tailscale/util/deephash/deephash.go

390 lines
11 KiB
Go
Raw Normal View History

// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Package deephash hashes a Go value recursively, in a predictable order,
// without looping. The hash is only valid within the lifetime of a program.
// Users should not store the hash on disk or send it over the network.
// The hash is sufficiently strong and unique such that
// Hash(x) == Hash(y) is an appropriate replacement for x == y.
//
util/deephash: remove unnecessary formatting for structs and slices (#2571) The index for every struct field or slice element and the number of fields for the struct is unncessary. The hashing of Go values is unambiguous because every type (except maps) encodes in a parsable manner. So long as we know the type information, we could theoretically decode every value (except for maps). At a high level: * numbers are encoded as fixed-width records according to precision. * strings (and AppendTo output) are encoded with a fixed-width length, followed by the contents of the buffer. * slices are prefixed by a fixed-width length, followed by the encoding of each value. So long as we know the type of each element, we could theoretically decode each element. * arrays are encoded just like slices, but elide the length since it is determined from the Go type. * maps are encoded first with a byte indicating whether it is a cycle. If a cycle, it is followed by a fixed-width index for the pointer, otherwise followed by the SHA-256 hash of its contents. The encoding of maps is not decodeable, but a SHA-256 hash is sufficient to avoid ambiguities. * interfaces are encoded first with a byte indicating whether it is nil. If not nil, it is followed by a fixed-width index for the type, and then the encoding for the underlying value. Having the type be encoded first ensures that the value could theoretically be decoded next. * pointers are encoded first with a byte indicating whether it is 1) nil, 2) a cycle, or 3) newly seen. If a cycle, it is followed by a fixed-width index for the pointer. If newly seen, it is followed by the encoding for the pointed-at value. Removing unnecessary details speeds up hashing: name old time/op new time/op delta Hash-8 76.0µs ± 1% 55.8µs ± 2% -26.62% (p=0.000 n=10+10) HashMapAcyclic-8 61.9µs ± 0% 62.0µs ± 0% ~ (p=0.666 n=9+9) TailcfgNode-8 10.2µs ± 1% 7.5µs ± 1% -26.90% (p=0.000 n=10+9) HashArray-8 1.07µs ± 1% 0.70µs ± 1% -34.67% (p=0.000 n=10+9) Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-08-04 03:35:57 +00:00
// The definition of equality is identical to reflect.DeepEqual except:
// * Floating-point values are compared based on the raw bits,
// which means that NaNs (with the same bit pattern) are treated as equal.
// * Types which implement interface { AppendTo([]byte) []byte } use
// the AppendTo method to produce a textual representation of the value.
// Thus, two values are equal if AppendTo produces the same bytes.
//
// WARNING: This package, like most of the tailscale.com Go module,
// should be considered Tailscale-internal; we make no API promises.
package deephash
import (
"bufio"
"crypto/sha256"
"encoding/binary"
"encoding/hex"
"fmt"
"hash"
"math"
"reflect"
"sync"
"time"
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
"unsafe"
)
util/deephash: remove unnecessary formatting for structs and slices (#2571) The index for every struct field or slice element and the number of fields for the struct is unncessary. The hashing of Go values is unambiguous because every type (except maps) encodes in a parsable manner. So long as we know the type information, we could theoretically decode every value (except for maps). At a high level: * numbers are encoded as fixed-width records according to precision. * strings (and AppendTo output) are encoded with a fixed-width length, followed by the contents of the buffer. * slices are prefixed by a fixed-width length, followed by the encoding of each value. So long as we know the type of each element, we could theoretically decode each element. * arrays are encoded just like slices, but elide the length since it is determined from the Go type. * maps are encoded first with a byte indicating whether it is a cycle. If a cycle, it is followed by a fixed-width index for the pointer, otherwise followed by the SHA-256 hash of its contents. The encoding of maps is not decodeable, but a SHA-256 hash is sufficient to avoid ambiguities. * interfaces are encoded first with a byte indicating whether it is nil. If not nil, it is followed by a fixed-width index for the type, and then the encoding for the underlying value. Having the type be encoded first ensures that the value could theoretically be decoded next. * pointers are encoded first with a byte indicating whether it is 1) nil, 2) a cycle, or 3) newly seen. If a cycle, it is followed by a fixed-width index for the pointer. If newly seen, it is followed by the encoding for the pointed-at value. Removing unnecessary details speeds up hashing: name old time/op new time/op delta Hash-8 76.0µs ± 1% 55.8µs ± 2% -26.62% (p=0.000 n=10+10) HashMapAcyclic-8 61.9µs ± 0% 62.0µs ± 0% ~ (p=0.666 n=9+9) TailcfgNode-8 10.2µs ± 1% 7.5µs ± 1% -26.90% (p=0.000 n=10+9) HashArray-8 1.07µs ± 1% 0.70µs ± 1% -34.67% (p=0.000 n=10+9) Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-08-04 03:35:57 +00:00
// There is much overlap between the theory of serialization and hashing.
// A hash (useful for determing equality) can be produced by printing a value
// and hashing the output. The format must:
// * be deterministic such that the same value hashes to the same output, and
// * be parsable such that the same value can be reproduced by the output.
//
// The logic below hashes a value by printing it to a hash.Hash.
// To be parsable, it assumes that we know the Go type of each value:
// * scalar types (e.g., bool or int32) are printed as fixed-width fields.
// * list types (e.g., strings, slices, and AppendTo buffers) are prefixed
// by a fixed-width length field, followed by the contents of the list.
// * slices, arrays, and structs print each element/field consecutively.
// * interfaces print with a 1-byte prefix indicating whether it is nil.
// If non-nil, it is followed by a fixed-width field of the type index,
// followed by the format of the underlying value.
// * pointers print with a 1-byte prefix indicating whether the pointer is
// 1) nil, 2) previously seen, or 3) newly seen. Previously seen pointers are
// followed by a fixed-width field with the index of the previous pointer.
// Newly seen pointers are followed by the format of the underlying value.
// * maps print with a 1-byte prefix indicating whether the map pointer is
// 1) nil, 2) previously seen, or 3) newly seen. Previously seen pointers
// are followed by a fixed-width field of the index of the previous pointer.
// Newly seen maps are printed as a fixed-width field with the XOR of the
// hash of every map entry. With a sufficiently strong hash, this value is
// theoretically "parsable" by looking up the hash in a magical map that
// returns the set of entries for that given hash.
const scratchSize = 128
// hasher is reusable state for hashing a value.
// Get one via hasherPool.
type hasher struct {
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
h hash.Hash
bw *bufio.Writer
scratch [scratchSize]byte
visitStack visitStack
}
func (h *hasher) reset() {
if h.h == nil {
h.h = sha256.New()
}
if h.bw == nil {
h.bw = bufio.NewWriterSize(h.h, h.h.BlockSize())
}
h.bw.Flush()
h.h.Reset()
}
// Sum is an opaque checksum type that is comparable.
type Sum struct {
sum [sha256.Size]byte
}
func (s1 *Sum) xor(s2 Sum) {
for i := 0; i < sha256.Size; i++ {
s1.sum[i] ^= s2.sum[i]
}
}
func (s Sum) String() string {
return hex.EncodeToString(s.sum[:])
}
var (
once sync.Once
seed uint64
)
func (h *hasher) sum() (s Sum) {
h.bw.Flush()
// Sum into scratch & copy out, as hash.Hash is an interface
// so the slice necessarily escapes, and there's no sha256
// concrete type exported and we don't want the 'hash' result
// parameter to escape to the heap:
copy(s.sum[:], h.h.Sum(h.scratch[:0]))
return s
}
var hasherPool = &sync.Pool{
New: func() interface{} { return new(hasher) },
}
// Hash returns the hash of v.
func Hash(v interface{}) (s Sum) {
h := hasherPool.Get().(*hasher)
defer hasherPool.Put(h)
h.reset()
once.Do(func() {
seed = uint64(time.Now().UnixNano())
})
h.hashUint64(seed)
h.hashValue(reflect.ValueOf(v))
return h.sum()
}
// Update sets last to the hash of v and reports whether its value changed.
func Update(last *Sum, v ...interface{}) (changed bool) {
sum := Hash(v)
if sum == *last {
// unchanged.
return false
}
*last = sum
return true
}
var appenderToType = reflect.TypeOf((*appenderTo)(nil)).Elem()
type appenderTo interface {
AppendTo([]byte) []byte
}
func (h *hasher) hashUint8(i uint8) {
h.bw.WriteByte(i)
}
func (h *hasher) hashUint16(i uint16) {
binary.LittleEndian.PutUint16(h.scratch[:2], i)
h.bw.Write(h.scratch[:2])
}
func (h *hasher) hashUint32(i uint32) {
binary.LittleEndian.PutUint32(h.scratch[:4], i)
h.bw.Write(h.scratch[:4])
}
func (h *hasher) hashUint64(i uint64) {
binary.LittleEndian.PutUint64(h.scratch[:8], i)
h.bw.Write(h.scratch[:8])
}
var uint8Type = reflect.TypeOf(byte(0))
func (h *hasher) hashValue(v reflect.Value) {
if !v.IsValid() {
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
w := h.bw
if v.CanInterface() {
// Use AppendTo methods, if available and cheap.
if v.CanAddr() && v.Type().Implements(appenderToType) {
a := v.Addr().Interface().(appenderTo)
size := h.scratch[:8]
record := a.AppendTo(size)
binary.LittleEndian.PutUint64(record, uint64(len(record)-len(size)))
w.Write(record)
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
}
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
// TODO(dsnet): Avoid cycle detection for types that cannot have cycles.
// Generic handling.
switch v.Kind() {
default:
panic(fmt.Sprintf("unhandled kind %v for type %v", v.Kind(), v.Type()))
case reflect.Ptr:
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
if v.IsNil() {
h.hashUint8(0) // indicates nil
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
// Check for cycle.
ptr := pointerOf(v)
if idx, ok := h.visitStack.seen(ptr); ok {
h.hashUint8(2) // indicates cycle
h.hashUint64(uint64(idx))
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
h.visitStack.push(ptr)
defer h.visitStack.pop(ptr)
h.hashUint8(1) // indicates visiting a pointer
h.hashValue(v.Elem())
case reflect.Struct:
for i, n := 0, v.NumField(); i < n; i++ {
h.hashValue(v.Field(i))
}
case reflect.Slice, reflect.Array:
vLen := v.Len()
if v.Kind() == reflect.Slice {
h.hashUint64(uint64(vLen))
}
if v.Type().Elem() == uint8Type && v.CanInterface() {
if vLen > 0 && vLen <= scratchSize {
// If it fits in scratch, avoid the Interface allocation.
// It seems tempting to do this for all sizes, doing
// scratchSize bytes at a time, but reflect.Slice seems
// to allocate, so it's not a win.
n := reflect.Copy(reflect.ValueOf(&h.scratch).Elem(), v)
w.Write(h.scratch[:n])
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
fmt.Fprintf(w, "%s", v.Interface())
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
for i := 0; i < vLen; i++ {
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
// TODO(dsnet): Perform cycle detection for slices,
// which is functionally a list of pointers.
// See https://github.com/google/go-cmp/blob/402949e8139bb890c71a707b6faf6dd05c92f4e5/cmp/compare.go#L438-L450
h.hashValue(v.Index(i))
}
case reflect.Interface:
if v.IsNil() {
h.hashUint8(0) // indicates nil
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
v = v.Elem()
h.hashUint8(1) // indicates visiting interface value
h.hashType(v.Type())
h.hashValue(v)
case reflect.Map:
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
// Check for cycle.
ptr := pointerOf(v)
if idx, ok := h.visitStack.seen(ptr); ok {
h.hashUint8(2) // indicates cycle
h.hashUint64(uint64(idx))
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
return
}
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
h.visitStack.push(ptr)
defer h.visitStack.pop(ptr)
h.hashUint8(1) // indicates visiting a map
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
h.hashMap(v)
case reflect.String:
s := v.String()
h.hashUint64(uint64(len(s)))
w.WriteString(s)
case reflect.Bool:
if v.Bool() {
h.hashUint8(1)
} else {
h.hashUint8(0)
}
case reflect.Int8:
h.hashUint8(uint8(v.Int()))
case reflect.Int16:
h.hashUint16(uint16(v.Int()))
case reflect.Int32:
h.hashUint32(uint32(v.Int()))
case reflect.Int64, reflect.Int:
h.hashUint64(uint64(v.Int()))
case reflect.Uint8:
h.hashUint8(uint8(v.Uint()))
case reflect.Uint16:
h.hashUint16(uint16(v.Uint()))
case reflect.Uint32:
h.hashUint32(uint32(v.Uint()))
case reflect.Uint64, reflect.Uint, reflect.Uintptr:
h.hashUint64(uint64(v.Uint()))
case reflect.Float32:
h.hashUint32(math.Float32bits(float32(v.Float())))
case reflect.Float64:
h.hashUint64(math.Float64bits(float64(v.Float())))
case reflect.Complex64:
h.hashUint32(math.Float32bits(real(complex64(v.Complex()))))
h.hashUint32(math.Float32bits(imag(complex64(v.Complex()))))
case reflect.Complex128:
h.hashUint64(math.Float64bits(real(complex128(v.Complex()))))
h.hashUint64(math.Float64bits(imag(complex128(v.Complex()))))
}
}
type mapHasher struct {
h hasher
val valueCache // re-usable values for map iteration
iter reflect.MapIter // re-usable map iterator
}
var mapHasherPool = &sync.Pool{
New: func() interface{} { return new(mapHasher) },
}
type valueCache map[reflect.Type]reflect.Value
func (c *valueCache) get(t reflect.Type) reflect.Value {
v, ok := (*c)[t]
if !ok {
v = reflect.New(t).Elem()
if *c == nil {
*c = make(valueCache)
}
(*c)[t] = v
}
return v
}
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
// hashMap hashes a map in a sort-free manner.
// It relies on a map being a functionally an unordered set of KV entries.
// So long as we hash each KV entry together, we can XOR all
// of the individual hashes to produce a unique hash for the entire map.
func (h *hasher) hashMap(v reflect.Value) {
mh := mapHasherPool.Get().(*mapHasher)
defer mapHasherPool.Put(mh)
iter := mapIter(&mh.iter, v)
defer mapIter(&mh.iter, reflect.Value{}) // avoid pinning v from mh.iter when we return
var sum Sum
k := mh.val.get(v.Type().Key())
e := mh.val.get(v.Type().Elem())
mh.h.visitStack = h.visitStack // always use the parent's visit stack to avoid cycles
for iter.Next() {
key := iterKey(iter, k)
val := iterVal(iter, e)
mh.h.reset()
mh.h.hashValue(key)
mh.h.hashValue(val)
sum.xor(mh.h.sum())
}
h.bw.Write(append(h.scratch[:0], sum.sum[:]...)) // append into scratch to avoid heap allocation
}
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
// visitStack is a stack of pointers visited.
// Pointers are pushed onto the stack when visited, and popped when leaving.
// The integer value is the depth at which the pointer was visited.
// The length of this stack should be zero after every hashing operation.
type visitStack map[pointer]int
func (v visitStack) seen(p pointer) (int, bool) {
idx, ok := v[p]
return idx, ok
}
func (v *visitStack) push(p pointer) {
if *v == nil {
*v = make(map[pointer]int)
}
util/deephash: improve cycle detection (#2470) The previous algorithm used a map of all visited pointers. The strength of this approach is that it quickly prunes any nodes that we have ever visited before. The detriment of the approach is that pruning is heavily dependent on the order that pointers were visited. This is especially relevant for hashing a map where map entries are visited in a non-deterministic manner, which would cause the map hash to be non-deterministic (which defeats the point of a hash). This new algorithm uses a stack of all visited pointers, similar to how github.com/google/go-cmp performs cycle detection. When we visit a pointer, we push it onto the stack, and when we leave a pointer, we pop it from the stack. Before visiting a pointer, we first check whether the pointer exists anywhere in the stack. If yes, then we prune the node. The detriment of this approach is that we may hash a node more often than before since we do not prune as aggressively. The set of visited pointers up until any node is only the path of nodes up to that node and not any other pointers that may have been visited elsewhere. This provides us deterministic hashing regardless of visit order. We can now delete hashMapFallback and associated complexity, which only exists because the previous approach was non-deterministic in the presence of cycles. This fixes a failure of the old algorithm where obviously different values are treated as equal because the pruning was too aggresive. See https://github.com/tailscale/tailscale/issues/2443#issuecomment-883653534 The new algorithm is slightly slower since it prunes less aggresively: name old time/op new time/op delta Hash-8 66.1µs ± 1% 68.8µs ± 1% +4.09% (p=0.000 n=19+19) HashMapAcyclic-8 63.0µs ± 1% 62.5µs ± 1% -0.76% (p=0.000 n=18+19) TailcfgNode-8 9.79µs ± 2% 9.88µs ± 1% +0.95% (p=0.000 n=19+17) HashArray-8 643ns ± 1% 653ns ± 1% +1.64% (p=0.000 n=19+19) However, a slower but more correct algorithm seems more favorable than a faster but incorrect algorithm. Signed-off-by: Joe Tsai <joetsai@digital-static.net>
2021-07-22 22:22:48 +00:00
(*v)[p] = len(*v)
}
func (v visitStack) pop(p pointer) {
delete(v, p)
}
// pointer is a thin wrapper over unsafe.Pointer.
// We only rely on comparability of pointers; we cannot rely on uintptr since
// that would break if Go ever switched to a moving GC.
type pointer struct{ p unsafe.Pointer }
func pointerOf(v reflect.Value) pointer {
return pointer{unsafe.Pointer(v.Pointer())}
}
// hashType hashes a reflect.Type.
// The hash is only consistent within the lifetime of a program.
func (h *hasher) hashType(t reflect.Type) {
// This approach relies on reflect.Type always being backed by a unique
// *reflect.rtype pointer. A safer approach is to use a global sync.Map
// that maps reflect.Type to some arbitrary and unique index.
// While safer, it requires global state with memory that can never be GC'd.
rtypeAddr := reflect.ValueOf(t).Pointer() // address of *reflect.rtype
h.hashUint64(uint64(rtypeAddr))
}