Both ran scram verify_plain (~1s PBKDF2 at 1.2M iters) inline under the engine
lock, freezing the whole daemon per login. Add NetAction::DeferAuthenticate +
AuthThen continuation + ctx.defer_authenticate, mirroring DeferPassword: the
module/SASL path fetches the verifier cheaply and defers; the link layer runs
verify_plain on spawn_blocking, then Engine::complete_authenticate finishes the
login (IDENTIFY reuses the same ctx helpers, so its login/AJOIN/vhost/memo
side-effects are unchanged). Tests resolve the defer inline (cfg(test)). The
gRPC web-login path was already fixed; no login path stalls services now.
(GHOST/DROP/CERT/GROUP still verify inline but are rare account ops, not logins.)
The Authenticate RPC held the engine Mutex across scram verify_plain, which at
production iteration counts (1.2M) is ~0.8s of PBKDF2 on the async reactor —
freezing the whole daemon (IRC link, gossip, every RPC) for the duration and
serializing all logins. Fetch the verifier under the lock (cheap), then run
verify_plain on spawn_blocking, lock released. 6 concurrent logins: 823ms (was
~4460ms serialized). set_password/register already offload this way; drop the
now-unused authority_authenticate.
!deop with no target deops the caller, but peace_blocks compared the target's
access rank to the caller's — equal for self — so PEACE refused it with 'you
can't act against someone with equal or higher access.' Exempt the self case
(target uid == caller uid) in peace_blocks: acting on yourself isn't acting
against anyone, so you can always drop your own status. Covers DEOP/DEVOICE and
the kick/ban paths too.
The earlier StatsSet fix persisted only the shared namespaced counters. The
per-channel BOTSTATS view (lines seen + top talkers, from /statserv #chan) lived
on the LiveChannel struct, which is rebuilt from each burst — so it reset to zero
on restart. Move it to a name-keyed Network.chan_activity map (independent of live
membership, seedable before the burst) and carry it in the StatsSet snapshot
(new serde-default `channels` field) alongside the counters. Reworded the
notice from 'seen this session' now that it persists.
Problem 1: an assigned bot joined via IJOIN with no status mode, so it had no
+o. Join it with the 4-param IJOIN form (ts 1, mode o) so it's opped, matching
Anope's default bot modes (+o), and exempt bots from SECUREOPS/RESTRICTED so the
op is never stripped.
Problem 2: the shared stat counters lived only in memory, so a restart wiped
them. Snapshot them to the log (StatsSet event, folded into NetData + compaction)
periodically and on shutdown, and reseed the live registry from the log at
startup. Live gauges (online counts) stay re-derived, not stored.
An Anope import creates accounts (channel ownership, vhosts, certs) but can't
carry the one-way password hash, so they exist without a verifier. provision
then refused them as AlreadyExists, so the website's backfill could never give
migrated members a credential — every one was locked out of password login.
Make provision update the verifier in place on an existing account (logged as
AccountPasswordSet), and treat an empty scram512 as absent consistently in
apply(). Keyed by name, so imported channel/vhost/ban data stays attached.
Mode locks were char-only, so Anope FLOOD/JOINFLOOD locks were dropped on
migration. Carry a value per locked param-mode through the event log, MLOCK
parsing/display, drift enforcement, compaction, and the Anope importer, so a
migrated network keeps its flood protection verbatim.
New 'echo import <anope.json> <out.log> [node]' subcommand builds a fresh event
log from an Anope db_json: accounts (identity only — Anope's one-way password
hashes can't become SCRAM verifiers, so migrated users auth by certfp/external
authority), grouped nicks, cert fingerprints, vhosts, channels (+desc/topic/
assigned bot), access mapped to op/voice by each channel's AUTOOP/AUTOVOICE
thresholds, param-less mode locks, BotServ bots (service pseudo-clients skipped),
and memos. Read-only on the source; self-verifies by replaying the produced log.
Verified against the live backup (11 accounts, 10 channels, 10 bots) + a unit
test over a synthetic fixture. The prerequisite for a safe cutover.
A fresh peer that backfills a node's log via the version vector must reach the
same global state (accounts, akills, forbids, groups); channels/jupes are Local
and correctly don't replicate. Guards the core federation guarantee — a
distinct invariant from same-node replay (exercises missing_for + version_vector
+ cross-node ingest).
Extends build_broad_state with vhost/suspend/akill/forbid/news/cert/access/
group-member removals and a memo read — the asymmetric removal apply arms are
exactly where the drop-purge bug hid. Both the fold-parity and compaction
round-trip tests pass over the full ~55-operation sequence, so live==replay
==compacted holds across adds, mutations and removals.
compact() rebuilds channels field-by-field, and it omitted several fields, so
automatic log compaction (should_compact -> compact) silently deleted them:
channel url, email, successor, noexpire, the restricted/noautoop/bot_greet/
nobot settings, and every jupe. Emits them all now (settings gated on !=
default via a new PartialEq). Adds a compaction round-trip property test
(compact()+reopen == before) alongside the fold-parity one; both share the
broad build/snapshot helpers.
Extends the round-trip live-vs-replay test to ~45 event types (bots, news,
reports, help tickets, oper grants, session exceptions, vhost forbid/template,
entrymsg, noexpire, oper notes, group founder, default bot, account email).
All pass — jupes was the only divergence — so this now guards essentially
every event's live mutation against its apply() arm.
A fold-parity property test caught it: jupes were plain in-memory Db fields
that logged no event, so after any services restart every juped server was
silently un-juped and could relink. Moved jupes into NetData (like akills/
forbids/groups) and event-sourced them with JupeAdded/JupeRemoved (Local
scope), so a jupe now replays from the log. Adds the round-trip property test
(live state == replayed state) as a permanent guard against this bug class.
REGISTER was rate-limited but RESEND and RESETPASS weren't — anyone could
repeatedly request confirmation/reset codes for any account with an address
on file, flooding the victim's inbox and burning the service's sender
reputation. A per-account 60s cooldown (code_issue_wait, mirroring the vhost
request throttle) now gates both paths; guessing was already infeasible
(2^40 code + 5-try limit), so this closes the abuse, not an auth hole.
