Abort (SIGABRT) in nvhttp thread on first run — capture initialises fine, no ports ever bind (Wayland/COSMIC, VAAPI)

[jon-m-craig]

Title

Abort (SIGABRT) in nvhttp thread on first run — capture initialises fine, no ports ever bind (Wayland/COSMIC, VAAPI)

Body

Description

On a fresh install, Sunshine completes capture and encoder initialisation successfully, then
aborts in the nvhttp::47984 thread before binding any of its ports. No exception message is
printed — the process calls abort() directly, so there is no terminate called after throwing…
text to go on.

Because it dies before the web UI binds, https://localhost:47990 is never reachable and initial
setup cannot be completed. Under systemd it crash-loops until StartLimitBurst stops it.

Everything up to that point works — including the xdg-desktop-portal path on COSMIC, DRM
screencasting, and VAAPI encoder selection — so this does not look like a capture or compositor
problem.

Steps to reproduce

1. Fresh install of sunshine (AUR, package maintained by LizardByte)
2. Ensure no prior state: rm -rf ~/.config/sunshine/credentials ~/.config/sunshine/portal_token
3. Run sunshine from a terminal
4. Accept the COSMIC portal prompt (both displays, "always allow")

Expected

Sunshine finishes starting, binds 47984/47989/47990, web UI reachable for first-run setup.

Actual

Aborts immediately after logging that sunshine_state.json does not exist:

[info]: Found H.264 encoder: h264_vaapi [vaapi]
[info]: Found HEVC encoder: hevc_vaapi [vaapi]
[info]: Open the Web UI to set your new username and password and getting started
[info]: File /home/<user>/.config/sunshine/sunshine_state.json doesn't exist
[1]  <pid> abort (core dumped)  sunshine

ss -ltn shows nothing listening on 47984, 47989 or 47990 at any point.

Coredump

PID: <pid> (sunshine)
TID: <tid> (nvhttp::47984)
UID: 1000
Signal: 6 (ABRT) si_code: SI_TKILL

Stack trace of faulting thread:
#0  n/a (libc.so.6 + 0x9a11c)
#1  raise (libc.so.6 + 0x3e5d0)
#2  abort (libc.so.6 + 0x25685)
#3  n/a (sunshine + 0xcba04)
#4  n/a (sunshine + 0x12b6359)
#5  n/a (libgcc_s.so.1 + 0x22043)
#6  _Unwind_RaiseException (libgcc_s.so.1 + 0x227ae)
#7  __cxa_throw (libstdc++.so.6 + 0xb5ac7)
#8  n/a (sunshine + 0x210ad)
#9  n/a (sunshine + 0xe8f6d)

So an exception is thrown and propagates out of the nvhttp thread, and the handler aborts
without surfacing what().

Reproducibility

Consistent. Reproduced across roughly a dozen starts, both under systemd --user and directly
from a terminal. Deleting credentials/ (certs regenerate cleanly) and portal_token does not
change the outcome.

What does work

Worth stating, since it rules a lot out:

• xdg-desktop-portal on COSMIC prompts, both displays granted, restore token saved and reloaded
• [portalgrab] Found stream for display id/name: 'DP-1' ... 2560x1440
• [portalgrab] Found stream for display id/name: 'HDMI-A-1' ... 1920x1080
• Found monitor for DRM screencasting / Found connector ID / Found cursor plane
• Found H.264 encoder: h264_vaapi and Found HEVC encoder: hevc_vaapi

Environment

Sunshine	2026.914.233613, commit 63d35f702ee9e362e43263742981836ec0710384
Install	AUR sunshine package
OS	Arch Linux, kernel 7.2.7-arch1-1
Desktop	COSMIC 1.9.0, Wayland (XDG_SESSION_TYPE=wayland)
GPU	AMD Radeon RX 550 (polaris11), amdgpu, /dev/dri/card1
Mesa	26.2.3-arch1.1, radeonsi
Displays	DP-1 2560x1440 @ 0x0, HDMI-A-1 1920x1080 @ 2560x360

Notes

• getcap /usr/bin/sunshine → cap_sys_admin,cap_sys_nice=p, and the log confirms Sunshine
  drops both at startup by design.
• Setting capture = kms makes the KMS monitor list come back empty (Couldn't find monitor [0]),
  which appears to be a consequence of that capability drop rather than a separate fault — the
  default auto-detect path finds the monitors fine.
• No other process is bound to 47984/47989/47990.