diff --git a/AGENTS.md b/AGENTS.md index 77f530d74..5e4034c7b 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -115,6 +115,7 @@ Each entry gives the gotcha, then the reason or the fix. Details are in the link - **ABL ignores the `vendor_boot` cmdline:** put kernel arguments in `/chosen/bootargs`. ABL appends `console=null`, so a panic is invisible: use `panic=10`. [BRINGUP-M1](device/cheetah/BRINGUP-M1.md) - **ABL leaves both cluster watchdogs running:** the phone resets at about 60 s of uptime unless `s3c2410_wdt` takes them over at probe. - **ABL picks the base DTB by `soc_id`:** `bin/omamobile images` ships both silicon revisions. It also needs an empty `/ect` node and `/chosen/bootargs`, and applies a DPM overlay that needs label stubs (build with `-@`). +- **Pixel 9a: a fallback to slot A costs the Linux root.** `fastboot --set-active=b` (which `bin/omamobile flash` runs) clears B's successful flag and gives B three ABL tries; every boot uses one, and the initramfs USB safety net (no USB host in about 65 s: panic) uses them fast. At zero ABL boots stock on slot A, and stock overwrote `userdata` on 10-28, and most likely on 10-26, before anyone could choose anything in recovery. So: `flash` only a known-good image set, and on the 9a ask the maintainer before any `flash`, because it unmarks B. Marking B successful is a `devinfo` write, and on the 9a it is on hold until the maintainer decides how it's done: don't mark B, and don't run `omamobile-bootctl` or anything that makes `devinfo` writable. With B marked, try new images with `push`, which doesn't call `--set-active` and leaves the mark alone; ABL then no longer falls back, so a broken pushed image means fastboot by the buttons. With B unmarked, every boot spends one of the three tries: keep reboots few and don't test new images. [tegu NOTES](device/tegu/NOTES.md#2026-10-26-and-2026-10-28-slot-a-fallbacks-that-cost-userdata); the operation approved on 2026-10-22, now on hold: [NOTES](device/tegu/NOTES.md#approved-b-success-operation-after-layout-review-2026-10-22). - **`reboot -f` and busybox `reboot` hang:** use `reboot-to ` (`LINUX_REBOOT_CMD_RESTART2`). ABL often ignores the `bootloader` reason anyway. - **Reserved memory on gs201 is `#address-cells = <2>`, `#size-cells = <1>`,** so each `reg`/`alloc-ranges` entry is 12 bytes. A 16-byte value is silently ignored, and the pool lands anywhere. The display pool's `alignment` must be two cells too, or `of_reserved_mem` rejects it. `/proc/device-tree` shows the raw property, so it *looks* right. [WIFI-BATTERY-PLAN](device/cheetah/WIFI-BATTERY-PLAN.md) 9.x, [NOTES](device/cheetah/NOTES.md) - **Check every stock `reserved-memory` region against `/proc/iomem`:** five were plain System RAM in our kernel, and firmware writes into them at runtime. diff --git a/device/tegu/MODEM-PLAN.md b/device/tegu/MODEM-PLAN.md index 59c880441..19d7460ff 100644 --- a/device/tegu/MODEM-PLAN.md +++ b/device/tegu/MODEM-PLAN.md @@ -514,9 +514,10 @@ startNmea 25; callbacks location 4, NMEA 5). POWER_OFF), Kepler still faulted on engine start; after six faults gnssd gave up ("chppPlatformNotifierWait in STATE_PERMANENT_FAILURE"). - Correction (2026-10-27, night): a stock reference can't come from this - palmtop. Arch is on `userdata`, so stock Android on slot A can't mount its - data and stops in recovery (HANDOFF, 10-26 slot A fallback). It needs a - second 9a, or a factory reset and a reinstall. Was: boot slot A once with + palmtop. Arch is on `userdata`, and stock Android on slot A overwrites + `userdata` before it reaches recovery + ([NOTES, 2026-10-26 and 2026-10-28](NOTES.md#2026-10-26-and-2026-10-28-slot-a-fallbacks-that-cost-userdata)), + so booting it loses the Linux root. It needs a second 9a. Was: boot slot A once with root and capture `logcat -b all` and dmesg while an app gets a fix (gnssd, the HAL, CPIF's gnss lines, the timemark device it opens), then compare; or decode the fault dump. Then the NMEA FIFO into omamobile-gnss-nmea, as cheetah. diff --git a/device/tegu/NOTES.md b/device/tegu/NOTES.md index 158f35618..53be7f2b6 100644 --- a/device/tegu/NOTES.md +++ b/device/tegu/NOTES.md @@ -35,6 +35,88 @@ resets) and are in [MODEM-PLAN](MODEM-PLAN.md) under "G3, twenty-fourth to thirty-fourth tries" and "The CP link hang". Not a GNSS matter: if a boot with nothing but Bluetooth starting resets again at that time, start here. +## 2026-10-26 and 2026-10-28: slot A fallbacks that cost userdata + +Both times slot B was active but **not marked successful**, so each boot +used up one ABL retry. After `fastboot --set-active=b` the 9a's ABL gives +B **three** tries (10-28: retry 3 after the flash, 1 after two boots, from +the devinfo dumps below). A boot that never reaches the USB host is crashed +by the initramfs safety net about 65 s in, so a bad image uses up the tries +within minutes, and ABL then boots stock on slot A. + +**2026-10-26.** The failing boot was a comparison run of Google's own +kernel (`6.1.134-android14-11-g2bae8aca899a-ab13800907`), written to slot B +by `omamobile flash` (so `--set-active=b`). It ran with a RAM-only +filesystem assembled from our initramfs files, with PID 1 replaced by a +static supervisor that opens no block device. Stock's own first stage didn't +run on slot B. It never brought up USB. +Fastboot (buttons) then showed A selected and B at retry 0. Stock's own +boot wasn't captured. Afterwards `userdata`'s ext4 and F2FS magics were +gone, with non-zero data in their place, `e2fsck -fn` failed, and `persist` +had changed. [I] Stock on slot A wrote over the start of `userdata`; nobody +saw its boot. At 16:35 MST the palmtop sat in stock recovery (USB 18d1:4ee7, +sideload-only adbd: every adb command answered `error: closed`). Nobody +chose anything in recovery. A fresh read-only backup of efs, efs_backup, +modem_userdata, persist, modem_a/b, devinfo and metadata was taken +(off-repo). Slot B came back by `omamobile flash` of the last good five +images (which set it active again). The root came back by streaming the +verified 10-25 recovery image over USB SSH into `userdata` from our +initramfs, with a readback hash and a read-only `e2fsck`. B was then on its +last try (byte 0x34 `00` in the "before" row below): one more unmarked +reboot would have fallen back again. + +**2026-10-28.** Kernel `7.3.0-rc6-omamobile-tegu-s2mpg15-m1-g991042e2eff3-dirty` +(the release the day's run kit records; the string stays the same across +that tree's builds, so it names the tree, not one build) with our initramfs +(`bin/omamobile initramfs`, `initramfs/init`, packed by the kit), and a new +DTB (the TCPC as a dual-role port with a link to dwc3's role switch). In the morning, slot B was +written by `omamobile flash`, recovering from a test boot of kernel #55's +head that hung on the Google logo, and it was **not marked** afterwards. +Every `push` round after that ran unmarked, so the tries ran out on the +first round that failed. That's the situation the rule is there to prevent. The DTB round lost USB on its boot (`push`: "no SSH after +reboot"), and a few minutes later the palmtop was on stock. adb answered +`error: closed` again, the same sideload-only adbd as 10-26. The owner +doesn't remember what the screen showed, and nobody tapped anything. +Afterwards `userdata` read as non-zero data with no filesystem magic (the +first 64 KiB looks random), and the root wouldn't mount. Fastboot by the +buttons, `omamobile flash` of the last good images, the 10-25 recovery +image streamed to `userdata` and read back against its sha256, then a boot. +Fresh read-only backup of the same eight partitions, taken before the boot: +efs, efs_backup, modem_userdata, persist, devinfo and metadata differed +from the 10-26 backup, modem_a/b didn't. `fastboot getvar slot-*` wasn't +captured either day, apart from `current-slot: a` on 10-28. + +So stock doesn't just stop at a recovery prompt: on both days `userdata` +was overwritten before anyone could choose anything. Treat a fallback to +slot A as losing the Linux root. + +**Marking B both days** was the approved operation below, `omamobile-bootctl +mark-successful` for B. The binary was built statically from +`packages/omamobile-device-cheetah/src/omamobile-bootctl.c` (source sha256 +`89d03460…`, binary `f44936d5…`), driven by a laptop script that: + +1. reads a fresh full `devinfo` snapshot and checks its sha256 against live + storage right before the write; +2. runs the tool's dry run, which must plan exactly one byte; +3. arms a 30 s timed read-only restore, writes, reads back all 8192 bytes + against the predicted image, and checks the read-only flags over a fresh + SSH login. + +systemd's 10 s runtime watchdog was running throughout (the bring-up checks +for it). The script didn't arm anything beyond that. + +The laptop keeps the command log. The 10-28 harness copy differed from the +10-26 one only in its baseline paths (slot A's bytes are read live). + +| Moment | Bytes 0x30–0x37 | +| --- | --- | +| 10-26, before | `02 1a 00 00 00 14 00 00` | +| 10-26, after | `02 1a 00 00 00 16 00 00` | +| 10-28, before | `02 1a 00 00 01 14 00 00` | +| 10-28, after | `02 1a 00 00 01 16 00 00` | + +Only byte 0x35 changed (bit 0x02), and slot A was untouched. + ## 2026-10-26: audio (speakers, microphones, A2DP) with omamobile-audio 0.5.0 Kernel `7.3.0-rc6-omamobile-tegu-s2mpg15-m1-g991042e2eff3-dirty`: the private