Multiple places in the pay lifecycle depend on mods to be set. By
setting the mods directly after the first listpeerchannels call,
subsequent calls to listpeerchannels are avoided.
Changelog-Fixed: pay-plugin: only call listpeerchannels once during a
payment lifecycle.
For nodes with many channels this is a tremendous improvement in pay
performance. PR #7611 improves payment performance from 15 seconds to
13.5 seconds on one of our nodes. This commit improves payment
performance from 13.5 seconds to 5.7 seconds.
Changelog-Fixed: Improved pathfinding speed for large nodes.
This minimizes the need to convert back and forth from and to sat
values, and it also removes a new instance of sats in the public
interface (`channel_hints`).
Suggested-By: Rusty Russell <@rustyrussell>
We were using `channel_hint` to temporarily tweak the graph inside of
a payment. However, these ad-hoc `channel_hints` are stickier than
their predecessors, in that they outlive the payment attempt itself,
and interfere with later ones.
Changelog-Changed: pay: Discarding an overly long or expensive route does not blocklist channels anymore.
We need to know the overall channel capacity, i.e., the amount_msat
that the channel was funded with, in order to relax the channel_hint
to refill over time.
It was weird not to have a capacity associated with localmods channels, and
fixing it has some very nice side effects.
Now the gossmap_chan_get_capacity() call never fails (we prevented reading
of channels from gossmap in the partially-written case already), so we
make it return the capacity. We do this in msat, because that's what
all the callers want.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The `pay` plugin, as well as other plugins making use of the tree-pay
executor, will now emit their observations as they see them. The
notifications are sent on the `channel_hint_updated` topic, and any
subscriber will get them.
We also added a `timestamp` to the `struct channel_hint`, as these
observations now outlive the `pay` call, and have to be attenuated /
relaxed as they age, until we can eliminate them completely (when the
restriction is equal to the structural information gathered from the
gossip).
Changelog-Added: pay: Payments now emit `channel_hint_updated` notification to share inferred balances and observations across multiple payments.
Changelog-Added: pay: The pay plugin now checks whether we have enough spendable capacity before computing a route, returning a clear error message if we don't
This allows us to directly returnan error code based on where we
decided to abort, rather than attemtping to infer it from the parts.
Changelog-Added: pay: The pay plugin now returns better error codes
Changelog-Added: Protocol: pay can now pay to bolt12 invoices if entry to blinded hop is specified as a short_channel_id (rather than node id).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
For bolt12, we have blinded paths so we route to the head of the blinded
path, which may not be the same as the final payment destination.
This matters mainly for detecting self-pay.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This can't happen because we go the self-pay path in this case, but
once we fix that for bolt12, this can be reached.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The spec has moved a bit here: the `outgoing_cltv_value` in the final onion
is basically the blockheight now (plus the 1 block delta we give ourselves).
Also, we were doubling ours, since p->min_final_cltv_expiry was already set
to p->blindedpay->cltv_expiry_delta above.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We don't actually support it yet, but this threads through the type change,
puts it in "decode" etc.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
The `chainlag` is defined as the positive difference between the
height of the last block processed by the node and the best height
known by the bitcoin backend. The chainlag is positive when we are
still catching up with the blockchain, and `0` otherwise.
The `chainlag` is used as an additional offset to the CLTV values when
sending payments, allowing payments to be sent even before the chain
sync completes.
Changelog-Fixed: Plugins: pay now correctly estimates channel capacity
after payment failure, fixing some retry cases.
The `estimated_capacity` was properly substracted from the channel
hint, but adding the amount back did not work. The amount was added back
and then immediately substracted again. This caused the
`estimated_capacity` to ever decrease. This commit makes sure re-adding
the amount to the `estimated_capacity` works as expected.
It's a u64, we should pass by copy. This is a big sweeping change,
but mainly mechanical (change one, compile, fix breakage, repeat).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
This has the benefit of being shorter, as well as more reliable (you
will get a link error if we can't print it, not a runtime one!).
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
I noticed that run-route-infloop chose some worse-looking paths after
routing was fixed, eg the second node:
Before:
Destination node, success, probability, hops, fees, cltv, scid...
02b3aa1e4ed31be83cca4bd367b2c01e39502cb25e282a9b4520ad376a1ba0a01a,1,0.991856,2,1004,40,2572260x39x0/1,2131897x45x0/0
After:
Destination node, success, probability, hops, fees, cltv, scid...
02b3aa1e4ed31be83cca4bd367b2c01e39502cb25e282a9b4520ad376a1ba0a01a,1,0.954540,3,1046,46,2570715x21x0/1,2346882x26x14/1,2131897x45x0/0
This is because although the final costs don't reflect it, routing was taking
into account local channels, and 2572260x39x0/1 has a base fee of 2970.
There's an easy fix: when we the pay plugin creates localmods for our
gossip graph, add all local channels with delay and fees equal to 0.
We do the same thing in our unit test. This improves things across
the board:
Linear success probability (when found): min-max(mean +/- stddev)
Before: 0.487040-0.999543(0.952548+/-0.075)
After: 0.486985-0.999750(0.975978+/-0.053)
Hops:
Before: 1-5(2.98374+/-0.77)
After: 1-5(2.09593+/-0.63)
Fees:
Before: 0-50848(922.457+/-2.7e+03)
After: 0-50041(861.621+/-2.7e+03)
Delay (blocks):
Before: 0-196(65.8081+/-60)
After: 0-190(60.3285+/-60)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Changelog-Changed: Plugins: `pay` route algorithm doesn't bias against our own "expensive" channels any more.
The "path_score" callback was supposed to evaluate the *entire path*,
but that was counter-intuitive and opened the door to a cost function
bug which caused this path cost to be less than the closer path.
In particular, the capacity bias code didn't understand this at all.
1. Rename the function to `channel_score` and remove the "distance"
parameter (always "1" since you're supposed to be evaluating a
single hop).
2. Rename "cost" to the more specific "fee": "score" is our
actual cost function result (we avoid the word "cost" as it
may get confused with satoshi amounts).
3. For capacity biassing, we do want to know the amount, but
explicitly hand that as a separate parameter "total".
4. Fix a minor bug where total handed to scoring function previously
included channel fee (this is wrong: fee is paid before sending into
channel).
5. Remove the now-unused total_delay member from the dijkstra
struct.
Here are the results of our test now (routing 4194303 msat, which
didn't crash the old code, so we could compare). In both cases
we could find routes to 615 nodes:
Linear success probability (when found): min-max(mean +/- stddev)
Before: 0.484764-0.999750(0.9781+/-0.049)
After: 0.487040-0.999543(0.952548+/-0.075)
Hops:
Before: 1-5(2.13821+/-0.66)
After: 1-5(2.98374+/-0.77)
Fees:
Before: 0-50041(2173.75+/-5.3e+03)
After: 0-50848(922.457+/-2.7e+03)
Delay (blocks):
Before: 0-294(83.1642+/-68)
After: 0-196(65.8081+/-60)
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Fixes: https://github.com/ElementsProject/lightning/issues/7092
Changelog-Fixed: Plugins: `pay` would occasionally crash on routing.
Changelog-Fixed: Plugins: `pay` route algorithm fixed and refined to balance fees and capacity far better.
It gave 0. A lot.
Firstly, rmsat was often very small, because delays are often small. Much
smaller than the actual fee.
We really just want to offset the bias and multiply it.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
We attempted to introduce a "capacity bias" so we would penalize
channels where we were using a large amount of their total capacity.
Unfortunately, this was both naive, and applied wrongly:
-log((capmsat + 1 - amtmsat) / (capmsat + 1));
As an integer gives 0 up to about 65% capacity. We then applied this
by multiplying:
(cmsat * rmsat * bias) / (cmsat + rmsat + bias + 1);
Giving a total score of 0 in these cases! If we're using the
arithmetic mean we really need to add 1 to bias. We might as well
use a double the whole way through, for a slightly more fine-grained
approach.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
So far we would call `preapproveinvoice` once for each payment split,
i.e., at least once per HTLC, and potentially more often. There is no
point in doing so repeatedly, and especially in remote signer setup
this is actually counterproductive due to the additional roundtrips.
Changelog-Changed pay: Improved performance by removing repeated `preapproveinvoice` calls
CI revealed one:
```
cc plugins/libplugin-pay.c
plugins/libplugin-pay.c: In function ‘payment_getroute’:
plugins/libplugin-pay.c:888:17: error: ‘errstr’ may be used uninitialized [-Werror=maybe-uninitialized]
888 | payment_fail(p, "%s", errstr);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~
plugins/libplugin-pay.c:851:21: note: ‘errstr’ was declared here
851 | const char *errstr;
| ^~~~~~
cc1: all warnings being treated as errors
```
My local compiler gave another:
```
channeld/channeld.c: In function ‘resume_splice_negotiation’:
channeld/channeld.c:3734:23: error: ‘final_tx’ may be used uninitialized [-Werror=maybe-uninitialized]
3734 | msg = towire_channeld_splice_confirmed_signed(tmpctx, final_tx,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
3735 | chan_output_index);
| ~~~~~~~~~~~~~~~~~~
channeld/channeld.c:3461:28: note: ‘final_tx’ was declared here
3461 | struct bitcoin_tx *final_tx;
| ^~~~~~~~
cc1: all warnings being treated as errors
make: *** [Makefile:298: channeld/channeld.o] Error 1
```
So fix both.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
