`ChannelDetails::outbound_capacity_msat` describes the total amount
available for sending across several HTLCs, basically just our
balance minus the reserve value maintained by our counterparty.
However, when routing we use it to guess the maximum amount we can
send in a single additional HTLC, which it is not.
There are numerous reasons why our balance may not match the amount
we can send in a single HTLC, whether the HTLC in-flight limit, the
channe's HTLC maximum, or our feerate buffer.
This commit splits the `outbound_capacity_msat` field into two -
`outbound_capacity_msat` and `outbound_htlc_limit_msat`, setting us
up for correctly handling our next-HTLC-limit in the future.
This also addresses the first of the reasons why the values may
not match - the max-in-flight limit. The inaccuracy is ultimately
tracked as #1126.
Default to creating tlv onions for nodes for which we haven't received
any features through node announcements or which aren't in the
`network_graph`, and where no other features are known such as invoice
features nor features in the init msg for nodes we have channels to.
When we start getting a numerator and divisor particularly close to
each other, the log approximation starts to get very noisy. In
order to avoid applying scores that are basically noise (and can
range upwards of 2x the default per-hop penalty), simply consider
such cases as having a success probability of 100%.
When we send values over channels of rather substantial size, the
imprecision of our log lookup tables creates a rather substantial
non-linearity between values that round up or down one bit.
For example, with the default scoring values, sending 100k sats
over channels with 1m, 2m, 3m, and 4m sats of capacity score
rather drastically differently: 3645, 2512, 500, and 1442 msat.
Here we expand the precision of our log lookup tables rather
substantially by: (a) making the multiplier 2048 instead of 1024,
which still fits inside a u16, and (b) quadrupling the size of the
lookup table to look at the top 6 bits after the most-significant
bit of an input instead of the top 4.
This makes the scores of the same channels substantially more
linear, with values of 3613, 1977, 1474, and 1223 msat.
The same channels would be scored at 3611, 1972, 1464, and 1216
msat with a non-approximating scorer.
The cost of large payments tends to be dominated by the channel fees. To
avoid this, add an amount penalty to ProbabilisticScorer with a user
configurable multiplier. The multiplier is applied for every 2^20th of
the amount weighted by the negative log10 of the channel's success
probability for the payment.
In ProbabilisticScorer, the channel liquidity balance is reduced
whenever a payment fails at the corresponding channel. The payment may
still be retried through the channel, however, because the liquidity
penalty is capped. Use u64::max_value instead in this situation to avoid
retrying over the same path. This effectively makes u64::max_value the
penalty for amounts exceeding the upper bound, as well.
As an edge case, avoid using u64::max_value on attempts where the amount
is equal to the effective capacity, which may be the HTLC maximum when
the channel capacity is unknown.
During the first pass of path finding, we seek a single path with the
exact payment amount, and only seek additional paths if (a) no single
path can carry the entire balance of the payment or (b) we found a good
path, but along the way we found candidate paths with the potential to
result in a lower total fee. This commit fixes the behavior of (b) -- we
were previously considering some paths to be candidates for a lower fee
when in fact they never would have worked. This caused us to re-run
Dijkstra's when it might not have been beneficial.
As we add new supported channel types, inbound channels which use
new features may cause backwards-compatibility issues for clients.
If a new channel is opened using new features while a client still
wishes to ensure support for downgrading to a previous version of
LDK, that new channel may cause the `ChannelManager` to fail
deserialization due to unsupported feature flags.
By exposing the channel type flags to the user in channel requests,
users wishing to support downgrading to previous versions of LDK
can reject channels which use channel features which previous
versions of LDK do not understand.
When a routing hint is given in an invoice, the effective capacity of
the channel is assumed to be infinite (i.e., u64::max_value) if the hop
is private. Adding 1 to this in the success probability calculation will
cause an overflow and ultimately an `index out of bounds panic` in
log10_times_1024. This was not an issue with using log10 because the use
of f64 would give infinite which casts to 0 for u64.
ProbabilisticScorer tends to prefer longer routes to shorter ones. Make
the default scoring behavior include a customizable base penalty to
avoid longer routes.
`channel_update` messages already have their signatures checked
with the network graph write lock held, so there's no reason to
check the signatures before doing other quicker checks first,
including checking if we're already aware of a newer update for the
channel.
This reduces common-case CPU usage as `channel_update`s are sent
rather liberally over the p2p network to gossip them.
When we have many channels to the same first-hop, many of which do
not have sufficient balance to make the requested payment, but when
some do, instead of simply using the available channel balance we
may switch to MPP, potentially with many, many paths.
Instead, we should seek to use the smallest channel which can
easily handle the requested payment, which we do here by sorting
the first_hops in our router before beginning the graph search.
Note that the "real" fix for this should be to instead decide which
channel to use at HTLC-send time, as most other nodes do during
relay, but this provides a minimal fix without needing to do the
rather-large work of refactoring our HTLC send+relay pipelines.
Issues with overly-aggressive MPP on many channels were reported by
Cash App.
On connection, if our peer supports gossip queries, and we never
send a `gossip_timestamp_filter`, our peer is supposed to never
send us gossip outside of explicit queries. Thus, we'll end up
always having stale gossip information after the first few
connections we make to peers.
The solution is to send a dummy `gossip_timestamp_filter`
immediately after connecting to peers.
