model-writtenrun 6966 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Michael Banck opened the thread with two patches needed to build and test PostgreSQL on Debian GNU/Hurd: portable fsync descriptor assertions co-authored with Samuel Thibault and an IOV_MAX fallback co-authored with Christoph Berg, while noting high-resolution timer requirements and parallel-query testing limitations #1.
- Tom Lane committed the fsync fix as 3a2617e4f0 and the IOV_MAX fix as d25d392e89, each taking one component of #1, preserving its code while editing comments and commit messages, backpatching it, and requesting a Hurd buildfarm animal #2.
- Michael Banck pointed out that POSIX permits IOV_MAX to be omitted, and Tom acknowledged that his new parenthetical criticism of Hurd was inaccurate but questioned whether changing the comment was worthwhile #3 #4.
- After nearly three months, Alexander Lakhin reproduced fruitcrow's test_shm_mq hangs and supplied diagnostic logging showing processes stopping around poll and latch handling; Michael confirmed similar hangs on 32-bit Hurd #6 #7.
- Michael Paquier warned that frequent platform failures could obscure PostgreSQL bugs, while Tom Lane accepted continued buildfarm testing provided the failure rate remained low #8 #9.
- Michael reported isolation-test crashes and zero function-runtime measurements, and Alexander reproduced the zero-duration behavior with a small timing program; a GNU Mach timer fix did not eliminate it #10 #11 #13.
- Tom distinguished legitimate equal consecutive clock readings from backward clock movement and identified the stats test's empty functions as the problem, recommending a short sleep; Michael submitted that fix and initially reported over a thousand successful iterations #16 #17 #20.
- To establish realistic timer-resolution expectations across platforms, Tom proposed logging complete pg_test_timing output and implemented that discussion directly in 7ccbf6d8b5, without an attached patchset #19.
- Alexander also reproduced timestamp-sensitive failures in subscription.sql and pg_stat_statements entry_timestamp, establishing that the stats isolation patch addressed only one affected test #21 #23.
- After Michael reported ten thousand successful stats-test iterations, Tom reviewed timing data from 32 buildfarm animals and reduced the proposed sleep from 1ms to 10us; fa42213d4a landed #20 with that adjustment and explanatory formatting, and Michael subsequently confirmed thousands more successful iterations #25 #26 #27.
- Thomas Munro opened a separate investigation into invalid signal numbers on fruitcrow, and Alexander's instrumentation showed that NSIG was correctly defined but handlers sometimes received enormous invalid values #28 #29.
- Thomas proposed experiments around nested signal handling and filesystem operations, while Michael described the reduced concurrency and VM configuration used to stabilize fruitcrow; Alexander found that higher concurrency made the signal failure easier to reproduce and also observed incorrect query results #30 #31 #32.
- After fruitcrow reproduced the pg_stat_statements timing failure, Michael submitted a separate patch adding sleeps and more complex queries to increase execution and planning time, but acknowledged that it still failed occasionally and asked whether reducing timing coverage would be preferable #35 #36.
- Alexander's next experiments found that blocking all signals during handlers suppressed the invalid-signal assertion but left hangs, whereas replacing SIGURG with SIGXCPU did not help; additional corruption symptoms and the unresolved minimum timer-resolution question remained #37 #39.
- Thomas suggested a minimal signal-stress reproducer and floated making reset timestamps strictly increase, while Alexander emphasized that several failures appeared unrelated and Michael reproduced a glibc/Mach error with a standalone signal test #40 #42 #43.
- The discussion broadened to native asynchronous I/O when Samuel Thibault confirmed that Hurd exposes asynchronous Mach-port requests; Thomas outlined a possible PostgreSQL I/O method, and Samuel explained that vectored requests and O_DIRECT were not yet provided #46 #47 #48.
- Thomas offered to prototype a Hurd I/O method after stability problems were resolved and vectored operations were available, while also identifying the conservative IOV_MAX fallback and its inaccurate comment as unfinished work; this remained a design discussion without a submitted implementation #49 #50.
- By January, Alexander had catalogued 113 fruitcrow failures across hangs, timing failures, isolation crashes, and invalid signals, and announced that he would stop individually tracking them until a new Hurd release #51.
- In February, Michael reported that Samuel had fixed missing fork-related debugger information in glibc, enabling a useful backtrace that exposed a cancellation-lock deadlock during signal-handler writes; a second glibc fix then survived five million test_shm_mq iterations without hanging #53 #52.
- Samuel confirmed the cancellation critical-section fix and Tom welcomed the result, closing the thread with the two original portability fixes, timing-output logging, and stats isolation hardening landed, while the pg_stat_statements proposal and other timing, signal, and corruption reports had no demonstrated resolution in this thread #54 #55 #36 #37.