[bpf-ci-bot] VM crash (exit code 2) silently skips all test result reporting

[kernel-patches-review-bot]

Summary

When vmtest exits non-zero (VM crash, kernel panic, timeout), the ERR trap in
run.sh fires immediately, skipping all post-processing: exitstatus collection,
kernel splat checks, JSON summaries, and the final results table. Developers see
only "Process completed with exit code 2" with zero diagnostic information.

Failure Details

• Test / Component: ci/run-vmtest/run.sh (CI infrastructure)
• Frequency: Observed in 5+ independent CI runs across multiple unrelated PRs
• Failure mode: VM crash / timeout causing silent exit code 2
• Affected architectures: x86_64, s390x (observed); potentially all
• CI runs observed:
  • https://github.com/kernel-patches/bpf/actions/runs/33405699443 ("Validate program attach type")
  • https://github.com/kernel-patches/bpf/actions/runs/33308375841 ("struct_ops trampoline v2")
  • https://github.com/kernel-patches/bpf/actions/runs/33391013382 ("allow terminal gotox")
  • https://github.com/kernel-patches/bpf/actions/runs/33285840714 ("[test] bpf-next_test")

Root Cause Analysis

run.sh line 4 sets trap 'exit 2' ERR, which causes immediate script
termination on any unhandled command failure. At line 115, vmtest is called
without the && true guard that is already used elsewhere in the same file
(line 72 for bpftool checks) and in run-bpf-selftests.sh (line 54 for
individual test execution).

When vmtest returns non-zero (because the VM crashed, the kernel panicked, or
the guest timed out), the ERR trap fires at line 115 and the script exits with
code 2 immediately. Lines 120-166 — which handle kernel splat reporting,
exitstatus file parsing, JSON summary generation, and the human-readable results
table — are never reached.

This means:

1. The exitstatus file is never read, so per-test pass/fail is lost
2. kernel_splats.log is never checked, so splat annotations are never emitted
3. test_*.json summaries are never printed to GITHUB_STEP_SUMMARY
4. The final "Test Results" table is never printed
5. The exit code is always 2 (ERR trap), not the actual test exit status

The && true pattern is the established idiom in this codebase for preventing
the ERR trap from firing while preserving the exit code in $?. The bpftool
check block (lines 68-79) demonstrates the exact same pattern with an
explanatory comment.

Proposed Fix

Add the && true guard to the vmtest invocation and capture its exit code.
If vmtest failed, emit a GitHub Actions warning annotation and write the failure
to the exitstatus file so the existing awk-based status collection correctly
reports the failure.

See: output/0001-ci-run-vmtest-preserve-test-results-on-VM-crash.patch

The fix is 6 lines of added code and follows existing patterns in the file
exactly. No behavioral change occurs when vmtest succeeds (exit code 0).

Impact

Without this fix, any VM crash, kernel panic, or guest timeout produces a
completely opaque "exit code 2" failure. Developers must manually re-run the
job or dig through raw logs to find what happened. With the fix, the existing
post-processing pipeline runs normally: per-test results are reported, kernel
splats are checked, and the JSON summary is written — even when the VM itself
had problems.

References

• ci/run-vmtest/run.sh:4 — ERR trap definition
• ci/run-vmtest/run.sh:72 — existing && true pattern (bpftool checks)
• ci/run-vmtest/run-bpf-selftests.sh:54 — existing && true pattern (test execution)
• ci/run-vmtest/run.sh:115 — bare vmtest call (the bug)

[kernel-patches-review-bot]

0001-ci-run-vmtest-preserve-test-results-on-VM-crash.patch

From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: BPF CI Bot <bot+bpf-ci@kernel.org>
Date: Sun, 31 Aug 2026 00:00:00 +0000
Subject: [PATCH] ci/run-vmtest: preserve test results on VM crash

When vmtest exits non-zero (VM crash, kernel panic, timeout), the ERR
trap at line 4 fires immediately and the script exits with code 2,
skipping all post-processing: exitstatus collection, kernel splat
checks, JSON summaries, and the human-readable results table.

Apply the same `&& true` guard already used for bpftool checks (line 72)
and individual test execution in run-bpf-selftests.sh (line 54). Capture
the vmtest exit code, emit a GitHub Actions warning annotation, and
write the failure to the exitstatus file so the existing awk-based
status collection reports it correctly.

No behavioral change when vmtest succeeds (exit code 0).

Generated-by: BPF CI Bot (claude-opus-4-6-v1) <bot+bpf-ci@kernel.org>
---
 run-vmtest/run.sh | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/run-vmtest/run.sh b/run-vmtest/run.sh
index 3fc30b8..f3a0f40 100755
--- a/run-vmtest/run.sh
+++ b/run-vmtest/run.sh
@@ -112,11 +112,17 @@ foldable start vmtest_toml "$VMTEST_TOML"
 cat $VMTEST_TOML
 foldable end vmtest_toml

-vmtest -c $VMTEST_TOML
+vmtest -c $VMTEST_TOML && true
+vmtest_rc=$?
 rm -f $VMTEST_TOML

 foldable end vmtest

+if [[ $vmtest_rc -ne 0 ]]; then
+  echo "::warning::vmtest exited with code $vmtest_rc — VM may have crashed or timed out"
+  echo "vmtest:${vmtest_rc}" >> exitstatus
+fi
+
 if grep -q '^kernel_splats:1$' exitstatus; then
   splat_error="kernel splat check failed"
   if [[ -s kernel_splats.log ]]; then
--
2.43.0