[erikwb]

Follow-ups to #1314 for the J713's deep sleep.

• USB4/Thunderbolt devices no longer block suspend. pcie_apple.cio_s2r and thunderbolt_apple.cio_s2r now default to on.
  • Before: with them off and deep the J713 default, a dock or other USB4/Thunderbolt device in a port made the T8132 PHY refuse every system sleep. DWC3 stops its host at PM_SUSPEND_PREPARE, the PHY keeps its CIO mode, and nothing hands the link to the NHI.
  • Now: the NHI takes the host router, the CIO PHY and the tunneled PCIe-C port through the deep sleep.
  • Refusals kept: the NHI still refuses a deep sleep it can't take over (no PCIe-C tunnel, XDomain rings open, the cable state changing), and the PHY then refuses the sleep as before.
  • Scope: both checks stay limited to the J713 (apple,j713 with the T8132 ACIO and PCIe-C), so other machines are unaffected. cio_s2r=0 on either module restores the old behaviour.
  • Docked s2idle: with the handoff on, the tunneled port keeps its link for an s2idle and refuses the s2idle if it can't keep the link, instead of stopping the port.
  • Older DT: a DART that an older device tree gives no sleep path no longer fails the PCIe-C probe. The port keeps the stock resume ordering, and a docked deep sleep is refused.
• brcmfmac: reset firmware that died during a sleep.
  • Before: while the platform owns the radio for a system sleep, reset work is blocked. A pending reset is cancelled when the lease starts, a new one isn't queued, and brcmf_pcie_reset() refuses. Nothing asked for the reset again when the lease ended, so firmware that died during the sleep stayed dead until the driver was reloaded.
  • Now: the firmware-death handler records that it wants a reset, a reset that runs clears it, and the lease end queues the reset again if it is still wanted. The recovery limit still applies.
  • Tests: a replay case in brcmfmac-sleep-lifetime-tests with a mutation that drops it. The recovery test checks the flag per reset and past the limit, with two new mutations.

Validation on this branch:

• Build and host tests: full arm64 Image, modules and DTBs. checkpatch is clean on all three commits. The brcmfmac host suites pass, and every mutation is rejected.
• J713 setup: 24 GB, with a CalDigit TS5 (USB4 at 20 Gb/s, PCIe tunnel with the dock's Ethernet, two DisplayPort monitors). Aura 1.0.0b11. No cio_s2r option on the command line; both parameters read Y.
• J713 run: one suspend test at the platform stage, then one real deep sleep with an RTC wake on the same boot.
• J713 results: every NHI, PHY, DART and PCIe-C stop and restore step ran, and the Wi-Fi and Bluetooth recovery after power loss completed. The dock came back at 20 Gb/s with Ethernet up and both monitors connected. No kernel warnings, no taint.

Not covered:

• Repeated sleeps on one boot.
• Wake sources other than the RTC.
• Docks other than the TS5.
• A firmware death during a sleep on hardware: the brcmfmac change is covered by the host tests only.
• A docked freeze while deep is the default: it is still refused at the PHY, because DWC3 stops its host and the NHI handoff only runs for a deep sleep.

These changes were developed and tested by Claude (Anthropic), working with me on the J713.

[DjDeveloperr]

Reviewed 324fb8579a02; approving. Thanks, Erik.

• cio_s2r default on. Both uses stay behind the T8132 ACIO/PCIe-C hardware and apple,j713, so other machines take apple_pcie_tunnel_add_link() as before. On the J713, an older DT whose DART has no sleep path now keeps the tunnel with a warning instead of failing the PCIe-C probe.
• An undocked J713 sleep is unaffected. The NHI exists only between apple_cio_start() and apple_cio_stop(), so with nothing plugged in its prepare never runs. A USB4 device without a PCIe tunnel is still refused, as before.
• brcmfmac.
  • reset_wanted is set only after the recovery-limit check, so the limit still holds.
  • A reset that runs clears it.
  • It is replayed only from brcmf_pcie_sleep_finish(), the J713 radio lease's end. Elsewhere it is a flag that is set and cleared with no other effect.
• CI. host-tests passed on the merge with ccc5a9d41de7. Since then integration/aurora gained only asm/mmu_context.h (#1509) and g17/sleep.rs (#1839). Neither is near these files, and the merge with the current tip is clean.
• J713 evidence: the docked deep sleep in the description. A follow-up we'd welcome: several docked sleeps on one boot, and a wake source other than the RTC.

[DjDeveloperr]

Merged as 4e5be40d8101 onto integration/aurora 03f060b251c8. The merge tree is the clean merge-tree I checked against that tip; the base had moved only by aop.rs (#1645) since the review. The functional change is J713-only, so the J713 run in the description is its hardware evidence. It also goes into the J700 kit that verifies #1314 after its merge. Thanks, Erik.