We currently copy the features objects in each channel as we walk
the graph during route calculation. This implies a significant
amount of malloc traffic as the features flags object are stored
on the heap.
Instead, because they features being referenced are in the network
graph which we hold a reference to, we can simply store references
to them.
This nontrivially improves our get_route benchmark by around 5%.
While walking the graph doing Dijkstra's, we may decrease the
amount being sent along one path, and not others, based on the
htlc_minimum_msat value. This may result in a lower relative fees
on that path in comparison to others. In the extreme, this may
result in finding a second path to a node with a lower fee than the
first path found (normally impossible in a Dijkstra's
implementation, as we walk next hops in fee-order).
In such a case, we end up with parts of our state invalid as
further hops beyond that node have been filled in using the
original total fee information.
Instead, we simply track which nodes have been processed and ignore
and further updates to them (as it implies we've reduced the amount
we're sending to find a lower absolute fee, but a higher relative
fee, which isn't a better route).
We check that we are in exactly this case in test builds with new
in-line assertions. Note that these assertions successfully detect
the bug in the previous commit.
Sadly this requires an extra HashMap lookup every time we pop an
element off of our heap, though we can skip a number of heap pushes
during the channel walking.
This is the same code as was recently failing in our benchmarks,
adapted to use a random starting seed instead of a fixed one and
a smaller iteration to reduce runtime.
When walking the network graph to calculate a route, we always
calculate the minimum fee which is required to make one further
hop. This avoids some extra hop processing at the end of each path
selection loop (saving around 10% runtime in our benchmarks).
However, if we use the real value which we expect
to send over a channel in that calculation, we may find an
alternate path to the same node which is more expensive but
capacity-constrained, resulting in us considering it cheaper as the
relative fee paid will be lower.
Instead, we can use the `minimal_value_contribution_msat`, which is
a constant through an entire path finding iteration, as the amount,
preventing any basis change in the relative fee paid.
Currently, the "best source" for a given node tracked during
Dijkstra's is updated with a different critera from the heap
sorting, resulting in loops in calculated paths.
This adds a test for the specific failure currently seen, utilizing
the new path-htlc-minimum tracking in the heap entries in place of
the per-hop htlc-minimum values which the MPP changeset partially
used.
If we walk the network graph and find a route that meets are
payment amount and has a liquidiy limit far in excess of our
payment amount, we may select the same path several times in the
main path-gathering loop.
Instead, if the path we selected was not limited by the available
liquidity, we set the middle hop in the path's liquidity available
to zero, disabling that channel in future path finding steps.
Previously, we'd happily send funds through a path where, while
generating the path, in some middle hope we reduce the value being
sent to meet an htlc_maximum, making a later hop invalid due to it
no longer meeting its htlc_minimum. Instead, we need to track the
path's htlc-minimum while we're transiting the graph.
The new MPP routing algorithm attempts to build paths with a higher
value than our payment to find paths which may allow us to pay a
fee to meet an htlc_minimum limit. This is great if we're
min-bounded, however it results in us rejecting paths where we are
bounded by a maximum near the end of the path, with fees, and then
bounded by a minimum earlier in the path. Further, it means we will
not find the cheapest path where paths have a lower relative fee
but a higher absolute fee.
Instead, we calculate routes using the actual amount we wish to
send. To maintain the previous behavior of searching for cheaper
paths where we can "pay the difference" to meet an htlc_minimum, we
detect if we were minimum-bounded during graph walking and, if we
are, we walk the graph again with a higher value.
Useful for constructing route hints for private channels in invoices.
Co-authored-by: Valentine Wallace <vwallace@protonmail.com>
Co-authored-by: Antoine Riard <ariard@student.42.fr>
Modify NetGraphMsgHandler::handle_query_channel_range to always use
first_blocknum=0 in replies. This is spec compliant after changes to
make sequence completion explicity using sync_complete.
Modifies NetGraphMsgHandler::handle_query_channel_range to use a constant
max value in replies. Modifies tests to generate 8000 channels instead
of making this value configurable.
We currently only use it to override the graph-specific features
returned in the route, though we should also use it to enable or
disable MPP.
Note that tests which relied on MPP behavior have had all of their
get_route calls upgraded to provide the MPP flag.
Initial implementation of handling query_channel_range messages in
NetGraphMsgHandler. Enqueues a sequence of reply message in the pending
message events buffer.
We could have possibly constructed a slightly inconsistent
path: since we reduce value being transferred all the way, we
could have violated htlc_minimum_msat on some channels
we already passed (assuming dest->source direction). Here,
we recompute the fees again, so that if that's the case, we
match the currently underpaid htlc_minimum_msat with fees.
Sadly rust upstream never really figured out the benchmark story,
and it looks like the API we use here may not be long for this
world. Luckily, we can switch to criterion with largely the same
API if that happens before upstream finishes ongoing work with the
custom test framework stuff.
Sadly, it requires fetching the current network graph, which I did
using Val's route-testing script written to test the MPP router.
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 (finally) exposes `ChannelManager`/`ChannelMonitor` _write
methods, which were (needlessly) excluded as the structs themselves
have generic parameters. Sadly, we also now need to parse
`(C-not exported)` doc comments on impl blocks as we otherwise try
to expose _write methods for `&Vec<RouteHop>`, which doesn't work
(and isn't particularly interesting for users anyway). We add such
doc comments there.
Our bindings generator is braindead with respect to the idents
used in a trait definition - it treats them as if they were used
where the trait is being used, instead of where the trait is
defined. Thus, if the idents used in a trait definition are not
also imported the same in the files where the traits are used, we
will claim the idents are bogus.
I spent some time trying to track the TypeResolvers globally
through the entire conversion run so that we could use the original
file's TypeResolver later when using the trait, but it is somewhat
of a lifetime mess. While likely possible, import consistency is
generally the case anyway, so unless it becomes more of an issue in
the future, it likely makes the most sense to just keep imports
consistent.
This commit keeps imports consistent across trait definition files
around `MessageSendEvent` and `MessageSendEventsProvider`.
This method was used to set the initial_routing_sync flag when sending
an outbound Init message to a peer. Since we are now relying on
gossip_queries instead of initial_routing_sync, synchronization can be
fully encapsulate into RoutingMessageHandler via sync_routing_table.
This commit removes should_request_full_sync from the trait
RoutingMessageHandler. The implementation is still used in
NetGraphMsgHandler and has been converted into a private method instead
of a trait function.
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 commit modifies sync_routing_table in RoutingMessageHandler to
accept a reference to the Init message received by the peer. This allows
the method to use the Peer's features to drive the operations of the
gossip_queries routing table 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.
This commit simplifies the sync process for routing gossip messages. When
a sync is initiated, the process is handled statelessly by immediately
issuing SCID queries as channel range replies are received. This greatly
simplifies the state machine at the cost of fully validating and
conforming to the current spec.
