The way PeerHandler was written, it was supposed to remove from
self.peers iff the API docs indicate that disconnect_event should
NOT be called (and otherwise rely on disconnect_event to do so).
Sadly, the implementation was way out of whack with reality - in
the implementation, essentially anywhere where PeerHandler
originated the disconnection, the peer was removed and no
disconnect_event was expected. The docs, however, indicated that
disconnect_event should nearly only be called, only not doing so
when the initial handshake message never completed.
We opt to change the docs, mostly, as well as clean up the
ping/pong handling somewhat and rename a few functions to clarify
what they actually do.
This is important for a number of reasons:
* Firstly, I hit this trying to implement rescan in the demo
bitcoinrpc client - if individual ChannelMonitors are out of
sync with each other, we cannot add them all into a
ManyChannelMonitor together and then rescan, but need to rescan
them individually without having to do a bunch of manual work.
Of the three return values in ChannelMonitor::block_connected,
only the HTLCsource stuff that is moved here makes no sense to
be exposed to the user.
* Secondly, the logic currently in ManyChannelMonitor cannot be
reproduced by the user! HTLCSource is deliberately an opaque
type but we use its data to decide which things to keep when
inserting into the HashMap. This would prevent a user from
properly implementing a replacement ManyChannelMonitor, which is
unacceptable.
* Finally, by moving the tracking into ChannelMonitor, we can
serialize them out, which prevents us from forgetting them when
loading from disk, though there are still other races which need
to be handled to make this fully safe (see TODOs in
ChannelManager).
This is safe as no two entries can have the same HTLCSource across
different channels (or, if they did, it would be a rather serious
bug), though note that, IIRC, when this code was added, the
HTLCSource field in the values was not present.
We also take this opportunity to rename the fetch function to match
our other event interfaces, makaing it clear that by calling the
function the set of HTLCUpdates will also be cleared.
Previously, if we have a live ChannelManager (that has seen blocks)
and we open a new Channel, if we serialize that ChannelManager
before a new block comes in, we'll fail to deserialize it. This is
the result of an overly-ambigious last_block_connected check which
would see 0s for the new channel but the previous block for the
ChannelManager as a whole.
We add a new test which catches this error as well as hopefully
getting some test coverage for other similar issues in the future.
Previously, when attempting to write out a channel with some
RemoteAnnounced pending inbound HTLCs, we'd write out the count
without them, but write out some of their fields. We should drop
them as intended as they will need to be reannounced upon
reconnection.
This was found while attempting to simply reproduce a different
bug by adding tests for ChannelManager serialization rount-trip at
the end of each functional_test (in Node::drop). That test is
included here to prevent some classes of similar bugs in the future.
This tests, after each functional test, that if we serialize and
reload all of our ChannelMonitors we end up tracking the same set
of outputs as before.
Upon deserialization/reload we need to be able to register each
outpoint which spends the commitment txo which a channelmonitor
believes to be on chain. While our other internal tracking is
likely sufficient to regenerate these, its much easier to simply
track all outpouts we've ever generated, so we do that here.
Previously, if new ouputs were found to be watched as part
of channel operations, the block was rescan which triggers
again parser and generation of transactions already issued.
This commit first modifies the test framework without
altering further ChannelMonitor.
ChannelMonitor refactoring is introduced in a latter commit.
This implements the new TLV variable-length encoding for onion hop
data, opting to send it if the RouteHop's node_features indicates
support. It also uses the new process_inline method in ChaCha20 to
optimize a few things (though it grows a new TODO for a
probably-important optimization).
There's quite a bit of machinery included here, but it neatly
avoids any dynamic allocation during TLV deserialization, and the
calling side looks nice and simple. The macro-generated code is
pretty nice, though has some redundant if statements (I haven't
checked if they get optimized out yet, but I can't imagine they
don't).
This adds a number of new stream adapters to track and/or calculate
the number of bytes read/written to an underlying stream, as well
as wrappers for the two (?!) variable-length integer types that TLV
introduces.
This prepares for variable-length per-hop-data by wrapping the full
hop_data field in a decrypting stream, with a few minor
optimizations and redundant allocations to boot.
This, as it should be, restricts OnionHopData to only being able to
represent valid states, while still allowing for tests to generate
bogus hop data fields to test deserialization.
Its a bit awkward to have an hmac field covering the struct that
its in, and there is little difference in removing it, so just pull
it out and use a [u8; 32] where we care about the hmac.
