Instead of having one ring buffer per network, each network has one ring
buffer per entity (channel or nick). This allows history to be more
fair: if there's a lot of activity in a channel, it won't prune activity
in other channels.
We now track history sequence numbers per client and per network in
networkHistory. The overall list of offline clients is still tracked in
network.offlineClients.
When all clients have received history, the ring buffer can be released.
In the future, we should get rid of too-old offline clients to avoid
having to maintain history for them forever. We should also add a
per-user limit on the number of ring buffers.
Any SendMessage call after Close could potentially block forever if the
outgoing channel was filled up. Now the channel is drained before the
writer goroutine exits.
This commit adds support for downstream LIST messages from multiple
concurrent downstreams to multiple concurrent upstreams, including
support for multiple pending LIST requests from the same downstream.
Because a unique RPL_LISTEND message must be sent to the requesting
downstream, and that there might be multiple upstreams, each sending
their own RPL_LISTEND, a cache of RPL_LISTEND replies of some sort is
required to match RPL_LISTEND together in order to only send one back
downstream.
This commit adds a list of "pending LIST" structs, which each contain a
map of all upstreams that yet need to send a RPL_LISTEND, and the
corresponding LIST request associated with that response. This list of
pending LISTs is sorted according to the order that the requesting
downstreams sent the LIST messages in. Each pending set also stores the
id of the requesting downstream, in order to only forward the replies to
it and no other downstream. (This is important because LIST replies can
typically amount to several thousands messages on large servers.)
When a single downstream makes multiple LIST requests, only the first
one will be immediately sent to the upstream servers. The next ones will
be buffered until the first one is completed. Distinct downstreams can
make concurrent LIST requests without any request buffering.
Each RPL_LIST message is forwarded to the downstream of the first
matching pending LIST struct.
When an upstream sends an RPL_LISTEND message, the upstream is removed
from the first matching pending LIST struct, but that message is not
immediately forwarded downstream. If there are no remaining pending LIST
requests in that struct is then empty, that means all upstreams have
sent back all their RPL_LISTEND replies (which means they also sent all
their RPL_LIST replies); so a unique RPL_LISTEND is sent to downstream
and that pending LIST set is removed from the cache.
Upstreams are removed from the pending LIST structs in two other cases:
- when they are closed (to avoid stalling because of a disconnected
upstream that will never reply to the LIST message): they are removed
from all pending LIST structs
- when they reply with an ERR_UNKNOWNCOMMAND or RPL_TRYAGAIN LIST reply,
which is typically used when a user is not allowed to LIST because they
just joined the server: they are removed from the first pending LIST
struct, as if an RPL_LISTEND message was received
Split user.register into two functions, one to make sure the user is
authenticated, the other to send our current state. This allows to get
rid of data races by doing the second part in the user goroutine.
Closes: https://todo.sr.ht/~emersion/soju/22
In a later commit, we'll be able to move part of downstreamConn.register
into the user goroutine to prevent races.
References: https://todo.sr.ht/~emersion/soju/22
