[BUG] NVIDIA NVML ioctl stalls (/dev/nvidiactl) trigger runner-stall loop and abort

[akoumjian]

Describe the bug

btop repeatedly logs:

ERROR: Stall in Runner thread, restarting!

and then aborts with:

terminate called without an active exception

Environment

• btop package version: 1.3.0-1 (apt, Ubuntu noble)
• Also tested: source build v1.4.5 with GPU_SUPPORT=true (same behavior)
• GPU: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
• NVIDIA driver: 570.195.03

Reproduction

1. Start btop with GPU enabled (show_gpu_info=Auto/On).
2. Leave running for a while.
3. Observe repeated runner stall messages.
4. Process eventually aborts with terminate called without an active exception.

Expected behavior

• If GPU polling hangs/fails, btop should degrade/skip those metrics and continue running.
• It should not enter repeated runner restarts and abort.

Actual behavior

• Repeated Stall in Runner thread restarts.
• Eventual abort (SIGABRT) and process exit.

Key evidence

strace -f shows worker threads blocking in NVIDIA ioctls on /dev/nvidiactl before abort.

Relevant pattern observed:

• ioctl(9, _IOC(_IOC_READ|_IOC_WRITE, 0x46, 0x2a, 0x20), ...) (fd for /dev/nvidiactl)
• thread then emits terminate called without an active exception
• sends SIGABRT (tgkill(...))

This strongly suggests NVML/GPU polling path stalls (same family as prior NVML-related reports).

What I already tested

• show_gpu_info=Off => stable (no runner stalls in multi-minute soak).
• show_gpu_info=On + nvml_measure_pcie_speeds=False => still stalls.
• shown_boxes excluding proc but keeping GPU => still stalls.
• Upgrading from apt 1.3.0 to locally built 1.4.5 did not eliminate stalls on this system.

Possibly related issues

• [BUG] Crash in btop 1.3.2 (nvidia-related?) #844
• [BUG] Crash in btop 1.3.2 #843
• btop core dumps now #1536
• [Bug] btop 1.4.4 black screen / hangs at startup when GPU support is enabled (on Fedora 42 @ WSL2 + Windows 11 build 26200) #1311
• [BUG] ERROR: Exception in runner thread -> Cpu:: -> graphs, clock, meter : unordered_map::at #899 / PR Don't terminate on GPU access failure #1247 (helped one path, but this stall still reproduces here)

Ask

Could GPU/NVML collection be isolated with hard timeouts/fallback so blocked NVML calls cannot stall the runner and cascade into abort?