[erikwb]

The J713 (MacBook Air 13" M4, T8132) audio on the AOP: the internal microphones, the headphone jack and the speakers, plus the ASoC fixes the J713 speakers need. It's one ASoC series, and it replaces #437, #439, #663, #580, #759, #760, #765 and #799, plus the audio half of #434. The review points raised on those PRs are answered below.

Depends on #819 (soc: apple: aop: T8132). Its two commits come first here; this PR adds the other 19.

Before this can ship

• Loader: J713: hand the AOP over with its DAPF, firmware segments and windows m1n1#47 (with the window-union fix, 1b9b611ba3ef) and arm: apple: reference m1n1's J713 AOP reservations from the OS tree u-boot#13 are merged on those repos' aurora-wip, but not yet on main. With an older m1n1 the AOP gets no memory-region, and the driver refuses to start it ("No firmware memory-region from the bootloader").

  • m1n1's bundled tree takes the AOP node from this series' DTBs. That path, and the per-kernel DTB through u-boot#13, were both tested in Omarchy's packaged layout (m1n1#47).

• Speakers (the ASoC: apple: J713 (M4 MacBook Air) speakers on the AOP (depends on #437) #439 hold): they must not ship without J713 protection and the crossover. The open pieces:

  • Add the MacBook Air 13" M4 (J713), measured and qualified on the machine speakerguardd#10, the J713 profile, which goes on Thunderbolt: M1/M2 Pro display, M1 PCIe and resume recovery #8;
  • macaudio, aop_audio: J713 (MacBook Air 13" M4) speakers, headphone jack and microphones alsa-ucm-conf-asahi#6;
  • j713: MacBook Air 13" M4 speakers and microphones asahi-audio#10.

  Without a protection daemon, the kernel keeps the speaker volumes at the locked maximum, so the speakers stay silent.

• Shared code: I could only boot a J713. These commits change code that other machines run, and need your regression boots:

  • J616s:
    • aop_audio registering each component where its own channels exist (commit 6). The J616s lists all of them, so it takes the same path as before.
    • accepting an already attached device at probe (18).
    • the volume restore after sleep (17). Jonathan confirmed the bug on a J616s in ASoC: apple: macaudio: Set the speaker volumes again after a system sleep #580 and reviewed the fix, read only.
  • M1 Pro and M2 (TAS2764):
    • the idle slot rewrite after reset (19);
    • the fault IRQ free in remove (20);
    • the SDZ undo on a failed probe (21);
    • the restore after sleep (17).
  • Opt-in only: deferred shutdown and the unmute latch clear (13–16) run only on amps with apple,deferred-shutdown. Other amps only take the new fault_lock in the IRQ handler.

Commits

Microphones (3–5)

• The T8132 AOP audio binding. dma-names allows the microphones alone, the microphones with the jack, and the two speaker layouts.
• The T8132 AOP and audio ADMACs. The channel map is the T6041 one, from the ADT iop-audio-controller: speaker base-ns 0/1, jack 4/5, mics on LEAP.
• The AOP enabled on the J713, with the microphones (3 channels, 48 kHz).

Headphone jack (6–9)

• aop_audio registers the jack where the node lists its own channels, without a speaker stream. Creating the unused services doesn't reach the firmware.
• The J713 jack reuses macaudio_j616_jack_cfg, including its −6 dB headphone DAC cap. This board's analog full scale is unmeasured too.
• DT:
  • the CS42L84 on i2c1 at 0x4b, with reset on nub GPIO 17 and the interrupt on AP GPIO 98, level-low;
  • the jack's ADMAC channels and its own card;
  • ps_i2c1 enabled. Without that the codec never probes (i2c-apple −110).

Speakers (10–12)

• macaudio_j713_cfg:
  • SN012776 amps, 1W1T per side, amp gain 15 and safe volume −20.
  • J616s-style AOP front-ends.
  • A new speaker_fe_channels (6). The AOP's speaker front-end carries six channels, the last two silent, and one channel per amp made the open fail with EINVAL.
• DT:
  • the four amps on i2c3, in provider order 0x38 LW, 0x3b RW, 0x39 LT, 0x3c RT;
  • their sense slots (amp i: imon at RX byte 4i, vmon at 4i+2);
  • the shared SDZ line (AP GPIO 15) as the fail-closed SDZ group's supply, and the shared IRQ (AP GPIO 14);
  • the amp-tx-zd-config idle masks 0xf0f0 / 0x0f0f;
  • the speaker and sense channels;
  • ps_i2c3 enabled.
• The ADT's speaker entry lists 4 playback and 8 sense slots, but the stream is the J616s one: 6 playback and 12 sense, of which the amps use the first 4 and 8.

Clean sense at stream ends (13–16)

• A new apple,deferred-shutdown on ti,tas27xx, set on the four J713 amps. The amp is muted at the stream's end as before, and goes to software shutdown 150 ms later. The AOP kept the bus running for up to 3520 frames (73 ms) after a stream ended.
  • A stream that starts in between cancels the pending shutdown.
  • Suspend and removal shut the amp down at once.
  • A clock-only fault while the amp is stopped is expected, and no longer logged.
• At unmute, a latch that holds only the clock fault picked up while stopped is cleared before the amp counts as running. It's judged under a lock shared with the IRQ handler. Any other latched fault is left for the handler.

Speaker volumes after a system sleep (17)

• Since 1cbce20, the speaker volumes stay clipped at −100 dB after the first sleep. alsactl saves them that way. After a reboot, they stay clipped unless the daemon's cold start sets them after alsactl's restore; Jonathan's J616s boots clean.
• At suspend-prepare, the volumes are kept before the lock clips them. They're set again at the first unlock after the resume, which is the daemon's first ping.
  • A second sleep before that ping keeps what the first saved.
  • The card's controls are held for reading across the keep/clip and in the unlock work, so a newer daemon volume can't be overwritten.
  • A new lock owner drops what was kept.
• The card's remove cancels the lock update work while the snd_card still exists. That work takes controls_rwsem, and a back-end stop during teardown can queue it.

Card rebinds and module reloads (18–21)

• aop_audio accepts a device that is still attached from a previous load (kIOReturnBusy). The speaker and jack services already did. Before, reloading snd-soc-aop left the machine with no audio until a reboot.
• tas2764 writes the idle slot setup again after the probe's reset. Before, a rebind left the idle-filling amps without the force-0 enable (SDOUT_HIZ_9 bit 7). That corrupted the right amps' V-sense, and the protection daemon then held the speakers at the safe level.
• tas2764 frees its fault IRQ in remove, and requests it last in the probe. Before, each rebind added four more handlers on the shared line.
• A failed probe now lowers SDZ and drops the supply enable it took. ASoC doesn't call the remove after a failed probe.

Testing

This series, on integration/aurora 41c4055

On the J713. The test kernel also carried #429, for the J713 display.

• Boot: untainted. All three cards register: microphones, jack, and speakers with their sense PCM. The AOP starts (RTKit: Initializing (protocol version 12)) with no DART faults.

• Speaker protection: speakerguardd takes the lock and unlocks the volumes. Playback through PipeWire's UCM speaker route runs.

• Amps: 4 handlers on the shared fault line; force-0 is 0x80 on both idle-filling amps; the SDZ supply holds one enable per amp.

• Microphones: they record room noise at about −50 dBFS.

• Silent restarts: 0.3 s silent streams with gaps of 40 ms to 1 s before each start. 0 of 64 starts logged an amp fault, and nothing was logged outside a stream.

• Rebind and reload: after a snd-soc-macaudio rebind and then a full snd-soc-aop reload, all three cards come back. Force-0 is still 0x80, there are still 4 IRQ handlers, the microphones record, the speakers play and the daemon governs again.

• AOP start failure and rebind (asked for in soc: apple: aop: T8132 (M4 MacBook Air) AOP and microphones #434): unbinding apple_aop itself times out waiting for the endpoints to acknowledge their shutdown (5 s, mask 0xfff). On the rebind, RTKit endpoint discovery fails (−62). The failure path then behaves as designed:

  • the AOP is halted again;
  • its buffers are kept until reboot;
  • the probe fails cleanly, with no oops and an untainted kernel.

  Audio comes back after a reboot. So the AOP core can't be rebound on J713, but its failure path is safe. Consumer reloads, as above, are fine.

On our J713 tree when each part was first posted

Each part ran on integration/aurora with the earlier parts of this series. The commits are unchanged since, except commit 17's card-remove cancel.

Microphones and jack

• cs42l84 0-004b: Cirrus Logic CS42L84, Revision: 36; the jack card registers.
• The cap holds: Jack DAC Playback Volume tops out at step 244, i.e. −6 dB.
• Playback: a 440 Hz tone at −20 dBFS is audible, the DMA runs in real time, and the DAPM path from Headphone Playback Mux through Jack HP to Headphone is on.
• Stereo: a left-only tone and then a right-only tone came out of the left and then the right earcup.
• Plug detection: Headphone Jack and Headset Mic Jack flip on unplug and replug, with one codec interrupt per event.
• Headset microphone: speech at about −10 dBFS over a −50 dBFS floor, at 0 dB PGA and +10 dB preamp.
• No DART or AOP errors.

Speakers

These ran with speakersafetyd 3.0.0 holding the lease, on the M3 Air's j613 profile with 12 channels. The ADT's per-unit DC resistances (4.00/4.10 Ω woofers, 3.55/3.68 Ω tweeters) match j613's nominals.

• Probe: all four amps bind; the SDZ regulator has 4 users.
• Interlock: without the daemon, the volumes stay at the locked maximum (silent). With it, Speaker volumes unlocked. Stopping it gives Speaker volumes locked: Lock timeout.
• Channel → amp → sense map: a −30 dBFS chord on channels 0–3 (500 / 700 / 1100 / 1300 Hz), with each tone located by Goertzel in the raw sense buffer. Each tone appears in exactly one amp's I/V pair: ch 0 → left woofer, ch 1 → right woofer, ch 2 → left tweeter, ch 3 → right tweeter. Sense channels 8–11 are silent, with no cross-talk.
• Power per speaker: 15.0 / 15.8 / 12.8 / 13.9 mW.
• By ear: woofers low and tweeters high, on the right sides.
• Timing: with 6/12 channels the stream runs in real time; a 4/8 wiring ran 1.5× fast.

Deferred shutdown and the stale clock fault (13–16)

• Sense tails, from three silent streams and a chord with the volumes locked: about 90 full-scale words per run before, all on the right tweeter's V slot at each stop. None after. Every amp still ends in software shutdown (PWR_CTRL 0x2).

• Restarts: pairs of short streams 90, 120 and 150 ms apart. All four amps play, the sense stays valid, and there are no pops.

• Faults at restarts:

  	Starts that logged a fault
  Without commit 16	9 of 64, all on the left woofer: 9 of 32 with gaps up to 140 ms, 0 of 32 with gaps of 170 ms or more
  With commit 16	0 of 96 (and 0 of 96 with tracing, where 11 stale latches were cleared at unmute)

• Under speakerguardd: a recording of 13 streams, including those restarts, has no full-scale sense words, and its replay certifies.

Volumes after sleep (17)

	Before	After
Speaker volumes after a suspend-to-idle	all four at 1 (−100 dB), for good	1 until the daemon pings, then all four back at 201 (0 dB)
Two suspends, the second with no ping before playback	—	still 1 after the second resume; back at 201 when playback starts
Playback	PCM runs, nothing audible	audible
speakersafetyd restart needed	yes	no

Jonathan reproduced the bug on a J616s in #580: six volumes at 1 after one s2idle cycle, with the daemon's PID unchanged.

Reloads and rebinds (18–21)

• Reload (18): before, rmmod/insmod snd_soc_aop failed the probe on the pdm0 attach. After: five reload cycles, and each time the three cards are back, the speakers and jack play, the mics record at the same level, and the daemon governs.

• Idle slots (19): SDOUT_HIZ_9 on the idle-filling amps.

  	Fresh module	After a macaudio rebind	After an snd-soc-aop reload
  Without	0x80	0x00	0x00
  With	0x80	0x80	0x80

  Without it, after a reload, the right woofer's and right tweeter's V channels each had about 11,700 of 48,000 words with the top bit set. With it, none.

• IRQ (20): 12 handlers on the fault line after two rebinds without the fix; 4 with it, fresh, after two rebinds and after a reload. The IRQ still fires and is handled.

• SDZ (21): I made the first setup write fail, then bound the card twice and then cleanly. Enables on sn012776-sdz: without the fix, 1-0038 holds 2 after the failed binds and 3 after the good one. With it, none and then 1 per amp.

Checks

• checkpatch --strict: no errors. The one warning and one check are the MAINTAINERS note, which the existing sound/soc/apple/* entry covers, and the kunit_test_suite() blank line, which follows the admission test header.
• Universal WERROR build (defconfig + asahi.config): all 133 fragment settings survive. Every file this series touches builds with WERROR: aop.o, aop_audio.o, macaudio.o, tas2764.o and the J713 DTB. The tree as a whole stops in drivers/nvme/host/apple.c, mt7932_bt_pcie.c and the asahi Rust core, which are untouched toolchain cases on clang 23 / Rust 1.99 (CI pins clang 18 / Rust 1.96). Not run: the Fedora-base recipe.
• KUnit, arm64 QEMU with KASAN and lockdep: apple-speaker-pm-volumes (3) and apple-speaker-admission (1) pass, with no reports.
• run-host-tests: 132/133, the same as the base. apple-dcp fails on bare integration/aurora as well.
• dt_binding_check is clean for apple,t6041-aop-audio.yaml, apple,macaudio.yaml and ti,tas27xx.yaml.
• CHECK_DTBS on t8132-j713.dtb gives the same diagnostics as the base.

Not tested: other machines (see Shared code above). Headphone impedance detection, which this driver doesn't do, as on the J616s.

[erykwieliczko]

Rebased onto the current integration/aurora, with three fixes (not pushed yet)

#820 no longer merges: on integration/aurora 21f7ffc, 7915d686 (the J614s cards) adds macaudio_j614_cfg at the same place as macaudio_j713_cfg. Rebasing also exposes two changes in the base that git merges without a conflict, and one bug in the series itself.

I have the rebased series locally (j713/audio-integ-next, on the rebased #819). Erik stays the author of every commit. The four I changed carry an [eryk: …] note and my Signed-off-by. I haven't pushed it: the PR allows maintainer edits, so if you're fine with it I'd force-push it to j713/audio-integ, and #819's to j713/aop-t8132.

Changes

ASoC: apple: macaudio: Add the J713 speakers

• Conflict: kept the base's macaudio_j614_cfg and put macaudio_j713_cfg after it. The jack comment keeps the base's "J614s and J616s".
• Route-only front-end: f0fb0a51b1e5 ("no PCM for the J616s route-only front-ends") leaves the J616s speaker card's first front-end unbound. The J713 card uses the same front-ends and has no headphone codec either, so it gets the same flag:

   	.fe_cpu_is_platform = true,
  +	/* No headphone codec on this card: the jack is a card of its own. */
  +	.route_only_fes = BIT(0),
   	.speaker_fe_channels = 6,

  The AOP component numbers front-end PCMs by link (use_dai_pcm_id). So the speakers stay hw:AppleJ713,1 and their sense hw:AppleJ713,2, which UCM and speakerguardd open; only the dead hw:AppleJ713,0 goes away.

ASoC: tas2764: Power the amp down again when the component probe fails

• The base's 33ab1c4a made the SDZ supply optional (devm_regulator_get_optional()), so sdz_reg is NULL on every amp whose SDZ is a GPIO (the M1 and M2 machines).
• The new error path called regulator_disable(tas2764->sdz_reg) unconditionally: a NULL dereference on exactly the failure this commit handles. git merges it without a conflict.

  -	if (!tas2764_group_failed(tas2764))
  +	/* As in the remove: no supply to drop with a GPIO SDZ, none left in a failed group */
  +	if (tas2764->sdz_reg && !tas2764_group_failed(tas2764))
   		regulator_disable(tas2764->sdz_reg);

ASoC: tas2764: Free the fault IRQ when the component is removed

• The IRQ was a devres of the I2C device, requested at component probe, after devm_snd_soc_register_component().
• When the I2C device is unbound while the card is bound (through sysfs, or with the adapter going away), devres releases in reverse order: the IRQ first, then the component. The component's remove then calls devm_free_irq() for a devres that no longer exists, which hits the WARN_ON in devm_free_irq() and taints the kernel.
• Card rebinds and snd-soc-aop reloads don't hit it.
• The component owns the IRQ now, so it uses a plain request and free:

  -		ret = devm_request_threaded_irq(tas2764->dev, tas2764->irq, NULL, tas2764_irq,
  +		ret = request_threaded_irq(tas2764->irq, NULL, tas2764_irq,
   ...
  -		devm_free_irq(tas2764->dev, tas2764->irq, tas2764);
  +		free_irq(tas2764->irq, tas2764);

The other 17 commits are unchanged apart from the rebase; one has a new DT context line.

Testing

• Build: the rebased soc: apple: aop: Add the T8132 (M4 MacBook Air) AOP #819 and ASoC: apple: J713 (M4 MacBook Air) microphones, headphone jack and speakers #820 alone on 21f7ffc, with W=1.
  • aop.o, the AOP audio driver, macaudio.o, tas2764.o, cs42l84.o and t8132-j713.dtb build cleanly. The only warning from this series is the DTC simple_bus_reg note on aop@390600000, which the T6040 AOP node triggers too.
  • The three changed commits also build at their own position in the series.
• On a J713: 21f7ffc with this series plus arm64: keep J713 GPU/DCP non-coherent when CLIDR_EL1.LoC == 0 (fixes J713 graphics on 7.3) #429, cpuidle: apple: add T8132 and enable it on J713 (s2idle groundwork) #432, J713: G16 performance governor, and the PMP idle floor it needs #455, J713: Touch ID (SEP profile for the V3 sensor, SoC and board device tree) #457, J713: enable the camera (T8132 ISP) #531, J713: hold the memory and fabric floor for GPU work, not whenever GFX is powered #582, drm/apple: m4: Power the display down for system sleep, while blanked and while static #830, usb: typec: tipd: Take a partner identity that arrives late on CD321x #883 and our AOP motion patches. The suspend-to-idle and USB4 series weren't included.
  • Boots untainted, with no warnings. The AOP gets its 17 regions from m1n1 and u-boot, and starts.
  • All three cards come up. The speaker card now has only pcm1p and pcm2c. Four tas2764 handlers share the fault line.
  • The microphones record on all three channels.
  • The speakers play through PipeWire's UCM route with speakerguardd governing: "Speaker volumes unlocked", and no amp faults.
  • The CS42L84 and the jack card come up. I didn't check headphone playback in this run.
• Graphics on this base: the GPU's DRM driver is now named aurora (a97f7b8be0eb), so graphics needs aurora-silicon/mesa 6edf62bcfd02 or later. I used aurora-wip 05ddb00f5ca9. An older Mesa gives a black screen.

Separately (not in this update)

On the previous stack, which had the suspend-to-idle series, one system suspend logged tas2764_irq_locked: 15208 callbacks suppressed and interrupt latch N unreadable. The shared fault line keeps firing while I2C is already suspended, and every handler returns IRQ_NONE. Disabling the IRQ across system sleep in the codec's suspend and resume would probably fix it. That's worth a follow-up once #831 is rebased.

---

🤖 Written by Claude Opus 5.5 (in Claude Code) for Eryk; the rebase, fixes and tests above were done the same way.

[erykwieliczko]

Merged. This comment is from Claude, working for Eryk, not from Eryk himself.

Rebased onto integration/aurora 565c8025e715 and force-pushed first, as Erik agreed. Each changed commit carries an [eryk: …] note. Erik stays the author of his commits.

Changes

• ASoC: apple: macaudio: Add the J713 speakers: resolved the conflict with the new J614s config, and marked the first front-end route-only (.route_only_fes = BIT(0)), as on the J616s, because the J713 jack is a card of its own.
• ASoC: tas2764: Free the fault IRQ when the component is removed: the IRQ is now requested without devres. On an I2C unbind, the devres release freed it before the component was removed, and the devm_free_irq() in the remove then warned.
• ASoC: tas2764: Power the amp down again when the component probe fails: regulator_disable() is skipped when the amp has no SDZ supply.
• New commit, not in my earlier comment: ASoC: tas2764: Keep the fault IRQ masked across system sleep. From the component suspend until its resume, the handler can't read or clear the latches. A fault latched in that window holds the shared level-low line asserted, and one J713 suspend logged 15208 suppressed callbacks. The IRQ now stays disabled across that window.

Testing

• Eryk confirmed audio by ear on a kernel with integration/aurora 565c8025e715, these commits, other J713 PRs and some local patches. Everything under sound/ in that kernel is byte-identical to what was merged.
• The merged tip on its own has not been built or booted.
• The new tas2764 commit has not been through a real suspend on hardware yet.

---

🤖 Written by Claude Opus 5.5 (in Claude Code) for Eryk.