mirror of
https://github.com/rustfs/rustfs.git
synced 2026-10-11 23:32:48 +00:00
唐小鸭
f3965512c1
rio no longer depends on `md-5` for production code. `EtagReader` and the `x-amz-checksum-md5` shell `Md5Hasher` both go through `rustfs_utils::hash::Md5Stream`, so the ETag path, the Content-MD5 check and the additional-checksum path can never disagree about which MD5 implementation is in use. `EtagReader` keeps the hasher as `Option<Md5Stream>` and consumes it exactly once on the poll that observes EOF; `get_etag` becomes `&self -> Option<String>` and is `None` before EOF (no partial digest is ever handed out). The tree has no pre-EOF caller: `try_resolve_etag` only asks after `finished`. `BadDigest`, the comparison point and the hex encoding are unchanged. `Md5Hasher::finalize(&mut self)` is now `finalize_reset`, which `HashReader` calls exactly once at EOF. `md-5` moves to dev-dependencies so test expectations still come from an implementation independent of the one under test. Tests: EtagReader with a tiny-capacity inner reader, reads after EOF, a transient inner `Err` that must not count as EOF nor corrupt the digest, `Pending` interleaved between chunks, and the calculated digest staying observable after a mismatch; `Md5Hasher` finalize/restart/reset contract through both the concrete type and `ChecksumType::MD5.hasher()`; Content-MD5 match and typed `BadDigest` mismatch through the full `HashReader` stack; MD5 added to the rio/checksums shared known-answer vector table.