Include tests for requires_unknown_bits and supports_unknown_bits when
an unknown even bit, odd bit, or neither is set. Refactor bit clearing
such that tests and production code share the same code path. Fix a
potential spec incompatibility (currently only exposed in testing code)
where trailing zero bytes are not removed after a bit is cleared.
Converting from InitFeatures to other Features is accomplished using
Features::with_known_relevant_init_flags. Define a more general
to_context method which converts from Features of one Context to
another.
Additionally, ensure the source context only has known flags before
selecting flags for the target context.
Refactoring the features module allowed for making code specific to
certain contexts generalizable. Specifically, KNOWN_FEATURE_MASK
is defined on Context instead of hardcoded in each method
specialization. Thus, such methods are no longer required.
Features for a given context are duplicated throughout the features
module. Use a macro for defining a Context and the applicable features
such that features only need to be defined for a Context in one place.
The Context provides bitmasks for selecting known and unknown feature
flags.
BOLT 1 and BOLT 9 refer to features as "known" if a peer understands
them. They also use the term "supported" to mean either optional or
required.
Update the features module to use similar terminology.
- Define contexts in terms of required and optional features rather than
just supported features
- Define known features as those that are optional or required
- Rename supported() constructor to known()
For completeness, clear_optional_bit for each feature is now called
clear_bits and clears both optional and required bits.
Each feature is represented by two bits within Features' flags field.
Working with these flags requires bitwise operations, which can be error
prone. Rather than directly checking and manipulating bits, encapsulate
the bits within each feature trait and provide mechanisms for doing so.
This removes the need to comment on which features correspond to bitwise
expressions since the expressions use feature trait identifiers instead.
With this approach, byte literals and expressions can be evaluated at
compile time still. However, for these cases, knowing which byte within
the flags that a feature corresponds to still must be determined by the
implementor.
Remove the special case where initial_routing_sync has no even bit. Now,
it (bit 2) is considered known by the implementation.
The initial_routing_sync feature is set by peer_handler whenever a full
sync of the network graph is desired. It is not explicitly set when
creating features with InitFeatures::supported().
An upcoming refactor will change supported() to known(), which will
return all features known by the implementation. Thus, the
initial_routing_sync flag will need to be set by default. This commit
makes the behavior change ahead of the refactor.
The test_upfront_shutdown_script functional test clears this feature
flag. However, the method used to clear the flag is implemented by bit
toggling. Thus, if the flag is not set the method would actually set it.
Implement the method using bit clearing instead.
Instead of adding signatures to LocalCommitmentTransactions, we
instead leave them unsigned and use them to construct signed
Transactions when we want them. This cleans up the guts of
LocalCommitmentTransaction enough that we can, and do, expose its
state to the world, allowing external signers to have a basic
awareness of what they're signing.
1107ab06c3 introduced an API to have a
ChannelKeys implementer sign HTLC transactions by calling into the
LocalCommitmentTransaction object, which would then store the tx.
This API was incredibly awkward, both because it required an
external signer trust our own internal interfaces, but also because
it didn't allow for any inspection of what was about to be signed.
Further, it signed the HTLC transactions one-by-one in a somewhat
inefficient way, and there isn't a clear way to resolve this (as
the which-HTLC parameter has to refer to something in between the
HTLC's arbitrary index, and its index in the commitment tx, which
has "holes" for the non-HTLC outputs and skips some HTLCs).
We replace it with a new function in ChannelKeys which allows us
to sign all HTLCs in a given commitment transaction (which allows
for a bit more effeciency on the signers' part, as well as
sidesteps the which-HTLC issue). This may also simplify the signer
implementation as we will always want to sign all HTLCs spending a
given commitment transaction at once anyway.
We also de-mut the LocalCommitmentTransaction passed to the
ChanKeys, instead opting to make LocalCommitmentTransaction const
and avoid storing any new HTLC-related data in it.
This cleans up sign_local_commitment somewhat by returning a
Result<Signaure, ()> over the local commitment transaction instead
of modifying the struct which was passed in.
This is the first step in making LocalCommitmentTransaction a
completely pub struct, using it just to communicate enough
information to the user to allow them to construct a signaure
instead of having it contain a bunch of logic.
This should make it much easier to implement a custom ChannelKeys
by disconnecting the local commitment transaction signing from our
own datastructures.
In e46e183084 we began tracking
whether a local commitment transaction had been signed and
broadcast in OnchainTxHandler, refusing to update the local
commitment transaction state in the ChannelMonitor on that basis.
This is fine, except that it doesn't make a lot of sense to store
the full local transaction state in OnchainTxHandler - we should be
providing it the unsigned local transaction at the time we wish to
broadcast and no more (just like we do all other transaction data).
As channel_value last usage was for computing feerate but as this
one is static per-commitment and will always-be following specification,
we remove it.
We want to avoid a third-party channel closure, where a random node
by sending us a payment expiring at current height, would trigger our
onchain logic to close the channel due to a near-expiration.
It seems we've recently been seeing sporadic long-running
full_stack_target cases when running honggfuzz in CI. These
shouldn't be killed (as its possible they hit an error or a deadlock,
especially since the longest-running tests probably have the most
coverage).
Relatively simple test that, after a monitor update fails, we get
the right return value and continue with the bits of the MPP that
did not send after the monitor updating is restored.
This is a key test for our automatic HTLC time-out logic, as it
ensures we don't allow an HTLC which indicates we should wait for
additional HTLCs before responding to cause us to force-close a
channel due to HTLC near-timeout.
expect_payment_failed!() was introduced after many of the tests
which could use it were written, so we take this opportunity to
switch them over now, increasing test coverage slightly by always
checking the payment hash expected.
We only do this for incoming HTLCs directly as we rely on channel
closure and HTLC-Timeout broadcast to fail any HTLCs which we
relayed onwards where our next-hop doesn't update_fail in time.
Same setup than Travis except for removing
`rm -f target/debug/lightning-*` as I do not believe
such file would exist on a fresh run.
I have not setup caching at this stage. The library is
small so I don't think it'd be that necessary/helpful.
I'd recommend to let both CI run for a bit to compare
performance and stability. The CI setup is straightforward
so I do not foresee any issue with GitHub actions.
Once happy, Travis file can be removed and branch
protection checks can be updated to block on the GitHub
actions.
You can also check the [Coverage report](752a58bc04/lightning/src/chain) to ensure it is as expected.
