Allow DisplayPort detach bookkeeping when the sink is unplugged

[martinstark]

Power-cycling a Dell U2725QE on 615.71.09 can leave DPLib tracking detached heads, breaking subsequent 4K120 setup. This patch restores detach cleanup and same-connector recovery in the tested setup.

Cause and fix

All four previousPlugged modeset guards were introduced in 615.71.09 (dpPreModeset() guard); 610.57.04 has none. dpPreModeset() rejects unplugged detach, leaving dpPostModeset() nothing to complete. NVKMS detaches hardware, clears its head bit and removes group devices; DPLib retains its per-head pointer and an empty, attached activeGroups entry.

With stale groups present, ConnectorImpl2x::enableFlush() proceeds past its early-return checks and attempts flush phase 1. Phase 1 returns 0x40 (NV_ERR_INVALID_STATE), aborting trainLinkOptimized() before hardware training or sink DSC enable. The firmware-side rejection reason is unavailable.

Every selected head is checked before mutation: allow HPD-low detach-only requests; reject any request containing an attachment target. Connector/discovery guards, forced-connected/DDS exceptions and DSC/flush/training policy remain unchanged. Existing detach callbacks and unplug discovery completion already permit cleanup around the hardware update.

Reproduction

RTX 5090, direct DisplayPort, Arch Linux, Sway/Wayland, kernel 7.2.4-arch1-2, NVIDIA modules/userspace 615.71.09. Both modes use 3840×2160/four lanes: 60 Hz at 5.4 Gbit/s without DSC/FEC; 120 Hz at 8.1 Gbit/s with DSC/FEC.

1. Fresh boot at 60 Hz; power off for ~10 seconds, then on. The picture can return with stale state.
2. Select 120 Hz: no signal. Another power cycle fails; moving the cable to another GPU connector recovers it.

Power-cycling from working 120 Hz also fails on stock 615.71.09. Expected: same-connector recovery.

Evidence and limits

Pinned evidence: sanitized timestamped logs, diagnostic patches, control-flow harness and build/reproduction instructions. Instrumentation is excluded from this PR.

Capture	Result
Before: display 0x800, SOR 1	Stale head 2 remains attached; head 3 requests DSC. Flush phase 1 returns 0x40; attach-training result is false.
Fixed: display 0x200, SOR 0	HPD-low detach (headMask=0x8) completes; group/transition lists are empty at rediscovery. Head 3 reattaches at 120 Hz; first training attempt, FEC and sink DSC enable succeed.

Captures use separate boots/connector assignments. The fixed capture covers one instrumented 120 Hz cycle on the same connector: no phase-1 failure or 0x40, but AUX errors persist during power-off/rediscovery. The nv-modeset-kernel.o build and control-flow checks passed. Not tested on hardware: diagnostic-free build, MST, UHBR, eDP/DDS, multi-head/head reassignment, other GPUs and repeated-cycle reliability.

Potentially related: NVIDIA bug 6760883, RTX 5090/Samsung G95NC wake regression (tracking acknowledgement).

AI tools assisted with investigation, patch development and this write-up. I performed the hardware tests.

[CLAassistant]

All committers have signed the CLA.