Repository navigation
Conversation
The CHPP link lost bits because spi-s3c64xx's flush cleared the packet count before turning the channels off (linux-omarchy-mobile#150) and because the GNSS SPI's pins ran on the bootloader's drive 0 (derekross#151); with stock's SPI configuration the receiver gets a fix. Holding the CP link up after ONLINE ends the link hang under GNSS traffic. Adds the plan section with the sweep that found it, three agent gotchas, a notes entry, the GNSS SPI bring-up overlays, the session script, and the start script's new module settings.
…s a minute per boot
|
Correction, pushed as a second commit: the "clean 3- and 10-minute sessions" were not. With cpif's keep_link_up set after ONLINE, the CP's PHONE_ACTIVE dropped 103-108 s after ONLINE in all six boots that held the link tonight (ONLINE -> CRASH_EXIT, the CP's own crash, CHPP healthy), gnssd went into crash handling, and the fixes stopped about a minute in; my fixed-of-total counting hid it. Left to cycle, the link hang of 10-29 comes back about 170 s after ONLINE. So the GNSS fix stands (SPI flush order, pins), the link hold is a diagnostic only, and the CP link under GNSS traffic is the open item (G3-L1). The plan, the notes entry and the AGENTS rule now say that; the session script holds the link only with HOLD=1. |
…ocation switch omamobile-gnss.service loads a GNSS chip driver and runs the gnss vendor profile, which is right for the 7 Pro's BCM4775x. On the 9a the receiver is the modem's GNSS core, so the stack is the modem with GNSS plus the gnss-slsi profile. These are the two drop-ins and the start script the palmtop runs, so the switch starts the receiver and GeoClue's clients get its NMEA.
…t didn't stop the hang
|
Pushed two more commits: the hand units that put the 9a's GNSS behind Omarchy's Location switch (device/tegu/bringup/location: two drop-ins so omamobile-gnss.service runs the modem's GNSS core plus the gnss-slsi profile instead of the 7 Pro's chip driver, and the start script), and the plan's record of them. The switch works: a client's connect on omamobile-gnss-nmea.socket starts the stack, the HAL serves NMEA on the socket and the fixes arrive (first fix about 80 s after the connect, 46 of 52 GGA fixed). Two things to know: on the 9a the package's units are masked by the boot command line (systemd.mask=omamobile-gnss.service in package-startup.dtso), so the switch needs that mask dropped; and a session still ends when the CP link does, which is the open item. Also recorded there: the doorbell-under-the-link-lock fix I built tonight (kernel branch tegu-cpif-doorbell) did not stop the hang, so that access is ruled out. |
|
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. |
…L1.1 alone does not help
… TPowerOn was my wrong guess
…runs measured nothing
|
This is a great result. The flush-order and pins findings are exactly the kind of root cause I hoped for, and thank you for correcting the "clean sessions" within the hour. Three things:
And until the link is settled, "GNSS gets a fix" fits the title better than "resolved". |
…-line gotcha, align the status with 10-31, attended-only install for the Location drop-ins
|
All three in, on today's main, and the title now says "gets a fix".
|
The 9a's GNSS cell goes from "no" to WIP: a position fix on 2026-10-30 and Omarchy's Location switch serving it on 2026-10-31, started by hand (46 of 52 GGA sentences fixed, 9-13 satellites, first fix about 80 s after a client connects). It needs kernel #150 (merged) and #114, and a session lasts as long as the modem's PCIe link, one to five minutes, which is the open item. The record lands with #202.
…ss#150 alone fixes GNSS The maintainer asked whether the CRC improvement holds with the SPI flush fix alone at the bootloader's drive 0. It does: 50 of 53 GGA fixed, every received frame at bit offset 0 with a valid CRC32, the same as at drive 3. Kernel derekross#151 is closed; the plan's cause 2, the notes and the AGENTS gotcha now say so, and the gotcha keeps the lesson: one change per boot, and check whose pins they are first.
|
Pushed a correction: kernel #151 (the SPI21 pin drive) is withdrawn and closed. With #150 alone at the bootloader's drive 0 the palmtop gave 50 of 53 GGA fixed and every received frame at bit offset 0 with a valid CRC32, the same as at drive 3, so the improvement was the flush fix's all along. Cause 2 in the plan, the notes and the AGENTS gotcha now say so; the pins overlay stays in the tree as a record and is out of the palmtop's boot set. |
…ains A traced controller build shows the substates enabled on both ends at every link-up. The link still cycles because the modem driver powers it off whenever the CP sleeps, as on stock, and the palmtop hung about 630 seconds after ONLINE instead of 170.
|
Thanks, all three points are in and they read well. #202 now conflicts with main in tegu's NOTES: when you merge main, please keep both entries and correct the 10-29 bullet's "drive 3". Four small things:
One question: did the install steps work on a boot with |
…of 10-31 Both NOTES entries kept, and the 10-29 bullet no longer credits drive 3. The status line and the script headers drop "resolved", linux-omarchy-mobile#150 is merged, derekross#150 and derekross#151 are written in full, HOLD=1 is the attended default, and the link section has today's substate result.
|
Merged main and pushed (edb3bda). Both NOTES entries are kept, and the 10-29 bullet no longer lists drive 3 among the pin settings: it says that boot also had the pins at drive 3, which the 10-31 check at drive 0 showed was not needed, with linux-omarchy-mobile#151 closed. The four points:
I also brought the link paragraph up to today. The L1 substates take effect on the real offset, but they don't stop the cycling, because cpif powers the link off every time the CP says it sleeps, as on stock. The hang moved to about 630 s after ONLINE on the one boot with them on, and the last trace puts it in cpif's power-down path. On the mask question: you're right that |
… link settles With the controller's steps traced, the session ran 299 power cycles and hung after cycle 292's power-up trained, before the controller's next access to its own registers. The point keeps moving, which suggests something outside the two drivers racing them.
… DBI read of LNKSTA With Wi-Fi down for the boot the session still hung at the same line. Finer traces put the hang on the controller's own LNKSTA read right after a good link-up.
|
Thanks for working through all four points and answering the mask question. The precondition line is exactly what I wanted, and the overnight traces (following the hang from the config save to the doorbell to the controller's own registers) are careful work. One thing before I merge. The Location section now says two different things about the boot mask. The precondition says your daily bootargs mask none of the GNSS units, but the next paragraph says they're masked by Smaller things, if you're in the file anyway:
I've merged #180, so this will need a rebase at the top of the 9a's notes (this evening's section goes above #180's morning one). |
…wn; not CPU sleep, not the PHY isolation The pstore console log shows the SoC stalling piece by piece after the CP's wake interrupt with no controller access after it. Holding the link up crashes the CP, and neither disabling CPU sleep nor keeping the PHY in bypass as stock does stops the resets.
The modem's AP2CP lines are on PERIC0, whose GPIO bus clock was gated around every pin access. Kept on (linux-omarchy-mobile derekross#172), four 600 s GNSS sessions ran over 1300 link cycles each with no reset.
A writable, manual-mode GPIO bus clock gate is switched off after every pin access, which hung the palmtop under the modem's GPIO traffic; GPIO gates stay critical or read-only on such a CMU.
3759 link power cycles with no reset and 1763 of 1765 sentences fixed, on the daily kernel with the PERIC0 GPIO clock kept on.
The open item at the top was the SoC hang under GNSS traffic; it now says what fixed it and points at the kernel pull request.
|
The GNSS session hang is found and fixed: the PERIC0 GPIO bus clock was a manual gate that the pin controller switched off after every access, and the modem's AP2CP lines on PERIC0 toggle about twice a second under GNSS. Kernel fix: derekross/linux-omarchy-mobile#172. On the palmtop's daily kernel with that change, two 10-minute sessions and one 30-minute session ran with no reset (1333-3759 link cycles, 558-1763 sentences fixed); every session before it reset. The plan's G3-L1 section has each step. |
The 9a's link driver began as a copy of the gs201 controller in kernel derekross#144, so it goes in as a second compatible stacked on that; the section lists what zumapro adds.
The zumapro compatible of the gs201 modem link driver and its binding, stacked on kernel derekross#144; the DT nodes follow once that stack is on omamobile.
# Conflicts: # device/tegu/NOTES.md
…aily boot, G3-L1 oldest first, the release for each run The daily boot masks none of the GNSS units, so the safety nets are the removal step and omamobile-gnss 0.2.5's hang marker. The G3-L1 runs are in date order, the Wi-Fi candidate is marked superseded, and each run names its kernel release.
|
Done, with main merged in (this evening's NOTES section above #180's morning one):
|
|
Congratulations on finding the hang. Following it from the config save to the doorbell, to the controller's own registers, and finally to a single GPIO write on PERIC0 was patient, careful work, and a gate the pin controller switches off after every access is a satisfying answer. Thanks for the three review points too: the Location paragraph now says what the daily boot has, the G3-L1 runs are oldest first with the superseded paragraph marked, and the notes give 170 s to about 26 minutes and point to linux-omarchy-mobile#172. One thing before I merge, about the Location path after the fix.
Could you tell me which path and which
One Location-switch session on the fixed kernel would then be the evidence the matrix's GNSS cell can cite. Smaller things:
|
…ink; HOLD=0 is the Location default The three sessions on the clock-fix kernel went through the Location switch with omamobile-gnss 0.2.5's start, which never writes keep_link_up, so the link cycled. With linux-omarchy-mobile#172 that is the working setting, so tegu-gnss-start now defaults to HOLD=0 and keeps 1 and 2 as diagnostics; holding the link would crash the CP after about 105 s. The plan names the path, the clock file's blob in the tested Image next to derekross#172's, and drops the private kit path.
|
Thanks, and you were right to ask. The 30-minute session, and the two 600 s sessions before it, used neither So I've made On which clock fix ran: Image |
…y default, fingerprint opt-in, GNSS to testing (#248) With omamobile-device-tegu 0.2.3 the charge limit, the AoC and NFC start at boot from the package, so the Pixel 9a's Charging, AoC, Speakers/mics, Sensors and NFC cells move to "works" (on by default), and Fingerprint to "works (opt-in)", as on the 7 Pro. Each was verified on a boot of its kernel pull request's head built from the shared config fragment; the record is #244's entry, which gives each boot's uname -r. The Boot cell records the 10-31 unmark and re-mark, and GNSS moves to testing with the modem started by hand (#114, #202, kernel #172). The kernel parts are in review, as each cell says. Closes #248 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Landed on main as 43a4b73 (squashed, authored by you). My review notes from before the merge, for the record: Thank you for the clear answer and for following through. Knowing that the three fixed-kernel sessions went through the package's start with the link cycling settles it: You're right about the hang-marker sentence. The drop-ins replace |
) The 9a's GNSS cell goes from "no" to WIP: a position fix on 2026-10-30 and Omarchy's Location switch serving it on 2026-10-31, started by hand (46 of 52 GGA sentences fixed, 9-13 satellites, first fix about 80 s after a client connects). It needs kernel derekross#150 (merged) and derekross#114, and a session lasts as long as the modem's PCIe link, one to five minutes, which is the open item. The record lands with derekross#202.
…y default, fingerprint opt-in, GNSS to testing (derekross#248) With omamobile-device-tegu 0.2.3 the charge limit, the AoC and NFC start at boot from the package, so the Pixel 9a's Charging, AoC, Speakers/mics, Sensors and NFC cells move to "works" (on by default), and Fingerprint to "works (opt-in)", as on the 7 Pro. Each was verified on a boot of its kernel pull request's head built from the shared config fragment; the record is derekross#244's entry, which gives each boot's uname -r. The Boot cell records the 10-31 unmark and re-mark, and GNSS moves to testing with the modem started by hand (derekross#114, derekross#202, kernel derekross#172). The kernel parts are in review, as each cell says. Closes derekross#248 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
) The 9a's GNSS gets a position fix: up to 13 satellites used, HDOP 0.8-1.2, GPS, Galileo, BeiDou and GLONASS tracked, CHPP clean. Two kernel fixes did it: spi-s3c64xx's FIFO flush order (kernel derekross#150) and SPI21's pinctrl (kernel derekross#151). With them stock's SPI configuration works and the byte swap is done in gnss_spi. MODEM-PLAN gains a "G3, resolved" section (causes, configuration, the link-hold answer, what's left); tegu-gnss-start now defaults to HOLD=0, with HOLD=1/2 kept as diagnostics, and three later sessions went through omamobile-gnss 0.2.5's start with the modem link left to cycle. AGENTS gains the flush-order and missing-pinctrl gotchas and the keep_link_up rule. NOTES gets a dated entry, and the run kit's bring-up overlays and session script are recorded, each marked bring-up only. Nothing locating the owner is recorded. Maintainer edit at merge: the Location paragraph no longer counts omamobile-gnss 0.2.5's hang marker as a safety net for the hand drop-ins, which replace the package's start and write no marker. Closes derekross#202 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n load the display controller, hang-safe (#438) A boot unit reads the bootloader's `exynos_drm.panel_name=` and sets `gs_dsim_cmd`'s `allow_tx` only for the S6E3FC5 that linux-omarchy-mobile#202 drives, so a Pixel 7a with an S6E3FC3 panel never receives its commands. `gs_decon` is blacklisted for autoload, because gs-decon decides the panel's rates at probe; the same start loads it after the write and sends the saved brightness again. The load is hang-safe, as on the Pixel 8 Pro: a synced marker written before the load keeps the next boot on simple-framebuffer after a load that never returned, until `--retry`. The unit runs only when the DECON node is enabled (device/lynx/overlays/display.dtso), so it does nothing on the published kernel. Run on the phone on 2026-11-16 on a build of #202 (`7.3.0-rc6-omamobile-push1-g61e15599a196`). Signed-off-by: cx3301J <cx3301J@users.noreply.github.com> Assisted-by: DeepSeek
The 9a's GNSS gets a position fix (2026-10-30, evening, on my kernel): up to 13 satellites used, HDOP 0.8-1.2, GPS, Galileo, BeiDou and GLONASS tracked, CHPP clean over a 3- and a 10-minute session with no reset. This records it.
Two fixes did it, both open as kernel PRs:
With those, stock's configuration works (mode 0, 32-bit words, feedback delay 0, automatic chip select, 64-byte reads), and the byte swap can be done in gnss_spi, so Google's
swap-modeDT property isn't needed. The CP link hang of 10-29 came back once under long GNSS traffic; holding the link up after ONLINE (cpif's bring-upkeep_link_up, which at load stops the CP from booting) ends it, answering G3-L1's first step.In this PR:
Nothing locating the owner is recorded: fix counts, satellites and HDOP only. tools/pr-check passes.