Repository navigation
Conversation
Record the baseline (311 mA awake with the screen off, 302 mA asleep, cable out), the proof that the CPUs sleep, the radio elimination (nothing), the MIF-in-sleep saving (60-75 mA, module loaded at boot by a hand unit) and the coulomb-counter method. Note that a direct rtcwake freeze failed to wake again, so every sleep test goes through systemctl suspend; add that to the agent gotchas. Retire the 2026-10-27 140 mA figure.
# Conflicts: # device/tegu/NOTES.md
|
Thanks for this. It's a clear S1, and retiring a figure that didn't hold up is exactly what the notes should do. The 302-against-311 mA result is a useful answer: the floor is below the CPUs, so MIF, SICD and the rails are the right next targets. I checked the systemd count against the 10-27 record, and it matches. A few things before I merge, all docs:
Smaller things: "What's measured" and S1's Expect still say about 140 mA. The AGENTS line says two direct freezes out of about six, while the plan's numbers give two of five. And the hand unit that loads #100 merged today, so this now needs a small hand-merge with |
# Conflicts: # AGENTS.md # device/tegu/NOTES.md
… rule points at systemd S1 gets the buck-off, dead-man, read-back and coulomb-counter steps with Expect, Stop and Rollback; the 140 mA expectation becomes the S1 figure; the hand unit for tegu-mif-sleep names its commit and rollback; the direct-freeze count reads two of five and says the Wi-Fi hook may be the real difference. AGENTS's NFC clause no longer allows a direct write, and power-ab runs with --suspend systemd on the 9a.
|
All in at 510a4e6, on top of a merge of main (AGENTS keeps every bullet, main's NFC line first and the new rule after it; NOTES keeps the new S1 entry at the top, then the sections #131, #125 and #100 added). Mergeable again.
|
…o the rule is attendance, not the entry path The 2026-10-30 cable-out run with tegu-mif-sleep loaded never woke from its 30-minute systemctl suspend. Three no-wakes now, on both entry paths, with and without the module: the plan tallies them, the hand unit is disabled, and the agent rule asks for attended sleeps and no boot-time module until the hang is understood.
|
A correction at b265200, from this morning: a 30-minute |
… factor is still open
|
One more line at 485be76: the cheap test ran. A cable-in 30-minute systemctl suspend with tegu-mif-sleep loaded, alarm logged as armed, woke on the RTC to the second (IRQ 56, the PMIC). So neither the module nor the 30-minute systemd sleep is enough on its own to reproduce the no-wake; the cable being out is the one difference left to this morning's failed run, and that's the next attended probe. |
|
The cable-out figure with the module is in at 97858da: the run repeated this afternoon with tegu-mif-sleep loaded woke on its alarm, 273 mA awake with the screen off and 259.7 mA asleep (130.0 mAh of charge_now over 1802 s), about 40 mA under yesterday's 302 mA without the module, in line with the coulomb-counter estimate for a gauge this coarse. It also means none of module, 30-minute systemd sleep or cable-out reproduces the no-wake by itself; sleeps stay attended and the boot unit stays off. |
…leep, 260 mA The repeated cable-out run with tegu-mif-sleep loaded woke on its alarm: 273 mA awake with the screen off, 260 mA asleep, about 40 mA below the previous day's figure without the module. None of the suspected factors reproduces the no-wake alone.
|
Thank you, everything from this morning is in, and the method block is exactly what I hoped for: S2 can now reuse it as written. And thanks for the two follow-ups the same day. Retracting "the systemd path always wakes" as soon as it didn't is what keeps these notes trustworthy. The correction reached SLEEP-PLAN's S1 and AGENTS, but not two older spots, so they now say the opposite: the new NOTES entry still says One question: does your palmtop have the |
… every spot The NOTES entry no longer says the systemd path has woken every time or that the module loads at boot; a 2026-10-30 bullet carries the cable-out no-wake, the disabled unit, the cable-in wake and the afternoon's 260 mA. The status line names the three no-wakes and attended sleeps, the S1 repeat and Next say attended, the AGENTS count is seven longer systemd wakes, the nightly auto-sleep without the module is said to be fine, and the NFC hook is in the repository since derekross#152.
|
All in at 80e6659, point by point:
Mergeable against main; the NOTES top is yours at merge, as you said. |
…p since the shell reinstall
|
A correction to my toggle answer at c8d10bf: checking the palmtop just now, omamobile-sleep is off (omarchy-toggle-enabled returns 1), since this afternoon's shell reinstall shipped it off; so the palmtop isn't sleeping unattended at present, and the plan and the AGENTS line now say that, with the on-period record (systemd-path sleeps without the module all woke) kept and turning it back on left to the owner. |
|
One more data point at 6a3a489, from the #183 proof run: cable in, tegu-mif-sleep still loaded from the drain run, a forced 10-minute systemd sleep; the owner's power-key press and tap did not wake it, the kernel came up once at about 112 s and went back to s2idle, and only the RTC alarm ended it. The power key woke s2idle on 10-27 without the module, so the next cheap test is a key press during a sleep without the module and then with it. A key wake that doesn't register would look just like an alarm that doesn't, which would tie the module to the three no-wakes after all. |
# Conflicts: # AGENTS.md
|
Merged main after #137: its NFC line is taken as merged (my small clause edit on it dropped, the new sleep rule follows it), so this is mergeable again. |
…nd the root went read-only The forced 10-minute sleep on 2026-10-30 woke on its alarm but the UFS link never came back (link startup failed, Host init failed, ufshcd_wl_resume failed), every disk read then failed and ext4 went emergency read-only; a sysrq reset recovered it. A palmtop that seems never to wake may have woken with dead storage, so the plan and the agent rules now say to check dmesg for the UFS failure after every sleep and reset from the kernel if it shows.
|
The afternoon's second data point turned into the finding of the day, at f3047ae. The forced 10-minute sleep did wake on its RTC alarm, but UFS did not: pstore's console has So a "never wakes, power key dead" palmtop may be awake with dead storage, and the three no-wake records can't tell the two apart (none had a console). It's 1 failure in about 50 systemd-path resumes after kernel #83's 32 of 32. The plan and the AGENTS line now say: after every s2idle check dmesg for the UFS failure first, reset from the kernel if it shows, and keep unattended sleep off. The AGENTS NFC/sleep lines you'll hand-merge got one more bullet from this. |
… USB bus on the sixth
|
Pushed tonight's cycles onto this branch. Attended 30-second RTC-woken s2idle cycles with the cable in, on the daily boot with the modem never loaded: cycles 1 to 5 all woke with The sixth cycle of that boot didn't come back, and this time the palmtop left the USB bus entirely, so it isn't the dead-storage shape of the afternoon (which kept answering SSH from memory while nothing on disk could be read). It needs a forced restart before anything can be read; the console should be in pstore afterwards. Two things that follow. The afternoon failure was also late in a long boot (uptime 10801 s) and tonight's hang came after five sleeps, so the next question is whether the failure follows the count of sleeps in a boot rather than any single sleep. And the three dark no-wakes left no ext4 recovery line in their next boot's journal (checked across 25 boots tonight: no boot recovered a journal), so they weren't dead storage either, which fits a hang like this one. Also in this push: the open-questions list now names the sixth-sleep question and what to run next, which is the cycles again from a fresh boot with the kmsg stream to the laptop and the sleeps counted. |
…, and the ACPM link is withdrawn
|
A correction to my comment above, from reading the afternoon's own console rather than my memory of it. The failing suspend was the third of that boot (entries at 387.7 s, 4746.9 s and 10798.6 s; the third one's resume failed), not a late one, and the 25 seconds before it hold nothing but PAM and audit lines from an SSH login: no ACPM timeout, no thermal error, no UFS warning. I had tied that failure to ACPM timeouts; those are from a different boot of the same day and the link is withdrawn. So the sleep-count idea doesn't fit either: the afternoon was the third sleep of its boot at about 3 h of uptime, tonight's hang the sixth at about 10 minutes. All that both share is that neither was the first sleep of its boot. The plan now says this. |
…es but cannot be the policy
|
One more from reading the running device tree, which matters for S2 now that a GNSS session works: the three GNSS rails and the fingerprint rail are declared with labels and no consumer at all, so nothing names them as a supply and that is exactly why they carry regulator-always-on. Dropping that property therefore doesn't move the cost to whoever uses the rail; it just lets the core's unused-rail cleanup switch all four off about 30 s after late init. Still a valid one-boot measurement and still worth doing, but it can't be the shipped policy: GNSS needs its three for a session and the sensor needs LDO15S for an operation. The policy S2 should lead to is a consumer, each path claiming its rails while in use, which needs a supply property and driver support on both sides. So the order is: measure, then decide whether the number justifies that work. The plan says this now. |
…E set, so it is link startup
|
Lined the three resumes of that boot up on their suspend entry, and the failure is narrower than "UFS didn't resume".
So it is link startup itself: it never reached the first DME set, each attempt took about 430-490 ms (the UIC timeout) and five in a row used up That points the next run at the link-startup path, which at |
… the cheap test for it
|
Added one more open question with its answer half-done: the record says we have never read which PMIC event ends an early RTC cycle, and the driver names the register. Noticed while reading kernel #149, which fixes the same readback for gs101's S2MPG10. Worth saying for anyone following that PR: zumapro's and gs201's cases are already correct in our tree, so #149 doesn't change what the 9a or the 7 Pro report. |
|
The failing Tests job here is the paused-tick check in test/owl-idle, not anything in this branch: it is false for one second in every ten by construction. #207 fixes it, and this should go green once that is in. |
… the log through the freeze
|
Two things from this morning, pushed. Ten of ten 30-second RTC-woken cycles on the daily boot woke with the storage counters, the ACPM timeout count and the watchdog count all zero, and More useful than the count: the reason every earlier hang left no kernel evidence is that printk suspends the consoles at the freeze, so nothing reached ramoops past that line. It needs no boot argument, |
|
Thanks for the steady stream of runs, and especially for reading your own console back and withdrawing the ACPM link, and for the console_suspend find. Before I merge, could you do one consistency pass?
There's also a doubled "the shell's the shell's" in AGENTS. |
|
Consistency pass done and pushed, on today's main.
The doubled "the shell's" I could not find after merging main; if it is still there in a line you see, point me at it and I'll fix it. |
Ten 30-second s2idle cycles with idle clock gating off all woke with no UFS errors. The run with gating on was also ten of ten, so the hypothesis is neither confirmed nor ruled out yet.
|
Thank you, this is a thorough pass, and console_suspend=N is a real find. I'll fix the last bit at merge: the AGENTS sleep line still says "three" and "both direct freezes", from before the sixth-cycle hang. The doubled phrase was "The shell's" with a capital T. |
|
Landed on main authored by you (the PR shows as closed because the commit went in directly). |
…s on the 7 Pro (#254) Kernel #153 (husky touch), #166 and #133 and omarchy-mobile #149 (panther touch) merged on 2026-10-31 and 11-01, and omamobile-audio 0.5.8 raised the 7 Pro's default speaker gains under verified protection. The website's device pages are built from this file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…oor and sleep tests stay attended (derekross#166) S1-S3 of the Pixel 9a's sleep plan: the drain is the SoC floor, the sleeps that didn't wake and the UFS resume failure are recorded with what pstore held and how each cycle entered sleep, the CLKGATE test is a runbook step, and sleep tests stay attended. Maintainer edit: the AGENTS sleep line counts four no-wakes, adds the sixth-cycle systemctl suspend hang, and drops the doubled phrase. Closes derekross#166 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s on the 7 Pro (derekross#254) Kernel derekross#153 (husky touch), derekross#166 and derekross#133 and omarchy-mobile derekross#149 (panther touch) merged on 2026-10-31 and 11-01, and omamobile-audio 0.5.8 raised the 7 Pro's default speaker gains under verified protection. The website's device pages are built from this file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
S1 of the sleep plan (#144) ran tonight on the palmtop. The cable-out baseline is 311 mA awake with the screen off and 302 mA asleep (two 30-minute systemctl suspend rounds through power-ab --suspend systemd, #157): sleep saves nothing over the awake screen-off state. A 60 s test shows the CPUs do sleep (timekeeping suspended for the whole minute, one deep-state entry per CPU), and pulling the radios apart with the cable in and the charger's buck off shows Bluetooth, NFC and Wi-Fi cost nothing measurable. What stays on is MIF, the buses and the always-on rails; our kernel has no system idle state for this SoC. MIF at 421 MHz in sleep (tegu-mif-sleep.ko) saves 60-75 mA by the gauge's coulomb counter, matching the 10-28 figure, and the palmtop loads it at boot from a hand unit until the policy lands properly.
Two corrections: the 2026-10-27 "140 mA asleep" figure doesn't hold up against 10-28 (434-445 mA as-is, cable out) or tonight and is retired; and a direct
rtcwake -m freezefailed to wake for the second time in two days while everysystemctl suspendcycle has woken, so the plan, the notes and the agent gotchas now say to sleep only through systemd. Next is S2 (the four rails) and a cable-out power-ab run with the module loaded.Docs only: SLEEP-PLAN, NOTES and AGENTS.md. tools/pr-check passes.