Multi-monitor layout reset to defaults after amdgpu runtime resume ("Output has no active mode" → "Broken config, all outputs disabled")

[vladngogo]

Summary

On an AMD RDNA3 desktop with two DisplayPort monitors, cosmic-comp fails to
re-apply the saved output configuration whenever the dGPU has runtime-suspended
(BACO) and the screens are woken. Its recovery path then disables all outputs
and resets them to defaults:

• rotation, scale and position are lost;
• one monitor often stays dark until the layout is re-applied from
  Settings → Displays.

It happened on 4 of 4 screen wakes that followed a GPU runtime suspend, and
0 of 8 once GPU runtime PM was disabled. There is a workaround (below), but
the all-or-nothing recovery seems worth fixing regardless.

Environment

• Pop!_OS 26.04, COSMIC (Wayland)
• cosmic-comp 0.1~1789989464~26.04~cd881a4
• Kernel 7.1.5-76070105-generic, Mesa 26.1.6, libdrm 2.4.134
• GPU: AMD Radeon RX 7700 XT / 7800 XT (Navi 32, amdgpu; runtime PM via BACO)
• Kernel cmdline extras: amdgpu.dcdebugmask=0x20002 amdgpu.sg_display=0 (for an
  unrelated suspend issue)
• Displays:
  • DP-1: LG HDR WQHD, 3440×1440@59.973, normal, 100%, Xwayland primary
  • DP-2: LG HDR 4K, 3840×2160@59.997, rotate270, 125%
  • Both report Adaptive Sync Support: false.
• System suspend is disabled; no S3/s2idle is involved.

Steps to reproduce

1. Two DP monitors, one rotated and scaled. GPU runtime PM enabled
   (/sys/bus/pci/devices/<gpu>/power/control = auto, the default).
2. Leave the machine idle until COSMIC's screen-off turns the outputs off. With
   every CRTC off, amdgpu runtime-suspends the dGPU (BACO) after its 5 s
   autosuspend delay.
3. Move the mouse or press a key to wake the screens. The dGPU runtime-resumes
   and the restore fails as below.

In my case the idle period came from switching a DisplayPort KVM to another
computer. The KVM keeps the connectors connected: switches without a screen-off
never trigger this, and no hotplug happens on a switch. So it should reproduce
without a KVM, though I haven't verified that.

Expected

The saved layout is restored when the screens wake.

Actual

The restore fails. The recovery disables all outputs and resets them to defaults.
DP-2 comes back landscape at default scale; DP-1 often stays dark until the
configuration is re-applied from Settings → Displays.

Logs

One occurrence (all four look identical):

15:28:39.229 kernel: amdgpu 0000:09:00.0: PSP is resuming...
15:28:39.648 kernel: amdgpu 0000:09:00.0: SMU is resumed successfully!
15:28:39.657 kernel: amdgpu 0000:09:00.0: [drm] DMUB hardware initialized: version=0x07002F00
15:28:40.041 cosmic-comp: Failed to destroy old mode property blob: No such file or directory (os error 2)
15:28:40.149 cosmic-comp: Failed to destroy old mode property blob: No such file or directory (os error 2)
15:28:40.253 cosmic-comp: Failed to set new config.
15:28:40.253 cosmic-comp: Unable to load output config: Failed to reset config
15:28:40.253 cosmic-comp: Failed to set new config.
15:28:40.485 cosmic-comp: Broken config, all outputs disabled. Resetting... [OutputConfig { mode: ((3440, 1440), Some(59973)), vrr: Enabled, scale: 1.0, transform: Normal, position: (0, 0), enabled: Disabled, max_bpc: Some(16), xwayland_primary: false }, OutputConfig { mode: ((3840, 2160), Some(59997)), vrr: Enabled, scale: 2.0, transform: Normal, position: (3440, 0), enabled: Disabled, max_bpc: Some(16), xwayland_primary: false }]
15:28:40.493 cosmic-comp: Failed to destroy old mode property blob: No such file or directory (os error 2)
15:28:40.585 cosmic-comp: Failed to set xwayland primary output: The provided output wasn't known to Xwayland

The underlying error only appears in the structured journal fields
(journalctl -o verbose), as F_ERR on the two Failed to set new config.
entries (src/config/mod.rs:482):

Failed to switch primary-plane scanout flags

Caused by:
    0: Failed to render outputs
    1: Output has no active mode

Failed to render outputs

Caused by:
    Output has no active mode

These are followed by Unable to load output config: Failed to reset config
(src/backend/kms/mod.rs:347) and Broken config, all outputs disabled
(src/config/mod.rs:435). The same sequence and F_ERR appear on all four
occurrences.

Observations

• It only happens on screen wake after a GPU runtime resume. A screen-off/on
  with the GPU kept powered (amdgpu.runpm=0) logs the mode-blob message (benign,
  per the discussion in cosmic-comp loses DRM access after suspend/resume on AMD RDNA3 #2191) and restores the layout correctly.
• The failure is "Output has no active mode" while re-applying the config right
  after the device re-initializes. The recovery then treats it as a broken config
  and disables every output. Retrying once the device has finished resuming, or
  falling back per output instead of resetting everything, would avoid the
  visible breakage.
• Both outputs carry vrr: Enabled even though both report adaptive sync as
  unsupported. It's probably unrelated; noting it in case it isn't.

Workaround

amdgpu.runpm=0 on the kernel command line keeps the dGPU out of BACO, so no
runtime resume happens.

Related

• cosmic-comp loses DRM access after suspend/resume on AMD RDNA3 #2191: same AMD RDNA3 stack, different failure (DRM commit Permission denied
  after S3 resume).
• cosmic-comp fails to start with previously saved display settings and resets to default settings #1193: the same reset-to-defaults behaviour when the saved config fails to apply
  at startup.