You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
tailscale/net/art/stride_table.go

232 lines
7.8 KiB
Go

net/art: implement the stride table building block of ART A stride table is an 8-bit routing table implemented as an array binary tree, with a special tree updating function (allot) that enables lightning fast address lookups and reasonably fast insertion and deletion. Insertion, deletion and lookup are all allocation-free. Updates #7781 │ sec/op │ StrideTableInsertion/10/random_order 16.79n ± 2% StrideTableInsertion/10/largest_first 16.83n ± 1% StrideTableInsertion/10/smallest_first 16.83n ± 0% StrideTableInsertion/50/random_order 17.84n ± 1% StrideTableInsertion/50/largest_first 20.04n ± 1% StrideTableInsertion/50/smallest_first 16.39n ± 0% StrideTableInsertion/100/random_order 14.63n ± 0% StrideTableInsertion/100/largest_first 17.45n ± 4% StrideTableInsertion/100/smallest_first 12.98n ± 0% StrideTableInsertion/200/random_order 12.51n ± 4% StrideTableInsertion/200/largest_first 18.36n ± 3% StrideTableInsertion/200/smallest_first 9.609n ± 3% StrideTableDeletion/10/random_order 19.50n ± 1% StrideTableDeletion/10/largest_first 19.34n ± 0% StrideTableDeletion/10/smallest_first 19.43n ± 0% StrideTableDeletion/50/random_order 14.58n ± 1% StrideTableDeletion/50/largest_first 14.27n ± 2% StrideTableDeletion/50/smallest_first 15.51n ± 0% StrideTableDeletion/100/random_order 12.02n ± 3% StrideTableDeletion/100/largest_first 10.64n ± 0% StrideTableDeletion/100/smallest_first 13.21n ± 3% StrideTableDeletion/200/random_order 14.05n ± 4% StrideTableDeletion/200/largest_first 9.288n ± 5% StrideTableDeletion/200/smallest_first 18.51n ± 1% StrideTableGet 0.5010n ± 0% │ routes/s │ StrideTableInsertion/10/random_order 59.55M ± 2% StrideTableInsertion/10/largest_first 59.42M ± 1% StrideTableInsertion/10/smallest_first 59.43M ± 0% StrideTableInsertion/50/random_order 56.04M ± 1% StrideTableInsertion/50/largest_first 49.91M ± 1% StrideTableInsertion/50/smallest_first 61.00M ± 0% StrideTableInsertion/100/random_order 68.35M ± 0% StrideTableInsertion/100/largest_first 57.32M ± 3% StrideTableInsertion/100/smallest_first 77.06M ± 0% StrideTableInsertion/200/random_order 79.93M ± 4% StrideTableInsertion/200/largest_first 54.47M ± 3% StrideTableInsertion/200/smallest_first 104.1M ± 3% StrideTableDeletion/10/random_order 51.28M ± 1% StrideTableDeletion/10/largest_first 51.70M ± 0% StrideTableDeletion/10/smallest_first 51.48M ± 0% StrideTableDeletion/50/random_order 68.60M ± 1% StrideTableDeletion/50/largest_first 70.09M ± 2% StrideTableDeletion/50/smallest_first 64.45M ± 0% StrideTableDeletion/100/random_order 83.21M ± 3% StrideTableDeletion/100/largest_first 94.03M ± 0% StrideTableDeletion/100/smallest_first 75.69M ± 3% StrideTableDeletion/200/random_order 71.20M ± 5% StrideTableDeletion/200/largest_first 107.7M ± 5% StrideTableDeletion/200/smallest_first 54.02M ± 1% StrideTableGet 1.996G ± 0% Signed-off-by: David Anderson <danderson@tailscale.com>
2 years ago
// Copyright (c) Tailscale Inc & AUTHORS
// SPDX-License-Identifier: BSD-3-Clause
package art
import (
"bytes"
"fmt"
"io"
"math/bits"
"strconv"
"strings"
)
// strideEntry is a strideTable entry.
type strideEntry[T any] struct {
// prefixIndex is the prefixIndex(...) value that caused this stride entry's
// value to be populated, or 0 if value is nil.
//
// We need to keep track of this because allot() uses it to determine
// whether an entry was propagated from a parent entry, or if it's a
// different independent route.
prefixIndex int
// value is the value associated with the strideEntry, if any.
value *T
// child is the child strideTable associated with the strideEntry, if any.
child *strideTable[T]
}
// strideTable is a binary tree that implements an 8-bit routing table.
//
// The leaves of the binary tree are host routes (/8s). Each parent is a
// successively larger prefix that encompasses its children (/7 through /0).
type strideTable[T any] struct {
// entries is the nodes of the binary tree, laid out in a flattened array.
//
// The array indices are arranged by the prefixIndex function, such that the
// parent of the node at index i is located at index i>>1, and its children
// at indices i<<1 and (i<<1)+1.
//
// A few consequences of this arrangement: host routes (/8) occupy the last
// 256 entries in the table; the single default route /0 is at index 1, and
// index 0 is unused (in the original paper, it's hijacked through sneaky C
// memory trickery to store the refcount, but this is Go, where we don't
// store random bits in pointers lest we confuse the GC)
entries [lastHostIndex + 1]strideEntry[T]
// refs is the number of route entries and child strideTables referenced by
// this table. It is used in the multi-layered logic to determine when this
// table is empty and can be deleted.
refs int
}
const (
// firstHostIndex is the array index of the first host route. This is hostIndex(0/8).
firstHostIndex = 0b1_0000_0000
// lastHostIndex is the array index of the last host route. This is hostIndex(0xFF/8).
lastHostIndex = 0b1_1111_1111
)
// getChild returns the child strideTable pointer for addr (if any), and an
// internal array index that can be used with deleteChild.
func (t *strideTable[T]) getChild(addr uint8) (child *strideTable[T], idx int) {
idx = hostIndex(addr)
return t.entries[idx].child, idx
}
// deleteChild deletes the child strideTable at idx (if any). idx should be
// obtained via a call to getChild.
func (t *strideTable[T]) deleteChild(idx int) {
t.entries[idx].child = nil
t.refs--
}
// getOrCreateChild returns the child strideTable for addr, creating it if
// necessary.
func (t *strideTable[T]) getOrCreateChild(addr uint8) *strideTable[T] {
idx := hostIndex(addr)
if t.entries[idx].child == nil {
t.entries[idx].child = new(strideTable[T])
t.refs++
}
return t.entries[idx].child
}
net/art: implement the Table type, a multi-level art route table. Updates #7781 │ sec/op │ TableInsertion/ipv4/10 1.562µ ± 2% TableInsertion/ipv4/100 2.398µ ± 5% TableInsertion/ipv4/1000 2.097µ ± 3% TableInsertion/ipv4/10000 2.756µ ± 4% TableInsertion/ipv4/100000 2.473µ ± 13% TableInsertion/ipv6/10 7.649µ ± 2% TableInsertion/ipv6/100 12.09µ ± 3% TableInsertion/ipv6/1000 14.84µ ± 5% TableInsertion/ipv6/10000 14.72µ ± 8% TableInsertion/ipv6/100000 13.23µ ± 41% TableDelete/ipv4/10 378.4n ± 5% TableDelete/ipv4/100 366.9n ± 3% TableDelete/ipv4/1000 418.6n ± 3% TableDelete/ipv4/10000 609.2n ± 11% TableDelete/ipv4/100000 679.2n ± 28% TableDelete/ipv6/10 504.2n ± 4% TableDelete/ipv6/100 959.5n ± 12% TableDelete/ipv6/1000 1.436µ ± 6% TableDelete/ipv6/10000 1.772µ ± 15% TableDelete/ipv6/100000 1.172µ ± 113% TableGet/ipv4/10 32.14n ± 11% TableGet/ipv4/100 38.58n ± 2% TableGet/ipv4/1000 45.03n ± 2% TableGet/ipv4/10000 52.90n ± 7% TableGet/ipv4/100000 135.2n ± 11% TableGet/ipv6/10 41.55n ± 1% TableGet/ipv6/100 44.78n ± 2% TableGet/ipv6/1000 49.03n ± 2% TableGet/ipv6/10000 65.38n ± 5% TableGet/ipv6/100000 525.0n ± 39% │ avg-B/op │ TableInsertion/ipv4/10 25.18Ki ± 0% TableInsertion/ipv4/100 17.63Ki ± 0% TableInsertion/ipv4/1000 14.14Ki ± 0% TableInsertion/ipv4/10000 12.92Ki ± 0% TableInsertion/ipv4/100000 11.13Ki ± 0% TableInsertion/ipv6/10 76.87Ki ± 0% TableInsertion/ipv6/100 98.33Ki ± 0% TableInsertion/ipv6/1000 91.44Ki ± 0% TableInsertion/ipv6/10000 90.39Ki ± 0% TableInsertion/ipv6/100000 87.19Ki ± 0% TableDelete/ipv4/10 3.230 ± 0% TableDelete/ipv4/100 4.020 ± 0% TableDelete/ipv4/1000 3.990 ± 0% TableDelete/ipv4/10000 4.000 ± 0% TableDelete/ipv4/100000 4.000 ± 0% TableDelete/ipv6/10 16.00 ± 0% TableDelete/ipv6/100 16.00 ± 0% TableDelete/ipv6/1000 16.00 ± 0% TableDelete/ipv6/10000 16.00 ± 0% TableDelete/ipv6/100000 16.00 ± 0% │ avg-allocs/op │ TableInsertion/ipv4/10 2.900 ± 0% TableInsertion/ipv4/100 2.330 ± 0% TableInsertion/ipv4/1000 2.070 ± 0% TableInsertion/ipv4/10000 1.980 ± 0% TableInsertion/ipv4/100000 1.840 ± 0% TableInsertion/ipv6/10 6.800 ± 0% TableInsertion/ipv6/100 8.420 ± 0% TableInsertion/ipv6/1000 7.900 ± 0% TableInsertion/ipv6/10000 7.820 ± 0% TableInsertion/ipv6/100000 7.580 ± 0% TableDelete/ipv4/10 1.000 ± 0% TableDelete/ipv4/100 1.000 ± 0% TableDelete/ipv4/1000 1.000 ± 0% TableDelete/ipv4/10000 1.000 ± 0% TableDelete/ipv4/100000 1.000 ± 0% TableDelete/ipv6/10 1.000 ± 0% TableDelete/ipv6/100 1.000 ± 0% TableDelete/ipv6/1000 1.000 ± 0% TableDelete/ipv6/10000 1.000 ± 0% TableDelete/ipv6/100000 1.000 ± 0% │ routes/s │ TableInsertion/ipv4/10 640.3k ± 2% TableInsertion/ipv4/100 417.1k ± 5% TableInsertion/ipv4/1000 477.0k ± 3% TableInsertion/ipv4/10000 362.8k ± 5% TableInsertion/ipv4/100000 404.5k ± 15% TableInsertion/ipv6/10 130.7k ± 1% TableInsertion/ipv6/100 82.69k ± 3% TableInsertion/ipv6/1000 67.37k ± 5% TableInsertion/ipv6/10000 67.93k ± 9% TableInsertion/ipv6/100000 75.63k ± 29% TableDelete/ipv4/10 2.642M ± 6% TableDelete/ipv4/100 2.726M ± 3% TableDelete/ipv4/1000 2.389M ± 3% TableDelete/ipv4/10000 1.641M ± 12% TableDelete/ipv4/100000 1.472M ± 27% TableDelete/ipv6/10 1.984M ± 4% TableDelete/ipv6/100 1.042M ± 11% TableDelete/ipv6/1000 696.5k ± 6% TableDelete/ipv6/10000 564.4k ± 13% TableDelete/ipv6/100000 853.6k ± 53% │ addrs/s │ TableGet/ipv4/10 31.11M ± 10% TableGet/ipv4/100 25.92M ± 2% TableGet/ipv4/1000 22.21M ± 2% TableGet/ipv4/10000 18.91M ± 8% TableGet/ipv4/100000 7.397M ± 12% TableGet/ipv6/10 24.07M ± 1% TableGet/ipv6/100 22.33M ± 2% TableGet/ipv6/1000 20.40M ± 2% TableGet/ipv6/10000 15.30M ± 5% TableGet/ipv6/100000 1.905M ± 28% │ B/op │ TableGet/ipv4/10 4.000 ± 0% TableGet/ipv4/100 4.000 ± 0% TableGet/ipv4/1000 4.000 ± 0% TableGet/ipv4/10000 4.000 ± 0% TableGet/ipv4/100000 4.000 ± 0% TableGet/ipv6/10 16.00 ± 0% TableGet/ipv6/100 16.00 ± 0% TableGet/ipv6/1000 16.00 ± 0% TableGet/ipv6/10000 16.00 ± 0% TableGet/ipv6/100000 16.00 ± 0% │ allocs/op │ TableGet/ipv4/10 1.000 ± 0% TableGet/ipv4/100 1.000 ± 0% TableGet/ipv4/1000 1.000 ± 0% TableGet/ipv4/10000 1.000 ± 0% TableGet/ipv4/100000 1.000 ± 0% TableGet/ipv6/10 1.000 ± 0% TableGet/ipv6/100 1.000 ± 0% TableGet/ipv6/1000 1.000 ± 0% TableGet/ipv6/10000 1.000 ± 0% TableGet/ipv6/100000 1.000 ± 0% Signed-off-by: David Anderson <danderson@tailscale.com>
2 years ago
func (t *strideTable[T]) getValAndChild(addr uint8) (*T, *strideTable[T]) {
idx := hostIndex(addr)
return t.entries[idx].value, t.entries[idx].child
}
net/art: implement the stride table building block of ART A stride table is an 8-bit routing table implemented as an array binary tree, with a special tree updating function (allot) that enables lightning fast address lookups and reasonably fast insertion and deletion. Insertion, deletion and lookup are all allocation-free. Updates #7781 │ sec/op │ StrideTableInsertion/10/random_order 16.79n ± 2% StrideTableInsertion/10/largest_first 16.83n ± 1% StrideTableInsertion/10/smallest_first 16.83n ± 0% StrideTableInsertion/50/random_order 17.84n ± 1% StrideTableInsertion/50/largest_first 20.04n ± 1% StrideTableInsertion/50/smallest_first 16.39n ± 0% StrideTableInsertion/100/random_order 14.63n ± 0% StrideTableInsertion/100/largest_first 17.45n ± 4% StrideTableInsertion/100/smallest_first 12.98n ± 0% StrideTableInsertion/200/random_order 12.51n ± 4% StrideTableInsertion/200/largest_first 18.36n ± 3% StrideTableInsertion/200/smallest_first 9.609n ± 3% StrideTableDeletion/10/random_order 19.50n ± 1% StrideTableDeletion/10/largest_first 19.34n ± 0% StrideTableDeletion/10/smallest_first 19.43n ± 0% StrideTableDeletion/50/random_order 14.58n ± 1% StrideTableDeletion/50/largest_first 14.27n ± 2% StrideTableDeletion/50/smallest_first 15.51n ± 0% StrideTableDeletion/100/random_order 12.02n ± 3% StrideTableDeletion/100/largest_first 10.64n ± 0% StrideTableDeletion/100/smallest_first 13.21n ± 3% StrideTableDeletion/200/random_order 14.05n ± 4% StrideTableDeletion/200/largest_first 9.288n ± 5% StrideTableDeletion/200/smallest_first 18.51n ± 1% StrideTableGet 0.5010n ± 0% │ routes/s │ StrideTableInsertion/10/random_order 59.55M ± 2% StrideTableInsertion/10/largest_first 59.42M ± 1% StrideTableInsertion/10/smallest_first 59.43M ± 0% StrideTableInsertion/50/random_order 56.04M ± 1% StrideTableInsertion/50/largest_first 49.91M ± 1% StrideTableInsertion/50/smallest_first 61.00M ± 0% StrideTableInsertion/100/random_order 68.35M ± 0% StrideTableInsertion/100/largest_first 57.32M ± 3% StrideTableInsertion/100/smallest_first 77.06M ± 0% StrideTableInsertion/200/random_order 79.93M ± 4% StrideTableInsertion/200/largest_first 54.47M ± 3% StrideTableInsertion/200/smallest_first 104.1M ± 3% StrideTableDeletion/10/random_order 51.28M ± 1% StrideTableDeletion/10/largest_first 51.70M ± 0% StrideTableDeletion/10/smallest_first 51.48M ± 0% StrideTableDeletion/50/random_order 68.60M ± 1% StrideTableDeletion/50/largest_first 70.09M ± 2% StrideTableDeletion/50/smallest_first 64.45M ± 0% StrideTableDeletion/100/random_order 83.21M ± 3% StrideTableDeletion/100/largest_first 94.03M ± 0% StrideTableDeletion/100/smallest_first 75.69M ± 3% StrideTableDeletion/200/random_order 71.20M ± 5% StrideTableDeletion/200/largest_first 107.7M ± 5% StrideTableDeletion/200/smallest_first 54.02M ± 1% StrideTableGet 1.996G ± 0% Signed-off-by: David Anderson <danderson@tailscale.com>
2 years ago
// allot updates entries whose stored prefixIndex matches oldPrefixIndex, in the
// subtree rooted at idx. Matching entries have their stored prefixIndex set to
// newPrefixIndex, and their value set to val.
//
// allot is the core of the ART algorithm, enabling efficient insertion/deletion
// while preserving very fast lookups.
func (t *strideTable[T]) allot(idx int, oldPrefixIndex, newPrefixIndex int, val *T) {
if t.entries[idx].prefixIndex != oldPrefixIndex {
// current prefixIndex isn't what we expect. This is a recursive call
// that found a child subtree that already has a more specific route
// installed. Don't touch it.
return
}
t.entries[idx].value = val
t.entries[idx].prefixIndex = newPrefixIndex
if idx >= firstHostIndex {
// The entry we just updated was a host route, we're at the bottom of
// the binary tree.
return
}
// Propagate the allotment to this node's children.
left := idx << 1
t.allot(left, oldPrefixIndex, newPrefixIndex, val)
right := left + 1
t.allot(right, oldPrefixIndex, newPrefixIndex, val)
}
// insert adds the route addr/prefixLen to t, with value val.
func (t *strideTable[T]) insert(addr uint8, prefixLen int, val *T) {
idx := prefixIndex(addr, prefixLen)
old := t.entries[idx].value
oldIdx := t.entries[idx].prefixIndex
if oldIdx == idx && old == val {
// This exact prefix+value is already in the table.
return
}
t.allot(idx, oldIdx, idx, val)
if oldIdx != idx {
// This route entry was freshly created (not just updated), that's a new
// reference.
t.refs++
}
return
}
// delete removes the route addr/prefixLen from t.
func (t *strideTable[T]) delete(addr uint8, prefixLen int) *T {
idx := prefixIndex(addr, prefixLen)
recordedIdx := t.entries[idx].prefixIndex
if recordedIdx != idx {
// Route entry doesn't exist
return nil
}
val := t.entries[idx].value
parentIdx := idx >> 1
t.allot(idx, idx, t.entries[parentIdx].prefixIndex, t.entries[parentIdx].value)
t.refs--
return val
}
// get does a route lookup for addr and returns the associated value, or nil if
// no route matched.
func (t *strideTable[T]) get(addr uint8) *T {
return t.entries[hostIndex(addr)].value
}
// TableDebugString returns the contents of t, formatted as a table with one
// line per entry.
func (t *strideTable[T]) tableDebugString() string {
var ret bytes.Buffer
for i, ent := range t.entries {
if i == 0 {
continue
}
v := "(nil)"
if ent.value != nil {
v = fmt.Sprint(*ent.value)
}
fmt.Fprintf(&ret, "idx=%3d (%s), parent=%3d (%s), val=%v\n", i, formatPrefixTable(inversePrefixIndex(i)), ent.prefixIndex, formatPrefixTable(inversePrefixIndex((ent.prefixIndex))), v)
}
return ret.String()
}
// treeDebugString returns the contents of t, formatted as a sparse tree. Each
// line is one entry, indented such that it is contained by all its parents, and
// non-overlapping with any of its siblings.
func (t *strideTable[T]) treeDebugString() string {
var ret bytes.Buffer
t.treeDebugStringRec(&ret, 1, 0) // index of 0/0, and 0 indent
return ret.String()
}
func (t *strideTable[T]) treeDebugStringRec(w io.Writer, idx, indent int) {
addr, len := inversePrefixIndex(idx)
if t.entries[idx].prefixIndex != 0 && t.entries[idx].prefixIndex == idx {
fmt.Fprintf(w, "%s%d/%d (%d/%d) = %v\n", strings.Repeat(" ", indent), addr, len, addr, len, *t.entries[idx].value)
indent += 2
}
if idx >= firstHostIndex {
return
}
left := idx << 1
t.treeDebugStringRec(w, left, indent)
right := left + 1
t.treeDebugStringRec(w, right, indent)
}
// prefixIndex returns the array index of the tree node for addr/prefixLen.
func prefixIndex(addr uint8, prefixLen int) int {
// the prefixIndex of addr/prefixLen is the prefixLen most significant bits
// of addr, with a 1 tacked onto the left-hand side. For example:
//
// - 0/0 is 1: 0 bits of the addr, with a 1 tacked on
// - 42/8 is 1_00101010 (298): all bits of 42, with a 1 tacked on
// - 48/4 is 1_0011 (19): 4 most-significant bits of 48, with a 1 tacked on
return (int(addr) >> (8 - prefixLen)) + (1 << prefixLen)
}
// hostIndex returns the array index of the host route for addr.
// It is equivalent to prefixIndex(addr, 8).
func hostIndex(addr uint8) int {
return int(addr) + 1<<8
}
// inversePrefixIndex returns the address and prefix length of idx. It is the
// inverse of prefixIndex. Only used for debugging and in tests.
func inversePrefixIndex(idx int) (addr uint8, len int) {
lz := bits.LeadingZeros(uint(idx))
len = strconv.IntSize - lz - 1
addr = uint8(idx&(0xFF>>(8-len))) << (8 - len)
return addr, len
}
// formatPrefixTable formats addr and len as addr/len, with a constant width
// suitable for use in table formatting.
func formatPrefixTable(addr uint8, len int) string {
if len < 0 { // this happens for inversePrefixIndex(0)
return "<nil>"
}
return fmt.Sprintf("%3d/%d", addr, len)
}