Skip to contentSkip to conversation

drm/i915: Make fb user dirty operation to invalidate frontbuffer

PATCH In review Rodrigo Vivi 5 participants 20 messages Jun 30, 2015 – Jul 8, 2015

Subsystems (2)
lore
Share & tools
lore

Rodrigo Vivi

author maintainer
#0 lore raw

PATCH 1/3 drm/i915: Make fb user dirty operation to invalidate frontbuffer

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +18 -0
index 724b0e3..b55b1b6 100644
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
 	return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+					       struct drm_file *file,
+					       unsigned flags, unsigned color,
+					       struct drm_clip_rect *clips,
+					       unsigned num_clips)
+{
+	struct drm_device *dev = fb->dev;
+	struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+	struct drm_i915_gem_object *obj = intel_fb->obj;
+
+	mutex_lock(&dev->struct_mutex);
+	intel_fb_obj_invalidate(obj, ORIGIN_GTT);
+	mutex_unlock(&dev->struct_mutex);
+
+	return 0;
+}
+
 static const struct drm_framebuffer_funcs intel_fb_funcs = {
 	.destroy = intel_user_framebuffer_destroy,
 	.create_handle = intel_user_framebuffer_create_handle,
+	.dirty = intel_user_framebuffer_dirty,
 };
 
 static
Signature · 6 lines
2.1.0

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

PATCH 2/3 drm/i915: PSR: Remove Low Power HW tracking mask.

By Spec we should only mask memup and hotplug detection for hardware tracking cases. However we always masked LPSP because with power well always enabled on audio PSR was never being activated and residency was always zeroed.

Apparently audio driver is tying power well management and runtime PM for some reason. But with audio runtime PM working or with audio completely out of picture we should remove this mask, otherwise we have a high risk of miss screen updates as faced by Matthew.

WARNING: With this patch if snd_intel_hda driver is runing and not releasing power well properly PSR will constant Exit and Performance Counter will be 0.

But the best thing of this patch is that with one more HW tracking working the risks of missed blank screen are minimized at most.

This affects just core platforms where PSR exit are also helped by HW tracking: Haswell, Broadwell and Skylake for now.

v2: Fix commit message explanation. It has nothing to do with runtime PM on i915 as previously advertised.

Cc Daniel Vetter <[email protected]>
Cc Matthew Garrett <[email protected] via codon.org.uk>
---
 drivers/gpu/drm/i915/intel_psr.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_psr.c +1 -1
index d79ba58..cab5153 100644
@@ -400,7 +400,7 @@ void intel_psr_enable(struct intel_dp *intel_dp)
 
 		/* Avoid continuous PSR exit by masking memup and hpd */
 		I915_WRITE(EDP_PSR_DEBUG_CTL(dev), EDP_PSR_DEBUG_MASK_MEMUP |
-			   EDP_PSR_DEBUG_MASK_HPD | EDP_PSR_DEBUG_MASK_LPSP);
+			   EDP_PSR_DEBUG_MASK_HPD);
 
 		/* Enable PSR on the panel */
 		hsw_psr_enable_sink(intel_dp);
Signature · 6 lines
2.1.0

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

Re: [PATCH 2/3] drm/i915: PSR: Remove Low Power HW tracking mask.

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 20 lines

By Spec we should only mask memup and hotplug detection for hardware tracking cases. However we always masked LPSP because with power well always enabled on audio PSR was never being activated and residency was always zeroed.

Apparently audio driver is tying power well management and runtime PM for some reason. But with audio runtime PM working or with audio completely out of picture we should remove this mask, otherwise we have a high risk of miss screen updates as faced by Matthew.

WARNING: With this patch if snd_intel_hda driver is runing and not releasing power well properly PSR will constant Exit and Performance Counter will be 0.

But the best thing of this patch is that with one more HW tracking working the risks of missed blank screen are minimized at most.

Quoted 6 lines

This affects just core platforms where PSR exit are also helped by HW tracking: Haswell, Broadwell and Skylake for now.

v2: Fix commit message explanation. It has nothing to do with runtime PM on i915 as previously advertised.

But then, as far as I understood, if you remove the mask, our current IGT patches will start failing with "PSR disabled" errors, right? So can you please also provide IGT fixes?

If the only thing we really need is Audio RPM instead of full graphics RPM, then I suggest you extract igt_setup_audio_runtime_pm() from the current igt_setup_runtime_pm(), and then call it on the tests that need PSR.

Quoted 28 lines
Cc Daniel Vetter <[email protected]>
Cc Matthew Garrett <[email protected] via codon.org.uk>
---
 drivers/gpu/drm/i915/intel_psr.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_psr.c +1 -2
index d79ba58..cab5153 100644
@@ -400,7 +400,7 @@ void intel_psr_enable(struct intel_dp *intel_dp)
 
                /* Avoid continuous PSR exit by masking memup and hpd */
                I915_WRITE(EDP_PSR_DEBUG_CTL(dev), EDP_PSR_DEBUG_MASK_MEMUP |
-                          EDP_PSR_DEBUG_MASK_HPD | EDP_PSR_DEBUG_MASK_LPSP);
+                          EDP_PSR_DEBUG_MASK_HPD);
 
                /* Enable PSR on the panel */
                hsw_psr_enable_sink(intel_dp);
--

2.1.0

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

Signature · 5 lines
Paulo Zanoni
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
author maintainer
#3 lore raw

PATCH 3/3 drm/i915: Enable PSR by default.

In reply to Rodrigo Vivi

With a reliable frontbuffer tracking and all instability corner cases solved let's re-enabled PSR by default on all supported platforms.

---
 drivers/gpu/drm/i915/i915_params.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_params.c +2 -2
index 7983fe4..2eb7c65 100644
@@ -36,7 +36,7 @@ struct i915_params i915 __read_mostly = {
 	.enable_execlists = -1,
 	.enable_hangcheck = true,
 	.enable_ppgtt = -1,
-	.enable_psr = 0,
+	.enable_psr = 1,
 	.preliminary_hw_support = IS_ENABLED(CONFIG_DRM_I915_PRELIMINARY_HW_SUPPORT),
 	.disable_power_well = 1,
 	.enable_ips = 1,
@@ -118,7 +118,7 @@ MODULE_PARM_DESC(enable_execlists,
 	"(-1=auto [default], 0=disabled, 1=enabled)");
 
 module_param_named(enable_psr, i915.enable_psr, int, 0600);
-MODULE_PARM_DESC(enable_psr, "Enable PSR (default: false)");
+MODULE_PARM_DESC(enable_psr, "Enable PSR (default: true)");
 
 module_param_named(preliminary_hw_support, i915.preliminary_hw_support, int, 0600);
 MODULE_PARM_DESC(preliminary_hw_support,
Signature · 6 lines
2.1.0

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

Re: [PATCH 3/3] drm/i915: Enable PSR by default.

Task id: 6691 -------------------------------------Summary-------------------------------------
Platform Delta drm-intel-nightly Series Applied

ILK                                  302/302              302/302
SNB                                  312/316              312/316
IVB                                  343/343              343/343

BYT -2 287/287 285/287
HSW 380/380 380/380
-------------------------------------Detailed-------------------------------------
Platform Test drm-intel-nightly Series Applied
*BYT igt@gem_partial_pwrite_pread@reads PASS(1) FAIL(1)
*BYT igt@gem_tiled_partial_pwrite_pread@reads PASS(1) FAIL(1)
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

#5 lore raw
In reply to Rodrigo Vivi
Quoted 13 lines · Rodrigo Vivi

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

Quoted 6 lines

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

Will it ever grow the ability to handle clip rects? I can detect the presence of the syscall and call it appropriately, but I don't want to have to start tracking frontbuffer damage unless there's a significant advantage in doing so (to offset the cost of the tracking). -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
#6 lore raw
Quoted 18 lines · Chris Wilson
Quoted 17 lines · Rodrigo Vivi

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

Quoted 7 lines
Quoted 2 lines

Will it ever grow the ability to handle clip rects? I can detect the presence of the syscall and call it appropriately, but I don't want to have to start tracking frontbuffer damage unless there's a significant advantage in doing so (to offset the cost of the tracking).

For now this is just for generic userspace using the dumb mmap ioctls, which does already dirty everything. For gem/i915 userspace the existing frontbuffer tracking rules will still apply.

I think the damage rect will only be useful for psr2 single-shot uploads, but we need to do some serious hacks using debug rects to pull this off - the hw wants to be way to helpful and gets in the way. And I think for that usecase it only makes sense to supply the overall per-crtc damage rect in an atomic flip. Tracking damage for frontbuffer rendering probably won't be worth it anytime soon at all.

In short: Nothing to worry about yet for a long time ;-) -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
#7 lore raw
Quoted 22 lines · Daniel Vetter
Quoted 21 lines · Chris Wilson
Quoted 19 lines · Rodrigo Vivi

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

Quoted 8 lines
Quoted 4 lines

Will it ever grow the ability to handle clip rects? I can detect the presence of the syscall and call it appropriately, but I don't want to have to start tracking frontbuffer damage unless there's a significant advantage in doing so (to offset the cost of the tracking).

For now this is just for generic userspace using the dumb mmap ioctls, which does already dirty everything. For gem/i915 userspace the existing frontbuffer tracking rules will still apply.

But they are inadequate for the map/set-domain scanout once and write through the GTT for umpteen seconds, which can happen quite frequenctly.

In that situation, we behave exactly like fbdev/dumb fb. -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 28 lines · Chris Wilson
Quoted 27 lines · Daniel Vetter
Quoted 25 lines · Chris Wilson
Quoted 13 lines · Rodrigo Vivi

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

Quoted 6 lines

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

Will it ever grow the ability to handle clip rects? I can detect the presence of the syscall and call it appropriately, but I don't want to have to start tracking frontbuffer damage unless there's a significant advantage in doing so (to offset the cost of the tracking).

Quoted 8 lines
Quoted 3 lines

For now this is just for generic userspace using the dumb mmap ioctls, which does already dirty everything. For gem/i915 userspace the existing frontbuffer tracking rules will still apply.

But they are inadequate for the map/set-domain scanout once and write through the GTT for umpteen seconds, which can happen quite frequenctly.

In that situation, we behave exactly like fbdev/dumb fb.

Yeah you can use it to flush gtt of course too. And there I'd just defensively flush the entire fb until we've grown more clueful in the kernel. But for forntbuffer flushing I don't expect that to ever happen for i915. It makes more sense ofc for udl/qxl and others where uploads are really expensive.
-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
#9 lore raw
Quoted 36 lines · Daniel Vetter
Quoted 35 lines · Chris Wilson
Quoted 22 lines · Daniel Vetter
Quoted 21 lines · Chris Wilson
Quoted 19 lines · Rodrigo Vivi

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

Quoted 8 lines
Quoted 4 lines

Will it ever grow the ability to handle clip rects? I can detect the presence of the syscall and call it appropriately, but I don't want to have to start tracking frontbuffer damage unless there's a significant advantage in doing so (to offset the cost of the tracking).

For now this is just for generic userspace using the dumb mmap ioctls, which does already dirty everything. For gem/i915 userspace the existing frontbuffer tracking rules will still apply.

But they are inadequate for the map/set-domain scanout once and write through the GTT for umpteen seconds, which can happen quite frequenctly.

Quoted 7 lines
Quoted 1 line

In that situation, we behave exactly like fbdev/dumb fb.

Yeah you can use it to flush gtt of course too. And there I'd just defensively flush the entire fb until we've grown more clueful in the kernel. But for forntbuffer flushing I don't expect that to ever happen for i915. It makes more sense ofc for udl/qxl and others where uploads are really expensive.

Ha, I thought that MIPI was going off in this direction precisely because high pixel count displays only transferring the dirty regions is a big power saving. Likewise I expect at some point, there will be a chipset mode to only portions of the framebuffer out of the chip local cache across the bus.

I want to make sure we are not going to shoot ourselves in the foot and can forward-proof the design so that we can easily detect if we want to use cliprects.

An easy way would be to return the number of rectangles pushed by dirtyfb i.e. if dirtyfb(fb, .num_rects=1)== 0 I know that it just pushes the whole framebuffer everytime and need not track damage.

The ABI would be negative error return on failure, 0 or postive value on success, where the positive value is the number of rects pushed (which should be identical to the user request on success). Dumb users then don't need to care as they can either always request full fb flushes or always push rects.
-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
#10 lore raw
In reply to Rodrigo Vivi

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

Also, don't forget that switching to intel_fb_obj_flush() will allow you to kill the struct_mutex lock.

BTW, I'm also looking forward to see the IGT test suggested by Daniel :)

Quoted 19 lines
drivers/gpu/drm/i915/intel_display.c +6 -1
+       mutex_unlock(&dev->struct_mutex);
+
+       return 0;
+}
+
 static const struct drm_framebuffer_funcs intel_fb_funcs = {
        .destroy = intel_user_framebuffer_destroy,
        .create_handle = intel_user_framebuffer_create_handle,
+       .dirty = intel_user_framebuffer_dirty,
 };
 
 static
--

2.1.0

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

Signature · 5 lines
Paulo Zanoni
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
author maintainer
#11 lore raw
Quoted 40 lines · Paulo Zanoni

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

Quoted 17 lines

Also, don't forget that switching to intel_fb_obj_flush() will allow you to kill the struct_mutex lock.

BTW, I'm also looking forward to see the IGT test suggested by Daniel :)

Quoted 11 lines
drivers/gpu/drm/i915/intel_display.c +6 -0
+       mutex_unlock(&dev->struct_mutex);
+
+       return 0;
+}
+
 static const struct drm_framebuffer_funcs intel_fb_funcs = {
        .destroy = intel_user_framebuffer_destroy,
        .create_handle = intel_user_framebuffer_create_handle,
+       .dirty = intel_user_framebuffer_dirty,
 };
 
Quoted 11 lines
Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +0 -1
 static
--

2.1.0

_______________________________________________
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

#12 lore raw
Quoted 75 lines · Rodrigo
Quoted 40 lines · Paulo Zanoni

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around. -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
author maintainer
#13 lore raw
Quoted 76 lines · Daniel Vetter
Quoted 75 lines · Rodrigo
Quoted 40 lines · Paulo Zanoni

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 5 lines
Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around.

Well, the flush itself in the way it is defined is to flush frontbuffer bits and put power saving features back to work. PSR flush for instance doesn't exit on every flush. Only in the cases that we know HW tracking doesn't work.

One possibility would be change PSR to respect that all flush always is invalidate (exit) + flush fb_bits. But I'm against this since PSR on core platforms doesn't have SW trackking ext and it wasn't implemented to be fully enabled/disabled every time.

Another idea on this line is to make flushs know origins or at least a boolean to separate cases where flush must be handled as flush bits + continue working or invalidate + flush bits + start working again

Please let me know what you think.

But putting flushs on this dirty callback is currently useless because flush just tells psr to continue working without getting screen updates.

Well, this is the most important discussion for now, but also even we change flush interface or only in PSR I'm not sure this will work.

This dirty case happens in the middle of fbdev and when it gives control back to fbcon psr would be working without no one else invalidate or flushing it and we would miss screen updates anyway.
But in this case I understand this is a hack and if we put the invalidate there I should also put a FIXME XXX explaining that...

Quoted 1 line

-Daniel

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

#14 lore raw

2015-07-06 13:43 GMT-03:00 Vivi, Rodrigo <[email protected]>:

Quoted 102 lines · Daniel Vetter
Quoted 76 lines · Rodrigo
Quoted 75 lines · Paulo Zanoni
Quoted 40 lines

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 5 lines
Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around.

Well, the flush itself in the way it is defined is to flush frontbuffer bits and put power saving features back to work. PSR flush for instance doesn't exit on every flush. Only in the cases that we know HW tracking doesn't work.

One possibility would be change PSR to respect that all flush always is invalidate (exit) + flush fb_bits. But I'm against this since PSR on core platforms doesn't have SW trackking ext and it wasn't implemented to be fully enabled/disabled every time.

Another idea on this line is to make flushs know origins or at least a boolean to separate cases where flush must be handled as flush bits + continue working or invalidate + flush bits + start working again

Please let me know what you think.

But putting flushs on this dirty callback is currently useless because flush just tells psr to continue working without getting screen updates.

Quoted 8 lines

Well, this is the most important discussion for now, but also even we change flush interface or only in PSR I'm not sure this will work.

This dirty case happens in the middle of fbdev and when it gives control back to fbcon psr would be working without no one else invalidate or flushing it and we would miss screen updates anyway.
But in this case I understand this is a hack and if we put the invalidate there I should also put a FIXME XXX explaining that...

Last week I discussed some of this with Rodrigo on IRC and we have different interpretations on what the frontbuffer tracking calls mean. Daniel, can you please clarify a few things?

One of the things I've discussed with Rodrigo is that I consider invalidate+flush to be equivalent to just a flush call, while Rodrigo sees both cases as different. Daniel, can you please clarify the intentions of the frontbuffer tracking infrastructure here?

One contradiction I see with the current code is that PSR does nothing on flush() for BDW because it believes the HW tracking, but it does psr_exit() on invalidate for every single case. If it relies on the HW tracking, shouldn't it also do nothing on invalidate()? If it doesn't rely on HW tracking, shouldn't it do something on flush()? If it relies on just a few specific pieces of the HW tracking, shouldn't it check enum fb_op_origin? The conclusion here is that the PSR code is completely relying on having invalidate() calls before the flush() for cases where it needs help from software.

Also, we've been discussing a boot splash screen, and, according to what you said, it's supposed to be using the dumb interfaces, right? So I see that the dumb mmap is just a gem_mmap() to our driver. But PSR is supposed to get disabled after GTT mmaps. Why is it not disabled in this case? That is because we don't get the invalidate call after the mmap, right? If we require an invalidate on the dirty ioctl, it means that PSR is enabled even though it is GTT mmapped... That's supposed to be wrong, right? Shouldn't the mmap IOCTL force an invalidate?

Quoted 4 lines
Quoted 3 lines

-Daniel

Signature · 5 lines
Paulo Zanoni
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#15 lore raw
Quoted 112 lines · Paulo Zanoni

2015-07-06 13:43 GMT-03:00 Vivi, Rodrigo <[email protected]>:

Quoted 102 lines · Daniel Vetter
Quoted 76 lines · Rodrigo
Quoted 75 lines · Paulo Zanoni
Quoted 40 lines

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 5 lines
Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around.

Well, the flush itself in the way it is defined is to flush frontbuffer bits and put power saving features back to work. PSR flush for instance doesn't exit on every flush. Only in the cases that we know HW tracking doesn't work.

One possibility would be change PSR to respect that all flush always is invalidate (exit) + flush fb_bits. But I'm against this since PSR on core platforms doesn't have SW trackking ext and it wasn't implemented to be fully enabled/disabled every time.

Another idea on this line is to make flushs know origins or at least a boolean to separate cases where flush must be handled as flush bits + continue working or invalidate + flush bits + start working again

Please let me know what you think.

But putting flushs on this dirty callback is currently useless because flush just tells psr to continue working without getting screen updates.

Quoted 8 lines

Well, this is the most important discussion for now, but also even we change flush interface or only in PSR I'm not sure this will work.

This dirty case happens in the middle of fbdev and when it gives control back to fbcon psr would be working without no one else invalidate or flushing it and we would miss screen updates anyway.
But in this case I understand this is a hack and if we put the invalidate there I should also put a FIXME XXX explaining that...

Quoted 8 lines

Last week I discussed some of this with Rodrigo on IRC and we have different interpretations on what the frontbuffer tracking calls mean. Daniel, can you please clarify a few things?

One of the things I've discussed with Rodrigo is that I consider invalidate+flush to be equivalent to just a flush call, while Rodrigo sees both cases as different. Daniel, can you please clarify the intentions of the frontbuffer tracking infrastructure here?

Yeah essentially flush without an invalidate is just a instantaneous invalidate+flush. invalidate just means "someone started to frontbuffer render and I have no idea at all when that will stop". Flush means "someone has done some frontbuffer rendering, make sure those updates land on the screen". So yeah Paulo's idea is correct.

Problem is now that when we get a flip we also call flush, and for psr on big core that's probably not what we want since it can handle flips correctly, but not real frontbuffer flushes.

Quoted 9 lines

One contradiction I see with the current code is that PSR does nothing on flush() for BDW because it believes the HW tracking, but it does psr_exit() on invalidate for every single case. If it relies on the HW tracking, shouldn't it also do nothing on invalidate()? If it doesn't rely on HW tracking, shouldn't it do something on flush()? If it relies on just a few specific pieces of the HW tracking, shouldn't it check enum fb_op_origin? The conclusion here is that the PSR code is completely relying on having invalidate() calls before the flush() for cases where it needs help from software.

Calling invalidate before flush isn't imo the right fix since usually (except when you have working hw tracking like fbc for gtt mmaps) you need to stop psr/fbc/whatever on invalidate. Whereas on flush you just need to make sure screen contents get to the screen.

I think the best option would be to wire up the fb_op_origin for flushes too, and then filter out ORIGIN_FLIP for psr (except for sprites on hsw) on big core.

Quoted 9 lines

Also, we've been discussing a boot splash screen, and, according to what you said, it's supposed to be using the dumb interfaces, right? So I see that the dumb mmap is just a gem_mmap() to our driver. But PSR is supposed to get disabled after GTT mmaps. Why is it not disabled in this case? That is because we don't get the invalidate call after the mmap, right? If we require an invalidate on the dirty ioctl, it means that PSR is enabled even though it is GTT mmapped... That's supposed to be wrong, right? Shouldn't the mmap IOCTL force an invalidate?

Yeah dumb mmap users are supposed to call dirtyfb after each draw call.

i915 userspace could start doing the same to mark the gtt mmap invalidate as done since we don't have any point where we stop it right now with an i915 ioctl. And to be able to work together with set_domain(GTT) it really needs to be a flush in ->dirty()
-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
author maintainer
#16 lore raw
Quoted 122 lines · Daniel Vetter
Quoted 112 lines · Paulo Zanoni

2015-07-06 13:43 GMT-03:00 Vivi, Rodrigo <[email protected]>:

Quoted 102 lines · Daniel Vetter
Quoted 76 lines · Rodrigo
Quoted 75 lines · Paulo Zanoni
Quoted 40 lines

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 5 lines
Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around.

Well, the flush itself in the way it is defined is to flush frontbuffer bits and put power saving features back to work. PSR flush for instance doesn't exit on every flush. Only in the cases that we know HW tracking doesn't work.

One possibility would be change PSR to respect that all flush always is invalidate (exit) + flush fb_bits. But I'm against this since PSR on core platforms doesn't have SW trackking ext and it wasn't implemented to be fully enabled/disabled every time.

Another idea on this line is to make flushs know origins or at least a boolean to separate cases where flush must be handled as flush bits + continue working or invalidate + flush bits + start working again

Please let me know what you think.

But putting flushs on this dirty callback is currently useless because flush just tells psr to continue working without getting screen updates.

Quoted 8 lines

Well, this is the most important discussion for now, but also even we change flush interface or only in PSR I'm not sure this will work.

This dirty case happens in the middle of fbdev and when it gives control back to fbcon psr would be working without no one else invalidate or flushing it and we would miss screen updates anyway.
But in this case I understand this is a hack and if we put the invalidate there I should also put a FIXME XXX explaining that...

Quoted 8 lines

Last week I discussed some of this with Rodrigo on IRC and we have different interpretations on what the frontbuffer tracking calls mean. Daniel, can you please clarify a few things?

One of the things I've discussed with Rodrigo is that I consider invalidate+flush to be equivalent to just a flush call, while Rodrigo sees both cases as different. Daniel, can you please clarify the intentions of the frontbuffer tracking infrastructure here?

Quoted 5 lines

Yeah essentially flush without an invalidate is just a instantaneous invalidate+flush. invalidate just means "someone started to frontbuffer render and I have no idea at all when that will stop". Flush means "someone has done some frontbuffer rendering, make sure those updates land on the screen". So yeah Paulo's idea is correct.

Thanks for the clarification...

Quoted 15 lines

Problem is now that when we get a flip we also call flush, and for psr on big core that's probably not what we want since it can handle flips correctly, but not real frontbuffer flushes.

Quoted 9 lines

One contradiction I see with the current code is that PSR does nothing on flush() for BDW because it believes the HW tracking, but it does psr_exit() on invalidate for every single case. If it relies on the HW tracking, shouldn't it also do nothing on invalidate()? If it doesn't rely on HW tracking, shouldn't it do something on flush()? If it relies on just a few specific pieces of the HW tracking, shouldn't it check enum fb_op_origin? The conclusion here is that the PSR code is completely relying on having invalidate() calls before the flush() for cases where it needs help from software.

Quoted 8 lines

Calling invalidate before flush isn't imo the right fix since usually (except when you have working hw tracking like fbc for gtt mmaps) you need to stop psr/fbc/whatever on invalidate. Whereas on flush you just need to make sure screen contents get to the screen.

I think the best option would be to wire up the fb_op_origin for flushes too, and then filter out ORIGIN_FLIP for psr (except for sprites on hsw) on big core.

Ok, this approach here works here when I have splash screen taking place and X coming right after it. However we still have one issue that is when it goes back to console.

fbdev on set_par invalidated front buffer, but then splash screen came and flushed it, psr started working, then control gets back to fbdev and nothing invalidates more so from this point on we miss many (if not all) screen updates.

So we are going to a endless path where we could have many cases where fbdev will be flushed with no reliable way to get invalidated again.

So, right now I just see 3 paths:
1. invalidate instead of flush on fb dirty callback just to make sure everything gets invalidated when using dumb ioctls
2. change fbdev and all its clients to make a proper use of a dirty.
3. change fbdev, fbcon to propagate console switch (KD_TEXT info) where we would have a way to know exactly who is in control
4. Introdoce IOCTLs for PSR and enable/disable it on DDX when we know DDX has control.

I really don't like the idea to change fbdev scructures and its clients...

Please let me know if you have other/better ideas and also what path you prefer...

Thanks,
Rodrigo.

Quoted 11 lines
Quoted 10 lines

Also, we've been discussing a boot splash screen, and, according to what you said, it's supposed to be using the dumb interfaces, right? So I see that the dumb mmap is just a gem_mmap() to our driver. But PSR is supposed to get disabled after GTT mmaps. Why is it not disabled in this case? That is because we don't get the invalidate call after the mmap, right? If we require an invalidate on the dirty ioctl, it means that PSR is enabled even though it is GTT mmapped... That's supposed to be wrong, right? Shouldn't the mmap IOCTL force an invalidate?

Quoted 7 lines

Yeah dumb mmap users are supposed to call dirtyfb after each draw call.

i915 userspace could start doing the same to mark the gtt mmap invalidate as done since we don't have any point where we stop it right now with an i915 ioctl. And to be able to work together with set_domain(GTT) it really needs to be a flush in ->dirty()
-Daniel

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

#17 lore raw
Quoted 176 lines · Rodrigo
Quoted 122 lines · Daniel Vetter
Quoted 112 lines · Paulo Zanoni

2015-07-06 13:43 GMT-03:00 Vivi, Rodrigo <[email protected]>:

Quoted 102 lines · Daniel Vetter
Quoted 76 lines · Rodrigo
Quoted 75 lines · Paulo Zanoni
Quoted 40 lines

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 5 lines
Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around.

Well, the flush itself in the way it is defined is to flush frontbuffer bits and put power saving features back to work. PSR flush for instance doesn't exit on every flush. Only in the cases that we know HW tracking doesn't work.

One possibility would be change PSR to respect that all flush always is invalidate (exit) + flush fb_bits. But I'm against this since PSR on core platforms doesn't have SW trackking ext and it wasn't implemented to be fully enabled/disabled every time.

Another idea on this line is to make flushs know origins or at least a boolean to separate cases where flush must be handled as flush bits + continue working or invalidate + flush bits + start working again

Please let me know what you think.

But putting flushs on this dirty callback is currently useless because flush just tells psr to continue working without getting screen updates.

Quoted 8 lines

Well, this is the most important discussion for now, but also even we change flush interface or only in PSR I'm not sure this will work.

This dirty case happens in the middle of fbdev and when it gives control back to fbcon psr would be working without no one else invalidate or flushing it and we would miss screen updates anyway.
But in this case I understand this is a hack and if we put the invalidate there I should also put a FIXME XXX explaining that...

Quoted 8 lines

Last week I discussed some of this with Rodrigo on IRC and we have different interpretations on what the frontbuffer tracking calls mean. Daniel, can you please clarify a few things?

One of the things I've discussed with Rodrigo is that I consider invalidate+flush to be equivalent to just a flush call, while Rodrigo sees both cases as different. Daniel, can you please clarify the intentions of the frontbuffer tracking infrastructure here?

Quoted 5 lines

Yeah essentially flush without an invalidate is just a instantaneous invalidate+flush. invalidate just means "someone started to frontbuffer render and I have no idea at all when that will stop". Flush means "someone has done some frontbuffer rendering, make sure those updates land on the screen". So yeah Paulo's idea is correct.

Thanks for the clarification...

Quoted 15 lines

Problem is now that when we get a flip we also call flush, and for psr on big core that's probably not what we want since it can handle flips correctly, but not real frontbuffer flushes.

Quoted 9 lines

One contradiction I see with the current code is that PSR does nothing on flush() for BDW because it believes the HW tracking, but it does psr_exit() on invalidate for every single case. If it relies on the HW tracking, shouldn't it also do nothing on invalidate()? If it doesn't rely on HW tracking, shouldn't it do something on flush()? If it relies on just a few specific pieces of the HW tracking, shouldn't it check enum fb_op_origin? The conclusion here is that the PSR code is completely relying on having invalidate() calls before the flush() for cases where it needs help from software.

Quoted 8 lines

Calling invalidate before flush isn't imo the right fix since usually (except when you have working hw tracking like fbc for gtt mmaps) you need to stop psr/fbc/whatever on invalidate. Whereas on flush you just need to make sure screen contents get to the screen.

I think the best option would be to wire up the fb_op_origin for flushes too, and then filter out ORIGIN_FLIP for psr (except for sprites on hsw) on big core.

Ok, this approach here works here when I have splash screen taking place and X coming right after it. However we still have one issue that is when it goes back to console.

fbdev on set_par invalidated front buffer, but then splash screen came and flushed it, psr started working, then control gets back to fbdev and nothing invalidates more so from this point on we miss many (if not all) screen updates.

So we are going to a endless path where we could have many cases where fbdev will be flushed with no reliable way to get invalidated again.

So, right now I just see 3 paths:
1. invalidate instead of flush on fb dirty callback just to make sure everything gets invalidated when using dumb ioctls
2. change fbdev and all its clients to make a proper use of a dirty.
3. change fbdev, fbcon to propagate console switch (KD_TEXT info) where we would have a way to know exactly who is in control
4. Introdoce IOCTLs for PSR and enable/disable it on DDX when we know DDX has control.

Quoted 5 lines

I really don't like the idea to change fbdev scructures and its clients...

Please let me know if you have other/better ideas and also what path you prefer...

As I mentioned we're not the only ones with exactly this problem, udl & qxl have the same need for calling dirtyfb all over the place from fbdev hooks. qxl seems to have the more complete solution using a workqueue.

What I'm still wondering about though is why we don't get an fbdev invalidate when switching back to the console. boot splash should be doing it's drawing on its own buffer, fbcon would switch back (which means a set_par call) and that should invalidate. I have no idea why we don't get this call in your setup ...
-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
author maintainer
#18 lore raw
Quoted 180 lines · Daniel Vetter
Quoted 179 lines · Rodrigo
Quoted 122 lines · Daniel Vetter
Quoted 112 lines · Paulo Zanoni

2015-07-06 13:43 GMT-03:00 Vivi, Rodrigo <[email protected]>:

Quoted 102 lines · Daniel Vetter
Quoted 76 lines · Rodrigo
Quoted 75 lines · Paulo Zanoni
Quoted 40 lines

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 5 lines
Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around.

Well, the flush itself in the way it is defined is to flush frontbuffer bits and put power saving features back to work. PSR flush for instance doesn't exit on every flush. Only in the cases that we know HW tracking doesn't work.

One possibility would be change PSR to respect that all flush always is invalidate (exit) + flush fb_bits. But I'm against this since PSR on core platforms doesn't have SW trackking ext and it wasn't implemented to be fully enabled/disabled every time.

Another idea on this line is to make flushs know origins or at least a boolean to separate cases where flush must be handled as flush bits + continue working or invalidate + flush bits + start working again

Please let me know what you think.

But putting flushs on this dirty callback is currently useless because flush just tells psr to continue working without getting screen updates.

Quoted 8 lines

Well, this is the most important discussion for now, but also even we change flush interface or only in PSR I'm not sure this will work.

This dirty case happens in the middle of fbdev and when it gives control back to fbcon psr would be working without no one else invalidate or flushing it and we would miss screen updates anyway.
But in this case I understand this is a hack and if we put the invalidate there I should also put a FIXME XXX explaining that...

Quoted 8 lines

Last week I discussed some of this with Rodrigo on IRC and we have different interpretations on what the frontbuffer tracking calls mean. Daniel, can you please clarify a few things?

One of the things I've discussed with Rodrigo is that I consider invalidate+flush to be equivalent to just a flush call, while Rodrigo sees both cases as different. Daniel, can you please clarify the intentions of the frontbuffer tracking infrastructure here?

Quoted 5 lines

Yeah essentially flush without an invalidate is just a instantaneous invalidate+flush. invalidate just means "someone started to frontbuffer render and I have no idea at all when that will stop". Flush means "someone has done some frontbuffer rendering, make sure those updates land on the screen". So yeah Paulo's idea is correct.

Thanks for the clarification...

Quoted 15 lines

Problem is now that when we get a flip we also call flush, and for psr on big core that's probably not what we want since it can handle flips correctly, but not real frontbuffer flushes.

Quoted 9 lines

One contradiction I see with the current code is that PSR does nothing on flush() for BDW because it believes the HW tracking, but it does psr_exit() on invalidate for every single case. If it relies on the HW tracking, shouldn't it also do nothing on invalidate()? If it doesn't rely on HW tracking, shouldn't it do something on flush()? If it relies on just a few specific pieces of the HW tracking, shouldn't it check enum fb_op_origin? The conclusion here is that the PSR code is completely relying on having invalidate() calls before the flush() for cases where it needs help from software.

Quoted 8 lines

Calling invalidate before flush isn't imo the right fix since usually (except when you have working hw tracking like fbc for gtt mmaps) you need to stop psr/fbc/whatever on invalidate. Whereas on flush you just need to make sure screen contents get to the screen.

I think the best option would be to wire up the fb_op_origin for flushes too, and then filter out ORIGIN_FLIP for psr (except for sprites on hsw) on big core.

Ok, this approach here works here when I have splash screen taking place and X coming right after it. However we still have one issue that is when it goes back to console.

fbdev on set_par invalidated front buffer, but then splash screen came and flushed it, psr started working, then control gets back to fbdev and nothing invalidates more so from this point on we miss many (if not all) screen updates.

So we are going to a endless path where we could have many cases where fbdev will be flushed with no reliable way to get invalidated again.

So, right now I just see 3 paths:
1. invalidate instead of flush on fb dirty callback just to make sure everything gets invalidated when using dumb ioctls
2. change fbdev and all its clients to make a proper use of a dirty.
3. change fbdev, fbcon to propagate console switch (KD_TEXT info) where we would have a way to know exactly who is in control
4. Introdoce IOCTLs for PSR and enable/disable it on DDX when we know DDX has control.

I really don't like the idea to change fbdev scructures and its clients...

Quoted 6 lines
Quoted 2 lines

Please let me know if you have other/better ideas and also what path you prefer...

As I mentioned we're not the only ones with exactly this problem, udl & qxl have the same need for calling dirtyfb all over the place from fbdev hooks. qxl seems to have the more complete solution using a workqueue.

Yes, I agree. I just want to block PSR on this fbdev rework that is not simple, involve many maintainers and it is not the main PSR use case.

Quoted 6 lines

What I'm still wondering about though is why we don't get an fbdev invalidate when switching back to the console. boot splash should be doing it's drawing on its own buffer, fbcon would switch back (which means a set_par call) and that should invalidate. I have no idea why we don't get this call in your setup ...

You were right.... I wasn't relying on set_par but it is actually called on fbcon switch back, but the invalidate was happening only on the first time because write_domain was already GTT. So with the following patch we can solve the case where I press "Esc" key on the middle of splash screen:
http://cgit.freedesktop.org/~vivijim/drm-intel/commit/?h=psr&id=ce7226054085080c10e53b82743a76e0593b785a

But I was facing also another issue with crypto pass and without any splash and trace lead me to fbdev_restore_mode so another hack:
http://cgit.freedesktop.org/~vivijim/drm-intel/commit/?h=psr&id=9a7c3f7853a327a82551e335b6056b5e5362004a

All (an more) in
http://cgit.freedesktop.org/~vivijim/drm-intel/log/?h=psr
for now... while I'm not sure a flicker on fedora reported by community is not reliably solved:

https://bugzilla.redhat.com/show_bug.cgi?id=1236303

Thanks,
Rodrigo.

Quoted 1 line

-Daniel

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

#19 lore raw
Quoted 198 lines · Rodrigo
Quoted 180 lines · Daniel Vetter
Quoted 179 lines · Rodrigo
Quoted 122 lines · Daniel Vetter
Quoted 112 lines · Paulo Zanoni

2015-07-06 13:43 GMT-03:00 Vivi, Rodrigo <[email protected]>:

Quoted 102 lines · Daniel Vetter
Quoted 76 lines · Rodrigo
Quoted 75 lines · Paulo Zanoni
Quoted 40 lines

2015-06-30 20:42 GMT-03:00 Rodrigo Vivi <[email protected]>:

Quoted 28 lines

Let's do a frontbuffer invalidation on dirty fb.
To be used for DIRTYFB drm ioctl.

This patch solves the biggest PSR known issue, that is missed screen updates during boot, mainly when there is a splash screen involved like plymouth.

Plymoth will do a modeset over ioctl that flushes frontbuffer tracking and PSR gets back to work while it cannot track the screen updates and exit properly. However plymouth also uses a dirtyfb ioctl whenever updating the screen. So let's use it to invalidate PSR back again.

v2: Remove ORIGIN_FB_DIRTY and use ORIGIN_GTT instead since dirty
callback is just called after few screen updates and not on everyone as pointed by Daniel.

---
 drivers/gpu/drm/i915/intel_display.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)
drivers/gpu/drm/i915/intel_display.c +0 -0
index 724b0e3..b55b1b6 100644

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14400

Quoted 8 lines
drivers/gpu/drm/i915/intel_display.c +5 -0
@@ -14393,9 +14393,27 @@ static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb,
        return drm_gem_handle_create(file, &obj->base, handle);
 }
 
+static int intel_user_framebuffer_dirty(struct drm_framebuffer *fb,
+                                              struct drm_file *file,
+                                              unsigned flags, unsigned color,
+                                              struct drm_clip_rect *clips,
+                                              unsigned num_clips)

You don't need the white space on the lines above, just the tabs.

Quoted 15 lines

drivers/gpu/drm/i915/intel_display.c static int intel_user_framebuffer_create_handle(struct drm_framebuffer *fb, · line 14407

Quoted 7 lines
drivers/gpu/drm/i915/intel_display.c +7 -0
+{
+       struct drm_device *dev = fb->dev;
+       struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
+       struct drm_i915_gem_object *obj = intel_fb->obj;
+
+       mutex_lock(&dev->struct_mutex);
+       intel_fb_obj_invalidate(obj, ORIGIN_GTT);

As far as I understood from my brief investigation, the dirty IOCTL just says "hey, I'm done writing on the memory, you can flush things now", and if the user writes on the buffer again later, it will need to call the dirty IOCTL again. So shouldn't we call intel_fb_obj_flush(obj, false) here? It seems this was also pointed by Daniel on the first review. It would be better because it would allow us to actually keep PSR/FBC enabled.

The flush caused by the dumb modeset ioctl is exactly what I want to revert here.

Well, I just tested to double check here and flush makes me to miss screen updates. (triple checked with retired = true as well just in case)

fbdev comes to place and invalidated frontbuffer, then plymouth does a flush that makes psr start working when it should continue disabled. Continue flushing it doesn't solve the problem, just ratify it.

But beside the issue that it is solving I don't believe we want is a flush anyway. There is something writing directly to frontbuffer with no invalidation. The dirty call is supposed to be a damage call that actually tells something on the screen got written and needs to be updated if it hasn't still. In our cause this is exactly the frontbuffer invalidate, not the flush.

Quoted 5 lines
Quoted 2 lines

The flush place would be after it really stopped doing something, but since I don't trust it I prefer to let it invalidated until next flip.

See my review on the first round of this patch: dirtyfb _is_ a flush. If it's not enough then there's some other problems still around.

Well, the flush itself in the way it is defined is to flush frontbuffer bits and put power saving features back to work. PSR flush for instance doesn't exit on every flush. Only in the cases that we know HW tracking doesn't work.

One possibility would be change PSR to respect that all flush always is invalidate (exit) + flush fb_bits. But I'm against this since PSR on core platforms doesn't have SW trackking ext and it wasn't implemented to be fully enabled/disabled every time.

Another idea on this line is to make flushs know origins or at least a boolean to separate cases where flush must be handled as flush bits + continue working or invalidate + flush bits + start working again

Please let me know what you think.

But putting flushs on this dirty callback is currently useless because flush just tells psr to continue working without getting screen updates.

Quoted 8 lines

Well, this is the most important discussion for now, but also even we change flush interface or only in PSR I'm not sure this will work.

This dirty case happens in the middle of fbdev and when it gives control back to fbcon psr would be working without no one else invalidate or flushing it and we would miss screen updates anyway.
But in this case I understand this is a hack and if we put the invalidate there I should also put a FIXME XXX explaining that...

Quoted 8 lines

Last week I discussed some of this with Rodrigo on IRC and we have different interpretations on what the frontbuffer tracking calls mean. Daniel, can you please clarify a few things?

One of the things I've discussed with Rodrigo is that I consider invalidate+flush to be equivalent to just a flush call, while Rodrigo sees both cases as different. Daniel, can you please clarify the intentions of the frontbuffer tracking infrastructure here?

Quoted 5 lines

Yeah essentially flush without an invalidate is just a instantaneous invalidate+flush. invalidate just means "someone started to frontbuffer render and I have no idea at all when that will stop". Flush means "someone has done some frontbuffer rendering, make sure those updates land on the screen". So yeah Paulo's idea is correct.

Thanks for the clarification...

Quoted 15 lines

Problem is now that when we get a flip we also call flush, and for psr on big core that's probably not what we want since it can handle flips correctly, but not real frontbuffer flushes.

Quoted 9 lines

One contradiction I see with the current code is that PSR does nothing on flush() for BDW because it believes the HW tracking, but it does psr_exit() on invalidate for every single case. If it relies on the HW tracking, shouldn't it also do nothing on invalidate()? If it doesn't rely on HW tracking, shouldn't it do something on flush()? If it relies on just a few specific pieces of the HW tracking, shouldn't it check enum fb_op_origin? The conclusion here is that the PSR code is completely relying on having invalidate() calls before the flush() for cases where it needs help from software.

Quoted 8 lines

Calling invalidate before flush isn't imo the right fix since usually (except when you have working hw tracking like fbc for gtt mmaps) you need to stop psr/fbc/whatever on invalidate. Whereas on flush you just need to make sure screen contents get to the screen.

I think the best option would be to wire up the fb_op_origin for flushes too, and then filter out ORIGIN_FLIP for psr (except for sprites on hsw) on big core.

Ok, this approach here works here when I have splash screen taking place and X coming right after it. However we still have one issue that is when it goes back to console.

fbdev on set_par invalidated front buffer, but then splash screen came and flushed it, psr started working, then control gets back to fbdev and nothing invalidates more so from this point on we miss many (if not all) screen updates.

So we are going to a endless path where we could have many cases where fbdev will be flushed with no reliable way to get invalidated again.

So, right now I just see 3 paths:
1. invalidate instead of flush on fb dirty callback just to make sure everything gets invalidated when using dumb ioctls
2. change fbdev and all its clients to make a proper use of a dirty.
3. change fbdev, fbcon to propagate console switch (KD_TEXT info) where we would have a way to know exactly who is in control
4. Introdoce IOCTLs for PSR and enable/disable it on DDX when we know DDX has control.

I really don't like the idea to change fbdev scructures and its clients...

Quoted 6 lines
Quoted 2 lines

Please let me know if you have other/better ideas and also what path you prefer...

As I mentioned we're not the only ones with exactly this problem, udl & qxl have the same need for calling dirtyfb all over the place from fbdev hooks. qxl seems to have the more complete solution using a workqueue.

Yes, I agree. I just want to block PSR on this fbdev rework that is not simple, involve many maintainers and it is not the main PSR use case.

Quoted 6 lines

What I'm still wondering about though is why we don't get an fbdev invalidate when switching back to the console. boot splash should be doing it's drawing on its own buffer, fbcon would switch back (which means a set_par call) and that should invalidate. I have no idea why we don't get this call in your setup ...

Quoted 6 lines

You were right.... I wasn't relying on set_par but it is actually called on fbcon switch back, but the invalidate was happening only on the first time because write_domain was already GTT. So with the following patch we can solve the case where I press "Esc" key on the middle of splash screen:
http://cgit.freedesktop.org/~vivijim/drm-intel/commit/?h=psr&id=ce7226054085080c10e53b82743a76e0593b785a

Oops I've accidentally broken this even worse with

commit 031b698a77a70a6c394568034437b5486a44e868

Author: Daniel Vetter <[email protected]>
Date:   Fri Jun 26 19:35:16 2015 +0200

drm/i915: Unconditionally do fb tracking invalidate in set_domain

This commit fixes this exact bug (already in gtt write domain), but unfortunately only for the set_domain ioctl, not for fbcon.

We really just need an fb invalidate in there, no need for the if.

Quoted 3 lines

But I was facing also another issue with crypto pass and without any splash and trace lead me to fbdev_restore_mode so another hack:
http://cgit.freedesktop.org/~vivijim/drm-intel/commit/?h=psr&id=9a7c3f7853a327a82551e335b6056b5e5362004a

Excellent catch, we definitely need that one too.
-Daniel

Quoted 16 lines

All (an more) in
http://cgit.freedesktop.org/~vivijim/drm-intel/log/?h=psr
for now... while I'm not sure a flicker on fedora reported by community is not reliably solved:

https://bugzilla.redhat.com/show_bug.cgi?id=1236303

Thanks,
Rodrigo.

Quoted 1 line

-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