Skip to content

Pixel 7a (lynx, gs201 IPOP) port: stock facts and first light #36

Open

Description

@latentoperator

I have a Pixel 7a I can dedicate to this (a retired phone, fine to wipe) and I'm starting the lynx port. DEVICES.md lists it under Wanted ("as panther"); opening this so efforts don't overlap.

Device

  • Model: Pixel 7a, codename lynx (ro.product.device), SKU GWKK3, hardware MP1.0
  • SoC: gs201 (Tensor G2), ro.boot.hw.soc.rev 2, pkg_mode 1 (IPOP)
  • Stock build: CP2A.260705.006 (Android 17, 2026-07-05), bootloader lynx-17.0-15199429
  • Bootloader: unlocked. Stock is on slot A (Google's factory image, verified SHA-256), slot B is free for Omarchy Mobile
  • Root: Magisk on slot A's init_boot only
  • Display panel: sdc-s6e3fc5 (exynos_drm.panel_name)

Done so far (read-only, CONTRIBUTING §5.1)

  • bin/omamobile capture on stock, without root
  • fastboot getvar all: same boot image set as cheetah (boot, init_boot, vendor_boot, vendor_kernel_boot, dtbo)
  • Rooted capture: live DT, dmesg, /proc/iomem, interrupts, modules, vendor firmware list. Captures stay in ~/android/reference/
  • Board dtbo: ro.boot.dtbo_idx 9 → board_id 0x30506, board_rev 0x10000 (matches ro.boot.cdt_hwid); vendor_boot page size 0x800

One difference from cheetah and felix: DTB selection

lynx's SoC is the IPOP package. The stock vendor_kernel_boot carries three FDTs (soc_id 0x9855000, soc_id 0x9855001, and one with /soc_compatible/B0_IPOP { product_id = <0x9855>; major = <1>; minor = <0>; pkg_mode = <1>; }), and the live DT on this phone is the third, so ABL matched it by /soc_compatible, not /soc_id. That matches the b0_v2-ipop row in cheetah's BRINGUP-M1 §1.

bin/omamobile images only knows SOC_IDS today; tegu's descriptor has the same need as a commented-out SOC_COMPATIBLE idea. My plan is to propose that key (a /soc_compatible node written into the DTB copy per entry), so lynx and tegu can share it. Unless ABL turns out to fall back to the soc_id B0 DTB on an IPOP chip, which I'll check on slot B first since a failed boot just returns to slot A.

Plan

  1. Build setup: the small AOSP clang + mkbootimg + libufdt checkout, linux-omarchy-mobile (omamobile-int)
  2. gs201-lynx.dts: USB, the bootloader's framebuffer, watchdogs; everything else disabled (following felix's first light)
  3. SOC_COMPATIBLE support in bin/omamobile images (or confirm the soc_id fallback)
  4. device/lynx/device.conf, lynx-m1-overlay.dts, NOTES.md
  5. First light on slot B: SSH at 172.16.42.1

Questions for maintainers

  • Is a SOC_COMPATIBLE descriptor key the shape you'd want, or would you rather handle IPOP some other way?
  • Anything else you'd want captured while stock is on the phone?

I'm working with an AI coding assistant (Claude Code); I'll follow AGENTS.md and the kernel's coding-assistants.rst (human sign-off, Assisted-by:).

Activity

  1. latentoperator commented on Oct 24, 2026

    @latentoperator
    ContributorAuthor

    First light on lynx. Slot B boots the omamobile-int kernel (7.3.0-rc6) and SSH comes up at root@172.16.42.1 over USB NCM. Slot A is still stock.

    What it took:

    • DTB: one copy with /soc_compatible/B0_IPOP (omamobile images: SOC_COMPATIBLE and AVB_FOOTERS for zuma (and zumapro) #27's SOC_COMPATIBLE="b0-ipop:0x9855:1:0:1"). ABL took it. No soc_id copy was needed.
    • AVB footers on boot and init_boot. This was the actual blocker. Without them ABL never showed the orange warning for slot B: black screen, a few retries, then back to slot A, with nothing on USB and nothing in pstore. Putting stock's own IPOP DTB in our vendor_kernel_boot failed the same way, which ruled out the DTB. With omamobile images: SOC_COMPATIBLE and AVB_FOOTERS for zuma (and zumapro) #27's AVB_FOOTERS="boot:0x4000000 init_boot:0x800000", AVB_ROLLBACK_INDEX=1783209600 and AVB_SECURITY_PATCH=2026-07-05 (stock's values from avbtool info_image), it booted on the next try. Stock lynx signs boot and init_boot with their own RSA4096 keys, and vbmeta chains them. So omamobile images: SOC_COMPATIBLE and AVB_FOOTERS for zuma (and zumapro) #27's footer support is needed on gs201 too, not just zuma, at least with this bootloader (lynx-17.0-15199429).
    • Board dtbo: empty overlay, board_id 0x30506, board_rev 0x10000 (dtbo index 9).

    From the first boot:

    • Both cluster watchdogs are stopped by s3c2410_wdt at probe, so it stays up past 60 s.
    • ABL adds its own sec_dram (0xb7200000), sec_pt (0xe0000000) and pkvm_guest_firmware reservations to the mainline DTB. gs201.dtsi doesn't have them, so I also listed them in the board file, along with tui, mfc_fw and famodel from stock's tree. ABL placed pkvm_guest_firmware at 0xb5000000 on our boot, not 0x8b000000 as on stock.
    • DPP0's RDMA base (0x1c0b0040) reads 0xfac00000, same as cheetah. On stock that address is inside the wifi_region CMA pool, so the board file needs to reserve it before adding a simple-framebuffer. That's my next step, along with the DECON0 trigger.
    • cma: Failed to reserve 384 MiB from the shared config, which doesn't matter at this stage.

    PRs coming: device/lynx here (descriptor, empty dtbo, notes), stacked on #27, and gs201-lynx.dts against linux-omarchy-mobile omamobile.

  2. latentoperator commented on Oct 24, 2026

    @latentoperator
    ContributorAuthor

    PRs are up: derekross/linux-omarchy-mobile#43 for the board file and #37 here for device/lynx (stacked on #27). Both ran on the phone from slot B.

  3. latentoperator commented on Oct 25, 2026

    @latentoperator
    ContributorAuthor

    Status update, and a Wi-Fi question for you, Derek.

    Since first light (all on the phone, slot B): UFS with cheetah's nodes, an Arch root on userdata, Hyprland on simpledrm, touch (Goodix GT9895 over SPI, needs two S2MPG12 LDOs that the lynx bootloader leaves off), the PMIC RTC, and the boot-success marker. Branches for UFS, touch and the device package are ready and will follow #43/#44/#37/#38 once those land. Notes for every step are in device/lynx/NOTES.md.

    Wi-Fi is where I'm stuck. The 7a does not use cheetah's Broadcom part. It has a Qualcomm QCA6490 (17cb:1103, subsystem 17cb:0108) on PCIe channel 1. Linux sees it as WCN6855 hw2.1 and drives it with ath11k.

    What works, on a local branch omamobile-gs201-lynx-wifi that imports your gs201 PCIe host, PHY and SysMMU drivers from omamobile-int:

    • link Gen2 x1, enumeration, config space, BAR access, 32 MSI vectors through iMSI-RX
    • MHI power-on; the chip DMA-reads the first 0x1248 bytes of amss.bin (ELF header plus the MBN hash segment header) and reacts to their content

    What does not: the chip's boot ROM then stops at a fixed stage (TCSR_PBL_LOGGING = 0xef0a0002). BHI_STATUS stays 0, no MSI, no AER or completion timeout, no pending transaction on the endpoint, link in L0 the whole time. After 20 s MHI gives up with -110. Stock finishes the same load in 57 ms.

    Ruled out so far, each with a test on the phone: firmware content and signing (the Pixel's own amss20.bin stalls identically, and both images are unsigned, version 5 MBN with hashes only), the PCIe PHY rails (S2MPG13 LDO3S/LDO18S are on), the PMIC 32 kHz outputs, link power states, cache coherency, Relaxed Ordering and No Snoop, read request size, SysMMU bypass vs firewall, the S2MPU, power-up timing and a full power cycle, BT_EN, the code segment sizes (an image with 256-byte segments stalls the same), and every register write ath11k makes before the MHI power-up. On that last point: I compared Google's cnss2 with ath11k and cnss2 writes nothing to the chip before MHI, while ath11k runs a hot-reset workaround, interrupt clear, QFPROM power control, BHIe vector clear and an MHICTRL reset. Skipping all of them changes nothing.

    The same chip family boots fine with upstream ath11k on Qualcomm hosts (X13s, sm8450-hdk, and the sm8350-hdk WCN6851 series from June). So what is left is the gs201 host path.

    The question: compared with Google's pcie-exynos-rc.c, what does the imported pcie-gs201 driver leave out or do differently for an endpoint that boots its own firmware by DMA? Stock runs channel 1 with the SysMMU translating (image buffers around 0xfa880000), a 36-bit DMA mask, use-ia on, L1.2 on, use-cache-coherency = "false", and three MHI MSI vectors. If any of those rings a bell as something the host side has to get right for this chip, that is where I would look next. I can also dump any root port or PHY register you want to compare.

    No PR for Wi-Fi until it boots. The full write-up with the register bisect table is in device/lynx/NOTES.md on the lynx-device branch.

  4. latentoperator commented on Oct 25, 2026

    @latentoperator
    ContributorAuthor

    Status update, and a Wi-Fi question for you, Derek.

    Since first light (all on the phone, slot B): UFS with cheetah's nodes, an Arch root on userdata, Hyprland on simpledrm, touch (Goodix GT9895 over SPI, needs two S2MPG12 LDOs that the lynx bootloader leaves off), the PMIC RTC, and the boot-success marker. Branches for UFS, touch and the device package are ready and will follow #43/#44/#37/#38 once those land. Notes for every step are in device/lynx/NOTES.md.

    Wi-Fi is where I'm stuck. The 7a does not use cheetah's Broadcom part. It has a Qualcomm QCA6490 (17cb:1103, subsystem 17cb:0108) on PCIe channel 1. Linux sees it as WCN6855 hw2.1 and drives it with ath11k.

    What works, on a local branch omamobile-gs201-lynx-wifi that imports your gs201 PCIe host, PHY and SysMMU drivers from omamobile-int:

    • link Gen2 x1, enumeration, config space, BAR access, 32 MSI vectors through iMSI-RX
    • MHI power-on; the chip DMA-reads the first 0x1248 bytes of amss.bin (ELF header plus the MBN hash segment header) and reacts to their content

    What does not: the chip's boot ROM then stops at a fixed stage (TCSR_PBL_LOGGING = 0xef0a0002). BHI_STATUS stays 0, no MSI, no AER or completion timeout, no pending transaction on the endpoint, link in L0 the whole time. After 20 s MHI gives up with -110. Stock finishes the same load in 57 ms.

    Ruled out so far, each with a test on the phone: firmware content and signing (the Pixel's own amss20.bin stalls identically, and both images are unsigned, version 5 MBN with hashes only), the PCIe PHY rails (S2MPG13 LDO3S/LDO18S are on), the PMIC 32 kHz outputs, link power states, cache coherency, Relaxed Ordering and No Snoop, read request size, SysMMU bypass vs firewall, the S2MPU, power-up timing and a full power cycle, BT_EN, the code segment sizes (an image with 256-byte segments stalls the same), and every register write ath11k makes before the MHI power-up. On that last point: I compared Google's cnss2 with ath11k and cnss2 writes nothing to the chip before MHI, while ath11k runs a hot-reset workaround, interrupt clear, QFPROM power control, BHIe vector clear and an MHICTRL reset. Skipping all of them changes nothing.

    The same chip family boots fine with upstream ath11k on Qualcomm hosts (X13s, sm8450-hdk, and the sm8350-hdk WCN6851 series from June). So what is left is the gs201 host path.

    The question: compared with Google's pcie-exynos-rc.c, what does the imported pcie-gs201 driver leave out or do differently for an endpoint that boots its own firmware by DMA? Stock runs channel 1 with the SysMMU translating (image buffers around 0xfa880000), a 36-bit DMA mask, use-ia on, L1.2 on, use-cache-coherency = "false", and three MHI MSI vectors. If any of those rings a bell as something the host side has to get right for this chip, that is where I would look next. I can also dump any root port or PHY register you want to compare.

    No PR for Wi-Fi until it boots. The full write-up with the register bisect table is in device/lynx/NOTES.md on the lynx-device branch.

  5. derekross commented on Oct 26, 2026

    @derekross
    Owner

    Thanks, @latentoperator. UFS, the Arch root, touch and Hyprland on a second gs201 phone is great news, and the Wi-Fi bisect is very thorough.

    I can't name the missing piece yet. Here is what I can say from the code.

    Where our driver hasn't been tested. Cheetah's BCM4389 never fetches its own firmware: brcmfmac writes it into the chip's RAM from the host. So pcie-gs201 has only carried the chip's ring traffic, never a boot ROM's fetch. It also has no MSI at all. msi_init is a stub and cheetah runs on INTx, so the iMSI-RX path is all yours. It also leaves out L1 substates and Google's SysMMU DMA ops. The I/A sequencer is ported.

    Two things in Google's lynx sources:

    1. Everything below 4 GiB. For EP_QC_WIFI only, dwc-whi/pcie-exynos-rc.c drops the controller's DMA mask to 32 bits around dw_pcie_host_init() ("devices only work with 32-bit MSI"). Stock's wifi_region uses alloc-ranges = <0x0 0x80000000 0x80000000>. A restricted-dma-pool ignores ath11k's 32-bit coherent mask (swiotlb_alloc() never checks it), so if you carried cheetah's pool over, only its alloc-ranges keeps the BHI buffer low. Could you check that BHI_IMGADDR_HIGH and the MSI target's high word are both 0?
    2. wlan-buck-gpio, gph2-5. gs201-lynx-wlan.dtsi drives it high, and cnss2's wlan_buck_enable() (power.c) raises it before the rails and WLAN_EN. On cheetah, gph2-5 is the dev-wake input with a pull-down. If your board file kept cheetah's pin setup, the chip may be running without that buck. That's a guess, but it's cheap to rule out.

    Suggested order, one change per boot:

    • 0, read-only: gph2-5's pin function and level, BHI_IMGADDR_HIGH/LOW and the root port's MSI address. Read them through the driver or lspci. Use devmem only on blocks you know are powered.
    • 1: gph2-5 high (a regulator-fixed in the chip's supply chain), nothing else changed.
    • 2: the pool at stock's range, if step 0 showed anything above 4 GiB.
    • 3, after the BHI load passes: board data. linux-firmware's WCN6855/hw2.1/board-2.bin has 17cb:1103/17cb:0108 entries, but they're keyed by QMI board id, and the Pixel's antennas and regulatory settings are in its own board file. Take that from the owner's factory image, as you did with amss20.bin. It's never committed or hosted, and neither are the firmware files.

    Safety: the chip sits behind the same HSI2 S2MPU as UFS, so until it's confined it can DMA anywhere in RAM. Keep the node disabled and the modules blacklisted by default, and stream the log to the laptop with --watchdog armed. Run the backup checksum check before step 1, because that step turns on a rail. Nothing reads or writes persist or efs*.

    One note for your plan: now that the Arch root is on userdata, slot A's Android stops at recovery. So the fallback after three retries and omamobile fastboot (which needs adb from Android) can't get you to fastboot any more. From here on, the buttons are the way back: Power about 30 s, then Volume Down + Power. Please write that into lynx's "how to get back" section.

    I'd be glad to look at any register dumps you post.

  6. FlorencioJP commented on Oct 26, 2026

    @FlorencioJP
    Contributor

    Here is an update from testing on Pixel 7a hardware (lynx, GS201 IPOP, slot B running Omarchy Mobile / 7.3.0-rc6):

    Following up on @derekross's points regarding power rails, addressing, and PBL registers:

    1. Power Rails (gph2-5 / WLAN_BUCK & gph2-4 / WLAN_EN)

    • We applied a DT overlay defining the regulator chain (lynx_wifi_buck on gph2-5 driving lynx_wifi_enable on gph2-4).
    • Direct inspection via debugfs pinctrl confirms both pins are actively driven HIGH:
      • pin 4 (gph2-4): CON(0x1) DAT(0x1) (WLAN_EN = 1)
      • pin 5 (gph2-5): CON(0x1) DAT(0x1) (WLAN_BUCK = 1)
    • The chip remains powered with the buck active before and throughout link training and MHI power-up.

    2. DMA Addressing (< 4 GiB) & MSI

    • 0000:01:00.0 has a 32-bit MSI capability (64bit-). 32 MSI vectors are successfully allocated via iMSI-RX.
    • BHI_IMGADDR_HIGH is 0x0.
    • The coherent buffers (wifi-coherent-pool and wifi-dma-pool) are bounded in <0x80000000 0x80000000> (< 4 GiB).
    • Note: In mainline DT, wifi-coherent-pool is marked no-map; (unlike Android stock's standard reusable CMA), which results in MEMREMAP_WC / non-cached mappings.

    3. SysMMU, S2MPU & iATU Status

    • PCIe 1 SysMMU is set to bypass (CTRL = 0x24), with HSI2 clocks held active.
    • S2MPU fault status (0x145e2000) reports 0x0 (0 faults detected).
    • Inbound iATU Region 0 was explicitly programmed for 1:1 mapping of 0x80000000..0xffffffff.

    4. PBL Register Dump at Timeout

    The chip's boot ROM (PBL) successfully performs DMA reads of the first 0x1248 bytes of amss.bin (ELF header + MBN hash header), but halts immediately after:

    • TCSR_PBL_LOGGING (0x01b000f8) = 0xef0a0002
    • BHI_STATUS = 0x0 (remains 0 until 20s timeout -110)
    • PBL_READ 0x1403da0 = 0xef0a0002
    • PBL_READ 0x1403db4 = 0x42100000
    • PBL_READ 0x1403dc4 = 0xef000008

    Current hypothesis: the stall occurs either when attempting to load the main code segments (inbound iATU window sizing or DMA attribute mismatch) or due to cache coherency / AxCACHE settings on the GS201 PCIe root complex (stock uses use-cache-coherency = "false").

  7. derekross commented on Oct 26, 2026

    @derekross
    Owner

    Thanks, @FlorencioJP, this is a useful dump. With the buck high, MSI working and the addresses low, the three things I asked about are ruled out.

    What the boot ROM's read tells us. The ROM fetched 0x1248 bytes, the ELF header plus the hash segment, and then stopped. In a secure boot that's the point where it checks the signature before it loads any code segments. So I'd look at the image before the iATU or cache settings. Two cheap checks, both read-only:

    1. Is the image in the BHI buffer the one stock loads? Compare the sha256 of the file ath11k actually requested with the one in the owner's factory image (stock's amss20.bin and any other amss*/m3* files there). Then compare BHI_IMGSIZE with that file's size. On a full-boot-chain chip MHI hands the ROM only the first sbl_size bytes through BHI and the rest later through BHIe. A size that doesn't match the file, or a file from a different board revision, would also stop it right after the hash segment.
    2. What the ROM's log codes mean. 0xef0a0002 repeats in TCSR_PBL_LOGGING and at 0x1403da0. If you have stock's cnss2 log from a normal boot (dmesg on stock with root, or a bugreport), the same registers at the same stage would show which value means "authentication passed". That's the comparison I'd trust over a guess.

    Coherency is still possible: stock does set use-cache-coherency = "false" and our no-map pool maps write-combined. But a ROM that reads the header and hash segment correctly is getting coherent data at least that far.

    On how you got numbers 3 and 4, a safety question. "PCIe 1 SysMMU set to bypass", "inbound iATU region 0 explicitly programmed" and the S2MPU status read: were those done by the kernel driver, or from userspace with devmem? Please write it into the plan either way. The HSI2 SysMMU, the S2MPU and the PCIe controller are blocks where a userspace read has hung the SoC on cheetah. A write there also changes what the chip can reach while UFS shares the same path. If any of it was devmem, please move it into the driver (or an overlay) so it happens in one known place at probe, and leave the S2MPU alone until the chip is confined.

    Keep going one change per boot, with the node disabled by default. I'm glad to read the next dump.

  8. FlorencioJP commented on Oct 26, 2026

    @FlorencioJP
    Contributor

    Thanks @derekross. Here are the two read-only checks you asked for, plus a bisection of the ROM's log code done by feeding ath11k modified copies of the image (served through firmware_class.path, nothing written to the phone's firmware directory). All of it on the same Pixel 7a, native kernel, one change per load.

    1. Image and size

    • The file ath11k requests (ath11k/WCN6855/hw2.1/amss.bin) is byte-identical to stock's /vendor/firmware/amss20.bin, sha256 b4d284416002ba1aa85c5ca0d3e94a1b895f9f72aece3d59f74391f563facee4; m3.bin matches too. It is the only amss* in stock's /vendor/firmware, and stock's kernel cmdline has firmware_class.path=/vendor/firmware. All 17 segment hashes in the image verify. The hash segment is MBN version 5 with sig_size = 0 and cert_chain_size = 0.
    • BHI_IMGSIZE reads 0x80000 = sbl_size (512 KiB), the same value cnss2 sets; BHI_IMGADDR is the CRC-checked buffer at 0xefd00000. BHI_ERRCODE stays 0; ERRDBG1/ERRDBG3 just echo the address and size.

    2. What 0xef0a0002 means, empirically

    Change to the image (one byte unless noted) TCSR_PBL_LOGGING
    none (control, served the same way) 0xef0a0002
    ELF magic destroyed 0xef128100
    e_phnum = 0 0xef0e0200
    hash-segment phdr without its hash type flag 0xef120200
    hash segment shrunk to 16 / 256 bytes, hash[0] recomputed 0xef0a0125 / 0xef0a0118
    hash[0] corrupted; hash[2] corrupted; a phdr byte flipped; MBN version 5→0x55 0xef0a0002 (no change)
    hash segment moved into page 0 (p_offset 0x800), hash[0] recomputed 0xef0a0002
    hash segment p_paddr 0xf001f000 → 0x01424000 (SRAM), hash[0] recomputed 0xef0a0002, SRAM at 0x1424000 still zero

    So the ROM parses the ELF header and the program headers from our buffer correctly, checks the hash segment's size, then fails at stage 0x0a with status 0002 regardless of the hash segment's content or location.

    3. Does it read the hash segment at all? Inbound iATU region 2 programmed with a UR response code (DWC RESPONSE_CODE bits), everything else pass-through:

    • block [BHI+0x1000, +0x80000) → 0xef0a011f, root port DevSta: UnsupReq+
    • block [BHI, +0x1000) → 0xef128100 (same as bad magic), UnsupReq+
    • block [BHI+0x2000, +0x80000) → 0xef0a0002, no UR at all

    The ROM issues exactly one read beyond the header, into +0x1000..+0x1248, that read completes normally, and it never reads +0x2000 or writes any SRAM. Endpoint MaxReadReq forced to 128 or 256 bytes before the load: no change.

    4. Other things ruled out here: the seven S2MPG12/S2MPG13 rails stock keeps regulator-always-on but the bootloader leaves off (LDO20M, LDO9S, LDO10S, LDO11S, LDO15S, LDO19S, LDO20S; enabled through ACPM with voltages unchanged, then restored) — no change. Bluetooth: in stock's dmesg bt_power only turns BT on at 50.9 s, the Wi-Fi load succeeds at 13.4 s. Chip registers after the stall: WLAN_SECURE_BOOTN 0x11111111, QFPROM_RAW_OEM_SECURE_ROW0_LSB 0x0, WLAN_JTAG_ID 0x1017b0e1, PBL_WLAN_BOOT_CFG 0x0, PBL_BOOTSTRAP_STATUS 0x0, WLAON_QFPROM_PWR_CTRL 0x62; 0xf0000000/0xf001f000 read 0xffffffff through the window; a scan of SRAM 0x1400000-0x1800000 finds neither the ELF header nor the hash table.

    Stock's cnss2 does not print these registers on a successful boot, so I have no "authentication passed" value to compare with.

    5. Safety: the SysMMU bypass, the inbound iATU regions and the IOCC bits are written by a kernel module (sysmmu-identity.ko, loaded after link up, before ath11k), not from userspace. The S2MPU was only read, never written. One /dev/mem read of the iATU registers dropped the SSH session (the phone stayed up); it was not repeated and the notes say so.

    @latentoperator, you mentioned a register bisect table in device/lynx/NOTES.md on your lynx-device branch; it is not on GitHub. Could you push it, or paste the PBL codes you saw? In particular: did you ever see a code other than 0xef0a0002 with the unmodified image, and what did "SysMMU firewall" and "BT_EN" look like in your runs?

  9. FlorencioJP commented on Oct 26, 2026

    @FlorencioJP
    Contributor

    Follow-up, with what the ROM itself says. All read-only, kernel-side, node disabled by default.

    I kept going one change per boot and the host side is exhausted:

    • PERST# released ~20 ms after WL_REG_ON (dropped the enable regulator's startup-delay-us from 200 ms to 10 ms): link trains Gen2 x1, same stall.
    • Same, without ath11k's pre-MHI soc reset: same stall.
    • CMU_HSI2 automatic clock gating cleared before the controller (cmu_acg_off=1, as Google's driver does): same stall.

    So I read the code out of the chip. After the stall I copied 128 KiB of the boot ROM at 0x00800000 through the BAR window (kernel module, read-only, freed on reboot) and disassembled it as Hexagon v62. The PBL keeps its source file names, which label every stage: pbl_cache_mmu.c, pbl_auth.c, pbl_ram_init.c, pbl_hw_init.c, pbl_elf_loader.c, pbl_flash_pcie.c, pbl_common.c, plus the strings USE_SERIAL_NUMBER_IN_SIGNING and ROOT_CERT_SEL.

    The two ath11k_info stop sites I log map to pbl_auth.c: the "bad magic" path logs pbl_auth.c line 0x1c6 with 0x128100, and the normal path reaches pbl_auth.c line 0x1ba. In other words stage 0x0a in 0xef0a0002 is the authentication step in pbl_auth.c, not a DMA or ELF-load step. The routine that runs just before the stop reads WLAN_SECURE_BOOTN (0x01E22B24) twice, masks bit 0, and compares the two reads; nearby code also reads 0x1E22B2C/0x1E22B3C and selects a root cert. On this phone WLAN_SECURE_BOOTN reads 0x11111111 after the stall and QFPROM_RAW_OEM_SECURE_ROW0_LSB reads 0.

    Put together with the earlier bisection (the ROM DMA-reads exactly the header + hash segment +0x1000..+0x1248, that read completes on the PCIe side, and nothing past +0x2000 is ever read): the chip gets the hash segment fine and then makes an authentication decision that goes the other way than it does under stock, with the same unsigned image that stock loads in 57 ms. That points away from the host (iATU, <4 GiB, pool type, coherency, MRRS, PERST/clock ordering are all ruled out now) and toward the chip's own internal state before BHI — something cnss2 or the chip PMU sets up that our ath11k path doesn't.

    Two questions where you'd know better than me:

    1. On a stock boot, what does WLAN_SECURE_BOOTN (0x1E22B24) read before the first BHI? If it's 0x0/0x1 there and 0x11111111 here, that masked bit-0 compare in pbl_auth.c is the likely branch.
    2. Does cnss2 (or wlan_firmware_service/the PMU) write anything into the chip's WLAON/secure block before mhi_sync_power_up() that ath11k never issues? That's the gap I can't see from our side.

    I won't post the ROM or the dump; it's Qualcomm's. Happy to share specific disassembled routines or register reads if that helps.

  10. derekross commented on Oct 26, 2026

    @derekross
    Owner

    Thanks, @FlorencioJP. The bisection is very good work: the ROM changing its code with the hash segment's size but not its content, the UR test that pins down exactly which reads happen, and tracing stage 0x0a to pbl_auth.c. Thank you also for keeping the ROM and the dump private.

    I went through Google's cnss2 for lynx (gs-6.1 wlan/wcn6740/cnss2, the copy lynx's BUILD.bazel builds, and its own MHI core in cnss2/mhi/) for both your questions. Stock may run an older copy, but these paths are generic Qualcomm code.

    1. WLAN_SECURE_BOOTN on a stock boot. I can't give you the value: the source never logs it on a good boot. cnss2 reads it only in cnss_pci_dump_bl_sram_mem() (pci.c:1411-1421), and only from the 7 s boot-debug timer or a failure path. A 57 ms load never gets there, which is why your stock dmesg doesn't have it. Nothing in cnss2 writes it.

    Stock does have a way to read it before BHI: as root with debugfs, echo on and then echo linkup into cnss/dev_boot, followed by echo "0xFE 0x1E22B24 4" > cnss/reg_read (debug.c:212, :244, :381-389). That needs a working stock Android, though, and yours stops at recovery now that userdata holds Arch. Please don't wipe userdata for this. It's only worth doing if someone has a second, rooted 7a.

    A guess, clearly labelled as one: Google's fw-api headers have a SECURE_BOOT fuse register for another Qualcomm chip, the BT fuse in fw-api/hw/qca6750/v1/lithium_top_reg.h:66880. It uses one nibble per image, with bit 0 AUTH_EN and bit 1 USE_SERIAL_NUM, which matches the USE_SERIAL_NUMBER_IN_SIGNING/ROOT_CERT_SEL strings you found. If WLAN_SECURE_BOOTN uses that layout, 0x11111111 means "authentication required" for every image, and an image with sig_size = 0 would stop exactly at the auth step, whatever its hashes say. But if it's a real fuse, stock reads the same value and can't be loading an unsigned image either. So either the register isn't stable on your power-up, or the ROM authenticates against something other than the v5 header fields. I can't tell which from here.

    2. Does cnss2 write anything before mhi_sync_power_up()? No, your reading is right. On lynx:

    • Power: wlan_buck_enable() raises gph2-5 and sleeps 100 ms. WLAN_EN (gph2-4) goes high, then a 1 ms delay. The regulator and clock lists are empty under CONFIG_WCN_GOOGLE, and there is no BT_EN, XO-clock or bootstrap GPIO in lynx's DT.
    • Link: exynos_pcie_pm_resume(), then D0, config-space restore and bus master. Config space only.
    • WLAON: cnss_pci_set_wlaon_pwr_ctrl() is the only WLAON write, and it's gated on qcom,set-wlaon-pwr-ctrl, which lynx doesn't set (main.c:3521-3529). For QCA6490, cnss_pci_config_regs() only picks a dump table.
    • MHI: the writes before the ROM's fetch are BHI_INTVEC=0, the RDDM BHIe RX vector setup, and BHI_STATUS/IMGADDR/IMGSIZE/IMGTXDB. Mainline MHI with ath11k does the same set, with the same rddm_size (0x420000), sbl_size/seg_len (512K) and fbc_download. Downstream MHI also reads BHI_OEMPKHASH[0..15] before the load; mainline doesn't.
    • wlan_firmware_service: it speaks QMI over MHI channels, so it can't reach the chip until AMSS is up.

    One difference in the shape of the power-up: stock enumerates once at boot, powers the chip off again by dropping only WLAN_EN (the buck stays high), and does the real load on a second WLAN_EN cycle. If your overlay's chain drops the buck together with the enable, your power cycle isn't the same event as stock's.

    Cheap read-only checks I'd do next, one per boot:

    1. Read WLAN_SECURE_BOOTN twice back to back right after link-up, before MHI power-up, then twice again at the stall. A value that changes points at the chip's power-up state rather than at the image.
    2. At the stall, from your module, read the fields the PBL publishes in BHI: BHI_EXECENV (+0x28) and BHI_OEMPKHASH(0..15) (+0x64…). Please post only whether the hash is zero, not its value. Skip BHI_SERIALNU (+0x40), because it identifies the chip. Don't use /sys/bus/mhi/devices/*/oem_pk_hash after ath11k gives up: mainline sets mhi_cntrl->bhi = NULL on unprepare, and that file then reads through a NULL pointer.
    3. Grep your stock dmesg for cnss:. cnss2's debug lines are plain KERN_DEBUG printks, so you should see the BUCK, WLAN_EN and "Starting image download via BHI" lines with timestamps. That gives stock's real gap from WLAN_EN to the first BHI to compare with yours.

    The usual rules for anything that reads the QFPROM block: don't post raw fuse rows (they hold the serial and key hashes), keep the node disabled by default, and keep --watchdog armed.

    I'm glad to read whatever these show.

  11. FlorencioJP commented on Oct 26, 2026

    @FlorencioJP
    Contributor

    Thanks, Derek — this narrows it. Check 2, read-only at the stall, no reboot: BHI_EXECENV = 0 (still in PBL) and all 16 BHI_OEMPKHASH words read zero. With QFPROM_RAW_OEM_SECURE_ROW0_LSB = 0 and the unsigned sig_size=0 image, nothing here says this part demands a signed image. The outlier is WLAN_SECURE_BOOTN = 0x11111111: a repeated-nibble pattern that looks more like an uninitialised power-up value than a provisioned fuse, which fits your "the register isn't stable on your power-up" branch.

    Check 1 is where I have to stop: reading 0x1E22B24 through the register window right before mhi_sync_power_up() hangs the HSI2 bus and reboots the phone (watchdog returns it to slot B). At the stall that same read is fine; pre-power-up it is not. I won't retry it without a safer path, and I won't wipe userdata for a stock read. So check 1 needs a second rooted 7a, as you said.

    I also pulled the stock timing from the bugreport I already had (no new boot). From wlan_buck_enable: WLAN_EN at +109 ms, first BHI at +167 ms, SBL +270 ms, AMSS +935 ms — then power-off and a second WLAN_EN cycle at +8.8 s. So stock's very first cold cycle already passes auth; the second cycle is mission mode. That points at the chip's internal state at its first power-up rather than a missing prior cycle. If downstream MHI reads BHI_OEMPKHASH[0..15] before the load and mainline doesn't, could that read be a required handshake the PBL waits on? That's the one host-side difference left that I can act on.

  12. FlorencioJP commented on Oct 26, 2026

    @FlorencioJP
    Contributor

    Closing the one lead I left open above: I made our MHI read BHI_OEMPKHASH[0..15] right before the BHI download, as the downstream core does (mainline doesn't). The read runs (OEMPK_PREREAD done, zero=1, all sixteen words zero) and the PBL stalls exactly the same, -110, no SBL. So that pre-read is not a handshake the ROM waits on. With that, every host-side difference I can act on is tested and negative; what's left needs a stock-side read of WLAN_SECURE_BOOTN before the first BHI, which on this phone can't be done safely. Waiting on your next read.

  13. derekross commented on Oct 26, 2026

    @derekross
    Owner

    Thanks, @FlorencioJP, for both updates, and sorry about check 1: I asked for that read, and it was the wrong point to read from. Before MHI power-up the chip side of that window evidently isn't safe to touch, and you did the right things: the watchdog was armed, you stopped after one hang, and you didn't retry. Please put it in your notes as a stop condition ("no windowed chip register reads before mhi_sync_power_up()"), so nobody repeats it.

    The BHI_OEMPKHASH pre-read test closes the last host-side difference I knew of. Thank you for running it.

    I also owe you a correction to my guess about 0x11111111. I parsed linux-firmware's ath11k/WCN6855/hw2.0/amss.bin (WLAN.HSP.1.1-03125, the file hw2.1 links to and mainline ath11k loads on WCN6855 laptops). It's laid out exactly like stock's amss20.bin: hash segment at 0x1000, 0x248 bytes, MBN version 5, sig_size = 0 and cert_chain_size = 0. So ordinary parts boot unsigned v5 images, and "the fuse demands a signature" is less likely than I made it sound. That fits your reading that 0x11111111 looks like an uninitialised power-up value rather than a provisioned fuse.

    Where that leaves us, honestly: I don't have another source-backed difference to point you at. Three ways on, your choice:

    1. Park Wi-Fi here. Write the result up in the lynx plan: everything you've ruled out, the PBL stage in pbl_auth.c, and the hang. Keep the node disabled. This is a perfectly good place to stop, and the write-up is valuable on its own.

    2. A stock-side read, if anyone following this issue has a second, rooted Pixel 7a. The cnss debugfs read of 0x1E22B24 before the first BHI (from my last comment) would settle the fuse-or-power-up question. Please don't wipe your userdata for it.

    3. An optional single-boot experiment, if you want one more data point: serve linux-firmware's WCN6855 amss.bin and m3.bin in place of stock's, through firmware_class.path as before. If the PBL's code changes, the stall depends on the image; if it's identical, it doesn't. If you try it, please stop at the PBL/SBL result:

      • no board-2.bin or other board data, so the firmware can't bring the radio up with another device's calibration;
      • no association;
      • one change per boot, with --watchdog on.

      I'm not certain it's worth the boot, so skip it if you'd rather park.

    Either way, thank you. This is the most careful firmware-loading investigation anyone has posted here.

  14. 22 remaining items

  15. FlorencioJP commented on Oct 28, 2026

    @FlorencioJP
    Contributor

    @derekross, the 16 modules the CNSS profile adds on top of the 52-module USB profile, in load order (cnss.load of the diagnostic initramfs, runs 9 to 17): 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 had the same list without rfkill/btpower.

    Noted on the rest: no more variants here, the fresh read-only backup before any mainline Wi-Fi session, and the Bluetooth-node caveat for the bench (recorded in device/lynx/NOTES.md on omarchy-mobile#98).

  16. derekross commented on Oct 28, 2026

    @derekross
    Owner

    Thank you, @FlorencioJP, that's the list I asked for, and the load order with run 9's difference is exactly what a later comparison will need.

    One thing slipped through on our side: #88 merged this morning with WIFI-PLAN.md as it stood before the fs_ready boot. So on main the plan's status line and section 7 item 1 still say the fs_ready boot is the next step, and none of the last day's results are in it. Could one of you (@latentoperator as the plan's author, or @FlorencioJP since these are your runs) send a small follow-up PR to device/lynx/WIFI-PLAN.md with:

    • the status line: the fs_ready boot ran and stalled, so "Google's host stack and full Android's wait" is ruled out as the pair, and the variants are stopped;
    • section 7 item 1's result: the run as you described it on the 27th (Google's kernel and vendor_dlkm modules, 68 modules, Bluetooth off, fs_ready written at 10.8 s, IN_COLD_BOOT_CAL, the power-up 5 s later, BHI from 16.0 s, 0xfad00000 on every dump, -110 at 76.8 s);
    • item 2's module comparison as background, with your caveat about first-stage loads, and the 52 + 16 list from this comment, so the "52 modules" in item 1 matches the run;
    • the ramoops note: the console survives only a warm return, so the diagnostic's init writes warm to /sys/kernel/reboot/mode first;
    • section 5's "at every Linux boot gph2-5 reads output low": true only on a kernel without the Bluetooth node (device-lynx: Bluetooth on the APM UART #98), with the "unbind serial0-0 first" step and the request to record which kernel a run used.

    The chip-revision row is already in, thanks. Nothing else is needed from either phone right now: the cold stock read (section 7 item 3) stays @latentoperator's call, with the preconditions as written.

  17. FlorencioJP commented on Oct 29, 2026

    @FlorencioJP
    Contributor

    @derekross, the follow-up is #133, device/lynx/WIFI-PLAN.md only, on top of main after #88. Your five points, in order:

    • status line: the fs_ready boot ran and stalled, "Google's host stack and full Android's wait" ruled out as the pair, variants stopped, cold stock read (section 7 item 3) the one measurement left;
    • section 7 item 1: the run as posted on the 27th (Google's kernel and vendor_dlkm, 52 + 16 modules, Bluetooth off, fs_ready at 10.800 s, IN_COLD_BOOT_CAL, buck at 15.844 s, BHI 16.001 s, 0xfad00000 on every dump, -110 at 76.772 s);
    • item 2: the module comparison as background, the first-stage caveat, the 16 CNSS modules in load order with run 9's difference, so item 1's count matches the run;
    • section 8: ramoops survives only a warm return, so the diagnostic's init writes warm to /sys/kernel/reboot/mode first;
    • section 5 and the pin table: gph2-5 reads output low only on a kernel without the Bluetooth node; with device-lynx: Bluetooth on the APM UART #98 hci_qca raises and drops it, so unbind serial0-0 first, and each run records the kernel it used.

    tools/pr-check passes. @latentoperator, it is your plan, so reword anything in it as you see fit.

  18. cx3301J commented on Nov 5, 2026

    @cx3301J
    Contributor

    Working locally on the s6e3fc5 panel driver for lynx: a mainline drm_panel port on top of curated/3-display and the published omamobile branch, with 1080x2400 60/90 Hz timings, a runtime-generated DSC PPS, and a display_pool taken from the phone's /proc/iomem rather than a guessed address. No PR yet; I will post the branch once it builds so nobody duplicates it.

    Separately, #436 is the small devinfo 3.11 boot-success fix for Android 15 units.

  19. latentoperator commented on Nov 5, 2026

    @latentoperator
    ContributorAuthor

    The controlled cold-stock comparison in WIFI-PLAN.md section 7 item 3 is complete on this Pixel 7a (2026-11-05).

    Stock Android 17 CP2A.260705.006, kernel 6.1.157-android14-11-gbd23337e42e7-ab14791245, slot A; Wi-Fi, Bluetooth and background scanning off.

    After stock shutdown completed and off, the buck still read high. GPIO sysfs was unavailable, so a reviewed GPIO cdev v1 helper requested only gph2 line 5 (GPIO 103), held it physically low for 5.000384158 seconds, read 0 at both ends, and released ownership. Stock on then logged fresh BUCK GPIO: [0] and BUCK GPIO is LOW, drive it HIGH, followed by linkup.

    Register After GPIO drop, before BHI Warm cycle, no BHI After successful firmware
    WLAN_SECURE_BOOTN (both reads) 0 0 0
    TLMM_PBL_BOOTSTRAP_STATUS 0x40 0x40 0x40
    PBL_WLAN_BOOT_CFG 0 0 0x200

    All 18 selected readings were identical across the cold and no-firmware warm sets, with no new BHI during either. Firmware validation used linkdown, off, then stock powerup: fresh BHI, SBL, AMSS/mission and FW_READY. All 54 selected reads completed. The ten PBL log words changed from 0 to 0x10001 after firmware.

    This is a mixed result: BOOT_CFG changes with the successful load, while SECURE_BOOTN and bootstrap status already differ from stalled mainline before the new BHI. It does not identify the cause or when the values diverge. The mainline comparison was captured after a failed load; its pre-BHI windowed-read stop condition still stands.

    Limits: the GPIO level was verified, not the rail voltage. Firmware had run earlier in this Android boot, so state outside the GPIO-controlled domain may persist; this is not a first-ever-BHI measurement. The successful run had live kernel logging, an unchanged boot identity, the existing 30-second stock watchdog and no USB interruptions. Earlier USB losses stopped attempts before any GPIO drop or register reads.

    Arch is restored on slot B and touch works. A fresh read-only post-return backup of all six protected partitions passed size, checksum and remote before/after verification; the guard remains intact. Raw logs, calibration hashes and partition images remain private. No new mainline power-up variant was run.

    Documentation update: #439 (only device/lynx/WIFI-PLAN.md).

  20. derekross commented on Nov 5, 2026

    @derekross
    Owner

    Thank you, @latentoperator. That's a careful run: the verified 5 s hold with timestamps on both ends, one register at a time through stock's own driver, stopping when USB dropped, and a fresh backup afterwards. #439 records it in the plan, and it's merged.

    How I read it: in section 5's earlier stock reads, with the buck held, BOOT_CFG was already 0x200 before BHI. After your 5 s drop it's 0 before BHI and 0x200 after a good load. So the drop did reset that register, a good load writes it, and mainline reading 0 after its stall is just what a load that never got that far would show. I'd take BOOT_CFG off the list. SECURE_BOOTN and the bootstrap status came through the same drop unchanged, which is the plan's "latched at power-up" outcome for those two, with the bootstrap register as the lead. The limits you wrote down still apply: firmware had run earlier in that Android boot, and mainline's values were read after its stall. Nothing more is needed from the phone; Wi-Fi stays parked.

    @cx3301J, thanks for the heads-up about the panel driver; I've left notes on the draft (linux-omarchy-mobile#202), mainly about a GPU change that would also reach the Pixel 7 Pro, and about lynx shipping with more than one panel, and on #436 and #438.

  21. alfikpl commented on Nov 11, 2026

    @alfikpl

    Hello. Just to let you know - I am also working on mainstreaming the pixel 7a phone. I managed to get the wifi module working, to the state of scanning the wireless networks, with the ath11k driver. You have to first load the AoC firmware - that takes away the stall. Thanks for your great work on all of the phones. I hope to contribute to this project soon !

  22. FlorencioJP commented on Nov 11, 2026

    @FlorencioJP
    Contributor

    @alfikpl, welcome, and thank you: that is the first lead on the Wi-Fi stall since we parked it. On our side the WCN6855 stalls in the PBL after the PCIe link comes up (the runs are in device/lynx/WIFI-PLAN.md), and the AoC was never part of the picture, so your result changes the plan.

    Could you say a bit more, so we can reproduce it here?

    • Which AoC firmware (the file, and the stock build it came from), and how you load it: Google's aoc driver, something you wrote, or a userspace write? Does the AoC have to be up before the PCIe link, or only before ath11k probes?
    • Which kernel tree and branch (upstream mainline, this repo's omamobile, or your own), and the ath11k firmware/board file you use.
    • If you can, a dmesg of a boot where the scan works, and the branch if it is public.

    Even a rough list of the steps is enough to start. If you'd rather, a PR or a draft with the notes is welcome too; Derek keeps the device page and the plan files per phone, so it has a place to land.

  23. alfikpl commented on Nov 12, 2026

    @alfikpl

    Thanks. Here is what we did, short form:

    The finding. The QCA6490 reads its boot strap from the AoC. With the AoC firmware running before WLAN_EN rises, the chip leaves the ROM (SECURE_BOOTN reads 0, BOOTSTRAP 0x40) and the ath11k firmware boots; without it the PBL stalls exactly as in your WIFI-PLAN.md (SECURE_BOOTN 0x11111111). We first proved it with Google's downstream drivers under a bare initramfs: loading aoc.bin alone, then WLAN_EN, was enough; the modem is not involved. So the answer to your ordering question: the AoC has to be online before the chip is powered (before WLAN_EN and therefore before the PCIe link), not merely before ath11k probes.

    AoC firmware. The stock vendor/firmware/aoc.bin of the lynx factory image cp3a.261005.002.a1 (md5 5791dfe1fdadcd51b409c48cc66915df), copied unchanged; the running AoC reports version 14022723-polygon, 82 services. Loaded by the kernel: your tree's gs201 Trusty client, GSA driver and AoC driver (the aoc node autoloads the file with request_firmware, the GSA authenticates it over its mailbox, the start goes through Trusty's com.android.trusty.gsa.hwmgr.aoc). Four fixes were needed on mainline: (1) the Trusty virtio IPC buffers must be premapped (virtqueue_set_dma_premapped, sg_dma_address = the buffer id) because set_dma_ops() is a no-op without CONFIG_ARCH_HAS_DMA_OPS, so Trusty was getting physical addresses and answered "inbound msg is too big"; (2) the GSA driver defers probing until the Trusty IPC is online; (3) aoc.c must pass the GSA's -EPROBE_DEFER through instead of turning it into -EINVAL; (4) aoc.c maps the chip-id register with devm_ioremap without claiming it (exynos-chipid owns it). The AoC is online at 2.6 s.

    Ordering in the tree. The PCIe root complex node carries a google,aoc = <&aoc> phandle; the controller driver returns -EPROBE_DEFER until that device is bound and its firmware reports online, so WLAN_EN is first raised after the AoC is up (3.1 s in the log). One more fix on our side: the root complex's link soft reset wiped the ELBI interrupt enables, so the BHI MSI never arrived (MHI -110); the enables are re-armed after link-up.

    Kernel and ath11k firmware. Mainline 7.3.0-rc6 with our own generated board/driver changes (gsbootstrap); not published as a branch yet. ath11k firmware from the same factory image: amss20.bin, m3.bin, regdb.bin, and a board-2.bin packed from the four bdwlan*.elf files keyed on the chip's subsystem id 17cb:0108 and the variant string google,gs201-lynx. The chip reports soc_id 0x400c1211, firmware WLAN.HSP.2.0.c12-00244; the controller takes the bootloader's /chosen/config/wlan_mac1 as its MAC.

    Steps to reproduce. 1. Put the stock aoc.bin where the aoc node's request_firmware finds it. 2. Trusty, GSA, AoC drivers with the four fixes; check the kernel log for "firmware version ... did become online". 3. Keep WLAN_EN low until then (phandle or a probe order of your choice). 4. Re-arm the root complex's ELBI IRQ enables after link-up. 5. ath11k with the factory firmware and board-2.bin.


    I have not made the kernel sources public yet. I began the project before the Omarchy-mobile and the code still has to be prepared for upstream. I hope to do it in a week or two, and go public.


    dmesg-for-lynx.txt

  24. latentoperator commented on Nov 14, 2026

    @latentoperator
    ContributorAuthor

    Following the AoC-before-WLAN_EN report above, I reproduced the ordering result on the original lynx test phone and reached one guarded connection/traffic pass on 2026-11-13.

    Build: 7.3.0-rc6-omamobile-lynx-msi-20261113+, from published kernel omamobile at 2db2bc35beecbf2deccb7f2e9bc8c3843b1edc9c plus local Trusty/AoC, power-sequencing, PCIe/MSI and board changes. The source changes and test overlays are not published yet; the base commit alone does not reproduce this build. AoC firmware was copied unchanged from my factory image and reported 14022723-polygon, 82 services, zero restarts. The Wi-Fi files were the stock firmware and board set; the modem and Bluetooth were not started.

    Successful sequence: Trusty IPC -> GSA/AoC online -> WLAN power -> PCIe/MSI -> ath11k firmware. Passive scan returned 24 BSS entries, then one activation obtained association and DHCPv4. Three gateway pings passed. HTTPS explicitly bound to Wi-Fi returned HTTP 200 and 577 bytes with TLS verification enabled and its source matching the Wi-Fi lease. Fresh SSH confirmed USB access. The six protected hashes matched, the storage guard held and all six required smoke checks passed; watchdogs and rollback timers were disarmed after fresh healthy status.

    The failure remains part of the result. Before the final MSI edge-rearm candidate, an initial activation raced supplicant readiness; a readiness-corrected attempt then stalled the WMI/TX queues and lost USB. The bounded guard recovered the phone. The final candidate passed once with no new WMI/TX timeout, oops or panic in the connection capture. This does not establish the earlier stall's cause or prove the MSI change fixes it.

    Drivers remain blacklisted for boot and startup is still a guarded manual sequence. The phone has since been rebooted; home/touch work, but the Wi-Fi sequence has not been repeated on that boot. Reconnect, endurance, suspend/resume, Bluetooth coexistence and unattended startup remain unqualified. No new windowed chip-register reads were used for these results; the existing stop conditions remain.

    Documentation and full saved smoke results: #751. Firmware, raw logs and identifying data remain private. Kernel code will need separate subsystem review before any release enablement.

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