip->seqno = 0;
+			ring->irq_put(ring);
+		}
+	}
+	spin_unlock_irqrestore(&dev_priv->mmio_flip_lock, irq_flags);
+}
+
+static int intel_queue_mmio_flip(struct drm_device *dev,
+			struct drm_crtc *crtc,
+			struct drm_framebuffer *fb,
+			struct drm_i915_gem_object *obj,
+			struct intel_engine_cs *ring,
+			uint32_t flags)
+{
+	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
+	unsigned long irq_flags;
+	int ret;
+
+	if (WARN_ON(intel_crtc->mmio_flip.seqno)) {
+		ret = -EBUSY;
+		goto err;
+	}
+
+	ret = intel_postpone_flip(obj);
+	if (ret < 0) {
+		goto err;
+	} else if (ret == 0) {
+		intel_do_mmio_flip(intel_crtc);
+		return 0;
+	}
+
+	spin_lock_irqsave(&dev_priv->mmio_flip_lock, irq_flags);
+	intel_crtc->mmio_flip.seqno = obj->last_write_seqno;
+	intel_crtc->mmio_flip.ring_id = obj->ring->id;
+	spin_unlock_irqrestore(&dev_priv->mmio_flip_lock, irq_flags);
+
+	/* Double check to catch cases where irq fired before
+	 * mmio flip data was ready
+	 */
+	intel_notify_mmio_flip(obj->ring);
+	return 0;
+err:
+	return ret;
+}
+
 static int intel_default_queue_flip(struct drm_device *dev,
 				    struct drm_crtc *crtc,
 				    struct drm_framebuffer *fb,
@@ -11515,6 +11652,9 @@  static void intel_init_display(struct drm_device *dev)
 		break;
 	}
 
+	if (intel_use_mmio_flip(dev))
+		dev_priv->display.queue_flip = intel_queue_mmio_flip;
+
 	intel_panel_init_backlight_funcs(dev);
 }
 
diff --git a/drivers/gpu/drm/i915/intel_drv.h b/drivers/gpu/drm/i915/intel_drv.h
index 9bb70dc..206b577 100644
--- a/drivers/gpu/drm/i915/intel_drv.h
+++ b/drivers/gpu/drm/i915/intel_drv.h
@@ -358,6 +358,11 @@  struct intel_pipe_wm {
 	bool sprites_scaled;
 };
 
+struct intel_mmio_flip {
+	u32 seqno;
+	u32 ring_id;
+};
+
 struct intel_crtc {
 	struct drm_crtc base;
 	enum pipe pipe;
@@ -411,6 +416,7 @@  struct intel_crtc {
 	wait_queue_head_t vbl_wait;
 
 	int scanline_offset;
+	struct intel_mmio_flip mmio_flip;
 };
 
 struct intel_plane_wm_parameters {