model-writtenrun 7061 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul - Sep 2022 · 39 entries
- Thomas Munro opened the thread with six patches removing obsolete Unix fallbacks and proposing mandatory thread safety, using buildfarm configuration data to justify a POSIX.1-2001 baseline in #1.
- Tom Lane successfully smoke-tested the initial series on old macOS in #3, Greg Stark checked the portability concerns in #5, and Robert Haas supported the cleanups while questioning removal of pg_ prefixes in #6.
- The naming discussion favored standard names when replacements satisfy the standard's guarantees, with Thomas Munro, Peter Eisentraut and Tom Lane explaining that position in #7, #8 and #10, while Andres Freund raised broader replacement-function concerns in #11.
- The first landing took getrusage, sigwait and getpwuid_r cleanup from #1 as 86e5eb4f58, fb12becdfb and e757cdd6ad; getrusage also gained configure and ShowUsage cleanup.
- Thomas Munro carried the remaining work into a larger series in #13, adding obsolete configure probes and proposing a side-effect-free preadv/pwritev fallback.
- Tom Lane's review in #16 prompted investigation of Cygwin's WIN32 definitions, preservation of descriptive HAVE_* guards and Windows-specific implementation filenames; Thomas Munro documented the Cygwin findings in #17 and implemented the revised approach in #19.
- The revised design retained feature macros where Unix-only facilities were optional, distinguished full Unix symlinks from Windows junction support, and moved Windows-only implementations into clearly named files in #19; #20 additionally removed RTLD probes and redundant declarations.
- Thomas Munro explicitly transferred the fdatasync cleanup from thread #45231 into #19, adding the Windows-specific implementation filename; its subsequent version in #20 is later than that thread's source #511428.
- Andres Freund reviewed the final posted series in #24, and Tom Lane agreed that the old macOS buildfarm host should no longer block clock_gettime requirements in #25.
- The dlopen, getrlimit, shm_open and setsid components of #20 landed separately as ca1e85513e, bdb657edd6, 098f4d813b and adeef67834.
- The symlink/readlink, link, poll and environment components of #20 landed as 2b1f580ee2, 5963c9a154, b79ec732d2 and 71f5dc6dfb, with extra documentation and comment cleanup for symlinks.
- The scalar and vector positional-I/O components of #20 landed as cf112c1220 and a0dc827112, dropping unnecessary Windows function prefixes and unused includes while eliminating the lseek-based vector fallback.
- The clock_gettime and fdatasync components of #20 landed as 623cc67347 and d2e150831a, with the fdatasync library search retained; Thomas Munro reported that the batch was committed except for mandatory thread safety in #28.
- Andres Freund proposed removing obsolete strtoll/strtoull aliases in #23, Thomas Munro supplied overlooked ecpg and timezone removals in #33, and that revised patch landed with Andres credited as author in feb593506b.
- Andres Freund's gettimeofday cleanup in #35 removed one-argument detection and isolated the MSVC fallback; reviews in #37 and #38 led to filename-comment corrections before 922a8fa098.
- Robert Haas proposed simplifying process-group signaling in #30, which Thomas Munro implemented in #39; Tom Lane identified a race with children that had not yet called setsid in #40, prompting #44, but further concerns about other callers remained in #46.
- Thomas Munro's strtof patch in #45 landed unchanged as cbf4403134, retaining the workaround for buggy implementations; later MinGW testing in #56 and #57 showed that removing that workaround depended on runtime-linking details and was not part of this landing.
- Tom Lane and Thomas Munro identified further Windows gettimeofday simplifications in #41, #42 and #43; the attachment in #48 landed as 24c3ce8f1c with the null-tzp assertion and spelling correction suggested in #49.
- Andres Freund's next five-patch series in #50, reviewed by Thomas Munro in #51, produced the __func__, minidump, AIX, Solaris and header-prerequisite cleanups in 320f92b744, 9ddb870bd4, 8f12a4e7ad, 20c105c4db and a2a36d56b9.
- The hoverfly buildfarm failure exposed a mistaken AIX conditional simplification, which Andres Freund diagnosed in #58 and corrected directly in 7e29a79a46.
- Tom Lane reported obsolete timing comments in #59, approved Thomas Munro's inline rewrite from #60 in #61, and the correction landed as 01126dc8cf.
- Thomas Munro introduced a new system-header cleanup batch in #62, while noting that mandatory thread safety still needed work on LDAP library detection.
- Tom Lane initially wanted to preserve Unix-socket conditional compilation in #63, Andres Freund examined Windows testing implications in #64, and Peter Eisentraut explained why unsupported address families could fail at runtime and requested a separate macro-removal patch in #68.
- The sys/uio.h and sys/select.h components of #62 landed as 75357ab940 and 7e50b4e3c5, including the assumption that non-Windows systems define IOV_MAX.
- The IPC-header, resource-header and shl_load components of #62 landed as 37a65d1db1, 36b3d52459 and adee0df127, with small extra include and header-check cleanup in the resource-header commit.
- The sys/un.h portion of #62 landed separately as 077bf2f275, retaining HAVE_UNIX_SOCKETS temporarily so the disputed removal could be reviewed independently.
- John Naylor proposed raising the Bison minimum to match the remaining buildfarm in #67; the supplied material contains no matching candidate landing or subsequent resolution of that proposal.
- Tom Lane expanded his PPC assembler-probe removal from #71 into unconditional modern spinlock and atomic sequences in #72 after testing older hardware; Thomas Munro assessed the compatibility implications in #73, and the expanded patch landed as 8ded65682b.
- Thomas Munro resubmitted the Unix-socket remainder with sockaddr_storage and getaddrinfo cleanups in #75, then added obsolete gethostbyname_r probe removal in #76; Tom Lane withdrew his Unix-socket objection after closer inspection in #77.
- The networking series in #76 landed as f558088285, de42bc3ac8, 5579388d2d and 52ea29045b, with additional documentation corrections for the Unix-socket component.
- Windows experiments raised doubts about abstract Unix-domain sockets in #78 and #79, prompting the documentation patch in #80; Peter Eisentraut recalled earlier success in #82, but Andres Freund reproduced a current Windows client failure in #85.
- Thomas Munro's next networking batch in #81 removed obsolete Solaris interface enumeration and net/if.h probing in a717cddcac and 2f8d918359, with deleted IPv6 guards accounted for when the patches were reordered.
- The same batch from #81 corrected the Windows gai_strerror macro conflict in cce28f6268, removed the netinet/tcp.h probe with an added Windows header guard in 2492fe49dc, and enabled mstcpip.h for MinGW in f340f97a13.
- Thomas Munro retained and revised the IPv6 component in #90; after Andres Freund's approval in #93, it landed as bcc8b14ef6 with only a comment adjustment.
- The discussion of Windows signal-handler safety in #87, #88 and #89 led to WIP test-transport patches in #91, but securing temporary socket directories with owner-only Windows ACLs remained unresolved.
- Thomas Munro found that four sockaddr_storage member probes could be replaced by the existing sockaddr.sa_len test in #92, and that patch landed unchanged as 64ef572c06.
- Tom Lane proposed replacing obsolete individual-stream flushing in #95, Thomas Munro endorsed both that change and additional flushes before subprocess creation in #96, and the expanded implementation landed as 7fed801135.
- Thomas Munro listed mandatory thread safety, remaining vestigial feature macros and secure Windows Unix-socket tests as unfinished work in #94, then closed the large CommitFest entry and planned separate follow-up entries in #100.
- The abstract-socket documentation correction from #80 eventually landed in February 2023 as 117d2604c2, with the final wording specifying Linux only.
Jun 2025 · 4 entries
- Michael Banck revived the thread in June 2025 with a GNU Hurd build failure caused by undefined IOV_MAX in #101; Andres Freund questioned treating an untested platform as supported in #102, while Thomas Munro argued that POSIX permits omitting this limit macro in #104.
- Michael Banck separately traced intermittent stats and subscription failures to GNU Mach's 10ms timer resolution and posted added sleeps in #106, reporting more than ten successful test runs; Tom Lane discussed supported-platform timer precision in #107, but no supplied candidate commits that test patch.
- At the last supplied message, Thomas Munro agreed to restore the conservative IOV_MAX fallback and backpatch it, leaving larger Hurd I/O-combining limits for future work in #108.
- The IOV_MAX fix subsequently landed as d25d392e89, credited to Michael Banck and Christoph Berg and committed by Tom Lane, but its matching June 30 attachment #594324 belongs to thread #51831, superseding the earlier attribution to discussion in this thread.