On Thu, 7 Aug 2025 at 11:38, Geert Uytterhoeven [off-list ref] wrote:

Hi Ulf,

On Wed, 30 Jul 2025 at 12:29, Ulf Hansson [off-list ref] wrote:

quoted

On Wed, 30 Jul 2025 at 11:56, Geert Uytterhoeven [off-list ref] wrote:

quoted

On Wed, 9 Jul 2025 at 13:31, Ulf Hansson [off-list ref] wrote:

quoted

On Tue, 1 Jul 2025 at 13:47, Ulf Hansson [off-list ref] wrote:

quoted

Changes in v3:
        - Added a couple of patches to adress problems on some Renesas
        platforms. Thanks Geert and Tomi for helping out!
        - Adressed a few comments from Saravanna and Konrad.
        - Added some tested-by tags.

I decided it was time to give this a try, so I have queued this up for
v6.17 via the next branch at my pmdomain tree.

If you encounter any issues, please let me know so I can help to fix them.

Thanks for your series!  Due to holidays, I only managed to test
this very recently.

Unfortunately I have an issue with unused PM Domains no longer being
disabled on R-Car:
  - On R-Car Gen1/2/3, using rcar-sysc.c, unused PM Domains are never
    disabled.
  - On R-Car Gen4, using rcar-gen4-sysc.c, unused PM Domains are
    sometimes not disabled.
    At first, I noticed the IOMMU driver was not enabled in my config,
    and enabling it did fix the issue.  However, after that I still
    encountered the issue in a different config that does have the
    IOMMU driver enabled...

FTR, unused PM Domains are still disabled correctly on R/SH-Mobile
(using rmobile-sysc.c) and on BeagleBone Black. Note that these use
of_genpd_add_provider_simple(), while all R-Car drivers use
of_genpd_add_provider_onecell().  Perhaps there is an issue with
the latter?  If you don't have a clue, I plan to do some more
investigation later...

of_genpd_add_provider_onecell() has:

    if (!dev)
            sync_state = true;
    else
            dev_set_drv_sync_state(dev, genpd_sync_state);

    for (i = 0; i < data->num_domains; i++) {
            ...
            if (sync_state && !genpd_is_no_sync_state(genpd)) {
                    genpd->sync_state = GENPD_SYNC_STATE_ONECELL;
                    device_set_node(&genpd->dev, fwnode);
                    sync_state = false;
                    ^^^^^^^^^^^^^^^^^^^
            }
            ...
    }

As the R-Car SYSC drivers are not platform drivers, dev is NULL, and
genpd->sync_state is set to GENPD_SYNC_STATE_ONECELL for the first PM
Domain only.  All other domains have the default value of sync_state
(0 = GENPD_SYNC_STATE_OFF).  Hence when genpd_provider_sync_state()
is called later, it ignores all but the first domain.
Apparently this is intentional, as of_genpd_sync_state() tries to
power off all domains handled by the same controller anyway (see below)?

Right, this is intentional and mainly because of how fw_devlink works.

fw_devlink is limited to use only the first device - if multiple
devices share the same fwnode. In principle, we could have picked any
of the devices in the array of genpds here - and reached the same
result.

quoted

quoted

BTW, the "pending due to"-messages look weird to me.
On R-Car M2-W (r8a7791.dtsi) I see e.g.:

    genpd_provider ca15-cpu0: sync_state() pending due to e6020000.watchdog
    renesas-cpg-mssr e6150000.clock-controller: sync_state() pending
due to e6020000.watchdog

ca15-cpu0 is the PM Domain holding the first CPU core, while
the watchdog resides in the always-on Clock Domain, and uses the
clock-controller for PM_CLK handling.

Unfortunately the first PM Domain is "ca15-cpu0", which is blocked on
these bogus pending states, and no PM Domain is powered off.

I see, thanks for the details. I am looking closer at this.

In any case, this is the main issue, as it prevents the ->sync_state()
callback to be called. Hence the "genpd->stay_on" will also *not* be
cleared for any of the genpd's for the genpd-provider.

If I remove the "sync_state = false" above, genpd_provider_sync_state()
considers all domains, and does power down all unused domains (even
multiple times, as expected).

I think those are getting called because with the change above, there
is no device_link being tracked. As stated above, fw_devlink is
limited to use only one device - if multiple devices share the same
fwnode.

In other words, the ->sync_state() callbacks are called even if the
corresponding consumer devices have not been probed yet.

Upon closer look, all "pending due to" messages I see claim that the
first (index 0) PM Domain is pending on some devices, while all of
these devices are part of a different domain (usually the always-on
domain, which is always the last (32 or 64) on R-Car).

So I think there are two issues:
  1. Devices are not attributed to the correct PM Domain using
     fw_devlink sync_state,
  2. One PM Domain of a multi-domain controller being blocked should
     not block all other domains handled by the same controller.

Right, that's a current limitation with fw_devlink. To cope with this,
it's possible to enforce the ->sync_state() callback to be invoked
from user-space (timeout or explicitly) for a device.

Another option would be to allow an opt-out behavior for some genpd's
that are powered-on at initialization. Something along the lines of
the below.

From: Ulf Hansson <redacted>
Date: Tue, 29 Jul 2025 14:27:22 +0200
Subject: [PATCH] pmdomain: core: Allow powered-on PM domains to be powered-off
 during boot

Signed-off-by: Ulf Hansson <redacted>
---
 drivers/pmdomain/core.c   | 3 ++-
 include/linux/pm_domain.h | 7 +++++++
 2 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/drivers/pmdomain/core.c b/drivers/pmdomain/core.c
index 0006ab3d0789..ef0760824c92 100644
--- a/drivers/pmdomain/core.c
+++ b/drivers/pmdomain/core.c

@@ -187,6 +187,7 @@ static const struct genpd_lock_ops genpd_raw_spin_ops = {
 #define genpd_is_opp_table_fw(genpd)   (genpd->flags & GENPD_FLAG_OPP_TABLE_FW)
 #define genpd_is_dev_name_fw(genpd)    (genpd->flags & GENPD_FLAG_DEV_NAME_FW)
 #define genpd_is_no_sync_state(genpd)  (genpd->flags &

GENPD_FLAG_NO_SYNC_STATE)
+#define genpd_is_no_stay_on(genpd)     (genpd->flags & GENPD_FLAG_NO_STAY_ON)

 static inline bool irq_safe_dev_in_sleep_domain(struct device *dev,
                const struct generic_pm_domain *genpd)

@@ -2392,7 +2393,7 @@ int pm_genpd_init(struct generic_pm_domain *genpd,
        INIT_WORK(&genpd->power_off_work, genpd_power_off_work_fn);
        atomic_set(&genpd->sd_count, 0);
        genpd->status = is_off ? GENPD_STATE_OFF : GENPD_STATE_ON;
-       genpd->stay_on = !is_off;
+       genpd->stay_on = !genpd_is_no_stay_on(genpd) && !is_off;
        genpd->sync_state = GENPD_SYNC_STATE_OFF;
        genpd->device_count = 0;
        genpd->provider = NULL;

diff --git a/include/linux/pm_domain.h b/include/linux/pm_domain.h
index b9d3c7d5c4f8..61b81574efc4 100644
--- a/include/linux/pm_domain.h
+++ b/include/linux/pm_domain.h

@@ -109,6 +109,12 @@ struct dev_pm_domain_list {
  *                             genpd provider specific way, likely through a
  *                             parent device node. This flag makes genpd to
  *                             skip its internal support for this.
+ *
+ * GENPD_FLAG_NO_STAY_ON:      A powered-on PM domain at initialization is
+ *                             prevented by genpd from being powered-off until
+ *                             we receive a ->sync_state() or runs the
+ *                             late_initcall_sync. Use this flag to allow
+ *                             power-off without waiting for these conditions.
  */
 #define GENPD_FLAG_PM_CLK       (1U << 0)
 #define GENPD_FLAG_IRQ_SAFE     (1U << 1)

@@ -120,6 +126,7 @@ struct dev_pm_domain_list {
 #define GENPD_FLAG_OPP_TABLE_FW         (1U << 7)
 #define GENPD_FLAG_DEV_NAME_FW  (1U << 8)
 #define GENPD_FLAG_NO_SYNC_STATE (1U << 9)
+#define GENPD_FLAG_NO_STAY_ON   (1U << 10)

 enum gpd_status {
        GENPD_STATE_ON = 0,     /* PM domain is on */

-- 
2.43.0

Kind regards
Uffe