Before we only did that, when we basically didn't have the key yet. But
since we usually get sent a RoomKey when a new message is sent after we
sign in, we were discarding, that those messages should usually now be
trusted.
Especially the emoji array shrinks a lot with this, but adds a few extra
relocations on startup. But it removes a lot of exception handling code
at runtime, which is nice and possibly this is still faster.
Calling fsync everytime we save to the db is slow, which is actually
fairly noticeable in some larger E2EE rooms. Speed that up slightly by
batching the olm session persisting.
Correctly verify that the reply to a secrets request is actually coming
from a verified device. While we did verify that it was us who replied,
we didn't properly cancel storing the secret if the sending device was
one of ours but was maliciously inserted by the homeserver and
unverified. We only send secret requests to verified devices in the
first place, so only the homeserver could abuse this issue.
Additionally we protected against malicious secret poisoning by
verifying that the secret is actually the reply to a request. This means
the server only has 2 places where it can poison the secrets:
- After a verification when we automatically request the secrets
- When the user manually hits the request button
It also needs to prevent other secret answers to reach the client first
since we ignore all replies after that one.
The impact of this might be quite severe. It could allow the server to
replace the cross-signing keys silently and while we might not trust
that key, we possibly could trust it in the future if we rely on the
stored secret. Similarly this could potentially be abused to make the
client trust a malicious online key backup.
If your deployment is not patched yet and you don't control your
homeserver, you can protect against this by simply not doing any
verifications of your own devices and not pressing the request button in
the settings menu.
