j314s (MacBook Pro M1 Pro): s2idle draws ~1.9–2.0 W; measurements and what's ruled out #290

Mac and installed build

• Mac: MacBook Pro (14-inch, M1 Pro, 2021), apple,j314s / apple,t6000, 8 cores.
• Installed: Omarchy 4.5.0rc2 (Arch), linux-aurora 7.2.11.aurora9-1, m1n1-aurora 1.6.1.aurora3-1, uboot-asahi 2026.07.asahi2-1, GRUB 2.16.
• Also measured: the CI package of Make M3 Pro sleep use 1.4 W instead of 2.6 W, and add the RTC wake alarm #274 (7.2.11.aurora10-1.382, installed beside the regular kernel). No local kernel patches.

Both firmware versions and provisioning

asahi,os-fw-version: 13.5, asahi,system-fw-version: 26.6.2

Reproduction, expected result, and actual result

Close the lid on battery with nothing attached. s2idle draws about 1.9–2.0 W, so a full charge lasts about a day and a half asleep. I've been measuring it since early October with a systemd-sleep hook that logs battery state and kernel power-state snapshots around each sleep, and I want to hand over what's already ruled out.

Measurements

Short lid-closed runs (11–14 min), averaged from charge_counter:

Overnight (17–24 h, from energy_now):

Method notes:

• On this machine charge_counter counts mA·s, not µAh. Over 61 s awake it gave 12.3 W against energy_now's 12.1 W.
• energy_now rose 0.7–0.85 Wh during every short sleep, with no charger attached (the gauge re-estimating after a rest). So short runs need charge_counter, and night-to-night differences under ~50 mW are within gauge noise.
• Short runs include a few seconds awake at each end, roughly +100 mW at 12 min, so they sit above the overnight figures.

Where the power isn't

• CPUs: all 8 cores enter CPU PD in s2idle; s2idle/usage counts none in WFI. Each core logs about 3.6 entries per second of sleep, the rate of the AOP mailbox interrupt (293408000.mbox, IRQ 40), which fires about 3.6 times a second all night.
• AOP: unbinding 293c00000.aop stops that interrupt but didn't change sleep power. It also can't be rebound on t6000 without a reboot (RTKit discovery fails with -62).
• Radios: soft-blocking Wi-Fi and Bluetooth saves about 41 mW.
• USB-C: atc0–3_usb stay on in sleep only because the dwc3-apple controllers are wakeup-enabled (GENPD_FLAG_ACTIVE_WAKEUP). With their wakeup disabled for the sleep they switch off. The saving is somewhere between 40 and 85 mW, unresolved because the kernel changed between nights.
• Display PMP requests: pmp-disp0 and pmp-dispext0 are asserted all night, but withdrawing them changes nothing (the table above; details in Make M3 Pro sleep use 1.4 W instead of 2.6 W, and add the RTC wake alarm #274).
• The PMP is on for this board (pmgr-misc: PMP resume fix, idle fabric governor and PMP for j314s #91: "PMP manages fabric and DCS p-states"), so apple_pmgr_misc.idle_fabric_pstate doesn't apply.

What stays powered

From pmgr-pwrstate dynamic debug across s2idle:

• Display coprocessors: disp0_cpu0, disp0_fe, dispext0_cpu0 and dispext0_fe never switch off. pmp_report_disp0 and pmp_report_dispext0 are apple,always-on in t600x-die0.dtsi, and they pin their PMGR parents. (dispext1_* does switch off.)
• Wi-Fi PCIe link: apcie_gp and pcie0_ref never switch off either. They're parents of the apple,always-on apcie_gp_sys, which holds 590000000.pcie. Removing the Wi-Fi/BT PCI functions for the sleep doesn't free them.
• 73 domains are apple,always-on: DCS and AMCC, the AFNC fabric, AIC, nub/aon/SPMI, pms, pmp, sep, soc_dpe, the ATC common, pcie and usb_aon blocks, and debug.
• Everything else powers down: gfx, ans2, audio, i2c, spi, uart, sio, mca, ane, isp, avd, atc*_cio*, dispext1, and the SD reader's link (pcie1_ref).

So what's left looks like the display coprocessors kept up because their firmware can't be restarted, plus the always-on SoC blocks. Is there known work on either on T600x: a DCP low-power state for s2idle, or the SoC sleep sequence (memory and fabric) beyond #170's CPU context-loss diagnostic? Or something cheaper to try first?

I can test on J314s. I can install a PR's CI package beside the regular kernel (no boot-chain changes, one-shot boot) and return charge_counter sleep numbers plus a PMGR-domain and interrupt breakdown the same day.

Relevant error lines

None; nothing fails. Every resume logs brcmf_p2p_set_firmware: failed to update device address ret -52 and spi-hid-apple-of spi2.0: status message mismatch on both aurora9 and aurora10.

Related issues, pull requests, or code examples

• pmgr-misc: PMP resume fix, idle fabric governor and PMP for j314s #91: PMP for J314s, and probably most of the 7.1.12 → 7.2.10 drop.
• Make M3 Pro sleep use 1.4 W instead of 2.6 W, and add the RTC wake alarm #274: DISPEXT withdrawal (the T600x data point is in its comments).
• thunderbolt: apple: opt-in port power-off for s2idle with CD321x restore #92: Thunderbolt teardown; doesn't apply with nothing attached.
• soc: apple: add gated J314s CPU context-loss diagnostic #170: J314s CPU context-loss diagnostic.

What was not tested

• No external display or dock.
• No RTC-timed runs (no wake alarm on T600x), so every short run is a manual lid close.
• Overnight figures are energy_now only; no full night with charge_counter yet.
• No other T600x board.