You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
macOS e2e: leaked gvproxy holds port 7654 and fails every later run #2456
The macOS e2e job (VM integration harnesses) can leave gvproxy running on the self-hosted runner. The next run then fails before any test logic:
minvmd::guest_classifier_integration guest_classifier_table_present
minvmd run --detach failed: Error: hostname proxy port 7654 held by pid 37193 (gvproxy); bind refused; configured ports never redraw (free the port or unset MINVMD_NODE_PROXY_PORT)
Seen 2026-10-20 on runner jks-Mac-mini. Every PR from epic #2236 (#2451–#2455) failed e2e / ci-macos-success until the process was killed by hand. Main was green at 07:33.
Fix (any or all):
Add if: always() post-steps that terminate the harness's gvproxy/minvmd children, tracked by pid file instead of by name, so concurrent jobs aren't hit.
Add a pre-step that fails fast with a clear message, or reclaims the port, when 7654 is held by a stale gvproxy owned by the runner user.
Or give each run its own MINVMD_NODE_PROXY_PORT so a leak can't block the next job.
Same leak, second process: after gvproxy pid 37193 was killed, the rerun (job 118705302024, 15:11Z) got past the classifier test and failed vm_escape_integration vm_escape_bounded_to_resident_union:
could not give the box an address: port 127.0.0.1:7656, which box names are answered on, is in use by pid 37192 (/Users/jk/actions-runner/_work/_temp/testbed/bin/minvmd)
The crashed run leaves the whole testbed behind (minvmd plus its gvproxy), not just gvproxy. The cleanup step has to kill the testbed process group under _work/_temp/testbed, and the pre-step has to check every port the harness binds (7654, 7656, …), not just one.
Activity
norrietaylor commentedon Oct 20, 2026
Same leak, second process: after
gvproxypid 37193 was killed, the rerun (job 118705302024, 15:11Z) got past the classifier test and failedvm_escape_integration vm_escape_bounded_to_resident_union:The crashed run leaves the whole testbed behind (
minvmdplus itsgvproxy), not just gvproxy. The cleanup step has to kill the testbed process group under_work/_temp/testbed, and the pre-step has to check every port the harness binds (7654, 7656, …), not just one.