# Chromium/Brave SIGILL on video playback (NVIDIA) crashes Hyprland — hard reboot required

## System

| | |
|---|---|
| OS | Arch Linux |
| Kernel | 7.0.9-arch2-1 |
| CPU | AMD Ryzen 9 5900X |
| GPU | NVIDIA GeForce RTX 5070 |
| NVIDIA driver | nvidia-open-dkms 595.71.05-2 |
| libva-nvidia-driver | 0.0.17-1 |
| Chromium | 148.0.7778.178-1 |
| Brave | 1.92.108 beta |
| Hyprland | 0.55.2-1 |
| Omarchy | (current as of 2026-05-30 upgrade) |


## What happened

Playing video on social/commerce sites (confirmed: Facebook, Twitter, Amazon)
often causes Chromium or Brave to crash and Hyprland to freeze unrecoverably. I'm still
able to shell into the system but as for the physical console I can't even switch
to a different VT.

## Investigation

I shelled in to the system and asked Claude to attempt recovery and
investigate the cause. I rejected some implausible conclusions and now
present one which seems plausible to me (fwiw!).

The browser crashes with SIGILL in its internal Media decoder
thread. The crash leaves an NVDEC hardware decode session open without freeing
it. The NVIDIA GSP firmware then times out waiting for a cleanup RPC, enters
a faulted state, and Hyprland loses its GPU context and aborts. Hyprland's
watchdog restarts it in safe mode, but safe mode fails to bind its socket.
With Wayland mediating VT switching, there is no escape to a TTY.

This happened twice in one day: once with stock Chromium (Facebook),
once with Brave beta (Amazon), each time requiring a reboot.
I've already tried the "ignore the problem; maybe it will go away" solution
but am now filing this issue—it's hard to see how it could be affecting
only me.

## Relationship to existing issues

This is distinct from #2964 (AMD GPU ring timeout during Google Meet) and
#3891 (ANGLE/EGL video rendering broken after Mesa update). Those are GPU
rendering pipeline failures on AMD hardware. This is a crash in Chromium's
software-side media decode thread on NVIDIA hardware, which then wedges the
GPU driver.

The surface symptom — Chromium takes down Hyprland, reboot required — is the
same, but the root cause and affected hardware are different.

## Workaround (unconfirmed)

Add `--disable-features=VaapiVideoDecoder` to
`~/.config/chromium-flags.conf`. This disables only hardware video
decode (NVDEC via VA-API); GPU rasterization, compositing, WebGL, and
canvas acceleration are unaffected.

## Further unconfirmed analysis from Claude

This is Claude writing now:

### Hyprland

When Hyprland crashes (SIGABRT, triggered by GPU context loss after an NVIDIA
driver fault), its watchdog restarts it in safe mode. The safe mode instance
fails to bind its socket, producing a non-functional compositor. Because VT
switching on Wayland is compositor-mediated rather than kernel-handled, the
user has no means of escaping to a TTY. The only recovery is a hard power-off.

There are two bugs here:
1. Safe mode instance fails to bind its socket
2. A broken compositor fully blocks VT switching with no kernel-level escape hatch

### Chromium

Chromium's Media thread terminates with SIGILL (illegal instruction) when playing
video on high-media sites (confirmed: Facebook, Amazon; also reproduced with Brave
1.92.108 beta on Twitter). The crash is deterministic — the same code address is hit
on every occurrence. The process dies without releasing its NVDEC/VA-API resources,
which cascades into an NVIDIA GSP firmware timeout (Xid 175) that degrades the GPU
driver state, takes down the Hyprland compositor, and leaves the system in an
unrecoverable state requiring a hard reboot.

#### Crash detail

##### Kernel trap message (boot of 2026-05-30, Chromium 148.0.7778.96-1)

```
07:27:14  kernel: traps: Media[320806] trap invalid opcode
          ip:55a64c04e477 sp:7f39907fbf50 error:0
          in chromium[8d62477,55a644848000+10240000]
07:27:14  systemd-coredump: Process 320640 (chromium) of user 1000
          terminated abnormally with signal 4/ILL
```

The crashing process is a `--type=renderer` subprocess. The thread named "Media"
is Chromium's in-process media decoder thread. The crash address is identical across
multiple occurrences (different sessions, different browser versions), pointing to a
deterministic code path rather than memory corruption.

The same pattern reproduces with Brave 1.92.108 beta (same upstream Chromium media
pipeline):

```
15:37:55  traps: Media[38436] trap invalid opcode ip:5561296141e8 in brave[...]
15:41:59  traps: Media[41157] trap invalid opcode ip:5561296141e8  ← same address
15:42:28  traps: Media[41963] trap invalid opcode ip:5561296141e8  ← same address
```

##### Coredump summary

```
Signal:   4 (ILL)
Process:  chromium --type=renderer [...]
Thread:   "Media" (worker thread, not main)
Storage:  truncated core (symbols stripped in release build)
Stack:    #0  0x000055a64c04e477 n/a (n/a + 0x0)
```

The binary contains a large number of `ud2` sequences (478,277 in the current
build), consistent with `CHECK()`/`IMMEDIATE_CRASH()` usage throughout the codebase.
The most likely explanation is an assertion failure in the media decode pipeline —
possibly in the VA-API/NVDEC interaction path — rather than a CPU ISA incompatibility.


#### Cascade

After the renderer process dies with SIGILL, the NVIDIA GSP firmware times out
waiting for a `FREE` RPC that will never arrive (the client is dead):

```
07:28:43  kernel: NVRM: Xid (PCI:0000:0b:00): 175, pid=69435, name=chromium,
          Timeout after 75s of waiting for RPC response from GPU0 GSP!
          Expected function 10 (FREE) sequence 230682 (0xbeef0403 0x0).
07:28:46  kernel: NVRM: GPU0 _kgspLogGspTraceCrashBuffer: [dump follows]
```

This leaves the NVIDIA driver in a degraded state. When a subsequent Chromium/Brave
crash occurs (or as a delayed consequence), Hyprland loses its GPU context and
aborts:

```
15:47:30  Hyprland (PID 1852): SIGABRT
15:47:30  xdg-desktop-portal-hyprland (PID 2185): SIGSEGV
```

Hyprland's watchdog restarts it in safe mode, but safe mode fails to bind its
socket. Under Wayland, VT switching is compositor-mediated; with a non-functional
compositor, keyboard VT switching is completely unavailable. Hard reboot required.