Skip to contentSkip to conversation

drm/i915: Fix fbc + rc6 combination on SNB

PATCH In review Ben Widawsky 6 participants 10 messages Nov 2, 2013 – Nov 19, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 drm/i915: Fix fbc + rc6 combination on SNB

On Sandybridge we must set the "PPGTT Render Target Base Address Valid for FBC" bit as noted in the programming guide. We did this at clock gating init. Thisbit is not saved and restored with RC6 power context, so the resetting it at ring flush should fix that.

The effect of not doing this should be corruption, and not a hang - as has so often been the case.

Note that we should actually clear this bit as well when not blitting to the scanout (using the blitter for other things), or else all operations

Cc Stéphane Marchesin <[email protected]>
---
 drivers/gpu/drm/i915/intel_pm.c         |  2 --
 drivers/gpu/drm/i915/intel_ringbuffer.c | 25 +++++++++++++++++++++++++
 2 files changed, 25 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_pm.c +0 -2
index ca9a778..67f460b 100644
@@ -193,8 +193,6 @@ static void sandybridge_blit_fbc_update(struct drm_device *dev)
 	/* Make sure blitter notifies FBC of writes */
 	gen6_gt_force_wake_get(dev_priv);
 	blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
-	blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
-		GEN6_BLITTER_LOCK_SHIFT;
 	I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 	blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
 	I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
drivers/gpu/drm/i915/intel_ringbuffer.c +25 -0
index 2dec134..ddd7681 100644
@@ -278,6 +278,28 @@ gen7_render_ring_cs_stall_wa(struct intel_ring_buffer *ring)
 	return 0;
 }
 
+static int gen6_ring_fbc_flush(struct intel_ring_buffer *ring)
+{
+	int ret;
+
+	if (!ring->fbc_dirty)
+		return 0;
+
+	ret = intel_ring_begin(ring, 4);
+	if (ret)
+		return ret;
+
+	intel_ring_emit(ring, MI_NOOP);
+	intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
+	intel_ring_emit(ring, GEN6_BLITTER_ECOSKPD);
+	intel_ring_emit(ring,
+			_MASKED_BIT_ENABLE(GEN6_BLITTER_FBC_NOTIFY));
+	intel_ring_advance(ring);
+
+	ring->fbc_dirty = false;
+	return 0;
+}
+
 static int gen7_ring_fbc_flush(struct intel_ring_buffer *ring, u32 value)
 {
 	int ret;
@@ -1712,6 +1734,9 @@ static int gen6_ring_flush(struct intel_ring_buffer *ring,
 	intel_ring_emit(ring, MI_NOOP);
 	intel_ring_advance(ring);
 
+	if (IS_GEN6(dev) && flush)
+		return gen6_ring_fbc_flush(ring);
+
 	if (IS_GEN7(dev) && flush)
 		return gen7_ring_fbc_flush(ring, FBC_REND_CACHE_CLEAN);
 
Signature · 6 lines
1.8.4.2

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

PATCH 2/2 drm/i915: Disable blt tracking of fbc when not used

When not blitting to scanout, we can save some power by not tracking blits, and more importantly, unnecessarily invalidating lines which we don't care a bout.

These instructions are explicitly spelled out in the spec, but it is how I expect it to work. Unfortunately, I do not have power data for this.

Cc Stéphane Marchesin <[email protected]>
---
 drivers/gpu/drm/i915/intel_ringbuffer.c | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/intel_ringbuffer.c +18 -3
index ddd7681..56be8cb 100644
@@ -296,7 +296,8 @@ static int gen6_ring_fbc_flush(struct intel_ring_buffer *ring)
 			_MASKED_BIT_ENABLE(GEN6_BLITTER_FBC_NOTIFY));
 	intel_ring_advance(ring);
 
-	ring->fbc_dirty = false;
+	/* We'll mark the fbc clean only after the operation has completed so we
+	 * can track when to disable the bit above */
 	return 0;
 }
 
@@ -642,11 +643,13 @@ gen6_add_request(struct intel_ring_buffer *ring)
 	struct drm_device *dev = ring->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct intel_ring_buffer *useless;
-	int i, ret;
+	int i, ret, extra_dwords = 4;
 
+	if (ring->fbc_dirty && ring->id == BCS)
+		extra_dwords += 4;
 	ret = intel_ring_begin(ring, ((I915_NUM_RINGS-1) *
 				      MBOX_UPDATE_DWORDS) +
-				      4);
+				      extra_dwords);
 	if (ret)
 		return ret;
 #undef MBOX_UPDATE_DWORDS
@@ -661,6 +664,18 @@ gen6_add_request(struct intel_ring_buffer *ring)
 	intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
 	intel_ring_emit(ring, ring->outstanding_lazy_seqno);
 	intel_ring_emit(ring, MI_USER_INTERRUPT);
+
+	if (ring->fbc_dirty && ring->id == BCS) {
+		intel_ring_emit(ring, MI_NOOP);
+		intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
+		intel_ring_emit(ring, GEN6_BLITTER_ECOSKPD);
+		intel_ring_emit(ring,
+				_MASKED_BIT_DISABLE(GEN6_BLITTER_FBC_NOTIFY));
+		__intel_ring_advance(ring);
+
+		ring->fbc_dirty = false;
+	}
+
 	__intel_ring_advance(ring);
 
 	return 0;
Signature · 6 lines
1.8.4.2

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw

PATCH v2 2/2 drm/i915: Disable blt tracking of fbc when not used

When not blitting to scanout, we can save some power by not tracking blits, and more importantly, unnecessarily invalidating lines which we don't care a bout.

These instructions are explicitly spelled out in the spec, but it is how I expect it to work. Unfortunately, I do not have power data for this.

v2: Don't advance the ring twice (Stéphane)
Rename extra_dwords to dwrods (Stéphane)

Cc Stéphane Marchesin <[email protected]>
---
 drivers/gpu/drm/i915/intel_ringbuffer.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/intel_ringbuffer.c +17 -3
index ddd7681..8c6e9b2 100644
@@ -296,7 +296,8 @@ static int gen6_ring_fbc_flush(struct intel_ring_buffer *ring)
 			_MASKED_BIT_ENABLE(GEN6_BLITTER_FBC_NOTIFY));
 	intel_ring_advance(ring);
 
-	ring->fbc_dirty = false;
+	/* We'll mark the fbc clean only after the operation has completed so we
+	 * can track when to disable the bit above */
 	return 0;
 }
 
@@ -642,11 +643,13 @@ gen6_add_request(struct intel_ring_buffer *ring)
 	struct drm_device *dev = ring->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct intel_ring_buffer *useless;
-	int i, ret;
+	int i, ret, dwords = 4;
 
+	if (ring->fbc_dirty && ring->id == BCS)
+		dwords += 4;
 	ret = intel_ring_begin(ring, ((I915_NUM_RINGS-1) *
 				      MBOX_UPDATE_DWORDS) +
-				      4);
+				      dwords);
 	if (ret)
 		return ret;
 #undef MBOX_UPDATE_DWORDS
@@ -661,6 +664,17 @@ gen6_add_request(struct intel_ring_buffer *ring)
 	intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
 	intel_ring_emit(ring, ring->outstanding_lazy_seqno);
 	intel_ring_emit(ring, MI_USER_INTERRUPT);
+
+	if (ring->fbc_dirty && ring->id == BCS) {
+		intel_ring_emit(ring, MI_NOOP);
+		intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
+		intel_ring_emit(ring, GEN6_BLITTER_ECOSKPD);
+		intel_ring_emit(ring,
+				_MASKED_BIT_DISABLE(GEN6_BLITTER_FBC_NOTIFY));
+
+		ring->fbc_dirty = false;
+	}
+
 	__intel_ring_advance(ring);
 
 	return 0;
Signature · 6 lines
1.8.4.2

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Tested-by Reviewed-by on v2 2/2
#3 lore raw

Re: [PATCH 2/2] [v2] drm/i915: Disable blt tracking of fbc when not used

Quoted 4 lines · Ben Widawsky

When not blitting to scanout, we can save some power by not tracking blits, and more importantly, unnecessarily invalidating lines which we don't care a bout.

Quoted 8 lines

These instructions are explicitly spelled out in the spec, but it is how I expect it to work. Unfortunately, I do not have power data for this.

v2: Don't advance the ring twice (Stéphane)
Rename extra_dwords to dwrods (Stéphane)

Cc Stéphane Marchesin <[email protected]>

For both:

Stéphane

Quoted 55 lines
---
 drivers/gpu/drm/i915/intel_ringbuffer.c | 20 +++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/intel_ringbuffer.c +17 -4
index ddd7681..8c6e9b2 100644
@@ -296,7 +296,8 @@ static int gen6_ring_fbc_flush(struct intel_ring_buffer *ring)
                        _MASKED_BIT_ENABLE(GEN6_BLITTER_FBC_NOTIFY));
        intel_ring_advance(ring);
 
-       ring->fbc_dirty = false;
+       /* We'll mark the fbc clean only after the operation has completed so we
+        * can track when to disable the bit above */
        return 0;
 }
 
@@ -642,11 +643,13 @@ gen6_add_request(struct intel_ring_buffer *ring)
        struct drm_device *dev = ring->dev;
        struct drm_i915_private *dev_priv = dev->dev_private;
        struct intel_ring_buffer *useless;
-       int i, ret;
+       int i, ret, dwords = 4;
 
+       if (ring->fbc_dirty && ring->id == BCS)
+               dwords += 4;
        ret = intel_ring_begin(ring, ((I915_NUM_RINGS-1) *
                                      MBOX_UPDATE_DWORDS) +
-                                     4);
+                                     dwords);
        if (ret)
                return ret;
 #undef MBOX_UPDATE_DWORDS
@@ -661,6 +664,17 @@ gen6_add_request(struct intel_ring_buffer *ring)
        intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
        intel_ring_emit(ring, ring->outstanding_lazy_seqno);
        intel_ring_emit(ring, MI_USER_INTERRUPT);
+
+       if (ring->fbc_dirty && ring->id == BCS) {
+               intel_ring_emit(ring, MI_NOOP);
+               intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
+               intel_ring_emit(ring, GEN6_BLITTER_ECOSKPD);
+               intel_ring_emit(ring,
+                               _MASKED_BIT_DISABLE(GEN6_BLITTER_FBC_NOTIFY));
+
+               ring->fbc_dirty = false;
+       }
+
        __intel_ring_advance(ring);
 
        return 0;
--

1.8.4.2

#4 lore raw
In reply to Ben Widawsky
Quoted 10 lines · Ben Widawsky

On Sandybridge we must set the "PPGTT Render Target Base Address Valid for FBC" bit as noted in the programming guide. We did this at clock gating init. Thisbit is not saved and restored with RC6 power context, so the resetting it at ring flush should fix that.

The effect of not doing this should be corruption, and not a hang - as has so often been the case.

Note that we should actually clear this bit as well when not blitting to the scanout (using the blitter for other things), or else all operations

This needs to be enabled during the invalidate phase, so that the tracking is active for the batch. However, the fbc_dirty flag is currently set later.
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#5 lore raw
In reply to Ben Widawsky
Quoted 23 lines · Ben Widawsky

On Sandybridge we must set the "PPGTT Render Target Base Address Valid for FBC" bit as noted in the programming guide. We did this at clock gating init. Thisbit is not saved and restored with RC6 power context, so the resetting it at ring flush should fix that.

The effect of not doing this should be corruption, and not a hang - as has so often been the case.

Note that we should actually clear this bit as well when not blitting to the scanout (using the blitter for other things), or else all operations

Cc Stéphane Marchesin <[email protected]>
---
 drivers/gpu/drm/i915/intel_pm.c         |  2 --
 drivers/gpu/drm/i915/intel_ringbuffer.c | 25 +++++++++++++++++++++++++
 2 files changed, 25 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_pm.c +0 -0
index ca9a778..67f460b 100644

drivers/gpu/drm/i915/intel_pm.c static void sandybridge_blit_fbc_update(struct drm_device *dev) · line 198

Quoted 8 lines
drivers/gpu/drm/i915/intel_pm.c +0 -2
@@ -193,8 +193,6 @@ static void sandybridge_blit_fbc_update(struct drm_device *dev)
 	/* Make sure blitter notifies FBC of writes */
 	gen6_gt_force_wake_get(dev_priv);
 	blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
-	blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
-		GEN6_BLITTER_LOCK_SHIFT;
 	I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 	blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
 	I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);

Why leave the other FBC_NOTIFY frobbing in place? Since you don't set the mask bit anymore the rest isn't guaranteed to do anything.

Quoted 34 lines
drivers/gpu/drm/i915/intel_ringbuffer.c +22 -0
index 2dec134..ddd7681 100644
@@ -278,6 +278,28 @@ gen7_render_ring_cs_stall_wa(struct intel_ring_buffer *ring)
 	return 0;
 }
 
+static int gen6_ring_fbc_flush(struct intel_ring_buffer *ring)
+{
+	int ret;
+
+	if (!ring->fbc_dirty)
+		return 0;
+
+	ret = intel_ring_begin(ring, 4);
+	if (ret)
+		return ret;
+
+	intel_ring_emit(ring, MI_NOOP);
+	intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
+	intel_ring_emit(ring, GEN6_BLITTER_ECOSKPD);
+	intel_ring_emit(ring,
+			_MASKED_BIT_ENABLE(GEN6_BLITTER_FBC_NOTIFY));
+	intel_ring_advance(ring);
+
+	ring->fbc_dirty = false;
+	return 0;
+}
+
 static int gen7_ring_fbc_flush(struct intel_ring_buffer *ring, u32 value)
 {
 	int ret;

drivers/gpu/drm/i915/intel_ringbuffer.c static int gen6_ring_flush(struct intel_ring_buffer *ring, · line 1739

Quoted 6 lines
drivers/gpu/drm/i915/intel_ringbuffer.c +3 -0
@@ -1712,6 +1734,9 @@ static int gen6_ring_flush(struct intel_ring_buffer *ring,
 	intel_ring_emit(ring, MI_NOOP);
 	intel_ring_advance(ring);
 
+	if (IS_GEN6(dev) && flush)
+		return gen6_ring_fbc_flush(ring);
+

What Chris said about doing this before the batch is dispatched.

Afer a bit of thought I think the following logic would work nicely:

execbuffer() {
	ring->new_fbc_obj = NULL;
	for_each_obj() {
		if (is_crtc_fb(obj) && obj.write_domains)
			ring->new_fbc_obj = obj;
	if (gen >= 7) {
		if (ring->new_fbc_obj)
			ring->fbc_dirty = true;
	} else {
		if (ring->new_fb_obj != ring->fbc_obj) {
			ring->fbc_obj = ring->new_fbc_obj;
			ring->fbc_dirty = true;
		}
	}

	invalidate() {
		if (gen < 7 && ring->fbc_dirty) {
			if (ring->fbc_obj)
				enable_tracking;
			else
				disable_tracking;
		}
	}

	dispatch()

	flush()	{
		if (gen >= 7 && ring->fbc_dirty)
			fbc_nuke();
		ring->fbc_dirty = false;
	}
}

I think that same logic would work for both blitter and render. The difference between the two is that for render we also need to update the address, for blitter we just need to set the notify bit.

Also since we could update the render tracking for every batch, the problem of having the render fbc tracking address in the context would also be solved by simply setting fbc_dirty=true on context switch.

I don't recall excatly how we're supposed to do blitter tracking on on gen7+. I seem to recall that it also had a nuke mechanism, but I don't see it being used in out code ATM.

Signature · 2 lines
Ville Syrjälä
Intel OTC
#6 lore raw

I'm just on going with another -collector update and since this patch fixes a bug I think it would be a good one to include.

But since it was bikeshedded it is better to ask Ville and Chris if their comments was a NAck or I can consider to get for -collector.

Thanks

Quoted 134 lines · Ville Syrjälä
Quoted 23 lines · Ben Widawsky

On Sandybridge we must set the "PPGTT Render Target Base Address Valid for FBC" bit as noted in the programming guide. We did this at clock gating init. Thisbit is not saved and restored with RC6 power context, so the resetting it at ring flush should fix that.

The effect of not doing this should be corruption, and not a hang - as has so often been the case.

Note that we should actually clear this bit as well when not blitting to the scanout (using the blitter for other things), or else all operations

Cc Stéphane Marchesin <[email protected]>
---
 drivers/gpu/drm/i915/intel_pm.c         |  2 --
 drivers/gpu/drm/i915/intel_ringbuffer.c | 25 +++++++++++++++++++++++++
 2 files changed, 25 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_pm.c +0 -0
index ca9a778..67f460b 100644

drivers/gpu/drm/i915/intel_pm.c static void sandybridge_blit_fbc_update(struct drm_device *dev) · line 198

Quoted 8 lines
drivers/gpu/drm/i915/intel_pm.c +0 -2
@@ -193,8 +193,6 @@ static void sandybridge_blit_fbc_update(struct drm_device *dev)
      /* Make sure blitter notifies FBC of writes */
      gen6_gt_force_wake_get(dev_priv);
      blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
-     blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
-             GEN6_BLITTER_LOCK_SHIFT;
      I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
      blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
      I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);

Why leave the other FBC_NOTIFY frobbing in place? Since you don't set the mask bit anymore the rest isn't guaranteed to do anything.

Quoted 34 lines
drivers/gpu/drm/i915/intel_ringbuffer.c +22 -0
index 2dec134..ddd7681 100644
@@ -278,6 +278,28 @@ gen7_render_ring_cs_stall_wa(struct intel_ring_buffer *ring)
      return 0;
 }
 
+static int gen6_ring_fbc_flush(struct intel_ring_buffer *ring)
+{
+     int ret;
+
+     if (!ring->fbc_dirty)
+             return 0;
+
+     ret = intel_ring_begin(ring, 4);
+     if (ret)
+             return ret;
+
+     intel_ring_emit(ring, MI_NOOP);
+     intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
+     intel_ring_emit(ring, GEN6_BLITTER_ECOSKPD);
+     intel_ring_emit(ring,
+                     _MASKED_BIT_ENABLE(GEN6_BLITTER_FBC_NOTIFY));
+     intel_ring_advance(ring);
+
+     ring->fbc_dirty = false;
+     return 0;
+}
+
 static int gen7_ring_fbc_flush(struct intel_ring_buffer *ring, u32 value)
 {
      int ret;

drivers/gpu/drm/i915/intel_ringbuffer.c static int gen6_ring_flush(struct intel_ring_buffer *ring, · line 1739

Quoted 6 lines
drivers/gpu/drm/i915/intel_ringbuffer.c +3 -0
@@ -1712,6 +1734,9 @@ static int gen6_ring_flush(struct intel_ring_buffer *ring,
      intel_ring_emit(ring, MI_NOOP);
      intel_ring_advance(ring);
 
+     if (IS_GEN6(dev) && flush)
+             return gen6_ring_fbc_flush(ring);
+

What Chris said about doing this before the batch is dispatched.

Afer a bit of thought I think the following logic would work nicely:

execbuffer() {
        ring->new_fbc_obj = NULL;
        for_each_obj() {
                if (is_crtc_fb(obj) && obj.write_domains)
                        ring->new_fbc_obj = obj;
        if (gen >= 7) {
                if (ring->new_fbc_obj)
                        ring->fbc_dirty = true;
        } else {
                if (ring->new_fb_obj != ring->fbc_obj) {
                        ring->fbc_obj = ring->new_fbc_obj;
                        ring->fbc_dirty = true;
                }
        }

        invalidate() {
                if (gen < 7 && ring->fbc_dirty) {
                        if (ring->fbc_obj)
                                enable_tracking;
                        else
                                disable_tracking;
                }
        }

        dispatch()

        flush() {
                if (gen >= 7 && ring->fbc_dirty)
                        fbc_nuke();
                ring->fbc_dirty = false;
        }
}

I think that same logic would work for both blitter and render. The difference between the two is that for render we also need to update the address, for blitter we just need to set the notify bit.

Also since we could update the render tracking for every batch, the problem of having the render fbc tracking address in the context would also be solved by simply setting fbc_dirty=true on context switch.

I don't recall excatly how we're supposed to do blitter tracking on on gen7+. I seem to recall that it also had a nuke mechanism, but I don't see it being used in out code ATM.

--
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Signature · 2 lines
Rodrigo Vivi
Blog: http://blog.vivi.eng.br
#7 lore raw
Quoted 5 lines · Rodrigo Vivi

I'm just on going with another -collector update and since this patch fixes a bug I think it would be a good one to include.

But since it was bikeshedded it is better to ask Ville and Chris if their comments was a NAck or I can consider to get for -collector.

It's more a meh, since as long as we don't enable fbc by default it won't really benefit upstream much at all.
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#8 lore raw
Quoted 9 lines · Daniel Vetter
Quoted 5 lines · Rodrigo Vivi

I'm just on going with another -collector update and since this patch fixes a bug I think it would be a good one to include.

But since it was bikeshedded it is better to ask Ville and Chris if their comments was a NAck or I can consider to get for -collector.

It's more a meh, since as long as we don't enable fbc by default it won't really benefit upstream much at all.

These two patches fix real FBC issues here, letting us enable it and save power. I think it's worth considering them for inclusion.

Stéphane

Quoted 9 lines

-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#9 lore raw
In reply to Rodrigo Vivi
Quoted 5 lines · Rodrigo Vivi

I'm just on going with another -collector update and since this patch fixes a bug I think it would be a good one to include.

But since it was bikeshedded it is better to ask Ville and Chris if their comments was a NAck or I can consider to get for -collector.

My FBC series makes this obsolete. I think I have a few more updates to that series that I didn't post yet. I'll try to get those out today. And I also have a crc based igt for this in the works.

Quoted 145 lines

Thanks

Quoted 134 lines · Ville Syrjälä
Quoted 23 lines · Ben Widawsky

On Sandybridge we must set the "PPGTT Render Target Base Address Valid for FBC" bit as noted in the programming guide. We did this at clock gating init. Thisbit is not saved and restored with RC6 power context, so the resetting it at ring flush should fix that.

The effect of not doing this should be corruption, and not a hang - as has so often been the case.

Note that we should actually clear this bit as well when not blitting to the scanout (using the blitter for other things), or else all operations

Cc Stéphane Marchesin <[email protected]>
---
 drivers/gpu/drm/i915/intel_pm.c         |  2 --
 drivers/gpu/drm/i915/intel_ringbuffer.c | 25 +++++++++++++++++++++++++
 2 files changed, 25 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/intel_pm.c +0 -0
index ca9a778..67f460b 100644

drivers/gpu/drm/i915/intel_pm.c static void sandybridge_blit_fbc_update(struct drm_device *dev) · line 198

Quoted 8 lines
drivers/gpu/drm/i915/intel_pm.c +0 -2
@@ -193,8 +193,6 @@ static void sandybridge_blit_fbc_update(struct drm_device *dev)
      /* Make sure blitter notifies FBC of writes */
      gen6_gt_force_wake_get(dev_priv);
      blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
-     blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
-             GEN6_BLITTER_LOCK_SHIFT;
      I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
      blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
      I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);

Why leave the other FBC_NOTIFY frobbing in place? Since you don't set the mask bit anymore the rest isn't guaranteed to do anything.

Quoted 34 lines
drivers/gpu/drm/i915/intel_ringbuffer.c +22 -0
index 2dec134..ddd7681 100644
@@ -278,6 +278,28 @@ gen7_render_ring_cs_stall_wa(struct intel_ring_buffer *ring)
      return 0;
 }
 
+static int gen6_ring_fbc_flush(struct intel_ring_buffer *ring)
+{
+     int ret;
+
+     if (!ring->fbc_dirty)
+             return 0;
+
+     ret = intel_ring_begin(ring, 4);
+     if (ret)
+             return ret;
+
+     intel_ring_emit(ring, MI_NOOP);
+     intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
+     intel_ring_emit(ring, GEN6_BLITTER_ECOSKPD);
+     intel_ring_emit(ring,
+                     _MASKED_BIT_ENABLE(GEN6_BLITTER_FBC_NOTIFY));
+     intel_ring_advance(ring);
+
+     ring->fbc_dirty = false;
+     return 0;
+}
+
 static int gen7_ring_fbc_flush(struct intel_ring_buffer *ring, u32 value)
 {
      int ret;

drivers/gpu/drm/i915/intel_ringbuffer.c static int gen6_ring_flush(struct intel_ring_buffer *ring, · line 1739

Quoted 6 lines
drivers/gpu/drm/i915/intel_ringbuffer.c +3 -0
@@ -1712,6 +1734,9 @@ static int gen6_ring_flush(struct intel_ring_buffer *ring,
      intel_ring_emit(ring, MI_NOOP);
      intel_ring_advance(ring);
 
+     if (IS_GEN6(dev) && flush)
+             return gen6_ring_fbc_flush(ring);
+

What Chris said about doing this before the batch is dispatched.

Afer a bit of thought I think the following logic would work nicely:

execbuffer() {
        ring->new_fbc_obj = NULL;
        for_each_obj() {
                if (is_crtc_fb(obj) && obj.write_domains)
                        ring->new_fbc_obj = obj;
        if (gen >= 7) {
                if (ring->new_fbc_obj)
                        ring->fbc_dirty = true;
        } else {
                if (ring->new_fb_obj != ring->fbc_obj) {
                        ring->fbc_obj = ring->new_fbc_obj;
                        ring->fbc_dirty = true;
                }
        }

        invalidate() {
                if (gen < 7 && ring->fbc_dirty) {
                        if (ring->fbc_obj)
                                enable_tracking;
                        else
                                disable_tracking;
                }
        }

        dispatch()

        flush() {
                if (gen >= 7 && ring->fbc_dirty)
                        fbc_nuke();
                ring->fbc_dirty = false;
        }
}

I think that same logic would work for both blitter and render. The difference between the two is that for render we also need to update the address, for blitter we just need to set the notify bit.

Also since we could update the render tracking for every batch, the problem of having the render fbc tracking address in the context would also be solved by simply setting fbc_dirty=true on context switch.

I don't recall excatly how we're supposed to do blitter tracking on on gen7+. I seem to recall that it also had a nuke mechanism, but I don't see it being used in out code ATM.

--
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Signature · 2 lines
Rodrigo Vivi
Blog: http://blog.vivi.eng.br
Signature · 2 lines
Ville Syrjälä
Intel OTC