Replace ChaCha20 implementation with an optimized version

.travis.yml

@@ -7,6 +7,7 @@ before_install:
   - go get golang.org/x/tools/cmd/cover
   - go get github.com/onsi/gomega
   - go get github.com/onsi/ginkgo
+  - go get github.com/aead/chacha20
 
 script:
   - go test -tags json github.com/v2ray/v2ray-core/...

common/crypto/chacha20.go

@@ -3,9 +3,27 @@ package crypto
 import (
 	"crypto/cipher"
 
-	"github.com/v2ray/v2ray-core/common/crypto/internal"
+	"github.com/aead/chacha20"
 )
 
+func makeNonce(nonce *[chacha20.NonceSize]byte, iv []byte) {
+	switch len(iv) {
+	case 8:
+		copy(nonce[4:], iv)
+	case 12:
+		copy(nonce[:], iv)
+	default:
+		panic("bad nonce length")
+	}
+}
+
 func NewChaCha20Stream(key []byte, iv []byte) cipher.Stream {
-	return internal.NewChaCha20Stream(key, iv, 20)
+	var Key [32]byte
+	var Nonce [12]byte
+	if len(key) != 32 {
+		panic("bad key length")
+	}
+	copy(Key[:], key)
+	makeNonce(&Nonce, iv)
+	return chacha20.NewCipher(&Nonce, &Key)
 }

common/crypto/internal/chacha.go

@@ -1,80 +0,0 @@
-package internal
-
-//go:generate go run chacha_core_gen.go
-
-import (
-	"encoding/binary"
-)
-
-const (
-	wordSize  = 4                    // the size of ChaCha20's words
-	stateSize = 16                   // the size of ChaCha20's state, in words
-	blockSize = stateSize * wordSize // the size of ChaCha20's block, in bytes
-)
-
-type ChaCha20Stream struct {
-	state  [stateSize]uint32 // the state as an array of 16 32-bit words
-	block  [blockSize]byte   // the keystream as an array of 64 bytes
-	offset int               // the offset of used bytes in block
-	rounds int
-}
-
-func NewChaCha20Stream(key []byte, nonce []byte, rounds int) *ChaCha20Stream {
-	s := new(ChaCha20Stream)
-	// the magic constants for 256-bit keys
-	s.state[0] = 0x61707865
-	s.state[1] = 0x3320646e
-	s.state[2] = 0x79622d32
-	s.state[3] = 0x6b206574
-
-	for i := 0; i < 8; i++ {
-		s.state[i+4] = binary.LittleEndian.Uint32(key[i*4 : i*4+4])
-	}
-
-	switch len(nonce) {
-	case 8:
-		s.state[14] = binary.LittleEndian.Uint32(nonce[0:])
-		s.state[15] = binary.LittleEndian.Uint32(nonce[4:])
-	case 12:
-		s.state[13] = binary.LittleEndian.Uint32(nonce[0:4])
-		s.state[14] = binary.LittleEndian.Uint32(nonce[4:8])
-		s.state[15] = binary.LittleEndian.Uint32(nonce[8:12])
-	default:
-		panic("bad nonce length")
-	}
-
-	s.rounds = rounds
-	ChaCha20Block(&s.state, s.block[:], s.rounds)
-	return s
-}
-
-func (s *ChaCha20Stream) XORKeyStream(dst, src []byte) {
-	// Stride over the input in 64-byte blocks, minus the amount of keystream
-	// previously used. This will produce best results when processing blocks
-	// of a size evenly divisible by 64.
-	i := 0
-	max := len(src)
-	for i < max {
-		gap := blockSize - s.offset
-
-		limit := i + gap
-		if limit > max {
-			limit = max
-		}
-
-		o := s.offset
-		for j := i; j < limit; j++ {
-			dst[j] = src[j] ^ s.block[o]
-			o++
-		}
-
-		i += gap
-		s.offset = o
-
-		if o == blockSize {
-			s.offset = 0
-			s.state[12]++
-			ChaCha20Block(&s.state, s.block[:], s.rounds)
-		}
-	}
-}

common/crypto/internal/chacha_core.go

@@ -1,124 +0,0 @@
-// GENERATED CODE. DO NOT MODIFY!
-package internal
-
-import "encoding/binary"
-
-func ChaCha20Block(s *[16]uint32, out []byte, rounds int) {
-	var x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15 = s[0], s[1], s[2], s[3], s[4], s[5], s[6], s[7], s[8], s[9], s[10], s[11], s[12], s[13], s[14], s[15]
-	for i := 0; i < rounds; i += 2 {
-		var x uint32
-
-		x0 += x4
-		x = x12 ^ x0
-		x12 = (x << 16) | (x >> (32 - 16))
-		x8 += x12
-		x = x4 ^ x8
-		x4 = (x << 12) | (x >> (32 - 12))
-		x0 += x4
-		x = x12 ^ x0
-		x12 = (x << 8) | (x >> (32 - 8))
-		x8 += x12
-		x = x4 ^ x8
-		x4 = (x << 7) | (x >> (32 - 7))
-		x1 += x5
-		x = x13 ^ x1
-		x13 = (x << 16) | (x >> (32 - 16))
-		x9 += x13
-		x = x5 ^ x9
-		x5 = (x << 12) | (x >> (32 - 12))
-		x1 += x5
-		x = x13 ^ x1
-		x13 = (x << 8) | (x >> (32 - 8))
-		x9 += x13
-		x = x5 ^ x9
-		x5 = (x << 7) | (x >> (32 - 7))
-		x2 += x6
-		x = x14 ^ x2
-		x14 = (x << 16) | (x >> (32 - 16))
-		x10 += x14
-		x = x6 ^ x10
-		x6 = (x << 12) | (x >> (32 - 12))
-		x2 += x6
-		x = x14 ^ x2
-		x14 = (x << 8) | (x >> (32 - 8))
-		x10 += x14
-		x = x6 ^ x10
-		x6 = (x << 7) | (x >> (32 - 7))
-		x3 += x7
-		x = x15 ^ x3
-		x15 = (x << 16) | (x >> (32 - 16))
-		x11 += x15
-		x = x7 ^ x11
-		x7 = (x << 12) | (x >> (32 - 12))
-		x3 += x7
-		x = x15 ^ x3
-		x15 = (x << 8) | (x >> (32 - 8))
-		x11 += x15
-		x = x7 ^ x11
-		x7 = (x << 7) | (x >> (32 - 7))
-		x0 += x5
-		x = x15 ^ x0
-		x15 = (x << 16) | (x >> (32 - 16))
-		x10 += x15
-		x = x5 ^ x10
-		x5 = (x << 12) | (x >> (32 - 12))
-		x0 += x5
-		x = x15 ^ x0
-		x15 = (x << 8) | (x >> (32 - 8))
-		x10 += x15
-		x = x5 ^ x10
-		x5 = (x << 7) | (x >> (32 - 7))
-		x1 += x6
-		x = x12 ^ x1
-		x12 = (x << 16) | (x >> (32 - 16))
-		x11 += x12
-		x = x6 ^ x11
-		x6 = (x << 12) | (x >> (32 - 12))
-		x1 += x6
-		x = x12 ^ x1
-		x12 = (x << 8) | (x >> (32 - 8))
-		x11 += x12
-		x = x6 ^ x11
-		x6 = (x << 7) | (x >> (32 - 7))
-		x2 += x7
-		x = x13 ^ x2
-		x13 = (x << 16) | (x >> (32 - 16))
-		x8 += x13
-		x = x7 ^ x8
-		x7 = (x << 12) | (x >> (32 - 12))
-		x2 += x7
-		x = x13 ^ x2
-		x13 = (x << 8) | (x >> (32 - 8))
-		x8 += x13
-		x = x7 ^ x8
-		x7 = (x << 7) | (x >> (32 - 7))
-		x3 += x4
-		x = x14 ^ x3
-		x14 = (x << 16) | (x >> (32 - 16))
-		x9 += x14
-		x = x4 ^ x9
-		x4 = (x << 12) | (x >> (32 - 12))
-		x3 += x4
-		x = x14 ^ x3
-		x14 = (x << 8) | (x >> (32 - 8))
-		x9 += x14
-		x = x4 ^ x9
-		x4 = (x << 7) | (x >> (32 - 7))
-	}
-	binary.LittleEndian.PutUint32(out[0:4], s[0]+x0)
-	binary.LittleEndian.PutUint32(out[4:8], s[1]+x1)
-	binary.LittleEndian.PutUint32(out[8:12], s[2]+x2)
-	binary.LittleEndian.PutUint32(out[12:16], s[3]+x3)
-	binary.LittleEndian.PutUint32(out[16:20], s[4]+x4)
-	binary.LittleEndian.PutUint32(out[20:24], s[5]+x5)
-	binary.LittleEndian.PutUint32(out[24:28], s[6]+x6)
-	binary.LittleEndian.PutUint32(out[28:32], s[7]+x7)
-	binary.LittleEndian.PutUint32(out[32:36], s[8]+x8)
-	binary.LittleEndian.PutUint32(out[36:40], s[9]+x9)
-	binary.LittleEndian.PutUint32(out[40:44], s[10]+x10)
-	binary.LittleEndian.PutUint32(out[44:48], s[11]+x11)
-	binary.LittleEndian.PutUint32(out[48:52], s[12]+x12)
-	binary.LittleEndian.PutUint32(out[52:56], s[13]+x13)
-	binary.LittleEndian.PutUint32(out[56:60], s[14]+x14)
-	binary.LittleEndian.PutUint32(out[60:64], s[15]+x15)
-}

common/crypto/internal/chacha_core_gen.go

@@ -1,70 +0,0 @@
-// +build generate
-
-package main
-
-import (
-	"fmt"
-	"log"
-	"os"
-)
-
-func writeQuarterRound(file *os.File, a, b, c, d int) {
-	add := "x%d+=x%d\n"
-	xor := "x=x%d^x%d\n"
-	rotate := "x%d=(x << %d) | (x >> (32 - %d))\n"
-
-	fmt.Fprintf(file, add, a, b)
-	fmt.Fprintf(file, xor, d, a)
-	fmt.Fprintf(file, rotate, d, 16, 16)
-
-	fmt.Fprintf(file, add, c, d)
-	fmt.Fprintf(file, xor, b, c)
-	fmt.Fprintf(file, rotate, b, 12, 12)
-
-	fmt.Fprintf(file, add, a, b)
-	fmt.Fprintf(file, xor, d, a)
-	fmt.Fprintf(file, rotate, d, 8, 8)
-
-	fmt.Fprintf(file, add, c, d)
-	fmt.Fprintf(file, xor, b, c)
-	fmt.Fprintf(file, rotate, b, 7, 7)
-}
-
-func writeChacha20Block(file *os.File) {
-	fmt.Fprintln(file, `
-func ChaCha20Block(s *[16]uint32, out []byte, rounds int) {
-  var x0,x1,x2,x3,x4,x5,x6,x7,x8,x9,x10,x11,x12,x13,x14,x15 = s[0],s[1],s[2],s[3],s[4],s[5],s[6],s[7],s[8],s[9],s[10],s[11],s[12],s[13],s[14],s[15]
-	for i := 0; i < rounds; i+=2 {
-    var x uint32
-    `)
-
-	writeQuarterRound(file, 0, 4, 8, 12)
-	writeQuarterRound(file, 1, 5, 9, 13)
-	writeQuarterRound(file, 2, 6, 10, 14)
-	writeQuarterRound(file, 3, 7, 11, 15)
-	writeQuarterRound(file, 0, 5, 10, 15)
-	writeQuarterRound(file, 1, 6, 11, 12)
-	writeQuarterRound(file, 2, 7, 8, 13)
-	writeQuarterRound(file, 3, 4, 9, 14)
-	fmt.Fprintln(file, "}")
-	for i := 0; i < 16; i++ {
-		fmt.Fprintf(file, "binary.LittleEndian.PutUint32(out[%d:%d], s[%d]+x%d)\n", i*4, i*4+4, i, i)
-	}
-	fmt.Fprintln(file, "}")
-	fmt.Fprintln(file)
-}
-
-func main() {
-	file, err := os.OpenFile("chacha_core.go", os.O_WRONLY|os.O_TRUNC|os.O_CREATE, 0644)
-	if err != nil {
-		log.Fatalf("Failed to generate chacha_core.go: %v", err)
-	}
-	defer file.Close()
-
-	fmt.Fprintln(file, "// GENERATED CODE. DO NOT MODIFY!")
-	fmt.Fprintln(file, "package internal")
-	fmt.Fprintln(file)
-	fmt.Fprintln(file, "import \"encoding/binary\"")
-	fmt.Fprintln(file)
-	writeChacha20Block(file)
-}