[...]

quoted

quoted

+
+       list_for_each_entry(de, &pd->dev_list, node) {
+               i += 1;
+               pm_clk_resume(pd->dev);

Do you really need to call pm_clk_resume() number of times that there
are devices in power domain? Did you want it to be

                pm_clk_resume(de->dev);

by any chance?

I was just about to ask the similar question as Dmitry did. :-)

You are right.I will modify in the next version.

Now, does that also mean you would like to assign the ->start|stop()
callbacks in the struct gpd_dev_ops to pm_clk_suspend|resume()? Or do
you intend to handle that from each driver instead?

quoted

quoted

+       }
+
+       /* no clk, set power domain will fail */
+       if (i == 0) {
+               pr_err("%s: failed to on/off power domain!", __func__);
+               spin_unlock_irq(&pd->dev_lock);
+               return ret;
+       }

Instead of counting I'd do

        if (list_empty(&pd->dev_list)) {
                pr_waen("%s: no devices in power domain\n", __func__);
                goto out;
        }

in the beginning of the function.

This is a good idea.

quoted

quoted

+
+       ret = rockchip_pmu_set_power_domain(pd, power_on);
+
+       list_for_each_entry(de, &pd->dev_list, node) {
+               pm_clk_suspend(pd->dev);

Same here?

quoted

+       }
+
+       spin_unlock_irq(&pd->dev_lock);
+
+       return ret;
+}
+
+static int rockchip_pd_power_on(struct generic_pm_domain *domain)
+{
+       struct rockchip_domain *pd = to_rockchip_pd(domain);
+
+       return rockchip_pd_power(pd, true);
+}
+
+static int rockchip_pd_power_off(struct generic_pm_domain *domain)
+{
+       struct rockchip_domain *pd = to_rockchip_pd(domain);
+
+       return rockchip_pd_power(pd, false);
+}
+
+void rockchip_pm_domain_attach_dev(struct device *dev)
+{
+       int ret;
+       int i = 0;
+       struct clk *clk;
+       struct rockchip_domain *pd;
+       struct rockchip_dev_entry *de;
+
+       pd = (struct rockchip_domain *)dev->pm_domain;
+       ret = pm_clk_create(dev);
+       if (ret) {
+               dev_err(dev, "pm_clk_create failed %d\n", ret);
+               return;
+       };

Stray semicolon.

quoted

+
+       while ((clk = of_clk_get(dev->of_node, i++)) && !IS_ERR(clk)) {
+               ret = pm_clk_add_clk(dev, clk);
+               if (ret) {
+                       dev_err(dev, "pm_clk_add_clk failed %d\n", ret);
+                       goto clk_err;
+               };
+       }
+
+       de = devm_kcalloc(pd->dev, 1,
+                       sizeof(struct rockchip_dev_entry *), GFP_KERNEL);

Why devm_calloc for a single element and not devm_kzalloc? Also, I am a
bit concerned about using devm_* API here. They are better reserved fir
driver's ->probe() paths whereas we are called from
dev_pm_domain_attach() which is more general API (yes, currently it is
used by buses probing code, but that might change in the future).

Using the devm_*API is supposed to work from here. I have kept this in
mind, while we added the new dev_pm_domain_attach|detach() API. The
buses also handles -EPROBE_DEFER.

Now, I just realized that while Geert added attach|detach_dev()
callbacks for the generic PM domain, those are both "void" callbacks.
It means the deferred probe error handling is broken for these
callbacks. We should convert the attach_dev() callback into an int, I
will cook a patch immediately.

quoted

Also, where is OOM error handling?

Ok,I will change the use  devm_kzalloc.
Register to pm domain devices, the number is not a lot.

[...]

Kind regards
Uffe