Matt Corallo [Tue, 20 Aug 2024 02:22:22 +0000 (02:22 +0000)]
Add a test of gossip message buffer limiting in `PeerManager`
This adds a simple test that the gossip message buffer in
`PeerManager` is limited, including the new behavior of bypassing
the limit when the broadcast comes from the
`ChannelMessageHandler`.
Matt Corallo [Tue, 20 Aug 2024 01:57:06 +0000 (01:57 +0000)]
Add a constructor for the test `SocketDescriptor` and `hang_writes`
In testing, its useful to be able to tell the `SocketDescriptor` to
pretend the system network buffer is full, which we add here by
creating a new `hang_writes` flag. In order to simplify
constructing, we also add a new constructor which existing tests
are moved to.
Matt Corallo [Mon, 24 Jun 2024 20:24:36 +0000 (20:24 +0000)]
Reliably deliver gossip messages from our `ChannelMessageHandler`
When our `ChannelMessageHandler` creates gossip broadcast
`MessageSendEvent`s, we generally want these to be reliably
delivered to all our peers, even if there's not much buffer space
available.
Here we do this by passing an extra flag to `forward_broadcast_msg`
which indicates where the message came from, then ignoring the
buffer-full criteria when the flag is set.
Matt Corallo [Mon, 24 Jun 2024 20:21:08 +0000 (20:21 +0000)]
Use a `MessageSendEvent`-handling fn rather than a single lopp
Rather than building a single `Vec` of `MessageSendEvent`s to
handle then iterating over them, we move the body of the loop into
a lambda and run the loop twice. In some cases, this may save a
single allocation, but more importantly it sets us up for the next
commit, which needs to know from which handler the
`MessageSendEvent` it is processing came from.
Matt Corallo [Thu, 12 Sep 2024 15:17:15 +0000 (15:17 +0000)]
Call `ChannelMessageHandler::message_received` without peer lock
While `message_received` purports to be called on every message,
prior to the message, doing so on `Init` messages means we have to
call `message_received` while holding the per-peer mutex, which
can cause some lock contention.
Instead, here, we call `message_received` after processing `Init`
messages (which is probably more useful anyway - the peer isn't
really "connected" until we've processed the `Init` messages),
allowing us to call it unlocked.
Matt Corallo [Thu, 12 Sep 2024 15:13:11 +0000 (15:13 +0000)]
Check that we aren't reading a second message in BOLT 12 retry test
`creates_and_pays_for_offer_with_retry` intends to check that we
re-send a BOLT 12 `invoice_request` in response to a
`message_received` call, but doesn't actually test that there were
no messages in the outbound buffer after the initial send, which we
do here.
Matt Corallo [Wed, 11 Sep 2024 23:36:29 +0000 (23:36 +0000)]
Validate `channel_update` signatures without holding a graph lock
We often process many gossip messages in parallel across different
peer connections, making the `NetworkGraph` mutexes fairly
contention-sensitive (not to mention the potential that we want to
send a payment and need to find a path to do so).
Because we need to look up a node's public key to validate a
signature on `channel_update` messages, we always need to take a
`NetworkGraph::channels` lock before we can validate the message.
For simplicity, and to avoid taking a lock twice, we'd always
validated the `channel_update` signature while holding the same
lock, but here we address the contention issues by doing a
`channel_update` validation in three stages.
First we take a read lock on `NetworkGraph::channels` and check if
the `channel_update` is new, then release the lock and validate the
message signature, and finally take a write lock, (re-check if the
`channel_update` is new) and update the graph.
DefaultRouter::create_blinded_payment_paths may creat a one-hop blinded
path with the recipient as the introduction node. Update the privacy
section of DefaultRouter's docs to indicate this as is done in the docs
for DefaultMessageRouter.
ChannelManager is parameterized by a Router, which must also implement
MessageRouter. Instead, add a MessageRouter parameter such that the
Router and MessageRouter traits can be de-coupled. This simplifies using
something other than DefaultMessageRouter, which DefaultRouter currently
delegates to.
Matt Corallo [Thu, 12 Sep 2024 18:36:58 +0000 (18:36 +0000)]
Add a `PaymentId` for inbound payments
We expect our users to have fully idempotent `Event` handling as we
may replay events on restart for one of a number of reasons. This
isn't a big deal as long as all our events have some kind of
identifier users can use to check if the `Event` has already been
handled.
For outbound payments, this is the `PaymentId` they provide in the
send methods, however for inbound payments we don't have a great
option.
`PaymentHash` largely suffices - users can simply always claim in
response to a `PaymentClaimable` of sufficient value and treat a
`PaymentClaimed` event as duplicate any time they see a second one
for the same `PaymentHash`. This mostly works, but may result in
accepting duplicative payments if someone (incorrectly) pays twice
for the same `PaymentHash`.
Users could also fail for duplicative `PaymentClaimable` events of
the same `PaymentHash`, but doing so may result in spuriously
failing a payment if the `PaymentClaimable` event is a replay and
they never saw a corresponding `PaymentClaimed` event.
While none of this will result in spuriously thinking they've been
paid when they have not, it does result in some pretty awkward
semantics which we'd rather avoid our users having to deal with.
Instead, here, we add a new `PaymentId` which is simply an HMAC of
the HTLCs (as Channel ID, HTLC ID pairs) which were included in the
payment.
Matt Corallo [Sun, 8 Sep 2024 16:38:22 +0000 (16:38 +0000)]
Add an `inbound_payment_id_secret` to `ChannelManager`
In the next commit we'll start generating `PaymentId`s for inbound
payments randomly by HMAC'ing the HTLC set of the payment. Here we
start by defining the HMAC secret for these HMACs.
This requires one small test adaptation and a full_stack_target
fuzz change because it changes the RNG consumption.
Matt Corallo [Sun, 8 Sep 2024 16:09:12 +0000 (16:09 +0000)]
Do not check the ordering of HTLCs in `PaymentClaim[able,ed]`
In the next commit we'll change the order of HTLCs in
`PaymentClaim[able,ed]` events. This shouldn't break anything, but
our current functional tests check that the HTLCs are provided in
the order they expect (the order they were received). Instead, here
we only validate that each claimed HTLC matches one expected path.
Matt Corallo [Wed, 18 Sep 2024 18:03:11 +0000 (18:03 +0000)]
Only attempt to `rustfmt` files checked into git
This avoids `rustfmt` failing on Rust files generated by dependent
crates in `target`, eg
```
+ rustfmt --edition 2021 --check ./target/debug/build/thiserror-8230374e07b5c05a/out/probe.rs
Diff in /home/matt/rust-lightning-3/target/debug/build/thiserror-8230374e07b5c05a/out/probe.rs at line 1:
Elias Rohrer [Wed, 4 Sep 2024 09:10:13 +0000 (11:10 +0200)]
Check workspace members with default features individually in CI
Previously, we would only check the workspace as a whole. This however
would mean that we would check/test crates with `lightning`'s default
features enabled, allowing failures-to-build under the crates own
default features to slip through, if they didn't explicitly enable
`lightning/std`, for example.
Here, we extend the CI to check the workspace as a whole but then run
checks, tests, and doc generation on the workspace members individually,
asserting that all of them build even when not built as part of the same
workspace as `lightning`.
Fix bug where we double-pay an offer due to stale manager
This fixes the following bug:
- An outbound payment is AwaitingInvoice
- We receive an invoice and lock the HTLCs into the relevant ChannelMonitors
- The monitors are successfully persisted, but the ChannelManager fails to
persist, so the outbound payment remains AwaitingInvoice
- We restart, causing the channels to close due to a stale ChannelManager
- We receive a duplicate invoice, and attempt to pay it again due to the
payment still being AwaitingInvoice in the stale ChannelManager
After the fix for this, we will notice that the payment is already locked into
the monitor on startup and transition the incorrectly-AwaitingInvoice payment
to Retryable, which prevents double-paying on duplicate invoice receipt.
Jeffrey Czyz [Thu, 1 Aug 2024 15:56:51 +0000 (10:56 -0500)]
Rename Offer::signing_pubkey to Offer::issuer_signing_pubkey
The spec was recently changed to use offer_issuer_id instead of
offer_node_id. LDK always used signing_pubkey to avoid confusion with a
node_id. Rename it to issuer_signing_pubkey now as InvoiceRequest and
Bolt12Invoice will have similarly named methods in upcoming commits.
Move the code that ensures that HTLCs locked into ChannelMonitors are
synchronized with the ChannelManager's OutboundPayments store to the
outbound_payments module.
This is useful both because ChannelManager::read is very long/confusing method,
so it's nice to encapsulate some of its functionality, and because we need to
fix an existing bug in this logic where we may risk double-paying an offer due
to outbound_payments being stale on startup. See the next commit for this
bugfix.
Currently we don't have any visibility if BackgroundProcessor
takes considerably more of time to process events, adding logs
to help debug such issues.
While these variants may sound similar, they are very different. One is so
temporary it's never even persisted to disk, the other is a state we will stay
in for hours or days. See added docs for more info.
Add new Bolt12PaymentError for failed blinded path creation.
Currently used when initiating an async payment via held_htlc_available OM. This
OM needs a reply path back to us, so use this error for our invoice_error OM if
we fail to create said reply path.
See AsyncPaymentsContext::hmac, but this prevents the recipient from
deanonymizing us. Without this, if they are able to guess the correct payment
id, then they could create a blinded path to us and confirm our identity.
We also move the PAYMENT_HASH_HMAC_INPUT const to use &[7; 16], which is safe
because this const was added since the last release. This ordering reads more
smoothly.