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 51616e6
Browse filesBrowse the repository at this point in the historyBrowse files
R7a — fixed, and the reviewer was right: the flagged string at 850554f
had ~30 literal spaces with no backslash (rustfmt had collapsed the
original continuation). Now a single space.
R8 — guest-init can no longer panic or hang silently:
- an EXIT trap (log rc, sync, poweroff -f) makes every PID 1 exit path —
including any set -e failure before the child launch — end the VMI
instead of kernel-panicking or hanging ready forever;
- the two FATAL paths (config disk will not mount, no payload) return 1
into the trap instead of exec sleep infinity.
Live-verified both ways: a secret with the userData key renamed boots,
the init logs the failure, the VMI goes Succeeded and the sandbox goes
Error; a payload with a shell syntax error fires the trap directly
("init exiting (rc=2); powering off"), same terminal outcome.
R9 — start no longer restarts a healthy booting VM:
kubevirt_start_plan now decides from the run strategy, the VM's ready
flag and printable status, AND the paired VMI's phase. A live or booting
VMI (Pending/Scheduling/Scheduled/Running/WaitingForSync) makes start
idempotent (AlreadyRunning — the gateway admits retries from Starting,
and a retry must not tear down a healthy boot); unknown phases are
conservatively treated as live. The bounce fires only for a terminal VMI
(Succeeded/Failed) or a gone VMI under Once with printableStatus Stopped.
Unit tests cover all three plans and every VMI phase.
Live-verified: two back-to-back starts during boot leave the VMI uid
unchanged (idempotent); the dead-VMI bounce race (start fired the moment
the VMI goes terminal, before the sticky Error snapshot lands) was
attempted and honestly reported as missed — the Error snapshot won, the
gateway rejected the start, and the demo prints that outcome instead of
faking a pass.
R10 — the demo's start NVIDIA#1 now fails on rejection (routed through the
same classifier as start NVIDIA#2) and asserts the VMI uid is non-empty before
any comparison, so a rejected start NVIDIA#1 can no longer fake a bounce proof.
README: the start-semantics section and demo output updated to match
(idempotence verified live; the bounce documented as the narrow
pre-Error race it is).
Signed-off-by: Jesse Jaggars <[email protected]>
0 commit comments