Niri idle screen-off HDMI hotplug storm OOMs the session

[xdvi]

Summary

On niri, idle power-off-monitors can drop HDMI HPD. The external display then reconnects in a loop, GPU GTT fills, and the OOM killer SIGKILLs niri (taking the whole session with it).

Reproduced on an AMD GPU with HDMI. The same class of bug can hit other GPUs if DPMS tears down the HDMI link the same way.

Steps to reproduce

1. niri + iNiR, internal panel + HDMI on an AMD GPU.
2. Leave the session idle until idle.screenOffTimeout fires.
3. Watch journalctl for HDMI disconnecting connector / new connector repeating every few seconds.
4. After enough cycles: kswapd0 invoked oom-killer and Out of memory: Killed process (niri).

Expected behavior

Screen-off blanks the displays without tearing down the HDMI link. Idle must not grow unreclaimable GPU memory.

Logs

niri: disconnecting connector: "HDMI-A-1"
niri: new connector: HDMI-A-1
inir: attempted to use dangling screen object
...
gpu_active:7106288kB gpu_reclaim:0kB
Free swap = 204kB
Out of memory: Killed process ... (niri) oom_score_adj:200

Dozens of HDMI disconnects in a few minutes after screen-off. iNiR RSS stayed in the hundreds of MB; the missing RAM was GPU GTT.

Environment

iNiR 2.29.1
Quickshell 0.3.1
niri 26.04
AMD GPU (iGPU, shared GTT)

Panel family

ii

Additional context

brightness sleepBegin already ignores hotplug for brightness objects (PR #193). That does not stop niri from connecting the output again, so the DRM buffers still leak. Screen-off should not call niri msg action power-off-monitors when that drops HDMI HPD.

[snowarch]

Confirmed against current prerelease: this is not fixed yet. services/Idle.qml still uses niri msg action power-off-monitors for the idle screen-off path.

This also lines up with an existing upstream Niri bug: niri-wm/niri#3295 is still open and reports VRAM/GTT growth specifically while power-off-monitors is active, eventually consuming unified memory. Related Niri DPMS/output issues include #3950 and #4031.

So the iNiR-side owner is the idle screen-off policy, not the brightness hotplug workaround: suppressing brightness reloads cannot prevent Niri/DRM from repeatedly reconnecting the output. No iNiR fix commit is linked yet.

[snowarch]

PR #259 by @xdvi is the active proposed fix for this issue. I reviewed it against current prerelease; the issue remains open while the PR addresses lifecycle/order regressions and completes final hardware/topology retesting. Keeping the fix and attribution on the contributor PR rather than duplicating the implementation locally.