mirror of
https://github.com/xcat2/xcat-core.git
synced 2026-09-05 04:27:55 +00:00
Daniel Hilst
ce709cbeb0
supervise() unblocks SIGCHLD and then hands the pid back for the caller to assign, so the
caller's ($mon_respawn, $pid_MON) = ... still lands after the signal is let back in -- the
exact window e0b0ac6 closed, reopened by moving the sequence into a function. A monitor that
dies in that gap is compared by ssl_reaper against a $pid_MON still holding 0, missed, and
the caller then writes the dead pid back over the reaper's work: !$pid_MON never fires again
and the monitor is never respawned. That is this PR's own failure, reached through the
respawn rather than through startup, and nothing in the file notices it.
Add a subtest that watches the window from inside. Racing a real death into it is not
something a test can arrange reliably, so it arranges a certainty instead: a decoy child is
forked and exits with SIGCHLD blocked, leaving the signal pending, so the handler is
guaranteed to run the moment supervise() unblocks -- inside supervise(), before it has
returned. What the handler sees there is what ssl_reaper would see: it must find the live pid
and a pacing state that already knows a child was forked.
Separately, the fork-and-port subtest's closing assertion could not fail. It asks that the
respawn after a healthy monitor's death lands within 2 seconds, with max_interval set to 2 --
so a backoff pinned at the ceiling satisfies it too, and deleting the healthy-run reset from
exited() leaves the subtest green. Raise the ceiling to 8, where only the reset can produce a
prompt respawn, and assert first that the monitor being killed had actually been up long
enough to count as having served, which is the premise the assertion rests on.
Verified: against the current supervise() the new subtest fails on both assertions ("got 0,
expected <pid>"), and the raised ceiling turns the closing assertion red (8 <= 2) when the
healthy branch of exited() is removed -- where at a ceiling of 2 it stayed green.
Signed-off-by: Daniel Hilst <392820+dhilst@users.noreply.github.com>