Johnathan Corgan
f89447d7ee
Every link peer's rekey now runs in its own rekey machine. The machine holds the link's current, pending and previous sessions with their indices and key-epoch bits, the drain, the peer's startup epoch, the dampening, the link's age and the record of answered msg1s, and decides every rekey step; the link handlers carry out what it decides. The deployed link rules are gone, and tests fail if one of them comes back. Nothing on the wire changes. What a peer or an operator can see: - Retire a still-draining session before a link cutover demotes the current one. - End an unanswered link rekey at 30 s and back off before the next. - Recognise link msg1 copies by their ephemeral, keep refusals and the record across re-peering, and hold own rekeys past 31 s. - Prompt the peer at a link rekey cutover and take the new key-epoch bit from the peer's. - Keep a link's previous session until the peer is seen to move, and count its frames for liveness. - Reset what a link holds for a peer when its rekey reveals the peer restarted. - Treat a dial whose msg2 shows a new startup epoch as the peer's restart. - Close the link rekey machine at a cross-connection move and build a new one on the winning session. - Send the bloom filter again after every link rekey this node answers. - Move to the later dial's session when a peer answers a second dial of ours. The core's slow test suites, the models of the deployed releases and the guard-switch meta-test, compile only with fips-core's slow-tests feature. GitHub CI's Linux job runs them on every push and lints them alongside the core's other clippy passes; local CI runs them with --slow-tests. Every other platform's test run skips them, and fips-core is built at opt-level 2 for dev and test builds, which brings the core's ordinary test run from about 210 s to 30 s. A link setup test backdates a pending by 150 s rather than an hour, since an Instant on Windows cannot go back past the host's boot.
2.6 KiB
Migrate an Embedder to v0.6
For code that uses the fips library crate (published as fips-mesh)
rather than running the fips daemon. Each section names one change in
v0.6.0 that can stop such code from building, what to do about it, and
the CHANGELOG entry that records it. Changes that touch only the daemon,
its configuration or the wire are in the CHANGELOG and the release
notes, not here.
ActivePeer no longer exposes its session slots or its rekey state
What changed. The methods of fips::peer::ActivePeer that read or
change its Noise sessions, their indices, the K-bit epoch, the session
time base, the stored link-setup msg2, the peer's startup epoch and the
rekey cycle are no longer public. These are now private to the crate:
with_session, has_session, noise_session, our_index,
their_index, session_elapsed_ms, set_handshake_msg2,
handshake_msg2, remote_epoch, session_established_at,
link_established_at, rekey_jitter_secs, current_k_bit,
rekey_in_progress, pending_our_index, pending_their_index,
previous_our_index, previous_session, pending_new_session,
is_draining and rekey_our_index. These are gone, because the rekey
state machine now holds the state they reached: noise_session_mut,
replace_session, session_start, is_rekey_dampened,
record_peer_rekey, previous_session_mut, pending_new_session_mut,
set_pending_session, cutover_to_new_session, handle_peer_kbit_flip,
drain_expired, complete_drain, abandon_rekey, set_rekey_state,
awaits_msg2, complete_rekey_msg2, needs_msg1_resend, rekey_msg1
and rekey_msg1_resend_count. Four methods nothing called are also
gone: set_their_index, clear_handshake_msg2,
set_rekey_in_progress and set_msg1_next_resend.
Why. The node drives these through its own handshake and rekey handlers, and calling them from outside could leave a peer's sessions out of step with the node's index tables. Keeping them private lets the rekey code change in later releases without breaking the library's interface again.
What to do. Code that calls none of the methods above needs no
change: the identity, addressing, tree position, filter, link statistics,
MMP and heartbeat accessors are unchanged. For the session and
rekey figures a monitoring tool shows, read the node's show_peers
control query, which reports for each peer its session index, its
session's send and receive counters, the K-bit, and whether a rekey or a
drain is in progress. No replacement is offered for driving a rekey from
outside the node.
Recorded in the CHANGELOG's v0.6.0 Removed section: the entry on
ActivePeer's session-slot and rekey methods.