Skip to contentSkip to conversation

drm/i915: Redo WMs when cursor size changes

PATCH In review Joe Konno 4 participants 6 messages Feb 26, 2015 – Feb 27, 2015

Subsystems (2)
lore
Share & tools
lore

Joe Konno

author
#0 lore raw

PATCH 1/1 drm/i915: Redo WMs when cursor size changes

From: Joe Konno <[email protected]>

In instances where cursor sizes change, as in Chromium Ozone/Freon, watermarks should be recomputed. There should be no hard-coded assumptions about cursor widths. This was corrected originally here:

commit 64f962e3e38bf6f40bbd2462f8380dee0369e1bf
Author: Chris Wilson <[email protected]>
Date: Wed Mar 26 12:38:15 2014 +0000

drm/i915: Recompute WM when the cursor size changes

However, it seems the recompute logic got lost in refactoring. Re-introduce the relevant WM re-compute code from the original patch.

Cc Chris Wilson <[email protected]>
Cc Matt Roper <[email protected]>
Fixes 32b7eeec4d1e ("drm/i915: Refactor work that can sleep out of commit (v7)")

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89346

---
 drivers/gpu/drm/i915/intel_display.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_display.c +6 -1
index b0f113d9daab..c5cbfea0551e 100644
@@ -12234,6 +12234,7 @@ intel_commit_cursor_plane(struct drm_plane *plane,
 	struct intel_plane *intel_plane = to_intel_plane(plane);
 	struct drm_i915_gem_object *obj = intel_fb_obj(state->base.fb);
 	uint32_t addr;
+	unsigned old_width;
 
 	crtc = crtc ? crtc : plane->crtc;
 	intel_crtc = to_intel_crtc(crtc);
@@ -12243,6 +12244,7 @@ intel_commit_cursor_plane(struct drm_plane *plane,
 	crtc->cursor_y = state->base.crtc_y;
 
 	intel_plane->obj = obj;
+	old_width = intel_crtc->cursor_width;
 
 	if (intel_crtc->cursor_bo == obj)
 		goto update;
@@ -12260,8 +12262,11 @@ update:
 	intel_crtc->cursor_width = state->base.crtc_w;
 	intel_crtc->cursor_height = state->base.crtc_h;
 
-	if (intel_crtc->active)
+	if (intel_crtc->active) {
+		if (old_width != intel_crtc->cursor_width)
+			intel_update_watermarks(crtc);
 		intel_crtc_update_cursor(crtc, state->visible);
+	}
 }
 
 static struct drm_plane *intel_cursor_plane_create(struct drm_device *dev,
Signature · 6 lines
2.3.0

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Tested-by on v1 1/1
#1 lore raw
Quoted 11 lines · Joe Konno

From: Joe Konno <[email protected]>

In instances where cursor sizes change, as in Chromium Ozone/Freon, watermarks should be recomputed. There should be no hard-coded assumptions about cursor widths. This was corrected originally here:

commit 64f962e3e38bf6f40bbd2462f8380dee0369e1bf
Author: Chris Wilson <[email protected]>
Date: Wed Mar 26 12:38:15 2014 +0000

Quoted 8 lines

drm/i915: Recompute WM when the cursor size changes

However, it seems the recompute logic got lost in refactoring. Re-introduce the relevant WM re-compute code from the original patch.

Cc Chris Wilson <[email protected]>
Cc Matt Roper <[email protected]>
Fixes 32b7eeec4d1e ("drm/i915: Refactor work that can sleep out of

commit (v7)")

Review

Quoted 7 lines

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89346

---
 drivers/gpu/drm/i915/intel_display.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/i915/intel_display.c

Quoted 22 lines

index b0f113d9daab..c5cbfea0551e 100644

drivers/gpu/drm/i915/intel_display.c +2 -0
@@ -12234,6 +12234,7 @@ intel_commit_cursor_plane(struct drm_plane *plane,
        struct intel_plane *intel_plane = to_intel_plane(plane);
        struct drm_i915_gem_object *obj = intel_fb_obj(state->base.fb);
        uint32_t addr;
+       unsigned old_width;
 
        crtc = crtc ? crtc : plane->crtc;
        intel_crtc = to_intel_crtc(crtc);
@@ -12243,6 +12244,7 @@ intel_commit_cursor_plane(struct drm_plane *plane,
        crtc->cursor_y = state->base.crtc_y;
 
        intel_plane->obj = obj;
+       old_width = intel_crtc->cursor_width;
 
        if (intel_crtc->cursor_bo == obj)
                goto update;
@@ -12260,8 +12262,11 @@ update:
        intel_crtc->cursor_width = state->base.crtc_w;
        intel_crtc->cursor_height = state->base.crtc_h;

drivers/gpu/drm/i915/intel_display.c update: · line 12272

Quoted 10 lines
drivers/gpu/drm/i915/intel_display.c +4 -1
 
-       if (intel_crtc->active)
+       if (intel_crtc->active) {
+               if (old_width != intel_crtc->cursor_width)
+                       intel_update_watermarks(crtc);
                intel_crtc_update_cursor(crtc, state->visible);
+       }
 }
 
 static struct drm_plane *intel_cursor_plane_create(struct drm_device

*dev,

Quoted 3 lines

--
2.3.0

Quoted 4 lines

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

--

James Ausmus
Sr. Software Engineer
SSG-OTC ChromeOS Integration

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

#2 lore raw
In reply to Joe Konno
Quoted 6 lines · Joe Konno

From: Joe Konno <[email protected]>

In instances where cursor sizes change, as in Chromium Ozone/Freon, watermarks should be recomputed. There should be no hard-coded assumptions about cursor widths. This was corrected originally here:

Quoted 8 lines

commit 64f962e3e38bf6f40bbd2462f8380dee0369e1bf
Author: Chris Wilson <[email protected]>
Date: Wed Mar 26 12:38:15 2014 +0000

drm/i915: Recompute WM when the cursor size changes

However, it seems the recompute logic got lost in refactoring. Re-introduce the relevant WM re-compute code from the original patch.

I don't believe the recompute logic got lost, it just got moved around a bit so that it wouldn't happen during our time-sensitive vblank evasion while interrupts are disabled. The old_width != intel_crtc->cursor_width check that you're re-adding below already happens in intel_check_cursor_plane():

   finish: 
           if (intel_crtc->active) {
                   if (intel_crtc->cursor_width != state->base.crtc_w)
                           intel_crtc->atomic.update_wm = true;

We just set a flag so that we know whether we'll need to update watermarks later or not. Then when we get to the commit phase, in intel_begin_crtc_commit(), we do the following before entering vblank evasion:

        if (intel_crtc->atomic.update_wm)
                intel_update_watermarks(crtc);

Some of the watermark code we have today sleeps, so we can't call it under vblank evasion (i.e., during the commit phase while interrupts are disabled).

Our high-level atomic workflow (which legacy plane ioctls also flow through internally) looks something like this:

- Some userspace ioctl (SetCursor, SetPlane, etc.)
- intel_check_TYPE_plane()
- check validity of request, bail if invalid
- set flags in intel_crtc->atomic for various stuff we will need to do before/after vblank evasion (e.g., wm's)
- intel_begin_crtc_commit()
- Perform sleepable operations noted in intel_crtc->atomic
- intel_pipe_update_start (interrupts are now disabled!)
- intel_commit_TYPE_plane()
- intel_finish_crtc_commit()
- intel_pipe_update_end (interrupts are now reenabled)
- Perform sleepable operations noted in intel_crtc->atomic

So your patch below could result in sleeps happening while vblanks are disabled, which is bad (IIRC, most of those sleeps are in the SKL codepath right now, but I think there's a workaround-related wait for IVB as well).

Our watermark code needs a lot of work to beat it into proper shape for atomic and that's what I'm working on at the moment.

Matt

Quoted 50 lines
Cc Chris Wilson <[email protected]>
Cc Matt Roper <[email protected]>
Fixes 32b7eeec4d1e ("drm/i915: Refactor work that can sleep out of commit (v7)")

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89346

---
 drivers/gpu/drm/i915/intel_display.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_display.c +6 -1
index b0f113d9daab..c5cbfea0551e 100644
@@ -12234,6 +12234,7 @@ intel_commit_cursor_plane(struct drm_plane *plane,
 	struct intel_plane *intel_plane = to_intel_plane(plane);
 	struct drm_i915_gem_object *obj = intel_fb_obj(state->base.fb);
 	uint32_t addr;
+	unsigned old_width;
 
 	crtc = crtc ? crtc : plane->crtc;
 	intel_crtc = to_intel_crtc(crtc);
@@ -12243,6 +12244,7 @@ intel_commit_cursor_plane(struct drm_plane *plane,
 	crtc->cursor_y = state->base.crtc_y;
 
 	intel_plane->obj = obj;
+	old_width = intel_crtc->cursor_width;
 
 	if (intel_crtc->cursor_bo == obj)
 		goto update;
@@ -12260,8 +12262,11 @@ update:
 	intel_crtc->cursor_width = state->base.crtc_w;
 	intel_crtc->cursor_height = state->base.crtc_h;
 
-	if (intel_crtc->active)
+	if (intel_crtc->active) {
+		if (old_width != intel_crtc->cursor_width)
+			intel_update_watermarks(crtc);
 		intel_crtc_update_cursor(crtc, state->visible);
+	}
 }
 
 static struct drm_plane *intel_cursor_plane_create(struct drm_device *dev,
Signature · 6 lines
2.3.0

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
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
#3 lore raw
Quoted 55 lines · Matt Roper
Quoted 6 lines · Joe Konno

From: Joe Konno <[email protected]>

In instances where cursor sizes change, as in Chromium Ozone/Freon, watermarks should be recomputed. There should be no hard-coded assumptions about cursor widths. This was corrected originally here:

Quoted 8 lines

commit 64f962e3e38bf6f40bbd2462f8380dee0369e1bf
Author: Chris Wilson <[email protected]>
Date: Wed Mar 26 12:38:15 2014 +0000

drm/i915: Recompute WM when the cursor size changes

However, it seems the recompute logic got lost in refactoring. Re-introduce the relevant WM re-compute code from the original patch.

I don't believe the recompute logic got lost, it just got moved around a bit so that it wouldn't happen during our time-sensitive vblank evasion while interrupts are disabled. The old_width != intel_crtc->cursor_width check that you're re-adding below already happens in intel_check_cursor_plane():

   finish: 
           if (intel_crtc->active) {
                   if (intel_crtc->cursor_width != state->base.crtc_w)
                           intel_crtc->atomic.update_wm = true;

We just set a flag so that we know whether we'll need to update watermarks later or not. Then when we get to the commit phase, in intel_begin_crtc_commit(), we do the following before entering vblank evasion:

        if (intel_crtc->atomic.update_wm)
                intel_update_watermarks(crtc);

Some of the watermark code we have today sleeps, so we can't call it under vblank evasion (i.e., during the commit phase while interrupts are disabled).

Our high-level atomic workflow (which legacy plane ioctls also flow through internally) looks something like this:

- Some userspace ioctl (SetCursor, SetPlane, etc.)
- intel_check_TYPE_plane()
- check validity of request, bail if invalid
- set flags in intel_crtc->atomic for various stuff we will need to do before/after vblank evasion (e.g., wm's)
- intel_begin_crtc_commit()
- Perform sleepable operations noted in intel_crtc->atomic
- intel_pipe_update_start (interrupts are now disabled!)
- intel_commit_TYPE_plane()
- intel_finish_crtc_commit()
- intel_pipe_update_end (interrupts are now reenabled)
- Perform sleepable operations noted in intel_crtc->atomic

Quoted 7 lines

So your patch below could result in sleeps happening while vblanks are disabled, which is bad (IIRC, most of those sleeps are in the SKL codepath right now, but I think there's a workaround-related wait for IVB as well).

Our watermark code needs a lot of work to beat it into proper shape for atomic and that's what I'm working on at the moment.

Alternative issue is that wm recompute happens before we update all the legacy state. Tvrkto just stumbled over that for some of his skl watermark code, he had to switch a few places in the wm code from looking at plane->fb to look at plane->state->fb.

Similar changes might be needed for the cursor wm code - cursor_width/height is kinda redundant since universal planes support. -Daniel

Signature · 7 lines
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
author
#4 lore raw
Quoted 9 lines · Daniel Vetter
Quoted 7 lines · Matt Roper

So your patch below could result in sleeps happening while vblanks are disabled, which is bad (IIRC, most of those sleeps are in the SKL codepath right now, but I think there's a workaround-related wait for IVB as well).

Our watermark code needs a lot of work to beat it into proper shape for atomic and that's what I'm working on at the moment.

Quoted 7 lines

Alternative issue is that wm recompute happens before we update all the legacy state. Tvrkto just stumbled over that for some of his skl watermark code, he had to switch a few places in the wm code from looking at plane->fb to look at plane->state->fb.

Similar changes might be needed for the cursor wm code - cursor_width/height is kinda redundant since universal planes support.

Thanks for the insights, Matt and Dan. It's clear to me from Matt's analysis that my patch is a hack at best and nowhere near a proper fix.

This patch was submitted in the hopes of fixing a substantial graphical regression. Shall my patch be retooled as a "better" work-around, or will the regression stand until the watermark code is beat into proper shape?
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#5 lore raw
Quoted 18 lines · Joe Konno
Quoted 9 lines · Daniel Vetter
Quoted 7 lines · Matt Roper

So your patch below could result in sleeps happening while vblanks are disabled, which is bad (IIRC, most of those sleeps are in the SKL codepath right now, but I think there's a workaround-related wait for IVB as well).

Our watermark code needs a lot of work to beat it into proper shape for atomic and that's what I'm working on at the moment.

Quoted 7 lines

Alternative issue is that wm recompute happens before we update all the legacy state. Tvrkto just stumbled over that for some of his skl watermark code, he had to switch a few places in the wm code from looking at plane->fb to look at plane->state->fb.

Similar changes might be needed for the cursor wm code - cursor_width/height is kinda redundant since universal planes support.

Quoted 6 lines

Thanks for the insights, Matt and Dan. It's clear to me from Matt's analysis that my patch is a hack at best and nowhere near a proper fix.

This patch was submitted in the hopes of fixing a substantial graphical regression. Shall my patch be retooled as a "better" work-around, or will the regression stand until the watermark code is beat into proper shape?

I have a couple patches already written that I think might solve your problem. I'll extract those from my in-progress tree and post them shortly for you to try.

Matt

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