[sithglan]

Base: integration/aurora (5186d2e). Branch thomas/j813-sound-int (114ebfd), 16 commits:
13 for sound and 3 that add KUnit tests.

The MacBook Air M5 (J813) DT in integration/aurora (t8142-j813-audio.dtsi) already describes
the AOP, its audio service, the four SN012776 amps, the CS42L84 and the macaudio card. Ryan's
eab3e66 also added the T8152-generation AOP and audio bindings. The drivers behind those
nodes never reached integration: they went into aurora-wip with #1000. On integration a J813
therefore has no sound card at all.

This ports #1000 onto integration/aurora:

• soc: apple: aop: T8152 support, and booting the T8142 AOP through its setup port, whose
  EPMAP names the service endpoints.
• The T8152 AOP audio driver, with the T8142 (speakers, jack) and the J813 microphone array.
• tas2764: a fixed BCLK/FS ratio (ti,bclk-ratio, which the J813 DT already sets).
  cs42l84: SDOUT1 routed from the ASP input.
• apple-admac: restart a paused capture at a frame boundary (the speakers' V/I sense).
• macaudio: J873 and J813.
• asahi.config: build the T8152 AOP audio driver, beside the T8140 one.

Differences from #1000:

• "dt-bindings: apple: Add the T8152 AOP and its audio service" is dropped. Integration has
  Ryan's version of both files (eab3e66), and the J813 DT follows it.
• "soc: apple: rtkit: Take a crash buffer the firmware maps itself" is dropped: integration
  already has it (ecbb337).
• The macaudio binding commit now adds J813 as well as J873. Integration's J813 DT names
  apple,j813-macaudio, which no binding listed.
• The T8152 and T8142 AopHwConfig entries take integration's fields in the commits that add
  them: ncal: None (the firmware keeps its own; neither entry patches boot arguments) and
  bootargs_word_access: false. J813: AOP sound on the MacBook Air M5 (speakers, jack, microphones) #1000 did this in a separate last commit. Here every commit
  builds on its own.
• The J813 and J873 macaudio tables carry no amp_gain. On integration the SN012776 takes each
  amp's level from its apple,speaker-config, which aura passes on the J813 (level 0x0f on all
  four amps). J873 has no safe level, so macaudio_speakers_allowed() keeps its speaker off.
  please_blow_up_my_speakers cannot override that on integration.
• The asahi.config line is new.

Not included: Acelogic's two J813 audio commits from #1048 (ad4910c "ASoC: apple:
validate J813 audio topology and use the default worker queue" and 12165bc "ASoC: apple:
initialize J813 output standby before microphone registration"). They have no Signed-off-by,
and the speakers, jack and microphones work without them (below). @Acelogic, if you want
them in integration, they apply on top of this.

The M1-M4 machines

The changes to shared drivers do nothing without the J813 or J873:

• tas2764: tas2764_parse_bclk_ratio() sets a ratio code only from ti,bclk-ratio, and
  tas2764_apply_bclk_ratio() writes nothing without one. The code starts at -1 in the I2C
  probe, before the device tree is read.
• apple-admac: admac_config_realign_bytes() takes a frame size only from a 4-byte
  peripheral_config. Only t8152_aop_audio.rs passes one, for its speaker sense capture.
  mca_hw_params() zeroes its slave config. admac_resume_realigns() holds for capture
  channels only.
• macaudio: the new spk_dai_fmt and spk_secondary_only fields are 0 in every other table,
  and the driver reads them only when they are set.
• cs42l84: two more DAPM routes into "SDOUT1 Select" (ASP RX CH1/CH2). The default route is
  still the ADC.

KUnit tests:

• tas2764-bclk-ratio (new):
  • Two amp nodes without ti,bclk-ratio, one with the ADT blocks the MCA machines get and one
    bare. Both keep the ratio auto-detection.
  • Their probe step makes no register write. The register map records every write, and
    TAS2764_CLK_CFG and TDM_CFG0 keep their values.
  • 250 BCLKs writes code 0x30 and sets only TDM_CFG0's fixed-ratio bit.
  • A ratio without a code fails the probe.
• apple-admac-realign (new):
  • A slave config built as mca_hw_params() builds it gets no frame size, and neither does a
    peripheral_config of another size.
  • A resume realigns only a capture channel that has a frame size and a transfer, never a
    playback channel.
• apple-speaker-admission: J813 is admitted (it has a safe level) and J873 is not (it has
  none).
• apple-speaker-amp-level: neither J813 nor J873 carries an SN012776 gain.
• apple-speaker-interface, a new case on the J813 table:
  • The volumes start at the volume lock's ceiling (1).
  • A daemon that holds the lock without pinging lifts nothing; after its ping the ceiling is
    201.
  • Under "Speaker Protection" the volumes are at 0. Without an owner a grant is -EPERM and no
    start is admitted. The owner's lease gives 201, and the release puts them back at 0.

Run in arm64 QEMU with this kunitconfig:

CONFIG_KUNIT=y
CONFIG_ARCH_APPLE=y
CONFIG_OF=y
CONFIG_OF_DYNAMIC=y
CONFIG_OF_OVERLAY=y
CONFIG_PM=y
CONFIG_PM_SLEEP=y
CONFIG_I2C=y
CONFIG_GPIOLIB=y
CONFIG_REGULATOR=y
CONFIG_COMMON_CLK=y
CONFIG_DMADEVICES=y
CONFIG_APPLE_ADMAC=y
CONFIG_SOUND=y
CONFIG_SND=y
CONFIG_SND_SOC=y
CONFIG_SND_SOC_APPLE_MACAUDIO=y
CONFIG_SND_SOC_TAS2764=y
CONFIG_SND_SOC_TAS2764_KUNIT_TEST=y

tools/testing/kunit/kunit.py run --arch=arm64 --make_options LLVM=1 --kunitconfig=<that file>
passes 84 / 84 tests in 11 suites. These suites are not in CI: its kunit job runs the SEP
kunitconfig only. As a check that the new cases catch regressions, I broke the drivers in
three ways:

• the TAS2764 writes CLK_CFG without the property;
• the ADMAC realigns playback;
• the J813 table loses its safe level.

Each of these fails the matching cases (6 of 84). At each test commit the run passes too (81
/ 81, 83 / 83). One run at the ADMAC test commit failed macaudio_spk_lease_expires, an
existing test with a 10 ms lease, while the host was building at full load. It passed in the
four runs after that.

Testing

On the J813 (macOS 27.0 stub), booted through aura-try:

• Kernel 7.3.0-aurora-j813-oni4s+: this branch merged with drm/asahi, drm/apple: J813 (M5 MacBook Air) GPU and display on the macOS 27.0 firmware #1413 (both on integration/aurora
  5186d2e). Built with CONFIG_SND_SOC_APPLE_T8152_AOP_AUDIO=m.
• Without these commits there is no sound card. With them there are AppleJ813 (Primary,
  Secondary, Speaker Sense) and AppleJ813HPAI (the microphone array), and dmesg has no AOP or
  audio errors.
• Speaker protection, with speakersafetyd running (its j813.conf): "Speaker Volume Unlock"
  reads 1, and each Speaker Volume's maximum is 201.
• After systemctl stop speakersafetyd:
  • macaudio logs "Speaker volumes locked: Lock timeout".
  • All four Speaker Volumes read max 1, value 1. Each Amp Gain Volume reads max 15, the
    level in its apple,speaker-config (00 0f 00 02, 01 0f 04 06, 02 0f 08 0a,
    03 0f 0c 0e).
  • Root setting a Speaker Volume to 201 gets 1.
• "Speaker Protection Interface" set to "Speaker Protection", with no owner:
  • All four Speaker Volumes are at 0, and "Speaker Protection" reads 0.
  • aplay of zeros to hw:AppleJ813,1 (the speakers) fails with "Operation not permitted",
    and speaker_starts_refused goes from 0 to 1.
  • The jack (device 0) still plays.
• Interface set back to "Speaker Volume Unlock" and speakersafetyd restarted: the volumes are
  unlocked again.
• The speakers play. A 3 s 1 kHz tone at 0.05 on hw:AppleJ813,1 reaches the three microphones
  at -50.1, -52.1 and -45.7 dBFS at 1 kHz, against -133 to -140 dBFS before the tone. The
  previous round measured -48.0, -49.1 and -42.9 dBFS RMS; J813: AOP sound on the MacBook Air M5 (speakers, jack, microphones) #1000 measured -50.5 dBFS.
• Integration's rtkit now logs "RTKit: crashlog endpoint not in the endpoint map" (info) for
  the AOP. The aurora-wip kernel did not log this line; it changes nothing.

CI checks, run locally:

• source: the aurora-sep tests pass on this base. On the old base, 064054d,
  test_diag_policy and test_diag_sensor fail on diag_shim.c's linux/bitmap.h, which
  Fix SEP diagnostics source checks and KUnit compilation #1438 fixed.
• Per-commit build with the J813 config (modules and dtbs, clang 22): 16 / 16 build.
• scripts/dtc/apple-dt-check.sh integration/aurora..this branch: 0 new warnings, 10 fewer.
  The fixed ones are t8142-j813.dtb's apple,j813-macaudio compatible and ti,bclk-ratio on
  its four amps, which are now documented. apple,macaudio.yaml and ti,tas27xx.yaml pass
  dt_binding_check.

Reviewers: @Acelogic @DjDeveloperr @RyanTheTide @LiamRay10

[RyanTheTide]

Needed before merge:

1. The failing source job is the diag_shim harness issue (linux/bitmap.h without a stub) that integration/aurora fixed in #1438; merging or rebasing onto current integration/aurora clears it.
2. tas2764.c, apple-admac.c and macaudio.c are shared with the M1/M2 machines: the ti,bclk-ratio path and the paused-capture restart need to be no-ops when the property is absent and for playback streams. Please add a KUnit or host test that covers the existing (no ti,bclk-ratio) TAS2764 configuration, or show it is unchanged.
3. For J813, show that the speakers stay behind the kernel speaker-protection interface (Speaker Volume Unlock / Speaker Protection controls) as on the other macaudio machines, so they are not driven above the safe ceiling without a protection daemon.

[sithglan]

Thanks, @RyanTheTide. All three are addressed. I force-pushed 3679438 -> 114ebfd, which is now 16 commits on integration/aurora 5186d2e. The description has the details.

1. Rebased onto 5186d2e, which is past Fix SEP diagnostics source checks and KUnit compilation #1438. The rebase changed none of the 13 commits (git range-diff: all =), and integration has not touched the files they change since 064054d. I ran the source job's aurora-sep tests locally: they pass now. On 064054d, test_diag_policy and test_diag_sensor fail on diag_shim.c's linux/bitmap.h, as you said.

2. The changes to shared drivers do nothing on M1/M2. I added KUnit tests for that (new commits):

   • tas2764: the property handling moved into tas2764_parse_bclk_ratio() and tas2764_apply_bclk_ratio() (squashed into the original commit). The new tas2764-bclk-ratio suite runs a register map that records every write:

     • Two amp nodes without ti,bclk-ratio, one with the ADT blocks the MCA machines get and one bare. Both keep the code at -1 and get no register write: CLK_CFG and TDM_CFG0 keep their values.
     • 250 BCLKs writes CLK_CFG = 0x30 and adds only TDM_CFG0's fixed-ratio bit.
     • 64 fails the probe.

     While doing this I also fixed the error path: a failed write now goes through err_sdz like the rest of tas2764_codec_probe(). It used to return with SDZ still high.

   • apple-admac: the two checks are now admac_config_realign_bytes() and admac_resume_realigns(). The new apple-admac-realign suite covers these cases:

     • A slave config built as mca_hw_params() builds it (zeroed, no peripheral_config) gets no frame size, and neither does a peripheral_config of the wrong size.
     • A resume realigns only a capture channel that has a frame size and a transfer, never a playback channel.

     Only t8152_aop_audio.rs passes the frame size, for its speaker sense capture.

   • macaudio: spk_dai_fmt and spk_secondary_only are 0 in every other table, and the driver reads them only when they are set.

   • Negative control: I made the TAS2764 write CLK_CFG without the property, then made the ADMAC realign playback. Each fails its new cases.

3. J813 speaker protection: macaudio_j813_cfg has safe_vol -20 on the SN012776. macaudio_speakers_allowed() therefore admits it, has_safety is set, and both interfaces apply from the bind. The "Speaker volumes unlocked" in my earlier test was speakersafetyd (it has a j813.conf on that machine), not the driver.

   • KUnit: apple-speaker-interface has a new case on the J813 table, using the fake card whose four volumes carry the J813 amps' prefixes:

     • At the bind the volumes are at the lock's ceiling (1).
     • A daemon that holds the lock without pinging lifts nothing; after its ping the ceiling is 201.
     • Under "Speaker Protection" the volumes are at 0. Without an owner a grant is -EPERM and nothing is admitted. The owner's lease gives 201, and the release puts them back at 0.

     apple-speaker-admission now expects J813 to be admitted and J873 not.

   • On the J813, this branch plus drm/asahi, drm/apple: J813 (M5 MacBook Air) GPU and display on the macOS 27.0 firmware #1413:

     • systemctl stop speakersafetyd gives "Speaker volumes locked: Lock timeout". All four Speaker Volumes then read max 1, value 1, and root setting 201 gets 1. Each Amp Gain Volume reads max 15, the level in its apple,speaker-config from aura.
     • With "Speaker Protection" as the interface and no owner, the volumes are at 0. aplay of zeros to the speaker PCM fails with "Operation not permitted", and speaker_starts_refused goes from 0 to 1. The jack still plays.
     • Back to "Speaker Volume Unlock" with the daemon restarted, the volumes are unlocked again.

   • Two table fixes:

     • The J813 and J873 tables dropped amp_gain = 15. It is dead on the SN012776, which takes its level from apple,speaker-config (0x0f on all four J813 amps). Both are now in apple-speaker-amp-level's no-table check.
     • J873's comment and commit message said its speaker could be enabled with please_blow_up_my_speakers, which on integration cannot bypass admission. They now say it stays off because it has no safe level.

KUnit: 84 / 84 in arm64 QEMU at the head, with the kunitconfig from the description. These suites are not in CI. At each test commit it is 81 / 81 and 83 / 83. One run at the ADMAC test commit failed macaudio_spk_lease_expires, an existing test with a 10 ms lease, while the host was building at full load. It passed in four runs after that.

Per-commit build: 16 / 16. apple-dt-check: 0 new warnings, 10 fewer. The speakers still play: a 1 kHz tone at 0.05 reaches the microphones at -50 / -52 / -46 dBFS.

cc @Acelogic @DjDeveloperr @LiamRay10

[DjDeveloperr]

Reviewed before merging. I found no correctness problems: the setup-port EPMAP endpoint discovery, the ADMAC capture realignment (only for a client that passes a frame size), the ti,bclk-ratio path (TAS2764_CLK_CFG is cached and writeable, so a resume's cache sync restores it), the new cs42l84 SDOUT1 routes, the J813 and J873 macaudio tables and the bindings all check out, and CI was green.

I pushed two small fixups as separate commits on top of yours, without changing your commits:

• 8579c33950cb ASoC: apple: t8152-aop-audio: Describe T8142 and the microphones. The Kconfig help and the module description still named only the Mac mini jack and speaker.
• de172b95d676 ASoC: tas2764: Keep the IRQZ_CLR bit next to INT_CLK_CFG. TAS2764_CLK_CFG had landed between that register and its field. No functional change.

I'll merge once CI is green on the new head.

[DjDeveloperr]

Codex (GPT-6.1 Sol), reviewing on @DjDeveloperr’s behalf.

No remaining blocking source findings in the reviewed/integrated net change. The source DMA/work lifetime gaps are now fixed on the contributor branch, with author cosmetics and original ancestry preserved. Microphone STOP/close flushes pending HPAI start before DMA/channel reuse, MIC_PCM_OPS has sync_stop, and the ASoC component mirrors the established T8140 FE/BE sync_stop behavior. Only the same Rust file changes; no6amp/DT/protection behavior was added. The EXACT final tree63b92ef compiled sound/soc/apple/t8152_aop_audio.o successfully through linux-build/build with matching generated Rust bindings, CONFIG_RUST=y and the module=m, with no object diagnostics. SEP rust-build compiles only apple-sep.o and is NOT cited as this audio proof. Author's existing84-test shared ADMAC/TAS/admission evidence and J813 protected-audio reports remain attributed. Fresh Hosted CI and the existing kernel interlock/safe-ceiling checks must still hold before merge; no playback/hardware actions were performed here.

Approval does not waive pending CI, dependency order, hardware-safety gates or source-versus-release qualification. No source merge is an installed-loader write or a new hardware test.

[DjDeveloperr]

I pushed 18fa06f, a merge of current integration/aurora. The earlier base merge into this branch (375393a) had carried its own version of the tools/aurora-sep/tests/test_policy_window.py fix. The agreed fix is #1561, now on integration/aurora, and the two versions merge without a conflict into a test that fails: the early continue keeps #1561's deferred-call count at 0. This merge takes integration/aurora's test file unchanged, so this PR no longer touches tools/aurora-sep. The audio, AOP, ADMAC and codec diffs are byte-for-byte what was reviewed (same patch-ids). The negative-control tests from 375393a can go to #1561's area as a separate PR if they're wanted.

[DjDeveloperr]

Note for any lane that pushes to this branch: please don't push further changes to #1423.

• What happened: 375393a merged the base and carried its own version of the policy-window test fix. That conflicted in effect with soc: apple: sep: policy window test: admit the initrd ref-key unseal #1561, now merged as ea10c4d: the combined result fails the test.
• Current state: 18fa06f re-merges the base and keeps integration/aurora's test file, so J813: AOP sound on integration/aurora (speakers, jack, microphones) #1423 no longer touches tools/aurora-sep. Its audio diff is unchanged from review, and CI is running on 18fa06f.
• Negative-control tests: if you want the extra negative-control tests from 375393a, please open them as their own small PR on top of integration/aurora.
• ASoC: apple: add J815 six-speaker AOP audio #1533: J815's ASoC: apple: add J815 six-speaker AOP audio #1533 can keep stacking on J813: AOP sound on integration/aurora (speakers, jack, microphones) #1423 as it stands now.

Assisted by Claude Code

[DjDeveloperr]

Codex (GPT-6.1 Sol), reviewing on @DjDeveloperr’s behalf.

Current-head compile follow-up for 18fa06f (tracked tree 794d2a6c554bf0d3cbfaea23ed89be14418110a2): both sound/soc/apple/t8152_aop_audio.o and the merged drivers/soc/apple/aop.o compiled successfully as AArch64 objects through linux-build/build, with matching generated Rust bindings and both modules enabled. This closes the specific current-provider/current-consumer compilation gap; the older sound-only build is not being substituted for it.

Clang23.1.1, Rust1.99.0, corrected bindgen0.73.2. Audio object261696bytes/SHA256 d717ef8cde88b38dc21e0e3888ea54152de0cf596d5bba77f8c017cb50425ecd; AOP object807208bytes/SHA256 6717d1e5ab043b094d2a367d2794c7a893a7607bde0f9b2d758fd5b139ba44f9. Tracked source remained clean and unchanged throughout; prior build/rollback evidence retained.

Existing source approval stands. Current host and DT workflows passed; SEP workflow43068470145 is still queued at this read, so merge remains pending its actual result. This is object compilation, not a full kernel/module-link, hardware audio or playback qualification. Existing protected-audio evidence and kernel interlocks keep their original scope. No further branch changes, boot/default-loader changes or hardware operations were performed.