Re: [PATCH v3 2/5] pmdomain: core: Allow a non-CPU device in a CPU PM domain to do power on

[Ulf Hansson]

On Thu, Sep 10, 2026 at 8:31 PM Dhruva G <goledhruva@xxxxxxxxx> wrote:
>
> On 07-09-2026 16:46, Ulf Hansson wrote:
> > A driver for a non-CPU device that is attached to a CPU PM domain (the
> > genpd has the GENPD_FLAG_CPU_DOMAIN configuration set), is currently not
> > able to power on the PM domain. More precisely, to power on a CPU PM domain
> > one of its corresponding CPUs needs to be woken up if they are idle.
> >
> > The current support for a non-CPU device is that its driver can only
> > prevent an already powered on CPU PM domain from being powered off. This
> > leads to problems for a driver while probing its device or when it needs to
> > call pm_runtime_get_sync() to turn on the power for it. From the driver
> > point of view it looks like it all works fine, but when accessing the
> > device it may end up with various errors as the device may not be fully
> > powered on.
> >
> > To fix the behavior for these types of devices, let's adjust the behaviour
> > in genpd_power_on() to wake up an idle CPU that belongs to it, in cases
> > when it's needed.
> >
> > Link: https://lore.kernel.org/all/CAPx+jO-sCierYj8jnoKQHckJG16dOBxnNrsZVYO=38R2cLV8nw@xxxxxxxxxxxxxx/
> > Reviewed-by: Abel Vesa <abel.vesa@xxxxxxxxxxxxxxxx>
> > Tested-by: Yuanfang Zhang <yuanfang.zhang@xxxxxxxxxxxxxxxx>
> > Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxxxxxxxx>
> > ---
> >
> > Changes in v3:
> >       - Moved to atomic polling, pointed out by Dhruva.
>
> Thanks, but even in v3 we still have potential issues.
> It does not fully address the consequences of polling for five seconds in that context,
> nor the parent-lock nesting issue.

Well, I assume we will not be polling for 5s, as it would be an error
and it means that we fail to wake up the CPU. But, I get your point,
5s is really an unnecessary long timeout.

Ideally the timeout should map towards the deepest domain idle state's
entry+exit-latency-us, but rather than looking at what is actually
available for the PM domain(s) in question, I think it's easier (and
good enough) if we just pick a common value. Usually these values are
in the range of a couple milliseconds and in some cases up to
~15-20ms. I suggest we decrease the timeout to 300ms and see how that
plays out.

Also note that, at this point I don't know of any use cases similar to
what you describe, where the device in question is in an irqsafe child
domain. Hence the polling would not be done in an atomic context at
all, so we should be safe. Anyway, if this doesn't work we would
simply have to limit the support to non irqsafe child domains.

In regards to the parent-lock nesting issue. I don't think it's a
problem as genpd_wakeup_cpu() is not being called recursively, but let
me double check this to be sure.

Kind regards
Uffe

>
> This seems to have been pointed out by this corresponding sashiko review as well [1]
>
> [1] https://sashiko.dev/#/patchset/20260907111659.263324-1-ulf.hansson%40oss.qualcomm.com
>
> >
> > Changes in v2:
> >       - Rename a function according to Abel's suggestion.
> > ---
> >  drivers/pmdomain/core.c | 80 ++++++++++++++++++++++++++++++++++++++---
> >  1 file changed, 75 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/pmdomain/core.c b/drivers/pmdomain/core.c
> > index 6abe8b198949..b62d4e544bc5 100644
> > --- a/drivers/pmdomain/core.c
> > +++ b/drivers/pmdomain/core.c
> > @@ -10,6 +10,7 @@
> >  #include <linux/idr.h>
> >  #include <linux/kernel.h>
> >  #include <linux/io.h>
> > +#include <linux/iopoll.h>
> >  #include <linux/platform_device.h>
> >  #include <linux/pm_opp.h>
> >  #include <linux/pm_runtime.h>
> > @@ -19,11 +20,14 @@
> >  #include <linux/slab.h>
> >  #include <linux/err.h>
> >  #include <linux/sched.h>
> > +#include <linux/smp.h>
> >  #include <linux/suspend.h>
> >  #include <linux/export.h>
> >  #include <linux/cpu.h>
> >  #include <linux/debugfs.h>
> >
> > +#include <trace/events/ipi.h>
> > +
> >  /* Provides a unique ID for each genpd device */
> >  static DEFINE_IDA(genpd_ida);
> >
> > @@ -32,7 +36,9 @@ static const struct bus_type genpd_provider_bus_type = {
> >       .name           = "genpd_provider",
> >  };
> >
> > -#define GENPD_RETRY_MAX_MS   250             /* Approximate */
> > +#define GENPD_RETRY_MAX_MS           250             /* Approximate */
> > +#define GENPD_CPU_ON_POLL_PERIOD_US  100             /* 100us */
> > +#define GENPD_CPU_ON_TIMEOUT_US              5000000         /* 5s */
> >
> >  #define GENPD_DEV_CALLBACK(genpd, type, callback, dev)               \
> >  ({                                                           \
> > @@ -1026,15 +1032,75 @@ static void genpd_power_off(struct generic_pm_domain *genpd, bool one_dev_on,
> >       }
> >  }
> >
> > +static bool genpd_status_on(struct generic_pm_domain *genpd)
> > +{
> > +     bool is_on;
> > +
> > +     genpd_lock(genpd);
> > +     is_on = genpd_status_on_unlocked(genpd);
> > +     genpd_unlock(genpd);
> > +
> > +     return is_on;
> > +}
> > +
> > +static int genpd_wakeup_cpu(struct generic_pm_domain *genpd)
> > +{
> > +     unsigned int cpu;
> > +     bool is_on;
> > +     int ret;
> > +
> > +     /* Find the first online CPU in the genpd's cpumask. */
> > +     cpu = cpumask_first_and(genpd->cpus, cpu_online_mask);
> > +     if (cpu >= nr_cpu_ids)
> > +             return -EAGAIN;
> > +
> > +     genpd_unlock(genpd);
> > +
> > +     /* Send a IPI to wakeup the selected CPU. */
> > +     smp_send_reschedule(cpu);
> > +
> > +     /* Poll to wait for it to complete the power on sequence. */
> > +     ret = readx_poll_timeout_atomic(genpd_status_on, genpd, is_on, is_on,
> > +                                     GENPD_CPU_ON_POLL_PERIOD_US,
> > +                                     GENPD_CPU_ON_TIMEOUT_US);
> > +
> > +     genpd_lock(genpd);
> > +
> > +     /* Re-check the status as we have released the lock in between. */
> > +     if (ret || !genpd_status_on_unlocked(genpd))
> > +             return -EAGAIN;
> > +
> > +     return 0;
> > +}
> > +
> > +static bool genpd_need_alive_cpu(struct generic_pm_domain *genpd,
> > +                              struct device *dev)
> > +{
> > +     if (!genpd_is_cpu_domain(genpd))
> > +             return false;
> > +
> > +     /* This is not for CPU devices as those are managed differently. */
> > +     if (to_gpd_data(dev->power.subsys_data->domain_data)->cpu >= 0)
> > +             return false;
> > +
> > +     /*
> > +      * If the current CPU doesn't belong to the genpd's cpumask, we need to
> > +      * wake up one of those idle CPUs to power on the CPU domain correctly.
> > +      */
> > +     return !cpumask_test_cpu(smp_processor_id(), genpd->cpus);
> > +}
> > +
> >  /**
> >   * genpd_power_on - Restore power to a given PM domain and its parents.
> >   * @genpd: PM domain to power up.
> > + * @dev: The device that needs the PM domain to power on.
> >   * @depth: nesting count for lockdep.
> >   *
> >   * Restore power to @genpd and all of its parents so that it is possible to
> >   * resume a device belonging to it.
> >   */
> > -static int genpd_power_on(struct generic_pm_domain *genpd, unsigned int depth)
> > +static int genpd_power_on(struct generic_pm_domain *genpd, struct device *dev,
> > +                       unsigned int depth)
> >  {
> >       struct gpd_link *link;
> >       int ret = 0;
> > @@ -1042,6 +1108,10 @@ static int genpd_power_on(struct generic_pm_domain *genpd, unsigned int depth)
> >       if (genpd_status_on_unlocked(genpd))
> >               return 0;
> >
> > +     /* Special case for a device attached to a CPU domain. */
> > +     if (genpd_need_alive_cpu(genpd, dev))
> > +             return genpd_wakeup_cpu(genpd);
> > +
> >       /* Reflect over the entered idle-states residency for debugfs. */
> >       genpd_reflect_residency(genpd);
> >
> > @@ -1056,7 +1126,7 @@ static int genpd_power_on(struct generic_pm_domain *genpd, unsigned int depth)
> >               genpd_sd_counter_inc(parent);
> >
> >               genpd_lock_nested(parent, depth + 1);
> > -             ret = genpd_power_on(parent, depth + 1);
> > +             ret = genpd_power_on(parent, dev, depth + 1);
> >               genpd_unlock(parent);
> >
> >               if (ret) {
> > @@ -1306,7 +1376,7 @@ static int genpd_runtime_resume(struct device *dev)
> >
> >       genpd_lock(genpd);
> >       genpd_restore_performance_state(dev, gpd_data->rpm_pstate);
> > -     ret = genpd_power_on(genpd, 0);
> > +     ret = genpd_power_on(genpd, dev, 0);
> >       genpd_unlock(genpd);
> >
> >       if (ret)
> > @@ -3410,7 +3480,7 @@ static int __genpd_dev_pm_attach(struct device *dev, struct device *base_dev,
> >
> >       if (power_on) {
> >               genpd_lock(pd);
> > -             ret = genpd_power_on(pd, 0);
> > +             ret = genpd_power_on(pd, dev, 0);
> >               genpd_unlock(pd);
> >       }
> >
>