Skip to contentSkip to conversation

drm/i915: Actually respect DSPSURF alignment restrictions

PATCH In review [email protected] 6 participants 10 messages Jun 11, 2015 – Jun 15, 2015

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

PATCH 1/3 drm/i915: Actually respect DSPSURF alignment restrictions

From: Ville Syrjälä <[email protected]>

Currently intel_gen4_compute_page_offset() simply picks the closest page boundary below the linear offset. That however may not be suitably aligned to satisfy any hardware specific restrictions. So let's make sure the page boundary we choose is properly aligned.

Also to play it a bit safer lets split the remaining linear offset into x and y values instead of just x. This should make no difference for most platforms since we convert the x and y offsets back into a linear offset before feeding them to the hardware. HSW+ are different however and use x and y offsets even with linear buffers, so they might have trouble if either the x or y get too big.

---
 drivers/gpu/drm/i915/intel_display.c | 37 ++++++++++++++++++++++--------------
 drivers/gpu/drm/i915/intel_drv.h     |  3 ++-
 drivers/gpu/drm/i915/intel_sprite.c  |  9 ++++++---
 3 files changed, 31 insertions(+), 18 deletions(-)
drivers/gpu/drm/i915/intel_display.c +23 -14
index 5cc2263..0ac213a 100644
@@ -2325,6 +2325,18 @@ intel_fill_fb_ggtt_view(struct i915_ggtt_view *view, struct drm_framebuffer *fb,
 	return 0;
 }
 
+static unsigned int intel_linear_alignment(struct drm_i915_private *dev_priv)
+{
+	if (INTEL_INFO(dev_priv)->gen >= 9)
+		return 256 * 1024;
+	else if (IS_BROADWATER(dev_priv) || IS_CRESTLINE(dev_priv))
+		return 128 * 1024;
+	else if (INTEL_INFO(dev_priv)->gen >= 4)
+		return 4 * 1024;
+	else
+		return 64 * 1024;
+}
+
 int
 intel_pin_and_fence_fb_obj(struct drm_plane *plane,
 			   struct drm_framebuffer *fb,
@@ -2342,14 +2354,7 @@ intel_pin_and_fence_fb_obj(struct drm_plane *plane,
 
 	switch (fb->modifier[0]) {
 	case DRM_FORMAT_MOD_NONE:
-		if (INTEL_INFO(dev)->gen >= 9)
-			alignment = 256 * 1024;
-		else if (IS_BROADWATER(dev) || IS_CRESTLINE(dev))
-			alignment = 128 * 1024;
-		else if (INTEL_INFO(dev)->gen >= 4)
-			alignment = 4 * 1024;
-		else
-			alignment = 64 * 1024;
+		alignment = intel_linear_alignment(dev_priv);
 		break;
 	case I915_FORMAT_MOD_X_TILED:
 		if (INTEL_INFO(dev)->gen >= 9)
@@ -2439,7 +2444,8 @@ static void intel_unpin_fb_obj(struct drm_framebuffer *fb,
 
 /* Computes the linear offset to the base tile and adjusts x, y. bytes per pixel
  * is assumed to be a power-of-two. */
-unsigned long intel_gen4_compute_page_offset(int *x, int *y,
+unsigned long intel_gen4_compute_page_offset(struct drm_i915_private *dev_priv,
+					     int *x, int *y,
 					     unsigned int tiling_mode,
 					     unsigned int cpp,
 					     unsigned int pitch)
@@ -2455,12 +2461,13 @@ unsigned long intel_gen4_compute_page_offset(int *x, int *y,
 
 		return tile_rows * pitch * 8 + tiles * 4096;
 	} else {
+		unsigned int alignment = intel_linear_alignment(dev_priv) - 1;
 		unsigned int offset;
 
 		offset = *y * pitch + *x * cpp;
-		*y = 0;
-		*x = (offset & 4095) / cpp;
-		return offset & -4096;
+		*y = (offset & alignment) / pitch;
+		*x = ((offset & alignment) - *y * pitch) / cpp;
+		return offset & ~alignment;
 	}
 }
 
@@ -2729,7 +2736,8 @@ static void i9xx_update_primary_plane(struct drm_crtc *crtc,
 
 	if (INTEL_INFO(dev)->gen >= 4) {
 		intel_crtc->dspaddr_offset =
-			intel_gen4_compute_page_offset(&x, &y, obj->tiling_mode,
+			intel_gen4_compute_page_offset(dev_priv,
+						       &x, &y, obj->tiling_mode,
 						       pixel_size,
 						       fb->pitches[0]);
 		linear_offset -= intel_crtc->dspaddr_offset;
@@ -2830,7 +2838,8 @@ static void ironlake_update_primary_plane(struct drm_crtc *crtc,
 
 	linear_offset = y * fb->pitches[0] + x * pixel_size;
 	intel_crtc->dspaddr_offset =
-		intel_gen4_compute_page_offset(&x, &y, obj->tiling_mode,
+		intel_gen4_compute_page_offset(dev_priv,
+					       &x, &y, obj->tiling_mode,
 					       pixel_size,
 					       fb->pitches[0]);
 	linear_offset -= intel_crtc->dspaddr_offset;
drivers/gpu/drm/i915/intel_drv.h +2 -1
index b28029a..c28d117 100644
@@ -1113,7 +1113,8 @@ void assert_fdi_rx_pll(struct drm_i915_private *dev_priv,
 void assert_pipe(struct drm_i915_private *dev_priv, enum pipe pipe, bool state);
 #define assert_pipe_enabled(d, p) assert_pipe(d, p, true)
 #define assert_pipe_disabled(d, p) assert_pipe(d, p, false)
-unsigned long intel_gen4_compute_page_offset(int *x, int *y,
+unsigned long intel_gen4_compute_page_offset(struct drm_i915_private *dev_priv,
+					     int *x, int *y,
 					     unsigned int tiling_mode,
 					     unsigned int bpp,
 					     unsigned int pitch);
drivers/gpu/drm/i915/intel_sprite.c +6 -3
index 8193a35..28f6972 100644
@@ -411,7 +411,8 @@ vlv_update_plane(struct drm_plane *dplane, struct drm_crtc *crtc,
 	crtc_h--;
 
 	linear_offset = y * fb->pitches[0] + x * pixel_size;
-	sprsurf_offset = intel_gen4_compute_page_offset(&x, &y,
+	sprsurf_offset = intel_gen4_compute_page_offset(dev_priv,
+							&x, &y,
 							obj->tiling_mode,
 							pixel_size,
 							fb->pitches[0]);
@@ -546,7 +547,8 @@ ivb_update_plane(struct drm_plane *plane, struct drm_crtc *crtc,
 
 	linear_offset = y * fb->pitches[0] + x * pixel_size;
 	sprsurf_offset =
-		intel_gen4_compute_page_offset(&x, &y, obj->tiling_mode,
+		intel_gen4_compute_page_offset(dev_priv,
+					       &x, &y, obj->tiling_mode,
 					       pixel_size, fb->pitches[0]);
 	linear_offset -= sprsurf_offset;
 
@@ -682,7 +684,8 @@ ilk_update_plane(struct drm_plane *plane, struct drm_crtc *crtc,
 
 	linear_offset = y * fb->pitches[0] + x * pixel_size;
 	dvssurf_offset =
-		intel_gen4_compute_page_offset(&x, &y, obj->tiling_mode,
+		intel_gen4_compute_page_offset(dev_priv,
+					       &x, &y, obj->tiling_mode,
 					       pixel_size, fb->pitches[0]);
 	linear_offset -= dvssurf_offset;
 
Signature · 6 lines
2.3.6

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

PATCH 2/3 drm/i915: Align DSPSURF to 128k on VLV/CHV

From: Ville Syrjälä <[email protected]>

VLV/CHV have problems with 4k aligned linear scanout buffers. The VLV docs got updated at some point to say that we need to align them to 128k, just like we do on gen4.

So far I've seen the problem manifest when the stride is an odd multiple of 512 bytes, and the surface address meets the following pattern '(addr & 0xf000) == 0x1000' (also == 0x2000 is problematic on VLV). The result is a starcase effect (so some pages get dropped maybe?), with a few pages here and there clearly getting scannout out at the wrong position.

I've not actually been able to reproduce this problem on gen4, so it's not clear of the issue is any way related to the 128k restrictions supposedly inherited from gen4. But let's hope the 128k alignment is sufficient to hide it all.

---
 drivers/gpu/drm/i915/intel_display.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_display.c +2 -1
index 0ac213a..2221323 100644
@@ -2329,7 +2329,8 @@ static unsigned int intel_linear_alignment(struct drm_i915_private *dev_priv)
 {
 	if (INTEL_INFO(dev_priv)->gen >= 9)
 		return 256 * 1024;
-	else if (IS_BROADWATER(dev_priv) || IS_CRESTLINE(dev_priv))
+	else if (IS_BROADWATER(dev_priv) || IS_CRESTLINE(dev_priv) ||
+		 IS_VALLEYVIEW(dev_priv))
 		return 128 * 1024;
 	else if (INTEL_INFO(dev_priv)->gen >= 4)
 		return 4 * 1024;
Signature · 6 lines
2.3.6

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 2/3
#2 lore raw

Re: [PATCH 2/3] drm/i915: Align DSPSURF to 128k on VLV/CHV

Quoted 28 lines · On 06/11/2015 06:31 AM, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

VLV/CHV have problems with 4k aligned linear scanout buffers. The VLV docs got updated at some point to say that we need to align them to 128k, just like we do on gen4.

So far I've seen the problem manifest when the stride is an odd multiple of 512 bytes, and the surface address meets the following pattern '(addr & 0xf000) == 0x1000' (also == 0x2000 is problematic on VLV). The result is a starcase effect (so some pages get dropped maybe?), with a few pages here and there clearly getting scannout out at the wrong position.

I've not actually been able to reproduce this problem on gen4, so it's not clear of the issue is any way related to the 128k restrictions supposedly inherited from gen4. But let's hope the 128k alignment is sufficient to hide it all.

---
  drivers/gpu/drm/i915/intel_display.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 0ac213a..2221323 100644

drivers/gpu/drm/i915/intel_display.c static unsigned int intel_linear_alignment(struct drm_i915_private *dev_priv) · line 2336

Quoted 9 lines
drivers/gpu/drm/i915/intel_display.c +2 -1
@@ -2329,7 +2329,8 @@ static unsigned int intel_linear_alignment(struct drm_i915_private *dev_priv)
  {
  	if (INTEL_INFO(dev_priv)->gen >= 9)
  		return 256 * 1024;
-	else if (IS_BROADWATER(dev_priv) || IS_CRESTLINE(dev_priv))
+	else if (IS_BROADWATER(dev_priv) || IS_CRESTLINE(dev_priv) ||
+		 IS_VALLEYVIEW(dev_priv))
  		return 128 * 1024;
  	else if (INTEL_INFO(dev_priv)->gen >= 4)
  		return 4 * 1024;

Fairly straight forward now that we know VLV and CHV require the 128k alignment.

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

Reviewed-by on v1 2/3
#3 lore raw

Re: [PATCH 2/3] drm/i915: Align DSPSURF to 128k on VLV/CHV

In reply to [email protected]
Quoted 12 lines · On Thursday 11 June 2015 07:01 PM, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

VLV/CHV have problems with 4k aligned linear scanout buffers. The VLV docs got updated at some point to say that we need to align them to 128k, just like we do on gen4.

So far I've seen the problem manifest when the stride is an odd multiple of 512 bytes, and the surface address meets the following pattern '(addr & 0xf000) == 0x1000' (also == 0x2000 is problematic on VLV). The result is a starcase effect (so some pages get dropped maybe?), with a few pages here and there clearly getting scannout out at the wrong position.

Quoted 6 lines

I've not actually been able to reproduce this problem on gen4, so it's not clear of the issue is any way related to the 128k restrictions supposedly inherited from gen4. But let's hope the 128k alignment is sufficient to hide it all.

Thanks and Regards,
Arun R Murthy
-------------------

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

#4 lore raw

PATCH 3/3 drm/i915: Drop the 64k linear scanout alignment on gen2/3

In reply to [email protected]

From: Ville Syrjälä <[email protected]>

The docs don't support the 64k linear scanout alignment we impose on gen2/3. And it really makes no sense since we have no DSPSURF register, so the only thing that the hardware will see is the linear offset which will be just pixel aligned anyway.

There is one case where 64k comes into the picture, and that's FBC. The start of the line length buffer corresponds to a 64k aligned address of the uncompressed framebuffer. So if the uncompressed fb is not 64k aligned, the first actually used entry in the line length buffer will not be byte 0. There are 32 extra entries in the line length buffer to account for this extra alignment so we shouldn't have to worry about it when mapping the uncompressed fb to the GTT.

---
 drivers/gpu/drm/i915/intel_display.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_display.c +1 -1
index 2221323..b48c237 100644
@@ -2335,7 +2335,7 @@ static unsigned int intel_linear_alignment(struct drm_i915_private *dev_priv)
 	else if (INTEL_INFO(dev_priv)->gen >= 4)
 		return 4 * 1024;
 	else
-		return 64 * 1024;
+		return 0;
 }
 
 int
Signature · 6 lines
2.3.6

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 3/3
#5 lore raw

Re: [PATCH 3/3] drm/i915: Drop the 64k linear scanout alignment on gen2/3

Quoted 15 lines · On Thu, Jun 11, 2015 at 04:31:16PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

The docs don't support the 64k linear scanout alignment we impose on gen2/3. And it really makes no sense since we have no DSPSURF register, so the only thing that the hardware will see is the linear offset which will be just pixel aligned anyway.

There is one case where 64k comes into the picture, and that's FBC. The start of the line length buffer corresponds to a 64k aligned address of the uncompressed framebuffer. So if the uncompressed fb is not 64k aligned, the first actually used entry in the line length buffer will not be byte 0. There are 32 extra entries in the line length buffer to account for this extra alignment so we shouldn't have to worry about it when mapping the uncompressed fb to the GTT.

Quoted 1 line

I could not find any rationale either, and it worked this way ~10 years ago, so

-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Tested-by on v1 3/3
#6 lore raw

Re: [PATCH 3/3] drm/i915: Drop the 64k linear scanout alignment on gen2/3

In reply to [email protected]

Task id: 6570 -------------------------------------Summary-------------------------------------
Platform Delta drm-intel-nightly Series Applied

PNV                                  276/276              276/276
ILK                                  303/303              303/303
SNB                                  312/312              312/312
IVB                                  343/343              343/343
BYT                                  287/287              287/287
BDW                                  321/321              321/321

-------------------------------------Detailed-------------------------------------
Platform Test drm-intel-nightly Series Applied
Note: You need to pay more attention to line start with '*'
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Reviewed-by on v1 1/3
#7 lore raw
In reply to [email protected]
Quoted 21 lines · On Thu, Jun 11, 2015 at 04:31:14PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently intel_gen4_compute_page_offset() simply picks the closest page boundary below the linear offset. That however may not be suitably aligned to satisfy any hardware specific restrictions. So let's make sure the page boundary we choose is properly aligned.

Also to play it a bit safer lets split the remaining linear offset into x and y values instead of just x. This should make no difference for most platforms since we convert the x and y offsets back into a linear offset before feeding them to the hardware. HSW+ are different however and use x and y offsets even with linear buffers, so they might have trouble if either the x or y get too big.

---

drivers/gpu/drm/i915/intel_display.c +0 -0
@@ -2455,12 +2461,13 @@ unsigned long intel_gen4_compute_page_offset(int *x, int *y,
 
 		return tile_rows * pitch * 8 + tiles * 4096;
 	} else {

drivers/gpu/drm/i915/intel_display.c unsigned long intel_gen4_compute_page_offset(int *x, int *y, · line 2470

Quoted 10 lines
drivers/gpu/drm/i915/intel_display.c +4 -3
+		unsigned int alignment = intel_linear_alignment(dev_priv) - 1;
 		unsigned int offset;
 
 		offset = *y * pitch + *x * cpp;
-		*y = 0;
-		*x = (offset & 4095) / cpp;
-		return offset & -4096;
+		*y = (offset & alignment) / pitch;
+		*x = ((offset & alignment) - *y * pitch) / cpp;
+		return offset & ~alignment;

Calculation looks solid. I presume we have a igt/kms test that combines linear/tiled, large surfaces and large offsets?

-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#8 lore raw
Quoted 25 lines · Chris Wilson
Quoted 24 lines · On Thu, Jun 11, 2015 at 04:31:14PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently intel_gen4_compute_page_offset() simply picks the closest page boundary below the linear offset. That however may not be suitably aligned to satisfy any hardware specific restrictions. So let's make sure the page boundary we choose is properly aligned.

Also to play it a bit safer lets split the remaining linear offset into x and y values instead of just x. This should make no difference for most platforms since we convert the x and y offsets back into a linear offset before feeding them to the hardware. HSW+ are different however and use x and y offsets even with linear buffers, so they might have trouble if either the x or y get too big.

---

drivers/gpu/drm/i915/intel_display.c +1 -0
@@ -2455,12 +2461,13 @@ unsigned long intel_gen4_compute_page_offset(int *x, int *y,
 
 		return tile_rows * pitch * 8 + tiles * 4096;
 	} else {
+		unsigned int alignment = intel_linear_alignment(dev_priv) - 1;
 		unsigned int offset;
 
Quoted 10 lines

drivers/gpu/drm/i915/intel_display.c unsigned long intel_gen4_compute_page_offset(int *x, int *y, · line 2470

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +3 -3
 		offset = *y * pitch + *x * cpp;
-		*y = 0;
-		*x = (offset & 4095) / cpp;
-		return offset & -4096;
+		*y = (offset & alignment) / pitch;
+		*x = ((offset & alignment) - *y * pitch) / cpp;
+		return offset & ~alignment;

Calculation looks solid. I presume we have a igt/kms test that combines linear/tiled, large surfaces and large offsets?

kms_plane has some kind of panning tests. Probably not as good as it could/should be. I have a few custom tests I created to hunt for the VLV/CHV bug, but those aren't really useable as regular igt tests as is. Would take a bit of extra effort to turn them into such.

Quoted 6 lines

-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#9 lore raw
Quoted 37 lines · Ville Syrjälä
Quoted 25 lines · Chris Wilson
Quoted 24 lines · On Thu, Jun 11, 2015 at 04:31:14PM +0300, [email protected] wrote:

From: Ville Syrjälä <[email protected]>

Currently intel_gen4_compute_page_offset() simply picks the closest page boundary below the linear offset. That however may not be suitably aligned to satisfy any hardware specific restrictions. So let's make sure the page boundary we choose is properly aligned.

Also to play it a bit safer lets split the remaining linear offset into x and y values instead of just x. This should make no difference for most platforms since we convert the x and y offsets back into a linear offset before feeding them to the hardware. HSW+ are different however and use x and y offsets even with linear buffers, so they might have trouble if either the x or y get too big.

---

drivers/gpu/drm/i915/intel_display.c +1 -0
@@ -2455,12 +2461,13 @@ unsigned long intel_gen4_compute_page_offset(int *x, int *y,
 
 		return tile_rows * pitch * 8 + tiles * 4096;
 	} else {
+		unsigned int alignment = intel_linear_alignment(dev_priv) - 1;
 		unsigned int offset;
 
Quoted 10 lines

drivers/gpu/drm/i915/intel_display.c unsigned long intel_gen4_compute_page_offset(int *x, int *y, · line 2470

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +3 -3
 		offset = *y * pitch + *x * cpp;
-		*y = 0;
-		*x = (offset & 4095) / cpp;
-		return offset & -4096;
+		*y = (offset & alignment) / pitch;
+		*x = ((offset & alignment) - *y * pitch) / cpp;
+		return offset & ~alignment;

Calculation looks solid. I presume we have a igt/kms test that combines linear/tiled, large surfaces and large offsets?

Quoted 4 lines

kms_plane has some kind of panning tests. Probably not as good as it could/should be. I have a few custom tests I created to hunt for the VLV/CHV bug, but those aren't really useable as regular igt tests as is. Would take a bit of extra effort to turn them into such.

We'd need a crc based testcase where we first draw a test picture at (0,0) and then with a hilariously large buffer at some offset. Evil values seems to be anything > 2048 iirc.

Testcase for this would be really good since I think skl is all broken too.
-Daniel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx