[wachi35676]

What

guest_system_load_exec_standard_error fails on every run on Linux hosts whose kernel lacks asynchronous userfaultfd write protection (before 6.7, e.g. WSL2's 6.6 or Debian 12's 6.1). It aborts without a message (Subprocess aborted), and several open PRs list it as failing on main in their containers (#10479, #10755, #10767, #11167).

The test requires the re-executed child's host standard error to be exactly started <mode>\n. On such a kernel the runtime first prints its start-up diagnostic [memory] write watch unavailable: asynchronous userfaultfd write protection failed (Invalid argument); guest memory is compared instead (GuestWriteWatch.cpp) to the same stream, so the exact comparison fails. strace of the child shows the diagnostic, then started list\n, then exit 0. With APS5_NO_WRITE_WATCH=1 the unchanged test passes.

The check now skips that one line, [memory] write watch unavailable: ... (matched by its exact prefix), when it comes first. The rest of the host stream must still be exactly started <mode>\n, and the guest stream must still be empty. Any other leading line still fails. This is a test-only change with no runtime code.

Tested

• Linux, WSL2 Ubuntu 24.04 (kernel 6.6.87.2), GCC 13.3, Release, BUILD_TESTING=ON, main 90f918f (re-run at fff18d2 with the narrower match: 3/3 pass):
  • main: the test aborts on every run.
  • This change: it passes 3/3 (--repeat until-fail:3).
  • With both dup2 calls that give the host standard error back to descriptor 2 removed (Export.cpp StartHostExecutable, LoadExec.cpp Replace), the updated test still fails, so it still catches the bug it guards.
  • Full ctest: 23 failures on main and 23 with this change. This test drops out of the list. The rest are the same lavapipe agc_* and pr_overlap_encoding failures in both runs, plus guest_memory's mapping-cost timing ratio, which failed once under -j 4 and passes alone 3/3.
• Windows 11, WinLibs GCC 15.2.0, Release: full build. The test is only registered on Linux and macOS. guest_system*, loadexec* and lnc_util* (22 tests) pass.
• python tools/check_conventions.py --base upstream/main passes.
• macOS: not tested.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• AI-assisted: yes (Claude Code). I ran the builds and tests above on my machine.

🤖 Generated with Claude Code

https://claude.ai/code/session_01En5cnt6RwCiNXDJwm9t8Tf

[MarcZX10]

The reported WSL2 reproduction explains the failure and the test still checks the child line and guest stream.

[MarcZX10]

One non-blocking test-hardening point: this skips every leading line beginning with [ rather than only the runtime's [memory] write watch unavailable: diagnostic. For example, [unexpected]\nstarted list\n is accepted as the exact expected output. Could this match the known startup-diagnostic prefix (or otherwise constrain accepted diagnostics) so unrelated child stderr cannot be hidden?

[wachi35676]

@MarcZX10 Thanks, fixed in fff18d2. The check now skips only a first line that begins with [memory] write watch unavailable: (the exact prefix GuestWriteWatch.cpp prints). Any other leading line, such as your [unexpected]\nstarted list\n, no longer equals started list\n, so the test fails. The rest of the host stream and the empty guest stream are checked as before.

Re-run on WSL2 Ubuntu 24.04 (kernel 6.6, GCC 13.3, Release) at fff18d2: guest_system_load_exec_standard_error passes 3/3. check_conventions.py passes.

AI-assisted: yes (Claude Code); results above were run on this machine.

[MarcZX10]

Re-reviewed exact head fff18d2f3b12e1654d266bb84f2b8337b5fea0b8. The test now skips only the leading host diagnostic with the emitted [memory] write watch unavailable: prefix; it requires a newline and still compares all remaining host output exactly, while requiring an empty guest stream. This addresses the earlier concern that arbitrary leading output could pass. The author reports the focused Linux test passes 3/3 on the new head and still fails if either host-stderr duplication is removed; Windows related tests and conventions also pass. I did not rerun those local tests. Hosted policy/convention checks pass; product jobs are path-skipped. No blocking issue remains.