This commit fixes a concurrency bug between the server’s htlc switch,
and an instance of a peer’s active htlc manager goroutine.
The deadlock condition was triggered when the htlc manager’s downstream
`htlcPacket` channel was full, causing the htlcSwitch’s main loop to
block on a send while the htlcManager was attempting to send a packet
to the switch for forwarding. This scenario created a circular
dependency resting in a deadlock.
The fix for this bug is relatively straight forward, if the destination
interface is found when handling an outgoing payment, then a new
goroutine is spawned to complete the request.
This commit updates the build, installation, and upgrade instructions to use
Glide which allows for reproducible builds using Golang 1.5 or 1.6.
As a result, if one wishes to ensure they've properly installed lnd, the `go
get` command can no longer be used to install lnd.
This commit adds glide integration in order to make lnd builds fully
reproducible. Rather than using “go get” users should now manually pull down
the repo, use glide to fetch+install the dependancies, then manually install
all related binaries.
With this change we no longer have to chase dependancies making breaking API
changes under us. We can manually update the managed dependancies once a new
stable release of a defendant package is released.
Additionally, reproducible builds are a strong requirement in order to securely
distribute future major releases of lnd.
This commit *significantly increases* the payment throughput per-core,
per-channel of the daemon.
With this commit updates are properly pipelined respecting the current
revocation window, htlc updates are batched, a timer is checked to push
chain convergence, and htlc update below the batch size are
periodically flushed to the remote chain.
The current pending update timer, trickle timer, and batch size have
been arbitrarily chosen based on my local tests. In the future these
parameters should be chosen to optimize response-time and throughput
after measurements are gathered.
This commit adds a new method, “PendingUpdates” to the channel state
machine which is intended to be a source to give callers a hint as to
when an additional commitment signature should be sent independent of
any request/response book keeping.
This commit patrons the state update logs properly within the channel
state machine. This change fixes a number of bugs caused by treating a
central log as two logically distinct logs. Rather than having a bit
indicating if the entry is incoming/outgoing, an entry is added to a
remote or local log depending on which modification method is used.
As a result the code is much easier to follow due to separation of
concerts.
Finally, when attempting to sign a new update with an exhausted
renovation window a distinct error is returned in order to allow higher
level callers to properly back-off and handle the protocol event.
This commit adds a ticker which attempts to print the total volume
sent/recv over the last 10 seconds if any updates took place during
that interval.
Additionally, when selecting a link to forward an htlcPacket insert a
break after selection in order to properly terminate the loop.
We now close the send payment stream after receiving a response.
Otherwise if the RPC server is asynchronously handling requests then,
closing the stream would result in the server returning an EOF error,
terminating the request.
This commit changes the current behavior around channeldb.Wipe().
Previously if a channel had never been closed, and a wipe was
attempted, then wipe operation would fail and the transaction would be
rolled back.
This commit fixes this behavior by checking for bolt.ErrBucketNotFound
error, and handling the specific error as a noop.
With this commit, calls to htlcSwitch.SendHTLC() are now synchronous,
only returning after the payment has been fully settled. This will
allow one to accurately measure the commitment update speed with the
current state machine implementation which is missing a number of
low-hanging optimizations.
The htlcManager for each channel now keeps a map of cleared HTLC’s
keyed by the index number of the add entry within the state machine’s
HTLC log. This map of HTLC’s will later be used to properly implement
time outs
Additionally, a slight refactoring has been executed w.r.t handling
upstream/downstream messages. This cleans up the main htlcManager loop,
freeing it up for the addition of future logic to properly observe
timeouts as well as, proper batching+trickling of HTLC updates, and a
commitment signature ticker.
This commit fixes a slight bug in the channel state machine’s code
executed when processing a revocation messages. With this commit after
processing a revocation, log entries which we should forward to the
downstream or upstream peer for settling/adding HTLC’s are now properly
returned.
The testa have also been updated to ensure to correct htlc’s are
returned “for forwarding”.
This commit adds two Dockerfiles, along with a docker-compose file
which links the two docker files together allowing for single-command
deployment.
Using the docker-compose file, two containers are deployed. One running
btcd, and the other running lnd. Both containers share the same shared
volume mounted to the file system in order to allow land to read btcd’s
certificates for the TLS RPC connections.
Additionally, the btcd instance comes will an automatic RPC
configuration generated allowing one to use btcctl out of the box via
calls to “docker-compose exec btcctl …”.
With this commit, support for changing the target network (testnet,
simnet, etc) has been finalized. Previously a command line option was
present to swap networks, but the RPC port wouldn’t automatically be
updated to reflect the network.
With this commitment, the daemon is now able to properly send+redeem
single-hop HTLC’s with another daemon running on the same network.
The htlcManager gains an additional channel which is reads from
receiving updates from either downstream peers, or the rpc server. An
htlcManager is spawned for each active channel with the remote peer. As
a result, the readHandler must now de-multiplex any messages which
update a known channel to the proper htlcManager.
Batching HTLC add updates with a trickle timer has not yet been
implemented, but will be in the near future along with several other
optimizations.
We no longer track HTLC’s by their r-hash within the log into the
index, as we may have multiple HTLC’s that can be redeemed by the same
pre-image. Instead we now use a separate index which is keyed by a
log-index.
Additionally, the SettleHTLC method now also returns the index of the
HTLC being settled which allows the remote party to quickly locate the
HTLC within their log.
This commit also introduces a few trace/debug log messages which will
likely be pruned in the near future
This commit introduces the invoice registry which is a central
repository of all outstanding invoices related to the daemon.
This registry will be used by the goroutines which manage the htlc’s
for a particular channel, and later by the point-to-point workflow
which negotiates the conditions for payment along with signed recipes.
This commit adds a new nested main function “lndMain”, within the
packages’s normal main function. This nesting is required in order to
properly execute all queued defer statements in the case of a forced
exit.
This commit fixes a bug caused by overriding the prefix key for storing
commitment keys with the first few bytes of a channel’s channel point.
Once a channel was deleted, then all future channels would result in a
panic due to a nil pointer deference since the prefix key was mutated,
causing all future stores/gets to fail.
Each active channel now gains its a dedicated htlcManager goroutine
which currently accepts to two golang channels, and a lightning
channel. The “downstream” channel will be used for dispatched multi-hop
payments sent from the htlcSwitch, while the “upstream” channel will be
used once the readHandler de-multiplexes hltc add/timeout/settle
messages.
Each time a new channel is fully created after N confirmations, the
peer’s channelManager registers the new link with the htlcSwitch. Once
the channel is closed either cooperatively or non-cooperatively, then
the link is unregistered.
This commit introduces the HtlcSwitch service into the daemon. The
purpose of the switch is to manage the message passing between the
goutiness within he peer managing each active channel. The switch
treats connected peers as interfaces, and their active channels as
links. All multi-hop, single-hop, and channel closure orchestration
will be managed by various goroutines within the HtlcSwitch.
The HtlcSwitch will also serve as the L2 contact point between the
yet-to-be-written LightningRouter and the switch, forwarding the new
htlc packet on the proper interface. Additionally, the switch will
handle dispatching htlc’s created locally via RPC, and multi-hop
forwarding initiated by peers.
