nim-libp2p/libp2p/protocols/pubsub/gossipsub/behavior.nim

617 lines
22 KiB
Nim

## Nim-LibP2P
## Copyright (c) 2021 Status Research & Development GmbH
## Licensed under either of
## * Apache License, version 2.0, ([LICENSE-APACHE](LICENSE-APACHE))
## * MIT license ([LICENSE-MIT](LICENSE-MIT))
## at your option.
## This file may not be copied, modified, or distributed except according to
## those terms.
import std/[tables, strutils, sequtils, sets, algorithm]
import random # for shuffle
import chronos, chronicles, metrics
import "."/[types, scoring]
import ".."/[pubsubpeer, peertable, timedcache, mcache, pubsub]
import "../rpc"/[messages]
import "../../.."/[peerid, multiaddress, utility, switch]
declareGauge(libp2p_gossipsub_cache_window_size, "the number of messages in the cache")
declareGauge(libp2p_gossipsub_peers_per_topic_mesh, "gossipsub peers per topic in mesh", labels = ["topic"])
declareGauge(libp2p_gossipsub_peers_per_topic_fanout, "gossipsub peers per topic in fanout", labels = ["topic"])
declareGauge(libp2p_gossipsub_peers_per_topic_gossipsub, "gossipsub peers per topic in gossipsub", labels = ["topic"])
declareGauge(libp2p_gossipsub_under_dlow_topics, "number of topics below dlow")
declareGauge(libp2p_gossipsub_under_dout_topics, "number of topics below dout")
declareGauge(libp2p_gossipsub_under_dhigh_above_dlow_topics, "number of topics below dhigh but above dlow")
declareGauge(libp2p_gossipsub_no_peers_topics, "number of topics without peers available")
declareCounter(libp2p_gossipsub_above_dhigh_condition, "number of above dhigh pruning branches ran", labels = ["topic"])
proc grafted*(g: GossipSub, p: PubSubPeer, topic: string) =
g.peerStats.withValue(p.peerId, stats):
var info = stats.topicInfos.getOrDefault(topic)
info.graftTime = Moment.now()
info.meshTime = 0.seconds
info.inMesh = true
info.meshMessageDeliveriesActive = false
# mgetOrPut does not work, so we gotta do this without referencing
stats.topicInfos[topic] = info
assert(g.peerStats[p.peerId].topicInfos[topic].inMesh == true)
trace "grafted", peer=p, topic
do:
g.initPeerStats(p)
g.grafted(p, topic)
proc pruned*(g: GossipSub, p: PubSubPeer, topic: string) =
let backoff = Moment.fromNow(g.parameters.pruneBackoff)
g.backingOff
.mgetOrPut(topic, initTable[PeerID, Moment]())
.mgetOrPut(p.peerId, backoff) = backoff
g.peerStats.withValue(p.peerId, stats):
if topic in stats.topicInfos:
var info = stats.topicInfos[topic]
if topic in g.topicParams:
let topicParams = g.topicParams[topic]
# penalize a peer that delivered no message
let threshold = topicParams.meshMessageDeliveriesThreshold
if info.inMesh and info.meshMessageDeliveriesActive and info.meshMessageDeliveries < threshold:
let deficit = threshold - info.meshMessageDeliveries
info.meshFailurePenalty += deficit * deficit
info.inMesh = false
# mgetOrPut does not work, so we gotta do this without referencing
stats.topicInfos[topic] = info
trace "pruned", peer=p, topic
proc handleBackingOff*(t: var BackoffTable, topic: string) =
let now = Moment.now()
var expired = toSeq(t.getOrDefault(topic).pairs())
expired.keepIf do (pair: tuple[peer: PeerID, expire: Moment]) -> bool:
now >= pair.expire
for (peer, _) in expired:
t.mgetOrPut(topic, initTable[PeerID, Moment]()).del(peer)
proc peerExchangeList*(g: GossipSub, topic: string): seq[PeerInfoMsg] =
var peers = g.gossipsub.getOrDefault(topic, initHashSet[PubSubPeer]()).toSeq()
peers.keepIf do (x: PubSubPeer) -> bool:
x.score >= 0.0
# by spec, larger then Dhi, but let's put some hard caps
peers.setLen(min(peers.len, g.parameters.dHigh * 2))
peers.map do (x: PubSubPeer) -> PeerInfoMsg:
PeerInfoMsg(peerID: x.peerId.getBytes())
proc handleGraft*(g: GossipSub,
peer: PubSubPeer,
grafts: seq[ControlGraft]): seq[ControlPrune] =
for graft in grafts:
let topic = graft.topicID
logScope:
peer
topic
trace "peer grafted topic"
# It is an error to GRAFT on a explicit peer
if peer.peerId in g.parameters.directPeers:
# receiving a graft from a direct peer should yield a more prominent warning (protocol violation)
warn "attempt to graft an explicit peer, peering agreements should be reciprocal", peer=peer.peerId, topic
# and such an attempt should be logged and rejected with a PRUNE
result.add(ControlPrune(
topicID: topic,
peers: @[], # omitting heavy computation here as the remote did something illegal
backoff: g.parameters.pruneBackoff.seconds.uint64))
let backoff = Moment.fromNow(g.parameters.pruneBackoff)
g.backingOff
.mgetOrPut(topic, initTable[PeerID, Moment]())
.mgetOrPut(peer.peerId, backoff) = backoff
peer.behaviourPenalty += 0.1
continue
if g.backingOff
.getOrDefault(topic)
.getOrDefault(peer.peerId) > Moment.now():
debug "attempt to graft a backingOff peer", peer=peer.peerId,
topic
# and such an attempt should be logged and rejected with a PRUNE
result.add(ControlPrune(
topicID: topic,
peers: @[], # omitting heavy computation here as the remote did something illegal
backoff: g.parameters.pruneBackoff.seconds.uint64))
let backoff = Moment.fromNow(g.parameters.pruneBackoff)
g.backingOff
.mgetOrPut(topic, initTable[PeerID, Moment]())
.mgetOrPut(peer.peerId, backoff) = backoff
peer.behaviourPenalty += 0.1
continue
# Notice this might not be necessary anymore
if peer.peerId notin g.peerStats:
g.initPeerStats(peer)
# not in the spec exactly, but let's avoid way too low score peers
# other clients do it too also was an audit recommendation
if peer.score < g.parameters.publishThreshold:
continue
# If they send us a graft before they send us a subscribe, what should
# we do? For now, we add them to mesh but don't add them to gossipsub.
if topic in g.topics:
if g.mesh.peers(topic) < g.parameters.dHigh or peer.outbound:
# In the spec, there's no mention of DHi here, but implicitly, a
# peer will be removed from the mesh on next rebalance, so we don't want
# this peer to push someone else out
if g.mesh.addPeer(topic, peer):
g.grafted(peer, topic)
g.fanout.removePeer(topic, peer)
else:
trace "peer already in mesh"
else:
trace "pruning grafting peer, mesh full", peer, score = peer.score, mesh = g.mesh.peers(topic)
result.add(ControlPrune(
topicID: topic,
peers: g.peerExchangeList(topic),
backoff: g.parameters.pruneBackoff.seconds.uint64))
else:
trace "peer grafting topic we're not interested in", topic
# gossip 1.1, we do not send a control message prune anymore
proc handlePrune*(g: GossipSub, peer: PubSubPeer, prunes: seq[ControlPrune]) =
for prune in prunes:
let topic = prune.topicID
trace "peer pruned topic", peer, topic
# add peer backoff
if prune.backoff > 0:
let
backoff = Moment.fromNow((prune.backoff + BackoffSlackTime).int64.seconds)
current = g.backingOff.getOrDefault(topic).getOrDefault(peer.peerId)
if backoff > current:
g.backingOff
.mgetOrPut(topic, initTable[PeerID, Moment]())
.mgetOrPut(peer.peerId, backoff) = backoff
trace "pruning rpc received peer", peer, score = peer.score
g.pruned(peer, topic)
g.mesh.removePeer(topic, peer)
# TODO peer exchange, we miss ambient peer discovery in libp2p, so we are blocked by that
# another option could be to implement signed peer records
## if peer.score > g.parameters.gossipThreshold and prunes.peers.len > 0:
proc handleIHave*(g: GossipSub,
peer: PubSubPeer,
ihaves: seq[ControlIHave]): ControlIWant =
if peer.score < g.parameters.gossipThreshold:
trace "ihave: ignoring low score peer", peer, score = peer.score
elif peer.iHaveBudget <= 0:
trace "ihave: ignoring out of budget peer", peer, score = peer.score
else:
# TODO review deduplicate algorithm
# * https://github.com/nim-lang/Nim/blob/5f46474555ee93306cce55342e81130c1da79a42/lib/pure/collections/sequtils.nim#L184
# * it's probably not efficient and might give preference to the first dupe
var deIhaves = ihaves.deduplicate()
for ihave in deIhaves.mitems:
trace "peer sent ihave",
peer, topic = ihave.topicID, msgs = ihave.messageIDs
if ihave.topicID in g.mesh:
for m in ihave.messageIDs:
let msgId = m & g.randomBytes
if msgId notin g.seen:
if peer.iHaveBudget > 0:
result.messageIDs.add(m)
dec peer.iHaveBudget
else:
return
# shuffling result.messageIDs before sending it out to increase the likelihood
# of getting an answer if the peer truncates the list due to internal size restrictions.
shuffle(result.messageIDs)
proc handleIWant*(g: GossipSub,
peer: PubSubPeer,
iwants: seq[ControlIWant]): seq[Message] =
if peer.score < g.parameters.gossipThreshold:
trace "iwant: ignoring low score peer", peer, score = peer.score
elif peer.iWantBudget <= 0:
trace "iwant: ignoring out of budget peer", peer, score = peer.score
else:
var deIwants = iwants.deduplicate()
for iwant in deIwants:
for mid in iwant.messageIDs:
trace "peer sent iwant", peer, messageID = mid
let msg = g.mcache.get(mid)
if msg.isSome:
# avoid spam
if peer.iWantBudget > 0:
result.add(msg.get())
dec peer.iWantBudget
else:
return
proc commitMetrics(metrics: var MeshMetrics) =
libp2p_gossipsub_under_dlow_topics.set(metrics.underDlowTopics)
libp2p_gossipsub_no_peers_topics.set(metrics.noPeersTopics)
libp2p_gossipsub_under_dout_topics.set(metrics.underDoutTopics)
libp2p_gossipsub_under_dhigh_above_dlow_topics.set(metrics.underDhighAboveDlowTopics)
libp2p_gossipsub_peers_per_topic_gossipsub.set(metrics.otherPeersPerTopicGossipsub, labelValues = ["other"])
libp2p_gossipsub_peers_per_topic_fanout.set(metrics.otherPeersPerTopicFanout, labelValues = ["other"])
libp2p_gossipsub_peers_per_topic_mesh.set(metrics.otherPeersPerTopicMesh, labelValues = ["other"])
proc rebalanceMesh*(g: GossipSub, topic: string, metrics: ptr MeshMetrics = nil) =
logScope:
topic
mesh = g.mesh.peers(topic)
gossipsub = g.gossipsub.peers(topic)
trace "rebalancing mesh"
# create a mesh topic that we're subscribing to
var
prunes, grafts: seq[PubSubPeer]
npeers = g.mesh.peers(topic)
if npeers < g.parameters.dLow:
if not isNil(metrics):
inc metrics[].underDlowTopics
trace "replenishing mesh", peers = npeers
# replenish the mesh if we're below Dlo
var candidates = toSeq(
g.gossipsub.getOrDefault(topic, initHashSet[PubSubPeer]()) -
g.mesh.getOrDefault(topic, initHashSet[PubSubPeer]())
).filterIt(
it.connected and
# avoid negative score peers
it.score >= 0.0 and
# don't pick explicit peers
it.peerId notin g.parameters.directPeers and
# and avoid peers we are backing off
it.peerId notin g.backingOff.getOrDefault(topic)
)
# shuffle anyway, score might be not used
shuffle(candidates)
# sort peers by score, high score first since we graft
candidates.sort(byScore, SortOrder.Descending)
# Graft peers so we reach a count of D
candidates.setLen(min(candidates.len, g.parameters.d - npeers))
trace "grafting", grafting = candidates.len
if candidates.len == 0:
if not isNil(metrics):
inc metrics[].noPeersTopics
else:
for peer in candidates:
if g.mesh.addPeer(topic, peer):
g.grafted(peer, topic)
g.fanout.removePeer(topic, peer)
grafts &= peer
else:
var meshPeers = toSeq(g.mesh.getOrDefault(topic, initHashSet[PubSubPeer]()))
meshPeers.keepIf do (x: PubSubPeer) -> bool: x.outbound
if meshPeers.len < g.parameters.dOut:
if not isNil(metrics):
inc metrics[].underDoutTopics
trace "replenishing mesh outbound quota", peers = g.mesh.peers(topic)
var candidates = toSeq(
g.gossipsub.getOrDefault(topic, initHashSet[PubSubPeer]()) -
g.mesh.getOrDefault(topic, initHashSet[PubSubPeer]())
).filterIt(
it.connected and
# get only outbound ones
it.outbound and
# avoid negative score peers
it.score >= 0.0 and
# don't pick explicit peers
it.peerId notin g.parameters.directPeers and
# and avoid peers we are backing off
it.peerId notin g.backingOff.getOrDefault(topic)
)
# shuffle anyway, score might be not used
shuffle(candidates)
# sort peers by score, high score first, we are grafting
candidates.sort(byScore, SortOrder.Descending)
# Graft peers so we reach a count of D
candidates.setLen(min(candidates.len, g.parameters.dOut))
trace "grafting outbound peers", topic, peers = candidates.len
for peer in candidates:
if g.mesh.addPeer(topic, peer):
g.grafted(peer, topic)
g.fanout.removePeer(topic, peer)
grafts &= peer
# get again npeers after possible grafts
npeers = g.mesh.peers(topic)
if npeers > g.parameters.dHigh:
if not isNil(metrics):
if g.knownTopics.contains(topic):
libp2p_gossipsub_above_dhigh_condition.inc(labelValues = [topic])
else:
libp2p_gossipsub_above_dhigh_condition.inc(labelValues = ["other"])
# prune peers if we've gone over Dhi
prunes = toSeq(g.mesh[topic])
# avoid pruning peers we are currently grafting in this heartbeat
prunes.keepIf do (x: PubSubPeer) -> bool: x notin grafts
# shuffle anyway, score might be not used
shuffle(prunes)
# sort peers by score (inverted), pruning, so low score peers are on top
prunes.sort(byScore, SortOrder.Ascending)
# keep high score peers
if prunes.len > g.parameters.dScore:
prunes.setLen(prunes.len - g.parameters.dScore)
# collect inbound/outbound info
var outbound: seq[PubSubPeer]
var inbound: seq[PubSubPeer]
for peer in prunes:
if peer.outbound:
outbound &= peer
else:
inbound &= peer
let
meshOutbound = prunes.countIt(it.outbound)
maxOutboundPrunes = meshOutbound - g.parameters.dOut
# ensure that there are at least D_out peers first and rebalance to g.d after that
outbound.setLen(min(outbound.len, max(0, maxOutboundPrunes)))
# concat remaining outbound peers
prunes = inbound & outbound
let pruneLen = prunes.len - g.parameters.d
if pruneLen > 0:
# Ok we got some peers to prune,
# for this heartbeat let's prune those
shuffle(prunes)
prunes.setLen(pruneLen)
trace "pruning", prunes = prunes.len
for peer in prunes:
trace "pruning peer on rebalance", peer, score = peer.score
g.pruned(peer, topic)
g.mesh.removePeer(topic, peer)
elif npeers > g.parameters.dLow and not isNil(metrics):
inc metrics[].underDhighAboveDlowTopics
# opportunistic grafting, by spec mesh should not be empty...
if g.mesh.peers(topic) > 1:
var peers = toSeq(g.mesh[topic])
# grafting so high score has priority
peers.sort(byScore, SortOrder.Descending)
let medianIdx = peers.len div 2
let median = peers[medianIdx]
if median.score < g.parameters.opportunisticGraftThreshold:
trace "median score below opportunistic threshold", score = median.score
var avail = toSeq(
g.gossipsub.getOrDefault(topic, initHashSet[PubSubPeer]()) -
g.mesh.getOrDefault(topic, initHashSet[PubSubPeer]())
)
avail.keepIf do (x: PubSubPeer) -> bool:
# avoid negative score peers
x.score >= median.score and
# don't pick explicit peers
x.peerId notin g.parameters.directPeers and
# and avoid peers we are backing off
x.peerId notin g.backingOff.getOrDefault(topic)
# by spec, grab only 2
if avail.len > 2:
avail.setLen(2)
for peer in avail:
if g.mesh.addPeer(topic, peer):
g.grafted(peer, topic)
grafts &= peer
trace "opportunistic grafting", peer
if not isNil(metrics):
if g.knownTopics.contains(topic):
libp2p_gossipsub_peers_per_topic_gossipsub
.set(g.gossipsub.peers(topic).int64, labelValues = [topic])
libp2p_gossipsub_peers_per_topic_fanout
.set(g.fanout.peers(topic).int64, labelValues = [topic])
libp2p_gossipsub_peers_per_topic_mesh
.set(g.mesh.peers(topic).int64, labelValues = [topic])
else:
metrics[].otherPeersPerTopicGossipsub += g.gossipsub.peers(topic).int64
metrics[].otherPeersPerTopicFanout += g.fanout.peers(topic).int64
metrics[].otherPeersPerTopicMesh += g.mesh.peers(topic).int64
trace "mesh balanced"
# Send changes to peers after table updates to avoid stale state
if grafts.len > 0:
let graft = RPCMsg(control: some(ControlMessage(graft: @[ControlGraft(topicID: topic)])))
g.broadcast(grafts, graft)
if prunes.len > 0:
let prune = RPCMsg(control: some(ControlMessage(
prune: @[ControlPrune(
topicID: topic,
peers: g.peerExchangeList(topic),
backoff: g.parameters.pruneBackoff.seconds.uint64)])))
g.broadcast(prunes, prune)
proc dropFanoutPeers*(g: GossipSub) =
# drop peers that we haven't published to in
# GossipSubFanoutTTL seconds
let now = Moment.now()
for topic in toSeq(g.lastFanoutPubSub.keys):
let val = g.lastFanoutPubSub[topic]
if now > val:
g.fanout.del(topic)
g.lastFanoutPubSub.del(topic)
trace "dropping fanout topic", topic
proc replenishFanout*(g: GossipSub, topic: string) =
## get fanout peers for a topic
logScope: topic
trace "about to replenish fanout"
if g.fanout.peers(topic) < g.parameters.dLow:
trace "replenishing fanout", peers = g.fanout.peers(topic)
if topic in g.gossipsub:
for peer in g.gossipsub[topic]:
if g.fanout.addPeer(topic, peer):
if g.fanout.peers(topic) == g.parameters.d:
break
trace "fanout replenished with peers", peers = g.fanout.peers(topic)
proc getGossipPeers*(g: GossipSub): Table[PubSubPeer, ControlMessage] {.gcsafe.} =
## gossip iHave messages to peers
##
libp2p_gossipsub_cache_window_size.set(0)
trace "getting gossip peers (iHave)"
let topics = toHashSet(toSeq(g.mesh.keys)) + toHashSet(toSeq(g.fanout.keys))
for topic in topics:
if topic notin g.gossipsub:
trace "topic not in gossip array, skipping", topicID = topic
continue
let mids = g.mcache.window(topic)
if not(mids.len > 0):
continue
var midsSeq = toSeq(mids)
libp2p_gossipsub_cache_window_size.inc(midsSeq.len.int64)
# not in spec
# similar to rust: https://github.com/sigp/rust-libp2p/blob/f53d02bc873fef2bf52cd31e3d5ce366a41d8a8c/protocols/gossipsub/src/behaviour.rs#L2101
# and go https://github.com/libp2p/go-libp2p-pubsub/blob/08c17398fb11b2ab06ca141dddc8ec97272eb772/gossipsub.go#L582
if midsSeq.len > IHaveMaxLength:
shuffle(midsSeq)
midsSeq.setLen(IHaveMaxLength)
let
ihave = ControlIHave(topicID: topic, messageIDs: midsSeq)
mesh = g.mesh.getOrDefault(topic)
fanout = g.fanout.getOrDefault(topic)
gossipPeers = mesh + fanout
var allPeers = toSeq(g.gossipsub.getOrDefault(topic))
allPeers.keepIf do (x: PubSubPeer) -> bool:
x.peerId notin g.parameters.directPeers and
x notin gossipPeers and
x.score >= g.parameters.gossipThreshold
var target = g.parameters.dLazy
let factor = (g.parameters.gossipFactor.float * allPeers.len.float).int
if factor > target:
target = min(factor, allPeers.len)
if target < allPeers.len:
shuffle(allPeers)
allPeers.setLen(target)
for peer in allPeers:
if peer notin result:
result[peer] = ControlMessage()
result[peer].ihave.add(ihave)
proc heartbeat*(g: GossipSub) {.async.} =
while g.heartbeatRunning:
try:
trace "running heartbeat", instance = cast[int](g)
# reset IWANT budget
# reset IHAVE cap
block:
for peer in g.peers.values:
peer.iWantBudget = IWantPeerBudget
peer.iHaveBudget = IHavePeerBudget
g.updateScores()
var meshMetrics = MeshMetrics()
for t in toSeq(g.topics.keys):
# remove expired backoffs
block:
handleBackingOff(g.backingOff, t)
# prune every negative score peer
# do this before relance
# in order to avoid grafted -> pruned in the same cycle
let meshPeers = g.mesh.getOrDefault(t)
var prunes: seq[PubSubPeer]
for peer in meshPeers:
if peer.score < 0.0:
trace "pruning negative score peer", peer, score = peer.score
g.pruned(peer, t)
g.mesh.removePeer(t, peer)
prunes &= peer
if prunes.len > 0:
let prune = RPCMsg(control: some(ControlMessage(
prune: @[ControlPrune(
topicID: t,
peers: g.peerExchangeList(t),
backoff: g.parameters.pruneBackoff.seconds.uint64)])))
g.broadcast(prunes, prune)
# pass by ptr in order to both signal we want to update metrics
# and as well update the struct for each topic during this iteration
g.rebalanceMesh(t, addr meshMetrics)
commitMetrics(meshMetrics)
g.dropFanoutPeers()
# replenish known topics to the fanout
for t in toSeq(g.fanout.keys):
g.replenishFanout(t)
let peers = g.getGossipPeers()
for peer, control in peers:
# only ihave from here
for ihave in control.ihave:
if g.knownTopics.contains(ihave.topicID):
libp2p_pubsub_broadcast_ihave.inc(labelValues = [ihave.topicID])
else:
libp2p_pubsub_broadcast_ihave.inc(labelValues = ["generic"])
g.send(peer, RPCMsg(control: some(control)))
g.mcache.shift() # shift the cache
except CancelledError as exc:
raise exc
except CatchableError as exc:
warn "exception ocurred in gossipsub heartbeat", exc = exc.msg,
trace = exc.getStackTrace()
for trigger in g.heartbeatEvents:
trace "firing heartbeat event", instance = cast[int](g)
trigger.fire()
await sleepAsync(g.parameters.heartbeatInterval)