keysources deps libfreemkv, now git-only, so keysources follows it off
crates.io. A committed [patch.crates-io] redirects the bare libfreemkv version
req to the git tag (local dev overrides to a path via the gitignored config).
The 'KeydbSource owns save+update' (Added) and 'Processing-Key
decryption restored' (Fixed) entries were stale carryover from the
never-released 1.1.0-beta.1; that code shipped in v1.1.0 (zero diff in
v1.1.0..HEAD). The 1.2.0 entry now lists only the real 1.2.0 changes
(hex unification, DiscInputs version stride, MKB cap).
The online source dropped any MKB over 10 MiB while libfreemkv's reader
captures up to 64 MiB — an MKB in that band was silently un-forwardable
(no key, no surfaced cause). Match the cap and log when it is exceeded.
unit_keys_from now loads all stored unit keys AND the VUK/MK-derived ones, deduping by key value instead of first-hit, so a disc with partial stored UKs plus a VUK yields every declared CPS unit key. Plus keydb.cfg format round-trip (to_keydb_cfg inverse of parse).
Found via a parse->serialize->parse->serialize byte-identical check over the full
182,535-disc OEM keydb:
- Title was reduced to a parenthesised substring, truncating real titles
("Lawrence of Arabia (Restored Version) - Disc 2 - 4K Ultra HD" -> "Restored
Version"). Keep the title VERBATIM (faithful copy); display prettification is a
separate layer.
- A disc titled exactly "M" (`= M | M | 0x...`) had its title eaten as the M
field tag, dropping the real media key. Field scan now skips parts[0] (always
the title), so a title that is a tag letter (M/I/V/U/D) can't shadow a field.
- Add to_keydb_cfg_is_idempotent_on_real_keydb (KEYDB_PATH-gated): proves parse
is lossless on its own output and the serializer is deterministic.
Central round-trippable codec: parse + to_keydb_cfg in one place.
Emits HC, DK, PK, then disc entries sorted by hash. Round-trip test
(parse -> serialize -> re-parse) covers HC/DK/PK/disc. disc_id -> vid
names the field for what it is (the I-field volume ID).
Make the parser public — parsing a keydb is not secret (freemkv uses it, and so
do tools that build a per-disc registry from it, e.g. a Volume-ID index). Purely
additive: a private module in this public crate becomes nameable; no behaviour
change, nothing removed.
Move the keydb save/validation/atomic-write path out of libfreemkv onto
KeydbSource. New KeydbSource::save(bytes) validates + decompresses (zip /
gz / plain, decompressed-size capped) and crash-safely writes to the
source's OWN path (sibling-temp + fsync + rename + parent-dir fsync) —
not a hardcoded default, so the caller chooses the destination.
KeydbSource::update(fetch, url) calls an INJECTED fetch closure then
save, keeping this crate transport-agnostic on the update path (the app
supplies its own TLS / SSRF-guarded transport). UpdateResult moves here
and is re-exported. Add flate2 + zip (already in the resolved graph via
libfreemkv) for decompression; no new HTTP stack.
unit_keys_from now resolves the Media Key in order: stored per-disc MK
-> keydb Processing Key pool (mk_from_pk vs this disc's own MKB) ->
device-key pool (mk_from_dk), then MK+VID -> VUK -> UK. MK/VUK entries
still honored directly; cross-disc MK-pool brute stays retired. Fixes the
factually-wrong justifying comment + adds PK-pool KATs.
The crate previously had only leak-guard.yml and release.yml, so its ~42
tests never ran on push/PR — only on tag. Mirror the sibling crates'
ci.yml (lint + test jobs, dtolnay/rust-toolchain@1.86.0, rust-cache).
No --locked: this library does not track Cargo.lock (release.yml notes
--locked would fail on a fresh runner).
Add tests/key_sources.rs exercising the published KeySource impls
end-to-end over real fixture files and the libfreemkv parsers:
- KeydbSource: disc-hash lookup from a real keydb.cfg (VUK hit ranks
ahead of the universal DK pool), hash miss yields only the pool,
missing file is silent/not errored, label + needs_samples, and
host-cert serving from a | HC | row (inherent and trait paths).
- paths: exe-local search list, default == search head, existing path
reflects on-disk state (local-only, no OS fallback).
- MapfileSource: persisted # freemkv-uk: keys read back as a terminal
Key::Unit, one-shot exhaustion, missing/keyless mapfile offers nothing.
- OnlineSource: unconfigured no-op (no network), one-shot latch,
metadata, and validate_keyserver_url scheme/SSRF gating.
- MultiSource: caller-supplied order/precedence (and its reverse),
empty-source skip, needs_samples/errored OR-aggregation, nesting, and
a real keydb-then-mapfile precedence chain over fixtures.
Move the encrypted sample reader and the key-resolution loop into libfreemkv
(they read the disc and validate keys — decryption mechanism, not lookup). A
key source now only looks a key up and hands it back.
A keydb can hand out a per-disc terminal Key::Unit (a UK entry keyed on
disc_hash). Unlike a derived key (Device/Processing/Media/Volume), a
terminal UK is applied as-is by Disc::decrypt_with: it is NOT re-derived
through the MKB-verified AACS resolver, so a UK entry whose hash matches
the disc but whose key bytes are wrong is only disproved by descrambling
real ciphertext.
KeydbSource inherited the default needs_samples() == false, so on the
autorip auto-resume / mux-worker path (which samples units only when some
source reports needs_samples()) a keydb-only resolve ran with empty
samples and committed a wrong UK as success, muxing undecryptable video
while reporting done. The CLI was unaffected because it always samples.
Override needs_samples() to true on KeydbSource so every consumer samples
encrypted units before resolving and a wrong keydb UK is rejected on all
paths. Regression test asserts the override.
- KeydbSource implements KeySource::host_certs(), delegating to the
inherent host_certs() — surfaces the | HC | / | HC2 | certs already
parsed from keydb.cfg by libfreemkv's parser, so the OEM cert route
collects them across the keysource layer. No new parsing.
- OnlineSource::host_certs() is a no-op stub: returns empty with zero
network access (no client fetch, no server endpoint). Online host-cert
serving is deferred. TODO(owner) marker left in place.
Tests: trait host_certs returns the keydb HC row; empty when keydb
missing; online host_certs is an empty no-op without network.
OnlineSource::query() now includes the disc's volume_label (UDF/ISO volume id)
as a plain-text `title` field in the /decode POST, so the key service can build
a disc_hash → title catalog from real rips. Depends on libfreemkv 0.30.
Each source implements next_key (a cursor over its candidates) instead of
returning them all at once. The keydb hands its per-disc candidates out
UK-first (UK > VK > MK > DK) so a stale/wrong VUK never pre-empts a good UK in
the same entry; online and mapfile are one-shot. MultiSource composes sources
in the caller's chosen order and resolve_and_apply drives the
next_key -> decrypt_with loop, stopping at the first key that decrypts.
read_sample_units moves here so the CLI and autorip share one content sampler.
Matthew Jackson
matthew