[erikwb]

Depends on #574 (the first 7 commits here are #574; the new commit is the last one).

The two SN201202x USB-C controllers on the MacBook Air (13-inch, M4) stay awake as long as the system does, even with nothing attached. The SMC reads about 200 mW on the 3.3 V rail that feeds them while they are awake, and about 1 mW once they sleep.

This lets an idle controller sleep, with runtime PM:

• A register access wakes it, and it sleeps again 1 s after the last one (power/autosuspend_delay_ms).
• Its event interrupt stays live while it sleeps, so an attach or a power change still reaches the handler, whose first read wakes it.
• It sleeps through charging and with a USB device attached. A DisplayPort, Thunderbolt or USB4 link keeps it awake, as does DBMa debug mode.
• After a few failed tries to sleep, it stays awake until the next access.
• System sleep wakes an idle controller before its own SLEEP handshake, and resume hands it back to runtime PM. If a controller does not wake, it sleeps on through system sleep without the wakeup watch, and resume wakes it.
• Only the T8132 controllers use idle sleep. Elsewhere the SPMI bus's runtime PM reference is kept and nothing changes.
• power/control = on turns idle sleep off.

Testing

On a MacBook Air (13-inch, M4):

• Power. The display is blanked and nothing is plugged in. Four alternating pairs compare power/control on against auto:
  • total system power is 1199 mW → 1113 mW on average, a saving of ~87 mW; each pair saved 64–116 mW;
  • the 3.3 V rail drops from ~209 mW to ~0.
• Plug tests, both controllers idle-asleep first:
  • Charger plug and unplug: the controller wakes for the PD negotiation, sleeps again 2 s later, and stays asleep for 90 s while the machine charges.
  • USB keyboard: it enumerates and works while the controller is asleep, and the unplug is seen.
  • The port with nothing attached never wakes.
• Suspend-to-idle with both controllers asleep: plugging in a USB keyboard wakes the system and the keyboard enumerates. Both controllers are asleep again within 5 s.
• KUnit: the sn201202x-pm suite passes 21/21 on arm64 QEMU. Register accesses go through the driver's own regmap bus to a simulated SPMI controller, and the fixture binds and unbinds the way the SPMI bus does. It covers:
  • idle sleep, woken by a register access;
  • accesses from the system sleep event path, which leave runtime PM alone;
  • retry, give-up, re-arm, and an access restarting the count;
  • an unconfirmed sleep with a failed wake;
  • the display link, an unplug, and DBMa found at probe or entered through debugfs;
  • unbind from idle sleep and rebind;
  • a controller without idle sleep;
  • system sleep from idle sleep, including a failed wake and the wakeup watch.
• Checks: checkpatch --strict is clean, and so is a W=1 build.

Known limitation

If a wake fails while runtime PM still counts the controller as active, register access fails until the next successful wake. Before this change that state was likewise unrecoverable.

This change was developed and tested by Claude (Anthropic), working with me on the J713.

[erikwb]

Rebased onto the current integration/aurora together with #574. host-tests failed here because the new host harnesses (neo-stability, apple-dcp) compile the dwc3-apple and tipd functions that #574 changes against stubs that lacked its new fields. #574 now updates those stubs in the commits that need them, and host-tests pass on both PRs. The sn201202x transport selftest is updated as well. The idle-sleep change itself is unchanged.

[erikwb]

Superseded by #833, which carries only the idle-sleep commit, stacked on #832.