Files
UniXNU/docs/devices/bluejay.md
2026-11-16 21:55:12 +00:00

25 KiB

Google Pixel 6a (Tensor G1)

tuxmanxd/UniXNU-bluejay

Google Pixel 6a (bluejay, Tensor G1) port scaffold. Hardware values come from the pixelios-kernel research port, which boots XNU 8796 with the iOS 16.7 userland on this phone; this repository restates them in UniXNU format. No UniXNU image has been built or booted for this board yet.

Source record: porting.toml, 3b0c11f78352d3399d1c603d951a6fcbebca2ab7.

This page is generated from the device record.

Area Status
Configuration implemented
Firmware and boot implemented
CPU and interrupts implemented
Memory and DMA implemented
Storage and root planned
Display planned
Input planned
GPU planned
Network planned
cellular not-applicable
Audio planned
Power planned
Image assembly implemented
usb planned
Diagnostics planned

Configuration

implemented

Components: ArmDTPlatform, ArmCpu, XnuPrelink, DarwinBoot, GopFramebuffer.

Hardware and dependencies

  • Google Pixel 6a (bluejay), Tensor G1 (GS101), 6 GiB DRAM (2 GiB low bank + 4 GiB high pair, measured), 128 GiB UFS, 1080x2400 panel.
  • pixelios-kernel hardware descriptions: soc/google/gs101/soc.json, device/google/bluejay/board.json, docs/research/05-fact-check.md.

Configuration

  • board.soc_source = gs101 pins UniXNU-gs101-common; board.os_profile = ipados-17.7.11-21H461.
  • config_fragments apply in order arch, soc, os; the resolved defconfig bluejay_defconfig applies last.
  • patch_series is arch, soc, os, device; every series except device is inherited and currently empty for this SoC.

Procedure

  • From a UniXNU checkout: ./unixnu defconfig bluejay_defconfig -D ../UniXNU-bluejay
  • ./unixnu fetch -D ../UniXNU-bluejay; ./unixnu status --json -D ../UniXNU-bluejay
  • python3 scripts/porting-record.py check --device ../UniXNU-bluejay

Development findings

  • UniXNU ships only iPadOS 17.7.11, iPadOS 26.5.2, iPadOS 27.0.1 and macOS 27.3 OS profiles. There is no iPhone/iOS 16.7-20H19 profile, so this scaffold selects the oldest shipped profile; a native iPhone profile is required before the result is a real phone port.
  • pixelios boots XNU 8796 (iOS 16 generation). UniXNU pins a newer XNU and its own ordered patch series, so the pixelios patch stack is not reused as-is.
  • The pixelios shim is a bare-metal EL2 loader, not UEFI. Its hardware facts (UFS cold re-init, HS-G4 link, memory map, device tree) are reusable; its code is not a UniXNU component.
  • The SoC commit pinned in feeds.toml is the first commit of UniXNU-gs101-common and must be replaced with the reviewed published commit.
  • The device owns no board driver submodules yet: Gs101Ufs, FTM5 touch and GS101 power are unassigned.

Completion criteria

  • Every source pin resolves; status shows the intended profiles, components and patch order; the porting record check passes.

Current limits

  • No component build has been attempted; UniXNU, uefi-bluejay and a working IPSW profile are missing.

Next action

  • Clone UniXNU beside this repository, run defconfig, fetch and status, and record the resolved output.

Evidence

Firmware and boot

implemented

Components: DarwinBoot.

Hardware and dependencies

  • tuxmanxd/uefi-bluejay: the Mu-Silicium GS101/Pixel 6a platform (bluejayPkg + GS101Pkg), installed with scripts/apply.sh.
  • Google Pixel 6a factory image for vendor firmware inputs.
  • ABL hands over a linear framebuffer at 0xfac00000 and a live FDT.

Configuration

  • [firmware] config = bluejay, dsc = Platforms/Google/bluejayPkg/bluejay.dsc, fd = Build/bluejayPkg/{target}_CLANGPDB/FV/SILICIUM_UEFI.fd.
  • [firmware.defines] DEVICE_MODEL = bluejay. [bootimg.image] header_version 4, kernel_compression none: ABL rejects a nonzero arm64 text_offset, so the shim is built with text_offset 0.
  • [bootimg] dtb is written by the firmware preflight: a 4 KiB FDT envelope followed by the real payload volume, because the pack stage appends only that file to the kernel field.
  • [payload] size_mib 54: the volume carries kernel.macho (49.5 MiB) plus the ADT; boot.img is checked against boot_partition_bytes 67108864.

Procedure

  • Install uefi-bluejay into a Mu-Silicium checkout with scripts/apply.sh, following https://unixnu.com/porting/uefi-firmware, then build the firmware volume with build_uefi.py -d bluejay.
  • Run scripts/package.sh --check --os 17.
  • Boot the artifact with fastboot boot and read the panel.

Development findings

  • The stock bootloader boots Android boot images only, so UniXNU must be packaged as a boot image first (layer 1) before the firmware volume is reachable; the pixelios fastboot boot flow proves this path on the phone.
  • ABL sets up the panel framebuffer and provides no framebuffer node in its FDT; the address and size are measured (0xfac00000, 0x9e3400).
  • The phone is read-only toward storage in the pixelios pre-kernel path; any UniXNU firmware must preserve that property until the userdata install is explicit.
  • Tensor ABL caps the boot image ramdisk at 32 MiB (measured 2026-11-15: 4 and 32 MiB boot, 36/40/42/44/46/48/54/56 MiB are refused, header v0 and v4 alike), while the prelinked kernel alone is 49.5 MiB. The payload therefore travels in the kernel field, which accepts 56 MiB.
  • ABL does not parse bytes appended after the FD: a kernel field with 54 MiB of 0xAA after the FD still boots the firmware, so the field can carry the payload volume behind a 4 KiB envelope.

Completion criteria

  • The firmware volume builds, UniXNU.efi runs from it, the menu appears on the panel and the memory map matches the DTS reservations.

Current limits

  • The menu scrolls without input on hardware; the xnu boot entry has not been reached yet. No firmware archive pin and no proprietary file hashes yet.

Next action

  • Boot the kernel-field payload image, confirm that UniXNU finds EFI/UniXNU/unixnu.toml on the payload volume, then fix the menu input/redraw bug.

Evidence

CPU and interrupts

implemented

Components: ArmCpu.

Hardware and dependencies

  • 8 cores: 4x Cortex-A55, 2x Cortex-A76, 2x Cortex-X1.
  • GICD 0x10400000 (64 KiB), GICR 0x10440000 (1 MiB, 8 redistributors, stride 0x20000).
  • CNTFRQ 24.576 MHz; virtual timer PPI 11 = INTID 27.

Configuration

  • cpus/timebase-frequency = <24576000>; psci method smc.
  • unixnu,cluster-type = E for cpu0-3 and P for cpu4-7 in the board DTS.
  • CONFIG_XNU_ARMCPU=y and CONFIG_XNU_ARM_DT_PLATFORM=y.

Procedure

  • Compile dts/bluejay.dts against soc:dts, boot, and read the CPU and interrupt bring-up lines from the kernel log.

Development findings

  • The big cluster mixes A76 and X1 cores; treating cpu4-7 as one P cluster is a scheduling policy choice, not a hardware boundary.
  • GICv3 needs both the distributor and all eight redistributor frames; a truncated GICR size leaves secondary CPUs unable to start.
  • pixelios boots all 8 CPUs with a custom boot-CPU driver (XNU patch 0018), which UniXNU ArmCpu replaces.

Completion criteria

  • All 8 CPUs come online, IPIs work, the timer ticks at 24.576 MHz and secondary CPUs survive idle.

Current limits

  • Not run on hardware in this configuration; cluster mapping has not been validated under UniXNU.

Next action

  • Boot the firmware and kernel and record the CPU startup and timer lines.

Evidence

Memory and DMA

implemented

Components: ArmDTPlatform.

Hardware and dependencies

  • DRAM banks measured: 0x80000000 (2 GiB, carve-outs), 0x880000000 and 0x900000000 (2 GiB each, contiguous, no reserved range inside).
  • XNU range: 0x884000000 size 0xf0000000 (the high pair after the firmware payload window, 3840 MiB, 137.9 MiB under the 0xF89F6000 run measured with the kernel image still held).
  • USB DMA carve-out: 0xf8800000 size 0x1000000 (ABL, no-map).

Configuration

  • One memory node for the high range only: reg = <0x8 0x84000000 0x0 0xf0000000>.
  • payload.adt.owned_memory_size = 0xf0000000; dram_size = 0xf0000000.
  • The first 64 MiB of the high range (0x880000000) is the firmware payload window: reserved in the UEFI map, outside XNU's node.
  • reserved-memory/framebuffer@fac00000 is no-map.

Procedure

  • Compare the firmware memory map with the DTS, then check that XNU boots with the full 4 GiB and no overlap.

Development findings

  • XNU maps [kernel, physBase+memSize) cacheable, so the range must contain no carve-out and its base and size must be multiples of 32 MiB (16K-granule L2 block). The high pair, minus the 64 MiB payload window, satisfies both.
  • The payload volume travels in the boot image kernel field because Tensor ABL caps the ramdisk at 32 MiB; BootShim copies it to 0x880000000 before the MMU is on, and the firmware maps that window as reserved.
  • The low bank cannot be handed over whole: ABL reserves several no-map ranges, including its 16 MiB USB event buffer. Its allocation moves between boots (0xF8B92000, 0xF8D12000 observed).
  • The 0xfac00000 framebuffer is below the XNU range, so it needs a reservation for firmware/DarwinBoot but does not fragment owned memory.

Completion criteria

  • XNU sees 0xf0000000 bytes of owned memory, boots past pmap setup, and the framebuffer, USB carve-outs and payload window stay untouched.

Current limits

  • The low bank is not adopted; memory outside the window is unavailable to the OS.

Next action

  • Measure the firmware memory map in the UniXNU boot log and confirm the DTS reservations.

Evidence

Storage and root

planned

Hardware and dependencies

  • UFS host controller 0x14700000, vendor HCI +0x1100, UNIPRO 0x14780000, PHY/PMA 0x14704000.
  • UFS LUN 0 GPT: userdata first 0x22F1E6, last 0x1DC03FA, 4096-byte LBAs.
  • Cold re-init after the bootloader shuts the controller down, then HS-G4 link.

Configuration

  • &ufs and &ufsphy status = okay; CONFIG_XNU_NATIVE_UFS=y.
  • [userdata] first_lba/last_lba/lba_bytes in board.toml.

Procedure

  • Implement a Gs101Ufs component (or a board driver) from the pixelios UFS cold-init sequence, then mount the iOS root read-only and verify the GPT geometry.

Development findings

  • The bootloader leaves the UFS controller off; a plain HCI reset is not enough and the pixelios shim reprograms vendor HCI, UNIPRO and the PHY before the link trains.
  • The link is switched to HS-G4 (12x read throughput); the gear change is part of bring-up, not a later optimisation.
  • The root lives in userdata, so the partition geometry is fixed by the phone GPT and is recorded in board.toml.
  • No shared UniXNU UFS component supports Exynos/GS101 UFS yet; ShengUfs is Qualcomm-specific and cannot be reused.

Completion criteria

  • The kernel enumerates the UFS LUNs, reads the GPT and mounts the root; a full read of userdata completes without CRC errors.

Current limits

  • No driver code; the pixelios implementation is bare-metal C in the shim, not an XNU kext.

Next action

  • Port shim ufsinit.c to a UniXNU storage component and prove a read-only userdata read.

Evidence

Display

planned

Components: GopFramebuffer.

Hardware and dependencies

  • Framebuffer 0xfac00000, size 0x9e3400, 1080x2400, stride 4320, a8r8g8b8, set up by ABL.
  • No framebuffer node in the firmware FDT.

Configuration

Procedure

  • Describe the scanout buffer as the GOP handoff, boot, and check that the boot logo and the first frames appear.

Development findings

  • ABL programs the panel and leaves a linear buffer; there is no display controller driver in the pixelios pre-kernel path.
  • 1080x2400 with a 4320-byte stride is 4 bytes per pixel with no padding beyond the width.
  • The pixelios kernel only patches an early framebuffer console (XNU patch 0003) and later hands off to a software-rendered UI.

Completion criteria

  • The firmware menu and the XNU boot text render, and the framebuffer is not reclaimed as free memory.

Current limits

  • No panel, backlight or DPU driver; brightness control is absent.

Next action

  • Map the measured buffer into GopFramebuffer and confirm the handoff on hardware.

Evidence

Input

planned

Hardware and dependencies

  • STMicro FTM5 touch, chip ID bytes 0x36 0x48, SPI11 / USI11 at 0x10D40000, USI mode register 0x10C2100C, max 10 MHz.
  • Pins: gpp23-0..2 SPI function 2, CS gpp23-3, IRQ gpa9-0 (XEINT_20), reset gpp26-0, TS_SPI_SEL gpp26-1.
  • Power: S2MPG10 LDO25M/LDO26M. Keys: power gpa10, vol+ gpa8, vol- gpa7, all active low.

Configuration

  • DTS touch node with st,max-coords and spi-max-frequency, currently disabled.
  • [adt.buttons] button-names = hold, voldown, volup.

Procedure

  • Implement the FTM5 SPI driver and pinctrl/rail setup, then verify contacts and the three buttons.

Development findings

  • Touch works on hardware under pixelios (2026-10-22): the lock screen and home screen respond, so the wiring and firmware protocol are proven.
  • The touch controller is behind USI in SPI mode; the USI mode register must be programmed before the SPI transfers.
  • GPIO keys are read directly through the pinctrl banks; uniXNU uses the ADT buttons node instead of the pixelios /sweet-keys node.

Completion criteria

  • Contacts reach the HID stack with correct coordinates and all three buttons generate events.

Current limits

  • No FTM5 driver or pinctrl component; the DTS node is disabled.

Next action

  • Write a board input driver from the pixelios touchpoke/touchprobe code and enable the DTS node.

Evidence

GPU

planned

Hardware and dependencies

  • Mali-G78 MP20 at 0x1C500000 size 0x5000, three SPIs (job 420, mmu 421, gpu 419).
  • Power: PMU G3D 0x17461E00/0x17461E04, CMU_G3D 0x1C400000, S2MPU windows 0x20080000-0x200B0000.
  • Firmware loaded by the shim/kbase; no Apple GPU stack.

Configuration

  • gpu node present but status = disabled; no GPU component selected.

Procedure

  • Decide between a native Mali component and no GPU, then enable the node only when a driver exists.

Development findings

  • UniXNU GPU components target Adreno (Adreno740/AdrenoMetal) and cannot drive Mali-G78.
  • pixelios has substantial GS101 GPU power and job-submission research, but the UI is software-rendered and slow; this is not a working Metal stack.
  • The framebuffer path does not depend on the GPU.

Completion criteria

  • Either a Mali component renders and presents, or the area stays not-applicable with the software path documented.

Current limits

  • No Mali driver component; GPU is not part of the first boot milestone.

Next action

  • Scope a Mali-G78 component or mark the area not-applicable for bring-up.

Evidence

Network

planned

Hardware and dependencies

  • Broadcom PCIe Wi-Fi behind the GS101 PCIe controller; firmware and NVRAM from the phone's vendor partition.
  • pixelios shim WLAN bring-up: link training, PMU/CMU sequences, firmware container packing.

Configuration

  • No wlan node or component selected yet.

Procedure

  • Add a wlan node, port the PCIe link and firmware loader, then run the iOS Wi-Fi stack.

Development findings

  • pixelios implemented PCIe link bring-up and a USB network log path, but Wi-Fi association is still listed as outstanding.
  • The WLAN register sequences are captured as data (soc/google/gs101/wlan_seq.json) with per-row citations; they are reusable facts, not a driver.
  • GS101 has no UniXNU network component.

Completion criteria

  • The interface associates, obtains an address and passes traffic under the target OS.

Current limits

  • No driver, no firmware hashes, no association evidence.

Next action

  • Port the PCIe/WLAN sequence into a component and capture firmware inputs.

Evidence

cellular

not-applicable

Hardware and dependencies

  • GS101 modem and the Pixel 6a baseband are out of scope for the first UniXNU bring-up.

Configuration

  • No modem node and no baseband component is selected.

Procedure

  • Revisit only after storage, display, input and network work.

Development findings

  • pixelios lists the modem as outstanding; no QMI/QRTR or IPA path exists for this board.
  • UniXNU has QcomBaseband for Qualcomm platforms; it does not cover the GS101 modem.

Completion criteria

  • Not applicable while the modem is outside the port scope.

Current limits

  • No cellular function; the area is deliberately empty.

Next action

  • Keep out of scope until the base platform boots.

Evidence

Audio

planned

Hardware and dependencies

  • GS101 audio subsystem: ACPM IPC, DSP firmware, S2MPG10/11 PMIC, speaker/headset codec paths.
  • pixelios shim ACPM protocol work and DVFS notes.

Configuration

  • No audio nodes or components selected.

Procedure

  • Describe the DSP/IPC nodes, port the firmware loader, then enable the OS audio stack.

Development findings

  • pixelios has ACPM host-tested protocol code but no audio playback on the phone; audio remains on the roadmap.
  • ACPM is the shared IPC transport for power and audio on GS101, so the power and audio areas share the same firmware interface.

Completion criteria

  • The OS audio server starts and produces playback through the speaker or headset.

Current limits

  • No DSP firmware, codec driver or mixer configuration.

Next action

  • Port ACPM transport first, then the codec path.

Evidence

Power

planned

Hardware and dependencies

  • S2MPG10/S2MPG11 PMICs over ACPM Speedy; PMU alive 0x17460000, G3D 0x17461E00.
  • pixelios regulator, DVFS and rail sequences, including the GPU rail suspend write.

Configuration

  • No PMIC or regulator nodes selected; no XNU PMGR replacement.

Procedure

  • Add PMIC/ACPM nodes, implement rail and clock control, then verify idle, suspend and thermal behaviour.

Development findings

  • GS101 has no Apple PMGR; the ADT reports vmm-present so AMFI does not wait for ApplePMGRNub.
  • The GPU rail sequence in pixelios performs a single guarded FUNC_UPDATE and refuses on unexpected mode or voltage, a pattern to preserve in any driver.
  • Charging, battery and thermal ownership are unresolved.

Completion criteria

  • Rails and clocks are owned deterministically; the OS boots without a PMGR timeout and can idle.

Current limits

  • No PMIC driver, no suspend/resume, no thermal policy.

Next action

  • Port the ACPM/PMIC rail layer and record the first idle-power measurement.

Evidence

Image assembly

implemented

Components: XnuPrelink, DarwinBoot.

Hardware and dependencies

  • Target IPSW and its disk images; the pixelios port used iOS 16.7 (20H19).
  • Android boot image geometry from the pixelios shim.
  • GPT userdata geometry from the phone.

Configuration

  • [userdata] first_lba 2290150, last_lba 31130618, lba_bytes 4096.
  • [payload] volume BLUEJAY, 54 MiB (kernel field, not ramdisk), loader firmware, dram_size 0xf0000000, kernel_args -v serial=3 debug=0x14e keepsyms=1.
  • [bootimg.image] base 0, kernel_offset 0x8000, tags_offset 0x100, pagesize 4096, header_version 4, kernel_compression none.

Procedure

  • scripts/package.sh --check --os 17
  • scripts/package.sh --os 17 --uefi ../uefi-bluejay

Development findings

  • The payload ADT uses chip id 0x8015, the Apple platform identity pixelios emits, not the GS101 id.
  • The pixelios port stores the iOS root in userdata and mounts it read-only; the UniXNU payload must keep the same partition contract.
  • No uefi-bluejay checkout exists, so package.yml names a repository placeholder and cannot run yet.

Completion criteria

  • A clean checkout produces boot.img, payload.img, kernel.macho and the manifest, and the images boot on the phone.

Current limits

  • The OS profile is iPadOS 17.7.11, not iOS 16.7; the IPSW pin and SHA-256 members are absent.

Next action

  • Create uefi-bluejay, then pin the IPSW, its member hashes and the payload runtime values.

Evidence

usb

planned

Hardware and dependencies

  • DWC_usb31 at 0x11110000 size 0x10000, GSNPSID 0x33313130, global block at +0xC100.
  • ABL USB DMA region 0xf8800000 size 0x1000000 with a moving event buffer.

Configuration

  • usb node present but disabled; no USB component selected.

Procedure

  • Enable the DWC3 node, port the XDCI/device mode path and verify enumeration as a network gadget.

Development findings

  • pixelios proves USB device mode on this phone: the running iOS appears as an Ethernet gadget and streams the kernel log over UDP.
  • The event buffer address changes between boots and must be checked against the 16 MiB ABL region before DMA is enabled.
  • UniXNU has QcomDwc3 for Qualcomm; a GS101 variant is not selected.

Completion criteria

  • The phone enumerates, a host sees the gadget, and the log transport works.

Current limits

  • No driver component; USB is not on the first boot path.

Next action

  • Port the DWC3 device-mode bring-up and reuse the USB log transport.

Evidence

Diagnostics

planned

Hardware and dependencies

  • pixelios log channels: on-screen console, RAM log, USB serial gadget, USB network log and QR-code log.

Configuration

  • kernel_args include -v, serial=3, debug=0x14e, keepsyms=1.

Procedure

  • Bring up one transport at a time, starting with the framebuffer console, then capture a full boot log with the image revision.

Development findings

  • The QR log channel exists because reading a phone screen by eye is unreliable; its encoder needed SHA3 removal because GS101 lacks those instructions.
  • A software-driven test loop (build, boot, read log, reset) works without button presses and is the model to keep.

Completion criteria

  • A boot log survives a failed boot and identifies the image revision.

Current limits

  • No log transport is integrated into UniXNU for this board.

Next action

  • Enable the framebuffer console first, then add USB or QR capture.

Evidence