model-writtenrun 4334 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut opened the discussion by requesting self-contained smgr API documentation, explicit blocknum and nblocks semantics, and clearer distinctions between smgrextend() and smgrzeroextend() #1.
- Bharath Rupireddy distinguished caller-supplied data from internally generated zeroes and separated the API from md.c implementation details, while Andres Freund explained the allocation and syscall savings behind the zero-extension implementation #2 #3.
- Peter demonstrated that hash indexes rely on smgrextend() leaving zero-reading gaps, and Greg Stark agreed that the documented API permits this and that the existing descriptions obscure the distinction between the functions #4 #5.
- Thomas Munro proposed using zero-extension for database copying; Andres warned about sparse-file allocation failures and reported little benefit in his Linux filesystem experiments, with possible double writes on systems lacking fallocate support, leaving that proposal unresolved #6 #7 #8.
- Peter submitted a combined cleanup that reindented comments, separated API documentation from implementation details, reordered functions, and clarified some arguments, while retaining a FIXME about the relationship between blocknum and nblocks #9.
- Andres requested a separate formatting commit, preferred keeping extension functions before reads and writes, and clarified that fallocate reserves actual storage while the starting-block behavior matches smgrextend() #10.
- Peter extracted and broadened the comment-indentation cleanup from #9 into 0b8ace8d77, then separately committed its mdwriteback() relocation as 4c9deebd37.
- Peter posted the remaining comment changes without the broader function reordering, leaving the documentation work unfinished and requesting explicit contracts for gaps, overlapping existing blocks, negative block counts, and consideration of copy-on-write filesystems #11.