Skip to contentSkip to conversation

drm/i915/skl: Increase ddb blocks to support large cursor sizes

PATCH Superseded Radhakrishna Sripada 6 participants 14 messages Dec 17, 2015 – Feb 18, 2016

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915/skl: Increase ddb blocks to support large cursor sizes

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#1 lore raw
Quoted 8 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700
Quoted 7 lines

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

Quoted 17 lines
---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 5 lines
Jani Nikula, Intel Open Source Technology Center
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw
In reply to Radhakrishna Sripada
Quoted 9 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Quoted 30 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw
Quoted 5 lines · Ville Syrjälä
Quoted 4 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

Quoted 9 lines
Quoted 5 lines

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function) and a full-screen primary plane, so the numbers don't actually change (for his use case).

Of course that bug is still worth fixing too; I was planning on writing up a patch for it later today.

Matt

Quoted 34 lines
Quoted 30 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 9 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw
Quoted 12 lines · Matt Roper
Quoted 11 lines · Ville Syrjälä
Quoted 9 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Quoted 8 lines
Quoted 3 lines

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Pop quiz: what does it use then?

Quoted 51 lines

and a full-screen primary plane, so the numbers don't actually change (for his use case).

Of course that bug is still worth fixing too; I was planning on writing up a patch for it later today.

Matt

Quoted 34 lines
Quoted 30 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 5 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw
Quoted 17 lines · Ville Syrjälä
Quoted 16 lines · Matt Roper
Quoted 14 lines · Ville Syrjälä
Quoted 9 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Quoted 6 lines
Quoted 4 lines

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here is bumping that number up.

Primary and sprite planes are supposed to divide up the remaining blocks proportional to their size, but the bug here causes them to all be considered full-screen size. If you're not actually using sprites and not windowing your primary plane, then the bug has no effect (which is probably why we didn't already catch and fix it). If you do use a sprite plane or window your primary plane, your proportions are probably wrong and you get non-optimal settings, although even then you usually won't have actual problems.

Matt

Quoted 56 lines
Quoted 52 lines

and a full-screen primary plane, so the numbers don't actually change (for his use case).

Of course that bug is still worth fixing too; I was planning on writing up a patch for it later today.

Matt

Quoted 34 lines
Quoted 30 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 5 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 9 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#6 lore raw
Quoted 23 lines · Matt Roper
Quoted 22 lines · Ville Syrjälä
Quoted 20 lines · Matt Roper
Quoted 5 lines · Ville Syrjälä
Quoted 4 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

Quoted 9 lines
Quoted 5 lines

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 7 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here is bumping that number up.

To elaborate on this some more, we have some suspicions about why the bspec-suggested 32 blocks might not be enough. There's some new workaround information in the bspec about sometimes needing to disable the system agent (SAGV) on non-BXT gen9, depending on the configuration and that's not something that's been implemented in our driver code yet. There's also another new workaround that checks raw system memory bandwidth and adjusts watermark programming that hasn't been implemented yet either. It could be that one of those two missing workarounds is the true culprit, but Radhakrishna's patch here looks like a reasonable short-term workaround; my main worry is that if he needs to bump the hardcoded 32 up to to 52 for the single pipe case, he might also need to bump the hardcoded 8 up as well for the multi-pipe case; I don't think anyone has tested that yet (and this seems to be a SKL-specific problem, so I can't reproduce it on my BXT).

Matt

Quoted 76 lines

Primary and sprite planes are supposed to divide up the remaining blocks proportional to their size, but the bug here causes them to all be considered full-screen size. If you're not actually using sprites and not windowing your primary plane, then the bug has no effect (which is probably why we didn't already catch and fix it). If you do use a sprite plane or window your primary plane, your proportions are probably wrong and you get non-optimal settings, although even then you usually won't have actual problems.

Matt

Quoted 56 lines
Quoted 52 lines

and a full-screen primary plane, so the numbers don't actually change (for his use case).

Of course that bug is still worth fixing too; I was planning on writing up a patch for it later today.

Matt

Quoted 34 lines
Quoted 30 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 5 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 5 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
Signature · 9 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#7 lore raw
Quoted 32 lines · Matt Roper
Quoted 23 lines · Matt Roper
Quoted 22 lines · Ville Syrjälä
Quoted 20 lines · Matt Roper
Quoted 5 lines · Ville Syrjälä
Quoted 4 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

Quoted 9 lines
Quoted 5 lines

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 7 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here is bumping that number up.

Quoted 14 lines

To elaborate on this some more, we have some suspicions about why the bspec-suggested 32 blocks might not be enough. There's some new workaround information in the bspec about sometimes needing to disable the system agent (SAGV) on non-BXT gen9, depending on the configuration and that's not something that's been implemented in our driver code yet. There's also another new workaround that checks raw system memory bandwidth and adjusts watermark programming that hasn't been implemented yet either. It could be that one of those two missing workarounds is the true culprit, but Radhakrishna's patch here looks like a reasonable short-term workaround; my main worry is that if he needs to bump the hardcoded 32 up to to 52 for the single pipe case, he might also need to bump the hardcoded 8 up as well for the multi-pipe case; I don't think anyone has tested that yet (and this seems to be a SKL-specific problem, so I can't reproduce it on my BXT).

This needs at least a very big comment that we're just hacking around with duct-tape. Also allocating DDB shouldn't matter for correctness, as long as we compute the WM values correctly. I wonder how this fixes anything really (except making it a bit harder to hit perhaps due to the enlarged buffering it affords for the cursor).

Imo better to just implement the other workarounds first, then see what's left.
-Daniel

Quoted 91 lines

Matt

Quoted 76 lines

Primary and sprite planes are supposed to divide up the remaining blocks proportional to their size, but the bug here causes them to all be considered full-screen size. If you're not actually using sprites and not windowing your primary plane, then the bug has no effect (which is probably why we didn't already catch and fix it). If you do use a sprite plane or window your primary plane, your proportions are probably wrong and you get non-optimal settings, although even then you usually won't have actual problems.

Matt

Quoted 56 lines
Quoted 52 lines

and a full-screen primary plane, so the numbers don't actually change (for his use case).

Of course that bug is still worth fixing too; I was planning on writing up a patch for it later today.

Matt

Quoted 34 lines
Quoted 30 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
 drivers/gpu/drm/i915/intel_pm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
 {
 	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
 	return 8;
 }
Signature · 6 lines
1.9.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 5 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
Signature · 2 lines
Ville Syrjälä
Intel OTC
Signature · 5 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
Signature · 9 lines
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
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
#8 lore raw
Quoted 48 lines · Daniel Vetter
Quoted 32 lines · Matt Roper
Quoted 23 lines · Matt Roper
Quoted 22 lines · Ville Syrjälä
Quoted 20 lines · Matt Roper
Quoted 5 lines · Ville Syrjälä
Quoted 4 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

Quoted 9 lines
Quoted 5 lines

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 7 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here is bumping that number up.

Quoted 14 lines

To elaborate on this some more, we have some suspicions about why the bspec-suggested 32 blocks might not be enough. There's some new workaround information in the bspec about sometimes needing to disable the system agent (SAGV) on non-BXT gen9, depending on the configuration and that's not something that's been implemented in our driver code yet. There's also another new workaround that checks raw system memory bandwidth and adjusts watermark programming that hasn't been implemented yet either. It could be that one of those two missing workarounds is the true culprit, but Radhakrishna's patch here looks like a reasonable short-term workaround; my main worry is that if he needs to bump the hardcoded 32 up to to 52 for the single pipe case, he might also need to bump the hardcoded 8 up as well for the multi-pipe case; I don't think anyone has tested that yet (and this seems to be a SKL-specific problem, so I can't reproduce it on my BXT).

Quoted 5 lines

This needs at least a very big comment that we're just hacking around with duct-tape. Also allocating DDB shouldn't matter for correctness, as long as we compute the WM values correctly. I wonder how this fixes anything really (except making it a bit harder to hit perhaps due to the enlarged buffering it affords for the cursor).

Agree Daniel, but in our detailed testing this is found to have fixed reliably all known flickers. I recommend to have a detailed comment and *put* the duct-tape. I can have a rerun of detailed tests again to ensure that its indeed fixing everything.

Quoted 3 lines

Imo better to just implement the other workarounds first, then see what's left.

I know there are few known workarounds pending. Lets continue implementing them and more. But with Matt's patch not fixing the issue, lets take this as an immediate short term fix and come back to remove the duct-tape when we have correct fix.

Regards
Shobhit

Quoted 94 lines

-Daniel

Quoted 91 lines

Matt

Quoted 70 lines

Primary and sprite planes are supposed to divide up the remaining blocks proportional to their size, but the bug here causes them to all be considered full-screen size. If you're not actually using sprites and not windowing your primary plane, then the bug has no effect (which is probably why we didn't already catch and fix it). If you do use a sprite plane or window your primary plane, your proportions are probably wrong and you get non-optimal settings, although even then you usually won't have actual problems.

Matt

Quoted 53 lines
Quoted 46 lines

and a full-screen primary plane, so the numbers don't actually change (for his use case).

Of course that bug is still worth fixing too; I was planning on writing up a patch for it later today.

Matt

Quoted 26 lines
Quoted 26 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate SAGV block time for the maximum possible cursor size.

---
  drivers/gpu/drm/i915/intel_pm.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -2
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
  static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
  {
  	if (config->num_pipes_active == 1)
-		return 32;
+		return 52;
 
  	return 8;
  }
--

1.9.1

Quoted 8 lines
Quoted 4 lines

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

--
Ville Syrjälä
Intel OTC

Quoted 6 lines

--
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795

Quoted 3 lines

--
Ville Syrjälä
Intel OTC

Quoted 6 lines

--
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795

--
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

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

#9 lore raw
Quoted 66 lines · Shobhit
Quoted 52 lines · Daniel Vetter
Quoted 36 lines · Matt Roper
Quoted 26 lines · Matt Roper
Quoted 25 lines · Ville Syrjälä
Quoted 23 lines · Matt Roper
Quoted 11 lines · Ville Syrjälä
Quoted 9 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of
256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700
Quoted 6 lines
Quoted 2 lines

drm/i915/skl: Eliminate usage of pipe_wm_parameters from
SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 8 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here
is bumping that number up.

Quoted 14 lines

To elaborate on this some more, we have some suspicions about why the bspec-suggested 32 blocks might not be enough. There's some new workaround information in the bspec about sometimes needing to disable the system agent (SAGV) on non-BXT gen9, depending on the configuration and that's not something that's been implemented in our driver code yet. There's also another new workaround that checks raw system memory bandwidth and adjusts watermark programming that hasn't been implemented yet either. It could be that one of those two missing workarounds is the true culprit, but Radhakrishna's patch here looks like a reasonable short-term workaround; my main worry is that if he needs to bump the hardcoded 32 up to to 52 for the single pipe case, he might also need to bump the hardcoded 8 up as well for the multi-pipe case; I don't think anyone has tested that yet (and this seems to be a SKL-specific problem, so I can't reproduce it on my BXT).

Quoted 6 lines

This needs at least a very big comment that we're just hacking around with
duct-tape. Also allocating DDB shouldn't matter for correctness, as long as we compute the WM values correctly. I wonder how this fixes anything really (except making it a bit harder to hit perhaps due to the enlarged buffering it affords for the cursor).

Agree Daniel, but in our detailed testing this is found to have fixed reliably all known flickers. I recommend to have a detailed comment and *put* the duct-tape. I can have a rerun of detailed tests again to ensure that its indeed fixing everything.

Quoted 1 line
Quoted 7 lines
Quoted 2 lines

Imo better to just implement the other workarounds first, then see what's left.

I know there are few known workarounds pending. Lets continue implementing them and more. But with Matt's patch not fixing the issue, lets take this as an immediate short term fix and come back to remove the duct-tape when we have correct fix.

I experimented and found that reverting this patch but having arbitrated display bandwidth based workaround to increase latency values for all levels is fixing the flicker. Will be working on this along with Mahesh, unless Matt, you have a patch for this already.

Regards
Shobhit

Quoted 106 lines

Regards
Shobhit

Quoted 102 lines

-Daniel

Quoted 99 lines

Matt

Quoted 78 lines

Primary and sprite planes are supposed to divide up the remaining blocks
proportional to their size, but the bug here causes them to all be considered full-screen size. If you're not actually using sprites and not windowing your primary plane, then the bug has no effect (which is probably why we didn't already catch and fix it). If you do use a sprite plane or window your primary plane, your proportions are probably
wrong and you get non-optimal settings, although even then you usually won't have actual problems.

Matt

Quoted 59 lines
Quoted 52 lines

and a full-screen primary plane, so the numbers don't actually change (for his use case).

Of course that bug is still worth fixing too; I was planning on writing
up a patch for it later today.

Matt

Quoted 31 lines
Quoted 31 lines

are used. Increasing the number of blocks to 52 to make cursor plane tolerate
SAGV block time for the maximum possible cursor size.

Signed-off-by: Radhakrishna Sripada
<[email protected]>

---
  drivers/gpu/drm/i915/intel_pm.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_pm.c +1 -1
index d385d99..137fb68 100644
@@ -2802,7 +2802,7 @@ skl_ddb_get_pipe_allocation_limits(struct
 drm_device *dev,
  static unsigned int skl_cursor_allocation(const struct
 intel_wm_config *config)
  {
      if (config->num_pipes_active == 1)
-        return 32;
+        return 52;
 
      return 8;

}
--
1.9.1

Quoted 8 lines
Quoted 4 lines

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

--
Ville Syrjälä
Intel OTC

Quoted 6 lines

--
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795

Quoted 3 lines

--
Ville Syrjälä
Intel OTC

Quoted 6 lines

--
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795

--
Matt Roper
Graphics Software Engineer
IoTG Platform Enabling & Development
Intel Corporation
(916) 356-2795
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Quoted 4 lines

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

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

#10 lore raw
Quoted 75 lines · Shobhit
Quoted 66 lines · Shobhit
Quoted 52 lines · Daniel Vetter
Quoted 36 lines · Matt Roper
Quoted 26 lines · Matt Roper
Quoted 25 lines · Ville Syrjälä
Quoted 23 lines · Matt Roper
Quoted 11 lines · Ville Syrjälä
Quoted 9 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of
256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700
Quoted 6 lines
Quoted 2 lines

drm/i915/skl: Eliminate usage of pipe_wm_parameters from
SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 8 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here
is bumping that number up.

Quoted 14 lines

To elaborate on this some more, we have some suspicions about why the bspec-suggested 32 blocks might not be enough. There's some new workaround information in the bspec about sometimes needing to disable the system agent (SAGV) on non-BXT gen9, depending on the configuration and that's not something that's been implemented in our driver code yet. There's also another new workaround that checks raw system memory bandwidth and adjusts watermark programming that hasn't been implemented yet either. It could be that one of those two missing workarounds is the true culprit, but Radhakrishna's patch here looks like a reasonable short-term workaround; my main worry is that if he needs to bump the hardcoded 32 up to to 52 for the single pipe case, he might also need to bump the hardcoded 8 up as well for the multi-pipe case; I don't think anyone has tested that yet (and this seems to be a SKL-specific problem, so I can't reproduce it on my BXT).

Quoted 6 lines

This needs at least a very big comment that we're just hacking around with
duct-tape. Also allocating DDB shouldn't matter for correctness, as long as we compute the WM values correctly. I wonder how this fixes anything really (except making it a bit harder to hit perhaps due to the enlarged buffering it affords for the cursor).

Agree Daniel, but in our detailed testing this is found to have fixed reliably all known flickers. I recommend to have a detailed comment and *put* the duct-tape. I can have a rerun of detailed tests again to ensure that its indeed fixing everything.

Quoted 1 line
Quoted 7 lines
Quoted 2 lines

Imo better to just implement the other workarounds first, then see what's left.

I know there are few known workarounds pending. Lets continue implementing them and more. But with Matt's patch not fixing the issue, lets take this as an immediate short term fix and come back to remove the duct-tape when we have correct fix.

Quoted 4 lines

I experimented and found that reverting this patch but having arbitrated display bandwidth based workaround to increase latency values for all levels is fixing the flicker. Will be working on this along with Mahesh, unless Matt, you have a patch for this already.

Excellent, really happy that we can go ahead with a reasonable looking fix right away here.

Thanks, 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
#11 lore raw
Quoted 76 lines · Daniel Vetter
Quoted 75 lines · Shobhit
Quoted 73 lines · Shobhit
Quoted 52 lines · Daniel Vetter
Quoted 36 lines · Matt Roper
Quoted 26 lines · Matt Roper
Quoted 25 lines · Ville Syrjälä
Quoted 23 lines · Matt Roper
Quoted 11 lines · Ville Syrjälä
Quoted 9 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of
256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700
Quoted 6 lines
Quoted 2 lines

drm/i915/skl: Eliminate usage of pipe_wm_parameters from
SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 8 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here
is bumping that number up.

Quoted 14 lines

To elaborate on this some more, we have some suspicions about why the bspec-suggested 32 blocks might not be enough. There's some new workaround information in the bspec about sometimes needing to disable the system agent (SAGV) on non-BXT gen9, depending on the configuration and that's not something that's been implemented in our driver code yet. There's also another new workaround that checks raw system memory bandwidth and adjusts watermark programming that hasn't been implemented yet either. It could be that one of those two missing workarounds is the true culprit, but Radhakrishna's patch here looks like a reasonable short-term workaround; my main worry is that if he needs to bump the hardcoded 32 up to to 52 for the single pipe case, he might also need to bump the hardcoded 8 up as well for the multi-pipe case; I don't think anyone has tested that yet (and this seems to be a SKL-specific problem, so I can't reproduce it on my BXT).

Quoted 6 lines

This needs at least a very big comment that we're just hacking around with
duct-tape. Also allocating DDB shouldn't matter for correctness, as long as we compute the WM values correctly. I wonder how this fixes anything really (except making it a bit harder to hit perhaps due to the enlarged buffering it affords for the cursor).

Agree Daniel, but in our detailed testing this is found to have fixed reliably all known flickers. I recommend to have a detailed comment and *put* the duct-tape. I can have a rerun of detailed tests again to ensure that its indeed fixing everything.

Quoted 3 lines

Imo better to just implement the other workarounds first, then see what's left.

I know there are few known workarounds pending. Lets continue implementing them and more. But with Matt's patch not fixing the issue, lets take this as an immediate short term fix and come back to remove the duct-tape when we have correct fix.

Quoted 7 lines
Quoted 4 lines

I experimented and found that reverting this patch but having arbitrated display bandwidth based workaround to increase latency values for all levels is fixing the flicker. Will be working on this along with Mahesh, unless Matt, you have a patch for this already.

Excellent, really happy that we can go ahead with a reasonable looking fix right away here.

Daniel, we implemented the memory latency increase for quick test and it worked, but in reality that situation was not hitting even with a QHD eDP and 4k@60 DP with a two channel 1867 MHz memory. So this was not a solution.

There have been another patch series to try to fix this that I posted -
https://patchwork.freedesktop.org/patch/71758/

and we have been trying with lot of Matt's patches for watermark, and along with what I posted but all of this has not yielded perfect under run error free results though improved a little.

But cursor DDB bump for both internal and external pipe is resulting in complete fix of the problem. As of now that is the only solution I can think off and has been extensively tested with QHD eDP/ 19x10 eDP + 4k DP/HDMI and working reliably. I think we should be dynamically allocating cursor DDB as well so that it will increase the DDB if cursor size is big.

But for now proposal is to statically increase these cursor DDB to 52 for single pipe as in this patch and additionally have it 32 in case of multi pipe scenario. Once we have all water mark thingy sorted out, we can come back and re-look at this.

Regards
Shobhit

Quoted 3 lines

Thanks, Daniel

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

#12 lore raw
Quoted 121 lines · Shobhit
Quoted 97 lines · Daniel Vetter
Quoted 90 lines · Shobhit
Quoted 80 lines · Shobhit
Quoted 63 lines · Daniel Vetter
Quoted 53 lines · Matt Roper
Quoted 28 lines · Matt Roper
Quoted 27 lines · Ville Syrjälä
Quoted 25 lines · Matt Roper
Quoted 11 lines · Ville Syrjälä
Quoted 9 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of
256x256 causing FIFO underruns when the reworked wm caluclations in

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700
Quoted 7 lines
Quoted 2 lines

drm/i915/skl: Eliminate usage of pipe_wm_parameters from
SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src
width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided
that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 9 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB
allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here
is bumping that number up.

To elaborate on this some more, we have some suspicions about why the bspec-suggested 32 blocks might not be enough. There's some new workaround information in the bspec about sometimes needing to disable
the system agent (SAGV) on non-BXT gen9, depending on the configuration
and that's not something that's been implemented in our driver code yet.
There's also another new workaround that checks raw system memory bandwidth and adjusts watermark programming that hasn't been implemented
yet either. It could be that one of those two missing workarounds is the true culprit, but Radhakrishna's patch here looks like a reasonable

Quoted 8 lines

short-term workaround; my main worry is that if he needs to bump the hardcoded 32 up to to 52 for the single pipe case, he might also need to
bump the hardcoded 8 up as well for the multi-pipe case; I don't think
anyone has tested that yet (and this seems to be a SKL-specific problem,
so I can't reproduce it on my BXT).

Quoted 9 lines

This needs at least a very big comment that we're just hacking around with
duct-tape. Also allocating DDB shouldn't matter for correctness, as long
as we compute the WM values correctly. I wonder how this fixes anything
really (except making it a bit harder to hit perhaps due to the enlarged
buffering it affords for the cursor).

Agree Daniel, but in our detailed testing this is found to have fixed reliably all known flickers. I recommend to have a detailed comment and *put* the duct-tape. I can have a rerun of detailed tests again to ensure that its indeed fixing everything.

Quoted 1 line
Quoted 8 lines
Quoted 3 lines

Imo better to just implement the other workarounds first, then see what's
left.

I know there are few known workarounds pending. Lets continue implementing them and more. But with Matt's patch not fixing the issue, lets take this as an immediate short term fix and come back to remove the duct-tape when we have correct fix.

Quoted 5 lines

I experimented and found that reverting this patch but having arbitrated display bandwidth based workaround to increase latency values for all levels
is fixing the flicker. Will be working on this along with Mahesh, unless Matt, you have a patch for this already.

Quoted 3 lines

Excellent, really happy that we can go ahead with a reasonable looking fix
right away here.

Daniel, we implemented the memory latency increase for quick test and it worked, but in reality that situation was not hitting even with a QHD eDP and 4k@60 DP with a two channel 1867 MHz memory. So this was not a solution.

There have been another patch series to try to fix this that I posted -
https://patchwork.freedesktop.org/patch/71758/

and we have been trying with lot of Matt's patches for watermark, and along with what I posted but all of this has not yielded perfect under run error free results though improved a little.

But cursor DDB bump for both internal and external pipe is resulting in complete fix of the problem. As of now that is the only solution I can think off and has been extensively tested with QHD eDP/ 19x10 eDP + 4k DP/HDMI and working reliably. I think we should be dynamically allocating cursor DDB as well so that it will increase the DDB if cursor size is big.

Quoted 4 lines

But for now proposal is to statically increase these cursor DDB to 52 for single pipe as in this patch and additionally have it 32 in case of multi pipe scenario. Once we have all water mark thingy sorted out, we can come back and re-look at this.

Daniel, any comments or suggestions ?

Regards
Shobhit

Quoted 6 lines

Regards
Shobhit

Quoted 1 line
Quoted 6 lines
Quoted 2 lines

Thanks, Daniel

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

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

#13 lore raw
In reply to Matt Roper
Quoted 23 lines · Matt Roper
Quoted 22 lines · Ville Syrjälä
Quoted 20 lines · Matt Roper
Quoted 5 lines · Ville Syrjälä
Quoted 4 lines · Radhakrishna Sripada

Original value of 32 blocks is not sufficient when using cursor size of 256x256 causing FIFO underruns when the reworked wm caluclations in

Quoted 9 lines
Quoted 5 lines

commit 024c9045221fe45482863c47c4b4c47d37f97cbf

Author: Matt Roper <[email protected]>
Date:   Thu Sep 24 15:53:11 2015 -0700

drm/i915/skl: Eliminate usage of pipe_wm_parameters from SKL-style WM (v4)

Well that commit is obviously incorrect. It's now using the pipe src width as the plane width for all planes.

Yeah, we already noted that bug in another email thread, but decided that it was unrelated to the problems Radhakrishna is facing. Radhakrishna is only using a cursor (which doesn't use that buggy function)

Quoted 6 lines
Quoted 1 line

Pop quiz: what does it use then?

All non-cursor planes (i.e., primary+sprite). Cursors use a fixed DDB allocation (currently 32 blocks as suggested by bspec, but Radhakrishna's testing has found this to be too small, so his patch here is bumping that number up.

Oh I'm talking about the WM calculation, not the DDB allocation.

Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx