mirror of
https://github.com/jmcorgan/fips.git
synced 2026-10-05 11:08:25 +00:00
Johnathan Corgan
cd1957c894
The FreeBSD package job had stalled seven consecutive runs, the last two killed by the thirty-minute bound and the five before it burning up to six hours each. The job builds and gets into the unit tests; libtest names the culprit itself in the run log, in a line the earlier reading of that log missed: test native::client::tests::an_empty_datagram_is_not_reported_as_a_closed_flow has been running for over 60 seconds That test sends a zero-length datagram and then reads it back with an unbounded blocking recv. FreeBSD accepts AF_UNIX SOCK_SEQPACKET and returns a socket that is not an atomic-record socket: seqpacketproto carries no PR_ATOMIC and shares its send and receive handlers with SOCK_STREAM. A zero-length send with no control data therefore queues nothing, wakes nobody, and returns success, so the recv waits for a message that was never delivered. Nothing bounds it: plain cargo test has no per-test deadline, so one blocked thread holds the whole binary open. The same kernel fact accounts for the five boundary tests that failed within a third of a second in the same run, which needed no separate explanation. macOS already takes SOCK_DGRAM here because it has no AF_UNIX SOCK_SEQPACKET at all. FreeBSD needs the same substitution for a different reason, so it joins that arm, along with the close-detection retry the arm carries. The cfg is written by exclusion so that a new unix target gets the Linux arm, which fails loudly by refusing the socketpair rather than quietly losing messages. Two things guard against a repeat rather than fixing this instance. Every blocking read in these tests now carries a deadline, so a platform that swallows a message fails with an assertion naming the flow instead of wedging the suite. And the FreeBSD job runs its tests under a wall-clock bound, so a hang is a named step failure in fifteen minutes rather than a job cancelled at the ceiling, with the output kept up to the kill because that output is what names the blocked test. What this does not establish, stated plainly because the code comments would otherwise imply otherwise: nobody ran any of this on FreeBSD. The mechanism is read out of the FreeBSD kernel source and inferred from a CI log whose shape matches it. The freebsd-gated probes added alongside are what would turn that into a measurement, and the next run of the job is the first real test.