Re: [PATCH v3 12/12] drm/panel: panel-simple: Use runtime pm to avoid excessive unprepare / prepare
Hi Doug, On Wed, Apr 14, 2021 at 06:22:57PM -0700, Doug Anderson wrote: > On Wed, Apr 14, 2021 at 5:58 PM Laurent Pinchart wrote: > > On Fri, Apr 02, 2021 at 03:28:46PM -0700, Douglas Anderson wrote: > > > Unpreparing and re-preparing a panel can be a really heavy > > > operation. Panels datasheets often specify something on the order of > > > 500ms as the delay you should insert after turning off the panel > > > before turning it on again. In addition, turning on a panel can have > > > delays on the order of 100ms - 200ms before the panel will assert HPD > > > (AKA "panel ready"). The above means that we should avoid turning a > > > panel off if we're going to turn it on again shortly. > > > > > > The above becomes a problem when we want to read the EDID of a > > > panel. The way that ordering works is that userspace wants to read the > > > EDID of the panel _before_ fully enabling it so that it can set the > > > initial mode correctly. However, we can't read the EDID until we power > > > it up. This leads to code that does this dance (like > > > ps8640_bridge_get_edid()): > > > > > > 1. When userspace requests EDID / the panel modes (through an ioctl), > > >we power on the panel just enough to read the EDID and then power > > >it off. > > > 2. Userspace then turns the panel on. > > > > > > There's likely not much time between step #1 and #2 and so we want to > > > avoid powering the panel off and on again between those two steps. > > > > > > Let's use Runtime PM to help us. We'll move the existing prepare() and > > > unprepare() to be runtime resume() and runtime suspend(). Now when we > > > want to prepare() or unprepare() we just increment or decrement the > > > refcount. We'll default to a 1 second autosuspend delay which seems > > > sane given the typical delays we see for panels. > > > > > > A few notes: > > > - It seems the existing unprepare() and prepare() are defined to be > > > no-ops if called extra times. We'll preserve that behavior. > > > > The prepare and unprepare calls are supposed to be balanced, which > > should allow us to drop this check. Do you have a reason to suspect that > > it may not be the case ? > > No, it was just code inspection. The old code definitely made an > effort to make enable of an already enabled panel a no-op and disable > of an already disabled panel a no-op. This is even before my > (somewhat) recent patch to make things timing based, though I did > touch the code. > > Can I maybe suggest that getting rid of the extra check should be a > separate patch after this one? Then if it breaks someone it's easy to > just revert that one and we can still keep the runtime pm? Sounds good to me. > > > - This is a slight change in the ABI of simple panel. If something was > > > absolutely relying on the unprepare() to happen instantly that > > > simply won't be the case anymore. I'm not aware of anyone relying on > > > that behavior, but if there is someone then we'll need to figure out > > > how to enable (or disable) this new delayed behavior selectively. > > > - In order for this to work we now have a hard dependency on > > > "PM". From memory this is a legit thing to assume these days and we > > > don't have to find some fallback to keep working if someone wants to > > > build their system without "PM". > > > > Sounds fine to me. > > > > The code looks good to me. Possibly with the prepared check removed, > > > > Reviewed-by: Laurent Pinchart -- Regards, Laurent Pinchart
Re: [PATCH v3 12/12] drm/panel: panel-simple: Use runtime pm to avoid excessive unprepare / prepare
Hi, On Wed, Apr 14, 2021 at 5:58 PM Laurent Pinchart wrote: > > Hi Doug, > > Thank you for the patch. > > On Fri, Apr 02, 2021 at 03:28:46PM -0700, Douglas Anderson wrote: > > Unpreparing and re-preparing a panel can be a really heavy > > operation. Panels datasheets often specify something on the order of > > 500ms as the delay you should insert after turning off the panel > > before turning it on again. In addition, turning on a panel can have > > delays on the order of 100ms - 200ms before the panel will assert HPD > > (AKA "panel ready"). The above means that we should avoid turning a > > panel off if we're going to turn it on again shortly. > > > > The above becomes a problem when we want to read the EDID of a > > panel. The way that ordering works is that userspace wants to read the > > EDID of the panel _before_ fully enabling it so that it can set the > > initial mode correctly. However, we can't read the EDID until we power > > it up. This leads to code that does this dance (like > > ps8640_bridge_get_edid()): > > > > 1. When userspace requests EDID / the panel modes (through an ioctl), > >we power on the panel just enough to read the EDID and then power > >it off. > > 2. Userspace then turns the panel on. > > > > There's likely not much time between step #1 and #2 and so we want to > > avoid powering the panel off and on again between those two steps. > > > > Let's use Runtime PM to help us. We'll move the existing prepare() and > > unprepare() to be runtime resume() and runtime suspend(). Now when we > > want to prepare() or unprepare() we just increment or decrement the > > refcount. We'll default to a 1 second autosuspend delay which seems > > sane given the typical delays we see for panels. > > > > A few notes: > > - It seems the existing unprepare() and prepare() are defined to be > > no-ops if called extra times. We'll preserve that behavior. > > The prepare and unprepare calls are supposed to be balanced, which > should allow us to drop this check. Do you have a reason to suspect that > it may not be the case ? No, it was just code inspection. The old code definitely made an effort to make enable of an already enabled panel a no-op and disable of an already disabled panel a no-op. This is even before my (somewhat) recent patch to make things timing based, though I did touch the code. Can I maybe suggest that getting rid of the extra check should be a separate patch after this one? Then if it breaks someone it's easy to just revert that one and we can still keep the runtime pm? > > - This is a slight change in the ABI of simple panel. If something was > > absolutely relying on the unprepare() to happen instantly that > > simply won't be the case anymore. I'm not aware of anyone relying on > > that behavior, but if there is someone then we'll need to figure out > > how to enable (or disable) this new delayed behavior selectively. > > - In order for this to work we now have a hard dependency on > > "PM". From memory this is a legit thing to assume these days and we > > don't have to find some fallback to keep working if someone wants to > > build their system without "PM". > > Sounds fine to me. > > The code looks good to me. Possibly with the prepared check removed, > > Reviewed-by: Laurent Pinchart Thanks!
Re: [PATCH v3 12/12] drm/panel: panel-simple: Use runtime pm to avoid excessive unprepare / prepare
Hi Doug, Thank you for the patch. On Fri, Apr 02, 2021 at 03:28:46PM -0700, Douglas Anderson wrote: > Unpreparing and re-preparing a panel can be a really heavy > operation. Panels datasheets often specify something on the order of > 500ms as the delay you should insert after turning off the panel > before turning it on again. In addition, turning on a panel can have > delays on the order of 100ms - 200ms before the panel will assert HPD > (AKA "panel ready"). The above means that we should avoid turning a > panel off if we're going to turn it on again shortly. > > The above becomes a problem when we want to read the EDID of a > panel. The way that ordering works is that userspace wants to read the > EDID of the panel _before_ fully enabling it so that it can set the > initial mode correctly. However, we can't read the EDID until we power > it up. This leads to code that does this dance (like > ps8640_bridge_get_edid()): > > 1. When userspace requests EDID / the panel modes (through an ioctl), >we power on the panel just enough to read the EDID and then power >it off. > 2. Userspace then turns the panel on. > > There's likely not much time between step #1 and #2 and so we want to > avoid powering the panel off and on again between those two steps. > > Let's use Runtime PM to help us. We'll move the existing prepare() and > unprepare() to be runtime resume() and runtime suspend(). Now when we > want to prepare() or unprepare() we just increment or decrement the > refcount. We'll default to a 1 second autosuspend delay which seems > sane given the typical delays we see for panels. > > A few notes: > - It seems the existing unprepare() and prepare() are defined to be > no-ops if called extra times. We'll preserve that behavior. The prepare and unprepare calls are supposed to be balanced, which should allow us to drop this check. Do you have a reason to suspect that it may not be the case ? > - This is a slight change in the ABI of simple panel. If something was > absolutely relying on the unprepare() to happen instantly that > simply won't be the case anymore. I'm not aware of anyone relying on > that behavior, but if there is someone then we'll need to figure out > how to enable (or disable) this new delayed behavior selectively. > - In order for this to work we now have a hard dependency on > "PM". From memory this is a legit thing to assume these days and we > don't have to find some fallback to keep working if someone wants to > build their system without "PM". Sounds fine to me. The code looks good to me. Possibly with the prepared check removed, Reviewed-by: Laurent Pinchart > Signed-off-by: Douglas Anderson > --- > > (no changes since v1) > > drivers/gpu/drm/panel/Kconfig| 1 + > drivers/gpu/drm/panel/panel-simple.c | 93 +--- > 2 files changed, 73 insertions(+), 21 deletions(-) > > diff --git a/drivers/gpu/drm/panel/Kconfig b/drivers/gpu/drm/panel/Kconfig > index 4894913936e9..ef87d92cdf49 100644 > --- a/drivers/gpu/drm/panel/Kconfig > +++ b/drivers/gpu/drm/panel/Kconfig > @@ -80,6 +80,7 @@ config DRM_PANEL_SIMPLE > tristate "support for simple panels" > depends on OF > depends on BACKLIGHT_CLASS_DEVICE > + depends on PM > select VIDEOMODE_HELPERS > help > DRM panel driver for dumb panels that need at most a regulator and > diff --git a/drivers/gpu/drm/panel/panel-simple.c > b/drivers/gpu/drm/panel/panel-simple.c > index be312b5c04dd..6b22872b3281 100644 > --- a/drivers/gpu/drm/panel/panel-simple.c > +++ b/drivers/gpu/drm/panel/panel-simple.c > @@ -27,6 +27,7 @@ > #include > #include > #include > +#include > #include > > #include > @@ -175,6 +176,8 @@ struct panel_simple { > bool enabled; > bool no_hpd; > > + bool prepared; > + > ktime_t prepared_time; > ktime_t unprepared_time; > > @@ -334,19 +337,31 @@ static int panel_simple_disable(struct drm_panel *panel) > return 0; > } > > +static int panel_simple_suspend(struct device *dev) > +{ > + struct panel_simple *p = dev_get_drvdata(dev); > + > + gpiod_set_value_cansleep(p->enable_gpio, 0); > + regulator_disable(p->supply); > + p->unprepared_time = ktime_get(); > + > + return 0; > +} > + > static int panel_simple_unprepare(struct drm_panel *panel) > { > struct panel_simple *p = to_panel_simple(panel); > + int ret; > > - if (p->prepared_time == 0) > + /* Unpreparing when already unprepared is a no-op */ > + if (!p->prepared) > return 0; > > - gpiod_set_value_cansleep(p->enable_gpio, 0); > - > - regulator_disable(p->supply); > - > - p->prepared_time = 0; > - p->unprepared_time = ktime_get(); > + pm_runtime_mark_last_busy(panel->dev); > + ret = pm_runtime_put_autosuspend(panel->dev); > + if (ret < 0) > + return ret; > + p->prepared = false; > > return 0; > } > @@ -376,22
[PATCH v3 12/12] drm/panel: panel-simple: Use runtime pm to avoid excessive unprepare / prepare
Unpreparing and re-preparing a panel can be a really heavy operation. Panels datasheets often specify something on the order of 500ms as the delay you should insert after turning off the panel before turning it on again. In addition, turning on a panel can have delays on the order of 100ms - 200ms before the panel will assert HPD (AKA "panel ready"). The above means that we should avoid turning a panel off if we're going to turn it on again shortly. The above becomes a problem when we want to read the EDID of a panel. The way that ordering works is that userspace wants to read the EDID of the panel _before_ fully enabling it so that it can set the initial mode correctly. However, we can't read the EDID until we power it up. This leads to code that does this dance (like ps8640_bridge_get_edid()): 1. When userspace requests EDID / the panel modes (through an ioctl), we power on the panel just enough to read the EDID and then power it off. 2. Userspace then turns the panel on. There's likely not much time between step #1 and #2 and so we want to avoid powering the panel off and on again between those two steps. Let's use Runtime PM to help us. We'll move the existing prepare() and unprepare() to be runtime resume() and runtime suspend(). Now when we want to prepare() or unprepare() we just increment or decrement the refcount. We'll default to a 1 second autosuspend delay which seems sane given the typical delays we see for panels. A few notes: - It seems the existing unprepare() and prepare() are defined to be no-ops if called extra times. We'll preserve that behavior. - This is a slight change in the ABI of simple panel. If something was absolutely relying on the unprepare() to happen instantly that simply won't be the case anymore. I'm not aware of anyone relying on that behavior, but if there is someone then we'll need to figure out how to enable (or disable) this new delayed behavior selectively. - In order for this to work we now have a hard dependency on "PM". From memory this is a legit thing to assume these days and we don't have to find some fallback to keep working if someone wants to build their system without "PM". Signed-off-by: Douglas Anderson --- (no changes since v1) drivers/gpu/drm/panel/Kconfig| 1 + drivers/gpu/drm/panel/panel-simple.c | 93 +--- 2 files changed, 73 insertions(+), 21 deletions(-) diff --git a/drivers/gpu/drm/panel/Kconfig b/drivers/gpu/drm/panel/Kconfig index 4894913936e9..ef87d92cdf49 100644 --- a/drivers/gpu/drm/panel/Kconfig +++ b/drivers/gpu/drm/panel/Kconfig @@ -80,6 +80,7 @@ config DRM_PANEL_SIMPLE tristate "support for simple panels" depends on OF depends on BACKLIGHT_CLASS_DEVICE + depends on PM select VIDEOMODE_HELPERS help DRM panel driver for dumb panels that need at most a regulator and diff --git a/drivers/gpu/drm/panel/panel-simple.c b/drivers/gpu/drm/panel/panel-simple.c index be312b5c04dd..6b22872b3281 100644 --- a/drivers/gpu/drm/panel/panel-simple.c +++ b/drivers/gpu/drm/panel/panel-simple.c @@ -27,6 +27,7 @@ #include #include #include +#include #include #include @@ -175,6 +176,8 @@ struct panel_simple { bool enabled; bool no_hpd; + bool prepared; + ktime_t prepared_time; ktime_t unprepared_time; @@ -334,19 +337,31 @@ static int panel_simple_disable(struct drm_panel *panel) return 0; } +static int panel_simple_suspend(struct device *dev) +{ + struct panel_simple *p = dev_get_drvdata(dev); + + gpiod_set_value_cansleep(p->enable_gpio, 0); + regulator_disable(p->supply); + p->unprepared_time = ktime_get(); + + return 0; +} + static int panel_simple_unprepare(struct drm_panel *panel) { struct panel_simple *p = to_panel_simple(panel); + int ret; - if (p->prepared_time == 0) + /* Unpreparing when already unprepared is a no-op */ + if (!p->prepared) return 0; - gpiod_set_value_cansleep(p->enable_gpio, 0); - - regulator_disable(p->supply); - - p->prepared_time = 0; - p->unprepared_time = ktime_get(); + pm_runtime_mark_last_busy(panel->dev); + ret = pm_runtime_put_autosuspend(panel->dev); + if (ret < 0) + return ret; + p->prepared = false; return 0; } @@ -376,22 +391,19 @@ static int panel_simple_get_hpd_gpio(struct device *dev, return 0; } -static int panel_simple_prepare_once(struct drm_panel *panel) +static int panel_simple_prepare_once(struct panel_simple *p) { - struct panel_simple *p = to_panel_simple(panel); + struct device *dev = p->base.dev; unsigned int delay; int err; int hpd_asserted; unsigned long hpd_wait_us; - if (p->prepared_time != 0) - return 0; - panel_simple_wait(p->unprepared_time, p->desc->delay.unprepare);