The only practical way to meet this requirement is to block
disconnect_socket until any pending events are fully processed,
leading to this trivial deadlock:
* Thread 1: select() woken up due to a read event
* Thread 2: Event processing causes a disconnect_socket call to
fire while the PeerManager lock is held.
* Thread 2: disconnect_socket blocks until the read event in
thread 1 completes.
* Thread 1: bytes are read from the socket and
PeerManager::read_event is called, waiting on the lock
still held by thread 2.
There isn't a trivial way to address this deadlock without simply
making the final read_event call return immediately, which we do
here. This also implies that users can freely call event methods
after disconnect_socket, but only so far as the socket descriptor
is different from any later socket descriptor (ie until the file
descriptor is re-used).
`Julian Knutsen <julianknutsen@users.noreply.github.com>` pointed
out in a previous discussion that `read_event` can reenter user
code despite the documentation stating explicitly that it will not.
This was addressed in #456 by simply codifying the reentrancy, but
its somewhat simpler to just drop the `do_attempt_write_data` call.
Ideally we could land most of Julian's work, but its still in need
of substantial git history cleanup to get it in a reviewable state
and this solves the immediate issue.
We can never assume that messages were reliably delivered whether
we placed them in the socket or not, so there isn't a lot of use in
explicitly handling the case that a peer was not connected when we
went to send it a message.
Two TODOs are left for the generation of a `FundingAbandoned` (or
similar) event, though it ultimately belongs in `ChannelManager`.
We were waiting for the initiator, but the spec doesn't guarantee that they will send Init first, so we might theoretically wait forever.
Also, lnprototest expects this behavior.
Previously, if a user simultaneously called
`PeerHandler::process_events()` from two threads, we'd race, which
ended up sending messages out-of-order in the real world.
Specifically, we first called `get_and_clear_pending_msg_events`,
then take the `peers` lock and push the messages we got into the
sending queue. Two threads may both get some set of messages to
send, but then race each other into the `peers` lock and send the
messages in random order.
Because we already hold the `peers` lock when calling most message
handler functions, we can simply take the lock before calling
`get_and_clear_pending_msg_events`, solving the race.
This will be used to expose forwarding info for route hints in the next commit.
Co-authored-by: Valentine Wallace <vwallace@protonmail.com>
Co-authored-by: Antoine Riard <ariard@student.42.fr>
This is largely useful for bindings, and the off-github discussion
around #814 concluded these should be pub, but the PR was not
updated to capture this. Now that the bindings support generation
for the structs, expose them.
Creates a MessageSendEvent for sending a reply_channel_range message.
This event will be fired when handling inbound query_channel_range
messages in the NetGraphMessageHandler.
We currently "support" not having a router or channel in memory by
forcing users to implement the same, but its trivial to provide our
own dummy implementations.
* Implemented protocol.
* Made feature optional.
* Verify that the default value is true.
* Verify that on shutdown,
if Channel.supports_shutdown_anysegwit is enabled,
the script can be a witness program.
* Added a test that verifies that a scriptpubkey
for an unreleased segwit version is handled successfully.
* Added a test that verifies that
if node has op_shutdown_anysegwit disabled,
a scriptpubkey with an unreleased segwit version on shutdown
throws an error.
* Added peer InitFeatures to handle_shutdown
* Check if shutdown script is valid when given upfront.
* Added a test to verify that an invalid test results in error.
* Added a test to check that if a segwit script with version 0 is provided,
the updated anysegwit check detects it and returns unsupported.
* An empty script is only allowed when sent as upfront shutdown script,
so make sure that check is only done for accept/open_channel situations.
* Instead of reimplementing a variant of is_witness_script,
just call it and verify that the witness version is not 0.
The only API change outside of additional derives is to change
the inner field in `DecodeError::Io()` to an `std::io::ErrorKind`
instead of an `std::io::Error`. While `std::io::Error` obviously
makes more sense in context, it doesn't support Clone, and the
inner error largely doesn't have a lot of value on its own.
This public method allows a client to easily disconnect peers while only
owning its node id. It will clean up peer state and disconnect properly
its descriptor.
This commit changes outbound routing table sync to use gossip_queries
instead of the effectively deprecated initial_routing_sync feature.
This change removes setting of initial_routing_sync in our outbound Init
message. Instead we now call sync_routing_table after receiving an Init
message from a peer. If the peer supports gossip_queries and
should_request_full_sync returns true, we initiate a full gossip_queries
sync.
This change modifies gossip_queries methods in RoutingMessageHandler to
move the message instead of passing a reference. This allows the message
handler to be more efficient by not requiring a full copy of SCIDs
passed in messages.
Defines message handlers for gossip_queries messages in the RoutingMessageHandler
trait. The MessageSendEventsProvider supertrait is added to RoutingMessageHandler
so that the implementor can use SendMessageEvents to send messages to a
peer at the appropriate time.
The trait methods are stubbed in NetGraphMsgHandler which implements
RoutingMessageHandler and return a "not implemented" error.
This change enables initiating gossip queries with a peer using the
SendMessageEvent enum. Specifically we add an event for sending
query_channel_range to discover the existance of channels and an event
for sending query_short_channel_ids to request routing gossip messages
for a set of channels. These events are handled inside the process_events
method of PeerManager which sends the serialized message to the peer.
To enable gossip_queries message decoding, this commit implements the
wire module's Encoding trait for each message type. It also adds these
messages to the wire module's Message enum and the read function to
enable decoding of a buffer.
Support for the gossip_queries feature flag (bits 6/7) is added to the
Features struct. This feature is available in the Init and Node
contexts. The gossip_queries feature is not fully implemented so this
feature is disabled when sent to peers in the Init message.
This changes the LICENSE file and adds license headers to most files
to relicense under dual Apache-2.0 and MIT. This is helpful in that
we retain the patent grant issued under Apache-2.0-licensed work,
avoiding some sticky patent issues, while still allowing users who
are more comfortable with the simpler MIT license to use that.
See https://github.com/rust-bitcoin/rust-lightning/issues/659 for
relicensing statements from code authors.
Tests use sources of randomness to produce seeds, preimages, secrets,
and ephemeral data. However, this makes comparing logs between different
test runs difficult. Remove uses of random number generators and the
current time in favor of fixed values in order to make the test output
deterministic.
This is a response to splitting #585 into smaller components. This extraction should allow the subsequent creation of a trait for all message handling, thereby enabling more flexibility in the state machine, particularly for bindings.
We also update to use single idents when referencing the Deref=*
types since the automated code generator is pretty braindead.
This also moves some test utils out of peer_handler.rs and into
util::test_utils to standardize things a little bit, which we need
to concretize the PeerHandler types used in testing.
This is a response to splitting #585 into smaller components. This extraction will permit a future extraction of all message handling into a separate method, and then later even into a separate trait that will be accessible through language bindings.
This caused a bunch of cascading changes, including
passing loggers down to Channels in function calls
rather than having each Channel have a pointer to the
ChannelManager's Logger (which was a circular reference).
Other structs that the Channel had passed its Logger to also
had their loggers removed. Other newly unused Loggers were
also removed, especially when keeping them would've caused
a bunch of extra test changes to be necessary, e.g. with
the ChainWatchInterfaceUtil's Logger.
