On 8/3/2021 10:08 AM, Rajendra Nayak wrote:

On 8/2/2021 6:29 PM, Ulf Hansson wrote:

quoted

On Tue, 20 Jul 2021 at 09:12, Rajendra Nayak [off-list ref] wrote:

quoted

Some devices within power domains with performance states do not
support DVFS, but still need to vote on a default/static state
while they are active. They can express this using the 'required-opps'
property in device tree, which points to the phandle of the OPP
supported by the corresponding power-domains.

Add support to parse this information from DT and then set the
specified performance state during attach and drop it on detach.
runtime suspend/resume callbacks already have logic to drop/set
the vote as needed and should take care of dropping the default
perf state vote on runtime suspend and restore it back on runtime
resume.

Signed-off-by: Rajendra Nayak <redacted>
---
  drivers/base/power/domain.c | 28 +++++++++++++++++++++++++---
  include/linux/pm_domain.h   |  1 +
  2 files changed, 26 insertions(+), 3 deletions(-)

diff --git a/drivers/base/power/domain.c b/drivers/base/power/domain.c
index a934c67..f454031 100644
--- a/drivers/base/power/domain.c
+++ b/drivers/base/power/domain.c

@@ -2598,6 +2598,12 @@ static void genpd_dev_pm_detach(struct device *dev, bool power_off)

         dev_dbg(dev, "removing from PM domain %s\n", pd->name);

+       /* Drop the default performance state */
+       if (dev_gpd_data(dev)->default_pstate) {
+               dev_pm_genpd_set_performance_state(dev, 0);
+               dev_gpd_data(dev)->default_pstate = 0;
+       }
+
         for (i = 1; i < GENPD_RETRY_MAX_MS; i <<= 1) {
                 ret = genpd_remove_device(pd, dev);
                 if (ret != -EAGAIN)

@@ -2635,9 +2641,10 @@ static void genpd_dev_pm_sync(struct device *dev)

  static int __genpd_dev_pm_attach(struct device *dev, struct device *base_dev,
                                  unsigned int index, bool power_on)
  {
+       struct device_node *np;
         struct of_phandle_args pd_args;
         struct generic_pm_domain *pd;
-       int ret;
+       int ret, pstate;

         ret = of_parse_phandle_with_args(dev->of_node, "power-domains",
                                 "#power-domain-cells", index, &pd_args);

@@ -2675,10 +2682,25 @@ static int __genpd_dev_pm_attach(struct device *dev, struct device *base_dev,

                 genpd_unlock(pd);
         }

-       if (ret)
+       if (ret) {
                 genpd_remove_device(pd, dev);
+               return -EPROBE_DEFER;
+       }
+
+       /* Set the default performance state */
+       np = base_dev->of_node;

Please use dev->of_node instead (it is set to the same of_node as
base_dev by the callers of __genpd_dev_pm_attach) as it's more
consistent with existing code.

quoted

+       if (of_parse_phandle(np, "required-opps", index)) {
+               pstate = of_get_required_opp_performance_state(np, index);
+               if (pstate < 0) {
+                       ret = pstate;
+                       dev_err(dev, "failed to set required performance state for power-domain %s: %d\n",
+                               pd->name, ret);
+               }
+               dev_pm_genpd_set_performance_state(dev, pstate);
+               dev_gpd_data(dev)->default_pstate = pstate;

This doesn't look entirely correct to me. If we fail to translate a
required opp to a performance state, we shouldn't try to set it.

yeah, that does not seem right at all :(

quoted

Perhaps it's also easier to call
of_get_required_opp_performance_state() unconditionally of whether a
"required-opps" specifier exists. If it fails with the translation,
then we just skip setting a default state and continue with returning
1.

Would that work?

Looks like calling of_get_required_opp_performance_state() unconditionally
makes it spit out a pr_err() in case the node is missing "required-opps" property,
so I posted a v6 [1] with the check in place and adding the missing else
condition.

[1] https://lore.kernel.org/patchwork/project/lkml/list/?series=510727

-- 
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation