Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.
C
Changelog
What changed, newest first. One line a change: the part of the system, then what is different for whoever runs it.
The commits say how; docs/ says why.
Sources
What a build is made of, and where each part is fetched from. Nothing of these is kept in the repositories but the changes to it; the scripts named fetch the release and put the changes on.
- the kernel: Apple's XNU, tag
xnu-8796.121.2(the release the Darwin 16.5 user space was built from) --xnu/apply.sh, the changes inxnu/hakios.patchandxnu/files/, the drivers indrivers/ - the data volume: Apple's HFS Plus, tag
hfs-627.100.6--xnu/hfs/hfs.sh, the changes inxnu/hfs/hfs.patch - the boot: U-Boot
v2026.07--hakios-uefi/u-boot/apply.sh, the changes inhakios-uefi/u-boot/hakios.patchandfiles/; the device's part inhakios-sm7325/u-boot/ - the UEFI (the other way to start): Project Silicium's Mu-Silicium at
e51cd877e31c--hakios-uefi/apply.sh - the GPU's OpenGL ES: Mesa 25.0.7, the freedreno driver for the Adreno --
tools/mk-mesa.sh, the changes intools/mesa/ - the GPU's Vulkan: the same Mesa's Turnip --
tools/mk-turnip.sh, the changes intools/mesa/turnip-darwin.patchandtools/mesa/hakios-drm/ - Metal, the direct path: hakios-metal, beside this tree --
userland/build.sh METAL_PLUGIN=ca, withuserland/src/HakiosMetalCA.m - Metal, on Vulkan: UniXNU/AdrenoMetal at
243d63cand steelbrain/metal2vulkan at624bab3with AdrenoMetal's 30 patches --tools/mk-m2v.sh, the changes intools/m2v/ - the build tools: Apple's libdispatch, libpthread and AvailabilityVersions, bootstrap_cmds, unifdef, the macOS 13.3 SDK and the iPhoneOS 16.5 SDK --
xnu-deps/setup.sh - not fetched by anything here: the Darwin 16.5 user space, which you lay out from your own IPSW (
sysimage/), and the phone's own firmware, pulled from its Android (fw-*)
2026-11-04
- metal: 120 frames a second is not reached: hakios-metal 101 to 103 in page swipes, about 90 in Settings' scroll, 89 to 92 at the home button, 4 to 6% of frames over 16.7 ms; after the finger lifts, 185 of a swipe's 205 frames take 8.3 ms, while it drags 75 to 80 a second;
docs/reference/metal-measurements.md - springboard:
wallpaper_freeze=1makes the wallpaper one picture, from the IOSurface on the panel (kern.hakios.fb_surface): home button 92 frames a second, against 28 withwallpaper_cache=1and 50 without either; the icons blink once, the wallpaper does not change until the next boot, and the lock screen shows the same picture - windows:
tools/metal-run.shsetswallpaper_freeze=1andmetal_defer=1; an argument inMOREtakes the place of a default of the same name - metal: 250 touch samples a second instead of 125 change nothing (100 to 102);
touch_poll_usstays 4000 intools/metal-run.sh - springboard: the status bar's legibility gradient (compositing filter
darkenSourceOver) came out solid white over the top eighth of the home screen, on both Metal paths; it is drawn without the filter now, the wallpaper reaches the top and the status bar reads; why that filter's shader gives white is not known - windows:
tools/metal-run.shrenders at the iPhone XR's screen by default, 828x1792 stretched to the panel: Settings' scroll 118 frames a second, swipes 110 to 115, home button 106, 1% of frames over 16.7 ms; the wallpaper is the picture a boot at the iPhone X's screen kept - display:
fb_beam=1, the default: each thread's rows are written behind the panel's scan line (2433 lines, 8338 us), and to the eye the moving bands are gone; it costs no frames (Settings' scroll 91 against 89, home button 89 to 92 against 81, swipes 101 against 99) - docs: a short README and a front page by what you came to do;
docs/reference/metal.mdand its measurements; every boot-argument of both Metal paths in the catalogue (tools/boot-args.pyreadshm_boot_arg_set) - licenses: Turnip, AdrenoMetal, metal2vulkan and the plug-in's origin in
THIRD_PARTY_LICENSES.md; a binary built withM2V=1has obligations and is to verify before any distribution
2026-11-03
- metal: Metal on Vulkan, as UniXNU runs it:
tools/mk-turnip.sh,tools/mk-m2v.sh,userland/build.sh TURNIP=1 M2V=1; nothing of it inprebuilt/, you fetch and build it - metal:
vk_testpasses on the phone (RC_JOBS=rc.jobs.vktest): Vulkan 1.3.305, a 4 MB fill checked word by word, a submission and its fence wait 2.0 ms on average - metal:
m2v_testpasses on the phone (RC_JOBS=rc.jobs.m2vtest): the device in 134 ms, a 1080 x 2400 clear checked pixel by pixel, a full-screen clear committed and waited for in 1.08 ms - metal:
metal_m2v=1: the window server draws through metal2vulkan, correctly since it opens QuartzCore's pipeline library and picks CoreAnimation's shader variants by label: swipes 50 to 57, scroll 41, the GPU 14 to 21 ms a frame;tu_debug=andm2v_utrace=1say why - gpu: HakiosGPU's request
0x7d: the client's own pages as a buffer; Turnip takes any page-aligned pointer throughVK_EXT_external_memory_host(grey icons and text before) - gpu: Turnip has
VK_NV_linear_color_attachmenton Darwin, unused so far: the window server's pass has a second attachment that is not linear - springboard:
wallpaper_cache=1rasterizes the wallpaper's window: page swipes 85-91 to 97-103, an app's close slower (22-30 against 40);wallpaper_freeze=2hides the wallpaper: home button 78 frames a second against 28, swipes unchanged;wallpaper_freeze=1's picture comes out one colour here, and is not used - metal:
metal_fps=mainpassprints the draws of a heavy pass,metal_fps=passeseach pass's attachments - display: surfaces let go of when their pixels are on the panel, not when the next frame comes: no change in frame rate
- metal: luma_boost is left out when QuartzCore's library has its own (a later system's, off a Mac): it was 83% of the GPU's timed passes;
metal_luma_boost=1draws it - metal:
metal_defer=1records the window server's command buffers on a thread of hakios-metal's own -- a draw costs its render thread ~5 us instead of ~40 - metal:
metal_fps=passesprints every pass of a slow frame; the full-screen pass of the home screen is 6 to 13 ms of the GPU's ~10 a frame - display:
fbsubtype=Nwithfbwidth,fbheightandfbscalereports another iPhone's screen class (the XR's 828 x 1792 at 2: no red rdar status bar) - springboard: vibrant views keep their filters when the window server draws through Metal;
ui_lite=1turns on Reduce Transparency and Reduce Motion,ui_lite=0turns them off - windows:
tools/metal-run.shruns on eight cores at 120 Hz with six scaling threads and without-v(VERBOSE=1 brings it back), and its summary shows the frames on the panel;tools/frames.pycounts the Metal road's frames and the share over 8.3 ms - loader:
xnuloader/mkidentity.shbuilds lzfse by its default target (there is nolzfseone)
2026-11-02
- metal: the window server's compiled shader stages outlive the boot -- kept in
/var/db/hakios-metal-cache(var_store left the old/var/tmpout), keyed by what the compiler is, so a rebuilt plug-in keeps them; and carried in the image (tools/metal-cache-seed.sh), so no boot compiles them - metal:
metal_fpssays why each pass stayed in sysmem, the slowest and the heavy one included - windows:
tools/metal-run.sh, a measured Metal run in one command from Git Bash (var_storeon by default);tools/ca-metallib.pyoff a Mac withCA_QC_METALLIB=
2026-10-30
- repositories:
mainis the branch ofhakios-xnu-kernel-- what washakiosand thenu-boot;hakiosis gone,git cloneneeds no-b - site: a Changelog page, this file
- site: the commands to start with clone the public repository (
hakios-xnu-kernel);oosexgang/hakiosis private and cloned for nobody - site: the documentation synced, with the guide to installing beside Android and the U-Boot page in the table of contents
2026-10-29
- u-boot: HakiOS's own U-Boot, started by the phone's bootloader from RAM; it starts the loader and the kernel from memory, nothing of the vendor's firmware in between
- u-boot: the loader's fastboot works under U-Boot (USB 2.0 without the link power states that dropped the cable after 3 s, endpoints and packet sizes as the loader asks, DWC3 transfers started when queued)
- u-boot: every download and every step says what it does on the screen, tagged
[oosexgang] - u-boot:
hakios sumchecks a flashed partition from the phone itself;hakios reboot bootloadergets to fastboot - install: HakiOS beside Android: U-Boot in
boot_a, Android in slot b untouched, a menu at every start (Android after 10 s, HakiOS (Darwin), fastboot, U-Boot's shell) - install: choosing Android marks HakiOS's slot unbootable and lets the bootloader switch; U-Boot never changes which slot is active (changing it by hand once left every image refused with "Device Error")
- install:
tools/userdata-carve.py --wipemakesdarwinat the end of userdata;tools/sparse-split.pycuts an 8 GB image into pieces Windows fastboot can send - storage: the data volume is HFS Plus in
darwinafter the system, about 12 GB, journaled, kept between starts; before,/private/varwas a memory volume shown as 1 GB and empty at every start - storage: HFS Plus built into the kernel; Apple's sources are fetched at
xnu/apply.sh, only this port's changes are kept - storage: the data volume's writes go straight to the flash and the cache is flushed when the journal asks (a flush and a read back after every write made the system crawl)
- storage: the loader takes any phone's
darwin(name, type, unit, at least 20 GiB), not only the first phone's place and identifier - storage:
tools/mkdatavol.pymakes the empty volume as a 50 KB image that leaves the system volume alone - graphics: the installed image draws with the GPU (Mesa,
GL=1): smooth, and the wallpaper drawn - graphics: the system's other services and the processes that draw parts of the interface are in the image (
REST=1): the wallpaper and the applications that need them - graphics: frames at 120 a second (
fb_hz=120), andfb_measure=1writes to the log how often the panel itself refreshes - loader: nothing printed over the system's screen while it runs; the log still takes every line
- loader: its log written under U-Boot too; a download the computer gave up on no longer hangs it
- wifi: the Wi-Fi controller is found where the system looks for its address, so registering applications no longer waits 5 s for each (found by the ios-sapphire port)
- tree: Darwin, not iOS, in the sources, the scripts, the file names and the documentation; no file of Apple's in the repositories
- docs: installing beside Android, step by step (
docs/dual-boot.md, in Russian too), and the U-Boot page with everything as it was done and checked