corebreaker / arbordb / 29

88%

main: 98%

Build Type

Pull #2

circleci

Committed by corebreaker

Commit Message

Address PR review: validate the scalar-write hint by structural epoch, reseal over buffered edits, and document the read memos

The `fetch_mut` scalar-write hint (an already-resolved a-node key that lets an in-place patch skip re-walking the directory tree) was trusted while it merely still named a live a-node; but an `AKey` is stable across renames, so a `mv` of the cursor's path elsewhere between opening the cursor and writing through it left the hint live-but-relocated and the patch silently landed on the moved entry (non-`permissions` build). The write-back buffer now carries a per-transaction structural-change counter that every `name → child` relink bumps (link/unlink, hence `store`/`mv`/`rm`/`cp`/`mkdir`); a `MutCursor` captures it at `fetch_mut` and the hint is honoured only while the counter is unchanged. A path resolves through those links alone, so an unchanged counter means the path still resolves to the hinted a-node; an interleaved relink drops the hint and re-resolves the path afresh, recreating the file if the path is now gone. A regression test relinks a hinted cursor's target and confirms the moved entry is left untouched while the path is recreated.

`reseal_integrity` read the entry straight from the engine, bypassing the write-back buffer, so an edit-then-chmod in one transaction sealed over the stale pre-edit bytes (benign today only because the commit-time flush re-seals over the correct bytes). It now reads buffer-first, so the seal covers the pending edited bytes, with no integrity re-check since it is recomputing exactly those tags. A regression test edits a file's value in place and then changes its ACL in one transaction, reopens, and reads the edit back.

Document that `ReadTxn`'s per-snapshot `enforced_paths`/`access_ok`/`verified` memos are intentionally uncapped — bounded by the snapshot's lifetime rather than an LRU capacity, and freed in full when the snapshot drops — so a read handle is not meant to be held open inde... (continued)

Pull Request Pull Request #2: Optimization

Coverage Stats

967 of 1017 new or added lines in 33 files covered. (95.08%)

18 existing lines in 7 files now uncovered.

8516 of 9731 relevant lines covered (87.51%)

372576.29 hits per line