On Fri, Sep 29, 2023 at 03:14:07PM +0200, Ulf Hansson wrote:
On Wed, 13 Sept 2023 at 14:26, Stephan Gerhold stephan@gerhold.net wrote:
On Wed, Sep 13, 2023 at 12:56:16PM +0200, Ulf Hansson wrote:
On Tue, 12 Sept 2023 at 11:40, Stephan Gerhold stephan.gerhold@kernkonzept.com wrote:
[...] However, at the moment nothing ever enables the virtual devices created in qcom-cpufreq-nvmem for the cpufreq power domain scaling, so they are permanently runtime-suspended.
Fix this by enabling the devices after attaching them and use dev_pm_syscore_device() to ensure the power domain also stays on when going to suspend. Since it supplies the CPU we can never turn it off from Linux. There are other mechanisms to turn it off when needed, usually in the RPM firmware or the cpuidle path.
Without this fix performance states votes are silently ignored, and the CPU/CPR voltage is never adjusted. This has been broken since 5.14 but for some reason no one noticed this on QCS404 so far.
Cc: stable@vger.kernel.org Fixes: 1cb8339ca225 ("cpufreq: qcom: Add support for qcs404 on nvmem driver") Signed-off-by: Stephan Gerhold stephan.gerhold@kernkonzept.com
drivers/cpufreq/qcom-cpufreq-nvmem.c | 21 ++++++++++++++++++++- 1 file changed, 20 insertions(+), 1 deletion(-)
diff --git a/drivers/cpufreq/qcom-cpufreq-nvmem.c b/drivers/cpufreq/qcom-cpufreq-nvmem.c index 84d7033e5efe..17d6ab14c909 100644 --- a/drivers/cpufreq/qcom-cpufreq-nvmem.c +++ b/drivers/cpufreq/qcom-cpufreq-nvmem.c @@ -25,6 +25,7 @@ #include <linux/platform_device.h> #include <linux/pm_domain.h> #include <linux/pm_opp.h> +#include <linux/pm_runtime.h> #include <linux/slab.h> #include <linux/soc/qcom/smem.h>
@@ -280,6 +281,7 @@ static int qcom_cpufreq_probe(struct platform_device *pdev) }
for_each_possible_cpu(cpu) {
struct device **virt_devs = NULL; struct dev_pm_opp_config config = { .supported_hw = NULL, };@@ -300,7 +302,7 @@ static int qcom_cpufreq_probe(struct platform_device *pdev)
if (drv->data->genpd_names) { config.genpd_names = drv->data->genpd_names;
config.virt_devs = NULL;
config.virt_devs = &virt_devs; } if (config.supported_hw || config.genpd_names) {@@ -311,6 +313,23 @@ static int qcom_cpufreq_probe(struct platform_device *pdev) goto free_opp; } }
if (virt_devs) {const char * const *name = config.genpd_names;int i;for (i = 0; *name; i++, name++) {ret = pm_runtime_resume_and_get(virt_devs[i]);if (ret) {dev_err(cpu_dev, "failed to resume %s: %d\n",*name, ret);goto free_opp;}Shouldn't we restore the usage count at ->remove() too?
/* Keep CPU power domain always-on */dev_pm_syscore_device(virt_devs[i], true);Is this really correct? cpufreq is suspended/resumed by the PM core during system wide suspend/resume. See dpm_suspend|resume(). Isn't that sufficient?
Moreover, it looks like the cpr genpd provider supports genpd's ->power_on|off() callbacks. Is there something wrong with this, that I am missing?
I think this question is a quite fundamental one. To explain this properly I will need to delve a bit into the implementation details of the two different GENPD providers that are applicable here:
Fundamentally, we are describing the main power supply for the CPU here. Consider a simple regulator with adjustable voltage. From the Linux point of view this regulator should be marked as "regulator-always-on". If we would turn off this regulator, the CPU would be immediately dead without proper shutdown done by firmware or hardware.
Representing the regulator as power domain does not change much, except that we now have abstract "performance states" instead of actual voltages. However, for power domains there is currently no generic mechanism like "regulator-always-on" in the DT, only drivers can specify GENPD_FLAG_ALWAYS_ON.
We have relied on genpd providers to act on their compatible strings to make the correct configuration. If that isn't sufficient, I don't see why we couldn't add a new DT property corresponding to GENPD_FLAG_ALWAYS_ON.
Right. It's not completely trivial though, since a DT node may provide many different power domains with #power-domain-cells = <N>. A regulator on the other hand has a dedicated DT node where you can just add "regulator-always-on". :')
The special situation on MSM8909 is that there are two possible setups for the CPU power supply depending on the PMIC that is used (see "[PATCH 4/4] cpufreq: qcom-nvmem: Add MSM8909"): CPR or RPMPD. Both are GENPD providers so in theory we can just have either
cpu@0 { power-domains = <&cpr>; }; // or cpu@0 { power-domains = <&rpmpd MSM8909_VDDCX_AO>; };
in the DT, without handling this specifically on the cpufreq side.
Looks like it would be nice to get a patch for the MSM8909 DTS too, as part of the series, to get a better picture of how this is going to be used. Would that be possible for you to provide?
Sure! Right now I cannot include it as working patch in this series since I don't have the base SoC DT (msm8909.dtsi) upstream yet. It's mostly a copy-paste of msm8916.dtsi so I was trying to finish up the SoC-specific parts before sending it.
I'm happy to provide links to the full DT and my changes though. Does that help? If you would like to comment inline I could copy paste the diffs in a mail or include some kind of RFC patch. It just wouldn't be possible to apply it successfully. :')
Here are the two commits with the my current DT changes (WIP): - MSM8909+PM8909 (RPMPD only): https://github.com/msm8916-mainline/linux/commit/791e0c5a3162372a0738bc7b0f4... - MSM8916 (CPR+RPMPD): https://github.com/msm8916-mainline/linux/commit/8880f39108206d7a60a0a8351c0... (- QCS404 (CPR only): already in mainline (see qcs404.dtsi))
The two GENPD providers behave quite differently though:
CPR: CPR is not really a power domain itself. It's more like a monitor on a power supply line coming from some other regulator. CPR provides suggestions how to adjust the voltage for best power/stability.
The GENPD .power_off() disables the CPR state machine and forwards this to the regulator with regulator_disable(). On QCS404 the regulator is marked regulator-always-on, so it will never be disabled from Linux. The SAW/SPM hardware component on Qualcomm SoCs will usually disable the regulator during deep cpuidle states.
Parts of this sound a bit odd to me. The CPR/CPUfreq shouldn't really need to vote for the CPU's power-rail(s) from a powered-on/off (CPU idle states) point of view, but only from a performance (voltage level) point of view.
If the enable/disable voting on the regulator really has an impact on some platforms, it sounds like it could prevent deeper CPU idle states too. That's probably not what we want, right?
I think this heavily depends on what exactly this "regulator" represents. Are we talking about a physical regulator with a binary enable/disable signal or actually some hardware/firmware magic that combines multiple independent "votes"?
If we are talking about a physical regulator then we can never disable it from Linux. Not even during CPU idle states. It would just cut off all power immediately and kill the CPU without proper shutdown. Instead, the platform might have special hardware/firmware functionality that will control the actual physical enable/disable signal of the regulator.
I also had a look at the existing CPR genpd provider's probe function/path (cpr_probe()) - and it turns out there is no call to regulator_enable(). Whatever that means to us.
In most (all?) setups the CPR genpd provider will manage the actual physical regulator. It could be part of the PMIC or even some off-the-shelf regulator with I2C control. It doesn't matter. There is nothing special about that regulator. You have the standard Linux regulator driver, set up the DT node for it and hook it up to CPR.
Now, to prevent the regulator driver in Linux from touching the physical enable signal (see above) we add "regulator-always-on". When Linux requests deep CPU idle states via PSCI the hardware will toggle the physical enable/disable signal of the regulator for us (after the CPU has been shut down).
On some platforms CPR is also used for the GPU or other power rails that are not critical for the CPU to run. In that case it's fine to disable the regulator directly from Linux. Just not for the CPU.
RPMPD: This is the generic driver for all the SoC power domains managed by the RPM firmware. It's not CPU-specific. However, as special feature each power domain is exposed twice in Linux, e.g. "MSM8909_VDDCX" and "MSM8909_VDDCX_AO". The _AO ("active-only") variant tells the RPM firmware that the performance/enable vote only applies when the CPU is active (not in deep cpuidle state).
The GENPD .power_off() drops all performance state votes and also releases the "enable" vote for the power domain.
Now, imagine what happens during system wide suspend/resume:
- CPR: The CPR state machine gets disabled. The voltage stays as-is.
- With "regulator-always-on": The CPU keeps running until WFI.
- Without: I would expect the CPU is dead immediately(?)
As I indicated above, I am starting to feel that this is a bit upside down. CPR/CPUfreq should vote on voltages to scale performance, but not for cpu idle states.
Perhaps what is missing is a synchronization point or a notification, to inform the CPR driver that its state machine (registers) needs to be saved/restored, when the power-rails get turned on/off. In fact, we have a couple mechanisms at hand to support this.
I think we can ignore this part of CPR for now. AFAICT Qualcomm's vendor driver does not explicitly disable the CPR state machine during CPU idle when the power rails are potentially turned off. They only do it during system wide suspend, for whatever reason. For that we don't need such a notification mechanism.
- RPMPD: The performance/enable vote is dropped. The power domain might go to minimal voltage or even turn off completely. However, the CPU actually needs to keep running at the same frequency until WFI! Worst case, the CPU is dead immediately when the power domain votes get dropped.
Since RPMPD is managing the voting for both performance and low power states for different kinds of devices, this certainly gets a bit more complicated.
On the other hand, the CPUfreq driver should really only vote for performance states for the CPUs and not for low power states. The latter is a job for cpuidle and other consumers of the RPMPD to manage, I think.
So, while thinking of this, I just realized that it may not always be a good idea for genpd to cache a performance state request, for an attached device and for which pm_runtime_suspended() returns true for it. As this is the default behaviour in genpd, I am thinking that we need a way to make that behaviour configurable for an attached device. What do you think about that?
Hm. This would be a bit of a special case of course. But I think this would be fine to solve the regression for CPR on QCS404.
Then we "just" need to solve the fundamental question from a few years ago: Who *will* actually vote for enabling the power domains/regulators required by the CPU? :D
I agree that enabling/disabling power supplies feels closer to cpuidle. But it's not a perfect fit either, given that we don't actually want to change our vote while entering CPU idle states. I think on all platforms I'm looking at here we need a permanent enable vote (effectively making the regulator/power domains always-on from the Linux point of view).
We could solve this by adding a "regulator-always-on" mechanism in the DT for power domains. This feels more like a workaround to me than an actual solution. With this the CPU won't appear as always-on consumer of the power domains in debugfs. There will just be a "suspended" consumer attributed to the CPU (from CPUfreq, since we don't have a dedicated device for CPUfreq).
While this patch is a bit strange from a conceptual perspective, on the implementation side it effectively makes that CPU consumer appear as active. So the end result is actually kind of the one we need. :'D
Thanks, Stephan