[DjDeveloperr]

Three wake follow-ups on #1374, from the Neo's final validation on the new Fedora install.

cpufreq: the PMP gets its 30 s again after a firmware sleep. Every wake logged PMP not running after 30 s (-11): fast die-temperature limit unavailable, CPUs 0,5 stay limited to 2688000 kHz. The PMP is stopped by the sleep and restarted only after the CPUs are back, so the first fast-die samples found it down. Since the boot was older than 30 s, that counted as a PMP that never came up, and sampling dropped to its 1 s give-up rate. The window now restarts at the suspend; jiffies don't count the sleep. Neo: 0 lines in 9 wakes on fin3, 0 on fin5 (before: one per boot, at the first wake). scaling_max_freq is back at 4044000 after the wake.

spmi: pending sources nothing uses are cleared, and the acks are flushed. After every firmware sleep the USB-PD SPMI controller fired its chained interrupt twice with nothing to run, and each time we got a bad_chained_irq dump for apple_spmi_irq_handler. Read at those interrupts (TEST kernels):

• the PD controllers' even lines were pending next to the odd ones tipd uses: ACK 0xd1c00/0x1450 against the enables 0x22000/0x28a0;
• another queue's response vector was pending in the vectors' word (0x80 next to ours);
• tipd's own line was pending while the resume still had it masked.

The handler now clears pending sources that no driver has mapped, and the other vectors' responses, as probe does. A mapped source that is pending while masked isn't reported. The ACK block is read back before the parent is unmasked. Neo: 2 dumps per wake before, 1 per wake after (fin5, and fin6 with the read-back). The last one is still open: at that interrupt nothing of ours is pending in the peripheral words.

tipd cd321x: get the PD contract back before rearming the debug link. On 11-09 (fin1 soak, cycle 2) a deep sleep woken by its own debug-link cut slept again as a dark wake, and that second sleep was refused: an early wake event, no device error. The GPU's sleep refused before the commit: GPU firmware never stopped was only a consequence: the refusal came before the commit, where the GPU firmware is stopped. After that the rearm failed all 40 tries at the debug VDM. The partner had dropped the contract (1-0c POWER_STATUS 0), both ends sat at default USB power, and Debug USB was gone until a replug. Now, after 4 failed tries, if the port has no contract, the rearm does what the cut does: a cold reset (Gaid), a wait for the contract, at most 3 times. Then it sets the controller up again as after the cut and continues the tries.

Limits, stated plainly:

• A synthetic repro (TEST knobs: the SMC warned before the cut so the cut wakes the sleep, then the dark-wake re-sleep aborted with an early wake event) refused the re-sleep as on 11-09. That time the contract survived and the link was rearmed on the first try, so the new recovery path did not run. It has not run on hardware yet.
• The same repro showed the Neo's own 1-0c typec partner missing after the refusal (failed to place Type-C mux in safe mode: -110, retrying; ATC Pipehandler lock not acked) until the next sleep brought it back. That is not fixed here.

Tested on the J700 with the kernel #1374 + #1377 + #1393 + this, Fedora 44's kernel config plus the sleep symbols: the rig RAM boot (stage2-15) and the distro boot (GRUB, LUKS, splash). Results are in the final-validation notes.

🤖 Generated with Claude Code

[DjDeveloperr]

Superseded by #1445 (merged as 2b05915); its changes are in integration/aurora. The PMP wait after a sleep is in drivers/cpufreq/apple-soc-cpufreq.c, the SPMI stray-interrupt handling in drivers/spmi/spmi-apple-controller.c, and the debug-link changes in drivers/usb/typec/tipd/.