Skip to content
Merged
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
95 changes: 75 additions & 20 deletions device/lynx/WIFI-PLAN.md
Original file line number Diff line number Diff line change
@@ -1,18 +1,23 @@
<!-- SPDX-License-Identifier: MIT -->
# Pixel 7a (lynx): Wi-Fi and Bluetooth

Status (2026-10-27): **parked.** The chip enumerates and reads host memory,
Status (2026-10-28): **parked.** The chip enumerates and reads host memory,
but its boot ROM stops at its authentication stage on the first firmware
load, with the same unsigned image stock loads in 57 ms. Every host-side
difference two testers could act on has been tried one at a time, one
change per boot, and all were negative, including a stock-shaped power
cycle with an 8 s and a 30 s wait (step 6). Bluetooth is not required for
a good load. No run yet has had Google's host stack and full Android's wait
before the load together; that is the next step (section 7). A stock-side
read of the register the ROM consults is done (section 5). The published kernel has no PCIe node for
lynx; the Wi-Fi branch stays local, and on the test phone `pcie_gs201` and
`ath11k_pci` stay blacklisted (section 8). Bluetooth is on the same chip and
waits on the same question. Tracking issue: #36.
cycle with an 8 s and a 30 s wait (step 6) and, on 2026-10-27, Google's
host stack with Google's own trigger and full Android's 5 s wait before
the load (section 7, item 1): it stalls the same way, so that pair is
ruled out and the variants are stopped. Bluetooth is not required for a
good load. The one measurement left is the cold stock read (section 7,
item 3), on this phone, at its owner's call. A stock-side read of the
register the ROM consults is done (section 5). The published kernel has no
PCIe node for lynx; the Wi-Fi branch stays local, and on the test phone
`pcie_gs201` and `ath11k_pci` stay blacklisted (section 8). Bluetooth, on
the same chip, scans on mainline through its own UART (linux-omarchy-mobile
#98, in review); section 5 says what its device tree changes at the bench.
Tracking issue: #36.

Two phones contributed: this one (`latentoperator`, sections 3 to 6) and
FlorencioJP's (#36, 2026-10-26: the ROM dump, the iATU bisection and the
Expand Down Expand Up @@ -189,11 +194,19 @@ What it says:
first BHI on a chip that has never passed under Linux, so it is what the
register shows after power-up with WLAN_EN high and no BHI yet, not a
"good run finished" marker. The warm/cold reading built on it is
withdrawn. Two things stay open: at every Linux boot `gph2-5` reads
output low and stock logs `BUCK GPIO: [0]` at 2.5 s, so if that GPIO
really controls the rail the chip's core domain starts off on both
sides; and no rail has been measured, so the 0.5 s "buck drop" in step 6
is a GPIO drop. A cold read before the first-ever BHI still needs the
withdrawn. Two things stay open: on a kernel without the Bluetooth node
(linux-omarchy-mobile #98) `gph2-5` reads output low at every Linux boot,
and stock logs `BUCK GPIO: [0]` at 2.5 s, so if that GPIO really controls
the rail the chip's core domain starts off on both sides; and no rail
has been measured, so the 0.5 s "buck drop" in step 6 is a GPIO drop.
With the Bluetooth node, a regulator owns `gph2-5`: `hci_qca` raises it
for the firmware download a few seconds into boot, drops it when
Bluetooth goes down, and the controller boots unconfigured (no address),
so the buck goes up, down, and up again once the address service runs.
Before a Wi-Fi harness takes the pin on such a kernel, unbind the
Bluetooth serdev (`echo serial0-0 > /sys/bus/serial/drivers/hci_uart_qca/unbind`),
and record for every run whether its kernel had the Bluetooth node and
how the buck was dropped before the load. A cold read before the first-ever BHI still needs the
calibration load held back on a stock phone (vendor init edited, or a
build whose init does not write `fs_ready`).
- Two registers that read 0 on mainline after the stall are nonzero under
Expand Down Expand Up @@ -254,10 +267,11 @@ below.
lines between the probe and the second assert in a full-Android log,
to confirm the first load goes through the calibration handler. This
phone's two stock boots already show it (section 2). Then the stock-kernel
diagnostic boot again, same 52 modules, Bluetooth off, with one change:
write `1` to `cnss2`'s `fs_ready` attribute, as `init.lynx.rc` does,
instead of `powerup`. Nothing is mounted and the calibration result stays
in memory, so nothing is written to storage. Stop conditions as before:
diagnostic boot again, the 52 modules of the USB profile plus the 16 of
the CNSS profile (item 2), Bluetooth off, with one change: write `1` to
`cnss2`'s `fs_ready` attribute, as `init.lynx.rc` does, instead of
`powerup`. Nothing is mounted and the calibration result stays in
memory, so nothing is written to storage. Stop conditions as before:
the watchdog armed, no windowed reads before `mhi_sync_power_up()`, the
same stall timeout.
- **Pass:** the first good load outside full Android, on a kernel and
Expand All @@ -266,10 +280,42 @@ below.
step compares that boot with step 6 at the same points.
- **Stall:** full Android does something else before the first BHI that
the diagnostic doesn't. Go to item 2.
- **Done 2026-10-27, stalled** (FlorencioJP's phone, run 17, log
`stock-reference-last-console-17.log` on #36). Google's `boot.img`
kernel and `vendor_dlkm` modules, 68 modules, nothing mounted,
Bluetooth off (`btpower` loaded, rfkill soft-blocked, BT_EN low),
the watchdog armed, nothing read before `mhi_sync_power_up()`.
Probe: `BUCK GPIO: [0]` / `drive it HIGH` 5.584 s, `cannot receive
L23_READY` 5.781 s. `fs_ready` written at 10.800 s, state
`IN_COLD_BOOT_CAL | PCI PROBE DONE` (`0x400008`), the handler's 5 s
wait, then `BUCK GPIO: [1]` 15.844 s (WLAN_EN low for 10 s, the buck
held), `Starting image download via BHI` 16.001 s, `BHI_EXECENV 0`,
`BHI_STATUS 0`, `BHI_ERRDBG1 0xfad00000`, `BHI_ERRDBG3 0x80000` on
every dump from 23.0 s to 73.2 s, `MHI did not load image over BHI,
ret: -110` at 76.772 s. So Google's host stack, Google's trigger and
the 5 s wait together are not the pair.
2. **What full Android does to the chip between the probe and the first
BHI.** Bluetooth was the first candidate and is ruled out (section 2).
Next, read-only: the diagnostic's 52 modules against what full Android
has loaded by the first BHI.
Read-only background, no boots spent on it (FlorencioJP, #36
2026-10-28): full Android's second-stage `init` logs 177 `Loaded kernel
module` lines before its first BHI (10.404 s in boot trace 1); 142 of
them are not in the diagnostic and 33 of the diagnostic's are not in
that list. The 33 include `cnss2`, `mhi`, `qrtr` and
`pinctrl-exynos-gs`, which first-stage init loads without a `Loaded
kernel module` line, so the Android side of that comparison is
incomplete. The 142 are mostly display, audio, charging, Trusty, AoC,
thermal and debug; the ones that touch power or buses before 10 s and
that the diagnostic lacks are `cpif`/`google_modemctl` (the modem on
PCIe0), `google_bcl`, `exynos-acme`, `gs_thermal`, `bts`,
`slc_acpm`/`slc_pt`, `s2mpg12-powermeter`, `gsa`/`gsa_gsc`,
`trusty-*` and `sjtag-driver`; no causality is read into any of them.
The diagnostic's own set: the USB profile's 52 modules, plus the CNSS
profile's 16 in load order (`exynos-dm`, `exynos_devfreq`, `rfkill`,
`btpower`, `qrtr`, `qmi_helpers`, `wlan_firmware_service`,
`cnss_plat_ipc_qmi_svc`, `acpm_flexpmu_dbg`, `exynos-pm`,
`pcie-exynos-gs201-rc-cal`, `exynos-pcie-iommu`, `itmon`,
`pcie-exynos-gs`, `mhi`, `cnss2`); run 9 (2026-10-26) had the same
list without `rfkill` and `btpower`.
3. **The cold stock read** (the maintainer's step 2 on #36, one host GPIO
write). Turn Bluetooth off first: `btpower` drives `gph2-5` too. On
stock, after `dev_boot shutdown` and `off`, drive `gph2-5` (GPIO 103) low
Expand All @@ -296,7 +342,7 @@ below.
| `gph2-0` | PERST | input, low, pull-down; sleep input / down | output, no pull |
| `gph2-1` | CLKREQ | input, high, no pull; sleep input / down | function 2, pull-up |
| `gph2-4` | WLAN_EN | input, low, pull-down (FlorencioJP); here already output, low, no pull, claimed by `regulator-wlan-en` | output, pull-up active, pull-down asleep |
| `gph2-5` | buck | output, low, no pull | output, pull-down |
| `gph2-5` | buck | output, low, no pull (kernel without the Bluetooth node; section 5) | output, pull-down |
| `gpa6-1` | host-wake | input, high, pull-down | EINT, pull-down |

If a chip strap shares WLAN_EN or PERST, the pull-down is what the
Expand Down Expand Up @@ -324,6 +370,15 @@ below.
`/usr/lib/firmware/updates/ath11k/WCN6855/hw2.1`: linux-firmware
`amss.bin` and `m3.bin`, plus the phone's own calibration as
`board.bin`. Nothing from the vendor image goes into this repository.
- **Console capture on a stock-kernel diagnostic boot:** the ramoops
console (`/sys/fs/pstore`, collected into
`/var/lib/systemd/pstore/console-ramoops-0` on the next Linux boot)
survives only a **warm** return. A cold return loses it: the first two
`fs_ready` runs came back with an empty pstore, and so does a plain cold
`systemctl reboot` of mainline. The diagnostic's init writes `warm` to
`/sys/kernel/reboot/mode` before `reboot(bootloader)`; do the same
anywhere a stock-kernel console is wanted this way (FlorencioJP, #36
2026-10-28; applies to both phones).
- **To resume:** read sections 6 and 7, re-check the live kernel, DT and
the six protected hashes, then run the next item in section 7 as a gated
step (expect, stop condition, rollback) and record the result here with
Expand Down