提交 0edb109f 编写于 作者: A Andy Polyakov

evp/e_chacha20_poly1305.c: further improve small-fragment TLS performance.

Improvement coefficients vary with TLS fragment length and platform, on
most Intel processors maximum improvement is ~50%, while on Ryzen - 80%.
The "secret" is new dedicated ChaCha20_128 code path and vectorized xor
helpers.
Reviewed-by: NRich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6638)
上级 2ce71b60
......@@ -196,14 +196,23 @@ static int chacha20_poly1305_init_key(EVP_CIPHER_CTX *ctx,
}
# if !defined(OPENSSL_SMALL_FOOTPRINT)
# if defined(POLY1305_ASM) && (defined(__x86_64) || defined(__x86_64__) || \
defined(_M_AMD64) || defined(_M_X64))
# define XOR128_HELPERS
void *xor128_encrypt_n_pad(void *out, const void *inp, void *otp, size_t len);
void *xor128_decrypt_n_pad(void *out, const void *inp, void *otp, size_t len);
static const unsigned char zero[4 * CHACHA_BLK_SIZE] = { 0 };
# else
static const unsigned char zero[2 * CHACHA_BLK_SIZE] = { 0 };
# endif
static int chacha20_poly1305_tls_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
const unsigned char *in, size_t len)
{
EVP_CHACHA_AEAD_CTX *actx = aead_data(ctx);
size_t i, tail, tohash_len, plen = actx->tls_payload_length;
unsigned char *buf, *tohash, *ctr, storage[2 * CHACHA_BLK_SIZE + 32];
size_t tail, tohash_len, buf_len, plen = actx->tls_payload_length;
unsigned char *buf, *tohash, *ctr, storage[sizeof(zero) + 32];
if (len != plen + POLY1305_BLOCK_SIZE)
return -1;
......@@ -212,9 +221,11 @@ static int chacha20_poly1305_tls_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
ctr = buf + CHACHA_BLK_SIZE;
tohash = buf + CHACHA_BLK_SIZE - POLY1305_BLOCK_SIZE;
if (plen <= CHACHA_BLK_SIZE) {
# ifdef XOR128_HELPERS
if (plen <= 3 * CHACHA_BLK_SIZE) {
actx->key.counter[0] = 0;
ChaCha20_ctr32(buf, zero, 2 * CHACHA_BLK_SIZE, actx->key.key.d,
buf_len = (plen + 2 * CHACHA_BLK_SIZE - 1) & (0 - CHACHA_BLK_SIZE);
ChaCha20_ctr32(buf, zero, buf_len, actx->key.key.d,
actx->key.counter);
Poly1305_Init(POLY1305_ctx(actx), buf);
actx->key.partial_len = 0;
......@@ -223,6 +234,31 @@ static int chacha20_poly1305_tls_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
actx->len.aad = EVP_AEAD_TLS1_AAD_LEN;
actx->len.text = plen;
if (plen) {
if (ctx->encrypt)
ctr = xor128_encrypt_n_pad(out, in, ctr, plen);
else
ctr = xor128_decrypt_n_pad(out, in, ctr, plen);
in += plen;
out += plen;
tohash_len = (size_t)(ctr - tohash);
}
}
# else
if (plen <= CHACHA_BLK_SIZE) {
size_t i;
actx->key.counter[0] = 0;
ChaCha20_ctr32(buf, zero, (buf_len = 2 * CHACHA_BLK_SIZE),
actx->key.key.d, actx->key.counter);
Poly1305_Init(POLY1305_ctx(actx), buf);
actx->key.partial_len = 0;
memcpy(tohash, actx->tls_aad, POLY1305_BLOCK_SIZE);
tohash_len = POLY1305_BLOCK_SIZE;
actx->len.aad = EVP_AEAD_TLS1_AAD_LEN;
actx->len.text = plen;
if (ctx->encrypt) {
for (i = 0; i < plen; i++) {
out[i] = ctr[i] ^= in[i];
......@@ -242,10 +278,12 @@ static int chacha20_poly1305_tls_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
memset(ctr + i, 0, tail);
ctr += i + tail;
tohash_len += i + tail;
} else {
}
# endif
else {
actx->key.counter[0] = 0;
ChaCha20_ctr32(buf, zero, CHACHA_BLK_SIZE, actx->key.key.d,
actx->key.counter);
ChaCha20_ctr32(buf, zero, (buf_len = CHACHA_BLK_SIZE),
actx->key.key.d, actx->key.counter);
Poly1305_Init(POLY1305_ctx(actx), buf);
actx->key.counter[0] = 1;
actx->key.partial_len = 0;
......@@ -300,7 +338,7 @@ static int chacha20_poly1305_tls_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
}
Poly1305_Update(POLY1305_ctx(actx), tohash, tohash_len);
OPENSSL_cleanse(buf, 2 * CHACHA_BLK_SIZE);
OPENSSL_cleanse(buf, buf_len);
Poly1305_Final(POLY1305_ctx(actx), ctx->encrypt ? actx->tag
: tohash);
......
......@@ -3753,6 +3753,110 @@ poly1305_emit_base2_44:
.size poly1305_emit_base2_44,.-poly1305_emit_base2_44
___
} } }
{ # chacha20-poly1305 helpers
my ($out,$inp,$otp,$len)=$win64 ? ("%rcx","%rdx","%r8", "%r9") : # Win64 order
("%rdi","%rsi","%rdx","%rcx"); # Unix order
$code.=<<___;
.globl xor128_encrypt_n_pad
.type xor128_encrypt_n_pad,\@abi-omnipotent
.align 16
xor128_encrypt_n_pad:
sub $otp,$inp
sub $otp,$out
mov $len,%r10 # put len aside
shr \$4,$len # len / 16
jz .Ltail_enc
nop
.Loop_enc_xmm:
movdqu ($inp,$otp),%xmm0
pxor ($otp),%xmm0
movdqu %xmm0,($out,$otp)
movdqa %xmm0,($otp)
lea 16($otp),$otp
dec $len
jnz .Loop_enc_xmm
and \$15,%r10 # len % 16
jz .Ldone_enc
.Ltail_enc:
mov \$16,$len
sub %r10,$len
xor %eax,%eax
.Loop_enc_byte:
mov ($inp,$otp),%al
xor ($otp),%al
mov %al,($out,$otp)
mov %al,($otp)
lea 1($otp),$otp
dec %r10
jnz .Loop_enc_byte
xor %eax,%eax
.Loop_enc_pad:
mov %al,($otp)
lea 1($otp),$otp
dec $len
jnz .Loop_enc_pad
.Ldone_enc:
mov $otp,%rax
ret
.size xor128_encrypt_n_pad,.-xor128_encrypt_n_pad
.globl xor128_decrypt_n_pad
.type xor128_decrypt_n_pad,\@abi-omnipotent
.align 16
xor128_decrypt_n_pad:
sub $otp,$inp
sub $otp,$out
mov $len,%r10 # put len aside
shr \$4,$len # len / 16
jz .Ltail_dec
nop
.Loop_dec_xmm:
movdqu ($inp,$otp),%xmm0
movdqa ($otp),%xmm1
pxor %xmm0,%xmm1
movdqu %xmm1,($out,$otp)
movdqa %xmm0,($otp)
lea 16($otp),$otp
dec $len
jnz .Loop_dec_xmm
pxor %xmm1,%xmm1
and \$15,%r10 # len % 16
jz .Ldone_dec
.Ltail_dec:
mov \$16,$len
sub %r10,$len
xor %eax,%eax
xor %r11,%r11
.Loop_dec_byte:
mov ($inp,$otp),%r11b
mov ($otp),%al
xor %r11b,%al
mov %al,($out,$otp)
mov %r11b,($otp)
lea 1($otp),$otp
dec %r10
jnz .Loop_dec_byte
xor %eax,%eax
.Loop_dec_pad:
mov %al,($otp)
lea 1($otp),$otp
dec $len
jnz .Loop_dec_pad
.Ldone_dec:
mov $otp,%rax
ret
.size xor128_decrypt_n_pad,.-xor128_decrypt_n_pad
___
}
$code.=<<___;
.align 64
.Lconst:
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册