The deferred regulator retrieval for Rockchip PM domains are causing some
weird dependencies. More precisely, if the power-domain is powered-on from
the HW perspective, its corresponding regulator must not be powered-off via
regulator_init_complete(), which is a late_initcall_sync.

Even on platforms that don't have the domain-supply regulator specified for
the power-domain provider, may suffer from these problems.

More precisely, things just happen to work before, because
genpd_power_off_unused() (also a late_initcall_sync) managed to power-off
the PM domain before regulator_init_complete() powered-off the regulator.

Ideally this fragile dependency must be fixed properly for the Rockchip PM
domains, but until then, let's fallback to the previous behaviour by using
the GENPD_FLAG_NO_STAY_ON flag.

Link: https://lore.kernel.org/all/20250902-rk3576-lockup-regression-v1-1-c4a0c9daeb00@collabora.com/
Reported-by: Nicolas Frattaroli <nicolas.frattaroli@collabora.com>
Cc: Heiko Stuebner <heiko@sntech.de>
Cc: Sebastian Reichel <sebastian.reichel@collabora.com>
Fixes: 0e789b491ba0 ("pmdomain: core: Leave powered-on genpds on until sync_state")
Fixes: 13a4b7fb6260 ("pmdomain: core: Leave powered-on genpds on until late_initcall_sync")
Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>
---
 drivers/pmdomain/rockchip/pm-domains.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/pmdomain/rockchip/pm-domains.c b/drivers/pmdomain/rockchip/pm-domains.c
index 242570c505fb..1955c6d453e4 100644
--- a/drivers/pmdomain/rockchip/pm-domains.c
+++ b/drivers/pmdomain/rockchip/pm-domains.c
@@ -865,7 +865,7 @@ static int rockchip_pm_add_one_domain(struct rockchip_pmu *pmu,
 	pd->genpd.power_on = rockchip_pd_power_on;
 	pd->genpd.attach_dev = rockchip_pd_attach_dev;
 	pd->genpd.detach_dev = rockchip_pd_detach_dev;
-	pd->genpd.flags = GENPD_FLAG_PM_CLK;
+	pd->genpd.flags = GENPD_FLAG_PM_CLK | GENPD_FLAG_NO_STAY_ON;
 	if (pd_info->active_wakeup)
 		pd->genpd.flags |= GENPD_FLAG_ACTIVE_WAKEUP;
 	pm_genpd_init(&pd->genpd, NULL,
-- 
2.43.0

On Tuesday, 9 September 2025 13:11:21 Central European Summer Time Ulf Hansson wrote:
> The deferred regulator retrieval for Rockchip PM domains are causing some
> weird dependencies. More precisely, if the power-domain is powered-on from
> the HW perspective, its corresponding regulator must not be powered-off via
> regulator_init_complete(), which is a late_initcall_sync.
> 
> Even on platforms that don't have the domain-supply regulator specified for
> the power-domain provider, may suffer from these problems.
> 
> More precisely, things just happen to work before, because
> genpd_power_off_unused() (also a late_initcall_sync) managed to power-off
> the PM domain before regulator_init_complete() powered-off the regulator.
> 
> Ideally this fragile dependency must be fixed properly for the Rockchip PM
> domains, but until then, let's fallback to the previous behaviour by using
> the GENPD_FLAG_NO_STAY_ON flag.
> 
> Link: https://lore.kernel.org/all/20250902-rk3576-lockup-regression-v1-1-c4a0c9daeb00@collabora.com/
> Reported-by: Nicolas Frattaroli <nicolas.frattaroli@collabora.com>
> Cc: Heiko Stuebner <heiko@sntech.de>
> Cc: Sebastian Reichel <sebastian.reichel@collabora.com>
> Fixes: 0e789b491ba0 ("pmdomain: core: Leave powered-on genpds on until sync_state")
> Fixes: 13a4b7fb6260 ("pmdomain: core: Leave powered-on genpds on until late_initcall_sync")
> Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>
> ---
>  drivers/pmdomain/rockchip/pm-domains.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/pmdomain/rockchip/pm-domains.c b/drivers/pmdomain/rockchip/pm-domains.c
> index 242570c505fb..1955c6d453e4 100644
> --- a/drivers/pmdomain/rockchip/pm-domains.c
> +++ b/drivers/pmdomain/rockchip/pm-domains.c
> @@ -865,7 +865,7 @@ static int rockchip_pm_add_one_domain(struct rockchip_pmu *pmu,
>  	pd->genpd.power_on = rockchip_pd_power_on;
>  	pd->genpd.attach_dev = rockchip_pd_attach_dev;
>  	pd->genpd.detach_dev = rockchip_pd_detach_dev;
> -	pd->genpd.flags = GENPD_FLAG_PM_CLK;
> +	pd->genpd.flags = GENPD_FLAG_PM_CLK | GENPD_FLAG_NO_STAY_ON;
>  	if (pd_info->active_wakeup)
>  		pd->genpd.flags |= GENPD_FLAG_ACTIVE_WAKEUP;
>  	pm_genpd_init(&pd->genpd, NULL,
> 

Tested-by: Nicolas Frattaroli <nicolas.frattaroli@collabora.com>

Fixes the full-SoC lockup I've observed on my RK3576 Radxa ROCK 4D,
which before this patch would occur when the vdd_npu_s0 regulator
was disabled.

Thank you!

Kind regards,
Nicolas Frattaroli

Am Dienstag, 9. September 2025, 13:11:21 Mitteleuropäische Sommerzeit schrieb Ulf Hansson:
> The deferred regulator retrieval for Rockchip PM domains are causing some
> weird dependencies. More precisely, if the power-domain is powered-on from
> the HW perspective, its corresponding regulator must not be powered-off via
> regulator_init_complete(), which is a late_initcall_sync.
> 
> Even on platforms that don't have the domain-supply regulator specified for
> the power-domain provider, may suffer from these problems.
> 
> More precisely, things just happen to work before, because
> genpd_power_off_unused() (also a late_initcall_sync) managed to power-off
> the PM domain before regulator_init_complete() powered-off the regulator.
> 
> Ideally this fragile dependency must be fixed properly for the Rockchip PM
> domains, but until then, let's fallback to the previous behaviour by using
> the GENPD_FLAG_NO_STAY_ON flag.
> 
> Link: https://lore.kernel.org/all/20250902-rk3576-lockup-regression-v1-1-c4a0c9daeb00@collabora.com/
> Reported-by: Nicolas Frattaroli <nicolas.frattaroli@collabora.com>
> Cc: Heiko Stuebner <heiko@sntech.de>
> Cc: Sebastian Reichel <sebastian.reichel@collabora.com>
> Fixes: 0e789b491ba0 ("pmdomain: core: Leave powered-on genpds on until sync_state")
> Fixes: 13a4b7fb6260 ("pmdomain: core: Leave powered-on genpds on until late_initcall_sync")
> Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>

NPU on rk3588 now again correctly gets its supplying regulator and
running said npu does not end up in a kernel panic :-)

Tested-by: Heiko Stuebner <heiko@sntech.de>
Acked-by: Heiko Stuebner <heiko@sntech.de>