Buffer locking is special (hints, checksums, AIO writes) #160
Description
--- As of 2026-02-09T22:19:06.000Z ---
This thread discusses improvements to PostgreSQL's buffer locking mechanism, particularly around hint bits and checksums. Andres Freund has been working on patches to optimize buffer hint bit operations, introducing BufferSetHintBits16() for single 16-bit value updates. Heikki Linnakangas provided detailed code review feedback on function documentation, comment clarity, and implementation details. The discussion covers technical aspects like LSN handling, race conditions in buffer writes, and proper locking semantics. Antonin Houska identified an issue with HeapTupleSatisfiesMVCCBatch() during logical decoding with historic snapshots, proposing a fix. Andres suggested an alternative solution restricting page-at-a-time scans to real MVCC snapshots. Kirill Reshke inquired about creating TAP tests for rare recovery scenarios involving torn pages. The patches appear to be progressing toward completion with most review feedback addressed.
这个讨论串涉及PostgreSQL缓冲区锁定机制的改进，特别是关于提示位和校验和的部分。Andres Freund一直在开发补丁来优化缓冲区提示位操作，引入了BufferSetHintBits16()函数用于单个16位值更新。Heikki Linnakangas对函数文档、注释清晰度和实现细节提供了详细的代码审查反馈。讨论涵盖了LSN处理、缓冲区写入竞争条件和适当锁定语义等技术方面。Antonin Houska发现了HeapTupleSatisfiesMVCCBatch()在逻辑解码期间使用历史快照时的问题，并提出了修复方案。Andres建议了一个替代解决方案，将页面批量扫描限制为真正的MVCC快照。Kirill Reshke询问了为涉及页面撕裂的罕见恢复场景创建TAP测试的问题。补丁似乎正在朝着完成方向发展，大部分审查反馈已得到解决。
participants - 参与者: