model-writtenrun 7210 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Nathan Bossart opened the thread with two backpatchable stopgap fixes: restrict immediate SIGTERM-driven exit to the system() call and prevent forked children from running the parent's exit callbacks, while leaving replacement of system() for longer-term work #1.
- Thomas Munro supplied a reproducer and verified that the initial patch prevented the assertion failure, then clarified that his demonstration used fork() rather than vfork(); Nathan confirmed the result and added the requested diagnostic newline #2 #3 #4.
- Andres Freund proposed checking the PID stored in PGPROC to catch incorrect or duplicate slot releases, which Nathan incorporated, and advocated centralizing signal-safe diagnostic output #5 #7 #8.
- Nathan initially reused write_stderr(), but Andres identified its signal-safety problem and outlined a dedicated helper; Nathan then added write_stderr_signal_safe(), with both agreeing to defer formatting support to separate work #9 #10 #12 #14 #15.
- After several months without landing, Andres reviewed the pending series and requested production PANIC checks because duplicate process cleanup could cause severe damage; Nathan incorporated those checks, argument naming, and documentation fixes in the final submission #20 #21.
- Discussion of the PANIC paths exposed the broader unresolved problem of calling proc_exit() from signal handlers: Andres called for removing such calls, Michael Paquier questioned what remained, and Nathan identified startup and other handlers still requiring work #22 #23 #24 #25.
- Nathan landed the first component of the final series with only comment reflow as 8fb13dd6ab, then landed its unchanged second component as 97550c0711, confirming that both had been committed and backpatched #21 #26.
- Although thread #47353 attributes both changes to its February submission #528038, the matching October submission #21 makes this thread the latest patch source for both commits.
- The backpatches broke Windows builds on older branches because STDERR_FILENO was unavailable, so Nathan proposed extending an existing portability fix and Tom Lane approved that approach #27 #28.
- Nathan reused Andres Freund's Windows descriptor definitions in 555bc89c90 and confirmed completion in the last message; both stopgap patches and the Windows follow-up were landed, while broader signal-handler cleanup and diagnostic formatting remained outside this work #29 #14 #25.