kit/linux-gts6l
1
0
Fork 0
Code Issues 18 Pull requests Projects Releases Packages Wiki Activity Actions

Power management: low-power sleep (s2idle) investigation #3

New issue
Open
opened 2026-08-20 13:17:10 +00:00 by kit · 1 comment
kit commented 2026-08-20 13:17:10 +00:00
Owner
Copy link

Goal: usable battery life via suspend. Constraints discovered so far:

  • Samsung's PSCI is nonstandard (psci: failed to set PC mode: -3 every boot), so deep PSCI suspend states may be off-limits; s2idle (suspend-to-idle) with cpuidle + RPMh sleep/wake state voting is the realistic first target.
  • cpuidle currently runs (cpuidle.off crutch long dropped); need to check which idle states actually enter (/sys/devices/system/cpu/cpu*/cpuidle/state*/usage).
  • RPMh sleep votes need the sync_state situation resolved (chronic sync_state() pending for gcc/rpmhpd/gpucc due to gmu/dispcc consumers) — related to the regulator_ignore_unused crutch still in the cmdline; proper rail ownership is a prerequisite.
  • Watchdogs must be suspend-aware (APSS watchdog + our boot/reachability layers) or a sleeping tablet gets rebooted.
  • Measure baseline idle power first (battery fuel gauge readable?) so improvements are quantifiable.
Goal: usable battery life via suspend. Constraints discovered so far: - Samsung's PSCI is nonstandard (`psci: failed to set PC mode: -3` every boot), so deep PSCI suspend states may be off-limits; s2idle (suspend-to-idle) with cpuidle + RPMh sleep/wake state voting is the realistic first target. - cpuidle currently runs (cpuidle.off crutch long dropped); need to check which idle states actually enter (`/sys/devices/system/cpu/cpu*/cpuidle/state*/usage`). - RPMh sleep votes need the `sync_state` situation resolved (chronic `sync_state() pending` for gcc/rpmhpd/gpucc due to gmu/dispcc consumers) — related to the `regulator_ignore_unused` crutch still in the cmdline; proper rail ownership is a prerequisite. - Watchdogs must be suspend-aware (APSS watchdog + our boot/reachability layers) or a sleeping tablet gets rebooted. - Measure baseline idle power first (battery fuel gauge readable?) so improvements are quantifiable.
kit commented 2026-08-25 10:51:06 +00:00
Author
Owner
Copy link

Correction: the sync_state prerequisite in the description is mostly a non-issue

The description lists "RPMh sleep votes need the sync_state situation resolved (chronic sync_state() pending for gcc/rpmhpd/gpucc due to gmu/dispcc consumers)" as a prerequisite. Investigated 2026-08-25 — it is much smaller than it looks, and two of the three parts are not real.

gcc and gpucc have no sync_state callback at all. gcc-sm8150.c and gpucc-sm8150.c define plain platform_drivers with only .probe; nothing in drivers/clk/qcom/ sets .sync_state except the ipq parts. So there is nothing being skipped for those two — the warning is noise.

Unused clocks and power domains are already being cleaned up, just by a different mechanism than sync_state. Both run normally at 0.65 s:

[    0.652567] clk: Disabling unused clocks
[    0.652840] PM: genpd: Disabling unused power domains

That is clk_disable_unused / genpd_power_off_unused, which are global late_initcalls and unrelated to sync_state. (Note the current debug image passes clk_ignore_unused pd_ignore_unused, which suppresses them — but that is a deliberate bring-up flag, not a defect.)

Only rpmhpd genuinely has a sync_state, and we deliberately stubbed it out — commit 11c730d573b7, with the reason recorded in the code:

/*
 * gts6l BRING-UP HACK - remove when real consumers exist. With this
 * minimal DT "all consumers probed" arrives while the display scanout
 * and most of the SoC still run on XBL's boot votes; releasing them
 * here collapses the corners mid-boot (crash right after the last
 * initcalls, matching the 6.14-era "dies after Btrfs loaded" note).
 */

So RPMh power-domain corners staying at XBL's boot values is our own choice, made to avoid a specific crash — not an accident of fw_devlink.

Why it never fires anyway: the pending consumers are 2c6a000.gmu and ad00000.clock-controller. Both are permanently unbound — 0xad00000 is camcc, which we have no driver for, and the a6xx GPU driver uses the GMU node directly without binding a driver to it. Neither will ever probe, so sync_state can never run regardless of the stub.

What this means for this issue

The real prerequisite is narrower than written: rpmhpd needs to be able to apply real corners, which means un-stubbing rpmhpd_sync_state() and having enough real consumers in the DT that releasing XBL's votes does not collapse the SoC. fw_devlink.sync_state=timeout would force it to fire, and is a cheap way to find out what breaks — worth trying once the display situation (#17) is settled, since the stub's comment blames the display scanout specifically.

regulator_ignore_unused in the cmdline is a separate crutch and still stands as written.

Also relevant to the baseline-power measurement in the description: the fuel gauge (#9) is a pure SM5705 I2C port — pm8150b is not fitted on this board, so there is no PMIC ADC route to battery current either.

## Correction: the `sync_state` prerequisite in the description is mostly a non-issue The description lists *"RPMh sleep votes need the `sync_state` situation resolved (chronic `sync_state() pending` for gcc/rpmhpd/gpucc due to gmu/dispcc consumers)"* as a prerequisite. Investigated 2026-08-25 — it is much smaller than it looks, and two of the three parts are not real. **gcc and gpucc have no `sync_state` callback at all.** `gcc-sm8150.c` and `gpucc-sm8150.c` define plain `platform_driver`s with only `.probe`; nothing in `drivers/clk/qcom/` sets `.sync_state` except the ipq parts. So there is nothing being skipped for those two — the warning is noise. **Unused clocks and power domains are already being cleaned up**, just by a different mechanism than sync_state. Both run normally at 0.65 s: ``` [ 0.652567] clk: Disabling unused clocks [ 0.652840] PM: genpd: Disabling unused power domains ``` That is `clk_disable_unused` / `genpd_power_off_unused`, which are global late_initcalls and unrelated to `sync_state`. (Note the current debug image passes `clk_ignore_unused pd_ignore_unused`, which suppresses them — but that is a deliberate bring-up flag, not a defect.) **Only `rpmhpd` genuinely has a `sync_state`, and we deliberately stubbed it out** — commit `11c730d573b7`, with the reason recorded in the code: ```c /* * gts6l BRING-UP HACK - remove when real consumers exist. With this * minimal DT "all consumers probed" arrives while the display scanout * and most of the SoC still run on XBL's boot votes; releasing them * here collapses the corners mid-boot (crash right after the last * initcalls, matching the 6.14-era "dies after Btrfs loaded" note). */ ``` So RPMh power-domain corners staying at XBL's boot values is our own choice, made to avoid a specific crash — not an accident of fw_devlink. **Why it never fires anyway:** the pending consumers are `2c6a000.gmu` and `ad00000.clock-controller`. Both are permanently unbound — `0xad00000` is **camcc**, which we have no driver for, and the a6xx GPU driver uses the GMU node directly without binding a driver to it. Neither will ever probe, so `sync_state` can never run regardless of the stub. ### What this means for this issue The real prerequisite is narrower than written: **rpmhpd needs to be able to apply real corners**, which means un-stubbing `rpmhpd_sync_state()` *and* having enough real consumers in the DT that releasing XBL's votes does not collapse the SoC. `fw_devlink.sync_state=timeout` would force it to fire, and is a cheap way to find out what breaks — worth trying once the display situation (#17) is settled, since the stub's comment blames the display scanout specifically. `regulator_ignore_unused` in the cmdline is a separate crutch and still stands as written. Also relevant to the baseline-power measurement in the description: the fuel gauge (#9) is a pure SM5705 I2C port — pm8150b is not fitted on this board, so there is no PMIC ADC route to battery current either.
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
gts6l/6.17
No results found.
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
kit/linux-gts6l#3
No description provided.