[niteria]

Deep sleep (suspend-to-RAM) for the M4 MacBook Air (T8132, J713) at EL2, on dj's generic sleep driver, with Erik Bourget's radio and docked-sleep series.

Base: integration/aurora (bf00add09e22; it merges cleanly with today's e661192638a3, whose two new commits touch none of these files), which has dj's T8140 suspend-to-RAM stack since #1445 merged. #1445 superseded #1334 and has the same generic apple-sleep driver. T8132 is that driver's second SoC, t8132_sleep_soc, with the fields it needs added to struct apple_sleep_soc.

What's in it, 52 commits:

• The T8132 port (from iommu: apple-dart: don't require DAPF ranges on a DART that has none to arm64: dts: apple: t8132-j713: name the SEP's domain in the PMGR policy), including the system level cache write-back, the SMC's T8132 numbering of its wake reasons, and the PMP thermal samplers paused across a deep sleep.
• Erik Bourget's native radio series, arm64: dts: apple: Restore DCP address cells for J713 IOMMU validation #1385, [1/5] iommu: apple-dart: Preserve J713 radio translations across S2R #1378 and [2/5] soc: apple: Resolve radio PERST through the sleep controller PCIe phandle #1381 to [5/5] Bluetooth: bcm4377: Restore private PCI windows before J713 resume #1384: Wi-Fi and Bluetooth through a deep sleep. Erik has closed those PRs. Ahead of his brcmfmac commit, wifi: brcmfmac: don't cancel a bus reset that was never set up.
• Erik Bourget's docked series, [1/7] pmdomain: apple: Let a consumer switch the J713 external display domains #1414 to [7/7] drm/apple: m4: Stop the J713 external DCPs for a deep sleep and restart them #1420, as he force-pushed it after dj's review: Thunderbolt host router, CIO PHY, tunneled PCIe and the external DCPs through a deep sleep, behind pcie_apple.cio_s2r and thunderbolt_apple.cio_s2r (both off by default). That's 18 commits, among them the four the earlier copy here lacked: rolling back a partial consumer power-off in pmgr-pwrstate, the apple,clock-protection binding and its J713 DT, and apple,policy-machine for the platform-handoff entries. His kernel code is as on his branches in all 18, and 10 are identical outright. The rest differ in context, three of them (the DART, ATC CIO and Thunderbolt apple.c ones) after conflict resolution that keeps both sides, and in the host-test changes the notes below list.
• Two binding commits for properties the radio and docked series' device trees use: a root port's clkreq-gpios and the USB4 router's cio PHY.
• tools/asahi: run the J713 sleep tests from the host-test runner: registers the 17 host tests the two series added, which run-host-tests.py requires, and makes them run under its python3 -I.

Erik's commits keep his authorship and sign-off. Twelve of them have a bracketed note after it: his Based-on-work-by: tags, which aren't kernel tags, moved into the text (five), checkpatch fixes in his host-test harnesses (four), older host tests' mocks updated for his changes (six, below), and the review fixes in item 2 (three). If this PR is where they land, #1414 to #1420 can be closed too.

Eight older host suites copy functions this series changes into mocks, and stop compiling against it. Each mock gets the new state, inert as on other machines, in the commit that first needs it:

• apple-dart-fw and apple-dart-range (apple-dart.c): the radio sleep record, the sleep writer gate and, with the dock series, cio_active and the gate's new return value;
• brcmf-suspend and bcm4377-resume, with their mutation runs: the radio sleep hooks;
• t6030-clock and apple-dcp (atc.c): the CIO sleep state;
• rtkit-oslog, with its mutation runs: it now also copies apple_rtkit_reset_session(), which the cold-restart commit split out of apple_rtkit_reinit();
• asahi-m4: G16's publish() checks the sleep flag.

asahi-m4 also gets a case for a publish during a sleep, and t6030-clock two for the stale ACK a CIO sleep leaves. Erik's own brcmfmac-sleep-lifetime test copies brcmf_pcie_pm_enter_D3(), which now has #1653's WoWLAN checks; its mocks get that state too, never armed.

Two fixes the J713 needed in integration/aurora itself are merged now: #1669 (the G16 probe overflowed its 32 KB stack) and #1668 (the scaler and JPEG DARTs had no power domain, so the DART probe panicked at boot).

Review items

1. Signed-off-by: on every commit; Erik's carry his and then mine as the submitter.

2. Changes from a review of 19de60a3ee67, first pushed as c53fb99165a4. Each fix is in the commit that had the problem, or in a new commit where marked:

   • arm64: dts: apple: t8132: add the sleep controller: the node is disabled in t8132.dtsi and enabled in t8132-j713.dts. It bound on every T8132 board and made deep sleep the default there, but the radio, DART and DCP recovery is the J713's alone, so every deep sleep was refused and s2idle was no longer the default.
   • iommu: apple-dart: accept a display DART the sleep engine rebuilt: its mark that the sleep engine rebuilt a display DART is only set at resume now, and only the DCP's next power-up clears it. With the panel off across a deep sleep and then an s2idle or a sleep refused late, the mark was cleared while the DART still had the engine's streams on, and the next power-up failed: DCP SID23 retired: power-up: state lost (-5), internal panel dead until a reboot. Reproduced on the J713 before the fix, gone after it.
   • New, thermal: apple-pmp: pause sampling across a deep sleep: sampling ran on through the stopped PMP. After every wake, "sensor error -19; CPU minimum caps retained" (about 100 lines per test run), the CPUs held at their minimum frequency for about a second, and the GPU's power level stepped back from 7 over about 15 s. Now no sensor or telemetry error after any wake.
   • drm/asahi: G16: ride out dark sensors while the GPU is idle: an idle sample keeps the last cool reading current, so the 2 s allowance runs from when work starts rather than from the last valid reading. After a deep sleep the GPU takes no idle sample before its first work, so that case remains (see Known limitations).
   • New, mfd: macsmc: number T8132's wake reasons as its SMC does: the wake reasons were named by T8140's numbering, where the lid is bit 10; the J713's SMC uses bit 8. A lid closed during a deep sleep woke the J713 fully, user space included. Now: "woken for the lid, which is closed", a dark wake with the outputs off, and straight back to sleep; opening the lid wakes it.
   • dt-bindings: pci: apple,pcie: give T8132 its own register list: the T8132 branch rejected the third-port layout of the J604, J623, J624 and J773g (12 new CHECK_DTBS warnings). It accepts it now, and those boards' older reg-names warnings are gone too.
   • soc: apple: mcc: write T8132's cache back for a deep sleep: the cache is also written back when probe left it alone and when the driver is unbound, and an enable that isn't confirmed turns it off again: nothing else writes it back before a sleep.
   • Erik's iommu: apple-dart: preserve native radio translations across system sleep: sleep-spin.h reads CNTFRQ_EL0 with arch_timer_get_cntfrq(), as arch_timer_get_rate() isn't exported: with CONFIG_APPLE_DART=m (arm64 defconfig) modpost failed. The spin also reads the register once more at its deadline before it reports a timeout. The radio port commit's host-test mock follows.
   • Erik's thunderbolt: keep PCIe and USB tunnels across a host router power cycle: tb_switch_suspend_cio() took tb_switch_clx_disable()'s positive return, the CL states it disabled, for an error.
   • New, wifi: brcmfmac: don't cancel a bus reset that was never set up: a suspend or a remove before the firmware had loaded ran cancel_work_sync() on a work with no function, and the workqueue code warned.

3. Rebased from f173f7c8537a onto 028d76246b00, then onto bf00add09e22. Three conflicts, all in Erik's commits, and one host test needed changes:

   • wifi: brcmfmac: pcie: keep Apple firmware in D0 across s2idle #1722 (Apple firmware kept in D0 across s2idle) and wifi: brcmfmac: coordinate retained firmware with J713 system sleep: only brcmf-suspend-tests' mock struct conflicted; it has both sides' fields. In pcie.c the new s2idle-D0 branch runs only for s2idle, so the J713's deep-sleep path (the lease, the D3 inform, apple_radio_sleep_endpoint_suspended()) is as before. brcmfmac-sleep-lifetime-tests copies the same function, and its mocks get the s2idle-D0 state, never taken in its deep-sleep cases.

   • PCI: apple: t8140: bring the radios back after a sleep off resume_noirq #1717 (T8140 radio resume off resume_noirq) and PCI: apple: stop and restart the J713 tunneled port for a deep sleep: both added an apple_pcie_config_read(). It's one function now: the Neo's D0 answer while its radio link comes back, then the J713 tunneled port's cio_cfg_lock. PCI: apple: t8140: bring the radios back after a sleep off resume_noirq #1717's T8140 resume is unchanged, and async resume stays the Neo host's: the J713 radio host still disables it.

   • T8132: DisplayPort alt mode on the J604's USB-C ports #1678 (DisplayPort alt mode on T8132) and phy: apple: atc: keep a T8132 CIO link when its USB host stops and phy: apple: atc: take the T8132 CIO PHY through a deep sleep: T8132: DisplayPort alt mode on the J604's USB-C ports #1678's t8132_dp_alt and the CIO sleep state are neighbouring fields, in atc.c and in the host tests' copies of its struct (atc-usb3-tests, and the DCP selftests' phy.py and dwc3/mocks.c). Each keeps both. They concern different modes: the CIO sleep is only prepared in a USB4 or Thunderbolt mode, and only the J604 marks a port for DP alt mode, and deep sleep stays off there.

   Nothing else the base gained (106 commits, 168 with merges) conflicted.

   Earlier: rebased from Apple SoC sleep: one driver for all Apple SoCs, matched only on tested ones #1334 onto integration/aurora after Apple SoC suspend-to-RAM on T8140 #1445 merged, with Erik's docked series replaced by his current one. Of the 31 other older patches, git range-diff against this PR's previous head shows 20 unchanged and three changed only in context. Five conflicted, and each keeps both sides:

   • soc: apple: sleep: add T8132 ...: Apple SoC suspend-to-RAM on T8140 #1445 now panics when a commit returns, and that check follows the per-SoC commit hook; T8132's commit doesn't return either. Apple SoC suspend-to-RAM on T8140 #1445's report of coprocessor stop errors sits next to T8132's report of power words still on, and the latched power words dj's later change records sit next to T8132's quiesce word. Ryan's AVE complex change (soc: apple: sleep: power the AVE complex as the encoders' parent domain #1663) skips .prepare's complex power-off once genpd owns the complex; T8132 has no complex, so that power-off also stays T8140's (has_ave), as before. A comment left over from removed bring-up code is gone.
   • iommu: apple-dart: accept a display DART the sleep engine rebuilt: a neighbouring bitfield.
   • Erik's iommu: apple-dart: preserve native radio translations ...: the DART mock's struct also gets Apple SoC suspend-to-RAM on T8140 #1445's fw_tree fields, and apple_dart_suspend() returns early for the live J613 MTP first, then for a radio DART.
   • Erik's soc: apple: resolve radio PERST through the sleep controller's PCIe host: it drops radio_pcie from the per-SoC data, next to soc: apple: sleep: power the AVE complex as the encoders' parent domain #1663's new AVE fields.
   • Erik's wifi: brcmfmac: coordinate retained firmware with J713 system sleep, against Ryan's WoWLAN in s2idle (wifi: brcmfmac: Wake-on-WLAN in s2idle for the BCM4387/BCM4388 Apple laptops #1653). The suspend refuses WoWLAN without wakeup permission or outside s2idle before the J713 sleep lease is taken, so a deep sleep never reaches the WoWLAN path; the resume checks the lease first, then takes WoWLAN's hot path. brcmf-suspend's mocks have both sides' fields, and it passes with all its mutations.

   Three changed on purpose:

   • phy: apple: atc: bring the T8132 PHY back after a deep sleep no longer sets the PIPE handler's dummy PHY enable itself. Friedrich's phy: apple: atc: enable the dummy PHY on T8132 SuperSpeed ports (phy: apple: atc: enable the dummy PHY on T8132 SuperSpeed ports #1492) now sets it at the same point for every T8132 port with its SuperSpeed banks, which both J713 ports have, and after a deep sleep this commit has the PIPE handler set up again.
   • arm64: dts: apple: t8132: describe what a deep sleep needs: its message said the driver falls back to s2idle without the APCIe tables. Since Apple SoC suspend-to-RAM on T8140 #1445 it refuses a deep sleep that stops ANS instead.
   • The registration commit registers the docked series' current host tests.

   The rest of the base's sleep changes need nothing for T8132. The SEP's SLEEP request goes out at the platform's last returnable step. The USB host's DRAM vote stays T8140's alone (apple,sleep-ddr). The latched power-word clearing and the T8140 domain changes take per-SoC tables T8132 doesn't fill. The refusal over an unbound running coprocessor and NVMe's shutdown notification at suspend apply to T8132 as well, and were in the tests below. So were the scaler and both JPEG engines that integration/aurora now enables on the J713: bound and runtime-suspended through every sleep. Its video decoder didn't bind in the tests (no apple/avd-fw-v5-t1.bin on this Air), so no sleep has run with it bound.

4. KUnit, built with GCC and run in QEMU: apple-sleep-soc-data (T8132 in its tested list) and apple-mcc pass, 7 of 7, and sep.kunitconfig's 287 apple_sep* tests pass with CONFIG_GPIOLIB added, which sensor_shim.c needs under GCC, on the base too. The SMC's macsmc-parser suite passes 15 of 16 with the new T8132 wake-reason cases; the one failure, macsmc_sleep_order_test, fails the same way on integration/aurora.

5. Bring-up removed: the DO NOT MERGE commit, the progress marks in the sleep driver, PM core and arm64 resume path, the panic and die notifiers, the comparison with macOS's power words, every bringup_* parameter (their defaults kept), and cpu-reset.o without KEXEC: the board's kernel config now has CONFIG_KEXEC=y. The T8132 commit no longer adds the PMU commands macOS sends after the commit; sleep and wake work without them.

6. t8132-j713-apcie-tunables.dtsi is gone: the bootloader passes ANS's APCIe tables. Aura 1.0.0b11 publishes them from the ADT on every chip: Publish APCIe ANS recovery tables with atomic packed-DT export aura#112's final revision opens a copy of the tree with the buffer's full capacity, so a tree prepare_dt() has already packed has room for them too. On the J713, 1.0.0b11 logs FDT: /arm-io/apcie tables published, 261 records, and the ten tables are byte-identical to the copy this PR used to carry. With an Aura that predates it (1.0.0b10 as tagged, say), nothing breaks, but every deep sleep is refused at the driver's last checks ("no tunables to bring ANS's PCIe block back") and resumes.

7. checkpatch: every commit is clean apart from two prototypes in .c files (the resume stub's, next to its asm(), and one the Rust side of G16 calls, like its neighbours) and the MAINTAINERS reminders for new files.

8. One copy: see above.

Known limitations

• USB devices disconnect at every deep sleep. For a deep sleep on a port without a fixed hub, dwc3-apple stops the USB host before the freezer runs and brings it back after the wake, as the base does; a mounted USB disk loses its block device, and no USB device can wake the Air.
• A deep sleep is refused unless brcmfmac and hci_bcm4377 are both bound: Erik's radio sleep needs both drivers to take the radio through its power loss.
• Docked with cio_s2r off, the default: a USB4 or Thunderbolt device in a port blocks every suspend, s2idle too. DWC3 stops its host for the default deep sleep, Erik's phy: apple: atc: keep a T8132 CIO link when its USB host stops keeps the PHY in a CIO mode, and atcphy_suspend() refuses. On integration/aurora that check still let a docked s2idle through.
• After some wakes the GPU runs at reduced clocks for up to about 16 s: 4 of 35 wakes on the pre-rebase test boot, 1 of 10 on af8707ff5d61's, none of 10 on this head's. Twice the GPU's power budget in apple-pmp-thermal started at its lowest level and released one step a second, and twice G16's own sensor backstop found the die sensors dark at its first sample after the wake. Neither came back in 18 sleeps looked at closely (the budget polled every 50 ms), so their causes aren't settled yet.
• Speakers with speakerguardd 0.1 (the Speaker Volume Unlock daemon the J713 runs today) come back almost silent after a deep sleep and take about 90 s to recover. At suspend, macaudio clips the speaker volumes ("Speaker volumes locked: System suspend"). A volume write from the daemon that lands in that window fails with EINVAL, and the daemon gives up ("leaving the speakers at the card's safe level"). After the wake it takes the lock back at about 97 dB of gain reduction and eases off from there. The kernel side behaves as intended: the pre-sleep volumes are restored at the unlock, and the amps keep their state through the sleep (their temperature registers survive it). speakerguardd v2 handles system sleep (packaging: orderly sleep, graph before logins, supervisor after sessions speakerguardd#84, Moved · 🧩 T6034 · M3 Max, 14-core — board index #85, PCI: apple: enable experimental J700 radio endpoint bootstrap #151) but doesn't support the J713 yet.

Tested

On the J713 (24 GB, Linux at EL2) as a full system build with CONFIG_KEXEC=y, radios, GPU driver and cache on, Aura 1.0.0b11 as Stage 2 (a one-boot aura-try trial from the installed b9 Stage 1, with this branch's DTB), with this head's kernel, all on one boot:

• 8 plain systemctl suspend sleeps from a lit desktop (40 s, and 3 min), each woken by its RTC alarm. In 7 of them a process rewrote and checked 256 MB up to the freeze: no bad blocks. After each, the radio recovered from the power loss, Wi-Fi carried 20 and 300 1400-byte pings to the gateway without loss, Bluetooth answered Read Local Version, and the GPU firmware was back and rendered. Then 2 sleeps refused at the last returnable step with apple_sep.sleep_refuse_before_send: each refused before SEP SLEEP was sent (-ECANCELED) and rolled back, and systemd's immediate second attempt then slept and woke normally.
• The panel off across a deep sleep, then a sleep refused late, then the panel on: it came back. On 19de60a3ee67 the same sequence left it dead until a reboot.
• Over the boot: the radio and the GPU back after every wake, no oops, panic, SError or WARNING, and no thermal sensor or power telemetry error after any wake (on 19de60a3ee67, about 100 "sensor error -19" lines in a run of 13 sleeps).
• The AOP's sensor services that the base now describes on T8132 (light, lid angle, motion and board temperature) were bound, and their drivers started again after every wake.

Mesa was built against this base's GPU UAPI, which has GEM userptr. The same run passed on the previous head, af8707ff5d61 (these patches on 028d76246b00).

Before these rebases, on c53fb99165a4 (the same patches on f173f7c8537a), one boot also had:

• The lid closed during a deep sleep: "woken for the lid, which is closed", a dark wake with the outputs off, and straight back to sleep; opening the lid woke the Air. On 19de60a3ee67 the same close woke it fully.
• 22 more sleeps, plain and with the 256 MB writer, with the thermal driver's power budget read every 50 ms.
• In all, 37 sleep attempts, 35 wakes counting the dark wake, the radio and the GPU back after every one, and no fault.

The speaker amps keep their state through a deep sleep: their temperature registers, which a power-on reset clears, still held their last readings after it, and sound played about 4 s after the wake.

The docked series' DT changes the apple,bridge-link format, so the bootloader has to pass this branch's DTB: with the previous one, two pmgr-pwrstate domains fail to probe ("bad bridge link").

Before the rebase, on #1334: the same 8 sleeps and 2 refusals (on a boot without apple_sep) with the tables in the device tree, and 8 more sleeps with them from the bootloader.

Also, at this head: it builds (Image, modules, DTBs) with no warning beyond the base's, the whole run-host-tests.py passes (310 of 310 runs; the base passes its 293), KUnit is as in item 4, and CHECK_DTBS on all six T8132 boards (J604, J623, J624, J713, J715, J773g) shows no new warnings and ten fewer than the base: the T8132 PCIe host's reg-names now match its binding (with ti,cd2e22.yaml, via,vl122.yaml and apple,pmp.yaml set aside: their uint8-array apple,tunable crashes dtschema's decoder, on the base too). Before these rebases, on c53fb99165a4, every commit built and passed the whole host suite, and the head also built with the DART, PCIe, NVMe, SMC, DWC3, MCC and ATC PHY drivers as modules (19de60a3ee67 failed modpost there: arch_timer_get_rate).

Not tested here

• The docked paths: this Air has a USB-C adapter, not a Thunderbolt dock, and this boot chain doesn't hand over the external DCPs (external DCP not handed over; Aura does that only with chosen.asahi,j713-dcpext=1). Since Erik's own dock test, the base gained Ryan's thunderbolt: apple: Restart the port when the peer link is lost (e847be1). It cycles a port whose peer stays gone for 3 s, from a freezable workqueue. The docked sleep path hasn't run with it.

🤖 Generated with Claude Code

[RyanTheTide]

What this branch needs before it can be merged, on head 715a022:

1. Signed-off-by. None of the 19 commits from 297d7a47d7c9 (iommu: apple-dart: don't require DAPF ranges on a DART that has none) to 715a0229e1b8 (soc: apple: mcc: write T8132's cache back for a deep sleep), excluding the six signed radio commits 6a720bcba1b3..b37219c61cbd, carries a Signed-off-by: line. Please add yours to each (git rebase -x 'git commit --amend --no-edit -s' feat/t8140-s2r) and push.

2. Rebase onto the current feat/t8140-s2r (31fd6d6). The branch forked at 5b1a671; the base has since merged #1311 and the merge now conflicts in drivers/pci/controller/pcie-apple.c (with b724b53 "PCI: apple: attach and retain all host power domains in s2idle" and f31c0a3 "PCI: apple: preserve probe without power-management support").

3. Build break. bringup_va_mark in arch/arm64/kernel/sleep.S (added by b6f4482) loads t8132_spmi_va, which is defined in drivers/soc/apple/t8140-sleep.c and only exists when CONFIG_APPLE_T8140_SLEEP=y. Any config without that driver built in fails to link (ld.lld: error: undefined symbol: t8132_spmi_va, see the kunit job of https://github.com/aurora-silicon/linux/actions/runs/42695306042, which builds this branch's code). This goes away with item 4.

4. Remove the bring-up parts listed in the description: 669e8db10c18 (DO NOT MERGE workarounds), the progress marks and panic/die notifiers in b6f448244d44, the bringup_* module parameters, and cpu-reset.o built without KEXEC (enable CONFIG_KEXEC instead).

5. arch/arm64/boot/dts/apple/t8132-j713-apcie-tunables.dtsi is 295 lines of calibration copied from the ADT. Please have the bootloader pass these on instead of carrying the table in the tree, as the description already plans.

6. checkpatch on every commit.

7. The radio commits 6a720bcba1b3..b37219c61cbd are also open as #1385, #1378, #1381, #1382, #1383 and #1384. Keep them in one place so the two copies cannot diverge.

This PR is based on #1208, which has to land first.

[erikwb]

Please merge PR 1378, 13{81-84}, 1414-20 into this branch before diverging it too much - should work smoothly on your machine, tested on what I believe is an identical machine!

[niteria]

Thanks for the list. Pushed an updated branch; where each item stands:

1. Signed-off-by: on every commit; Erik's carry his and then mine as the submitter.
2. Rebased, now onto Apple SoC sleep: one driver for all Apple SoCs, matched only on tested ones #1334 (feat/apple-soc-sleep-generic) rather than Deep sleep (suspend-to-RAM) for the MacBook Neo (T8140) at EL2 #1208 alone: T8132 is a struct apple_sleep_soc entry in Apple SoC sleep: one driver for all Apple SoCs, matched only on tested ones #1334's generic driver, so on Deep sleep (suspend-to-RAM) for the MacBook Neo (T8140) at EL2 #1208 it would need redoing once Apple SoC sleep: one driver for all Apple SoCs, matched only on tested ones #1334 lands. This PR's base is changed to that branch. Nothing here needs t8140-sleep: take the deep path when an SLP_DDR wake would stall (camera ran, AOP crashed) #1358 or later.
3. The t8132_spmi_va link failure went with the instrumentation. sep.kunitconfig builds on this branch and its 210 apple_sep* tests pass; apple-sleep-soc-data (now listing T8132 as tested) and apple-mcc pass too. Built with GCC here, which on Apple SoC sleep: one driver for all Apple SoCs, matched only on tested ones #1334 too needs CONFIG_GPIOLIB for sensor_shim.c.
4. Bring-up removed: the DO NOT MERGE commit, the progress marks (sleep driver, PM core, arm64 resume path), the panic and die notifiers, the power-word comparison, and every bringup_* parameter, with their defaults kept as code. cpu-reset.o is back under KEXEC_CORE; the J713 board config now sets CONFIG_KEXEC=y.
5. The APCIe tables now come from the bootloader and t8132-j713-apcie-tunables.dtsi is gone. aura#112 already publishes them, but off T8140 prepare_dt() has packed the tree before the publisher runs, so it found no room ("APCIe tables not published"). The fix reopens the tree around it, as dt_disable_j713_pcie() does (kboot: publish the APCIe tables into a reopened tree off T8140 aura#145, stacked on Moved · J414C · MacBook Pro 14-inch (M2 Max / T6021) #112). With it the J713 gets all ten tables, byte-identical to the copy this PR carried, and sleeps on them.
6. checkpatch: every commit is clean apart from two prototypes in .c files (the resume stub's, next to its asm() block, and one G16's Rust side calls, like its neighbours) and the MAINTAINERS reminders. Erik's dock commits needed fixes in their host-test harnesses and their Based-on-work-by: tags; both are in his commits with a note.
7. Erik's radio (arm64: dts: apple: Restore DCP address cells for J713 IOMMU validation #1385, [1/5] iommu: apple-dart: Preserve J713 radio translations across S2R #1378, [2/5] soc: apple: Resolve radio PERST through the sleep controller PCIe phandle #1381-[5/5] Bluetooth: bcm4377: Restore private PCI windows before J713 resume #1384) and docked ([1/7] pmdomain: apple: Let a consumer switch the J713 external display domains #1414-[7/7] drm/apple: m4: Stop the J713 external DCPs for a deep sleep and restart them #1420) series are in this branch with his authorship kept: conflict resolution, and the item 6 fixes. Two new binding commits document what his device trees use (a root port's clkreq-gpios, the routers' cio PHY). Erik, if this is to be the one place, those PRs can be closed.

Also fixed on the way: the 17 host tests the radio and dock series added weren't listed in run-host-tests.py, which fails a run on an unlisted test; three couldn't import their helper under python3 -I, and one needed libfdt. T8132's sleep binding lacked ps-prot, and its PCIe host's register names never matched the binding. The dt-check job can't show any of that: its dtschema crashes decoding apple,tunable (typed uint8-array by ti,cd2e22.yaml and via,vl122.yaml) on base and head alike.

Tested on the J713, as a full system with radios, GPU and cache on: 8 of 8 sleeps with a 256 MB check (no bad blocks), radios and GPU back after each, 2 of 2 last-moment refusals rolled back; then 8 of 8 again with the APCIe tables from the bootloader. Every commit builds; the host tests, the sleep, MCC and apple_sep KUnit suites, and CHECK_DTBS on the T8132 boards (no new warnings) pass.

[niteria]

A correction to my reply above. "The host tests ... pass" covered only the 17 tests the radio and dock series added and the suites that read the sleep driver; I hadn't run the whole tools/asahi/run-host-tests.py. On the previous head (d9ec74bc) it fails 17 runs in seven older suites. Each copies a function this series changes into a mock that lacks something the new code uses:

• apple-dart-fw and apple-dart-range (apple-dart.c): the radio sleep record, the sleep writer gate, cio_active.
• brcmf-suspend and bcm4377-resume, with their mutation runs: the radio sleep hooks.
• t6030-clock and apple-dcp (atc.c): the CIO sleep state.
• asahi-m4: G16's publish() now checks the sleep flag.

None of them is a behaviour change. Each mock now has the new state, inert as on other machines, in the commit that first needs it. asahi-m4 also gets a case for a publish during a sleep, and t6030-clock two for the stale ACK a CIO sleep leaves. Only host-test files changed; the kernel is the one tested on the J713.

The whole runner now passes on #1334 (228 of 228 runs) and at every commit of this branch: 228 of 228 up to the last commit, which registers the series' 17 new tests, and 245 of 245 at the head. Before the registration the runner refuses to start, so for those per-commit runs that check was relaxed and the unregistered tests were skipped. Without the registration (on #1420 alone, say) it stops before running anything, which is why none of this showed earlier. CI doesn't run it here either: host-tests.yml only runs on PRs into integration/aurora and aurora-wip, and this one targets #1334's branch.

[DjDeveloperr]

Review of af8707ff5d61, mostly the code shared with other machines (T8140, T6050, M1/M2/M3, the other T8132 boards). File and line references are to af8707ff5d61. This is a code review; the Neo (T8140) rows and the M5 smoke on the merge with integration/aurora are still to come.

Changes outside the J713

1. Every Apple USB4 host now does sleep bookkeeping with the defaults (cio_s2r=0, resume_recovery=0). drivers/thunderbolt/apple.c:2734-2747, in a085c6d "thunderbolt: apple: power-cycle the J713 host router for a deep sleep".

• On the base, apple_nhi_prepare() returned straight away unless resume_recovery was set.

• Now every sleep of every ACIO host does three things:

  • pauses the DP-IN retries;
  • sets pcie_pm_prepared, so the PCIe-C quiesce at a tunnel teardown is deferred to .complete and isn't done synchronously;
  • latches pcie_resume_expected |= pcie_tunnel_requested.

  This covers T8103, T8112, T600x, T602x, T6030, T6040, T6050, T8142 and every T8132 board.

• The latch is cleared only by a successful apple_cio_check_connection() or a fresh NHI probe.

• Scenario: an M5 Pro (or M1/M2 Pro) sleeps docked with a PCIe tunnel, and after the wake that tunnel isn't requested again (the rebuild failed, or the device was deauthorized). The next same-state CD321x update then gets -ENOLINK (apple.c:3369), which cycles ACIO and the PHY. With resume_recovery=0 the base never did that.

• Also, when apple_nhi_cio_prepare() fails, the unwind clears pcie_pm_prepared but leaves pcie_resume_expected set.

• The commit message says the parameter defaults to off and doesn't mention this. If this is meant for every host, please say so in the message. If not, gate it on the J713 native path (cio_s2r) or on resume_recovery, as before.

2. The J715 gets the J713's SEP and external-display domain changes. t8132-j713-j715.dtsi:142-155 (3b1139f, "t8132-j713: enable the SEP for system sleep") is included by t8132-j715.dts.

• So on the J715 the SEP and its DART are enabled, with ps_sep/ps_gpio as domains.
• Through the same file and t8132-j713-pmgr.dtsi (3c3a89d, e2a5b72), the J715 also gets:
  • ps_sep as firmware-owned;
  • the dispext domains as consumer-switched, with the retention workaround;
  • the ACIO hosts' cio PHY.
• The base had the SEP off on every T8132 board, and the sleep controller is still J713-only.
• If the J715 is meant to have these, please say so (it hasn't run on one). Otherwise please move them to t8132-j713.dts.

3. The cluster power gates are enabled on every T8132 board. t8132.dtsi:586 (de972d0).

• The cluster_pwrgate node has no status, so apple-cluster-pwrgate.c's early initcall sets bit 31 in both clusters' words on the J604, J623, J624, J715 and J773g as well. The node's comment notes that "iBoot sets E's only".
• It's the J713's SoC, so this is probably fine. Please confirm it's meant for every M4 board rather than only the J713 with the sleep controller.

Smaller changes on every T8132 board, which look deliberate (just noting them):

• an all-zero ATC PHY is clamped and cold-started instead of refused with -EBUSY (fb147bf);
• the PHY stays in CIO mode when DWC3 stops in USB4 mode (724ca90, which says so);
• unbinding the MCC now turns the cache off and writes it back, and ignores a write-back timeout (2574603, mcc.c:265).

J713 only

4. A refused external-DCP stop leaves that display dark until a reboot. iomfb_m4.c:5500 (27a6e19).

• iomfb_m4_suspend() bumps ext_pm_epoch and sets ext_dart_down before iomfb_m4_ext_stop(), which can now fail.
• When .suspend fails, the PM core never calls this device's .resume, so the matching increment at :5561 doesn't happen and the epoch stays odd.
• From then on, the D572 handler (:1670) treats every powerUpDART(1) as not fresh (!(now & 1)), so ext_dart_down never clears. The external output then keeps returning -EAGAIN.
• Every later sleep adds 2, so the parity never recovers.
• Suggested fix: do the stop before the increment, or undo the increment and ext_dart_down on error.

5. After a refused sleep, the G16 fallback stop leaves every GPU job hanging until the next real deep sleep. g16/sleep.rs:220 and :311 (0230837).

• If leave_for_commit() fails, .prepare stops the firmware itself.
• Any later refusal, such as a wakeup event or another device's suspend error, ends in sleep_hold() with sleeping still set. Meanwhile publish() returns Ok without publishing.
• The PMP driver refuses the sleep in the same situation ("no stop at the sleep's commit: refusing"). Refusing in .prepare here too would avoid a stuck GPU.

6. Lower confidence, from reading the code:

• tb-cio-sleep.c:212-231 (63d7d13):
  • router_sleep can fail after ROUTER_CS_5.SLP or the vendor sleep request was written, e.g. when the controller-state poll times out.
  • The unwind then calls tb_cio_wake_locked() with cio_host_slept still false, so the controller isn't restarted, which per the commit is the only way out of the vendor sleep.
  • Result: a panic, or SLP left set (only with cio_s2r=1).
• iomfb_m4.c:5404 (27a6e19): the comment "Its CPU still runs, or was never left" is reached after iomfb_m4_cpu_run(false) succeeded. A session quarantined there keeps r->done set, so the J713 refuses every later deep sleep while that display stays connected.
• pcie.c brcmfmac (127b5d2): a bus reset whose work runs after the radio host's .prepare took the lease returns -EBUSY and is dropped. The crashed firmware then refuses the sleep at the D3 inform, and nothing reschedules the reset afterwards.
• apple-dart.c:6429 (511d77d): host_off_slept is set in apple_dart_resume(), which is also the runtime-resume callback, and pm_resume_via_firmware() stays true until the next sleep. So a runtime resume after the wake also skips the streams-off check.
• sleep.c:3139 (34c435f): t8132_clean_ram() runs before sleep_ps_save() fills ps_commit[]/ps_aon_commit[] and before aic_sleep_mask_all(). If the explicit clean matters, those are read back stale after the wake.

Checked: no change off the J713

• pcie-apple.c (the #1717 resolution): with cio_dart NULL, apple_pcie_config_read() is #1717's line for line, and cio_dart is set only with cio_s2r on the J713's PCIe-C.
  • The other new paths are gated on managed_sids (accepted only on apple,t8132-pcie), the J713 radio record, or cio_dart.
  • The Neo's suspend_noirq/resume_noirq branches and its async resume are unchanged.
  • The new .prepare returns 0 and .complete does nothing elsewhere.
• Sleep driver: T8140's per-SoC values reproduce the base.
  • Covered: cpu_p_first_bit, has_ave (#1663's AVE paths), the display order and pmc_dcp writes, wake_enable_count, the empty ps_aon_macos_on, a NULL rewake.
  • apple,apcie resolves to the same radio host as before.
• Unchanged elsewhere:
  • rtkit: apple_rtkit_reinit()'s behaviour is unchanged.
  • macsmc: non-T8132 compatibles keep the old table.
  • AIC: only T8132 gets fw_suspend.
  • apple-dart: DARTs without apple,managed-sids or apple,clock-protection take the base paths, including the DAPF and fw_tree DARTs.
  • pmgr-pwrstate: no in-tree apple,bridge-link users outside the J713.
  • brcmfmac and hci_bcm4377: every hook does nothing without the J713 lease; #1653/#1722's s2idle-D0 and WoWLAN branches are as on the base.
  • G13-G17: unchanged.
  • DCP: other DCPs unchanged.
  • apple-pmp-thermal: T8132 firmware path only.
  • NVMe: other SoCs unchanged.

I'd like 1 and 2 answered (a gate or a sentence in the commit message, and the J715 question) and 4 and 5 fixed before merging. The rest can be follow-ups if you prefer. The Neo rows and the M5 smoke results will follow here.

🤖 Generated with Claude Code

[DjDeveloperr]

@niteria, a clarification on item 1 in the review above (the Thunderbolt sleep bookkeeping that every Apple USB4 host now does with the default settings).

The M5 smoke on the merge result is one boot with no sleep, so this path gets no measurement on a non-J713 host before the merge. A note in the commit message won't be enough, then. Either of these code changes would settle it:

• the bookkeeping in apple_nhi_prepare() (apple_dpin_pause_retries(), pcie_pm_prepared, pcie_resume_expected) does nothing unless cio_s2r is on, or resume_recovery is set as on the base; or
• pcie_resume_expected can't lead to the post-wake -ENOLINK cycle unless the sleep actually power-cycled the router.

Items 2 (J715), 4 and 5 are as in the review.

🤖 Generated with Claude Code

[niteria]

integration/aurora moved again after af8707ff5d61, and #1678 (DisplayPort alt mode on T8132) conflicted, so I rebased once more: d0ac23c1cfe4, the same 52 commits on bf00add09e22. It also merges cleanly with today's e661192638a3.

Against af8707ff5d61, git range-diff shows 49 commits identical. Three differ only in conflict context, all T8132:

• Erik's phy: apple: atc: keep a T8132 CIO link when its USB host stops and phy: apple: atc: take the T8132 CIO PHY through a deep sleep. T8132: DisplayPort alt mode on the J604's USB-C ports #1678's t8132_dp_alt and the CIO sleep state are neighbouring fields in struct apple_atcphy and in four host-test copies of it, and each copy keeps both. No code changed.
• The host-test registration commit: context only.

So the patches touch T8140 exactly as before, and T8140 results from af8707ff5d61 still apply to them.

Tested at this head:

• The whole run-host-tests.py: 310 of 310. The base passes its 293.
• KUnit as before: 7 of 7 and 287 of 287. macsmc-parser is 15 of 16, with the same failure as on the base.
• CHECK_DTBS on the six T8132 boards: no new warnings.
• It builds, with no warning beyond the base's.
• On the J713, one boot of exactly this head:
  • 8 deep sleeps, 7 of them with the 256 MB memory check: no bad blocks.
  • 2 sleeps refused late, each followed at once by a sleep that woke normally.
  • The panel off across a deep sleep, then a late refusal: the panel came back.
  • After every wake, the radio, Wi-Fi, Bluetooth and the GPU came back, with no GPU slow start this time.
  • No oops, SError or WARNING, and no thermal sensor error.
  • The AOP sensor drivers the new base binds on T8132 (light, lid angle, motion, temperature) were bound through all of it.

Good to merge from my side once CI is green on this head.

[niteria]

I missed this review when I posted the rebase above, so please disregard my "good to merge from my side" there. The review applies unchanged to d0ac23c1cfe4: the rebase only touched neighbouring fields in atc.c and the host tests' copies of its struct.

I'll answer 1 to 3 and fix 4 and 5 before asking for a merge.