Skip to contentSkip to conversation

drm/stm: Enable RPM during fbdev registration

PATCH In review Marek Vasut 3 participants 5 messages Nov 4, 2020 – Nov 6, 2020

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/stm: Enable RPM during fbdev registration

Enable runtime PM before registering the fbdev emulation and disable it afterward, otherwise register access to the LTDC IP during the fbdev emulation registration might hang the system.

The problem happens because RPM is activated at the end of ltdc_load(), but the fbdev emulation registration happens only after that, and ends up calling ltdc_crtc_mode_set_nofb(), which checks whether RPM is active and only if it is not active, calls pm_runtime_get_sync() to enable the clock and so on. If the clock are not enabled, any register access in ltdc_crtc_mode_set_nofb() could hang the platform completely.

This patch makes sure that ltdc_crtc_mode_set_nofb() is called within pm_runtime_get_sync(), so with clock enabled.

Cc Alexandre Torgue <[email protected]>
Cc Benjamin Gaignard <[email protected]>
Cc Philippe Cornu <[email protected]>
Cc Yannick Fertré <[email protected]>
---
 drivers/gpu/drm/stm/drv.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/stm/drv.c +2 -0
index 411103f013e2..d8921edc83db 100644
@@ -197,7 +197,9 @@ static int stm_drm_platform_probe(struct platform_device *pdev)
 	if (ret)
 		goto err_put;
 
+	pm_runtime_get_sync(ddev->dev);
 	drm_fbdev_generic_setup(ddev, 16);
+	pm_runtime_put_sync(ddev->dev);
 
 	return 0;
 
Signature · 7 lines
2.28.0


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
#1 lore raw
Quoted 14 lines · Marek Vasut

Enable runtime PM before registering the fbdev emulation and disable it afterward, otherwise register access to the LTDC IP during the fbdev emulation registration might hang the system.

The problem happens because RPM is activated at the end of ltdc_load(), but the fbdev emulation registration happens only after that, and ends up calling ltdc_crtc_mode_set_nofb(), which checks whether RPM is active and only if it is not active, calls pm_runtime_get_sync() to enable the clock and so on. If the clock are not enabled, any register access in ltdc_crtc_mode_set_nofb() could hang the platform completely.

This patch makes sure that ltdc_crtc_mode_set_nofb() is called within pm_runtime_get_sync(), so with clock enabled.

Quoted 8 lines
Cc Alexandre Torgue <[email protected]>
Cc Benjamin Gaignard <[email protected]>
Cc Philippe Cornu <[email protected]>
Cc Yannick Fertré <[email protected]>

This looks like you're papering over a bug in your modeset code. If userspace later on does a setpar on the fbdev chardev, the exact same thing could happen. You need to fix your modeset code to avoid this, not sprinkle temporary rpm_get/put all over some top level entry points, because you can't even patch those all.
-Daniel

Quoted 25 lines
---
 drivers/gpu/drm/stm/drv.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/stm/drv.c +2 -0
index 411103f013e2..d8921edc83db 100644
@@ -197,7 +197,9 @@ static int stm_drm_platform_probe(struct platform_device *pdev)
 	if (ret)
 		goto err_put;
 
+	pm_runtime_get_sync(ddev->dev);
 	drm_fbdev_generic_setup(ddev, 16);
+	pm_runtime_put_sync(ddev->dev);
 
 	return 0;
 
Signature · 6 lines
2.28.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
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
#2 lore raw
Quoted 14 lines · Daniel Vetter
Quoted 13 lines · Marek Vasut

Enable runtime PM before registering the fbdev emulation and disable it afterward, otherwise register access to the LTDC IP during the fbdev emulation registration might hang the system.

The problem happens because RPM is activated at the end of ltdc_load(), but the fbdev emulation registration happens only after that, and ends up calling ltdc_crtc_mode_set_nofb(), which checks whether RPM is active and only if it is not active, calls pm_runtime_get_sync() to enable the clock and so on. If the clock are not enabled, any register access in ltdc_crtc_mode_set_nofb() could hang the platform completely.

This patch makes sure that ltdc_crtc_mode_set_nofb() is called within pm_runtime_get_sync(), so with clock enabled.

[...]

Quoted 5 lines

This looks like you're papering over a bug in your modeset code. If userspace later on does a setpar on the fbdev chardev, the exact same thing could happen. You need to fix your modeset code to avoid this, not sprinkle temporary rpm_get/put all over some top level entry points, because you can't even patch those all.

I have a feeling all those pm_runtime_active() checks in the driver might be the root cause of this ? I wonder why the code doesn't use pm_runtime_{get,put}_sync() only when accessing registers. Thoughts?

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#3 lore raw

Hi Marek,

Quoted 24 lines · Marek Vasut
Quoted 14 lines · Daniel Vetter
Quoted 13 lines · Marek Vasut

Enable runtime PM before registering the fbdev emulation and disable it afterward, otherwise register access to the LTDC IP during the fbdev emulation registration might hang the system.

The problem happens because RPM is activated at the end of ltdc_load(), but the fbdev emulation registration happens only after that, and ends up calling ltdc_crtc_mode_set_nofb(), which checks whether RPM is active and only if it is not active, calls pm_runtime_get_sync() to enable the clock and so on. If the clock are not enabled, any register access in ltdc_crtc_mode_set_nofb() could hang the platform completely.

This patch makes sure that ltdc_crtc_mode_set_nofb() is called within pm_runtime_get_sync(), so with clock enabled.

[...]

Quoted 5 lines

This looks like you're papering over a bug in your modeset code. If userspace later on does a setpar on the fbdev chardev, the exact same thing could happen. You need to fix your modeset code to avoid this, not sprinkle temporary rpm_get/put all over some top level entry points, because you can't even patch those all.

Quoted 3 lines

I have a feeling all those pm_runtime_active() checks in the driver might be the root cause of this ? I wonder why the code doesn't use pm_runtime_{get,put}_sync() only when accessing registers. Thoughts?

First line of function ltdc_crtc_mode_set_nofb check the pm_runtime to avoid to access registers without clock enabled.

static void ltdc_crtc_mode_set_nofb(struct drm_crtc *crtc) {
...

	if (!pm_runtime_active(ddev->dev)) {
		ret = pm_runtime_get_sync(ddev->dev);

I test the fb with framebuffer console, & it works fine on my side. Do you test fb on a old kernel?
How can I reproduce your issue?

Best regards

Yannick
_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#4 lore raw
Quoted 1 line · Yannick FERTRE

Hi Marek,

Hi,

Quoted 37 lines · Marek Vasut
Quoted 24 lines · Daniel Vetter
Quoted 14 lines · Marek Vasut
Quoted 13 lines

Enable runtime PM before registering the fbdev emulation and disable it afterward, otherwise register access to the LTDC IP during the fbdev emulation registration might hang the system.

The problem happens because RPM is activated at the end of ltdc_load(), but the fbdev emulation registration happens only after that, and ends up calling ltdc_crtc_mode_set_nofb(), which checks whether RPM is active and only if it is not active, calls pm_runtime_get_sync() to enable the clock and so on. If the clock are not enabled, any register access in ltdc_crtc_mode_set_nofb() could hang the platform completely.

This patch makes sure that ltdc_crtc_mode_set_nofb() is called within pm_runtime_get_sync(), so with clock enabled.

[...]

Quoted 5 lines

This looks like you're papering over a bug in your modeset code. If userspace later on does a setpar on the fbdev chardev, the exact same thing could happen. You need to fix your modeset code to avoid this, not sprinkle temporary rpm_get/put all over some top level entry points, because you can't even patch those all.

Quoted 3 lines

I have a feeling all those pm_runtime_active() checks in the driver might be the root cause of this ? I wonder why the code doesn't use pm_runtime_{get,put}_sync() only when accessing registers. Thoughts?

First line of function ltdc_crtc_mode_set_nofb check the pm_runtime to avoid to access registers without clock enabled.

static void ltdc_crtc_mode_set_nofb(struct drm_crtc *crtc) {
...

Quoted 6 lines
	if (!pm_runtime_active(ddev->dev)) {
		ret = pm_runtime_get_sync(ddev->dev);

I test the fb with framebuffer console, & it works fine on my side. Do you test fb on a old kernel?
How can I reproduce your issue?

I observed sporadic hangs and tracked it down to the fbdev registration, which calls this code. Note that pm_runtime_active() here will return 0, because it was already activated in ltdc_load().

My question in reply to Daniel was more geared toward why do we even have all these checks in the driver, wouldn't it be better to just always call pm_runtime_get_sync()/pm_runtime_put_sync() in the code which requires the hardware to be active ?

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel