mirror of
https://github.com/PerpetualSoftware/pad.git
synced 2026-09-20 17:43:26 +00:00
xarmian
de4d28d576
* feat(attachments): AttachmentStore interface + FSStore (TASK-870) Introduces the storage backend abstraction described in DOC-865 and ships its first concrete implementation. No call sites yet — TASK-871 (upload API) wires it in. internal/attachments/store.go AttachmentStore interface (Put/Get/Stat/Delete) and ErrNotFound sentinel. Put is documented as idempotent — concurrent Puts of the same hash converge — and required to verify that the streamed bytes actually hash to the supplied value. internal/attachments/registry.go Registry routes "<prefix>:<rest>" keys to the store registered for that prefix (Phase 1 = "fs"; Phase 2 will register "s3" alongside). Convenience Get/Stat/Delete helpers resolve + forward in one call so callers don't have to spell out the two-step pattern everywhere. Register panics if the prefix contains ':' since that would make the store unreachable. internal/attachments/fs_store.go FSStore writes to <baseDir>/<aa>/<bb>/<full-hash> with the first 4 hex chars sharding the directory tree two levels deep. Atomic writes: stream + hash to a randomized .tmp in the destination dir, fsync, then intra-directory rename. The streaming sha256 is verified against the supplied hash before the rename, so a mismatch never leaves a visible file. Idempotent fast path: if the canonical file already exists Put short-circuits (and drains the reader so callers don't get a stuck stream). Get returns wrapped ErrNotFound on missing keys; Delete on a missing key is a no-op (matches what the orphan GC needs). Tests cover put/get/stat/delete, hash mismatch, invalid hash format, idempotency, 16-goroutine concurrent Put of the same hash converging to one on-disk file with no orphan tmp files, registry routing, forward-error semantics, and the prefix-with-colon panic. Parent: PLAN-866. * fix(attachments): validate hash on every FSStore key + verify on fast path per Codex review (round 1) Round 1 raised two issues — both real, both fixed. 1. Path traversal in Get/Stat/Delete. extractHash only checked that the key began with "fs:" and the suffix was non-empty before passing it to pathFor(), which used the suffix as a path component. A key like "fs:../../etc/passwd" would escape baseDir for reads/stats/deletes. Fix: extractHash now requires the canonical 64-char lowercase-hex sha256 form via validHash. Same gate that Put already used; now it covers every public method. 2. Idempotent Put fast path skipped hash verification. If the canonical target file already existed, Put returned the key without checking that the supplied reader's bytes hashed to the supplied hash — violating the AttachmentStore.Put contract that implementations MUST verify on every call. A buggy upload path could associate the wrong bytes with an existing hash and silently succeed. Fix: stream r through a hasher when the target exists (no disk I/O), compare against the supplied hash, and reject on mismatch. Also dropped the dead "_short" branch in pathFor — every caller now goes through validHash. Tests added: - TestFSStore_GetStatDeleteRejectBadKeys covers empty/wrong-prefix/empty- hash/non-hex/wrong-length/path-traversal/path-separator/uppercase keys across all three read methods. - TestFSStore_PutFastPathStillVerifiesHash confirms the contract holds on the fast path: a second Put that lies about the hash is rejected with no corruption of the existing file.