[PATCH v16 09/16] pmdomain: rockchip: add optional per-domain power-on settle delay

Vladimir Bely vlwwwwww at gmail.com
Fri Nov 13 05:23:22 PST 2026


On Sun, Nov 08, 2026 at 07:59:10PM +1300, Jiaxing Hu wrote:
> The NPU power domain is switched off when the driver probes, so the first
> power-on goes through this wait even when the bootloader left the NPU on.

[...]

> +	[RK3576_PD_NPU]		= DOMAIN_RK3576_R("npu",  0x0, BIT(0),  BIT(0), 0,       0x0, 0,       0,       0,       0,    false),

Hi Jiaxing,

(reply to https://lore.kernel.org/all/20261108065917.135194-10-gahing@gahingwoo.com/ )

This probe-time power-off hangs a board that does not enable the NPU.
I hit it with v14 as Armbian carries it on 7.3.3; v16 has the same
DOMAIN_RK3576_R("npu", ...), so I expect the same there, but I have not
run v16.

Board: FriendlyElec NanoPi R76S. Its DTS has no domain-supply on pd_npu
and the NPU nodes stay disabled. vdd_npu_s0 is RK806 dcdc-reg2,
regulator-boot-on, no consumers.

On about every second boot the SoC hangs hard ~30 s after start: no
oops, no answer to SysRq over the serial console, only the watchdog
gets it out. The last console line is "vdd_npu_s0: disabling" from
regulator_late_cleanup(). The kernel without the series (7.2.8) does
not hang. 10 boots each, no hang, with any of:

  - regulator_ignore_unused
  - regulator-always-on on vdd_npu_s0 only
  - the NPU enabled as in your ROCK 4D patch (domain-supply and
    npu-supply on vdd_npu_s0, cores and MMUs okay)

PMU registers after boot (regmap debugfs of 27380000.power-management):

               NPU not enabled   NPU enabled as on ROCK 4D
  0x110 req    0x000013e0        0x000013fe
  0x128 idle   0x000313e0        0x000313fe
  0x214 pwr    0x0000003e        0x000001fe
  0x230 stat   0x003e4181        0x01fe4181

With the NPU not enabled, nputop/npu0/npu1 (bits 6..8 at 0x214) are
still powered and their idle requests (bits 1..4 at 0x110) are not set,
while pm_genpd_summary lists npu, nputop, npu0 and npu1 as off-0 from
early boot.

My reading: rockchip_pm_add_one_domain() powers the parent "npu" domain
off at probe because of need_regulator, but the children U-Boot left on
are neither idled nor powered off first, and genpd then treats them as
off, so nothing cleans them up. When vdd_npu_s0 is cut later, the bus
hangs on those NIUs. I have not dumped the PMU on a kernel without the
series and have not tested a kernel fix, so this is a report rather
than a diagnosis. Idling and powering off the children that are on
before the parent at probe, or skipping the probe power-off when there
is no domain-supply, look like possible directions.

This would hit every RK3576 board that does not enable the NPU, which
is most of them.

More detail: https://github.com/gahingwoo/linux-rk3576-npu/issues/4
I am happy to test patches on the R76S; it has a serial console.

Thanks,
Vladimir



More information about the Linux-rockchip mailing list