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
Commit d3c3820
Browse filesBrowse the repository at this point in the historyBrowse files
smoke: one Xvfb per run, and the EGL libs a runner lacks
The worst CI run in the repo's history was one push with two host-side
wiring bugs and a latent third, all mine, none of them guest code:
- install, iso, dead-disk and boot all died at qemu's first line:
'Couldn't open libEGL.so.1' — the runner image never had EGL libs
because no job before this one asked qemu for a GL display. The kvm
action now installs libegl1 and libgl1 beside xvfb.
- xvfb-run ran qemu as a child, not exec, so the smoke's $! named the
wrapper shell and every kill below would have orphaned the VM —
port still bound, cleanup traps blind. Fixed by starting one Xvfb
for the whole run and exporting DISPLAY: qemu is the direct child
again, and the kill contract is what it was before.
- the shellcheck findings (unquoted $QEMU_XVFB, three sites) dissolve
with the wrapper they named.
Verified locally before push: shellcheck v0.11.0 clean (the pinned
version), actionlint clean at the digest the workflow pins, and the
underlying combination — Xvfb :77, DISPLAY, gtk,gl=on, virtio-gpu-gl —
is the one measured end to end on Oct 05, greeter alive indefinitely.
0 commit comments