model-writtenrun 7483 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Craig Ringer proposes identifying individual LWLocks through lock pointers and numeric tranche IDs, adding acquisition and variable-wait probes, removing a false acquisition event, and documenting tranche allocation; he explicitly acknowledges that DSM mappings prevent universal cross-backend pointer comparisons #1.
- Craig demonstrates the proposed instrumentation with a SystemTap script reporting lock hold and wait durations and identifying blockers, while Dmitry Dolgov questions the tranche-initialization assertion and tracing overhead #4 #5.
- Craig explains the assertion and lock-order analysis use case, and identifies unnecessary tranche-name computation as a separate optimization opportunity #7.
- Peter Eisentraut verifies that LWLockWaitForVar should not emit an acquire event and commits that component of the original series unchanged as 09418bed67, leaving the broader instrumentation pending #6 #9.
- After a rebase request, Craig resubmits the three remaining components and adds probes for backend type, postmaster PID and application-name changes; the original remaining components retain their implementation #10 #11.
- Peter questions the release-probe relocation, Craig explains that firing after wakeup incorrectly extends the apparent lock-holding interval, and Peter confirms that the documentation requires firing before waiters awaken #12 #13 #14.
- Peter extracts that relocation from the revised series as 96ae658e62, preserving the existing probe arguments and leaving the expanded interfaces pending #11 #17.
- Peter finds that the new uint64 probe arguments break macOS builds, requests documentation and a more consistent argument order, and asks whether application-name tracing should become generic GUC tracing; Craig promises fixes while defending the utility of a dedicated application-name probe #18 #21.
- Peter establishes that the build already supports probe-enabled semaphore checks, and Craig agrees to use them to avoid evaluating unused probe arguments #15 #19 #20.
- Andres Freund objects to runtime and maintenance costs from widespread static probes, corrects Craig's understanding of tranche-ID consistency across backends, and suggests shared tranche-name storage; Craig becomes willing to move name resolution into tracing tools #22 #24 #25.
- Craig supplies disassembly showing that simple inactive probes become no-ops but computed arguments still incur work, and that semaphore guards avoid GetLWTrancheName calls at the cost of branches and larger functions #26 #27.
- Robert Haas explains why the former array-offset lock identity scheme could not represent multiple tranche instances, while Andres discusses compact wait-state structures and dynamic probing; Craig proposes retaining names only on wait probes and adding tranche-registration events, but submits no revised implementation of that design here #28 #29 #31.
- Peter defers the broader feature work to PostgreSQL 15 while requesting a separate guard fix for PostgreSQL 14 and 13; Robert offers to review tranche documentation, and Craig later plans to post that work separately #33 #35 #38.
- Peter posts the standalone twelve-guard patch, Craig verifies its application, builds and behavior on master and REL_13_STABLE, and Peter commits it unchanged as b94409a02f #36 #37 #40.
- Andres reports empty-body compiler warnings in builds without tracing, and Craig responds with a patch making disabled probe macros expand to do-while dummy statements instead of disappearing entirely #41 #43.
- Tom Lane confirms the warnings across buildfarm members and Andres identifies the compiler options involved; Peter commits Craig's dummy-statement fix as fa8fbadb93, adding the corresponding Perl-generator change #45 #46 #48.
- The closing discussion resolves buildfarm compiler-upgrade problems, with Andres reporting all clear; the four focused fixes have landed, while the broader lock-identification, additional-probe and tranche-documentation work remains deferred without another patchset in this thread #49 #52 #33 #38.