model-writtenrun 2255 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andrew Dunstan reopened the long-disabled SSL log checks in #1; Tom Lane suggested borrowing synchronization ideas in #2, while Andrew stressed that the problem concerned server log contents in #3.
- Tom's first patch #4 restored the assertions and waited for a DEBUG2 backend-exit message before inspecting logs, updating other vulnerable callers and replacing obsolete fixed certificate serial numbers; Jacob Champion endorsed the serial-number change in #5.
- Testing revealed that an older backend's delayed exit could release the wait prematurely in #6, so Tom revised the expression to correlate backend creation and exit by PID in #7.
- Andrew questioned the fragility of the logging prerequisites in #8; Tom retained them, explaining in #9 that missing DEBUG2 logging would cause a timeout and that identifying the correct failed backend was the harder problem.
- The revised patch #7 landed unchanged as e0f373ee42; Tom subsequently explained in #10 that OpenSSL could notify the client before PostgreSQL logged the failure, creating the observed race.
- Thomas Munro reported that the restored intermediate-certificate assertion failed under LibreSSL on OpenBSD in #11, and Tom reproduced it alongside previously hidden RSA-PSS SCRAM failures in #12.
- Tom traced the RSA-PSS failure to LibreSSL's signature-algorithm handling in #13 and questioned the old SSL context API in #15; Daniel Gustafsson supplied a TLS_method() cleanup in #17, but Tom established that it was only an alias change in #18, and Daniel described a separate libtls backend as a different project in #19.
- After LibreSSL developers confirmed incomplete RSA-PSS support, Tom proposed skipping that test in #20; Daniel pointed to existing LibreSSL-specific exclusions in #21, while Tom's further certificate-callback investigation remained inconclusive in #22.
- Tom submitted two test workarounds in #23: skip RSA-PSS on LibreSSL and accept its reported certificate for the intermediate-chain failure; Daniel approved them and suggested centralizing library detection and citing the upstream discussion in #24.
- The API discussion considered expanding ssl_library(), distinguishing backends from implementations, and returning a structured result in #26, #27, and #28; Tom instead proposed dedicated detection methods in #29 to avoid extra complexity and unnecessary detection work.
- Daniel reviewed the dedicated helper in #30, raising future compatibility concerns, and Tom explained in #31 that unrelated backends could return false and that the immediate goal was keeping library checks coordinated.
- The helper additions from #29 landed as 1ddb9e14ea, with an extra compatibility comment but without the attachment's conversion of the existing test caller.
- The certificate-diagnostic portion of #23 landed separately as 75d73331d0, simplifying duplicated test calls into a conditional expected pattern.
- A missing v15 configuration prerequisite required 00811a96ac, which added SSL_CTX_set_cert_cb detection without a matching attached patch.
- The RSA-PSS exclusion from #23 landed as 793aa989f8, using the centralized LibreSSL helper; this workaround and the helper were backpatched through v15.
- Tom reported the work pushed and the SSL tests passing on OpenBSD 7.7 in #32, and Daniel acknowledged completion in #33; the supplied commits resolve the test failures, while the TLS_method() cleanup has no demonstrated landing here and future library-version detection remains an idea from #29.