model-writtenrun 9648 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Chris Travers proposed replacing direct interrupt-flag checks with an ordered interrupt-level enum and PENDING_INTERRUPT_LEVEL() macro in regex compilation and dynamic shared memory allocation #1.
- Michael Paquier objected that the hierarchy obscured the actual conditions and constrained future interrupt types, prompting Chris to reconsider the abstraction #2 #5 #6.
- Discussion explored combined bit flags and a struct of flags, but Chris Travers and Tom Lane explained the lost-update risk of read-modify-write operations in signal handlers, while Andres Freund warned that atomics could use interrupt-unsafe locks #7 #9 #13.
- Michael Paquier isolated a smaller standards-correctness improvement: converting five pending-interrupt flags from bool to sig_atomic_t while retaining separate variables #14 #15 #16.
- Tom Lane identified the missing PGDLLIMPORT declaration on ClientConnectionLost, Andres Freund supported fixing it for extension portability, and Chris Travers endorsed the conversion #17 #19 #24.
- Michael Paquier committed the flag conversion from #16 with the additional PGDLLIMPORT annotation as ba16aade33 and announced completion #26.
- Chris Travers redirected the original effort toward documenting the existing signal-handling design and safe flag checks, submitting a README after concluding that further abstraction lacked support #25 #27 #28.
- Peter Eisentraut questioned the README's location and C89 framing and preferred code comments to reduce documentation drift; Dmitry Dolgov agreed, and Chris subsequently replaced the README with comments in globals.c #30 #33 #35.
- Andres Freund found those comments redundant or inaccurate because CHECK_FOR_INTERRUPTS also depends on interrupt holdoffs and Windows behavior, but supported an overview README because no single source file naturally housed the topic #36 #38.
- Chris Travers restored the earlier README under src/backend/utils/misc/README.SIGNAL_HANDLING, changing its location but retaining the original document text #40.
- Andres Freund challenged the README's standards claims, signal semantics, omission of errno preservation, and overly broad cleanup restrictions; Chris's next revision adjusted the flag list and interrupt-handling guidance but left substantial issues unresolved #42 #43.
- Tom Lane reviewed the revised README in July, proposing a better location and README.signals name, correcting signal meanings, and explaining that transaction-abort cleanup makes CHECK_FOR_INTERRUPTS permissible in more places than the draft suggested #44.
- Craig Ringer raised related concerns about custom signal handlers in worker_spi and walsender, while Alvaro Herrera clarified that those code issues need not block the documentation #46 #47.
- At the thread's end, the sig_atomic_t conversion had landed, the original abstraction had been abandoned, and the documentation remained waiting on Chris Travers to address Tom Lane's review ba16aade33 #28 #45 #47.