[Bug]: Renderer crashes with SIGILL (deliberate CHECK trap) on 'Empty cache and hard reload' (Linux, 0.21.2.1) #2927

Closed
opened 2026-11-17 12:52:59 +00:00 by DerDjamel · 0 comments
DerDjamel commented 2026-11-17 12:52:59 +00:00 (Migrated from github.com)

Operating system

Arch Linux x86_64 (kernel 7.2.15), Wayland (Hyprland), AMD Ryzen 5 5600G, root on btrfs over LUKS.

Version

0.21.2.1 (Chromium 157.0.8103.50), official Linux x86_64 build (AUR helium-browser-bin 0.21.2.1-1)

Also re-tested on 0.21.3.1 (Chromium 157.0.8103.78) — see "Additional context" for what happens there.

Have you tested that this is not an upstream issue or an issue with your configuration?

  • I have tried reproducing this issue in Chrome and it could not be reproduced there
  • I have tried reproducing this issue in ungoogled-chromium and it could not be reproduced there
  • I have tried reproducing this issue in Helium with a new and empty profile using --user-data-dir command line argument and it could not be reproduced there

Not yet — happy to run these comparisons on request. Filing now because the crash is fully deterministic on my machine (3/3 attempts) and the core-dump forensics below show it is an in-binary CHECK failure trap, not a config/environment issue.

Description

Using the reload-button context-menu item "Empty cache and hard reload" deterministically kills the tab's renderer process with SIGILL. It reproduced 3 out of 3 times in a single session (Nov 14, 19:10 / 19:13 / 19:15), always at the same instruction.

Crash ID

None — no crashpad report was retained for these crashes (the Crash Reports directory only contains an unrelated report from October). The crashes were captured by systemd-coredump instead; I can share the compressed core, coredumpctl info output, or gdb output from it if that helps.

How to Reproduce?

  1. Browse normally for a while so the disk cache is populated.
  2. Right-click the reload button and choose Empty cache and hard reload.
  3. The renderer of the current tab dies (crash page); repeat attempts crash again the same way.

Actual behavior

All three renderer processes died with SIGILL, si_code: ILL_ILLOPN, at the identical instruction pointer (renderers fork from the same zygote, so the load base is constant):

kernel: traps: helium[691750] trap invalid opcode ip:555cb2a74e38 sp:7ffd80245f90 error:0 in helium[c1a4e38,555ca9b0a000+df02000]
kernel: traps: helium[725194] trap invalid opcode ip:555cb2a74e38 sp:7ffd80246d20 error:0 in helium[c1a4e38,555ca9b0a000+df02000]
kernel: traps: helium[726352] trap invalid opcode ip:555cb2a74e38 sp:7ffd80246ce0 error:0 in helium[c1a4e38,555ca9b0a000+df02000]

The crashing process is --type=renderer:

/opt/helium-browser-bin/helium --type=renderer --ozone-platform=wayland --lang=en-US
  --num-raster-threads=4 --enable-main-frame-before-activation --renderer-client-id=223 ...

Forensics (I kept the 0.21.2.1 package and analyzed the core):

  • The faulting address corresponds to vaddr 0xC1A5E38 in the 0.21.2.1 binary (ELF file offset 0xC1A4E38).
  • Disassembling the official 0.21.2.1 binary at exactly that address shows a ud2 — the compiler's encoding of a deliberate crash (IMMEDIATE_CRASH() after a failed CHECK). So this is an intentional renderer-side CHECK failure, not a stray jump into data and not an unsupported CPU instruction (the CPU is a Ryzen 5 5600G with AVX2, and the faulting page is the binary's own text segment, not JIT memory).
  • Because the crash is deterministic and always the same offset, it should be identifiable from build symbols on your side: ud2 at binary offset 0xC1A4E38 in the 0.21.2.1 Linux x86_64 build.

Each crash left a ~9 GB uncompressed core on my machine (large renderer), so the systemd-coredump processing amplified each tab crash into a multi-minute system-wide stall — that part is my distro's coredump configuration, not the browser, but it's why the crash feels like a full system freeze here.

Expected behavior

"Empty cache and hard reload" clears the cache and reloads the page without the renderer dying.

Additional context

  • Possibly related: #2720 reports the same signature (renderer SIGILL / ILL_ILLOPN, i.e. the same kind of deliberate trap) with a different trigger.
  • Follow-up on 0.21.3.1 (Chromium 157.0.8103.78): repeating the exact same action no longer crashes the renderer — the tab survives. However, I captured what happens instead with a system sampler I run for freeze diagnosis: immediately after the "Empty cache and hard reload" click, the network.mojom.NetworkService process sustained a write burst of 30–160 MB per 5 seconds for ~4.5 minutes (~2.2 GB total, near-zero reads), keeping the disk at 100% busy and pushing I/O pressure (PSI some) to 87.9%, which makes the whole desktop unresponsive on my setup (btrfs on LUKS). The profile's cache directory afterwards is only ~14 MB, so most of those writes appear to be rewritten/deleted churn rather than retained cache entries. I'm not claiming this is the same root cause as the 0.21.2.1 crash — reporting the measurement in case it's a useful clue about what the cache-clear path is doing.
  • Environment details: Helium launched via --ozone-platform=wayland; GPU is the Radeon iGPU of the 5600G (amdgpu); 14 GB RAM; crash times and journal excerpts above are from journalctl -k.
### Operating system Arch Linux x86_64 (kernel 7.2.15), Wayland (Hyprland), AMD Ryzen 5 5600G, root on btrfs over LUKS. ### Version 0.21.2.1 (Chromium 157.0.8103.50), official Linux x86_64 build (AUR `helium-browser-bin` 0.21.2.1-1) Also re-tested on 0.21.3.1 (Chromium 157.0.8103.78) — see "Additional context" for what happens there. ### Have you tested that this is not an upstream issue or an issue with your configuration? - [ ] I have tried reproducing this issue in Chrome and it could not be reproduced there - [ ] I have tried reproducing this issue in ungoogled-chromium and it could not be reproduced there - [ ] I have tried reproducing this issue in Helium with a new and empty profile using `--user-data-dir` command line argument and it could not be reproduced there Not yet — happy to run these comparisons on request. Filing now because the crash is fully deterministic on my machine (3/3 attempts) and the core-dump forensics below show it is an in-binary `CHECK` failure trap, not a config/environment issue. ### Description Using the reload-button context-menu item **"Empty cache and hard reload"** deterministically kills the tab's renderer process with SIGILL. It reproduced 3 out of 3 times in a single session (Nov 14, 19:10 / 19:13 / 19:15), always at the same instruction. ### Crash ID None — no crashpad report was retained for these crashes (the `Crash Reports` directory only contains an unrelated report from October). The crashes were captured by systemd-coredump instead; I can share the compressed core, `coredumpctl info` output, or gdb output from it if that helps. ### How to Reproduce? 1. Browse normally for a while so the disk cache is populated. 2. Right-click the reload button and choose **Empty cache and hard reload**. 3. The renderer of the current tab dies (crash page); repeat attempts crash again the same way. ### Actual behavior All three renderer processes died with `SIGILL`, `si_code: ILL_ILLOPN`, at the identical instruction pointer (renderers fork from the same zygote, so the load base is constant): ``` kernel: traps: helium[691750] trap invalid opcode ip:555cb2a74e38 sp:7ffd80245f90 error:0 in helium[c1a4e38,555ca9b0a000+df02000] kernel: traps: helium[725194] trap invalid opcode ip:555cb2a74e38 sp:7ffd80246d20 error:0 in helium[c1a4e38,555ca9b0a000+df02000] kernel: traps: helium[726352] trap invalid opcode ip:555cb2a74e38 sp:7ffd80246ce0 error:0 in helium[c1a4e38,555ca9b0a000+df02000] ``` The crashing process is `--type=renderer`: ``` /opt/helium-browser-bin/helium --type=renderer --ozone-platform=wayland --lang=en-US --num-raster-threads=4 --enable-main-frame-before-activation --renderer-client-id=223 ... ``` Forensics (I kept the 0.21.2.1 package and analyzed the core): - The faulting address corresponds to vaddr `0xC1A5E38` in the 0.21.2.1 binary (ELF file offset `0xC1A4E38`). - Disassembling the official 0.21.2.1 binary at exactly that address shows a **`ud2`** — the compiler's encoding of a deliberate crash (`IMMEDIATE_CRASH()` after a failed `CHECK`). So this is an intentional renderer-side CHECK failure, not a stray jump into data and not an unsupported CPU instruction (the CPU is a Ryzen 5 5600G with AVX2, and the faulting page is the binary's own text segment, not JIT memory). - Because the crash is deterministic and always the same offset, it should be identifiable from build symbols on your side: `ud2` at binary offset `0xC1A4E38` in the 0.21.2.1 Linux x86_64 build. Each crash left a ~9 GB uncompressed core on my machine (large renderer), so the systemd-coredump processing amplified each tab crash into a multi-minute system-wide stall — that part is my distro's coredump configuration, not the browser, but it's why the crash feels like a full system freeze here. ### Expected behavior "Empty cache and hard reload" clears the cache and reloads the page without the renderer dying. ### Additional context - Possibly related: #2720 reports the same signature (renderer SIGILL / `ILL_ILLOPN`, i.e. the same kind of deliberate trap) with a different trigger. - **Follow-up on 0.21.3.1 (Chromium 157.0.8103.78):** repeating the exact same action no longer crashes the renderer — the tab survives. However, I captured what happens instead with a system sampler I run for freeze diagnosis: immediately after the "Empty cache and hard reload" click, the `network.mojom.NetworkService` process sustained a **write burst of 30–160 MB per 5 seconds for ~4.5 minutes (~2.2 GB total, near-zero reads)**, keeping the disk at 100% busy and pushing I/O pressure (PSI `some`) to 87.9%, which makes the whole desktop unresponsive on my setup (btrfs on LUKS). The profile's cache directory afterwards is only ~14 MB, so most of those writes appear to be rewritten/deleted churn rather than retained cache entries. I'm not claiming this is the same root cause as the 0.21.2.1 crash — reporting the measurement in case it's a useful clue about what the cache-clear path is doing. - Environment details: Helium launched via `--ozone-platform=wayland`; GPU is the Radeon iGPU of the 5600G (amdgpu); 14 GB RAM; crash times and journal excerpts above are from `journalctl -k`.
Sign in to join this conversation.
No milestone
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
imputnet/helium#2927
No description provided.