When a block is disconnected due to a reorg, DisconnectBlockAtHeight
is called for the block height. Prior to this patch, it would delete
every SCID in the graph with a block height greater than the
disconnected height. This meant that a reorg would delete every
zero-conf channel edge from the graph. The fix simply iterates up
until the StartingAlias and deletes every SCID between the
disconnected height and the StartingAlias height.
feature-bit channels
This allows opening zero-conf chan-type, scid-alias chan-type, and
scid-alias feature-bit channels. scid-alias chan-type channels are
required to be private. Two paths are available for opening a zero-conf
channel:
* explicit chan-type negotiation
* LDK carve-out where chan-types are not used, LND is on the
receiving end, and a ChannelAcceptor is used to enable zero-conf
When a zero-conf channel is negotiated, the funding manager:
* sends a FundingLocked with an alias
* waits for a FundingLocked from the remote peer
* calls addToRouterGraph to persist the channel using our alias in
the graph. The peer's alias is used to send them a ChannelUpdate.
* wait for six confirmations. If public, the alias edge in the
graph is deleted and replaced (not atomically) with the confirmed
edge. Our policy is also read-and-replaced, but the counterparty's
policy won't exist until they send it to us.
When a scid-alias-feature channel is negotiated, the funding manager:
* sends a FundingLocked with an alias:
* calls addToRouterGraph, sends ChannelUpdate with the confirmed SCID
since it exists.
* when six confirmations occurs, the edge is deleted and re-inserted
since the peer may have sent us an alias ChannelUpdate that we are
storing in the graph.
Since it is possible for a user to toggle the scid-alias-feature-bit
to on while channels exist in the funding manager, care has been taken
to ensure that an alias is ALWAYS sent in the funding_locked message
if this happens.
An OptionalMsgField has been added that allows outside subsystems
to provide a short channel id we should insert into a ChannelUpdate
that we then sign and send to our peer.
When the gossiper receives a ChannelUpdate, it will query the
alias manager by the passed-in FindBaseByAlias function to determine
if the short channel id in the ChannelUpdate points to a known
channel. If this lookup returns an error, we'll fallback to using
the original id in the ChannelUpdate when querying the router.
The lookup and potential fallback must occur in order to properly
lock the multimutex, query the correct router channels, and rate
limit the correct short channel id. An unfortunate side effect of
receiving ChannelUpdates from our peer that reference on of our
aliases rather than the real SCID is that we must store this policy.
Yet it is not broadcast-able. Care has been taken to ensure the
gossiper does not broadcast *any* ChannelUpdate with an alias SCID.
The cachedNetworkMsg uses the new processedNetworkMsg struct. This
is necessary so that delete-and-reinsert in the funding manager
doesn't process a ChannelUpdate twice and end up in a deadlock since
the err chan is no longer being used.
This commit was previously split into the following parts to ease
review:
- 2d746f68: replace imports
- 4008f0fd: use ecdsa.Signature
- 849e33d1: remove btcec.S256()
- b8f6ebbd: use v2 library correctly
- fa80bca9: bump go modules
In this commit, we modify the implementation of ForEachChannel to
utilize the new kvdb method ForAll. This greatly reduces the number of
round-trips to the database needed to iterate over all channels
in the graph.
In case the channeldb package is used as a library in external tools, it
can be useful to allow read-only access to a DB. This allows such a
tool to access a DB even if not all migrations were executed, which can
be useful for recovery purposes.
To make it possible to even start the DB with a read-only backend, we
need to disable the automatic migration step.
This commit, adds a new ForEachNode method to the channel graph cache
that assumes the contents won't be modified. This is generally useful,
and will be used in a later commit to optimize some heavy RPC calls.
This commit partially reverts bf27d05a.
To avoid creating multiple database transactions during a single path
finding operation, we create an explicit transaction when the cached
graph is instantiated.
We cache the source node to avoid needing to look that up for every path
finding session.
The database transaction will be nil in case of the in-memory graph.
With this commit we forward the config option for disabling the channel
graph cache as a boolean to the channeldb. But we invert its meaning to
make the flag easier to understand.
Fixes#5830.
When a channel for a node is announced before the node itself is
announced on the network, it's possible that we have channels for a node
but no features defined yet. This was previously logged as a warning
which spammed the log unnecessarily.
With this commit we use an optimized version of the node iteration that
causes fewer memory allocations by only loading the part of the graph
node that we actually need to know for the cache.
To avoid the channel map needing to be re-grown while we fill the cache
initially, we might as well pre-allocate it with a somewhat sane value
to decrease the number of grow events.
To further separate the channel graph from the channel state, we
refactor the AddrsForNode method to use the graphs's public methods
instead of directly accessing any buckets. This makes sure that we can
have the channel state cached with just its buckets while not using a
kvdb level cache for the graph.
At the same time we refactor the graph's test to also be less dependent
upon the channel state DB.
Depends on btcsuite/btcwallet#757.
Pulls in the updated version of btcwallet and walletdb that have the DB
interface enhanced by their own View() and Update() methods with the
reset callback/closure supported out of the box. That way the global
package-level View() and Update() functions now become pure redirects.
In this commit, we add a new method that allows us to mark a channel as
being a zombie on the fly without needing to go through the normal
channel deletion process.
In this commit, we add strict zombie pruning as a config level param.
This allow us to add the option for those that want a tighter graph, and
not change the default composition of the channel graph for most users
over night.
In addition, we expand the test case slightly by testing that the self
node won't be pruned, but also that if there's a node with only a single
known stale edge, then both variants will prune that edge.
This change was largely motivated by an increase in high disk usage as a
result of channel update spam. With an in memory graph, this would've
gone mostly undetected except for the increased bandwidth usage, which
this doesn't aim to solve yet. To minimize the effects to disks, we
begin to rate limit channel updates in two ways. Keep alive updates,
those which only increase their timestamps to signal liveliness, are now
limited to one per lnd's rebroadcast interval (current default of 24H).
Non keep alive updates are now limited to one per block per direction.
Similarly as with kvdb.View this commits adds a reset closure to the
kvdb.Update call in order to be able to reset external state if the
underlying db backend needs to retry the transaction.
This commit adds a reset() closure to the kvdb.View function which will
be called before each retry (including the first) of the view
transaction. The reset() closure can be used to reset external state
(eg slices or maps) where the view closure puts intermediate results.
The explicit `bbolt` dep is gone, as we depend on `kvdb`, which is
actually `walletdb`, which has its own module that defines the proper
`bbolt` version.
In this commit, we migrate all the code in `channeldb` to only reference
the new `kvdb` package rather than `bbolt` directly.
In many instances, we need to add two version to fetch a bucket as both
read and write when needed. As an example, we add a new
`fetchChanBucketRw` function. This function is identical to
`fetchChanBucket`, but it will be used to fetch the main channel bucket
for all _write_ transactions. We need a new method as you can pass a
write transaction where a read is accepted, but not the other way around
due to the stronger typing of the new `kvdb` package.
Previously we would return nil features when we didn't have a node
announcement or a given node. With this change, we can always assume the
feature vector is populated during pathfinding.
