Merge e15517ee53 into d6edcbd946

proposals/4104-auth-lock-soft-failure-be-gone.md

@@ -0,0 +1,204 @@
+# MSC4104: Auth Lock: Soft-failure-be-gone!
+
+This proposal aims to make soft failure redundant, by introducing
+`m.auth_lock` as a mechanism to stop authorising events that are
+crafted to reference previous state. All without disrupting the
+history of the room for new users.
+
+This MSC introduces a new authorization event, the `m.auth_lock` event.
+When a server issues an authorization event that supersedes an existing
+event, for example by banning a user, the admin can choose to
+canonicalise their version of the room history by issuing an auth-lock.
+This can be thought of as a thread lock from forum software applied
+to a specific portion of the DAG. The auth-lock event not only
+specifies the authorization event to lock, but all forward
+extremities that reference the locked authorization event. Thus
+canonicalising the history of that part of the DAG.
+
+Any further events that reference an authorization event that has been
+*locked* will then be rejected outright during authorization.
+
+## Proposal
+
+### The auth lock event `m.auth_lock`
+
+The `m.auth_lock` event can be issued in scenarios where room admins
+or their tooling detect or anticipate events referring to previous
+authorization events. This should only be done eagerly when there
+is reasonable risk. For example, it would be inappropriate to
+issue an `m.auth_lock` event after banning a spammer that
+resided on `matrix.org` as it is unlikely for matrix.org
+to attempt to maliciously reference old authorization events.
+By delaying the application of `m.auth_lock`, room administrators
+reduce the risk of divergence. This is something that currently
+occurs frequently with [soft failure](https://github.com/element-hq/synapse/issues/9329).
+
+#### Properties
+- `content`:
+  + `locked_event_id`: The event_id of the authorization event to lock.
+    This is protected from redaction.
+  + `extremities`: A list of extremities that should be shared to other
+   servers representing the canonicalised version of the room history.
+   This is protected from redaction.
+
+- `state_key`: A sha256 of the `locked_event_id` concatenated with the
+  mxid of the event sender. We do this so that different room admins
+  can optionally specify their own versions of the `extremities`
+  that will combine to form a complete set to be considered during
+  authorization. This also keeps `extremities` conflict free.
+
+### Changes to auth_rules
+
+#### Issuing `m.auth_lock`
+
+If type is `m.auth_lock`:
+  1. if `state_key` does not match a sha256 of `locked_event_id` and
+     `sender`, reject.
+
+#### Authorizing other events
+
+Considering the event's `auth_events`:
+  1. For each `auth_event` all `m.auth_lock` events with a
+     `locked_event_id` matching the `event_id` of the `auth_event`
+     are found from the room's current state.
+     From each `m.auth_lock` event relevant to one `auth_event`, the
+     `extremities` fields are combined to form a set of all `extremities`
+     relevant to the `auth_event`.
+     If the considered auth_event's `event_id` matches any
+     `m.auth_lock`'s `locked_event_id` field AND the `event_id` of the
+     event being authorized is absent from both the `extremities` set
+     AND the `event_id` of any connected event referenced via each
+     extremity's `prev_events`, reject.
+
+## Potential issues
+
+### Multiple auth-lock events
+
+Imagine that there are two admins in a room, Alice and Bob.
+Imagine that there is a malicious user Chelsea.
+Alice and Bob both ban Chelsea.
+Alice sees Chelsea send events B -> A.
+Bob sees Chelsea send events C -> B -> A.
+Alice and Bob both issue auth_lock events for Chelsea's prior membership.
+Alice's `extremities` field includes `B`, but excludes `C`.
+Bob's `extremities` field includes `C`.
+
+#### Solution
+
+The receiving server must combine the `extremities` from both events
+such that the canonical history becomes C -> B -> A.
+This is already explicitly written within the authorization rules
+in this MSC.
+
+### Diverging `extremities`
+
+Imagine Bob's server receives an `m.auth_lock` event and Bob's server
+has authorized events referencing the `locked_event_id` that are not
+present within any relevant `m.auth_lock`'s `extremities`.
+Bob's server must decide what to do with the events it has already
+authorized.
+
+This is a problem that occurs in the following scenarios:
+
+#### Scenario A: Alice deliberately encodes less `extremities` than Bob has
+Alice issues an `m.auth_lock` to deliberately encode less or no
+extremities, in an attempt permanently to diverge Bob's DAG.
+Joining user David would then also not be able to see a large portion
+of the history that Bob can, because it wouldn't be authorized in
+David's DAG when David backfills.
+
+#### Scenario B: Alice issues a lock quickly after changing an auth event
+
+Imagine that Alice issues an `m.room.power_levels` event P.
+Alice revises this event so that it becomes P'.
+Alice issues an auth_lock for P, Lock-P.
+Imagine that Alice receives an event B from Bob that references P.
+Alice and every other server with Lock-P is forced to reject event B.
+
+#### Scenario C: Bob's events are mixed with Chelsea's
+
+Chelsea has sent `X -> A` to Alice, but `Y -> A` to Bob.
+Bob has then appended `B -> Y -> A` and in turn Chelsea
+has appended `Z -> B -> Y -> A` and sent `Z` to Bob.
+
+Alice has banned Chelsea and locked Chelsea's prior membership.
+Only now has Bob received this `m.auth_lock` event, and it only
+contains the extremity `X`.
+
+Alice can get `B` if Bob sends it to her,
+but new joiner David will probably be unable to by backfilling `B`,
+at least if David's server doesn't have special code to account
+for the connected events being rejected.
+
+#### Effects on participating servers 
+
+Servers are required to maintain all their existing `extremities`
+and their `prev_events`, regardless of whether they are missing from
+the `extremities` field of any `m.auth_lock` event.
+
+These servers should still incorporate these diverged `extremities`
+into `prev_events` so that they can be discovered and repaired later.
+
+#### Effects on joining servers
+
+Joining servers will no longer be able to authorize large parts of the
+room's history.
+
+#### Solution A: Dedicated API for showing diverged events to room admins
+
+A room administrator could be made aware of all the extremities that
+are diverged. This can be done because a server can detect when
+authorized events contain references to unauthorized events
+within `prev_events` of authorized events in the DAG.
+
+A room administrator can then preview the effects of incorporating
+diverged extremities and make a decision to do so via the
+`extremities` field of `m.auth_lock`.
+
+#### Solution B: Room administrators on other servers issue their own `m.auth_lock`
+
+Room admins residing on different servers with diverged extremities
+can issue their own `m.auth_lock` event to include diverged extremities.
+When this event is received by other servers, the diverged extremities
+can be authorized and the DAG will converge again.
+
+#### Solution C: Unprivileged users propose changes to the `extremities` of `m.auth_lock`
+
+Participants in the room can propose changes to the `extremities`
+contained within `m.auth_lock` via a dedicated event. Room administrators
+can then review the effects and merge them.
+
+#### Solution D: Recrafting PDUs
+
+Take scenario B. Bob receives Lock-P, and notices that B is absent
+from the chain specified in Lock-P's `extremities`.
+
+We would need a new MSC for bob to resend B as B' such that servers who
+have accepted B would recognise B' to be a duplicate that references
+more recent authorization events.
+
+This would allow Alice to see `B` in Scenario C. But it wouldn't
+allow Alice to redact `Z` or `Y`, which Bob my find harmful.
+
+### Locking all current auth events
+
+It would be possible to lock all auth_events and stop any further
+developments in the room.
+
+## Alternatives
+
+None recognized.
+
+
+## Security considerations
+
+See potential issues.
+
+
+## Unstable prefix
+
+None issued.
+
+## Dependencies
+
+None.