- introduce PaymentFeeBudget, which contains limits for fee budget and cltv budget
- when splitting a payment,
- the fee budget is linearly distributed between the parts
- this resolves a FIXME in lnrouter ("FIXME in case of MPP")
- the cltv budget is simply copied
- we could also add other kinds of budgets later, e.g. for the num in-flight htlcs
- resolves TODO in lnworker ("todo: compare to the fee of the actual route we found")
The convention is that edges (start_node -> edge_node) store
the policy/fees for the *start_node*.
This is what the non-trampoline edges were already using (for a long time),
but the trampoline ones were off-by-one (policy was for end_node),
which was then worked around in multiple places, to correct for...
i.e. I think because of all the workarounds, there was no actual bug,
but it was just very confusing.
Also note that the prior usage of trampoline edges would not work if
we (sender) were not directly connected to a TF (trampoline-forwarder)
but had extra edges in the route to even get to the first TF.
Having the policy corresponding to the start_node of the edge would work
even in that case.
Electrum clients that have non-trampoline channels.
- if an invoice supports trampoline, but provides only
non-trampoline nodes in its routing hints, we must use
a legacy trampoline payment
- when we create a trampoline onion for a legacy payment,
the invoice_features field is a u64, so we need to remove
high feature bits.
- when creating an invoice, if we had multiple chans with the same trampoline,
were putting duplicate t-hints into the invoice
- when paying an invoice that had duplicate t-hints, we would sometimes
construct invalid paths (SRC>T1->T1->DST)
Upon receiving UNKNOWN_NEXT_PEER, TEMPORARY_NODE_FAILURE
or TEMPORARY_CHANNEL_FAILURE, remember the trampoline route
that was used, and try other routes before raising the
trampoline fee.
Before this commit, we used to raise the trampoline fee
upon receiving any error message, in order to ensure
termination of the payment loop.
Note that we will still retry failing routes after we have
raised the trampoline fee. This choice is questionable, it
is unclear if doing so significantly increases the probability
of success.
Tests: add a test for trampoline handling of UNKNOWN_NEXT_PEER
(This was added in 3a7f5373ac. Not sure what the reason is, but
it triggers UNKNOWN_NEXT_PEER errors, that in turn affect routing
choices, e.g. fallback to single trampoline route)
* Refactor `create_trampoline_route`.
* Enables end-to-end multi-trampoline multipart payments.
Trampoline-to-legacy payments are still not enabled, as this is
currently not supported by Eclair.
* Reverts to a global trampoline fee level, as trampoline failures
are currently not handled properly, see (#7648), which doubles
fee rates.
- Separates the trampoline and local routing multi-part payment cases.
- Ask only for splits that don't send over a single channel (those have
been tried already in the single-part case).
- Makes sure that create_routes_for_payment only yields partial routes
that belong to a single split configuration.
- Tracks trampoline fee levels on a per node basis, previously, in the
case of having two channels with a trampoline forwarder, the global
fee level would have increased by two levels upon first try.
It is the last Trampoline Forwarder that should be checked, not the
first one.
Consider route (of Trampolines only):
Alice-electrum -> T_ACINQ -> T_Hodlister -> Bob-electrum
Even if Bob has a transport open with ACINQ or even if Bob has a channel open with ACINQ,
Alice can safely use end-to-end trampoline for this route: ACINQ will not know who
the recipient is, so they will not try to do pay-to-open (and hold up the payment for minutes...).
related: https://github.com/ACINQ/lightning-kmp/pull/237
- better error handling: previously we stopped all attempts on any of
TRAMPOLINE_EXPIRY_TOO_SOON, UNKNOWN_NEXT_PEER, TEMPORARY_NODE_FAILURE.
Instead we should retry (but see code comments).
- previously payments failed if ALL of the following criteria applied:
- sender is paying via trampoline, but not via the ACINQ node (which is
special cased)
- receiver only has private channels and has put r_tags into invoice, along
with setting the trampoline feature bit in the invoice, however the receiver
is not connected to any trampoline forwarders directly
The sender would then assume that the private routing hints in the invoice
correspond to trampoline forwarders.
- also, previously if both the sender and the recipient used trampoline and
they shared a trampoline forwarder (that they were both connected to), the
private channels the recipient had (with nodes other than the shared TF)
would never be attempted.
- fix regression in create_routes:
fwd_trampoline_onion was not added to the tuple
- fix onion structure for e2e
- maybe_fulfill_htlc:
check the mpp_status of the outer onion,
return trampoline_onion to be forwarded
We pass the private edges to lnrouter, and let it find routes end-to-end.
Previously the edge_cost heuristics didn't apply to the private edges
and we were just randomly picking one of the route hints and use that.
So e.g. cheaper private edges were not preferred, but they are now.
PathEdge now stores both start_node and end_node; not just end_node.
- trampoline node is the final recipient of MPP
- each trampoline receives a bucket of HTLCs
- if a HTLC from a bucket fails, wait for the entire bucket to fail
- move trampoline route and onion code into trampoline module
ThomasV
SomberNight
bitromortac
wakiyamap
Benoit Verret