mirror of
https://github.com/zerotier/ZeroTierOne.git
synced 2025-06-05 20:13:44 +02:00
Crypto fixes and benchmarks.
This commit is contained in:
parent
4af3467757
commit
5d0c6db19c
3 changed files with 222 additions and 211 deletions
294
node/AES.cpp
294
node/AES.cpp
|
@ -14,6 +14,8 @@
|
|||
#include "Constants.hpp"
|
||||
#include "AES.hpp"
|
||||
|
||||
#include <cstdio>
|
||||
|
||||
namespace ZeroTier {
|
||||
|
||||
// GMAC ---------------------------------------------------------------------------------------------------------------
|
||||
|
@ -482,6 +484,7 @@ void AES::CTR::crypt(const void *const input,unsigned int len) noexcept
|
|||
if (!len) {
|
||||
_ctr[0] = c0;
|
||||
_ctr[1] = Utils::hton(c1);
|
||||
_len = totalLen;
|
||||
return;
|
||||
}
|
||||
--len;
|
||||
|
@ -514,167 +517,6 @@ void AES::CTR::crypt(const void *const input,unsigned int len) noexcept
|
|||
out += totalLen;
|
||||
_len = (totalLen + len);
|
||||
|
||||
#if 0
|
||||
// This is the largest chunk size that will fit in SSE registers with four
|
||||
// registers left over for round key data and temporaries.
|
||||
while (len >= 192) {
|
||||
__m128i d0,d1,d2,d3,d4,d5,d6,d7,d8,d9,d10,d11;
|
||||
if (likely(c1 < 0xfffffffffffffff4ULL)) {
|
||||
d0 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
d1 = _mm_set_epi64x((long long)Utils::hton(c1 + 1ULL),(long long)c0);
|
||||
d2 = _mm_set_epi64x((long long)Utils::hton(c1 + 2ULL),(long long)c0);
|
||||
d3 = _mm_set_epi64x((long long)Utils::hton(c1 + 3ULL),(long long)c0);
|
||||
d4 = _mm_set_epi64x((long long)Utils::hton(c1 + 4ULL),(long long)c0);
|
||||
d5 = _mm_set_epi64x((long long)Utils::hton(c1 + 5ULL),(long long)c0);
|
||||
d6 = _mm_set_epi64x((long long)Utils::hton(c1 + 6ULL),(long long)c0);
|
||||
d7 = _mm_set_epi64x((long long)Utils::hton(c1 + 7ULL),(long long)c0);
|
||||
d8 = _mm_set_epi64x((long long)Utils::hton(c1 + 8ULL),(long long)c0);
|
||||
d9 = _mm_set_epi64x((long long)Utils::hton(c1 + 9ULL),(long long)c0);
|
||||
d10 = _mm_set_epi64x((long long)Utils::hton(c1 + 10ULL),(long long)c0);
|
||||
d11 = _mm_set_epi64x((long long)Utils::hton(c1 + 11ULL),(long long)c0);
|
||||
c1 += 12;
|
||||
} else {
|
||||
d0 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d1 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d2 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d3 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d4 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d5 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d6 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d7 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d8 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d9 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d10 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
d11 = _mm_set_epi64x((long long)Utils::hton(c1),(long long)c0);
|
||||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
}
|
||||
|
||||
{
|
||||
__m128i k0 = _aes._k.ni.k[0];
|
||||
__m128i k1 = _aes._k.ni.k[1];
|
||||
d0 = _mm_xor_si128(d0,k0);
|
||||
d1 = _mm_xor_si128(d1,k0);
|
||||
d2 = _mm_xor_si128(d2,k0);
|
||||
d3 = _mm_xor_si128(d3,k0);
|
||||
d4 = _mm_xor_si128(d4,k0);
|
||||
d5 = _mm_xor_si128(d5,k0);
|
||||
d6 = _mm_xor_si128(d6,k0);
|
||||
d7 = _mm_xor_si128(d7,k0);
|
||||
d8 = _mm_xor_si128(d8,k0);
|
||||
d9 = _mm_xor_si128(d9,k0);
|
||||
d10 = _mm_xor_si128(d10,k0);
|
||||
d11 = _mm_xor_si128(d11,k0);
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
d4 = _mm_aesenc_si128(d4,k1);
|
||||
d5 = _mm_aesenc_si128(d5,k1);
|
||||
d6 = _mm_aesenc_si128(d6,k1);
|
||||
d7 = _mm_aesenc_si128(d7,k1);
|
||||
d8 = _mm_aesenc_si128(d8,k1);
|
||||
d9 = _mm_aesenc_si128(d9,k1);
|
||||
d10 = _mm_aesenc_si128(d10,k1);
|
||||
d11 = _mm_aesenc_si128(d11,k1);
|
||||
for (int r=2;r<14;r+=2) {
|
||||
k0 = _aes._k.ni.k[r];
|
||||
k1 = _aes._k.ni.k[r+1];
|
||||
d0 = _mm_aesenc_si128(d0,k0);
|
||||
d1 = _mm_aesenc_si128(d1,k0);
|
||||
d2 = _mm_aesenc_si128(d2,k0);
|
||||
d3 = _mm_aesenc_si128(d3,k0);
|
||||
d4 = _mm_aesenc_si128(d4,k0);
|
||||
d5 = _mm_aesenc_si128(d5,k0);
|
||||
d6 = _mm_aesenc_si128(d6,k0);
|
||||
d7 = _mm_aesenc_si128(d7,k0);
|
||||
d8 = _mm_aesenc_si128(d8,k0);
|
||||
d9 = _mm_aesenc_si128(d9,k0);
|
||||
d10 = _mm_aesenc_si128(d10,k0);
|
||||
d11 = _mm_aesenc_si128(d11,k0);
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
d4 = _mm_aesenc_si128(d4,k1);
|
||||
d5 = _mm_aesenc_si128(d5,k1);
|
||||
d6 = _mm_aesenc_si128(d6,k1);
|
||||
d7 = _mm_aesenc_si128(d7,k1);
|
||||
d8 = _mm_aesenc_si128(d8,k1);
|
||||
d9 = _mm_aesenc_si128(d9,k1);
|
||||
d10 = _mm_aesenc_si128(d10,k1);
|
||||
d11 = _mm_aesenc_si128(d11,k1);
|
||||
}
|
||||
k0 = _aes._k.ni.k[14];
|
||||
d0 = _mm_aesenclast_si128(d0,k0);
|
||||
d1 = _mm_aesenclast_si128(d1,k0);
|
||||
d2 = _mm_aesenclast_si128(d2,k0);
|
||||
d3 = _mm_aesenclast_si128(d3,k0);
|
||||
d4 = _mm_aesenclast_si128(d4,k0);
|
||||
d5 = _mm_aesenclast_si128(d5,k0);
|
||||
d6 = _mm_aesenclast_si128(d6,k0);
|
||||
d7 = _mm_aesenclast_si128(d7,k0);
|
||||
d8 = _mm_aesenclast_si128(d8,k0);
|
||||
d9 = _mm_aesenclast_si128(d9,k0);
|
||||
d10 = _mm_aesenclast_si128(d10,k0);
|
||||
d11 = _mm_aesenclast_si128(d11,k0);
|
||||
}
|
||||
|
||||
{
|
||||
__m128i p0 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in));
|
||||
__m128i p1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 16));
|
||||
__m128i p2 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 32));
|
||||
__m128i p3 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 48));
|
||||
p0 = _mm_xor_si128(d0,p0);
|
||||
p1 = _mm_xor_si128(d1,p1);
|
||||
p2 = _mm_xor_si128(d2,p2);
|
||||
p3 = _mm_xor_si128(d3,p3);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out),p0);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 16),p1);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 32),p2);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 48),p3);
|
||||
p0 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 64));
|
||||
p1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 80));
|
||||
p2 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 96));
|
||||
p3 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 112));
|
||||
p0 = _mm_xor_si128(d4,p0);
|
||||
p1 = _mm_xor_si128(d5,p1);
|
||||
p2 = _mm_xor_si128(d6,p2);
|
||||
p3 = _mm_xor_si128(d7,p3);
|
||||
d0 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 128));
|
||||
d1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 144));
|
||||
d2 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 160));
|
||||
d3 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 176));
|
||||
d0 = _mm_xor_si128(d8,d0);
|
||||
d1 = _mm_xor_si128(d9,d1);
|
||||
d2 = _mm_xor_si128(d10,d2);
|
||||
d3 = _mm_xor_si128(d11,d3);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 64),p0);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 80),p1);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 96),p2);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 112),p3);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 128),d0);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 144),d1);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 160),d2);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(out + 176),d3);
|
||||
}
|
||||
|
||||
in += 192;
|
||||
len -= 192;
|
||||
out += 192;
|
||||
}
|
||||
#endif
|
||||
|
||||
while (len >= 64) {
|
||||
__m128i d0,d1,d2,d3;
|
||||
if (likely(c1 < 0xfffffffffffffffcULL)) {
|
||||
|
@ -694,40 +536,86 @@ void AES::CTR::crypt(const void *const input,unsigned int len) noexcept
|
|||
if (unlikely(++c1 == 0ULL)) c0 = Utils::hton(Utils::ntoh(c0) + 1ULL);
|
||||
}
|
||||
|
||||
{
|
||||
__m128i k0 = _aes._k.ni.k[0];
|
||||
__m128i k1 = _aes._k.ni.k[1];
|
||||
d0 = _mm_xor_si128(d0,k0);
|
||||
d1 = _mm_xor_si128(d1,k0);
|
||||
d2 = _mm_xor_si128(d2,k0);
|
||||
d3 = _mm_xor_si128(d3,k0);
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
for (int r=2;r<14;r+=2) {
|
||||
k0 = _aes._k.ni.k[r];
|
||||
k1 = _aes._k.ni.k[r+1];
|
||||
d0 = _mm_aesenc_si128(d0,k0);
|
||||
d1 = _mm_aesenc_si128(d1,k0);
|
||||
d2 = _mm_aesenc_si128(d2,k0);
|
||||
d3 = _mm_aesenc_si128(d3,k0);
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
}
|
||||
k0 = _aes._k.ni.k[14];
|
||||
d0 = _mm_aesenclast_si128(d0,k0);
|
||||
d1 = _mm_aesenclast_si128(d1,k0);
|
||||
d2 = _mm_aesenclast_si128(d2,k0);
|
||||
d3 = _mm_aesenclast_si128(d3,k0);
|
||||
}
|
||||
|
||||
__m128i k0 = _aes._k.ni.k[0];
|
||||
__m128i k1 = _aes._k.ni.k[1];
|
||||
__m128i k2 = _aes._k.ni.k[2];
|
||||
__m128i k3 = _aes._k.ni.k[3];
|
||||
d0 = _mm_xor_si128(d0,k0);
|
||||
d1 = _mm_xor_si128(d1,k0);
|
||||
d2 = _mm_xor_si128(d2,k0);
|
||||
d3 = _mm_xor_si128(d3,k0);
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
k0 = _aes._k.ni.k[4];
|
||||
k1 = _aes._k.ni.k[5];
|
||||
d0 = _mm_aesenc_si128(d0,k2);
|
||||
d1 = _mm_aesenc_si128(d1,k2);
|
||||
d2 = _mm_aesenc_si128(d2,k2);
|
||||
d3 = _mm_aesenc_si128(d3,k2);
|
||||
d0 = _mm_aesenc_si128(d0,k3);
|
||||
d1 = _mm_aesenc_si128(d1,k3);
|
||||
d2 = _mm_aesenc_si128(d2,k3);
|
||||
d3 = _mm_aesenc_si128(d3,k3);
|
||||
k2 = _aes._k.ni.k[6];
|
||||
k3 = _aes._k.ni.k[7];
|
||||
d0 = _mm_aesenc_si128(d0,k0);
|
||||
d1 = _mm_aesenc_si128(d1,k0);
|
||||
d2 = _mm_aesenc_si128(d2,k0);
|
||||
d3 = _mm_aesenc_si128(d3,k0);
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
k0 = _aes._k.ni.k[8];
|
||||
k1 = _aes._k.ni.k[9];
|
||||
d0 = _mm_aesenc_si128(d0,k2);
|
||||
d1 = _mm_aesenc_si128(d1,k2);
|
||||
d2 = _mm_aesenc_si128(d2,k2);
|
||||
d3 = _mm_aesenc_si128(d3,k2);
|
||||
d0 = _mm_aesenc_si128(d0,k3);
|
||||
d1 = _mm_aesenc_si128(d1,k3);
|
||||
d2 = _mm_aesenc_si128(d2,k3);
|
||||
d3 = _mm_aesenc_si128(d3,k3);
|
||||
k2 = _aes._k.ni.k[10];
|
||||
k3 = _aes._k.ni.k[11];
|
||||
d0 = _mm_aesenc_si128(d0,k0);
|
||||
d1 = _mm_aesenc_si128(d1,k0);
|
||||
d2 = _mm_aesenc_si128(d2,k0);
|
||||
d3 = _mm_aesenc_si128(d3,k0);
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
k0 = _aes._k.ni.k[12];
|
||||
k1 = _aes._k.ni.k[13];
|
||||
d0 = _mm_aesenc_si128(d0,k2);
|
||||
d1 = _mm_aesenc_si128(d1,k2);
|
||||
d2 = _mm_aesenc_si128(d2,k2);
|
||||
d3 = _mm_aesenc_si128(d3,k2);
|
||||
__m128i p0 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in));
|
||||
__m128i p1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 16));
|
||||
d0 = _mm_aesenc_si128(d0,k3);
|
||||
d1 = _mm_aesenc_si128(d1,k3);
|
||||
d2 = _mm_aesenc_si128(d2,k3);
|
||||
d3 = _mm_aesenc_si128(d3,k3);
|
||||
k2 = _aes._k.ni.k[14];
|
||||
d0 = _mm_aesenc_si128(d0,k0);
|
||||
d1 = _mm_aesenc_si128(d1,k0);
|
||||
d2 = _mm_aesenc_si128(d2,k0);
|
||||
d3 = _mm_aesenc_si128(d3,k0);
|
||||
__m128i p2 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 32));
|
||||
__m128i p3 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(in + 48));
|
||||
d0 = _mm_aesenc_si128(d0,k1);
|
||||
d1 = _mm_aesenc_si128(d1,k1);
|
||||
d2 = _mm_aesenc_si128(d2,k1);
|
||||
d3 = _mm_aesenc_si128(d3,k1);
|
||||
d0 = _mm_aesenclast_si128(d0,k2);
|
||||
d1 = _mm_aesenclast_si128(d1,k2);
|
||||
d2 = _mm_aesenclast_si128(d2,k2);
|
||||
d3 = _mm_aesenclast_si128(d3,k2);
|
||||
|
||||
p0 = _mm_xor_si128(d0,p0);
|
||||
p1 = _mm_xor_si128(d1,p1);
|
||||
p2 = _mm_xor_si128(d2,p2);
|
||||
|
@ -836,11 +724,6 @@ void AES::CTR::finish() noexcept
|
|||
// Encrypt any remaining bytes as indicated by _len not being an even multiple of 16.
|
||||
if (rem) {
|
||||
uint8_t tmp[16];
|
||||
for (unsigned int i = 0,j = _len - rem;i < rem;++i)
|
||||
tmp[i] = _out[j];
|
||||
for (unsigned int i = rem;i < 16;++i)
|
||||
tmp[i] = 0;
|
||||
|
||||
__m128i d0 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(_ctr));
|
||||
d0 = _mm_xor_si128(d0,_aes._k.ni.k[0]);
|
||||
d0 = _mm_aesenc_si128(d0,_aes._k.ni.k[1]);
|
||||
|
@ -857,24 +740,19 @@ void AES::CTR::finish() noexcept
|
|||
d0 = _mm_aesenc_si128(d0,_aes._k.ni.k[12]);
|
||||
d0 = _mm_aesenc_si128(d0,_aes._k.ni.k[13]);
|
||||
d0 = _mm_aesenclast_si128(d0,_aes._k.ni.k[14]);
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(tmp),_mm_xor_si128(_mm_loadu_si128(reinterpret_cast<__m128i *>(tmp)),d0));
|
||||
|
||||
for (unsigned int i = 0,j = _len - rem;i < rem;++i)
|
||||
_out[j] = tmp[i];
|
||||
_mm_storeu_si128(reinterpret_cast<__m128i *>(tmp),d0);
|
||||
for (unsigned int i=0,j=_len-rem;i<rem;++i)
|
||||
_out[j+i] ^= tmp[i];
|
||||
}
|
||||
return;
|
||||
}
|
||||
#endif
|
||||
|
||||
if (rem) {
|
||||
uint8_t tmp[16],keyStream[16];
|
||||
for (unsigned int i = 0,j = _len - rem;i < rem;++i)
|
||||
tmp[i] = _out[j];
|
||||
for (unsigned int i = rem;i < 16;++i)
|
||||
tmp[i] = 0;
|
||||
_aes._encryptSW(reinterpret_cast<const uint8_t *>(_ctr),keyStream);
|
||||
for (unsigned int i = 0,j = _len - rem;i < rem;++i)
|
||||
_out[j] = tmp[i] ^ keyStream[i];
|
||||
uint8_t tmp[16];
|
||||
_aes._encryptSW(reinterpret_cast<const uint8_t *>(_ctr),tmp);
|
||||
for (unsigned int i=0,j=_len-rem;i<rem;++i)
|
||||
_out[j+i] ^= tmp[i];
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
@ -37,6 +37,12 @@ namespace ZeroTier {
|
|||
class Salsa20 : public TriviallyCopyable
|
||||
{
|
||||
public:
|
||||
#ifdef ZT_SALSA20_SSE
|
||||
static constexpr bool accelerated() noexcept { return true; }
|
||||
#else
|
||||
static constexpr bool accelerated() noexcept { return false; }
|
||||
#endif
|
||||
|
||||
ZT_ALWAYS_INLINE Salsa20() noexcept {}
|
||||
ZT_ALWAYS_INLINE ~Salsa20() { Utils::burn(&_state,sizeof(_state)); }
|
||||
|
||||
|
|
133
node/Tests.cpp
133
node/Tests.cpp
|
@ -42,10 +42,37 @@
|
|||
#include <cstdint>
|
||||
#include <cstring>
|
||||
#include <cstdio>
|
||||
#include <ctime>
|
||||
#include <set>
|
||||
|
||||
#ifdef __UNIX_LIKE__
|
||||
#include <unistd.h>
|
||||
#include <sys/time.h>
|
||||
#include <sys/types.h>
|
||||
#endif
|
||||
|
||||
using namespace ZeroTier;
|
||||
|
||||
static volatile unsigned int foo = 0;
|
||||
|
||||
static int64_t now()
|
||||
{
|
||||
#ifdef __WINDOWS__
|
||||
FILETIME ft;
|
||||
SYSTEMTIME st;
|
||||
ULARGE_INTEGER tmp;
|
||||
GetSystemTime(&st);
|
||||
SystemTimeToFileTime(&st,&ft);
|
||||
tmp.LowPart = ft.dwLowDateTime;
|
||||
tmp.HighPart = ft.dwHighDateTime;
|
||||
return (int64_t)( ((tmp.QuadPart - 116444736000000000LL) / 10000L) + st.wMilliseconds );
|
||||
#else
|
||||
timeval tv;
|
||||
gettimeofday(&tv,nullptr);
|
||||
return ( (1000LL * (int64_t)tv.tv_sec) + (int64_t)(tv.tv_usec / 1000) );
|
||||
#endif
|
||||
};
|
||||
|
||||
// Cryptographic tests vectors ----------------------------------------------------------------------------------------
|
||||
|
||||
static const uint8_t ECC384_TV0_PUBLIC[49] = { 0x02,0xed,0xbc,0xbb,0x1f,0x23,0x9b,0xbd,0x9d,0x3d,0x7c,0xef,0x6b,0x37,0xa3,0x26,0x69,0xe9,0x4d,0xf4,0x26,0x64,0xfb,0xac,0x76,0x40,0xc2,0x22,0x21,0xa6,0xa3,0xdf,0x8c,0x96,0x81,0x76,0x0f,0x0e,0x67,0xab,0xd4,0x51,0x58,0xb3,0x15,0x63,0xfb,0x49,0x71 };
|
||||
|
@ -82,6 +109,9 @@ static const uint8_t AES_CTR_TEST_VECTOR_0_IV[16] = { 0xf0,0xf1,0xf2,0xf3,0xf4,0
|
|||
static const uint8_t AES_CTR_TEST_VECTOR_0_IN[64] = { 0x6b,0xc1,0xbe,0xe2,0x2e,0x40,0x9f,0x96,0xe9,0x3d,0x7e,0x11,0x73,0x93,0x17,0x2a,0xae,0x2d,0x8a,0x57,0x1e,0x03,0xac,0x9c,0x9e,0xb7,0x6f,0xac,0x45,0xaf,0x8e,0x51,0x30,0xc8,0x1c,0x46,0xa3,0x5c,0xe4,0x11,0xe5,0xfb,0xc1,0x19,0x1a,0x0a,0x52,0xef,0xf6,0x9f,0x24,0x45,0xdf,0x4f,0x9b,0x17,0xad,0x2b,0x41,0x7b,0xe6,0x6c,0x37,0x10 };
|
||||
static const uint8_t AES_CTR_TEST_VECTOR_0_OUT[64] = { 0x60,0x1e,0xc3,0x13,0x77,0x57,0x89,0xa5,0xb7,0xa7,0xf5,0x04,0xbb,0xf3,0xd2,0x28,0xf4,0x43,0xe3,0xca,0x4d,0x62,0xb5,0x9a,0xca,0x84,0xe9,0x90,0xca,0xca,0xf5,0xc5,0x2b,0x09,0x30,0xda,0xa2,0x3d,0xe9,0x4c,0xe8,0x70,0x17,0xba,0x2d,0x84,0x98,0x8d,0xdf,0xc9,0xc5,0x8d,0xb6,0x7a,0xad,0xa6,0x13,0xc2,0xdd,0x08,0x45,0x79,0x41,0xa6 };
|
||||
|
||||
// Key and IV are same as test vector 0, input is a buffer filled by iterating 0..777 and setting to these &0xff
|
||||
static const uint8_t AES_CTR_TEST_VECTOR_1_OUT[777] = { 0x0b,0xde,0x7f,0xf2,0x5d,0x12,0x10,0x34,0x56,0x93,0x81,0x1e,0xc4,0x6d,0xcb,0x0d,0x4a,0x7f,0x7b,0x8e,0x47,0x74,0x0f,0x11,0x4c,0x2a,0x9c,0x27,0x93,0x78,0x65,0x8b,0x3b,0xe0,0x0e,0xbf,0x25,0x44,0x2b,0x7a,0x25,0xa2,0xfc,0x88,0x1b,0xa3,0xe4,0x4d,0x19,0x67,0xd3,0xfb,0x5d,0x00,0x00,0x86,0x86,0xd0,0xa6,0x48,0x9f,0x28,0x48,0x89,0xcb,0x36,0xbd,0xa3,0x9d,0x39,0x4f,0xd5,0x9f,0xbe,0x44,0x57,0xa5,0x82,0x8d,0xf8,0x1c,0xba,0x35,0xd1,0x34,0x53,0x5e,0xd4,0x34,0xfb,0x91,0x1e,0x79,0xc7,0xdb,0xf2,0x79,0xaf,0x31,0x9f,0x54,0x70,0xe7,0x15,0xbd,0x3e,0x76,0x2b,0x82,0x2c,0x37,0x07,0x44,0x1e,0x5b,0x7a,0xca,0xb8,0x17,0x74,0x1c,0x5e,0xa6,0xe0,0x57,0xaa,0x13,0x99,0x5e,0x3c,0x11,0xfe,0xce,0xeb,0x6b,0x8e,0x5e,0xc6,0x79,0xa2,0xa7,0x4c,0x00,0xff,0x0a,0x5d,0xbb,0x3b,0xc5,0x76,0xe7,0x9a,0x53,0x76,0x67,0xee,0x8e,0x73,0x80,0xa8,0x6d,0xad,0x38,0x73,0x62,0x71,0x76,0x8a,0x3d,0x5e,0x42,0xae,0xb8,0x7d,0x61,0xd3,0x1d,0x97,0xd7,0xda,0x51,0xf5,0x89,0x26,0xed,0x45,0x75,0x88,0x65,0x69,0x06,0xe8,0x3d,0x31,0x82,0xe0,0xa4,0x82,0x79,0xf0,0x44,0x7c,0x36,0x4a,0xd0,0x25,0x13,0x30,0x1a,0x1e,0x52,0x5c,0x0e,0xe4,0x68,0x93,0xbf,0x1e,0x5a,0x43,0xc1,0xe6,0x7f,0x8a,0xf7,0xff,0xb8,0x95,0x41,0x6b,0x28,0xbc,0x2a,0xac,0x9e,0x8e,0x7f,0x7c,0xc1,0xd9,0xed,0x0b,0x52,0x55,0x28,0x6c,0xb4,0x15,0x96,0xbe,0xfc,0xe9,0xfd,0x3a,0x84,0xc9,0xd6,0x4f,0x40,0xe4,0xe0,0x59,0xc4,0xfe,0x62,0x8a,0x2c,0xbb,0xcd,0xc2,0x5b,0x4b,0x53,0x67,0x0d,0xc3,0x6d,0x0f,0x67,0xfe,0x20,0x77,0x43,0x98,0x92,0x13,0x15,0x0f,0x9b,0x2c,0x4b,0xa9,0x53,0x2a,0xb2,0xd0,0x72,0x64,0x5e,0x92,0x7f,0x25,0xde,0x0a,0x93,0x46,0x03,0x33,0xdc,0x43,0x4f,0xfa,0x00,0x8a,0xfc,0xcb,0x42,0x3d,0x5f,0x25,0xbc,0xa2,0x78,0xdd,0x3d,0x0c,0x82,0x52,0xd8,0x0a,0x32,0xee,0xc7,0xd8,0x46,0x24,0x63,0xc3,0x33,0xba,0x66,0x0b,0x20,0x74,0xc6,0x06,0x5b,0x2c,0xea,0x06,0xed,0x36,0xda,0xbc,0x3f,0x0a,0xb0,0xa5,0xad,0xd0,0xd6,0x9a,0x33,0x49,0xbe,0xb6,0x94,0xa2,0xef,0x82,0xa4,0x1b,0x81,0x71,0xb9,0xea,0x37,0xfe,0x43,0x48,0xa1,0x30,0x92,0x6f,0x69,0x45,0xc6,0xf7,0xdd,0x0d,0x10,0x3b,0x71,0x59,0x8c,0xfc,0x18,0xf2,0x48,0x21,0xf3,0x6c,0xa8,0xaa,0x33,0xff,0xf4,0x38,0x96,0x5b,0x34,0x43,0x7f,0xcc,0x9c,0x87,0x36,0x3b,0x96,0x3e,0x1a,0xbf,0x6d,0xa1,0x89,0x42,0xb3,0xb9,0x64,0x9e,0xa3,0xef,0x36,0x4a,0x41,0xb9,0xa3,0xb9,0xad,0xa4,0xd9,0x33,0xad,0xa5,0xba,0x41,0x83,0x12,0xc5,0x92,0xf8,0x6a,0x10,0x20,0x1e,0xe0,0xfd,0xe8,0x6d,0xfc,0x4a,0x7c,0x72,0x7f,0x54,0x4d,0x00,0xd5,0x3e,0x6a,0x28,0x94,0x11,0x8a,0x38,0xef,0xb5,0xb6,0xf3,0xbf,0xd2,0xbe,0xf3,0x1e,0x8b,0xe8,0x0c,0xf2,0x9d,0xaf,0xff,0x90,0x4c,0x8a,0x44,0xad,0xd9,0x8b,0x99,0x47,0x65,0x31,0x74,0xb2,0x24,0xb3,0x6d,0xd3,0x4a,0x4c,0x19,0xf6,0x2f,0x53,0x2d,0xb8,0x05,0xd3,0x7d,0x53,0xc9,0xc7,0x7e,0x03,0xeb,0xfc,0x18,0x36,0xe2,0x4b,0xcf,0xfd,0xe8,0x97,0xde,0xd8,0x42,0x80,0x05,0x77,0x8c,0xec,0x15,0xae,0x23,0x13,0xe8,0xa5,0x1c,0xca,0x30,0x60,0x8c,0x20,0x86,0xdf,0xa0,0xd7,0x88,0x2c,0x27,0xa0,0x8f,0x99,0x23,0x9c,0x11,0xa9,0xe6,0x4a,0xfc,0x0a,0x89,0xa4,0x60,0x94,0xa9,0x2e,0x40,0x62,0x22,0xd2,0xae,0x31,0x4b,0x52,0x2c,0x5f,0x9a,0xd6,0x54,0x58,0x08,0xa1,0x97,0x1c,0x28,0xef,0x53,0x5d,0x6a,0x4b,0x07,0xe7,0x62,0x06,0xb2,0xa8,0xb4,0x12,0x9a,0x10,0x12,0xa2,0xe3,0x02,0xce,0xc3,0xa7,0x73,0x47,0xf1,0xfb,0xcb,0x77,0xb5,0x33,0x81,0xb9,0xf7,0x79,0x1d,0x93,0x2d,0x2e,0x14,0x94,0x05,0x36,0xe1,0x41,0xcd,0xdc,0x83,0x05,0xea,0xac,0x61,0xe9,0xe5,0xc2,0x7d,0x53,0x44,0x65,0x8a,0x25,0x0d,0xb4,0x66,0x43,0x5c,0xbf,0x6d,0x7f,0xc0,0x46,0xab,0xab,0xfb,0x0e,0xd2,0x33,0xdf,0x67,0x0a,0x1f,0x29,0x29,0x6c,0x1d,0x32,0x7d,0x3a,0xff,0x10,0x59,0x10,0x79,0x68,0x0a,0x04,0x0d,0x1c,0x4e,0xe0,0x6d,0x1b,0x59,0xfc,0x23,0x6b,0x9c,0x6a,0xde,0x1f,0x1c,0x9f,0x0c,0x1c,0x8f,0x5e,0xd1,0x64,0x7e,0x33,0x2f,0xae,0xdf,0x76,0x87,0xab,0x64,0x04,0xd2,0xc3,0xfe,0x4f,0x95,0x47,0xf2,0x16,0x11,0xdb,0x00,0x56,0xb4,0x96,0x0f,0x1e,0x18,0xc6,0xcd,0xa3,0x29,0x8e,0xd0,0xf3,0x0f,0x85,0x2f,0xa0,0xe7,0x8c,0x12,0x2c,0xdc,0x85,0x0b,0xef,0xb3,0x7d,0x59,0x87,0xaa,0x1d,0xfc,0xde,0xd0,0xbc,0x4c,0xe8,0x49,0x11,0x50,0xf5,0x7f };
|
||||
|
||||
static const uint8_t AES_GMAC_VECTOR_0_KEY[32] = { 0xbb, 0x10, 0x10, 0x06, 0x4f, 0xb8, 0x35, 0x23, 0xea, 0x9d, 0xf3, 0x2b, 0xad, 0x9f, 0x1f, 0x2a, 0x4f, 0xce, 0xfc, 0x0f, 0x21, 0x07, 0xc0, 0xaa, 0xba, 0xd9, 0xb7, 0x56, 0xd8, 0x09, 0x21, 0x9d };
|
||||
static const uint8_t AES_GMAC_VECTOR_0_IV[12] = { 0x2f, 0x9a, 0xd0, 0x12, 0xad, 0xfc, 0x12, 0x73, 0x43, 0xfb, 0xe0, 0x56 };
|
||||
static const uint8_t AES_GMAC_VECTOR_0_IN[16] = { 0xdb, 0x98, 0xd9, 0x0d, 0x1b, 0x69, 0x5c, 0xdb, 0x74, 0x7a, 0x34, 0x3f, 0xbb, 0xc9, 0xf1, 0x41 };
|
||||
|
@ -531,9 +561,9 @@ extern "C" const char *ZTT_crypto()
|
|||
}
|
||||
|
||||
{
|
||||
uint8_t out[64];
|
||||
ZT_T_PRINTF("[crypto] Testing AES-CTR (hardware acceleration: %s)... ",AES::accelerated() ? "enabled" : "disabled");
|
||||
ZT_T_PRINTF("[crypto] Testing AES-CTR... ");
|
||||
{
|
||||
uint8_t out[64];
|
||||
AES aes(AES_CTR_TEST_VECTOR_0_KEY);
|
||||
AES::CTR ctr(aes);
|
||||
ctr.init(AES_CTR_TEST_VECTOR_0_IV,out);
|
||||
|
@ -544,6 +574,39 @@ extern "C" const char *ZTT_crypto()
|
|||
return "AES-CTR test vector 0 failed";
|
||||
}
|
||||
}
|
||||
{
|
||||
uint8_t out[777];
|
||||
AES aes(AES_CTR_TEST_VECTOR_0_KEY);
|
||||
AES::CTR ctr(aes);
|
||||
ctr.init(AES_CTR_TEST_VECTOR_0_IV,out);
|
||||
for(unsigned int i=0;i<777;++i)
|
||||
out[i] = (uint8_t)i;
|
||||
ctr.crypt(out,777);
|
||||
ctr.finish();
|
||||
if (memcmp(out,AES_CTR_TEST_VECTOR_1_OUT,777) != 0) {
|
||||
ZT_T_PRINTF("FAILED (test vector 1)" ZT_EOL_S);
|
||||
return "AES-CTR test vector 1 failed";
|
||||
}
|
||||
}
|
||||
{
|
||||
uint8_t out[777];
|
||||
AES aes(AES_CTR_TEST_VECTOR_0_KEY);
|
||||
AES::CTR ctr(aes);
|
||||
ctr.init(AES_CTR_TEST_VECTOR_0_IV,out);
|
||||
for(unsigned int i=0;i<777;++i)
|
||||
out[i] = (uint8_t)i;
|
||||
ctr.crypt(out,111);
|
||||
ctr.crypt(out + 111,111);
|
||||
ctr.crypt(out + 222,222);
|
||||
ctr.crypt(out + 444,111);
|
||||
ctr.crypt(out + 555,221);
|
||||
ctr.crypt(out + 776,1);
|
||||
ctr.finish();
|
||||
if (memcmp(out,AES_CTR_TEST_VECTOR_1_OUT,777) != 0) {
|
||||
ZT_T_PRINTF("FAILED (test vector 1, fragmented)" ZT_EOL_S);
|
||||
return "AES-CTR test vector 1 (fragmented) failed";
|
||||
}
|
||||
}
|
||||
ZT_T_PRINTF("OK" ZT_EOL_S);
|
||||
}
|
||||
|
||||
|
@ -614,6 +677,65 @@ extern "C" const char *ZTT_defragmenter()
|
|||
|
||||
extern "C" const char *ZTT_benchmarkCrypto()
|
||||
{
|
||||
try {
|
||||
uint8_t tmp[16384],tag[16];
|
||||
memset(tmp,0,sizeof(tmp));
|
||||
memset(tag,0,sizeof(tag));
|
||||
|
||||
{
|
||||
ZT_T_PRINTF("[crypto] Benchmarking AES-CTR... ");
|
||||
AES aes(AES_CTR_TEST_VECTOR_0_KEY);
|
||||
AES::CTR ctr(aes);
|
||||
int64_t start = now();
|
||||
for(long i=0;i<500000;++i) {
|
||||
ctr.init(AES_CTR_TEST_VECTOR_0_IV,tmp);
|
||||
ctr.crypt(tmp,sizeof(tmp));
|
||||
ctr.finish();
|
||||
}
|
||||
int64_t end = now();
|
||||
foo = tmp[0]; // prevent optimization
|
||||
ZT_T_PRINTF("%.8f MiB/sec" ZT_EOL_S,((16384.0 * 500000.0) / 1048576.0) / ((double)(end - start) / 1000.0));
|
||||
|
||||
ZT_T_PRINTF("[crypto] Benchmarking AES-GMAC... ");
|
||||
AES::GMAC gmac(aes);
|
||||
start = now();
|
||||
for(long i=0;i<500000;++i) {
|
||||
gmac.init(tag);
|
||||
gmac.update(tmp,sizeof(tmp));
|
||||
gmac.finish(tag);
|
||||
}
|
||||
end = now();
|
||||
foo = tag[0]; // prevent optimization
|
||||
ZT_T_PRINTF("%.8f MiB/sec" ZT_EOL_S,((16384.0 * 500000.0) / 1048576.0) / ((double)(end - start) / 1000.0));
|
||||
}
|
||||
|
||||
{
|
||||
ZT_T_PRINTF("[crypto] Benchmarking Poly1305... ");
|
||||
int64_t start = now();
|
||||
for(long i=0;i<500000;++i)
|
||||
poly1305(tag,tmp,sizeof(tmp),tag);
|
||||
int64_t end = now();
|
||||
foo = tag[0]; // prevent optimization
|
||||
ZT_T_PRINTF("%.8f MiB/sec" ZT_EOL_S,((16384.0 * 500000.0) / 1048576.0) / ((double)(end - start) / 1000.0));
|
||||
}
|
||||
|
||||
{
|
||||
ZT_T_PRINTF("[crypto] Benchmarking Salsa20/12 (using vector acceleration: %s)... ",Salsa20::accelerated() ? "yes" : "no");
|
||||
Salsa20 s20(tmp,tag);
|
||||
int64_t start = now();
|
||||
for(long i=0;i<250000;++i)
|
||||
s20.crypt12(tmp,tmp,sizeof(tmp));
|
||||
int64_t end = now();
|
||||
foo = tmp[0]; // prevent optimization
|
||||
ZT_T_PRINTF("%.8f MiB/sec" ZT_EOL_S,((16384.0 * 250000.0) / 1048576.0) / ((double)(end - start) / 1000.0));
|
||||
}
|
||||
} catch (std::exception &e) {
|
||||
ZT_T_PRINTF(ZT_EOL_S "[crypto] Unexpected exception: %s" ZT_EOL_S,e.what());
|
||||
return e.what();
|
||||
} catch ( ... ) {
|
||||
ZT_T_PRINTF(ZT_EOL_S "[crypto] Unexpected exception: unknown exception" ZT_EOL_S);
|
||||
return "an unknown exception occurred";
|
||||
}
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
|
@ -624,7 +746,12 @@ int main(int argc,char **argv)
|
|||
ok &= ZTT_general() == nullptr;
|
||||
ZT_T_PRINTF(ZT_EOL_S);
|
||||
ok &= ZTT_crypto() == nullptr;
|
||||
return ok ? 0 : 1;
|
||||
ok &= ZTT_benchmarkCrypto() == nullptr;
|
||||
if (!ok) {
|
||||
ZT_T_PRINTF(ZT_EOL_S "At least one error occurred!" ZT_EOL_S);
|
||||
return 1;
|
||||
}
|
||||
return 0;
|
||||
}
|
||||
#endif
|
||||
|
||||
|
|
Loading…
Add table
Reference in a new issue