Skip to content
Closed
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
121 changes: 43 additions & 78 deletions device/tegu/SLEEP-PLAN.md
Original file line number Diff line number Diff line change
Expand Up @@ -451,7 +451,7 @@ with the cable in.
**A watchdog through s2idle, tried 2026-11-01 (same release): it breaks
the sleep, so it can't catch the hang.** `s3c2410_wdt` stops both
watchdogs in its suspend callback, so a hang in s2idle never resets the
palmtop. A bring-up module (`tegu-wdt-s2idle`, in the owner's logs)
palmtop. A bring-up module (`tegu-wdt-s2idle`, source in [e761ba793](https://github.com/W4HO-ham/linux-omarchy-mobile/commit/e761ba793) on the owner's kernel fork, not for merge)
restarted the cluster 0 watchdog in the noirq phase with the PMU reset
path unmasked and the counter enabled, as the driver's own enable does,
and stopped it in resume_noirq. Armed, three sleeps of three ended in a
Expand All @@ -467,83 +467,48 @@ with the cable in.
stops answering after resume looks like, which matches the stall seen
in G3-L1's CPU-sleep run (MODEM-PLAN).

**The sleeps that woke were short: the PMIC wakes s2idle about a second
in (2026-11-01, same release).** In the disarmed sleeps above,
`PM: suspend exit` came 1.5 and 1.6 s after `PM: suspend entry`, not
at the 30 s RTC alarm, with `/sys/power/pm_wakeup_irq` 56, the
S2MPG14's own line (GICv3 106, `s2mpg14-pmic`). One more 30 s sleep
with `pm_debug_messages` on: "Triggering wakeup from IRQ 56" 0.6 s
after entry, exit 1.5 s after it. Across that sleep and the 40 s after
it, the PMIC line counted 2 and `rtc-alarm0`, its only mapped child,
1, which is the real alarm going off after the wake; awake, the same
`rtcwake -m no -s 30` alarm fired at 30 s as it should. So another PMIC
source, one without a mapped child, raises the line as the palmtop
goes to sleep. Earlier cycle logs only recorded wake IRQ 56 and the
time to SSH, which can't tell a short sleep from a full one, so S1's
wake counts may describe sleeps of about a second. Next: read the
S2MPG14 interrupt registers (INT1-INT5 through the ACPM regmap, from
the kernel, never from userspace) at that wake, and check the 30-minute
sleeps' records for their real length.

A bring-up module (`tegu-pmic-wakeread`) then read INT1-INT5 and their
masks through the PMIC's regmap in suspend_noirq and resume_noirq
(reads only; INT1-INT5 clear on read). Three sleeps, all woken about a
second in:

| Alarm set | INT1-5 at sleep | INT1-5 at wake | Wake IRQ |
| --- | --- | --- | --- |
| 30 s | 00 09 00 00 00 | 00 0c 00 00 00 | 56 (PMIC) |
| 120 s | 00 09 00 00 00 | 00 0d 00 00 00 | 56 (PMIC) |
| none (`rtcwake -m disable`) | 00 08 00 00 00 | 02 09 00 00 00 | 59 (Power key, gpa4-0) |

The masks read `ff fb ff ff ff` each time, so only INT2 bit 2, alarm 0,
is unmasked. With an alarm set, alarm 0 latched within 60 ms of the
suspend_noirq read, whether it was due in 30 or 120 s; with none set,
INT1 bit 1 (power key released) latched and the power key's GPIO woke
the SoC, with nobody touching the key. So something in the last stage
of the suspend makes the S2MPG14 raise events that didn't happen, and
every sleep since at least 2026-10-30 has probably lasted about a
second. The 1 s and 60 s RTC ticks (INT2 bits 3 and 0) were latched
but masked throughout. After the no-alarm sleep, SSH over USB took
about 8 minutes to come back, with nothing in the kernel log about
USB.

**With CPU core power-down off, the sleep lasted its full length.**
The only deeper cpuidle state on each CPU is state1
(`cpu-a520-sleep`, `cpu-a720-sleep`, `cpu-x4-sleep`, PSCI core
power-down; each was used 4-5 times by s2idle in the short sleeps).
With state1 disabled on all eight CPUs (WFI only) and a 30 s alarm,
`PM: suspend entry` at 472.7 s and the wake at 501.9 s, on IRQ 56 with
INT2 `0c` (alarm 0): the real alarm, 29 s in. So the events that end
every other sleep a second in come with the CPUs' power-down during
s2idle, most likely from the firmware's side of that power-down, not
from the PMIC drivers. One sleep so far; next: a series of longer
sleeps with state1 off (a sleep hook can turn it off before the sleep
and on after), their hang rate, and the drain against the short-sleep
figures.

Our idle states match stock's: the same three `c2` states with
`arm,psci-suspend-param = <0x10000>` (core power-down) and stock's
latencies. What stock has besides them is Samsung's `exynos-cpupm`
node: cluster power-down per cluster (`cpd_cl0` for CPUs 0-3, `cpd_cl1`
for 4-6, `cpd_cl2` for 7) and `sicd` across all eight (system idle
clock down), which its vendor driver coordinates with the APM. Mainline
has no such driver, so when every CPU sits in core power-down in s2idle,
the firmware sees a state stock's kernel would have announced to it
first. That the PMIC's false events come from there is an inference,
not a measurement. The device package's fix for now is a sleep hook
that allows only WFI for the length of a sleep (`omamobile-device-tegu`
0.2.5, #255).

**Ten sleeps with the hook (2026-11-01, same release, device-tegu 0.2.5
installed from its branch).** `systemctl suspend` with a 60 s RTC alarm,
ten times: nine slept 55-61 s, entry to exit, and woke on IRQ 56 (the
alarm); the seventh's resume took longer than the 8 s the script waited
before measuring, so its length isn't known, but it woke and the hook
restored the states. `suspend_stats` 6 to 15 successes, 0 failures; no
UFS errors, no ACPM timeouts. The screen stayed lit through these sleeps
(`systemctl suspend` skips the shell's blanking), so no drain figure
from this run.
**Correction (2026-11-01, same day): the sleeps were not short.** The
paragraphs that stood here said the PMIC woke every s2idle about a
second in, and that disabling CPU core power-down fixed it (a hook was
proposed as device-tegu 0.2.5, #255, now closed). That was a
measurement error: the length of each sleep was taken from
`PM: suspend entry` to `PM: suspend exit` in dmesg, and the kernel log's
clock does not advance while the SoC is suspended. The RTC shows the
real length. After a sleep with a 120 s alarm, `since_epoch` read 4 s
past the alarm, so the palmtop slept the whole 120 s and woke on it.
The PMIC registers read at each wake agree once read that way: INT2
bit 2 (alarm 0) for the alarm wakes, and INT1 bit 1 (power key
released) for the one sleep with no alarm, which the owner ended by
pressing the key minutes later. The journal's wall-clock times give
28-30 s for the 30 s cycles of 10-31. With core power-down disabled the
dmesg deltas grew to the alarm time only because with WFI alone the
s2idle loop doesn't freeze the tick, so timekeeping kept running: a
shallower sleep, not a fixed one. So S1's wake counts (and AGENTS' 9a
gotcha) stand as written, and the same error affects the watchdog
paragraph above: "within 60 ms of arming" and "after 1.5 and 1.6 s" are
dmesg deltas, not the sleeps' real lengths. The rest of that paragraph
(ACPM stopping after the wake with the watchdog armed, three of three)
stands. To measure a sleep, read the RTC
(`/sys/class/rtc/rtc0/since_epoch`) before and after, not dmesg (this
kernel has no `suspend_stats/last_hw_sleep`). Twenty more 120 s sleeps the same day with core power-down off
all woke, but they don't speak to the hang with core power-down on.

**Thirty 120 s sleeps with core power-down on (2026-11-01, same release,
device-tegu 0.2.4).** Screen off through Hyprland's DPMS before each
`systemctl suspend`, `printk.console_suspend=N`, NFC polling stopped by
its hook, Wi-Fi kept: thirty of thirty woke on the PMIC line (IRQ 56,
the S2MPG14's, which the RTC alarm shares with other PMIC events).
Measured by the RTC (`since_epoch` before and after, less about 15 s
of the cycle that isn't sleep), 27 lasted 118-123 s, the alarm's time;
one lasted 106 s, one 112 s and one 64 s, too early for the 120 s alarm.
The series recorded only the wake IRQ, not the `rtc-alarm0` count or
INT2, so what woke those three isn't known: count them as early PMIC
wakes, as on the 7 Pro, not as alarm wakes. The dmesg delta read 1.5-1.6 s
every time, so the correction above holds. `suspend_stats` 36 to 66
successes, 0 failures; no UFS errors and no ACPM timeouts. At the
earlier rate of about one hang in 21, thirty clean sleeps would still
happen by chance about one time in four, so this narrows the hang but
doesn't retire it.
- **What the five clean cycles rule out, and what they don't.** Five in a
row with the storage counters at zero says the UFS resume path is not
broken per se; the afternoon failure and this hang are both the sixth or
Expand Down