hiveiq-src/hiveops-guide
4
0
Fork 0
Code Issues 10 Pull Requests Actions 1 Packages Projects Releases Wiki Activity

docs(technical.browser): GPU must run out-of-process (SwiftShader int3 crash, #97)

Browse Source 
Document the ~hourly Linux crash: SwiftShader compositor int3 is fatal only under
--in-process-gpu. Add the load-bearing rule (never in-process-gpu), the HIVEOPS_GPU
override, diagnostics, and the not-an-OOM distinction.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Johannes
2026-07-22 13:09:22 -04:00
parent 4fccd30122
commit 668da77241
1 changed files with 21 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/src/main/resources/content/technical/browser/gotchas.md
+21
View File
@@ -27,6 +27,27 @@ self-prompt**, for any user/role — the fix (remove the beta channel, remove th
Bug introduced in `bca8f60` (2026-06-30, ~10 min before the 2.4.7 build); fixed in `8a5cd3a` + `cfbae6a`.
## Linux: the GPU must run OUT-OF-PROCESS (SwiftShader compositor `int3`)
On Linux the app forces **SwiftShader** software rendering (`use-gl=swiftshader` + `disable-gpu`) to
dodge hardware-GPU driver instability on AMD **gfx1150** and other new GPUs (the old exit-133 /
SIGTRAP renderer crash). But SwiftShader's *own* compositor periodically aborts under sustained use —
kernel log `traps: Compositor[...] trap int3 ... in hiveops-browser`.
**The load-bearing rule: never `--in-process-gpu`.** With the GPU in-process, that compositor abort
kills the **whole app** — symptom: *"the browser closes every ~hour, even sitting idle overnight."*
Running the GPU out-of-process makes the same abort a recoverable GPU-process restart instead. Fixed
in **#97 / 2.5.0** by dropping `--in-process-gpu` and `--disable-gpu-sandbox` (SwiftShader kept).
- **Override:** `HIVEOPS_GPU=hardware` uses the real GPU instead of SwiftShader (worth retrying as
Mesa stabilises gfx1150); unset / anything else keeps the safe SwiftShader default.
- **Diagnose:** `journalctl | grep 'Compositor.*trap int3'` (crash history), and confirm a
`--type=gpu-process` child exists (`ps -C hiveops-browser -o args | grep gpu-process`) — present =
out-of-process = good; absent = the fatal in-process config regressed.
- Not to be confused with an OOM: these crashes happen with plenty of free RAM and no oom-killer
events, and leave orphaned `~/.config/hiveops-browser/.org.chromium.Chromium.*` temp files (one per
unclean exit).
## "Self signed certificate" on login = DNS, not code
If the browser errors with a self-signed-certificate message, it's almost always a **DNS / `/etc/hosts`