| Message ID | 20190925213528.21515-3-khilman@kernel.org |
|---|---|
| State | New |
| Headers | show |
| Series | soc: amlogic: ee-pwrc: cleanup init state | expand |
On 25/09/2019 23:35, Kevin Hilman wrote: > From: Kevin Hilman <khilman@baylibre.com> > > During init, ensure that the driver on/off state as well as clock and > reset state matches the hardware state. Do this by always calling the > drivers 'on' function, and then callling the 'off' function if the > HW state was initially detected as off. > > Signed-off-by: Kevin Hilman <khilman@baylibre.com> > --- > drivers/soc/amlogic/meson-ee-pwrc.c | 29 ++++++++--------------------- > 1 file changed, 8 insertions(+), 21 deletions(-) > > diff --git a/drivers/soc/amlogic/meson-ee-pwrc.c b/drivers/soc/amlogic/meson-ee-pwrc.c > index dcce8e694a07..2e8eee0dc166 100644 > --- a/drivers/soc/amlogic/meson-ee-pwrc.c > +++ b/drivers/soc/amlogic/meson-ee-pwrc.c > @@ -323,6 +323,8 @@ static int meson_ee_pwrc_init_domain(struct platform_device *pdev, > struct meson_ee_pwrc *pwrc, > struct meson_ee_pwrc_domain *dom) > { > + bool is_off; > + > dom->pwrc = pwrc; > dom->num_rstc = dom->desc.reset_names_count; > dom->num_clks = dom->desc.clk_names_count; > @@ -356,27 +358,12 @@ static int meson_ee_pwrc_init_domain(struct platform_device *pdev, > dom->base.power_on = meson_ee_pwrc_on; > dom->base.power_off = meson_ee_pwrc_off; > > - /* > - * TOFIX: This is a special case for the VPU power domain, which can > - * be enabled previously by the bootloader. In this case the VPU > - * pipeline may be functional but no driver maybe never attach > - * to this power domain, and if the domain is disabled it could > - * cause system errors. This is why the pm_domain_always_on_gov > - * is used here. > - * For the same reason, the clocks should be enabled in case > - * we need to power the domain off, otherwise the internal clocks > - * prepare/enable counters won't be in sync. > - */ > - if (dom->num_clks && dom->desc.is_off && !dom->desc.is_off(dom)) { > - int ret = clk_bulk_prepare_enable(dom->num_clks, dom->clks); > - if (ret) > - return ret; > - > - pm_genpd_init(&dom->base, &pm_domain_always_on_gov, false); > - } else > - pm_genpd_init(&dom->base, NULL, > - (dom->desc.is_off ? > - dom->desc.is_off(dom) : true)); > + /* Ensure that driver state matches HW state */ > + is_off = dom->desc.is_off ? dom->desc.is_off(dom) : true; > + meson_ee_pwrc_on(&dom->base); > + if (is_off) > + meson_ee_pwrc_off(&dom->base); > + pm_genpd_init(&dom->base, NULL, is_off); > > return 0; > } > I don't see what you are trying to solve except simplifying the code. And the case is more that "matching the clock state" here, the pm_domain_always_on_gov was is a real case when booting from the Amlogic U-boot. The display power domain is complex and as been half solved by using "simple-framebuffer" on gx and is missing on g12a/g12b/sm1. For example, Debian installer runs without the modules, but will use the EFIfb set by U-Boot, but in this precise case : - the DRM driver isn't loaded - we can't hook this power domain with EFIfb When *not* in EFIfb, we use simple-framebuffer on GX, using this power domain, but it hasn't been copied to G12A. Personally I'll leave this code until we really tested and checked all uses cases, not only on the sei510/sei610 using mainline u-boot. Neil
Neil Armstrong <narmstrong@baylibre.com> writes: > On 25/09/2019 23:35, Kevin Hilman wrote: >> From: Kevin Hilman <khilman@baylibre.com> >> >> During init, ensure that the driver on/off state as well as clock and >> reset state matches the hardware state. Do this by always calling the >> drivers 'on' function, and then callling the 'off' function if the >> HW state was initially detected as off. [...] > I don't see what you are trying to solve except simplifying the code. Simplifying the code is a worthwhile goal on its own, but that's not the only thing I'm tring to accomplish. Part of the goal is make the init less VPU specific and thus make it more understandable/maintainable. But the more important part is to ensure that the driver state and HW state is consistent for all the domains (not just VPU.) I've had to debug lots of power-domain issues, and inconsistiences between HW state and driver state tend to be the primary cause of problems. Also I'm generally not a fan of the "don't mess with bootloader state" approach as that leads to subtle dependencies on specific bootloader versions that are also difficult to debug. Antother equally important goal is to actually be able to power-down the VPU when it's not used. Right now, we'll never power off the VPU if the bootloader enabled it, and that seems a bit extreme so I'm looking to improve that and be able to power off the VPU when not used. > And the case is more that "matching the clock state" here, the > pm_domain_always_on_gov was is a real case when booting from the Amlogic > U-boot. I'm not understanding what you mean about vendor uboot. I've done testing with vendor uboot too: I tried on g12a-u200, g12a-x96-max, and sm1-khadas-vim3l all of which have vendor uboot, and I tried before and after $SUBJECT patch. In all cases, I see the vendor uboot splash screen, and I see the framebuffer console from linux after kernel boot. I see the same behavior before and after my patch. I also tried on g12b-odroid-n2 (vendor uboot), and there is _no_ uboot spash screen, but I see the linux framebuffer console come up both before and after my patch. What's the specific case you're worried about with vendor uboot? Also... note something interesting I noticed on sm1-khadas-vim3l: before my patch, the framebuffer console appears, but the background is a bluish/green color. After my patch, the background is black (as expected.) > The display power domain is complex and as been half solved by using > "simple-framebuffer" on gx and is missing on g12a/g12b/sm1. > > For example, Debian installer runs without the modules, but will use > the EFIfb set by U-Boot, but in this precise case : > - the DRM driver isn't loaded > - we can't hook this power domain with EFIfb OK, so I agree that this case where we want the display to continue to work but linux DRM drivers never loaded is a special case. Is there a way to check if efifb is enabled? Is the the linux driver (drivers/video/fbdev/efifb.c) or something else only done by uboot? If we can detect efifb from the kernel (not just whether the domain is already on), then a simple addition to my patch like this: if (is_off) meson_ee_pwrc_off(&dom->base); + else if (efifb_is_enabled) + dom->base.flags |= GENPD_FLAG_ALWAYS_ON; would force the domain always-on in the case of efifb. In fact, now that I think of it, simply adding: if (is_off) meson_ee_pwrc_off(&dom->base); + else + dom->base.flags |= GENPD_FLAG_ALWAYS_ON; to my current patch would get back to the same behavior of the existing driver. But I still don't like the idea that we can *never* power off the VPU if the bootloader enables it. :( I'd rather see patches conditionally adding that flag with detailed explanations as to why it's needed. > When *not* in EFIfb, we use simple-framebuffer on GX, using this > power domain, but it hasn't been copied to G12A. I don't quite understand what problem simple-framebuffer is solving. Could you explain that in more detail. Assuming it is solving something, why can't it be used on g12[ab]/sm1 ? > Personally I'll leave this code until we really tested and checked all > uses cases, Right, I don't want to break anything on purpose, but I think the current state of this driver is fragile and difficult to understand/maintain, so I would be grateful for any help in understanding the corner cases better, as well as testing them (or explaining to me how to reproduce them.) Generally, with long-term maintenance in mind, if I'm forced to choose between understandable/maintainable code and "covers 100% of corner cases" I will most often chose the former. > not only on the sei510/sei610 using mainline u-boot. See above about all the other boards with vendor uboots also tested. Kevin
On 26/09/2019 21:08, Kevin Hilman wrote: > Neil Armstrong <narmstrong@baylibre.com> writes: > >> On 25/09/2019 23:35, Kevin Hilman wrote: >>> From: Kevin Hilman <khilman@baylibre.com> >>> >>> During init, ensure that the driver on/off state as well as clock and >>> reset state matches the hardware state. Do this by always calling the >>> drivers 'on' function, and then callling the 'off' function if the >>> HW state was initially detected as off. > > [...] > >> I don't see what you are trying to solve except simplifying the code. > > Simplifying the code is a worthwhile goal on its own, but that's not the > only thing I'm tring to accomplish. I still find it ugly to power_on a domain to power it off right afterwards. The issue is with the CCF enable handling which is not in sync with the HW, if you boot with an already enabled clock, it won't be marked enabled in CCF, and it's clearly bad when you want to have a fine-tuned gate state handling. > > Part of the goal is make the init less VPU specific and thus make it > more understandable/maintainable. But the more important part is to > ensure that the driver state and HW state is consistent for all the > domains (not just VPU.) I've had to debug lots of power-domain issues, > and inconsistiences between HW state and driver state tend to be the > primary cause of problems. > > Also I'm generally not a fan of the "don't mess with bootloader state" > approach as that leads to subtle dependencies on specific bootloader > versions that are also difficult to debug. > > Antother equally important goal is to actually be able to power-down the > VPU when it's not used. Right now, we'll never power off the VPU if the > bootloader enabled it, and that seems a bit extreme so I'm looking to > improve that and be able to power off the VPU when not used. > >> And the case is more that "matching the clock state" here, the >> pm_domain_always_on_gov was is a real case when booting from the Amlogic >> U-boot. > > I'm not understanding what you mean about vendor uboot. I've done > testing with vendor uboot too: > > I tried on g12a-u200, g12a-x96-max, and sm1-khadas-vim3l all of which > have vendor uboot, and I tried before and after $SUBJECT patch. > > In all cases, I see the vendor uboot splash screen, and I see the > framebuffer console from linux after kernel boot. I see the same > behavior before and after my patch. > > I also tried on g12b-odroid-n2 (vendor uboot), and there is _no_ uboot > spash screen, but I see the linux framebuffer console come up both > before and after my patch. Thanks for testing all these cases > > What's the specific case you're worried about with vendor uboot? It's an issue I got when bringing up mainline uboot and the vpu power controller driver, I had regressions on GXBB and GXL boards. But it seems it's no more the case on G12A/G12B, we'll see this when GX support will be added in this driver. > > Also... note something interesting I noticed on sm1-khadas-vim3l: > before my patch, the framebuffer console appears, but the background is > a bluish/green color. After my patch, the background is black (as > expected.) Yes it's an issue I have on my infinite todo list, it's a different init done by the latest vendor uboot. I have a fix but it seems it breaks display when booting other boards. > >> The display power domain is complex and as been half solved by using >> "simple-framebuffer" on gx and is missing on g12a/g12b/sm1. >> >> For example, Debian installer runs without the modules, but will use >> the EFIfb set by U-Boot, but in this precise case : >> - the DRM driver isn't loaded >> - we can't hook this power domain with EFIfb > > OK, so I agree that this case where we want the display to continue to > work but linux DRM drivers never loaded is a special case. > > Is there a way to check if efifb is enabled? Is the the linux driver > (drivers/video/fbdev/efifb.c) or something else only done by uboot? > > If we can detect efifb from the kernel (not just whether the domain is > already on), then a simple addition to my patch like this: > > if (is_off) > meson_ee_pwrc_off(&dom->base); > + else if (efifb_is_enabled) > + dom->base.flags |= GENPD_FLAG_ALWAYS_ON; > > would force the domain always-on in the case of efifb. > > In fact, now that I think of it, simply adding: > > if (is_off) > meson_ee_pwrc_off(&dom->base); > + else > + dom->base.flags |= GENPD_FLAG_ALWAYS_ON; > > to my current patch would get back to the same behavior of the existing > driver. But I still don't like the idea that we can *never* power off > the VPU if the bootloader enables it. :( I'd rather see patches > conditionally adding that flag with detailed explanations as to why it's > needed. > >> When *not* in EFIfb, we use simple-framebuffer on GX, using this >> power domain, but it hasn't been copied to G12A. > > I don't quite understand what problem simple-framebuffer is > solving. Could you explain that in more detail. simple-framebuffer has the power domain hooked in it's node, so when u-boot will boot linux with HDMI enabled it will enable this node and until the DRM driver loads the simple-framebuffer will live and keep the power domain enabled. > > Assuming it is solving something, why can't it be used on g12[ab]/sm1 ? It will, but I need to push the patches. > >> Personally I'll leave this code until we really tested and checked all >> uses cases, > > Right, I don't want to break anything on purpose, but I think the > current state of this driver is fragile and difficult to > understand/maintain, so I would be grateful for any help in > understanding the corner cases better, as well as testing them (or > explaining to me how to reproduce them.) > > Generally, with long-term maintenance in mind, if I'm forced to choose > between understandable/maintainable code and "covers 100% of corner > cases" I will most often chose the former. > >> not only on the sei510/sei610 using mainline u-boot. > > See above about all the other boards with vendor uboots also tested. Let's leave apart the vendor uboot issue for g12. Since display support for G12A will land soon in mainline U-boot, let me push the DT patches for simple-framebuffer to we have a fallback in case the DRM driver isn't loaded/built. We can consider that in case of EFIfb, the simple-framebuffer node would have been enabled and the power domain hooked and enabled. Neil > > Kevin >
Neil Armstrong <narmstrong@baylibre.com> writes: > On 26/09/2019 21:08, Kevin Hilman wrote: >> Neil Armstrong <narmstrong@baylibre.com> writes: >> >>> On 25/09/2019 23:35, Kevin Hilman wrote: >>>> From: Kevin Hilman <khilman@baylibre.com> >>>> >>>> During init, ensure that the driver on/off state as well as clock and >>>> reset state matches the hardware state. Do this by always calling the >>>> drivers 'on' function, and then callling the 'off' function if the >>>> HW state was initially detected as off. >> >> [...] >> >>> I don't see what you are trying to solve except simplifying the code. >> >> Simplifying the code is a worthwhile goal on its own, but that's not the >> only thing I'm tring to accomplish. > > I still find it ugly to power_on a domain to power it off right afterwards. > The issue is with the CCF enable handling which is not in sync with the > HW, if you boot with an already enabled clock, it won't be marked enabled > in CCF, and it's clearly bad when you want to have a fine-tuned gate state > handling. It's not just the clocks. The only thing we actually know is the HW state of the sleep bit (because we read it.) We don't know the state of all the mem_pd bits or the iso_reg, nor do we know the state of the reset lines. Calling 'on' ensures everything is where we expect, and we're not relying on the bootloader to do it. >> Part of the goal is make the init less VPU specific and thus make it >> more understandable/maintainable. But the more important part is to >> ensure that the driver state and HW state is consistent for all the >> domains (not just VPU.) I've had to debug lots of power-domain issues, >> and inconsistiences between HW state and driver state tend to be the >> primary cause of problems. >> >> Also I'm generally not a fan of the "don't mess with bootloader state" >> approach as that leads to subtle dependencies on specific bootloader >> versions that are also difficult to debug. >> >> Antother equally important goal is to actually be able to power-down the >> VPU when it's not used. Right now, we'll never power off the VPU if the >> bootloader enabled it, and that seems a bit extreme so I'm looking to >> improve that and be able to power off the VPU when not used. >> >>> And the case is more that "matching the clock state" here, the >>> pm_domain_always_on_gov was is a real case when booting from the Amlogic >>> U-boot. >> >> I'm not understanding what you mean about vendor uboot. I've done >> testing with vendor uboot too: >> >> I tried on g12a-u200, g12a-x96-max, and sm1-khadas-vim3l all of which >> have vendor uboot, and I tried before and after $SUBJECT patch. >> >> In all cases, I see the vendor uboot splash screen, and I see the >> framebuffer console from linux after kernel boot. I see the same >> behavior before and after my patch. >> >> I also tried on g12b-odroid-n2 (vendor uboot), and there is _no_ uboot >> spash screen, but I see the linux framebuffer console come up both >> before and after my patch. > > Thanks for testing all these cases > >> >> What's the specific case you're worried about with vendor uboot? > > It's an issue I got when bringing up mainline uboot and the vpu power controller > driver, I had regressions on GXBB and GXL boards. > But it seems it's no more the case on G12A/G12B, we'll see this when > GX support will be added in this driver. > >> Also... note something interesting I noticed on sm1-khadas-vim3l: >> before my patch, the framebuffer console appears, but the background is >> a bluish/green color. After my patch, the background is black (as >> expected.) > > Yes it's an issue I have on my infinite todo list, it's a different > init done by the latest vendor uboot. I have a fix but it seems it > breaks display when booting other boards. > >> >>> The display power domain is complex and as been half solved by using >>> "simple-framebuffer" on gx and is missing on g12a/g12b/sm1. >>> >>> For example, Debian installer runs without the modules, but will use >>> the EFIfb set by U-Boot, but in this precise case : >>> - the DRM driver isn't loaded >>> - we can't hook this power domain with EFIfb >> >> OK, so I agree that this case where we want the display to continue to >> work but linux DRM drivers never loaded is a special case. >> >> Is there a way to check if efifb is enabled? Is the the linux driver >> (drivers/video/fbdev/efifb.c) or something else only done by uboot? >> >> If we can detect efifb from the kernel (not just whether the domain is >> already on), then a simple addition to my patch like this: >> >> if (is_off) >> meson_ee_pwrc_off(&dom->base); >> + else if (efifb_is_enabled) >> + dom->base.flags |= GENPD_FLAG_ALWAYS_ON; >> >> would force the domain always-on in the case of efifb. >> >> In fact, now that I think of it, simply adding: >> >> if (is_off) >> meson_ee_pwrc_off(&dom->base); >> + else >> + dom->base.flags |= GENPD_FLAG_ALWAYS_ON; >> >> to my current patch would get back to the same behavior of the existing >> driver. But I still don't like the idea that we can *never* power off >> the VPU if the bootloader enables it. :( I'd rather see patches >> conditionally adding that flag with detailed explanations as to why it's >> needed. >> >>> When *not* in EFIfb, we use simple-framebuffer on GX, using this >>> power domain, but it hasn't been copied to G12A. >> >> I don't quite understand what problem simple-framebuffer is >> solving. Could you explain that in more detail. > > simple-framebuffer has the power domain hooked in it's node, so > when u-boot will boot linux with HDMI enabled it will enable > this node and until the DRM driver loads the simple-framebuffer > will live and keep the power domain enabled. Ah, thanks for the explanation. I didn't realize it was u-boot that was adding/enabling the simplefb node in the DT. >> >> Assuming it is solving something, why can't it be used on g12[ab]/sm1 ? > > It will, but I need to push the patches. > OK. >>> Personally I'll leave this code until we really tested and checked all >>> uses cases, >> >> Right, I don't want to break anything on purpose, but I think the >> current state of this driver is fragile and difficult to >> understand/maintain, so I would be grateful for any help in >> understanding the corner cases better, as well as testing them (or >> explaining to me how to reproduce them.) >> >> Generally, with long-term maintenance in mind, if I'm forced to choose >> between understandable/maintainable code and "covers 100% of corner >> cases" I will most often chose the former. >> >>> not only on the sei510/sei610 using mainline u-boot. >> >> See above about all the other boards with vendor uboots also tested. > > Let's leave apart the vendor uboot issue for g12. > > Since display support for G12A will land soon in mainline U-boot, let me > push the DT patches for simple-framebuffer to we have a fallback in > case the DRM driver isn't loaded/built. > > We can consider that in case of EFIfb, the simple-framebuffer node would have > been enabled and the power domain hooked and enabled. OK, that makes sense to me. So can I consider this an 'ack' from you for this patch, as long as I wait to apply it until the simplefb DT patches are also merged? Thanks, Kevin
On Fri 27 Sep 2019 at 08:37, Neil Armstrong <narmstrong@baylibre.com> wrote: > On 26/09/2019 21:08, Kevin Hilman wrote: >> Neil Armstrong <narmstrong@baylibre.com> writes: >> >>> On 25/09/2019 23:35, Kevin Hilman wrote: >>>> From: Kevin Hilman <khilman@baylibre.com> >>>> >>>> During init, ensure that the driver on/off state as well as clock and >>>> reset state matches the hardware state. Do this by always calling the >>>> drivers 'on' function, and then callling the 'off' function if the >>>> HW state was initially detected as off. >> >> [...] >> >>> I don't see what you are trying to solve except simplifying the code. >> >> Simplifying the code is a worthwhile goal on its own, but that's not the >> only thing I'm tring to accomplish. > > I still find it ugly to power_on a domain to power it off right afterwards. > The issue is with the CCF enable handling which is not in sync with the > HW, if you boot with an already enabled clock, it won't be marked enabled > in CCF, and it's clearly bad when you want to have a fine-tuned gate state > handling. > CCF should disable unused clock so, in theory, you should not have to call enable() then disable() to get things in sync. I suppose the clock in question has the flag CLK_IGNORE_UNUSED (one of the gates) ? If the CLK_INGORE_UNUSED is becoming a problem, it would be better to fix the clock tree rather than adding quirks in consumers.
Jerome Brunet <jbrunet@baylibre.com> writes: > On Fri 27 Sep 2019 at 08:37, Neil Armstrong <narmstrong@baylibre.com> wrote: > >> On 26/09/2019 21:08, Kevin Hilman wrote: >>> Neil Armstrong <narmstrong@baylibre.com> writes: >>> >>>> On 25/09/2019 23:35, Kevin Hilman wrote: >>>>> From: Kevin Hilman <khilman@baylibre.com> >>>>> >>>>> During init, ensure that the driver on/off state as well as clock and >>>>> reset state matches the hardware state. Do this by always calling the >>>>> drivers 'on' function, and then callling the 'off' function if the >>>>> HW state was initially detected as off. >>> >>> [...] >>> >>>> I don't see what you are trying to solve except simplifying the code. >>> >>> Simplifying the code is a worthwhile goal on its own, but that's not the >>> only thing I'm tring to accomplish. >> >> I still find it ugly to power_on a domain to power it off right afterwards. >> The issue is with the CCF enable handling which is not in sync with the >> HW, if you boot with an already enabled clock, it won't be marked enabled >> in CCF, and it's clearly bad when you want to have a fine-tuned gate state >> handling. >> > > CCF should disable unused clock so, in theory, you should not have to > call enable() then disable() to get things in sync. But CCF won't disabled unused clocks until late(ish) in the boot process, which is also when the unused PM domains will be disabled, so I think there's still a potential for race between the late "disable unused" features of clocks and pm-domains. Kevin
diff --git a/drivers/soc/amlogic/meson-ee-pwrc.c b/drivers/soc/amlogic/meson-ee-pwrc.c index dcce8e694a07..2e8eee0dc166 100644 --- a/drivers/soc/amlogic/meson-ee-pwrc.c +++ b/drivers/soc/amlogic/meson-ee-pwrc.c @@ -323,6 +323,8 @@ static int meson_ee_pwrc_init_domain(struct platform_device *pdev, struct meson_ee_pwrc *pwrc, struct meson_ee_pwrc_domain *dom) { + bool is_off; + dom->pwrc = pwrc; dom->num_rstc = dom->desc.reset_names_count; dom->num_clks = dom->desc.clk_names_count; @@ -356,27 +358,12 @@ static int meson_ee_pwrc_init_domain(struct platform_device *pdev, dom->base.power_on = meson_ee_pwrc_on; dom->base.power_off = meson_ee_pwrc_off; - /* - * TOFIX: This is a special case for the VPU power domain, which can - * be enabled previously by the bootloader. In this case the VPU - * pipeline may be functional but no driver maybe never attach - * to this power domain, and if the domain is disabled it could - * cause system errors. This is why the pm_domain_always_on_gov - * is used here. - * For the same reason, the clocks should be enabled in case - * we need to power the domain off, otherwise the internal clocks - * prepare/enable counters won't be in sync. - */ - if (dom->num_clks && dom->desc.is_off && !dom->desc.is_off(dom)) { - int ret = clk_bulk_prepare_enable(dom->num_clks, dom->clks); - if (ret) - return ret; - - pm_genpd_init(&dom->base, &pm_domain_always_on_gov, false); - } else - pm_genpd_init(&dom->base, NULL, - (dom->desc.is_off ? - dom->desc.is_off(dom) : true)); + /* Ensure that driver state matches HW state */ + is_off = dom->desc.is_off ? dom->desc.is_off(dom) : true; + meson_ee_pwrc_on(&dom->base); + if (is_off) + meson_ee_pwrc_off(&dom->base); + pm_genpd_init(&dom->base, NULL, is_off); return 0; }