Skip to content

Rebase onto Asahi 7.2.10: one kernel carrying integration/aurora and aurora-wip - #139

Merged
RyanTheTide merged 10000 commits into
omacom:integration/aurorafrom
aurora-silicon:unified/7.2.10
Oct 19, 2026
Merged

RyanTheTide merged 10000 commits into
omacom:integration/aurorafrom
aurora-silicon:unified/7.2.10

Conversation

@RyanTheTide

@RyanTheTide RyanTheTide commented Oct 18, 2026 •

Copy link
Copy Markdown
Collaborator

Ready for review. Build, KUnit, dtbs and hygiene gates pass, and the hardware is qualified on the MacBook Neo (J700), the M1 Pro (J316s) and the M2 (J413). One known regression is listed under Known issues. Any further updates will be fast-forward pushes to this branch. This branch moves integration/aurora to Asahi Linux asahi-wip v7.2.10 (6d20a3172ffe). On top of that base it carries both:

The result is one kernel for Apple Silicon Macs from M1 to the MacBook Neo (J700).

Layout

The branch is asahi-wip v7.2.10 plus one merge commit per topic, in dependency order (assemble.sh): at update #10 (b4c606fb1c51), the canonical assembly b4c606fb1c51^2 carried 2,306 non-merge commits over Asahi 6d20a3172ffe (2,208 unique subjects). Counted as non-merge commits in the head outside both Asahi and integration/aurora 168dc647cc51, the head adds 3,107; the extra 801 are earlier versions of rewritten commits that the fast-forward publishes keep in history. Each topic is reviewable on its own:

000-devicetree · 010-soc · 020-dart · 050-nvme · 060-power-thermal · 070-audio · 080-radio · 090-input · 100-sep · 170-usb · 200-dcp · 210-gpu · 220-userpages · 240-isp · 245-ane · 900-config-tooling

Topics with a hard dependency are based on that dependency: 200-dcp on 170-usb, and 080-radio on 020-dart. That way every commit builds.

Updates to this PR are fast-forward only: each re-assembly is published as a merge commit on top of the previous head, so the branch is never force-pushed.

Provenance

  • Every carried commit keeps its original author and date, and ends with (cherry picked from commit <sha>).
  • A commit that needed substantive adaptation to 7.2 carries a [ryan: adapted to 7.2.10 — …] note above its trailers.
  • Nothing from one author is squashed into another author's commit.
  • Third-party trailers, including their own co-author lines (one contributor commit credits an AI co-author), are left exactly as written.
  • Some older Aurora commits end with Signed-off-by alone. No trailer was added after the fact.
  • New commits by Aurora end with Assisted-by: LLM and Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>.
  • No upstream AI-instruction files (AGENTS.md, CLAUDE.md, GEMINI.md) were added.

What was carried, and how it was checked

Every commit in both source trees was classified by stable patch-id, subject and author, cherry-pick trailers, and file-level deltas for near-matches. A subject match alone was never treated as equality.

Source Commits Already present / older copy of ours Carried or ported Review fixes Dropped Held
aurora-wip-rebase (0bbaaa063195) vs aurora-wip 1,324 611 (412 present + 199 older copies) 702 (283 port + 2 delta + 417 overlapping, best version kept), plus the 9 commits with a provenance note — 4 —
integration/aurora delta beyond that (5d570a78a4ab) 1,388 1,309 — 64 (#136) 5 —

Rule on conflicts: Asahi is the base. Where two versions of the same change exist, the more correct and more complete one is kept, and nothing either side relied on is lost. J700/Neo code and the SEP keep Aurora's implementation unless the other version is demonstrably better. Every decision is recorded in the topic hand-off notes.

Per topic:

  • 000-devicetree: 35 carried. DT hunks from mixed commits are split here; bindings stay with their drivers.
  • 010-soc / 020-dart / 060-power-thermal: 28 carried. The T8122/T6031 display bring-up rows are ported in update media: apple: avd: update decoder and improve playback reliability #10: the PMP startup provider on 010-soc and the display gate on 170-usb (see "M3 native display"). The T6030 rows stay listed, not ported.
  • 080/090/240/245: 24 carried, plus 2 adaptations.
  • 100-sep (+ tools/aurora-sep): 192 carried, 15 of them adapted.
  • 170-usb: 59 ported, 35 already present or superseded; rebased onto HPD v2.
  • 200-dcp: 161 dispositioned (78 ported, 67 adapted, 13 present, 3 dropped).
  • 210-gpu: the M3/T8122 backend is compile-only and gated to M3 hardware, with no change to G13/G14/G17. The M3 runtime/GPUVM redesign is listed for its maintainer (see "Not in this PR").
  • 220-userpages: the Merge aurora-wip (carries #131) with review fixes #136 review fixes, carried mechanically.
  • 900-config-tooling: generated by make-900.sh; 15 carried, 8 adapted.

Carried with a provenance note

The following commits were carried on Aurora's maintainer decision. They are disclosed here so their authors can document sources or follow up on gaps:

  • 0aa39105: drm/apple: let a dock replug restore a parked dcpext scanout (Chris Kearney).
  • 7b1603a0, f18e6a5b: the T8122 GPU DT resources and UAT placeholders (garymeyerca), as disabled nodes.
  • 8b20dcee: gate T8122 G15G identity before startup (garymeyerca).
  • 84e34610, 8231fa35, 330a560b, 0d51a403: the T8122 register lists and HwData words (Satya Benson).
  • a0fe58a7: the T8122 remaining runtime InitData words (Chris Kearney).

The M3 code they feed is only reachable through M3 compatibles. The G13, G14 and G17 paths are unchanged.

Landed in the fast-forward updates (latest 644133b7799f)

  • Update Aurora kernel roadmap: priorities, current evidence and acceptance #12 (644133b7799f, review follow-ups and base alignment). It has integration/aurora 339ea0487dfc as a parent and carries ASoC: apple: select the Neo speaker codec with macaudio #149 (MAX98357A select for macaudio).
    • GPU UAPI matches integration/aurora. include/uapi/drm/asahi_drm.h now uses the merged numbering from 0c986f352044 (G17P feature bits 10–15, timestamp-copy arrays, VM_CREATE flags). A layout check proves all 499 of its symbols equal to 339ea0487dfc's header. The only diff is two comments describing this tree's behaviour (GEM_MADVISE, RSRC_SPEC_HI). Without this, merging would have reverted the base's UAPI and broken current Aurora Mesa (aurora-silicon/mesa aurora-wip). Older Mesa still runs on it through its existing fallbacks.
    • A DCP that doesn't start keeps the boot framebuffer. The panel's DCP has to come up before simplefb is removed; an external DCP never stands in for a failed panel DCP. A T8122 dcpext whose PMP request isn't acknowledged is left off.
    • Failed PMP start: documented recovery, debugfs start state, and KUnit fault injection (12 cases).
    • Touch ID readiness accounts for the xART guard. Diag gains xart_guard, xart_guard_list, xart_guard_refusals and matching (usable / unverified / blocked); aurora-sep-ready reports NOT ready for configurations the guard is known to block and names the xart_guard=0 opt-out. The diag ABI stays 2 (keys are only added).
    • apple-isp: surfaces the firmware releases are now freed at stream stop. With T8140's resident firmware they were only freed at unbind, so each camera session leaked ~100 MB of memory and ISP IOVA until capture failed with ENOMEM.
    • XART-SAFETY.md full-list wording and the sensor-wake test comment are corrected.
  • Update wifi: brcmfmac: AWDL support for the BCM4387 (AirDrop link layer) #11 (cdc7cd13b1b4).
    • drm/asahi (with the Rust shmem/gpuvm abstractions):
      • an imported dma-buf's SG table is left to its attachment, instead of being unmapped and freed by the importer;
      • unmaps that split two mappings preallocate enough GPUVAs;
      • a failed exec_lock() no longer leaks the VM.
    • Hardware result: on 7.2 without the first fix, every GStreamer camera → GL close on J700 oopses in vb2_dma_sg detach. With it, heap, udmabuf, ISP camera and AVD decode imports pass on a KASAN + kmemleak kernel, with 0 reports.
    • 25G83 runtime: iomfb_v26_6.c finds the display subsystem through disp0, so the 25G83 DCP passes its handoff check.
    • Bindings: apple,pmp, apple,pmp-report, apple,no-ack and apple,soc-device-record are documented. dtbs_check is 7,153 lines, below Support USB4/TB DisplayPort Tunneling #9.
    • Config: RESET_APPLE_CIO defaults to ARCH_APPLE, so a 7.1 config no longer loses the USB4 routers silently.
  • M3 native display on ordinary 14.x boots (update media: apple: avd: update decoder and improve playback reliability #10). On J613/J615, the display gate now completes the tree's own PMP topology with the boot loader's per-Mac values, and the PMP report driver starts the halted PMP and answers apple_pmp_report_wait_ready(), so the DCP takes over from the boot framebuffer. dcpext is admitted the same way. M3 Max (T6031) works the same way but is opt-in with apple_t6031_display.enable=1. Authorship of the carried gate and provider commits is kept.
    • Superseded: on T8122/T6031 the gate reuses the PMP report, DART and mailbox nodes this tree already describes instead of applying the PMP overlays (which this tree would refuse with -EEXIST), so those two overlays are dropped. T6030 keeps the overlay path unchanged. 25G83 coldplug stays in drm/apple; only the gate's 25G83 skip is carried.
    • Not carried: the M3 GPU hunks of those commits, because every T8122 tree keeps the GPU disabled. They go with the M3 GPU runtime.
  • The 64 review fixes from Merge aurora-wip (carries #131) with review fixes #136 (Chris Kearney), ported onto the 7.2 topics.
  • The commits listed under "Carried with a provenance note".
  • The T8112 SEP DART window plus the xART write guard (M2 Touch ID). The DT part was already carried, from 6d52684728eb.
  • Touch ID sensor attach: the serial and nonce are read before registration, and attempts are spaced and bounded with one late retry.
  • Touch ID ready reporting: the sensor's bind is reported to diag, so aurora-sep-ready says ready only once the enrolment is restored, and NOT ready with a reason otherwise.
  • Everything in aurora-wip not yet in this base is now ported, including the 5-cell/legacy DART window KUnit test, the SELinux label for /dev/sep-bio, the authselect PAM hint and the XART-SAFETY cold-boot note. A per-commit audit of all aurora-wip work against this branch found 4 commits missing, and they are the ports above. The other 256 are present under another subject, superseded, already upstream in Asahi 7.2.10, or intentionally dropped. A follow-up hunk-level audit of every file aurora-wip changes (1,176 files, 7,337 hunks) found nothing else missing: 6,684 hunks match exactly, 441 are adapted to 7.2, 205 are superseded with a cited replacement, and 7 are recorded drops.
    • the J700 AVE debugfs encode transport is not carried: it has no stable ABI, needs !SUSPEND so no shipped config can build it, and stays parked on research/ave-transport-7.2.10.
  • M1/M2 Touch ID: the sensor status is re-read after registration on every profile, as on 7.2.10. That removes the session-share refusals seen on M1 Pro boots.
  • integration/aurora arm64: handle Apple M4+ wait state without boot flags #147: Diljit Singh's "arm64: handle Apple M4+ wait state without boot flags", carried unchanged (the J700 keeps the register-preserving WFI path, and arm64.nowfxt is no longer needed).
  • Review fixes (Rebase onto Asahi 7.2.10: one kernel carrying integration/aurora and aurora-wip #139 review): the xART guard list is capped at the length the boot-time read accepts; a sensor bring-up stops between protocol stages when the system suspends; KUnit covers a refused CLEAR_STATE re-registration; XART-SAFETY.md documents the guard per platform; the J613/J615 25G83 display coldplug finds its nodes by alias and phandle.
  • Touch ID attach retries: after the serial is registered, a retry clears the enclave's session state (CLEAR_STATE) and repeats the whole sequence, including share and commit.
  • Display and USB fixes: the DCPEXT reservation-owner warning, every deprecated-workqueue user moved to system_dfl_wq, the M1/M2 Type-C replug relink, the ATC KUnit objects built as modules, and missing MODULE_DESCRIPTIONs.

Validation

Build gates (head 644133b7799f; base asahi-wip v7.2.10; compared with 7.2.10 rebase-wip/7.2.10 c6d6d0964f73)

Gate Result
J700 build (Aurora's published config, LLVM=1) PASS: 0 compiler warnings, 2,508 modules. It is running as the Neo's daily (see Hardware)
M1 Pro / M2 builds (Fedora stock-config recipe) PASS
W=1, J700, the 109 touched directories PASS: 0 new lines vs 7.2.10
KUnit, the union of every topic's suites (37 configs) PASS: 0 failures. SEP 132/132 (+1 skip), DCP 315/315 (+1 skip), GPU 625/625, Type-C/USB 296/296, DART 22/22 (+1 skip), and more. The known apple-drm-resume fixture still stalls under lockdep; every other DCP suite passes there
apple-dcp host selftests (run.py) PASS: 56/56
tools/aurora-sep tests PASS: 767 tests, 0 failing
dtbs_check Compared with this PR's base integration/aurora 5d570a78a4ab: no new warning comes from Aurora's commits, and J700 has no new line. The 337 new lines (after node moves) are all carried: 290 are Asahi 7.2.10's own warnings, 34 are Asahi t8122.dtsi warnings on the base's J613/J615 profiles, and 13 already exist in substance on the base. The T8122 GPU, mailbox and UAT nodes stay disabled
checkpatch --strict 0 errors in this PR's new and adapted commits. Inherited classes are listed: (cherry picked from …) references, and contributor commits without a nominal-author Signed-off-by
Hygiene 0 Co-Authored-By on Aurora commits (contributors' own trailers are kept) · every Aurora commit ends with Assisted-by: LLM + Signed-off-by · 0 AI-instruction files · 0 private paths or identifiers in diffs or messages

Hardware

MacBook Neo (J700), first boot of this tree, compared with the 7.2.10 daily:

  • Driver binding: identical on every platform/SPMI/I²C/SPI/PCI device: DCP×2, apple-drm, the G17P GPU, the SEP, ISP, AOP audio, ALS, PMP, and MT7932 Wi-Fi/BT.

  • External display, 180 Hz over the rear dock, software-driven:

    • DPMS ×10: picture back in 1.8 s median;
    • mode change ×5: 0.9 s;
    • s2idle with the monitor on ×6: 1.4 s;
    • 0 drops.
  • Short sweep: display, GPU (Vulkan/EGL/GLES via Zink/Honeykrisp), Wi-Fi, BT, Touch ID (restored 3 on the first attempt), camera, mic, USB, wake and thermal all pass.

    • Trusted-key seal/unseal works.
    • 3×60 s s2idle at 99.2% residency.
    • AOP capture with 0 xruns.
    • 0 WARN/BUG.
  • Regression sweep, essentials (same scripts as the 7.1 baseline):

    • DCP s2idle 6/6, with the mailbox re-entry rate at baseline;
    • 3 boots at ~14 s to userspace, the same as 7.1;
    • storage 15/0 with data intact across s2idle;
    • s2idle ×5;
    • input 12/0;
    • Wi-Fi e2e on 5 GHz and 2.4 GHz plus a BT reconnect;
    • power/thermal: trip points, IPA policies and cpufreq ladders identical to 7.1;
    • ISP streaming before and after s2idle;
    • speaker playback through the protected path, with the speaker lock held throughout;
    • HQ/LQ mic capture.
  • It was the Neo's daily kernel until the updates below replaced it.

  • The latest update (ffa615e26f29) on the Neo: it passes the same kit and is now the Neo's daily.

    • Driver binding is identical, with 0 WARN.
    • The DCPEXT reservation warnings and the deprecated-workqueue notice are gone.
    • The xART guard reports off on T8140, by design.
    • DPMS 10/10, mode change 5/5, s2idle with the monitor 6/6, the short sweep, and Touch ID restored 3.
  • Attended checks on the Neo, on this kernel: all pass.

    • USB keyboard wake with the lid open and closed (clamshell, external display);
    • Bluetooth keyboard wake;
    • real-lid clamshell (the internal mic and the Touch ID reader hide while closed and return on open);
    • Touch ID verify, lock/unlock and sudo;
    • a cold power-on to the desktop with the external display in about 27 s.
  • The current head (a9712ec4665e) on the Neo: it passes the full kit. Driver binding is identical, with 0 WARN. DPMS 10/10, mode change 5/5, s2idle with the monitor 6/6, the short sweep before and after sleep, about 99.2% s2idle residency, and a 5-minute idle check matching the previous head. Touch ID restored 3, including a first-attempt failure that the retry recovered.

  • M1 Pro and M2 on this tree's code (the head before the final M4+ wait-state commit, which does not apply to M1/M2): the full battery passes on both, and both run it as their daily. The M2 has the SEP on and the xART ledger unchanged (0 of 10).

M1 Pro (J316s) and M2 (J413) on 7.2.10 (the same code base minus the unified carries):

  • M1 Pro: the full battery passes. That covers tipd, USB4 routers, Type-C cycles, pm_test, the display DART attach, the GL smoke test and the ANE probe.

  • Touch ID verified on the M1 Pro, with the xART ledger unchanged.

  • M1 Pro on this tree (e0fa50d863ee, the Fedora stock-config recipe): the same battery passes, matching 7.2.10.

    • tipd 4/4 with nothing deferred;
    • all USB4 routers bound;
    • Type-C cycles 10/10;
    • pm_test=platform 3/3 with 0 DART faults;
    • 3 DCPs with eDP connected;
    • GL smoke test, ANE probe;
    • Touch ID restores its enrolment, with the xART ledger unchanged;
    • 0 Oops.

    The only new dmesg line is an intermittent brcmf_get_assoc_ies -52. It is now the M1 Pro's daily kernel, with 7.2.10, the 7.1 Touch ID kernel and stock as GRUB fallbacks.

  • M1 Pro, attended: Touch ID verify matches; a USB 3 camera in all three ports, both orientations, enumerates at SuperSpeed 7/7.

  • M1 Pro external display through a USB-C dock (DP alt mode), attended:

    • monitor power-cycle and input switch recover;
    • a dock replug recovers in 2 of 3 tries, through the Type-C relink path, in about 7 s.
    • Root-caused: in 1 of 3 replugs the port came back USB 2 only, without DP. A stale CRTC power-off was releasing the replacement Type-C display session. The fix extends the existing session-ownership guard to every Type-C output, and has landed.
  • M1 Pro on the 7.2.10 DTB set (one-shot boots; the daily runs the 7.1 DTB set, which this kernel also supports): 7 of 8 boots of this tree came up with display, Wi-Fi and tipd 4/4. One hung before the root filesystem, with no log, and the watchdog recovered it; it did not reproduce in 6 retries. The 7.2.10 baseline came up 4 of 4. Every boot shows the already-known late-IOMMU probe race, which moves between devices.

  • M2 on this tree: the same battery passes with 0 warnings, and the xART ledger is unchanged (0 of 10).

  • M2 with the SEP on (this tree plus the xART guard and the T8112 SEP DART window, now landed):

    • the SEP attaches with the guard at level 1, and all 10 existing session records are protected;
    • Touch ID restores its enrolled identity across reboots;
    • the xART ledger shows 0 of 10 records changed and 0 added, before, after, and across a plain reboot;
    • 0 warnings.
      It is now the M2's daily kernel.

Known issues and follow-ups

  • M3 native display (J613/J615), regression against integration/aurora.
    • 25G83 profile (J613; J615 only with asahi,j615-25g83-experimental): the coldplug now finds all ten display/PMP nodes through aliases and phandles, so the 25G83 native display can be admitted again.
    • Ordinary 14.x boots: native display is restored in update media: apple: avd: update decoder and improve playback reliability #10 (see "Landed"). T6031 is opt-in.
    • Verification: static only. Builds and W=1 are clean, KUnit passes 68/68 for the gate and validators, and a host test runs the unmodified gate and the provider's lookups against the compiled J613/J615/J514c/J516c DTBs (44/44, with mutated trees refused). Not tested on M3 hardware, and the PMP tables in that test are synthetic.
    • dtbs_check: the bindings landed in update wifi: brcmfmac: AWDL support for the BCM4387 (AirDrop link layer) #11, so this no longer adds lines.
    • 25G83 runtime: fixed in update wifi: brcmfmac: AWDL support for the BCM4387 (AirDrop link layer) #11 (disp0 lookup in iomfb_v26_6.c).
    • M3 PMP start failure. On native-display boots Linux starts the PMP that iBoot left halted. If the start fails after the PMP is enabled, it keeps running until reboot. There is no PMP shutdown handshake, and halting a running PMP would freeze the SoC power states it manages. Recovery: reboot. To boot without starting it, use apple_t8122_display.enable=0 (T8122) or nomodeset. T6031 only starts it with apple_t6031_display.enable=1. The display stays on the boot framebuffer in that case. Verified with KUnit fault injection only.
    • M3 hardware qualification (J613, J615 opt-in: cold boots, PMP start timeouts, DCP failure after the PMP acknowledges, s2idle, dcpext) is owed before rc.
    • Config: APPLE_T6030_DISPLAY_GATE now defaults to y on Apple. It returns immediately without an M3 handoff. The J700 release config pins it off, so its config is unchanged.
  • M1/M2 Touch ID first-attempt reliability (regression vs 7.2.10). On the M1 Pro (J316s), the first sensor bring-up after boot fails more often than on the 7.2.10 kernel: about 1 boot in 4 here, against 0 of 21 on 7.2.10 in an interleaved A/B. The attach retry now clears the enclave's session state and recovers most of these, so Touch ID came up on 28 of 31 recent boots, and a reboot recovers the rest. The xART records and the enrolment are never touched. The rate is bursty: in a later 32-boot interleaved run, the unmodified head failed its first attempt on 1 of 8 boots, and three candidate-revert kernels on none. Every boot restored Touch ID. This is accepted for an edge merge. Finding the cause needs a longer instrumented soak, tracked as a follow-up; the fix will follow as a fast-forward update. The M2 (J413) and the Neo (J700) have not shown an unrecovered boot.
  • J700 rear USB-C SuperSpeed: after a cold boot, a dock's SuperSpeed side doesn't retrain until it is physically replugged. Warm reboots keep it up (11/11), and after a replug it runs at full 10 Gb/s. A fix is still being worked on.
  • HPD v2 (CD321x DP hotplug) is inert on J700, because a DT displayport link has no matching DRM connector there; adding one would give no extra function. It applies to M1/M2 CD321x ports.
  • xART guard on M1-family Macs. With the default xart_guard=1, restoring Touch ID at boot works on every platform, because it only reads xART. Using it does not always work: a fingerprint match rewrites session records, and the guard allows that only for records Linux created while the guard was on.
    • Cold-boot Macs (M1, T8103; M1 Pro/Max, T6000/T6001): creating the keybag also rewrites a session record from before Linux, so setup needs xart_guard=0. Because setup ran unguarded, the matches that follow need it too.
    • M2 (T8112): keybag creation makes no xART writes.
    • Untested under the guard: M2 Pro/Max (T6020/T6021) and M3 (T8122).
      A refused write fails closed and is logged; tools/aurora-sep/XART-SAFETY.md covers restore, use, setup and recovery per platform.
  • Boot.bin DT set: boot.bin holds one DTB set for every GRUB entry. 7.1 kernels need aurora-wip's 5-cell DART backport (now merged there) to run on this tree's DTBs.
  • M1 Pro on the 7.2.10 DTB set: one unexplained early hang in 8 boots; see Hardware.
  • M3 GPU runtime (firmware-event lifetime, GPUVM): it needs its maintainer's redesign on 7.2. Only the compile-only, gated backend is carried here.
  • 240-isp: the AVD part of 107036c9c167 and the ISP part of 7f835188accc still need owner reconciliation.

How to review

Review the canonical assembly, the second parent of the head, not the first-parent chain. The first parents only link earlier fast-forward publishes, so their topic merges are stale copies. The canonical is Asahi 7.2.10 plus one merge per topic, in NNN order.

C=644133b7799f^2          # or <head>^2 after later updates
git log --first-parent --format='%h %s' 6d20a3172ffe..$C            # the topic merges
T=200-dcp                                                           # any topic
M=$(git log --first-parent --format=%H --grep="^Merge branch 'unified/bits/$T'" 6d20a3172ffe..$C | head -1)
git log --reverse --no-merges --format='%h %an %s' $M^1..$M^2        # the topic's commits, in order
git diff $M^1 $M                                                    # its net effect
git diff 6d20a3172ffe $C                                            # the whole kernel over Asahi 7.2.10

Because the canonical is reassembled for each update, git rev-list over the whole head counts every topic commit several times.

What from integration/aurora is carried, deferred or dropped: the head has integration/aurora as a parent, so git counts all of its commits as merged even where the content isn't carried. A per-commit list (1,389 commits checked; 153 deferred, most of them the M3 GPU runtime, 99 superseded, 3 dropped) is here: https://gist.github.com/RyanTheTide/980fce5c903a3b158b37a45432b47abd

Please merge with a merge commit (or a fast-forward), not squash or rebase, so the topic structure and authorship survive.

🤖 Generated with Claude Code

Ryan Murray and others added 30 commits October 18, 2026 13:48
The user-subpage pool, owner and retired-mapping frees, the BPF pin
release and the VFIO ioctl test worker are queued on system_unbound_wq,
which is deprecated in 7.2: the first item queued on it logs a warning
once per boot and hides any later user.

Queue them on system_dfl_wq, the same unbound queue under its current
name. Ordering, references and teardown are unchanged.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
…command encoders)

Carry only the M3 TA and 3D command encoders, m3_tiler_command.rs and
m3_fragment_command.rs, and the render geometry and state helpers they
use, agx_render.rs and agx_render_state.rs. The encoders were held out
of the earlier M3-only splits because they share files with the held
G15G register-list change; they are carried now so that it can follow.
The backend is compile-only: nothing constructs a command, and no
G13/G14/G17 path or UAPI is changed.

[ryan: adapted to 7.2.10 — minimal, pending DJ review: isolate these
modules under m3 (crate:: becomes super::). The encoders are as of the
second source commit below, which also holds their origin in the first,
without the depth-compression bindings that Eryk Wieliczko added before
it; those are carried next under their own authorship. The state helpers
are as of the third source commit. No runtime integration is claimed by
this split.]

Source: commit bbefc95 ("drm/asahi: add the M3 (G15S) GPU runtime")
Source: commit 91fd07a ("drm/asahi: excise unbound G16 and G17 runtimes")
Source: commit 65c2a80 ("drm/asahi: remove unused backend helpers and narrow layout visibility")
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
Populate native ZLS metadata registers and spill mirrors from the existing GPUVM-validated UAPI. Retain rejection of stencil compression and compressed multisample/layered jobs. Native 26A428 register snapshots identify the field locations; the opt-in single-layer D32 Minecraft replay renders pixel-identical frames on M3. All nine serializer tests pass, including the sixteen unchanged whole-command fixtures. Broader Mesa compression qualification remains separate.

[ryan: adapted to 7.2.10 — carry only the 3D command encoder hunk, now
under m3/; the GPUVM/UAPI side of these bindings stays with the M3
runtime port]

(cherry picked from commit a8b0d26998992c1db4f3bb8092c7bd405fbf2c1d on Eryk Wieliczko's eryk_m3_gpu branch; driver changes only, lab tooling omitted)

(cherry picked from commit 217822c)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
modpost warns that scan_timeout_test.o has no MODULE_DESCRIPTION() when
the brcmfmac KUnit tests are built as a module. Add one.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
modpost warns that dcp-fabric-test.o has no MODULE_DESCRIPTION() when
the fabric tests are built as a module. Describe it like the other DCP
test modules.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
[ryan: adapted to 7.2.10 — the M3 probe arm refuses before resources;
replace its pending-provenance T8122 refusal with this read-only
identity gate in m3::refuse_probe() and keep -ENODEV; Core gains its 7.2
lifetime; of_node_put() is called only with CONFIG_OF_DYNAMIC, as in
kernel::of::Node]

(cherry picked from commit 5b89559a4f424bf2ce5dcdd71ea572e86e770ec0 in iconidentify/m3-linux air/t8122-g15g-admission)

[Adapted to aurora-linux: driver.rs has no T8140 path here, so the T8122
gate is the first check in refuse_agx3_probe().]
Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
Claude-Session: https://claude.ai/code/session_01FwdG63dRaMrRbYrUBkSa1i

(cherry picked from commit 8b20dce)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
Take register windows, IO mappings, UAT geometry, runtime HwDataB
words and optional userspace features from the admitted SoC. Preserve
the T6030 mapping tables and configuration words. Reserved UAT tables
require the registered 42-bit hardware configuration.

Describe the T8122 C0 firmware layout and check its OPP table using
voltage-sorted primary and secondary frequencies. Leave missing T8122
hardware, IO and setup values unset so startup remains disabled. Require
the SMC die thermal zone for every T8122 board without widening board
admission.

Validate IO ownership, range bounds and firmware entropy-tag bounds
before mutation. Add host tests for admission, OPP shape, IO ownership,
firmware layouts and thermal-zone requirements.

[ryan: adapted to 7.2.10 — the rest of this change is carried as an
M3-only split; carry the held-file remainder: the operating-point shape
check of t8122_admission.rs, the compile-only subset of the per-SoC
table (SoC names and HwDataB words, with G15RuntimeHwDataB kept M3-local
instead of in the shared initdata.rs), and the host-test entry limited
to the carried modules]

Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
(cherry picked from commit 88e5636)

(cherry picked from commit b2a9b8f)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
The G15G firmware takes G15S's compute register list without its last
two entries (0xa599, 0xd411): the same 20 registers in the same order,
the same 12-byte entries at +0x20, count and length at +0x728 and the
same control word. Add a per-SoC RegisterSet and send the G15G list on
T8122. The T6030 list is unchanged.

[ryan: adapted to 7.2.10 — carry the compute command register lists and
the per-SoC register set into the compile-only m3 modules; the Compute
job and runtime call sites (m3_compute.rs, m3_runtime.rs) stay with the
M3 runtime port]

(cherry picked from commit 42f5431f326dd25ccb4185593f2dc934386ab077)

Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
Claude-Session: https://claude.ai/code/session_01GWRhUgND1qhew6aEd6vsj8

(cherry picked from commit 84e3461)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
The G15G firmware takes G15S's TA register list without 14 of its
registers and the 3D list without 4; the rest keep the G15S order and
entry layout, and the TA/3D command and microsequence layouts are
G15S's. Encode the G15G lists on T8122, from the SoC's RegisterSet.
The T6030 lists are unchanged.

[ryan: adapted to 7.2.10 — carry the TA and 3D command register lists
into the compile-only m3 encoders; the Render job and runtime call sites
(m3_render.rs, m3_runtime.rs) stay with the M3 runtime port]

(cherry picked from commit a1ba5a7da2abcb516987ebb9cafc7f957abb66b8)

Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
Claude-Session: https://claude.ai/code/session_01GWRhUgND1qhew6aEd6vsj8

(cherry picked from commit 8231fa3)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
The G15G firmware takes T6030's runtime HwDataB configuration words and
unit masks except for four words: +0xa2c is 2 (the chip revision's major
number), +0xb20 is 0x0a (ten core slots, where T6030 has twenty),
+0x17b8 is 4 and +0x1818 is 0xffffffff. Give the T8122 table those
words, and the T6030 unit masks, 0x700000003 and 0x7.

The T8122 start experiment now uses the table's words. Its unit-mask
defaults become the table's masks; before, the knobs refused mask A's
second bit, which the G15G firmware runs with. The T6030 words are
unchanged.

[ryan: adapted to 7.2.10 — carry the T8122 HwDataB words into the
compile-only per-SoC table; the start experiment and its unit-mask
defaults (t8122_start.rs, t8122_knobs.rs) stay with the M3 runtime port
and still need this table change]

(cherry picked from commit 562b15fac35243fc15b3d66e3cfc36b1b47dd9af)

Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
Claude-Session: https://claude.ai/code/session_01GWRhUgND1qhew6aEd6vsj8

(cherry picked from commit 330a560)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
A word-by-word comparison of the T8122 InitData that the runtime builds
from a J613's device tree against a J613's working InitData shows the
same voltage, frequency, IO-mapping and controller words, and a short
list of configuration words where the G15G firmware runs with other
values than the shared builder writes:

- HwDataB +0xa30 and +0xa34: 0 and 4 (the builder writes 1 and 0);
- HwDataB +0x17e0..+0x1860: the G15G flag words (the builder writes
  the T6030 ones, 1 at +0x17e8, +0x1804, +0x1814 and +0x1860 and
  all-ones at +0x1848);
- HwDataA +0x11e0, a shader-engine controller word: 30 (40);
- HwDataA +0x1290: 125, and HwDataA +0x424c: 24000000, the 24 MHz
  reference clock (both 0);
- Globals +0x9bc: the CDM backoff timeout 4 with the next three bytes
  0 (the builder writes a 1 after it);
- the Globals smart idle-off curve and count, which the T8122 hardware
  configuration left empty: the T6030 values.

Give the SoC table per-object word lists and the T8122 values. The
words are SoC configuration, not per-Mac values. T6030 has empty lists
and its hardware configuration is unchanged.

[ryan: adapted to 7.2.10 — carry the T8122 HwData and Globals word lists
into the compile-only per-SoC table; the InitData builder writes
(m3_adt_config.rs) and the smart idle-off curve of the T8122 hardware
configuration (hw/t8122.rs) stay with the M3 runtime port]

Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>
Claude-Session: https://claude.ai/code/session_01GWRhUgND1qhew6aEd6vsj8

(cherry picked from commit a0fe58a)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
[ryan: adapted to 7.2.10 — the rest of this change is carried as an
M3-only split; carry its held-file remainder, the named flush-stamp
offsets of the TA and 3D command encoders]

Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>

(cherry picked from commit 179d09e)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
Write the tiler sequence length at 0x778 and the fragment sequence length at 0x1c. Share those offsets with the command encoders so optional measurement packets preserve both command pointers and all other descriptor fields.

[ryan: adapted to 7.2.10 — the rest of this change is carried as an
M3-only split; carry its held-file remainder, the stage-specific
sequence-length offsets of the TA and 3D command encoders]

Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>

(cherry picked from commit 6ea5927)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
On an M3 MacBook Air (T8122) with the G15 14.8.3 GPU firmware, the GPU
start stalls at the first device control: the 0x13 acknowledgement
times out and the probe fails with -110. It did so in every boot
reported in omacom/linux-aurora#26.

The start succeeds when the 20-word block at HwDataA +0x418c..+0x41db
holds these little-endian u32 values, with nothing else changed:

  1, 500, 5, 12500 x5, 3000 x3, 12500 x8, 1

The values are not in the device tree. The range is covered by the
G13-era core_leak_coef[4..8], sram_leak_coef and aux_leak_coef fields,
which the G15 builder leaves at zero on T8122; the values are integers,
not floats. Compile-time asserts tie the block to exactly that range.

Write the block when the runtime backend builds the InitData on T8122
with the 14.8.3 firmware. Add asahi.t8122_hwdata_418c (0 or 1, default
1, read-only in sysfs): 0 leaves the range at zero for an A/B
comparison; any other value is rejected and the default stays. One
kernel log line says which applied. Other SoCs, the T8122 26.x (G16)
path and T8122 boots without asahi.t8122_start=1 are unchanged.

Tested on hardware (omacom/linux-aurora#26), with local builds that set
the same 20 words at the same offset:
- J613, 10 GPU cores, GPU firmware 14.8.3: the GPU starts, the first
  compute and render jobs complete, and OpenGL and Vulkan readback
  pass. The same kernel with the block off stalls at 0x13 again.
- J613, 8 GPU cores, same firmware: the GPU starts, OpenGL and Vulkan
  readback pass, and Hyprland renders on the GPU.
This commit itself has not been booted yet. J615 on the 14.8.3
firmware has not been tested.

Add host tests (tools/asahi/m3-runtime-tests.rs) for the values, their
exact placement with every other byte unchanged, a short buffer, and
the parameter parsing.

[Chris Kearney: rebased onto integration/aurora; comments, a test name and
the log line reworded to describe what the code does]

[ryan: adapted to 7.2.10 — carry the HwDataA +0x418c block
(t8122_hwdata.rs) and its host tests into the compile-only m3 modules;
the InitData write, its layout asserts (m3_adt_config.rs) and the
asahi.t8122_hwdata_418c parameter (m3_params.rs) stay with the M3
runtime port, so no module parameter is added for other GPUs]

Signed-off-by: Satya Benson <satya@benson.earth>
Signed-off-by: Chris Kearney <303316+iconidentify@users.noreply.github.com>

(cherry picked from commit 0d51a40)
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
…e driver

PHY_APPLE_ATC_KUNIT_TEST is a bool that builds the tests "into the
driver", and it allows KUNIT=m when the driver is a module. The
Makefile, however, adds atc-tunnel-test.o and atc-t8122-dp-test.o as
objects of their own under that bool. They are therefore built into
vmlinux while KUnit is a module, and the link fails with undefined
kunit_binary_ptr_assert_format and __kunit_abort, as with the Fedora
configuration (KUNIT=m, PHY_APPLE_ATC=m).

Include both files from atc.c next to atc-test.c instead, so that they
are built wherever the driver is, as the option says. Drop their own
module metadata, which the driver already provides. The tests came with
commit 2421d71 ("phy: apple: atc: add KUnit tests for the T8122 DisplayPort helpers")
and commit c30adcc ("thunderbolt: apple: Qualify display tunnels by connector wiring").

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
The integrated RTKit receiver dispatches the debug endpoint and acknowledges
admitted channel-zero OSLog counters outside the Neo/M5 and synchronous-log
policies. The host fixtures still describe the older opaque-only receiver,
so oslog-events and ordered-rx fail before exercising their controls.

Extract the current OSLog constants and observe counter acknowledgements,
channel admission and the platform no-ACK policies. Add the debug endpoint,
policy field and receiver stub to the ordered-RX fixture. Keep its threaded
reply-progress and deferred-dimensions assertions intact, and retain support
for the legacy receiver used by the 7.2.10 topics.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
DEBUG_WW_MUTEX_SLOWPATH can make an atomic state lookup return -EDEADLK.
The resume fixture asserts on that expected retry result while holding
modeset locks. KUnit then aborts the test thread with those locks held,
and teardown cannot make progress. The lockdep lane reaches this in the
fourth resume test even though the ordinary lane passes.

Build both the initial scanout and the DPMS-off transition through one
helper that releases each partial atomic state, backs off and retries.
Always finish the acquire context before returning to a KUnit assertion.
Keep all four resume controls and their framebuffer ownership checks.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
A dock can detach and reattach while userspace is still processing the old connector loss.  On M1/M2, that stale CRTC disable released the already-starting replacement DPTX session.  Firmware then dropped the sink while the CRTC-off gate suppressed relink recovery, leaving only USB2 and the dock billboard device.

Apply the existing replacement-session ownership check to every Type-C output, and refresh the CRTC token after a successful modeset so later intentional DPMS still releases the current link.  Cover the stale-disable and ordinary-DPMS sequences with a source-extracted host fixture.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
cd321x-pm.h now carries cd321x_dp_send(), which takes u32 arguments and
which the probe units do not call. The host units drop
<linux/types.h> and build with -Werror, so typec-probe and
probe-mux-pm-dispatch stopped compiling once the header gained it.

Give the host copy of the header a u32 typedef and mark
cd321x_dp_send() unused there. The header and the units' checks are
unchanged.

[ryan: integration fix for the unified assembly: the header change comes
 from 170-usb, the host units live here]

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
Fifteen call sites queue DCP work on system_wq or system_unbound_wq:
the backlight fade, the swap watchdog, the external stop retry, the
IOMFB command timeout and backlight retry, the DCPEXT scanout,
repattern and desktop work, the iBoot endpoint query, and the link idle
work of the v14.7 and v26.6 interfaces. Both queues are deprecated in
7.2. The first item queued on either logs a warning once per boot, as
dcp_backlight_fade_work does on J700, and hides any later user.

None of these workers needs the CPU that queued it: each takes the DCP
or link state through its own locks and references. Queue them all on
system_dfl_wq. For the work that was on system_unbound_wq this is the
same unbound queue under its current name.

The host selftest shim that compiles dcp_external_stop_retry() declares
the queue it names, so rename it there too.

Keep the delays, coalescing, cancellation and teardown unchanged.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
On one J700 boot in 29 (0ab402d0) Touch ID came up without its sensor.
The first attach-time attempt had REGISTER_SERIAL accepted and then
found the sensor silent when it read the session nonce. The second and
third attempts power-cycled the sensor, identified it, cleared the
enclave and were refused at REGISTER_SERIAL with status 0x16, and the
driver gave up: "authenticated attach incomplete; capture unavailable".
The Catacomb restore had worked, but nothing brought the sensor up
again until a verify, which the login screen never makes. Seven other
boots that day needed a second attempt.

A J314s was refused the same way after a failed sequence counter init,
as reported in

commit 01dd28c ("soc: apple: sep: retry a failed sensor bring-up from a fresh registration")

CLEAR_STATE does not end a serial registration whose session never
opened, and while the enclave holds one it refuses another with 0x16.
Re-registration after a completed session has not failed on an idle
machine. So:

- Read both sensor inputs, the serial and the nonce, before the enclave
  hears of the serial. A sensor that falls silent then leaves the
  enclave as it was. The bus sees the same commands in the same order.

- When the enclave may still hold an unfinished serial registration,
  because a later step failed or its reply was lost, a 0x16 refusal of
  the new one is the enclave saying so: open the session on the one it
  holds instead of giving up. Only a registration this boot submitted
  is resumed. If the enclave refuses SENSOR_NONCE too, the attempt
  fails as before.

- Space attach-time attempts 0.5, 1, 2 and 4 s apart, five at most and
  none starting more than 10 s after the first, instead of three back to
  back. A verify, which someone waits on, keeps three immediate ones.

- If the attach still falls short, try one more round 10 s after
  activation, once the boot-time bring-up is over, holding the Catacomb
  writer lock and the capture count as a verify does, so that it never
  overlaps an enrol, a verify, a deletion or a suspend. It runs at most
  once, is put off at most three times and is cancelled at unbind.

- Report the outcome in diag (ABI 2): sensor_attach (unknown, attached,
  retrying, not-ready) and restore (unknown, restored, empty, unproven).
  restore reads "restored" exactly when the driver logs "Touch ID:
  restored N", so userspace can wait for it rather than for the open.

The retry decisions are plain functions in registration.rs. KUnit runs
them against a model of the enclave that holds a serial registration
across CLEAR_STATE: a silent first attempt then success, repeated 0x16
refusals then success, a 0x16 refusal of an unfinished registration
resumed, and every attempt failing in both rounds, which ends in
"not-ready" within the budget. tools/aurora-sep/tests/test_verify_bringup.py
runs the real attach and verify methods with the new spacing.

The test-only sensor_test_fault parameter stops the first authentication
of a boot after REGISTER_SERIAL, so the recovery from an unfinished
registration can be exercised on purpose rather than once in 29 boots.

Nothing here logs SEP data: statuses, sensor states, step names and
timings only. The xART store and its anti-replay writes are untouched.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
aurora-sep-ready logged "Touch ID is ready after 0 s" on a boot whose
sensor attach had failed: the open of /dev/sep-bio succeeds as soon as
activation is done, however far the sensor got.

Decide from the driver's diag attributes instead. Ready means touchid
"ready", sensor_attach "attached" and restore "restored" (or "empty",
with nothing enrolled), the state in which the kernel logs "Touch ID:
restored N". While the driver retries the attach, wait for it, up to
50 s from the start. Otherwise log "Touch ID is NOT ready" with the
reason (the attach fell short, the fingerprints were not restored,
activation failed or the keybag did not load) and fail the unit; units
that only want it still start. A machine with no keybag provisioned is
reported as not set up, not as a failure, and a kernel without diag
ABI 2 is still handled.

tests/test_sep_ready.py runs the script against synthetic diag states,
including the one of boot 0ab402d0. Update the package checksums and
releases.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
…session

On an M1 Pro, 2 of 4 plain boots came up without the sensor: the
attach-time patch failed at the sequence counter init, with the sensor
silent about 1.3 s after power-on, and the driver logged "the sensor did
not wake at attach; the next verify retries the bring-up". The enrolled
fingerprints were never restored that boot, because nothing brought the
sensor up again before a verify, and the login screen makes none. The
next boots worked. The attach of these profiles made one attempt, and
the spaced and late retries added for nonce-session profiles did not
cover it.

Give every profile the same attach: up to five attempts, 0.5, 1, 2 and
4 s apart, and the late round 10 s after activation when they all fall
short, with diag's sensor_attach reporting it. An attach attempt on a
profile without a nonce session stays what it was: wake and patch the
sensor, without calibrating it.

The restart differs. A J314s that failed the same step, as reported in

commit 01dd28c ("soc: apple: sep: retry a failed sensor bring-up from a fresh registration")

refused the CLEAR_STATE and serial re-registration of every later
attempt for the rest of the boot, while a fresh registration recovered.
So on a profile without a nonce session each attempt after a failed one,
and the late round's first, starts from a fresh registration, as a
reboot does. Nonce-session profiles keep clearing the enclave first.
registration::retry_stage holds the decision, with KUnit cases for both
profiles and for a sensor that stays silent on its first attempts.

A verify on these profiles keeps its single attempt.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
…fied

Topic head 7beea60, 169 commits not already in the tree.
Top commit: arm64: dts: apple: j613-25g83: match the G16G GPU handoff layout
Topic head 6d6d919, 146 commits not already in the tree.
Top commit: selftests: asahi: cover the T8140 QSPI window in the host fixture
…ase-wip/7.2.10-unified

ad9d402 ("usb: typec: tipd: Descend into tipd/ whenever the core is enabled") fixes a link failure of
asahi-wip itself with the J700 configuration. Merged ahead of its topic so
that every first-parent commit of rebase-wip/7.2.10-unified links; the 170-usb merge brings
the rest of that topic.
Topic head a934b23, 42 commits not already in the tree.
Top commit: dt-bindings: iommu: apple,dart: allow DMA windows on all supported SoCs
Topic head 2314db8, 21 commits not already in the tree.
Top commit: nvme: apple: queue controller recovery on system_dfl_wq
…unified

Topic head 202286b, 48 commits not already in the tree.
Top commit: soc: apple: finish the merge of DART, QSPI and SMC hwmon
Topic head 085d961, 82 commits not already in the tree.
Top commit: ASoC: apple: t8140-aop-audio: queue work on system_dfl_wq
Ryan Murray added 13 commits October 19, 2026 14:19
Topic head cb7bda8, 83 commits not already in the tree.
Top commit: ASoC: apple: select the MAX98357A codec for macaudio
Topic head 7e1c8a0, 194 commits not already in the tree.
Top commit: wifi: brcmfmac: describe the scan timeout KUnit module
Topic head 78742b7, 51 commits not already in the tree.
Top commit: HID: magicmouse: document the host-driven click members
Topic head 9acd787, 382 commits not already in the tree.
Top commit: tools: aurora-sep: Refresh the Arch PKGBUILD sums for XART-SAFETY.md

Conflict resolved in MAINTAINERS (ARM/APPLE MACHINE SUPPORT): 060-power-thermal
widens the hwmon pattern to macsmc-hwmon*.c and 100-sep moves the
drivers/pmdomain/apple/ line to its sorted place; keep both, as aurora-wip has it.

Unified 900 review: retain the current macsmc-hwmon*.c coverage and drivers/pmdomain/apple/ entry in the sole MAINTAINERS conflict. Keep every nonconflicting SEP entry from the three-way merge. The narrower old hwmon filename and stale removal of pmdomain coverage are not improvements. Exact base/ours/theirs blobs are recorded in MAINTAINERS.sides.
Re-recorded when 010-soc added the pmp-report KUnit config line; the rule is unchanged.

drivers/soc/apple/aop.rs: keep the assembled side. 070-audio's
AOP service registration runs from its own registration queue behind the
removal gate; 100-sep's only change there is the system() ->
system_dfl() workqueue switch, which that queue already supersedes.
Topic head 2c399a1, 280 commits not already in the tree.
Top commit: reset: apple-cio: build the CIO reset driver by default on Apple

Unified assembly resolutions (900 lane):
- include/linux/soc/apple/pmp-report.h: keep both additions. 010-soc
  b56195a145e5 adds apple_pmp_report_wait_supplier_ready(); 170-usb
  2af69d67f3b3 (from 4336e52) adds apple_pmp_report_wait_ready().
  Both prototypes and both !CONFIG_APPLE_PMP_REPORT stubs are kept.
- include/linux/soc/apple/pmp.h (add/add): keep the 010-soc header. It
  declares the same apple_pmp_link_device() and
  apple_pmp_set_device_power() prototypes as 170-usb, plus their
  !CONFIG_APPLE_PMP stubs and the PMP driver's registration API, which
  170's subset omits.
Topic head 3b8f11d, 421 commits not already in the tree.
Top commit: drm/apple: do not start a T8122 dcpext the PMP has not acknowledged
Topic head 6d5b51d, 175 commits not already in the tree.
Top commit: drm/asahi: uapi: use integration/aurora's wording

Unified assembly: keep the current 010-soc RTKit file. The GPU topic changes only start_ep to return the send result relative to its merge base; current 010 already propagates that result and adds one retry only for an unsent -ETIMEDOUT. Retain all current OSLog Neo/M5/sync no-ACK policy, adopted session, kdebug and teardown changes. No GPU-specific RTKit delta is lost.
Re-recorded (900 lane) when the current side gained
apple_rtkit_flush_rx(); the resolution is still the current side verbatim.
The firmware allocates its working memory through the SHAREDMALLOC
channel, and frees it on the same channel by sending the surface's
iova. The driver acknowledges a free but keeps the surface on the gc
list, and frees that list only when the firmware stops, because the
firmware does not free everything it allocated and may still access a
surface after freeing it.

Resident firmware never stops while the driver is bound, so nothing
it frees is ever freed. On J700 every camera session allocates its
SHAREDMALLOC surfaces anew: MemAvailable falls by about 85 MB per
session, from 3.2 to 1.5 GB over 20 sessions, until the ISP's IOVA
range is used up. Then the firmware's 0x484000-byte requests fail
("SHAREDMALLOC: handler failed: -12"), IO commands time out, and
REQBUFS returns -ENOMEM because the capture buffers no longer fit in
the IOVA range. Firmware started for each stream is not affected,
since every stream ends with a firmware stop.

Handle the free request: move the surface with that iova to a list of
released surfaces, and free those when the stream stops, after the
channel has stopped and its buffers have been returned, or when the
firmware stops. Waiting for the end of the stream keeps the surface
mapped while the firmware may still be using it. Surfaces the firmware
never frees are still freed only when it stops, and a free for an iova
the driver does not know is still acknowledged. A released log or
buffer return surface is no longer used to read firmware messages.
The firmware itself stays resident, as before.

Test the free request handling with KUnit. Compile-tested and KUnit
only; not yet run on hardware.

Fixes: 16a04f7 ("media: apple: isp: Keep resident firmware running between streams")
Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
Topic head d58647b, 108 commits not already in the tree.
Top commit: arm64: handle Apple M4+ wait state without boot flags

Conflict resolved in drivers/gpu/drm/drm_gem_shmem_helper.c as in the 210/220
test merge 5bc189ffceb3: the shared drm_gem_shmem_object_offset() line once, and
in drm_gem_shmem_mmap() 220's byte-range validation before 210's revocation
check, which is aurora-wip's order.
Re-recorded after the 220 review fixes (#136): same rule, with 220's
validation now covering a partial native-page PRIME offset.
Topic head 282a6e1, 82 commits not already in the tree.
Top commit: media: apple: isp: Free the shared surfaces resident firmware releases
Topic head 0febca4, 29 commits not already in the tree.
Top commit: accel/ane: queue the boost release on system_dfl_wq
…-unified

Topic head 5810654, 91 commits not already in the tree.
Top commit: selftests: Register Aurora host tools and the AFKV2 target
…, MAX98357A

Bring publish/unified-7.2.10 to the tree of the canonical unified
assembly, as a fast-forward of the previous publish cdc7cd1. The
second parent is that canonical assembly; the tree is identical to it.
The third parent is integration/aurora 339ea04, so this branch
contains the integration head it targets and stays mergeable into it.

New since cdc7cd1:

- 210-gpu: include/uapi/drm/asahi_drm.h now matches integration/aurora
  symbol for symbol (the M3 and J613 runtime numbers, and the G17P
  numbers moved past them), with the runtimes themselves not carried.
  The M1/M2 queue ends each command at the new timestamp copies, so a
  command that uses them fails with EINVAL, as on G17P.
- 070-audio: integration/aurora's #149, the macaudio machine driver
  selects SND_SOC_MAX98357A, cherry-picked onto the audio topic.
- 200-dcp: a DCP that did not start leaves the boot framebuffer in
  place, and a T8122 dcpext whose PMP request was not acknowledged is
  not started.
- 010-soc: pmp-report documents a failed PMP start and its recovery,
  with KUnit fault injection for the start.
- 100-sep: diag reports the xART guard and whether matching can save
  state; aurora-sep-ready tells restored fingerprints from usable
  matching; wording fixes, and the Arch PKGBUILD sums refreshed.
- 240-isp: firmware-released SHAREDMALLOC surfaces are freed at stream
  stop on resident-firmware T8140.

Assisted-by: LLM
Signed-off-by: Ryan Murray <ryan@aurorasilicon.org>
@RyanTheTide

ghost commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

Thanks @maralcbr. Your follow-ups from both approvals are in updates #11 (cdc7cd13b1b4) and #12 (644133b7799f). Details:

From the #10 review

  1. 25G83 runtime path. iomfb_v26_6.c now finds the display subsystem through disp0 and checks its compatible. That was the last bad /soc/... path in drm/apple and soc/apple.

  2. DCP startup failure. simplefb now stays unless the DCP that drives the panel actually came up. "Any DCP ready" isn't enough on M3: dcpext reports ready straight away and would hide a failed panel DCP. A T8122 dcpext whose PMP request wasn't acknowledged is now left off instead of started.

  3. PMP after a failed start. Recovery is documented in the Kconfig help, the log and Known issues: reboot, or apple_t8122_display.enable=0 / nomodeset. On T6031 the PMP only starts with .enable=1. Two corrections:

    • nomodeset works because appledrm_register() returns before anything waits on the PMP. The gate still enables the nodes.
    • The SMC hwmon still runs, so the docs say "this driver does not check".

    There's a debugfs start state and 12 KUnit fault-injection cases. No safe stop exists: there's no shutdown handshake, and halting a running PMP would freeze the power states it owns.

  4. M3 hardware. Agreed. It stays under Known issues as owed before rc.

  5. Bindings. apple,pmp, apple,pmp-report, apple,no-ack and apple,soc-device-record are documented. dtbs_check is 7,153 lines, below Support USB4/TB DisplayPort Tunneling #9.

From the #9 approval

  1. Readiness vs the xART guard. The diag gains xart_guard, xart_guard_list, xart_guard_refusals and matching (usable / unverified / blocked). aurora-sep-ready reports NOT ready for setups the guard blocks and names the opt-out. An empty list is never read as "unguarded". The ABI stays 2 because keys are only added.
  2. XART-SAFETY.md full-list wording: fixed as you suggested.
  3. The test_sensor_wake.py comment: narrowed to what the mock proves.
  4. Counts: corrected in the description, with the revision and what's counted.

Two more things worth your eyes

Also in #12: an apple-isp fix for a per-session leak on T8140's resident firmware (~100 MB per camera session until ENOMEM).

Gates on #12:

Hardware: the Neo runs #11 as its daily. #12's Neo qualification, with the matching Mesa, is running now. I'll post the result.

The DRM_APPLE=y / SND=m link fix is still coming as a separate patch.

@RyanTheTide
RyanTheTide requested a review from maralcbr October 19, 2026 17:34
@RyanTheTide

ghost commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

Correction to my last comment: I said older Mesa keeps working on #12 through its fallbacks. It doesn't, and that applies to integration/aurora's own kernel too.

On the Neo, #12 with the pre-merged-ABI Mesa (agx3b-based) ran fine until a modeset about 100 s in. Then Zink got DEVICE_LOST and the kernel logged submit failed … EINVAL.

I traced the cause from the code; it isn't confirmed on hardware yet.

  • In the merged numbering, PER_COMMAND_INPUT_SYNCS moved from bit 3 to bit 10, so old Mesa no longer sees it and falls back to chaining every command onto the first.
  • Honeykrisp's chaining (hk_sync.c, first_sq) picks the subqueue from cmdbufs[0] only. A submit like [empty, compute, render] then gets a barrier on a subqueue with no earlier command, which g17/command.rs rejects with EINVAL.
  • 339ea0487dfc has the same check and the same bit layout, so this comes with the ABI change, not with our port.
  • Current aurora-wip Mesa has the same chaining bug, and hits it on kernels without the merged bits.

The fix is in Mesa: pick the first subqueue that has work, and only send the large ray-query compute command when the kernel advertises it. I'm building that now. #12 isn't the Neo's daily until it's qualified together with the fixed Mesa; the Neo stays on #11.

A kernel-side option would also keep old Mesa working: advertise bit 3 on G17P. @DjDeveloperr, that's your call, since it puts an integration bit to a new G17P use.

@RyanTheTide

ghost commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

#12 (644133b7799f) is qualified on the Neo and is now its daily kernel. It runs with Mesa from aurora-silicon/mesa#27: aurora-wip plus the first_sq fix and the ray-tracing gate.

  • GPU: GNOME and the display hotplug/unblank legs pass (10/10, 5/5, 6/6). There are 0 submit EINVAL and 0 GPU faults. VK_KHR_ray_query, acceleration_structure and ray_tracing_pipeline are present on the Apple GPU.
  • Same Mesa on the old-UAPI kernel (wifi: brcmfmac: AWDL support for the BCM4387 (AirDrop link layer) #11): it renders, with the RT extensions off, so the rollback kernels stay usable.
  • apple-isp fix: across 40 helper sessions plus 20 GStreamer camera sessions, memory drifts 0.43 MB per session, against roughly 100 MB before. All 3720 surfaces the firmware released were freed. The Zink camera import passes 20/20; it was 1/20 before the fix.
  • Touch ID stays ready across 5×30 s s2idle. Diag reads matching=usable with xart_guard=0.
  • One daily boot: the enclave refused SENSOR_ROOT_KEY (0x12e) on every attempt. The enrolment was intact, and aurora-sep-ready correctly reported NOT ready. The next two boots attached normally, and the host state, firmware and initrd were unchanged, so I'm treating it as transient and watching for it.

The M1 Pro and M2 haven't run #11 or #12 yet; they're still on the #7-era kernel.

@RyanTheTide

ghost commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

The ref Macs are on #12 as well. The M1 Pro (J316s, stock Fedora Asahi Mesa 26.1.3) and the M2 (J413, stock Mesa 26.2.3) both passed and run it as their daily kernel. GNOME, Vulkan and GL render on stock Mesa with the merged UAPI: no submit EINVAL, no device loss, 0 GPU errors. tipd replug held 10/10 on each, and 3× s2idle passed. Touch ID restores on both. Diag reads matching=usable on the M1 Pro (xart_guard=0) and matching=unverified on the M2 (guard level 1). The M2's xART ledger is unchanged, 0 of 10 records. All three machines (Neo, M1 Pro, M2) now run #12 daily.

@chriskearney

ghost commented Oct 22, 2026

Copy link
Copy Markdown

Thanks for the 7.2 port. We found a small merge regression while testing a two-display TS3 Plus on the M3 Pro: apple_dpin_m3 lost capacity_retry and defer_new_bringup from a29fbe8 in merge 44d5afc.

The second tunnel can reach DCP while the HDMI-capable processor is still in its settle window. DCP returns -EBUSY; without those flags, setup ends instead of waiting. Restoring just the two flags lets the tunnel stay alive and attach when the processor becomes available. The boot logs now show the wait followed by pipeline-free and tunnel-up. That is a routing result, not a claim that the physical monitor has a picture.

The fix and the original policy tests are in 068fd09 on #177. DP IN KUnit passes 35/35 and Thunderbolt 55/55; removing the flags fails exactly the four restored M3 cases. M1/M2 policy is unchanged. The larger PR is still a draft for separate hardware findings, but this should be useful as a small follow-up to the merge.

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

8 participants