rk817: correct the fuel gauge across boot and sleep

[Jacob-Matthew-Cook]

Summary

@macromorgan @porschemad911 — flagging early since 001 and 002 are Chris's patches (002 co-developed), and the rk817 driver is his upstream.

• 001 — NVRAM soc misread (Chris Morgan). The stored value runs 0..100000, but the bounds check truncates at 10000, so a full battery reads back as 10%. Fixes: the original driver commit.
• 002 — boot soc resolved from the counter (Chris Morgan, co-developed). The boot value comes from PWRON_VOL when OFF_CNT reads three or more and from the columb counter otherwise, and both fail in the field. PWRON_VOL samples at the power-on instant, so a power-on caused by plugging in a charger samples under charge current — on an RG353M it read above the pack's maximum design voltage while the pack sat at 74%, and an RG DS was seen holding 2.63V while running. A negative counter was clamped to zero and reported 0%. The counter keeps counting while the system is off, so resolve from it first.
• 005 — gauge corrected across sleep (mine). The counter is erratic while the system sleeps: one 15 minute sleep on an RG353M charging at 1.4A registered 69% of the metered charge, another 139%, and a charge that finished during sleep still showed the pre-sleep percentage at wake. The correcting signals are gone by the time the driver runs again — RELAX_STS clears once the wakeup load exceeds the relax threshold (300mA, less than an idle UI draws) and CHARGE_FINISH disappears while the charger re-detects its input. Stamp boottime, charge, current and charging state at suspend; recompute from that evidence at resume.

Split out of #3220 at @porschemad911's suggestion, so each change is independently reviewable rather than one oversized PR. Every file here is unchanged from #3220 and has been hardware-tested on the affected devices; that PR itemizes every change and carries the full test record. Authorship and sign-offs are carried in the patch files themselves.

Testing

• Kernel patches (001/002/005) validated on an RG353M, reading kernel values over SSH rather than the UI.
• This is v4 of the fuel gauge work, simplified after Chris's review — the boot ladder was reduced to counter-first and FG_INIT dropped.
• voltage_ocv semantics confirmed on hardware by @macromorgan; suspend/resume confirmed under Rockchip BL31.
• Sleep-blindness, the dead relax gate, invisible CHARGE_FINISH and the reboot-at-full re-anchor were each measured directly.
• Off-charge test: saved at 74% and charged 1.5 hours powered off, the RG353M booted with the counter at full — which is what disproved the earlier v3 premise that the counter cannot count while powered off.
• Patches apply at strict zero fuzz (-F0) to both consumer kernels of the shared Rockchip tree (7.0.2 RK3566/76, 7.1.2 RK3326).
• Full test record is in RK3326/RK3566: battery, SD card and suspend/resume fixes #3220, where this was developed; the files here are unchanged from it.

Additional Context

• mainline-rockchip/ numbering skips 003/004, which are the dwc2 patches in dwc2: fix resume crash, suspend hang, and vbus leak on role switch #3303.

---

AI Usage

Did you use AI tools to help write this code? PARTIALLY