Hi Klement, Dave, Could you please review/approve 46720 PS2? https://gerrit.fd.io/r/c/vpp/+/46720/2 CI is already Verified+1, and I found a recent real CI occurrence of the lock-poisoning cascade this patch addresses.
On 1 November, the CI for HTTP change 47386 PS1 failed on both Ubuntu 24.04 ARM64 and Ubuntu 22.04 ARM64. The traces show a worker stuck in an ip link del subprocess while holding the global multiprocessing lock. The runner terminates timed-out workers; subsequent workers then remain blocked in multiprocessing.synchronize.__enter__ -> delete_all_namespaces. I counted 15 distinct workers blocked there on Ubuntu 24.04 and 16 on Ubuntu 22.04, including TAP/AF_XDP, HTTP static and Prometheus setup. The lock-wait cascades persisted for more than an hour in each job. Ubuntu 24.04 ARM64: https://github.com/FDio/vpp/actions/runs/41346130867/job/124966682460 Ubuntu 22.04 ARM64: https://github.com/FDio/vpp/actions/runs/41346130867/job/124966682502 This is concrete evidence for preventing the secondary false failures: flock releases the lock when the Python owner dies, and host-network commands are bounded. It does not establish that the initial host-network fault would disappear; I have not replayed those historical CI jobs with the patch. I also exercised the exact PS2 helper on Linux: 23 targeted process/namespace cases pass, including SIGKILL/SIGTERM owner death, concurrent access, command timeouts, descriptor cleanup and a private network namespace lifecycle. The negative control against the parent reproduces the poisoned lock after owner termination. Thanks, Jerome
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#27324): https://lists.fd.io/g/vpp-dev/message/27324 Mute This Topic: https://lists.fd.io/mt/122074487/21656 Group Owner: [email protected] Unsubscribe: https://lists.fd.io/g/vpp-dev/leave/14379924/21656/631435203/xyzzy [[email protected]] -=-=-=-=-=-=-=-=-=-=-=-
