Files
tailscale/net/traffic/traffic.go
Amal Bansode fd07b9a2b6 net/traffic: fix rendezvous hashing (#20828)
The rendezvous hasher for traffic steering loadbalancing was flawed.
By plainly using the FNV-1a hash value, the result often reflected the
magnitude of the most significant bits in the hash seed, meaning the
hash function was not diffusive (aka missing the Avalanche Effect).

Popular wisdom seems to be that the output of FNV-1a should be mixed
with some large numbers to perturb more output bits. Borrow concepts
from other (Rust, Java) libraries by using the mix13 variant of 64-bit
finalizers by David Stafford.

Modify the fuzz test that asserts this fairness. Adjust a few
constants like client count and candidate count to more closely
reflect real-world scenarios and practical probabilities. Tighten
the bounds for distribution from 50% to +-20%.

Updates tailscale/corp#46471

Signed-off-by: Amal Bansode <amal@tailscale.com>
2026-08-12 14:13:42 -07:00

149 lines
4.4 KiB
Go
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
// Copyright (c) Tailscale Inc & contributors
// SPDX-License-Identifier: BSD-3-Clause
// Package traffic contains helpers for evaluating traffic steering scores and
// picking appropriate nodes.
package traffic
import (
"cmp"
"encoding/binary"
"hash/fnv"
"iter"
"maps"
"slices"
"tailscale.com/tailcfg"
"tailscale.com/util/mak"
)
// Score is a nodes traffic score, where any int could be a valid score.
// A higher traffic score suggests that the client should prefer that peer
// over one with a lower traffic score.
type Score int
// Scores is a memoization cache for the traffic scores of the current nodes peers.
type Scores struct {
self tailcfg.NodeID
hash NodeHasher
scores map[tailcfg.NodeID]Score
}
// ScoresFor returns a new [Scores] cache for the current nodes ID,
// after scoring the peer nodes and adding these scores to the cache.
func ScoresFor(self tailcfg.NodeID, peers []tailcfg.NodeView) Scores {
ss := Scores{
self: self,
hash: MakeRendezvousHasher(self),
}
ss.ScorePeers(peers)
return ss
}
// IsValid reports whether ss has been initialized with the current node ID.
func (ss Scores) IsValid() bool {
return !ss.self.IsZero()
}
// Score scores the given peer node and returns it after adding the score to the cache.
func (ss *Scores) Score(n tailcfg.NodeView) Score {
id := n.ID()
if s, ok := ss.scores[id]; ok {
return s
}
var s Score
if hi := n.Hostinfo(); hi.Valid() {
if loc := hi.Location(); loc.Valid() {
s = Score(loc.Priority())
}
}
mak.Set(&ss.scores, id, s)
return s
}
// ScorePeers scores the peer nodes and adds these scores to the cache.
func (ss *Scores) ScorePeers(peers []tailcfg.NodeView) {
if len(peers) == 0 {
return
}
if ss.scores == nil {
ss.scores = make(map[tailcfg.NodeID]Score, len(peers))
}
for _, n := range peers {
ss.Score(n)
}
}
// All returns an iterator over the scores for every peer in the cache.
// The iteration order is not specified and is not guaranteed to be the same
// from one call to the next.
func (ss Scores) All() iter.Seq2[tailcfg.NodeID, Score] {
return maps.All(ss.scores)
}
// SortNodes sorts the slice of nodes in descending order of [Scores.Score],
// using rendezvous hashing to break ties when both nodes have the same score.
// After sorting, the zeroth element is the preferred node.
func (ss Scores) SortNodes(nodes []tailcfg.NodeView) {
slices.SortFunc(nodes, func(a, b tailcfg.NodeView) int {
c := cmp.Compare(ss.Score(b), ss.Score(a)) // Highest score first.
if c == 0 {
return ss.hash.Compare(b.ID(), a.ID()) // Descending order.
}
return c
})
}
// NodeHasher returns a 64-bit hash of a node ID.
type NodeHasher func(tailcfg.NodeID) uint64
// MakeRendezvousHasher returns a function that hashes a node ID to a uint64.
// https://en.wikipedia.org/wiki/Rendezvous_hashing
func MakeRendezvousHasher(seed tailcfg.NodeID) NodeHasher {
en := binary.BigEndian
return func(n tailcfg.NodeID) uint64 {
var b [16]byte
en.PutUint64(b[:], uint64(seed))
en.PutUint64(b[8:], uint64(n))
// FNV-1a is more modern and distributes bits more evenly,
// so it is recommended by the designers.
//
// Note that we dont use a global hasher and h.Reset
// because this closure could be called concurrently.
// This is cheap because hash/fnv doesnt need to allocate.
h := fnv.New64a()
h.Write(b[:])
v := h.Sum64()
// After FNV-1a, finalize the result by mixing in some large
// numbers. This ensures a small change in the seed/input bits
// causes a large perturbation in the output bits, aka the
// "Avalanche Effect".
// We opted to use the mix13 variant described by David Stafford,
// a popular choice in other language libraries.
// https://web.archive.org/web/20260406221046/https://zimbry.blogspot.com/2011/09/better-bit-mixing-improving-on.html
v ^= v >> 30
v *= 0xbf58476d1ce4e5b9
v ^= v >> 27
v *= 0x94d049bb133111eb
v ^= v >> 31
return v
}
}
// Compare compares the node ID hashes of peers a and b, using the same convention as [cmp.Compare].
// Since h is seeded with the current nodes ID, the ordering between a and b will remain stable
// for this node; but the order may flip for when h is seeded for another node.
// This function should return zero, if and only if a and b have the same node ID.
func (h NodeHasher) Compare(a, b tailcfg.NodeID) int {
c := cmp.Compare(h(a), h(b))
if c == 0 {
// In the unlikely event of a hash collision, compare the actual IDs.
return cmp.Compare(a, b)
}
return c
}