Skip to content

Pixel 9a (tegu / Tensor G4): native bring-up #28

Open

Description

@W4HO-ham

Pixel 9a (tegu / Tensor G4): native bring-up

Tracking the Pixel 9a port requested in the review of #24. This is WIP:
public main does not yet build the complete port and there is no installable
release. The first-light record and staged, disabled descriptor are merged;
subsystem support is being reviewed separately.

Verified on one prototype

Native Arch/Omarchy boots from slot B with its root on internal userdata,
eight CPUs, guarded storage, USB networking/SSH, S3908 touch and a G715
Panthor/Mesa desktop. Later local tests added synchronized display scanout
and read-only battery measurements. Registered buttons have not yet had a
complete physical-button/wake qualification. A locally assembled kernel from the curated subsystem topics now also
boots native Arch/Omarchy. Public main alone still lacks the complete
published tegu source and install recipe.

The exact revised combined Mesa package in #32 also passes system-library
GLES 3.1 pixel readback on the prototype. Vulkan has not been qualified.

Curated-topic hardware qualification

Kernel 7.3.0-rc6-omamobile-tegu-g7e457a96fc45, with board source
79288b2a3574, starts native root, G715 rendering and the Omarchy session
locally, without a host GO, module load or confirmation. Guard/watchdog,
matching modules, six protected checksums and five B payloads passed.
Read-only gauge/UPower reporting and native key registration passed;
physical button/wake and the new build's physical touch/smoothness checks
remain open. A temporary recovery helper is retained. Helper-free reset,
unplugged cold boot, charging policy and suspend are not qualified.

Current bench status: healthy native Arch desktop. Physical recovery returned
straight to the known-good keys-enabled profile; fresh read-only backups,
all five B payload checks and the native root/GPU/touch/gauge checks pass.
The subsequent native USB cable-cycle gate passed and its recovery timer
is canceled. No new flash was needed for recovery.

Review checklist

  • Record source/stock facts, partition/DT selector evidence and a gated
    first-light plan: merged device-tegu: record Pixel 9a first light and staged descriptor #24. Selector tuples remain commented, so image
    tooling refuses the staged descriptor until public DTB/AVB support lands.
  • Equalise both slot firmware stacks with the official BP3A.250905.014
    full OTA twice. Both resident loaders report tegu-16.3-13642544; stock A
    boots. Both slots' checked vbmeta signatures are valid with flags 0.
    Fresh protected backups were taken after the stock boot.
  • Retest ordinary installed-B boot without the HCRX workaround, with
    KVM enabled. All eight CPUs and a bounded real ARM64 KVM guest passed.
  • Derive OS header values from the matched equalised stock init_boot:
    OS_VERSION=16.0.0, OS_PATCH_LEVEL=2025-09.
  • Obtain maintainer agreement for layout 3.15 B mark-successful only.
    Fresh backup/checksum, all 8192 bytes compared, only byte 0x35 OR 0x02,
    read-only protection restored and ABL successful=yes verified. No other
    devinfo operation is agreed. Future installs use the same approved routine.
    This records a completed qualification, not the flag state after later
    normal project flashes; no manual retry refresh or slot-A change is implied.
  • Submit separate subsystem kernel topics; initial dependency index:
    drm/panthor: Add Tensor G4 cache hash configuration linux-omarchy-mobile#18 (comment)
    GPU cache hashes and PLL guard are merged. Other topics remain in review.
  • Complete requested kernel review revisions and public board integration.
    The local reconstruction from reviewed topics matches the qualified C,
    headers and DTS exactly; only explanatory clock-binding wording differs.
    Twelve exact commits now have verified human DCO sign-offs; existing
    APM/USI6 certifications are reused. Independent reboot and display topics
    are submitted as kernel tools/phone-smoke: read the slot from bootconfig #45 and omamobile-shell 0.2.18: the touch audit's shell quick wins #46. Remaining review changes, publication
    and settled dependencies remain open.
  • Prove the shared SOC_COMPATIBLE/AVB packing path locally with the
    unchanged omamobile images: SOC_COMPATIBLE and AVB_FOOTERS for zuma (and zumapro) #27 tool: reproducible five-image packing, exact payload/footer
    and stock-root AVB checks, and the guarded curated-topic boot above.
  • Land the shared tooling (omamobile images: SOC_COMPATIBLE and AVB_FOOTERS for zuma (and zumapro) #27), remaining public DTB/board source and
    device recipe before enabling the staged descriptor or offering installs.
  • Qualify firmware-on HSI0 in curated rescue context: four secure kernel
    samples across actual VBUS loss/return and180 idle seconds read powered=1.
    MAX77779's known data-switch restoration returned same-boot USB/SSH, with
    fresh guard/protected/payload checks. Result already posted on kernel Pixel 11 Pro Fold (yogi, malibu / Tensor G6) port: stock facts, contributors welcome #40:
    clk: samsung: zumapro: Add CMU_HSI0 linux-omarchy-mobile#40 (comment)
  • Repeat the cable-cycle check in native Arch context: four secure HSI0
    samples across actual VBUS loss/return and 180 idle seconds all read 1.
    One guarded MAX77779 switch restoration returns same-boot USB/SSH;
    final root/GPU/touch/gauge/guard, six protected hashes and five B payloads
    pass. Result: Pixel 9a (tegu / Tensor G4): native bring-up #28 (comment)
  • Qualify the current-boot polling rescue service before adopting a
    default replug policy. Production TCPM/PD and charging remain open.
  • Remove the temporary reboot helpers and qualify ordinary no-argument
    normal and explicit bootloader resets independently on the native desktop.
    Both pass, with matching modules and final guard/protected/payload checks.
    Production packaging and the legacy project-push reason remain open.
    Result: Pixel 9a (tegu / Tensor G4): native bring-up #28 (comment)
  • Advance Wi-Fi beyond the resource/clock stages. Resource acquisition
    and fourteen clock observations passed after the OTA. The bounded PMU
    bit-0 test stopped on 0xc0 → 0x01 → 0xc0 instead of expected 0xc1;
    restoration and independent health passed. Bits 6/7 remain undefined.
    The strict assertion is retained; no further PMU write or peripheral
    register/radio access is authorized by that result. The node is unbound,
    with no live link or enumerated endpoint. The private test uses a secure
    masked-RMW override absent from the public PMU identity proposal.

Bluetooth, charging/thermal policy, power/suspend, audio/sensors, modem and
cameras remain further work. Working desktop graphics are not evidence
that those peripherals work. No captures, calibration, firmware or vendor
binaries will be attached here.

Activity

  1. W4HO-ham commented on Oct 23, 2026

    @W4HO-ham
    ContributorAuthor

    A focused Wi-Fi/PMU source question while the public board work continues.

    On the equalised firmware, normal-HCRX/KVM private kernel and freshly verified protected backups, the fixed read-only observers now pass again: all fourteen HSI2 PCIe clock gates have bit 21 set, and PCIe1's PMU control at offset 0x3ec8 reads 0x000000c0 (bypass bit 0 clear). The inert readers were removed, recovery timers cancelled and fresh native GPU/display/touch/battery/root health passed. No PCIe/PHY peripheral read or radio operation was involved.

    A controlled follow-up now repeats the original strict bit-0 test once on the equalised firmware, after the same-boot clock/PMU observations and freshly refreshed six-partition checksum checks. The unchanged probe uses the previously specifically approved bit-0-only scope, immediate masked restoration, streamed logs, watchdog and cleanup/reboot timers. It gives the same result as before the firmware update: all secure operation returns are zero, but the words read c0 → 01 → c0. The original assertion still requires intermediate c1, so it stops with -EIO. Exact original PMU word restoration, reader removal, recovery cancellation and same-boot native health all pass. No acceptance rule was relaxed and no further write or peripheral access follows.

    Google's PCIe isolation helper uses rmw_priv_reg first; its EL3 wrapper passes the physical address, RMW operation, mask and value. That explains the access route, but I have not found a definition of bits 6/7 that explains the readback change.

    Have you seen those bits behave as feedback/status on a Tensor PHY control word, or do you have a citable register-field reference? I want to distinguish hardware status changing from an unexpected secure-update effect before preparing another isolation or peripheral-access gate. The current hardware path stays at the read-only boundary meanwhile.

  2. derekross commented on Oct 23, 2026

    @derekross
    Owner

    Thanks for stopping at the read-only boundary and asking. I haven't seen
    bits 6 and 7 do this, and I don't have a citable field definition for
    them. Here's what I can say from the sources:

    • Google's tree uses only two bits of that word. exynos_pcie_phy_isolation()
      updates bit 0 alone (PCIE_PHY_CONTROL_MASK is 0x1 in
      pcie-exynos-rc.h), and exynos_pcie_rc_resumed_phydown() sets bit 10
      (rmw_priv_reg(pmu_alive_pa + pmu_offset, 0x400, 0x400)). Nothing names
      bits 6 and 7, and Google never checks the word after writing it. Its
      ITMON handler only logs the whole value. I checked zumapro-pcie.dtsi
      too: PCIe1's pmu-offset is 0x3ec8, so you have the right register.
    • On our Pixel 7 Pro (gs201) the PCIe ch0 word at 0x3ec0 only ever read
      0x0 or 0x1, so I have no gs201 equivalent of your 0xc0.
    • The two write paths aren't the same. Google's driver uses the RMW SMC
      (PRIV_REG_OPTION_RMW with mask 0x1). Mainline's gs101-pmu.c
      tensor_sec_update_bits() uses atomic set/clear for every PMU_ALIVE
      offset except two: an SMC write of the bit index to the register's
      set/clear alias (offset | 0xc000 or | 0x8000). Which one did your probe
      use, for the set and for the restore?

    The restore may already answer your question. If the restore cleared only
    bit 0 (a masked RMW or an atomic clear) and bits 6 and 7 still came back
    to c0, then nothing you wrote put them back. That points to hardware
    status, not something the secure call stored. If the restore wrote the
    whole saved word 0xc0, the test can't tell the two apart. Could you say
    which it was?

    Either way, I'd rather not have another PMU write gate for this until the
    secure-regmap identity PR (kernel #31) is reviewed, since it decides which
    path these writes take on zumapro. Your strict c1 assertion is stricter
    than anything Google's driver does, but please keep it until we know what
    those bits are. When convenient, could you refresh the checklist in this
    issue's description? Equalisation, the header values and the B marking
    are done now.

  3. W4HO-ham commented on Oct 23, 2026

    @W4HO-ham
    ContributorAuthor

    Both operations used secure masked RMW, mask 0x1, not the atomic set/clear aliases and not a whole-word restore. I checked the exact private source used by the running post-OTA kernel and the saved probe/result; no new phone operation was needed.

    The private kernel's G4-specific reg_update_bits override routes only offset 0x3ec8 to its tensor_sec_reg_rmw() helper. That invokes TENSOR_SMC_PMU_SEC_REG with the physical address, option 2, mask and value. The probe calls were:

    • Set: regmap_update_bits(pmu, 0x3ec8, BIT(0), BIT(0)) → secure RMW mask 0x1, value 0x1.
    • Restore: regmap_update_bits(pmu, 0x3ec8, BIT(0), original & BIT(0)) → secure RMW mask 0x1, value 0x0 because the original word was 0xc0.

    All four set/read/restore/read return values were 0. The observed sequence was 0xc0 → 0x01 → 0xc0. The restoration request did not contain 0xc0; the probe printed the intermediate result only after the restore/readback had completed. The strict expected intermediate 0xc1 failed with -EIO, and fresh native/guard/USB/touch/GPU health passed before recovery cancellation.

    This supports your hardware-status interpretation, but does not establish the meaning of bits 6/7. The strict assertion remains intact and no further PMU write, PHY/controller/BAR/ELBI access or radio operation has been attempted.

    One distinction to make explicit: the running private kernel includes that bounded PCIe1 RMW override; kernel PR31 does not. Its review revision is binding-only with the zuma fallback, as requested, and retains the public default update path. I am not treating the private masked-RMW result as qualification of the public fallback's atomic-alias behavior. Any later PCIe-specific callback proposal needs its own source review and hardware gate.

    I have also refreshed this issue's checklist for the completed firmware equalisation, stock header values, approved B-only successful marking, merged first-light record and submitted subsystem topics. Public board integration and working Wi-Fi remain open.

  4. W4HO-ham commented on Oct 23, 2026

    @W4HO-ham
    ContributorAuthor

    A focused source-design question for the public HSI0/USB dependency, while the PMU write gates remain held as requested.

    The prototype's cached live DT has no power-domains edge on DWC3, either PHY, or CMU_HSI0. Both PHY drivers are already bound; the actual resource gaps are the child's three identical 19.2 MHz placeholder clocks and absent supply phandles. Fixed read-only observations now establish the selected PLL/divider/mux/gate state and the three inherited USB rails (decoded 0.75/1.2/3.3 V). A separate kernel secure-regmap read returned HSI0_STATUS bit 0 on at PMU + 0x2a84. Every observer was removed with same-boot health and recovery cancellation verified. This is current-state evidence, not a lifetime power lease or cold-cycle test.

    I have a local, unpublished clock candidate with the real 19-output graph, five inputs, existing read-only Samsung PLL operations and the qualified USI2 controls. It drops the staged DWC3 reference, which would create a supplier cycle when USB requests these clocks. Source/object/schema, mocked lifetime tests and the combined Image/modules link pass. Its present guard requires an attached real domain before retaining the CMU's own runtime PM reference. No fake domain is supplied or board node enabled.

    I then checked the reference: gs201.dtsi's CMU_HSI0 has no domain edge either, and clk-gs201.c uses ordinary exynos_arm64_register_cmu. So making full G4 genpd support mandatory for first light is a stronger prerequisite than the current reference port has.

    For the first public G4 DTB, which direction do you prefer: an explicitly bounded firmware-on clock handoff following the reference port, with power-off/suspend still unqualified, or a complete HSI0 domain provider before enabling this clock node? The latter needs the GPL domain sequence, roughly 127 CMU context entries, rail ownership and S2MPU/security restoration; a local PM counter alone is insufficient. I am tracing those contracts offline meanwhile. This question does not change the PCIe strict assertion or propose a PMU write test.

  5. W4HO-ham commented on Oct 23, 2026

    @W4HO-ham
    ContributorAuthor

    A separate source-only APM clock question while the HSI0 design question and PMU write boundary remain unchanged.

    The pinned GPL CAL table defines the APM input signals directly with FIXEDRATE: oscillator 26 MHz, ALIVE-divider signals 400/200/50/100 MHz, and pad input 100 MHz. Its separate PLL_ALV entry is 787 MHz, so dividing that literal does not reproduce the APM input values. I found no programmable ALIVE PLL register model in that CAL register list. The divider-like names alone do not establish exact input frequencies.

    The public gs101 driver at the current base instead uses 393.216/196.608/49.152 MHz for its three APM signals. I am not transferring those sibling rates to G4 without evidence.

    Do you have a citable G4 rate reference or an established way to distinguish CAL's nominal signal values from the actual APM input frequencies? If the gs201 work established the corresponding rates independently, that method/source would also help.

    I have an uncommitted, unpublished 15-clock APM draft with the actual ordered mux parents, divider and six read-only gates. It has no selector/rate/gate writes or manual/suspend register list. Source field/ID checks, strict style, ARM64 W=1 objects, full binding/examples, 11 exact-source sanitizer probe cases and the combined Image/modules link pass. The nominal input-rate accuracy gate is explicitly still open, so neither it nor a production mailbox/PMIC node is enabled or submitted. No phone operation, new PMU write or radio step was run for this work.

  6. derekross commented on Oct 23, 2026

    @derekross
    Owner

    Thanks. Answers to both, HSI0 first.

    HSI0: firmware-on handoff, no domain, for first light.

    Do it the way the reference port does. Upstream gs101 has run USB on
    Pixel 6 this way for a while: cmu_hsi0 in gs101.dtsi has no
    power-domains, and neither do the PHY or DWC3. clk-gs101.c and our
    clk-gs201.c both probe with the plain exynos_arm64_register_cmu(), not
    the _pm variant, so there's no runtime PM on the CMU and no domain
    reference to hold. I'd drop the attached-domain guard rather than add a
    domain for it to find. Don't call exynos_arm64_register_cmu_pm() for
    HSI0 yet.

    Two things to keep in mind on that path:

    • The plain helper still writes at probe. With auto_clock_gate and an
      option_offset set, exynos_arm64_init_clocks() writes the CMU's
      controller option register and sets each listed gate register. Those
      probe-time writes are what to qualify, not a power domain. If you'd
      rather first light write nothing, leave option_offset and clk_regs
      out of the G4 HSI0 info for now. That's another thing to note as
      unqualified.
    • Say it in the DT, the way we did for G3D in gs201.dtsi: no
      power-domains because the bootloader leaves HSI0 on and nothing in
      Linux turns it off. Power-off, suspend and the S2MPU restore stay
      explicitly unqualified.

    The full HSI0 domain provider (the GPL sequence, CMU context, rail
    ownership, S2MPU) belongs in its own topic after first light, reviewed on
    its own. You're right that a local PM counter wouldn't stand in for it.
    One check that does matter for the firmware-on path: confirm HSI0 stays on
    across an unplug/replug and a few idle minutes on the bound USB stack. Use
    your existing kernel-side read for that, not anything from userspace.

    APM input rates: no citable G4 reference beyond CAL, as far as I know.

    What I can tell you about where the sibling numbers come from:

    • gs101's 393.216/196.608/49.152 MHz are CAL's nominal values too. At the
      commit you pinned, cal-if/gs101/cmucal-node.c has OSCCLK_APM
      24.576 MHz and PLL_ALV_DIV2/4/8/16_APM 393.216/196.608/98.304/49.152 MHz.
      Those are exact multiples of gs101's 24.576 MHz oscillator. The upstream
      commit (2c597bb7d66a, "clk: samsung: clk-gs101: Add cmu_top, cmu_misc and
      cmu_apm support") doesn't cite any other source. They don't carry over to
      a 26 MHz part. So don't move them to G4. You're right about that.
    • gs201 didn't establish the rates independently either. We took CAL's
      _APM values (394/197/99/50, pad 100) directly, and the commit says so
      (clk: samsung: gs201: add a read-only CMU_APM subset for the APM UART).
      Our only hardware check was functional: Bluetooth's HCI link runs at
      3 Mbaud off the 197 MHz APM UART clock. That bounds the error to UART
      tolerance (a couple of percent). It wouldn't separate 393.5 from 400.
    • In gs201's table, the _APM values match PLL_ALV 787 MHz divided by
      2/4/8/16, rounded up (393.5, 196.75, 98.4, 49.2). In zuma's table the
      same PLL_ALV 787 and PLL_ALV_DIV2/4/16 393/197/50 are still there,
      but the _APM entries read 400/200/100/50. zuma's cmucal-vclk.c LUTs
      (blk_apm 400000, APM UART and i3c_pmic 200000 kHz) follow those
      numbers, so they don't confirm anything independently. My guess is that
      zuma's _APM entries are round targets and the real inputs are either
      the 393.5 family or really 400. The source can't tell which. Note that
      Google's own kernel computes the rates from these FIXEDRATE values, so
      stock on tegu reports 400/200 either way.

    How much this matters depends on the consumer. The mailboxes, ACPM and
    the PMIC interface (run by the APM firmware) don't take a rate from Linux,
    so a nominal value with a comment ("CAL nominal, unmeasured") shouldn't
    hold up a read-only CMU description or a mailbox node. The rate is only
    real for a Linux consumer that derives timing from it: a UART's baud, an
    I2C/I3C SCL rate, a timer.

    For measuring it, the cleanest way I know uses such a consumer, timed
    against the arch timer:

    • Take an APM-domain USI UART that Linux owns. On cheetah, USI1 was
      Bluetooth's UART. Put it in internal loopback (TIOCM_LOOP, as our
      tools/bt-uart-test loop does, with no pins involved). Write a known
      long burst at a fixed baud, tcdrain(), and time it with
      CLOCK_MONOTONIC.
    • The driver sets the divider from the rate it believes. So the measured
      baud over the requested baud is the actual clock over the registered
      clock, after you correct for the divider's quantisation (computable from
      the driver's formula).
    • A few seconds of transfer puts scheduling jitter well under 0.1 %.
      That's enough to separate 393.5 from 400 (about 1.6 %).

    That writes only the UART's own registers, and in loopback, but it's still
    a gate of its own and your call to propose. I'd keep it off the critical
    path: land the APM description with nominal values and the open accuracy
    gate noted, as you have it.

  7. W4HO-ham commented on Oct 23, 2026

    @W4HO-ham
    ContributorAuthor

    Thanks—that resolves the two source-policy questions.

    I've prepared a separate local HSI0 revision removing the attached-domain requirement and local runtime-PM bookkeeping. It retains the bound secure PMU status check before CMU access, real input references and the existing clock/control fields. The binding now describes firmware-on handoff without a domain. The eventual board DTS will carry the explicit no-power-domains explanation too; its HSI0/USB nodes are not submitted yet.

    This provider currently uses device-managed clock registration rather than either common CMU helper. Its existing three owned USI2 manual-mode setup writes remain, so I am not calling it a zero-write probe. Those writes and the consumer gate/divider operations still need staged qualification. Power-off, suspend, rail/context ownership and S2MPU restoration remain outside first light. I will record the kernel-side HSI0 status observations across unplug/replug and idle on the bound stack separately; that hardware test has not run.

    The source revision passes strict style, an ARM64 W=1 object build, complete binding/style/example checks and exact-source secure-status/error/reference tests. It is unsigned and unpublished; no phone operation followed this revision.

    The APM draft already labels its input rates as nominal CAL descriptions, not measured rates. I'll curate that description without making independent rate measurement a prerequisite for the mailbox/ACPM/PMIC path, while keeping timing-sensitive consumers behind the accuracy gate. No sibling rates are being substituted. The PCIe strict c1 assertion and unknown-bit question are unchanged.

  8. W4HO-ham commented on Oct 24, 2026

    @W4HO-ham
    ContributorAuthor

    Further local curated-source qualification is complete on Pixel 9a, using the unchanged 7.3.0-rc6-omamobile-tegu-g8d09358f40dc Image, the existing drivers, the shared guarded initramfs and the unmodified project image/flash tool:

    • HSI0 actual-VBUS-gated unplug/replug + three idle minutes: four secure power samples all1; one guarded MAX77779 data-switch restoration returns the same-boot USB link. Detailed result.
    • DT-bound read-only MAX77779 measurements and passive scheduled-refresh event pass. Battery result.
    • The USI2/SPI20 host and reset-held default state pass. Enabling the existing S3908 child, CUSTOM_ALIVE IRQ provider and named regulator consumers identifies the controller and registers input at0.558s. The regulator API reports AVDD3.3V/DVDD1.8V. Owner swipes produce41 contacts,40 releases and685 synchronized frames with motion in both axes in a nonexclusive20s evdev capture. The gauge remains usable alongside it.

    Final same-boot checks pass USB, storage guard, all six protected backup/live hashes and all five installed B payloads. Watchdog, persistent console, laptop logging and independent recovery covered the gates; recovery was canceled after fresh SSH. No stock Android fallback or firmware/slotA/calibration write occurred.

    This remains a rescue-root qualification, not a curated desktop or released installer. The disabled touch DTS additions are still local and unsigned, outside the earlier DCO packets; they will need separate review/sign-off. GPU/display integration, production TCPC ownership, charger/thermal/suspend and the other roadmap stages remain open. The already requested #33 regression is separate.

  9. W4HO-ham commented on Oct 24, 2026

    @W4HO-ham
    ContributorAuthor

    Native Arch context gate passed on 7.3.0-rc6-omamobile-tegu-g7e457a96fc45, retaining the known-good GPU/display/touch/gauge/input profile and recovery helper. No image flash was needed after physical recovery.

    One actual five-second unplug/replug, followed by 180 idle seconds:

    Kernel secure HSI0 sample Powered bit
    Before unplug 1
    While VBUS absent 1
    After reconnect 1
    After 180 seconds idle 1

    MAX77779 (product 0x79, vendor 0x0456) opened its data switch on unplug. On VBUS return, the same previously qualified identity/ownership/VBUS/state-guarded helper performed exactly one USBSW_CTRL 0x93: 0x00 -> 0x09 restoration and readback. Fresh USB SSH returned on the same boot. Final native-root/GPU/gauge/guard checks, six protected hashes against freshly captured backups, and all five known B payloads passed. No kernel fault/UFS error. The diagnostic reader is removed, its recovery timer canceled after fresh health, and PID1's 10-second watchdog retained.

    Next policy gate: run this exact helper under a manually started, current-boot-only polling stand-in, following cheetah's existing rescue-loop pattern. It stops on a TCPC client, unexpected state or any error, does no write with VBUS absent, and has a persistent marker plus independent recovery. One further cable cycle, then stop and recheck; no boot enablement or production TCPM/PD/charging claim. This changes no HSI0 clock, power-domain, PHY or PMU policy. Source updates remain queued for their dependency/sign-off gates.

  10. W4HO-ham commented on Oct 24, 2026

    @W4HO-ham
    ContributorAuthor

    Helper-free native boot and both reset gates now pass on 7.3.0-rc6-omamobile-tegu-g7e457a96fc45, retaining the qualified keys/touch/G715/display/gauge profile.

    Only the standard-reboot opt-in DT flag changes. The Image, matching modules, shared guarded initramfs and hardware properties remain unchanged. Compatible startup files keep the qualified late warm-mode setup; there is no early reboot=warm argument. Both private recovery/normal modules are absent.

    • Local Arch/G715/Omarchy startup passes without a host GO or module load.
    • Ordinary no-argument systemctl reboot returns directly to a fresh native desktop, without any host boot/flash command.
    • A separate explicit bootloader reset reaches actual Fastboot. The unchanged project tool then flashes the same five checked B images and the native desktop returns.
    • Final native/root/USB/watchdog checks, storage guard, six protected backup hashes and all five installed payload hashes pass. No kernel fault or UFS error; local recovery deadlines/markers are canceled on success.

    No firmware, slot-A, devinfo or manual retry write. Physical button actions and a new finger/smoothness test are not claimed. Production packaging/default enablement, the project push caller's legacy omamobile reason, unplugged cold boot, charging and suspend remain open.

    The independent common zero-valued-mode fix and inherited-display support are submitted as kernel #45 and #46; both CI runs pass. Requested wording revisions for #35, #37 and #39 are published separately, preserving the human sign-offs.

  11. W4HO-ham commented on Oct 24, 2026

    @W4HO-ham
    ContributorAuthor

    A focused HSI1 input-clock question for the GPIO prerequisite (haptics reset
    is on this block), applying the nominal-rate guidance above.

    Guarded observations on native 7.3.0-rc6-omamobile-tegu-g7e457a96fc45:
    secure HSI1_STATUS is 1. The TOP path is shared0/4, gate on, divide by 1,
    but the local NoC mux actually selects ALT (0x1000 = 1). ALT_USER
    selects the alternate input (0x620 = 0x10) and its divider is 1
    (0x1808 = 0); the unused NOC_USER selects oscillator (0x600 = 0).
    The final mux recheck agrees. Each inert observer was removed; fresh
    root/guard, six protected hashes, five B payloads and recovery cancellation
    pass. No clock/power/GPIO/radio/PCIe write or peripheral access occurred.

    The pinned GPL node table
    declares I_CLK_HSI1_ALT as FIXEDRATE(400000000) without a parent. I have
    not established its physical origin or measured its rate, and have not
    aliased it to APM or HSI0. The
    local save/read sequence
    conditions these local controls on secure PMU 0x2b04 bit 0; the observer
    rechecks that and the qualified TOP state immediately before local access.

    For first GPIO support, would you accept a distinct input described as
    CAL nominal 400 MHz, unmeasured, with both local routes retained,
    read-only mux/divider/gate operations and no Linux power-off/suspend path?
    GPIO would not derive timing from that rate. Actual rate qualification and
    complete domain/context handling would remain separate; there would be no
    forced NoC switch or clock programming. Or should this HSI1 input wait for
    a citable physical-source identification before even that GPIO gate?

    The two-file, three-clock TOP prerequisite builds cleanly with ARM64 W=1;
    its two local commits remain unsigned/unpublished. No local CCF provider or
    GPIO node is enabled by these observations.

  12. W4HO-ham commented on Oct 24, 2026

    @W4HO-ham
    ContributorAuthor

    Following up on the PCIe qualification boundary now that kernel #31 is merged. That PR is binding-only; it did not add the private PCIe masked-RMW override described in my earlier reply.

    The strict c1 requirement is still intact. The saved test used mask 0x1 for both set and restore, with no whole-word restore, and observed c0 → 01 → c0. Neither the GPL field search nor the existing captures establish bits 6/7. No new isolation write or PCIe/PHY register access has followed.

    What concrete evidence would you want next to resolve that requirement? In particular, would a separate source-only review of the PCIe-specific secure-RMW callback be useful first, or should that also wait for an identified register-field reference? If the latter, is there a suggested upstream contact or non-writing diagnostic that could establish the missing semantics? The current public atomic-alias path and the private masked-RMW observation remain separate qualifications.

    The independent normal UART/GPIO work now completes ROM HCI identification (BCM4383A1) and external 3 Mbaud exchanges with verified power-off. That clears Bluetooth's transport prerequisite, but does not qualify the separate Wi-Fi PCIe path.

  13. W4HO-ham commented on Oct 26, 2026

    @W4HO-ham
    ContributorAuthor

    Wi-Fi status on the Pixel 9a, and three PRs that are ready now: linux-omarchy-mobile #54 (the S2MPG14 RTC), omarchy-mobile #43 (shell: dock under apps, home pill, icons, Power-key wake) and #44 (the tegu audio UCM profile).

    Wi-Fi works on the phone today: PCIe channel 1 trains at Gen2 x1 to the BCM4383 (14e4:4449), brcmfmac loads the owner's firmware set, iwd associates with WPA2 and the phone has an address, DNS and internet. What it took, so you can tell me how you want it shaped before I send it:

    1. A DesignWare host glue for the G4's PCIe channel 1 (google,zumapro-pcie): CMU_HSI2 gates, the PMU PHY isolation bit over the SMC regmap, Google's PHY power-up and config sequences as tables, PERST and WLAN_EN as GPIOs, MSI only (no INTx map). Three things were decisive: map ELBI without a region request (the core requests "elbi" itself), clear APP_XFER_PENDING after training (left set, config space works but every memory read answers 0xffffffff), and keep ASPM off. With L1 the link does not come back and config reads turn to 0xffffffff until the root port's ASPM bits are cleared, so L1/L1ss (CLKREQ, the vendor's set_l1ss) is still unimplemented. It lives as an out-of-tree probe in the bring-up tree for now.
    2. brcmfmac: BRCM_CC_4383_CHIP_ID firmware names and the 0x4449 PCI id, the versioned "join" gated on this chip too (the unversioned join and SET_SSID both fail with -52 on firmware 20.123.13), and feature_disable=0x182000 because the firmware rejects the supplicant offload (brcmf_set_wsec: failed to change PSK in firmware); iwd does the 4-way handshake. This sits on top of the api9 series (ring validation, shared startup, chip memory) that is not in omamobile yet.
    3. The base DT: the HSI2 pinctrl block enabled (a runtime overlay oopsed in samsung_pinctrl_probe, its bank tables are __initconst), gpa7/gpa8 EINT data, and the pcie node.

    Open before a PR: ASPM/L1 handling as above, and DMA isolation. The endpoint has unrestricted DMA right now; the HSI2 SysMMU is in the tree but, as on cheetah, PCIe and UFS share a VID, so I'd rather hear how you want that boundary drawn (restricted DMA pool as cheetah does, or the SysMMU with a VID split) before sending the host driver. Also: do you want the api9 brcmfmac commits as their own PR ahead of the BCM4383 bits?

  14. derekross commented on Oct 26, 2026

    @derekross
    Owner

    Great news, @W4HO-ham: Wi-Fi with internet on the 9a, and the APP_XFER_PENDING and ELBI findings are exactly the kind of thing that saves the next port a week. Please put them in tegu's notes as you have them here.

    DMA isolation: a restricted DMA pool, as on cheetah. On gs201 the HSI2 SysMMU and the S2MPU carry PCIe and UFS on the same VID, so confining the Wi-Fi chip confined storage too and wedged the phone (cheetah's WIFI-BATTERY-PLAN, W4). We ship a 16 MiB restricted-dma-pool on the Wi-Fi node instead, and our SysMMU driver refuses to enforce on that shared VID. Please do the same on tegu: a memory-region on the endpoint node, and no SysMMU enforcement. If you find evidence that zumapro puts PCIe and UFS on separate VIDs, that would be a real upgrade, but bring the evidence first and don't enforce anything on a VID that UFS might share.

    Order: yes, the api9 brcmfmac commits as their own PR first. That code is shared with cheetah's BCM4389, so it needs to show it changes nothing there (I'll boot cheetah with it). Then the BCM4383 support, then the host glue. For the BCM4383 bits, I'd rather have the join and the feature_disable bits as chip- or firmware-keyed quirks in the driver than a module parameter, so the phone works without extra config.

    ASPM off is fine for a first version, as long as the host driver says why, with the evidence you have: L1 doesn't come back, and config reads return 0xffffffff until the root port's ASPM bits are cleared. L1/L1ss can follow once the link basics are in.

    The __initconst pinctrl oops is worth recording too, since it rules out enabling HSI2's pin block from an overlay.

  15. W4HO-ham commented on Oct 26, 2026

    @W4HO-ham
    ContributorAuthor

    Done on tegu, tested today on the palmtop:

    • Base DT: wifi_dma_pool under /reserved-memory, restricted-dma-pool, 16 MiB, alignment = <0x0 0x200000>, alloc-ranges below 4 GiB (two address cells, one size cell, as the SoC declares; the 12-byte entry gotcha from cheetah applies here too).
    • PCIe1 overlay: root port pci@0,0 (pciclass,0604) with wifi@0,0 = pci14e4,4449, brcm,board-type = "google,zumapro-tegu", memory-region = <&wifi_dma_pool>.
    • Boot: software IO TLB: Reserved memory: created restricted DMA pool at 0x00000000ff000000, size 16 MiB; /sys/bus/pci/devices/0000:01:00.0/of_node/compatible is pci14e4,4449; after the join, /sys/kernel/debug/swiotlb/wifi-dma-pool/io_tlb_used is 1161 of 8192 slabs, so brcmfmac's DMA bounces through the pool. Scan, join (iwd) and traffic work as before. No SysMMU or S2MPU enforcement, as you said.
    • dt-validate: no new diagnostics against the schema for the base DTB or the overlaid one.

    The __initconst pinctrl oops is recorded in device/tegu/NOTES.md (a Samsung pinctrl block can't be enabled by a runtime overlay; HSI2 is in the base DT now).

    PR order, as you asked: the api9 brcmfmac commits first as their own PR (the three patches I have are against my m1 tree and don't apply to yours as they are, so I'm reworking them on a worktree off omamobile), then the BCM4383 bits keyed on the chip id (firmware names, PCI id, JOIN_V1 on 4389 || 4383, the PSK-offload refusal as a firmware-keyed feature disable rather than a module parameter), then the host glue with the ASPM rationale (L1 exit not handled; pcie_aspm.policy=performance until it is) and the DT pieces.

  16. W4HO-ham commented on Oct 26, 2026

    @W4HO-ham
    ContributorAuthor

    The brcmfmac part is up as derekross/linux-omarchy-mobile#56, in your order: the version 9 firmware support first (keyed on the reported version; the BCM4389 path is unchanged, that is the cheetah boot you wanted), then the BCM4383 with its join and supplicant-offload quirks keyed on the chip, no module parameters. It is based on omamobile-int because the version 9 code sits on the BCM4389 PCIe work that is in the integration snapshot but not on omamobile yet; retarget as you prefer. The palmtop runs that exact build now, loaded without parameters. The host glue and the DT come next, separately.

  17. W4HO-ham commented on Oct 28, 2026

    @W4HO-ham
    ContributorAuthor

    Status of the Pixel 9a's first eleven pieces. These are the kernel PRs that would make each one "merged upstream", and what each is waiting on. All of them sit on today's omamobile, with CI green.

    Piece PRs Waiting on
    1 Boot, USB, SSH derekross/linux-omarchy-mobile#80 (SoC and board DT) Your merge: the boot of this head is posted on derekross/linux-omarchy-mobile#80, and the "Tested on" line is in the board commit
    2 Clocks, pins, UFS derekross/linux-omarchy-mobile#99 (UFS idle clock gating) Your review. The clocks and the UFS reset are merged (derekross/linux-omarchy-mobile#57, derekross/linux-omarchy-mobile#79, derekross/linux-omarchy-mobile#83, derekross/linux-omarchy-mobile#102)
    3 Touch and buttons derekross/linux-omarchy-mobile#81 (touch DT), derekross/linux-omarchy-mobile#82 (sleep fix) derekross/linux-omarchy-mobile#81 lands with the stack through derekross/linux-omarchy-mobile#87; derekross/linux-omarchy-mobile#82 waits on your 7 Pro suspend run
    4 GPU derekross/linux-omarchy-mobile#85, derekross/linux-omarchy-mobile#87 The stack through derekross/linux-omarchy-mobile#87, after derekross/linux-omarchy-mobile#80
    5 Display derekross/linux-omarchy-mobile#88 (120 Hz), derekross/linux-omarchy-mobile#90 (display DT) derekross/linux-omarchy-mobile#88 waits on your 7 Pro run; derekross/linux-omarchy-mobile#90 is on the stack
    6 Battery and charging derekross/linux-omarchy-mobile#92 (charger), derekross/linux-omarchy-mobile#119 (gauge and USB-C DT) derekross/linux-omarchy-mobile#92 is answered and waits on your review. derekross/linux-omarchy-mobile#119 booted this morning in its exact form and is ready
    7 Wi-Fi derekross/linux-omarchy-mobile#56 Your BCM4389 series on omamobile (your plan on derekross/linux-omarchy-mobile#56)
    8 Bluetooth derekross/linux-omarchy-mobile#122 (DT; the driver derekross/linux-omarchy-mobile#94 is merged) Booted in its exact form (round bt2) and ready
    9 Haptics and torch derekross/linux-omarchy-mobile#96 (DT; the drivers are merged) On the stack after derekross/linux-omarchy-mobile#90
    10 ACPM, PMIC, thermal derekross/linux-omarchy-mobile#87 (DVFS and zones), derekross/linux-omarchy-mobile#101 (S2MPG15) derekross/linux-omarchy-mobile#86 merged; derekross/linux-omarchy-mobile#101 waits on your review
    11 CPU idle, UFS power derekross/linux-omarchy-mobile#99 Plus a measurement on my side: today's cable-out sleep run

    Order, if it helps: derekross/linux-omarchy-mobile#80 → derekross/linux-omarchy-mobile#81 → derekross/linux-omarchy-mobile#85 → derekross/linux-omarchy-mobile#87 → derekross/linux-omarchy-mobile#90 → derekross/linux-omarchy-mobile#96 → derekross/linux-omarchy-mobile#105 → derekross/linux-omarchy-mobile#106 → derekross/linux-omarchy-mobile#119 → derekross/linux-omarchy-mobile#122, then derekross/linux-omarchy-mobile#99, derekross/linux-omarchy-mobile#101, derekross/linux-omarchy-mobile#92, derekross/linux-omarchy-mobile#88, derekross/linux-omarchy-mobile#82 and derekross/linux-omarchy-mobile#56 independently.

    The DT stack's PRs only add the 9a's board file and zumapro nodes; they don't touch any other board's DTB.

  18. W4HO-ham commented on Oct 28, 2026

    @W4HO-ham
    ContributorAuthor

    Status of the 9a's 16 pieces, 2026-10-28:

    # Piece Works on the palmtop Merged
    1 Boot, USB, SSH ✅ ❌ PRs ready
    2 Clocks, pins, storage ✅ ❌ PRs ready
    3 Touch and buttons ✅ ❌ PRs ready
    4 GPU ✅ ❌ PRs ready
    5 Display (120 Hz) ✅ ❌ PRs ready
    6 Battery ✅ ❌ PRs ready
    7 Wi-Fi ✅ ❌ PRs ready
    8 Bluetooth ✅ ❌ PRs ready
    9 Haptics and torch ✅ ❌ PRs ready
    10 CPU DVFS and thermal ✅ ❌ PRs ready
    11 CPU idle, UFS power ✅ ❌ PRs ready
    12 USB-C (host mode, charging) ❌ a USB drive read once by hand; the role switch on attach is next ❌
    13 Suspend ✅ (drain still high) ❌
    14 AoC, sensors, audio ✅ ❌
    15 GNSS, NFC, modem, cameras ❌ GNSS gets no fix ❌
    16 Fingerprint ❌ ❌

    The merge order for 1-11 is in my earlier comment here. One note: the palmtop lost its root today after a failed DTB round. Slot B wasn't marked successful, so ABL fell back to stock A and stock reformatted userdata. It's restored and slot B is marked again. AGENTS.md now has the rule.

  19. derekross commented on Oct 28, 2026

    @derekross
    Owner

    Thank you, @W4HO-ham, this is a really useful map of where the 9a stands, and the merge order saves me working it out myself.

    A few things from reading it against the trees:

    1. Links. In this repository a bare #80 or #122 links to omarchy-mobile, not to the kernel, so in your two tables #119 points to a tools PR and #122 to the recents issue. Writing derekross/linux-omarchy-mobile#80 (or kernel #80 with a link) fixes that.

    2. A rebase for the device-tree stack. Kernel Speakers: the bottom CS35L41's protection model runs away with DRE on below stock's gain (media muted since 2026-10-26) #80 and the stack above it now conflict with omamobile, all in arch/arm64/boot/dts/exynos/google/Makefile: the Pixel 7 (panther) board landed this afternoon and adds its line where yours goes. It's a one-line resolution (keep both entries). Please rebase the top of the stack when you next push, so CI runs on the tree that will land.

    3. The slot A fallback today. Sorry you lost the root again, and thank you for saying so here. Three things before the next session:

    derekross/linux-omarchy-mobile#56 is mine to move, and I'll update it there.

  20. W4HO-ham commented on Oct 28, 2026

    @W4HO-ham
    ContributorAuthor

    Thanks. All three are done:

    1. Links. The merge-list table now links every PR as derekross/linux-omarchy-mobile#N. The status chart had no numbers.
    2. The DT stack is rebased onto today's omamobile, with panther's line kept beside the 9a's in the Makefile. All eleven branches (kernel Speakers: the bottom CS35L41's protection model runs away with DRE on below stock's gain (media muted since 2026-10-26) #80, omamobile-audio 0.5.3: the boost converter off below gain 17, DRE and the jump rule by stage; AOC-PLAN C9j #81, omamobile-shell: the home widgets collapse in landscape #85, The screen lock never engages on a phone installed from our image #87, Touch batch 4: landscape layouts for the home, call card, quick settings, drawer and recents (omamobile-shell 0.2.30) #90, README, DEVICES, ROADMAP, AOC-PLAN: the speakers work #96, tools/extract-camera-data: no S5KGN1, S5K3J1 or S5KGM5 lists in CP3A.261005.002.A1's camera HAL #105, tools/phone-provision: a phone with no earlier backup can't be provisioned (README quick start) #106, DEVICES: status after the 2026-10-28 merges #118, tools/kernel-pr-check: a base build that stops no longer fakes new device-tree warnings #119, omamobile.switcher: cards show the app id instead of its name (TOUCH-PLAN 3.7) #122) were force-pushed with leases. The three tips build zumapro-tegu.dtb and gs201-panther.dtb with W=1. The one warning is gs201's dsi node, which predates the stack.
    3. The fallback. The answers are on AGENTS, device/tegu: a fallback to slot A on the 9a costs the Linux root #125 and in a dated tegu NOTES entry there. On each day a fresh read-only backup of efs, efs_backup, modem_userdata, persist, modem_a/b, devinfo and metadata was taken before anything else, so the next modem step compares against today's. Slot B was marked with omamobile-bootctl mark-successful under the approved conditions, with the before and after bytes in the entry. Stock ended in recovery both times, but userdata had already been overwritten.
  21. den-kloot commented on Nov 7, 2026

    @den-kloot
    Contributor

    Heads-up to avoid duplicate work: I'm looking at the Pixel 9a main camera's autofocus, on top of linux-omarchy-mobile#164's device tree (local build, with an AI assistant).

    • The main module's focus/OIS controller is an LC898129 (the HAL calls it act-nessie / ois-nessie), on hsi2c_5 at 0x24. It shares the main sensor's supplies and reset and only answers while the sensor streams.
    • Writing 0xF01A ("Focus target set" in the HAL, value 0x0002xxxx) moves the lens. Close subjects come out sharp.
    • Still open: what the position value means (sharpness jumps exactly at 0x2000) and the valid range from the module calibration. Nothing is written to the EEPROM.
    • Plan: a small V4L2 lens driver with V4L2_CID_FOCUS_ABSOLUTE, so libcamera's contrast AF can drive it.

    If someone is already on this, tell me and I'll hand over my notes instead.

  22. W4HO-ham commented on Nov 10, 2026

    @W4HO-ham
    ContributorAuthor

    Where the 9a's three open rows stand after today (2026-11-09), in the order I think they merge, so it's in one place. Everything you asked for on these is in unless it says otherwise.

    Suspend

    1. linux-omarchy-mobile#224, the UFS pin state at resume. On the published tree with it: 36 of 36 with modules, 36 of 36 with Wi-Fi up, a Power-key wake, no UFS line in 73 sleeps.
    2. linux-omarchy-mobile#236, the main PMIC's interrupt. Without it the published kernel has no RTC alarm on the 9a.
    3. omamobile-device-tegu: keep the cores out of C2 through a sleep (the 9a's silent no-wake) #591, the sleep hook that keeps C2 out of the sleep. Your two points are in; 40 of 40 through the hook itself tonight.
    4. device/tegu SLEEP-PLAN, bringup: the two 9a suspend failures found and fixed (kernel #224) #409, the plan, the AGENTS wording and the series scripts.

    After 1 and 2 are on omamobile I'll run a series from that tip as published, for the cell.

    GNSS and the modem, from the published kernel's own drivers

    1. linux-omarchy-mobile#237, the GNSS nodes.
    2. kernel config: the zuma-family PCIe PHY (CONFIG_PHY_ZUMA_PCIE) #274, CONFIG_PHY_ZUMA_PCIE=m; the PHY driver it waited for landed with kernel device/husky: UFS-PLAN: U4, the Omarchy image boots #187.
    3. device/tegu, omamobile-device-tegu: the Pixel 9a's modem and GNSS from the published kernel's own drivers #594, the modem-cpif overlay, the GNSS link driver in the config and the module guards. With 5 to 7 on a published-style build: four boots by the boot unit on the in-tree drivers, confirmed stops, GNSS fixes in four sessions, a call, data, the NV partitions unchanged.
    4. omamobile-device-tegu 0.2.12: start the Pixel 9a's modem and phone service at boot, opt-in #372, the boot unit, opt-in. Waiting for your 7 Pro run of the shared stop path.
    5. omamobile-device-tegu: the 9a's modem start skips where it can't run, and the record of it running by itself (opt-in kept; stacked on #372) #590 on top of it: the skip conditions and the record, opt-in kept as you asked.
    6. reboot-to, omamobile-modem: a deliberate restart doesn't cost the next boot its modem #403 and omamobile-mmd 0.1.3: wait for confirmed call release #461: both waiting for your 7 Pro runs; reboot-to, omamobile-modem: a deliberate restart doesn't cost the next boot its modem #403 is merged up to main again, omamobile-mmd 0.1.3: wait for confirmed call release #461 has its real call on the 9a.

    On by default for the modem, and with it GNSS, waits for the slot B decision (#484, then #483). The palmtop is available as a second device for both whenever that helps.

  23. W4HO-ham commented on Nov 13, 2026

    @W4HO-ham
    ContributorAuthor

    A question about the 9a's row, so I work on the right things. Three of its cells read "testing": GNSS, the modem and suspend. By the legend, "works" is verified on the phone and on by default. Here is what I think stands between each cell and that; please correct what I have wrong or have missed.

    Modem. Calls, texts and data work; pictures come with #697. It isn't on by default: the boot unit is opt-in while marking slot B on the 9a is on hold, and that waits for #483, which waits for #484. Is that the whole of it for "works", with the call echo (#374) as a named gap, or does the echo have to go first?

    GNSS. It gets fixes through the Location switch once the modem is up (104 of 109 on kernel #247's test build today, 78 of 81 on the daily kernel). Two things I know of: the repository's package has no HAL client for the 9a (#748, with #749 so the next such package says so), and it depends on the modem's start, so it is on by default when the modem is. Anything else?

    Suspend. With kernel #224, #591 and now #247 the three known causes of a bad sleep are fixed, and #247's build woke 300 of 300. What I think is left: a series on the first published kernel that has #247; auto-sleep, which ships off; and sleep with the modem up, where one attempt leaves CPIF unable to reach the CP until a reboot ("Can't set D0 state"), so the modem holds its inhibitor. The 7 Pro's cell is "testing" too, so I may be misreading what this cell can reach on the 9a before the 7 Pro's does. What would you want to see for it?

    For the modem-up sleep I'd start with the CPIF resume path on the 9a, offline first, then attended tests that end in a reboot, as the rule says. If #484 would move faster with a second pair of hands (a rebase, a run on the 9a), say so.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions