[DjDeveloperr]

The G16 counterpart is in #1314: when the sleep's commit can't take the GPU's stop, G16 now refuses the sleep in .prepare. G17 still had a fallback there, which this removes.

Before. If asahi_g17_commit_prepare() failed and asahi_g17_commit_required() was false, G17's .prepare stopped the firmware itself ("no GPU stop at the commit: stopping it now"). Any later refusal of that sleep then took .complete into sleep_hold(): the firmware stayed stopped and publication stayed closed until the next real sleep. Every GPU job hung meanwhile.

After. .prepare only drains, checks for quiet and leaves the stop to the commit, or refuses the sleep with EBUSY if the commit can't take it. A sleep refused before the commit then always finds the firmware never stopped, and .complete just reopens publication.

Reach. The A18 Pro (T8140) is the only G17 chip with sleep_restart; the others refuse sleep outright. On it, the GPU's .prepare only runs for a firmware sleep (driver.rs). The only source of those is the sleep driver, which leaves the stop to its commit, so asahi_g17_commit_required() is true and the fallback was not reached. The change takes out a stop path that a future sleep path could reach, and with it the Rust stop sequence (down(), and the idle/wake/stop hardware steps it alone used) and their kunit tests. up(), hold() and lower_domain() keep theirs. asahi_g17_commit_required() stays in sleep.c/sleep.h for now, to stay clear of #1314's lines there.

Testing

• CI on this head.
• MacBook Neo (J700), on a one-shot test boot of this aurora.ko:
  • two late refusals forced through the PM core's pm_test (devices, then platform: every device suspended, then the sleep refused before the platform's enter, where the commit runs);
  • one normal DRAM-kept sleep;
  • a GPU smoke (dEQP triangle, gpubench, vkcube) and the kernel log checked for GPU faults after each.
  • Results to follow.

🤖 Generated with Claude Code

[DjDeveloperr]

Tested on the MacBook Neo (J700), 2026-11-13. A one-shot test boot ran this change as aurora.ko, built from the same g17/sleep.rs on a slightly older integration base (the file is identical there), against the Neo's default kernel. Then back to the default boot. PASS.

• Two late refusals, forced through the PM core's pm_test (devices, then platform): every device suspended, then the sleep was refused before the platform's enter, where the commit runs.
  • Each logged sleep refused before the commit: GPU firmware never stopped, and neither logged stays stopped until a real sleep.
  • After each, the GPU passed a fresh-process Vulkan triangle (dEQP) and a gpubench render.light row (status: ok).
• One normal DRAM-kept sleep, woken by the PMU RTC: the GPU firmware restarted, and nothing was refused or lost. The same smoke passed afterwards.
• Kernel log: no GPU fault or GPU driver warning over the whole run.
• vkcube: it couldn't reach the session's Wayland socket from the test harness, at the start too, so it didn't count either way.
• Before merge: CI (host-tests) green on this head. No new compiler warnings: the dead-code warning set equals that of a build of the same base without this change.