Hi Vladimir, Thanks for pointing this out. Your analysis is right, and I have reproduced it on a ROCK 4D.
The problem is that a power domain marked need_regulator is switched off as soon as add_one_domain() adds it, and its subdomains are added only after that. NPUTOP, NPU0 and NPU1, which U-Boot left on, are neither idled nor really powered off, but genpd treats them as off. So when the regulator core disables vdd_npu_s0 because it has no users, the system hangs. I reproduced it on the ROCK 4D with a DT that has the NPU disabled. All 14 boots showed the same wrong state (0x214 = 0x23e, 0x110 = 0x13e1), and 1 of them hung. The ROCK 4D hangs less often than your R76S, but the underlying state is the same. The fix: when the power domains are added, a domain that needs a regulator and all of its subdomains are marked as off from the start. Once the whole subtree has been added, the domains are switched off in order, from the leaves up to the root. With the fix, all 10 boots showed the correct state (0x214 = 0x3fe) and none hung. I also tested it on the board with the NPU enabled, and all 30 power-ups were fine. If you have time, could you test the patch on the R76S as well? 10 boots are enough, without any workaround. That would show whether the fix works on your board. Patch: https://github.com/gahingwoo/build/blob/rk3576-pmdomain-probe-off/patch/kernel/archive/rockchip64-7.3/rk3576-pmdomain-power-off-npu-subtree-at-probe.patch Thanks again, this really helped find the root cause. Jiaxing
