[samucarre]

Problem. On the T6031 the G15 is powered by its PMGR domain, gfx, alone, and once started its firmware session lasts until reboot: nothing starts it again. The G15 holds gfx with a runtime-PM reference, but in the noirq phase of system sleep genpd counts the GPU, the only device in gfx, as suspended whatever its runtime-PM state, and powers gfx off under the running firmware (genpd_finish_suspend(), genpd_sync_power_off()). genpd spares devices in the wakeup path only in GENPD_FLAG_ACTIVE_WAKEUP domains, and PMGR gives that flag only to domains it finds on when it probes them; gfx is off then. This is read from genpd, not observed: before this change no system sleep went past pm_test=devices here with the G15 running.

Change. While the G15 holds gfx, the GPU is a syscore device (dev_pm_syscore_device()), as dwc3-apple, the ISP and the SEP do for the domains of their resident firmware: the PM core skips it in every phase, genpd never counts it as suspended, and gfx stays on, under afr, which is always on. asahi_g15_release_power() clears the flag with the hold. Nothing changes while the system is awake: gfx is powered when it was before, and not at all in a boot without the G15. The admission now reads the ASC control word before it takes the hold, so that no failure there leaves the hold, or the flag, behind. T6031 only, whose configuration admits the J516C alone; g15_sleep_hold=0 restores the old behaviour, and the admission logs which of the two applies. T6030 is not changed.

Marking gfx apple,always-on from the display gate would keep it on in sleep too, but PMGR would then power gfx up when it probes it, before the PMP has started and well before the G15's admission (1.62 s, 2.17 s and 3.44 s in one boot), and keep it up in boots where the G15 never starts.

Nothing else changes for sleep: the GPU mailbox's interrupts are IRQF_NO_SUSPEND and neither the G15's poll work nor RTKit's is freezable, so the firmware is still served during the sleep, each message waking a CPU briefly without ending it; the G15's deadlines run on CLOCK_MONOTONIC, which s2idle does not advance here; and the firmware is not told about the sleep. With #258, its PMP AGX request is not withdrawn for the sleep either.

Checked on a MacBook Pro 16" M3 Max (J516C, 64 GB), integration/aurora 2f72554 plus local patches, among them the sleep work of #274 (now merged) with the J516C RTC alarm of #285, HDMI on dcpext0, #258 and #261, with the G15 running:

• boot log: G15: the GFX domain stays on through system sleep, as nothing starts the firmware again (g15_sleep_hold=0 lets genpd power it off)
• pm_test=platform, then three s2idle sleeps of 30 s, each ended by the RTC alarm: gfx was never powered off (pm_genpd/gfx/idle_states counted no new entry into its off state), the GPU mailbox raised 0 to 8 interrupts per sleep without ending it, and after each sleep the GPU ran compute work (12,296 computes in 4 s).
• with the G15 not loaded, the same local kernel without this change slept the same way (3/3), so the sleep itself comes from Make M3 Pro sleep use 1.4 W instead of 2.6 W, and add the RTC wake alarm #274 and T6031 RTC alarm for #274 (M3 Max J516C) #285 (with local patches), not from this one.
• builds on 2f72554 (arm64, clang 21) with no new warnings in drivers/gpu/drm/asahi/asahi.o; g15/glue.o is clean with W=1; the G15 host selftests pass. It touches none of the files of Make M3 Pro sleep use 1.4 W instead of 2.6 W, and add the RTC wake alarm #274 or T6031 RTC alarm for #274 (M3 Max J516C) #285, and merges cleanly with soc: apple: report the T6031 GPU to the PMP #258 and drm/asahi: g15: allow 8 prefixes in flight per VM (was 2), tunable with g15_max_inflight #261.

Not checked. This branch alone on hardware; an HDMI display attached through the sleep; long sleeps and their battery cost (the mailbox interrupts keep waking a CPU briefly); other T6031 boards.

🤖 Generated with Claude Code

[chriskearney]

Reviewed all four changed files at exact head ff2c986fd2db9ccd4af5ae616a5b3cd58594dd11 and the independent PM-core/genpd and lifetime analysis. The syscore flag is restricted to the existing T6031/J516C DeviceTree G15 admission, is cleared on the pre-firmware unwind, and the ASC read now precedes the power hold. T6030 remains outside this gate. No source blocker found; normal current CI and a full native package build/merge-group result are still required.

The author's three s2idle/compute observations are on integration plus local patches, not an exact-head hardware qualification. Long-sleep battery cost remains unmeasured. This review does not qualify a new board, firmware image or HDMI/sleep cohort, and no additional physical test is required for this protective scoped change. Small documentation nit: SEP is not a syscore precedent in this tree; dwc3-apple and ISP are.