Skip to contentSkip to conversation

drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

PATCH In review Emily Deng 4 participants 15 messages Aug 11, 2016 – Aug 16, 2016 11 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

Emily Deng

author
#0 lore raw

PATCH 1/1 drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

The adev->ddev->vblank[crtc].count couldn't be used here, so define another variable to compute the vblank count.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h |  1 +
 drivers/gpu/drm/amd/amdgpu/dce_virtual.c | 10 ++++------
 2 files changed, 5 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +1 -0
index 50aeccf..054653e 100644
@@ -370,6 +370,7 @@ struct amdgpu_mode_info {
 	int			disp_priority;
 	const struct amdgpu_display_funcs *funcs;
 	struct hrtimer vblank_timer;
+	u32			timer_vblank_count;
 	enum amdgpu_interrupt_state vsync_timer_enabled;
 };
 
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +4 -6
index 2ce5f90..d616ab9 100644
@@ -58,7 +58,7 @@ static u32 dce_virtual_vblank_get_counter(struct amdgpu_device *adev, int crtc)
 	if (crtc >= adev->mode_info.num_crtc)
 		return 0;
 	else
-		return adev->ddev->vblank[crtc].count;
+		return adev->mode_info.timer_vblank_count;
 }
 
 static void dce_virtual_page_flip(struct amdgpu_device *adev,
@@ -353,7 +353,6 @@ static int dce_virtual_crtc_init(struct amdgpu_device *adev, int index)
 static int dce_virtual_early_init(void *handle)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-
 	adev->mode_info.vsync_timer_enabled = AMDGPU_IRQ_STATE_DISABLE;
 	dce_virtual_set_display_funcs(adev);
 	dce_virtual_set_irq_funcs(adev);
@@ -361,6 +360,7 @@ static int dce_virtual_early_init(void *handle)
 	adev->mode_info.num_crtc = 1;
 	adev->mode_info.num_hpd = 1;
 	adev->mode_info.num_dig = 1;
+	adev->mode_info.timer_vblank_count = 0;
 	return 0;
 }
 
@@ -653,7 +653,7 @@ static enum hrtimer_restart dce_virtual_vblank_timer_handle(struct hrtimer *vbla
 	struct amdgpu_mode_info *mode_info = container_of(vblank_timer, struct amdgpu_mode_info ,vblank_timer);
 	struct amdgpu_device *adev = container_of(mode_info, struct amdgpu_device ,mode_info);
 	unsigned crtc = 0;
-	adev->ddev->vblank[0].count++;
+	adev->mode_info.timer_vblank_count++;
 	drm_handle_vblank(adev->ddev, crtc);
 	dce_virtual_pageflip_irq(adev, NULL, NULL);
 	hrtimer_start(vblank_timer, ktime_set(0, DCE_VIRTUAL_VBLANK_PERIOD), HRTIMER_MODE_REL);
@@ -680,8 +680,6 @@ static void dce_virtual_set_crtc_vblank_interrupt_state(struct amdgpu_device *ad
 		hrtimer_cancel(&adev->mode_info.vblank_timer);
 	}
 
-	if (!state || (state && !adev->mode_info.vsync_timer_enabled))
-		adev->ddev->vblank[0].count = 0;
 	adev->mode_info.vsync_timer_enabled = state;
 	DRM_DEBUG("[FM]set crtc %d vblank interrupt state %d\n", crtc, state);
 }
@@ -718,7 +716,7 @@ static int dce_virtual_crtc_irq(struct amdgpu_device *adev,
 	unsigned crtc = 0;
 	unsigned irq_type = AMDGPU_CRTC_IRQ_VBLANK1;
 
-	adev->ddev->vblank[crtc].count++;
+	adev->mode_info.timer_vblank_count++;
 	dce_virtual_crtc_vblank_int_ack(adev, crtc);
 
 	if (amdgpu_irq_enabled(adev, source, irq_type)) {
Signature · 6 lines
1.9.1

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

A message that is not in the archive (sent off-list or not yet received): 1470887198-3465-1-git-send-email-Emily.Deng-5C7GfCeVMHo@public.gmane.org

#2 lore raw

ping...
Anybody could help to review it?

Quoted 96 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Emily Deng
Sent: Thursday, August 11, 2016 11:47 AM
To: [email protected]
Cc: Deng, Emily <[email protected]>
Subject: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

The adev->ddev->vblank[crtc].count couldn't be used here, so define another variable to compute the vblank count.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h |  1 +
drivers/gpu/drm/amd/amdgpu/dce_virtual.c | 10 ++++------
 2 files changed, 5 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +1 -0
index 50aeccf..054653e 100644
@@ -370,6 +370,7 @@ struct amdgpu_mode_info {
 	int			disp_priority;
 	const struct amdgpu_display_funcs *funcs;
 	struct hrtimer vblank_timer;
+	u32			timer_vblank_count;
 	enum amdgpu_interrupt_state vsync_timer_enabled;  };
 
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
index 2ce5f90..d616ab9 100644
@@ -58,7 +58,7 @@ static u32 dce_virtual_vblank_get_counter(struct
 amdgpu_device *adev, int crtc)
 	if (crtc >= adev->mode_info.num_crtc)
 		return 0;
 	else
-		return adev->ddev->vblank[crtc].count;
+		return adev->mode_info.timer_vblank_count;
 }
 

static void dce_virtual_page_flip(struct amdgpu_device *adev, @@ -353,7
+353,6 @@ static int dce_virtual_crtc_init(struct amdgpu_device *adev, int index) static int dce_virtual_early_init(void *handle) {

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -1
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-
 	adev->mode_info.vsync_timer_enabled =

AMDGPU_IRQ_STATE_DISABLE;

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -0
 	dce_virtual_set_display_funcs(adev);
 	dce_virtual_set_irq_funcs(adev);
@@ -361,6 +360,7 @@ static int dce_virtual_early_init(void *handle)
 	adev->mode_info.num_crtc = 1;
 	adev->mode_info.num_hpd = 1;
 	adev->mode_info.num_dig = 1;
+	adev->mode_info.timer_vblank_count = 0;
 	return 0;
 }
 
@@ -653,7 +653,7 @@ static enum hrtimer_restart

dce_virtual_vblank_timer_handle(struct hrtimer *vbla
struct amdgpu_mode_info *mode_info = container_of(vblank_timer,
struct amdgpu_mode_info ,vblank_timer);
struct amdgpu_device *adev = container_of(mode_info, struct
amdgpu_device ,mode_info);

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
 	unsigned crtc = 0;
-	adev->ddev->vblank[0].count++;
+	adev->mode_info.timer_vblank_count++;
 	drm_handle_vblank(adev->ddev, crtc);
 	dce_virtual_pageflip_irq(adev, NULL, NULL);
 	hrtimer_start(vblank_timer, ktime_set(0,

DCE_VIRTUAL_VBLANK_PERIOD), HRTIMER_MODE_REL); @@ -680,8 +680,6 @@ static void dce_virtual_set_crtc_vblank_interrupt_state(struct amdgpu_device *ad

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -2
 		hrtimer_cancel(&adev->mode_info.vblank_timer);
 	}
 
-	if (!state || (state && !adev->mode_info.vsync_timer_enabled))
-		adev->ddev->vblank[0].count = 0;
 	adev->mode_info.vsync_timer_enabled = state;
 	DRM_DEBUG("[FM]set crtc %d vblank interrupt state %d\n", crtc,

state); } @@ -718,7 +716,7 @@ static int dce_virtual_crtc_irq(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -2
 	unsigned crtc = 0;
 	unsigned irq_type = AMDGPU_CRTC_IRQ_VBLANK1;
 
-	adev->ddev->vblank[crtc].count++;
+	adev->mode_info.timer_vblank_count++;
 	dce_virtual_crtc_vblank_int_ack(adev, crtc);
 
 	if (amdgpu_irq_enabled(adev, source, irq_type)) {
--

1.9.1

Quoted 4 lines

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

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

A message that is not in the archive (sent off-list or not yet received): DM5PR12MB1659D06D606338121D353F128F1F0-2J9CzHegvk/[email protected]

#4 lore raw

Ping ... again.

Quoted 110 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Deng, Emily
Sent: Friday, August 12, 2016 10:00 AM
To: [email protected]
Subject: RE: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

ping...
Anybody could help to review it?

Quoted 49 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Emily Deng
Sent: Thursday, August 11, 2016 11:47 AM
To: [email protected]
Cc: Deng, Emily <[email protected]>
Subject: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

The adev->ddev->vblank[crtc].count couldn't be used here, so define another variable to compute the vblank count.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h |  1 +
drivers/gpu/drm/amd/amdgpu/dce_virtual.c | 10 ++++------
 2 files changed, 5 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mode.h +1 -0
index 50aeccf..054653e 100644
@@ -370,6 +370,7 @@ struct amdgpu_mode_info {
 	int			disp_priority;
 	const struct amdgpu_display_funcs *funcs;
 	struct hrtimer vblank_timer;
+	u32			timer_vblank_count;
 	enum amdgpu_interrupt_state vsync_timer_enabled;  };
 
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
index 2ce5f90..d616ab9 100644
@@ -58,7 +58,7 @@ static u32 dce_virtual_vblank_get_counter(struct
 amdgpu_device *adev, int crtc)
 	if (crtc >= adev->mode_info.num_crtc)
 		return 0;
 	else
-		return adev->ddev->vblank[crtc].count;
+		return adev->mode_info.timer_vblank_count;
 }
 

static void dce_virtual_page_flip(struct amdgpu_device *adev, @@
-353,7
+353,6 @@ static int dce_virtual_crtc_init(struct amdgpu_device *adev, +int
index) static int dce_virtual_early_init(void *handle) {

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static int dce_virtual_crtc_init(struct amdgpu_device *adev, int index)

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -1
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-
 	adev->mode_info.vsync_timer_enabled =

AMDGPU_IRQ_STATE_DISABLE;

Quoted 16 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -0
 	dce_virtual_set_display_funcs(adev);
 	dce_virtual_set_irq_funcs(adev);
@@ -361,6 +360,7 @@ static int dce_virtual_early_init(void *handle)
 	adev->mode_info.num_crtc = 1;
 	adev->mode_info.num_hpd = 1;
 	adev->mode_info.num_dig = 1;
+	adev->mode_info.timer_vblank_count = 0;
 	return 0;
 }
 
@@ -653,7 +653,7 @@ static enum hrtimer_restart

dce_virtual_vblank_timer_handle(struct hrtimer *vbla
struct amdgpu_mode_info *mode_info = container_of(vblank_timer,
struct amdgpu_mode_info ,vblank_timer);
struct amdgpu_device *adev = container_of(mode_info, struct
amdgpu_device ,mode_info);

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static enum hrtimer_restart dce_virtual_vblank_timer_handle(struct hrtimer *vbla

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
 	unsigned crtc = 0;
-	adev->ddev->vblank[0].count++;
+	adev->mode_info.timer_vblank_count++;
 	drm_handle_vblank(adev->ddev, crtc);
 	dce_virtual_pageflip_irq(adev, NULL, NULL);
 	hrtimer_start(vblank_timer, ktime_set(0,

DCE_VIRTUAL_VBLANK_PERIOD),

Quoted 23 lines

HRTIMER_MODE_REL); @@ -680,8 +680,6 @@ static void dce_virtual_set_crtc_vblank_interrupt_state(struct amdgpu_device *ad

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -2
 		hrtimer_cancel(&adev->mode_info.vblank_timer);
 	}
 
-	if (!state || (state && !adev->mode_info.vsync_timer_enabled))
-		adev->ddev->vblank[0].count = 0;
 	adev->mode_info.vsync_timer_enabled = state;
 	DRM_DEBUG("[FM]set crtc %d vblank interrupt state %d\n", crtc,

state); } @@ -718,7 +716,7 @@ static int dce_virtual_crtc_irq(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -2
 	unsigned crtc = 0;
 	unsigned irq_type = AMDGPU_CRTC_IRQ_VBLANK1;
 
-	adev->ddev->vblank[crtc].count++;
+	adev->mode_info.timer_vblank_count++;
 	dce_virtual_crtc_vblank_int_ack(adev, crtc);
 
 	if (amdgpu_irq_enabled(adev, source, irq_type)) {
--

1.9.1

Quoted 8 lines
Quoted 4 lines

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

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

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

#5 lore raw
In reply to a message not in the archive

[ Adding the dri-devel list ]

Quoted 4 lines · Emily Deng

The adev->ddev->vblank[crtc].count couldn't be used here, so define another variable to compute the vblank count.

[...]

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static u32 dce_virtual_vblank_get_counter(struct amdgpu_device *adev, int crtc) · line 62

Quoted 11 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
index 2ce5f90..d616ab9 100644
@@ -58,7 +58,7 @@ static u32 dce_virtual_vblank_get_counter(struct amdgpu_device *adev, int crtc)
 	if (crtc >= adev->mode_info.num_crtc)
 		return 0;
 	else
-		return adev->ddev->vblank[crtc].count;
+		return adev->mode_info.timer_vblank_count;
 }

AFAIK the vblank_get_counter hook is supposed to always return 0 when there's no hardware frame counter which can be used. That's what drm_vblank_no_hw_counter was created for.

You mentioned internally that you ran into trouble when trying this though. Please provide more information about that, e.g.: Which base kernel version did you test it with? What values did the DRM_IOCTL_WAIT_VBLANK ioctl return to userspace? ...

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static int dce_virtual_crtc_init(struct amdgpu_device *adev, int index) · line 358

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -1
@@ -353,7 +353,6 @@ static int dce_virtual_crtc_init(struct amdgpu_device *adev, int index)
 static int dce_virtual_early_init(void *handle)
 {
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-
 	adev->mode_info.vsync_timer_enabled = AMDGPU_IRQ_STATE_DISABLE;
 	dce_virtual_set_display_funcs(adev);
 	dce_virtual_set_irq_funcs(adev);

BTW, this hunk would break the kernel coding style.

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
@@ -653,7 +653,7 @@ static enum hrtimer_restart dce_virtual_vblank_timer_handle(struct hrtimer *vbla
 	struct amdgpu_mode_info *mode_info = container_of(vblank_timer, struct amdgpu_mode_info ,vblank_timer);
 	struct amdgpu_device *adev = container_of(mode_info, struct amdgpu_device ,mode_info);
 	unsigned crtc = 0;
-	adev->ddev->vblank[0].count++;
+	adev->mode_info.timer_vblank_count++;
 	drm_handle_vblank(adev->ddev, crtc);
 	dce_virtual_pageflip_irq(adev, NULL, NULL);
 	hrtimer_start(vblank_timer, ktime_set(0, DCE_VIRTUAL_VBLANK_PERIOD), HRTIMER_MODE_REL);

[...]

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static int dce_virtual_crtc_irq(struct amdgpu_device *adev, · line 722

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
@@ -718,7 +716,7 @@ static int dce_virtual_crtc_irq(struct amdgpu_device *adev,
 	unsigned crtc = 0;
 	unsigned irq_type = AMDGPU_CRTC_IRQ_VBLANK1;
 
-	adev->ddev->vblank[crtc].count++;
+	adev->mode_info.timer_vblank_count++;
 	dce_virtual_crtc_vblank_int_ack(adev, crtc);
 
 	if (amdgpu_irq_enabled(adev, source, irq_type)) {

Wouldn't this result in adev->mode_info.timer_vblank_count getting incremented twice every time the virtual interrupt timer fires?

Anyway, this approach means that the virtual interrupt timer can never be disabled, even while userspace isn't using the vblank functionality, which is undesirable.

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 6b517057-1e45-d040-a51b-a1b21cbfee99-otUistvHUpPR7s880joybQ@public.gmane.org

#7 lore raw
Quoted 28 lines

-----Original Message-----
From: Michel Dänzer [mailto:[email protected]]
Sent: Monday, August 15, 2016 9:46 AM
To: Deng, Emily <[email protected]>
Cc: [email protected]; [email protected]
Subject: Re: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

[ Adding the dri-devel list ]

Quoted 4 lines · Emily Deng

The adev->ddev->vblank[crtc].count couldn't be used here, so define another variable to compute the vblank count.

[...]

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static u32 dce_virtual_vblank_get_counter(struct

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
index 2ce5f90..d616ab9 100644
@@ -58,7 +58,7 @@ static u32 dce_virtual_vblank_get_counter(struct

amdgpu_device *adev, int crtc)

Quoted 11 lines

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static u32 dce_virtual_vblank_get_counter(struct amdgpu_device *adev, int crtc) · line 62

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
 	if (crtc >= adev->mode_info.num_crtc)
 		return 0;
 	else
-		return adev->ddev->vblank[crtc].count;
+		return adev->mode_info.timer_vblank_count;
 }

AFAIK the vblank_get_counter hook is supposed to always return 0 when there's no hardware frame counter which can be used. That's what drm_vblank_no_hw_counter was created for.

[[EmilyD]] Sorry, I don't know much about drm_vblank_no_hw_counter, can it support vsync when return to 0?

Quoted 6 lines

You mentioned internally that you ran into trouble when trying this though. Please provide more information about that, e.g.: Which base kernel version did you test it with? What values did the DRM_IOCTL_WAIT_VBLANK ioctl return to userspace? ...

[[EmilyD]] I run the driver on kernel version 4.6, and run glxgears with vsync enabled. It is hard to detail the issue, I will try my best to description the issue I found.
After I double checked, it is not segment fault in libGL.so when run glxgears with vsync, but the glxgears will be stucked in waiting for the before swap buffers to complete. This is because when enable vsync, every time swap buffers, the DDX driver will call DRM_IOCTL_WAIT_VBLANK to queue the vblank event, and the vbl.request.sequence will be set to current_vblank_count + swap_interval. Then in kernel driver, when timer interrupt occurs, it will call drm_handle_vblank_events, it will call drm_vblank_count_and_time to get current seq, and only seq >= vbl.request.sequence, then will call send_vblank_event. So it will never call send_vblank_event.
For example, the DDX driver call DRM_IOCTL_WAIT_VBLANK , then kernel driver will call drm_queue_vblank_event, and current vblank_count is 1 (As we only return 0 in vblank_get_counter, so the vblank_count will never change except calling drm_reset_vblank_timestamp which will make adev->ddev->vblank[crtc].count++), and swap_interval is 1, then vbl.request.sequence will be 2, but the drm_vblank_count_and_time will always
return 1 except calling drm_reset_vblank_timestamp(The function drm_reset_vblank_timestamp will only be called in drm_vblank_post_modeset and drm_vblank_on ), so the send_vblank_event will never be called, and swap buffers won't complete, so the glxgears will be stucked.

Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static int dce_virtual_crtc_init(struct amdgpu_device *adev, int index)

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
@@ -353,7 +353,6 @@ static int dce_virtual_crtc_init(struct

amdgpu_device *adev, int index) static int
dce_virtual_early_init(void *handle) {

drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -1
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
-
 	adev->mode_info.vsync_timer_enabled =

AMDGPU_IRQ_STATE_DISABLE;

Quoted 5 lines

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static int dce_virtual_crtc_init(struct amdgpu_device *adev, int index) · line 358

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
 	dce_virtual_set_display_funcs(adev);
 	dce_virtual_set_irq_funcs(adev);

BTW, this hunk would break the kernel coding style.

[[EmilyD]] Thanks, I will modify this.

Quoted 27 lines

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static enum hrtimer_restart

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
@@ -653,7 +653,7 @@ static enum hrtimer_restart

dce_virtual_vblank_timer_handle(struct hrtimer *vbla

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static enum hrtimer_restart · line 653

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
 	struct amdgpu_mode_info *mode_info = container_of(vblank_timer,

struct amdgpu_mode_info ,vblank_timer);

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static enum hrtimer_restart · line 654

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
 	struct amdgpu_device *adev = container_of(mode_info, struct

amdgpu_device ,mode_info);

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static enum hrtimer_restart dce_virtual_vblank_timer_handle(struct hrtimer *vbla

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
 	unsigned crtc = 0;
-	adev->ddev->vblank[0].count++;
+	adev->mode_info.timer_vblank_count++;
 	drm_handle_vblank(adev->ddev, crtc);
 	dce_virtual_pageflip_irq(adev, NULL, NULL);
 	hrtimer_start(vblank_timer, ktime_set(0,

DCE_VIRTUAL_VBLANK_PERIOD),

Quoted 1 line

HRTIMER_MODE_REL);

[...]

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static int dce_virtual_crtc_irq(struct amdgpu_device

Quoted 1 line
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
@@ -718,7 +716,7 @@ static int dce_virtual_crtc_irq(struct amdgpu_device

*adev,

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
 	unsigned crtc = 0;
 	unsigned irq_type = AMDGPU_CRTC_IRQ_VBLANK1;
 
-	adev->ddev->vblank[crtc].count++;
+	adev->mode_info.timer_vblank_count++;
 	dce_virtual_crtc_vblank_int_ack(adev, crtc);
 
Quoted 6 lines

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static int dce_virtual_crtc_irq(struct amdgpu_device *adev, · line 722

Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
 	if (amdgpu_irq_enabled(adev, source, irq_type)) {

Wouldn't this result in adev->mode_info.timer_vblank_count getting incremented twice every time the virtual interrupt timer fires?

[[EmilyD]] Sorry, I don't think so, the dce_virtual_crtc_irq will only be called when hardware vsync timer interrupt occurs.

Quoted 4 lines

Anyway, this approach means that the virtual interrupt timer can never be disabled, even while userspace isn't using the vblank functionality, which is undesirable.

[[EmilyD]] No, it could be disabled by calling dce_virtual_set_crtc_vblank_interrupt_state.

Quoted 4 lines

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer

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

A message that is not in the archive (sent off-list or not yet received): DM5PR12MB165919E6217B9C49998009F08F120-2J9CzHegvk/[email protected]

#9 lore raw
Quoted 28 lines · Emily
Quoted 17 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Monday, August 15, 2016 9:46 AM

Quoted 4 lines · Emily Deng

The adev->ddev->vblank[crtc].count couldn't be used here, so define another variable to compute the vblank count.

[...]

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static u32 dce_virtual_vblank_get_counter(struct

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
index 2ce5f90..d616ab9 100644
@@ -58,7 +58,7 @@ static u32 dce_virtual_vblank_get_counter(struct

amdgpu_device *adev, int crtc)

Quoted 11 lines

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static u32 dce_virtual_vblank_get_counter(struct amdgpu_device *adev, int crtc) · line 62

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
 	if (crtc >= adev->mode_info.num_crtc)
 		return 0;
 	else
-		return adev->ddev->vblank[crtc].count;
+		return adev->mode_info.timer_vblank_count;
 }

AFAIK the vblank_get_counter hook is supposed to always return 0 when there's no hardware frame counter which can be used. That's what drm_vblank_no_hw_counter was created for.

Quoted 2 lines

[[EmilyD]] Sorry, I don't know much about drm_vblank_no_hw_counter, can it support vsync when return to 0?

That's its purpose. BTW, I realized in the meantime that we can't use drm_vblank_no_hw_counter directly, because there's a single struct drm_driver used by all amdgpu driver instances.

Quoted 5 lines
Quoted 5 lines

You mentioned internally that you ran into trouble when trying this though. Please provide more information about that, e.g.: Which base kernel version did you test it with? What values did the DRM_IOCTL_WAIT_VBLANK ioctl return to userspace? ...

Quoted 12 lines

[[EmilyD]] I run the driver on kernel version 4.6, and run glxgears with vsync enabled. It is hard to detail the issue, I will try my best to description the issue I found.
After I double checked, it is not segment fault in libGL.so when run glxgears with vsync, but the glxgears will be stucked in waiting for the before swap buffers to complete. This is because when enable vsync, every time swap buffers, the DDX driver will call DRM_IOCTL_WAIT_VBLANK to queue the vblank event, and the vbl.request.sequence will be set to current_vblank_count + swap_interval. Then in kernel driver, when timer interrupt occurs, it will call drm_handle_vblank_events, it will call drm_vblank_count_and_time to get current seq, and only seq >= vbl.request.sequence, then will call send_vblank_event. So it will never call send_vblank_event.
For example, the DDX driver call DRM_IOCTL_WAIT_VBLANK , then kernel driver will call drm_queue_vblank_event, and current vblank_count is 1 (As we only return 0 in vblank_get_counter, so the vblank_count will never change except calling drm_reset_vblank_timestamp which will make adev->ddev->vblank[crtc].count++), and swap_interval is 1, then vbl.request.sequence will be 2, but the drm_vblank_count_and_time will always
return 1 except calling drm_reset_vblank_timestamp(The function drm_reset_vblank_timestamp will only be called in drm_vblank_post_modeset and drm_vblank_on ), so the send_vblank_event will never be called, and swap buffers won't complete, so the glxgears will be stucked.

Looking at the drm_update_vblank_count() code, you also need to do the following in the virtual DCE case:

* Set dev->max_vblank_count = 0
* Make amdgpu_get_vblank_timestamp_kms either return values based on when the virtual vblank interrupt timer last fired, or just return a negative error code immediately, instead of calling drm_calc_vbltimestamp_from_scanoutpos
* Make amdgpu_get_vblank_counter_kms take the else path (or just return 0 immediately), not run any of the scanout position related code

Signature · 7 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer

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

A message that is not in the archive (sent off-list or not yet received): 6c149642-3e82-834a-9e8f-8a1affcc1ee7-otUistvHUpPR7s880joybQ@public.gmane.org

#11 lore raw

Hi Michel,
Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And
the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Best Wishes,
Emily Deng

Quoted 88 lines

-----Original Message-----
From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 11:12 AM
To: Deng, Emily <[email protected]>
Cc: [email protected]; [email protected]
Subject: Re: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

Quoted 28 lines · Emily
Quoted 17 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Monday, August 15, 2016 9:46 AM On 11/08/16 12:46 PM, Emily

Quoted 4 lines · Deng

The adev->ddev->vblank[crtc].count couldn't be used here, so define another variable to compute the vblank count.

[...]

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static u32 dce_virtual_vblank_get_counter(struct

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +0 -0
index 2ce5f90..d616ab9 100644
@@ -58,7 +58,7 @@ static u32 dce_virtual_vblank_get_counter(struct

amdgpu_device *adev, int crtc)

Quoted 11 lines

drivers/gpu/drm/amd/amdgpu/dce_virtual.c static u32 dce_virtual_vblank_get_counter(struct · line 62

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/dce_virtual.c +1 -1
 	if (crtc >= adev->mode_info.num_crtc)
 		return 0;
 	else
-		return adev->ddev->vblank[crtc].count;
+		return adev->mode_info.timer_vblank_count;
 }

AFAIK the vblank_get_counter hook is supposed to always return 0 when there's no hardware frame counter which can be used. That's what drm_vblank_no_hw_counter was created for.

Quoted 2 lines

[[EmilyD]] Sorry, I don't know much about drm_vblank_no_hw_counter, can it support vsync when return to 0?

That's its purpose. BTW, I realized in the meantime that we can't use drm_vblank_no_hw_counter directly, because there's a single struct drm_driver used by all amdgpu driver instances.

Quoted 7 lines
Quoted 5 lines

You mentioned internally that you ran into trouble when trying this though. Please provide more information about that, e.g.: Which base kernel version did you test it with? What values did the DRM_IOCTL_WAIT_VBLANK ioctl return to userspace? ...

[[EmilyD]] I run the driver on kernel version 4.6, and run glxgears with vsync enabled. It is hard to detail the issue, I will try my best to

description the issue I found.

Quoted 5 lines

After I double checked, it is not segment fault in libGL.so when run glxgears with vsync, but the glxgears will be stucked in waiting for the before swap buffers to complete. This is because when enable vsync, every time swap buffers, the DDX driver will call DRM_IOCTL_WAIT_VBLANK to queue the vblank event, and the

vbl.request.sequence will be set to current_vblank_count + swap_interval. Then in kernel driver, when timer interrupt occurs, it will call drm_handle_vblank_events, it will call drm_vblank_count_and_time to get current seq, and only seq >= vbl.request.sequence, then will call send_vblank_event. So it will never call send_vblank_event.

Quoted 9 lines

For example, the DDX driver call DRM_IOCTL_WAIT_VBLANK , then kernel driver will call drm_queue_vblank_event, and current vblank_count is 1 (As we only return 0 in vblank_get_counter, so the vblank_count will never change except calling drm_reset_vblank_timestamp which will make adev->ddev->vblank[crtc].count++), and swap_interval is 1, then adev->ddev->vbl.request.sequence will be 2, but the adev->ddev->drm_vblank_count_and_time will always
return 1 except calling drm_reset_vblank_timestamp(The function
drm_reset_vblank_timestamp will only be called in

drm_vblank_post_modeset and drm_vblank_on ), so the send_vblank_event will never be called, and swap buffers won't complete, so the glxgears will be stucked.

Looking at the drm_update_vblank_count() code, you also need to do the following in the virtual DCE case:

* Set dev->max_vblank_count = 0
* Make amdgpu_get_vblank_timestamp_kms either return values based on when the virtual vblank interrupt timer last fired, or just return a negative error code immediately, instead of calling drm_calc_vbltimestamp_from_scanoutpos
* Make amdgpu_get_vblank_counter_kms take the else path (or just return 0 immediately), not run any of the scanout position related code

Quoted 3 lines

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer

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

A message that is not in the archive (sent off-list or not yet received): DM5PR12MB16590BBC53DA15179AAA7DBD8F130-2J9CzHegvk/[email protected]

#13 lore raw
Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll probably run into issues sooner or later.

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): c43417fa-f177-e65c-76b2-2ed0296090d8-otUistvHUpPR7s880joybQ@public.gmane.org

#15 lore raw
Quoted 16 lines

-----Original Message-----
From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 12:01 PM
To: Deng, Emily <[email protected]>
Cc: [email protected]; [email protected]
Subject: Re: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Quoted 6 lines

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll probably run into issues sooner or later.

[[EmilyD]] Hi Michel, yes, you are right, we are pretending a hardware vblank counter in virtual DCE case to drm layer. But can you specific some cases that we must communicate with drm layer that we don't have the vblank counter.

Quoted 4 lines

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer

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

#16 lore raw
Quoted 10 lines · Emily
Quoted 10 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 12:01 PM

Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Quoted 8 lines
Quoted 6 lines

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll probably run into issues sooner or later.

[[EmilyD]] Hi Michel, yes, you are right, we are pretending a hardware vblank counter in virtual DCE case to drm layer. But can you specific some cases that we must communicate with drm layer that we don't have the vblank counter.

I described all the necessary steps (that I know of; there may be more) in https://lists.freedesktop.org/archives/amd-gfx/2016-August/001342.html .

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): 531e4e95-1350-3cc4-a354-89bdbb0c6309-otUistvHUpPR7s880joybQ@public.gmane.org

#18 lore raw
Quoted 27 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Michel D?nzer
Sent: Tuesday, August 16, 2016 2:43 PM
To: Deng, Emily <[email protected]>
Cc: [email protected]; [email protected]
Subject: Re: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

Quoted 10 lines · Emily
Quoted 10 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 12:01 PM On 16/08/16 12:49 PM, Deng,

Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Quoted 5 lines
Quoted 5 lines

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll

probably run into issues sooner or later.

Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 2 lines

[[EmilyD]] Hi Michel, yes, you are right, we are pretending a hardware vblank counter in virtual DCE case to drm layer. But can you specific some

cases that we must communicate with drm layer that we don't have the vblank counter.

I described all the necessary steps (that I know of; there may be more) in
https://lists.freedesktop.org/archives/amd-gfx/2016-August/001342.html .

[[EmilyD]] Hi Michel, I means can you detail the reasons or cases that we should communicate with drm layer when don't have the vblank counter instead of pretending a hardware vblank counter in virtual DCE case.

Quoted 8 lines

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

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

A message that is not in the archive (sent off-list or not yet received): DM5PR12MB1659A85C4EDDA59E5BF3CC7E8F130-2J9CzHegvk/[email protected]

#20 lore raw
Quoted 23 lines · Emily
Quoted 23 lines

From: amd-gfx [mailto:[email protected]] On Behalf Of Michel D?nzer
Sent: Tuesday, August 16, 2016 2:43 PM

Quoted 10 lines · Emily
Quoted 10 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 12:01 PM On 16/08/16 12:49 PM, Deng,

Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Quoted 5 lines
Quoted 5 lines

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll

probably run into issues sooner or later.

Quoted 3 lines
Quoted 1 line

[[EmilyD]] Hi Michel, yes, you are right, we are pretending a hardware vblank counter in virtual DCE case to drm layer. But can you specific some

Quoted 8 lines
Quoted 6 lines

cases that we must communicate with drm layer that we don't have the vblank counter.

I described all the necessary steps (that I know of; there may be more) in
https://lists.freedesktop.org/archives/amd-gfx/2016-August/001342.html .

[[EmilyD]] Hi Michel, I means can you detail the reasons or cases that we should communicate with drm layer when don't have the vblank counter instead of pretending a hardware vblank counter in virtual DCE case.

The DRM core code expects the get_vblank_counter hook to use a reliable hardware counter. If that is not the case, it should always return 0.

Similarly, drm_calc_vbltimestamp_from_scanoutpos expects the get_scanout_position hook to return accurate scanout position values based on reliable hardware registers.

If we pretend to return accurate values from these when we actually can't, the DRM core code may provide incorrect vblank counter / timestamp values to userspace, or worse.

Those are just off the top of my head, there may be more along the same lines.

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 3bbe3fc7-86cb-d077-6679-64bd37ecff12-otUistvHUpPR7s880joybQ@public.gmane.org

#22 lore raw

Hi Michel and Daniel,
Return the fake vblank count or return 0 in vblank_get_counter is only the virtual display feature's behavior, and the virtual display feature need to be enabled by set module parameter,
so won't affect normal case. And I think the vblank counter will be increased every time when the vsync timer interrupt occurs in virtual display feature is reasonable, so that the 3D app about vblank count could work normally. For the time stamp could be set to limitations for virtual display feature.

Best Wishes,
Emily Deng

Quoted 56 lines

-----Original Message-----
From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 3:18 PM
To: Deng, Emily <[email protected]>
Cc: [email protected]; [email protected]
Subject: Re: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

Quoted 22 lines · Emily
Quoted 22 lines

From: amd-gfx [mailto:[email protected]] On Behalf Of Michel D?nzer
Sent: Tuesday, August 16, 2016 2:43 PM On 16/08/16 03:28 PM, Deng,

Quoted 10 lines · Emily
Quoted 10 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 12:01 PM On 16/08/16 12:49 PM, Deng,

Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Quoted 6 lines
Quoted 6 lines

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll

probably run into issues sooner or later.

Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines

[[EmilyD]] Hi Michel, yes, you are right, we are pretending a hardware vblank counter in virtual DCE case to drm layer. But can you specific some

cases that we must communicate with drm layer that we don't have the vblank counter.

I described all the necessary steps (that I know of; there may be more) in https://lists.freedesktop.org/archives/amd-gfx/2016-

August/001342.html .

Quoted 3 lines
Quoted 1 line

[[EmilyD]] Hi Michel, I means can you detail the reasons or cases that we should communicate with drm layer when don't have the vblank counter

instead of pretending a hardware vblank counter in virtual DCE case.

The DRM core code expects the get_vblank_counter hook to use a reliable hardware counter. If that is not the case, it should always return 0.

Similarly, drm_calc_vbltimestamp_from_scanoutpos expects the get_scanout_position hook to return accurate scanout position values based on reliable hardware registers.

If we pretend to return accurate values from these when we actually can't, the DRM core code may provide incorrect vblank counter / timestamp values to userspace, or worse.

Quoted 6 lines

Those are just off the top of my head, there may be more along the same lines.

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer

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

A message that is not in the archive (sent off-list or not yet received): DM5PR12MB16597DC3495CA22D1FC61FA38F130-2J9CzHegvk/[email protected]

#24 lore raw

Hi Michel and Daniel,

     I just tryed Michel's advice: return 0 in vblank_get_counter, and set dev->max_vblank_count = 0 ,

and found the adev->ddev->vblank[crtc].count also can increase which makes the 3D app with vsync can work properly as well. But I don't know the principle. Anyway I will take Michel's advice about vblank count, and send a patch later.

Best Wishes,
Emily Deng

Quoted 81 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Deng, Emily
Sent: Tuesday, August 16, 2016 5:14 PM
To: Michel Dänzer <[email protected]>; Daniel Vetter <[email protected]>
Cc: [email protected]; [email protected]
Subject: RE: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

Hi Michel and Daniel,
Return the fake vblank count or return 0 in vblank_get_counter is only the
virtual display feature's behavior, and the virtual display feature need to be enabled by set module parameter, so won't affect normal case. And I think the vblank counter will be increased every time when the vsync timer interrupt occurs in virtual display feature is reasonable, so that the 3D app about vblank count could work normally. For the time stamp could be set to limitations for virtual display feature.

Best Wishes,
Emily Deng

Quoted 60 lines

-----Original Message-----
From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 3:18 PM
To: Deng, Emily <[email protected]>
Cc: [email protected]; [email protected]
Subject: Re: [PATCH] drm/amdgpu: For virtual_display feature, define a variable for vblank count of cpu vsync timer.

Quoted 22 lines · Emily
Quoted 22 lines

From: amd-gfx [mailto:[email protected]] On Behalf Of Michel D?nzer
Sent: Tuesday, August 16, 2016 2:43 PM On 16/08/16 03:28 PM, Deng,

Quoted 10 lines · Emily
Quoted 10 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 12:01 PM On 16/08/16 12:49 PM,

Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Quoted 6 lines
Quoted 6 lines

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll

probably run into issues sooner or later.

Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 8 lines
Quoted 3 lines

[[EmilyD]] Hi Michel, yes, you are right, we are pretending a hardware vblank counter in virtual DCE case to drm layer. But can you specific some

cases that we must communicate with drm layer that we don't have the vblank counter.

I described all the necessary steps (that I know of; there may be more) in https://lists.freedesktop.org/archives/amd-gfx/2016-

August/001342.html .

Quoted 4 lines
Quoted 1 line

[[EmilyD]] Hi Michel, I means can you detail the reasons or cases that we should communicate with drm layer when don't have the vblank counter

instead of pretending a hardware vblank counter in virtual DCE case.

The DRM core code expects the get_vblank_counter hook to use a reliable hardware counter. If that is not the case, it should always return 0.

Similarly, drm_calc_vbltimestamp_from_scanoutpos expects the get_scanout_position hook to return accurate scanout position values based on reliable hardware registers.

If we pretend to return accurate values from these when we actually can't, the DRM core code may provide incorrect vblank counter / timestamp values to userspace, or worse.

Those are just off the top of my head, there may be more along the same lines.

Quoted 7 lines
Quoted 3 lines

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer

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

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

#25 lore raw
In reply to Michel Dänzer
Quoted 11 lines · Michel Dänzer
Quoted 10 lines · Emily
Quoted 10 lines

From: Michel Dänzer [mailto:[email protected]]
Sent: Tuesday, August 16, 2016 12:01 PM

Quoted 6 lines · Emily

Hi Michel, Thanks, I still couldn't see the issue that use a variable to calculate the vblank count when vsync timer interrupt occurs. I just think it only emulates the hardware behavior. And the code change will only in virtual display files which won't touch drm layer. I incline to not modify the code in drm layer or amdgpu other codes, because it may lead to other issues .

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

I see it basically the other way around: We are currently pretending to the DRM core code that we have a reliable hardware vblank counter and timing even in the virtual DCE case, when we really don't. We should stop pretending and instead communicate the lack of these hardware facilities in the virtual DCE case as intended, otherwise we'll probably run into issues sooner or later.

Yes please don't lie to the vblank core. In the future we might depend more on this, and that might lead to surprises.

Quoted 5 lines
Quoted 2 lines

[[EmilyD]] Hi Michel, yes, you are right, we are pretending a hardware vblank counter in virtual DCE case to drm layer. But can you specific some cases that we must communicate with drm layer that we don't have the vblank counter.

I described all the necessary steps (that I know of; there may be more) in https://lists.freedesktop.org/archives/amd-gfx/2016-August/001342.html .

Yeah I missed the accurate timestamp part, since on i915 on gen2 we can still do that part. Your todo list looks accurate from what I can tell. -Daniel

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