Mirror
/
mergerfs
mirror of https://github.com/trapexit/mergerfs.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
Browse Source

change from fasthash64 to wyhash 

wyhash has very good performance for small keys (our usecase) and
works on platforms with alignment concerns.

A user had an issue where fasthash64 lead to misaligned reads and
signal exceptions. wyhash does not have such issues and is faster.
pull/805/head
Antonio SJ Musumeci 4 years ago
parent
58a8f8326d
commit
0371b047f9
5 changed files with 151 additions and 149 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 84
      src/fasthash.cpp
  2. 56
      src/fasthash.h
  3. 12
      src/fs_inode.cpp
  4. 4
      src/hashset.hpp
  5. 140
      src/wyhash.h

84
src/fasthash.cpp
View File

@ -1,84 +0,0 @@
/* The MIT License
Copyright (C) 2012 Zilong Tan (eric.zltan@gmail.com)
Permission is hereby granted, free of charge, to any person
obtaining a copy of this software and associated documentation
files (the "Software"), to deal in the Software without
restriction, including without limitation the rights to use, copy,
modify, merge, publish, distribute, sublicense, and/or sell copies
of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be
included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
*/
#include "fasthash.h"
// Compression function for Merkle-Damgard construction.
// This function is generated using the framework provided.
#define mix(h) ({ \
(h) ^= (h) >> 23; \
(h) *= 0x2127599bf4325c37ULL; \
(h) ^= (h) >> 47; })
uint64_t
fasthash64(const void *buf,
size_t len,
uint64_t seed)
{
const uint64_t m = 0x880355f21e6d1965ULL;
const uint64_t *pos = (const uint64_t *)buf;
const uint64_t *end = pos + (len / 8);
const unsigned char *pos2;
uint64_t h = seed ^ (len * m);
uint64_t v;
while (pos != end) {
v = *pos++;
h ^= mix(v);
h *= m;
}
pos2 = (const unsigned char*)pos;
v = 0;
switch (len & 7) {
case 7: v ^= (uint64_t)pos2[6] << 48;
case 6: v ^= (uint64_t)pos2[5] << 40;
case 5: v ^= (uint64_t)pos2[4] << 32;
case 4: v ^= (uint64_t)pos2[3] << 24;
case 3: v ^= (uint64_t)pos2[2] << 16;
case 2: v ^= (uint64_t)pos2[1] << 8;
case 1: v ^= (uint64_t)pos2[0];
h ^= mix(v);
h *= m;
}
return mix(h);
}
uint32_t
fasthash32(const void *buf,
size_t len,
uint32_t seed)
{
// the following trick converts the 64-bit hashcode to Fermat
// residue, which shall retain information from both the higher
// and lower parts of hashcode.
uint64_t h;
h = fasthash64(buf, len, seed);
return (h - (h >> 32));
}

56
src/fasthash.h
View File

@ -1,56 +0,0 @@
/* The MIT License
Copyright (C) 2012 Zilong Tan (eric.zltan@gmail.com)
Permission is hereby granted, free of charge, to any person
obtaining a copy of this software and associated documentation
files (the "Software"), to deal in the Software without
restriction, including without limitation the rights to use, copy,
modify, merge, publish, distribute, sublicense, and/or sell copies
of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be
included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
*/
#ifndef _FASTHASH_H
#define _FASTHASH_H
#include <stdint.h>
#include <stdio.h>
#ifdef __cplusplus
extern "C" {
#endif
/**
* fasthash32 - 32-bit implementation of fasthash
* @buf: data buffer
* @len: data size
* @seed: the seed
*/
uint32_t fasthash32(const void *buf, size_t len, uint32_t seed);
/**
* fasthash64 - 64-bit implementation of fasthash
* @buf: data buffer
* @len: data size
* @seed: the seed
*/
uint64_t fasthash64(const void *buf, size_t len, uint64_t seed);
#ifdef __cplusplus
}
#endif
#endif

12
src/fs_inode.cpp
View File

@ -18,8 +18,8 @@
#include "ef.hpp" #include "ef.hpp"
#include "errno.hpp" #include "errno.hpp"
#include "fasthash.h"
#include "fs_inode.hpp" #include "fs_inode.hpp"
#include "wyhash.h"
#include <string> #include <string>
@ -61,9 +61,10 @@ path_hash(const char *fusepath_,
const dev_t dev_, const dev_t dev_,
const ino_t ino_) const ino_t ino_)
{ {
return fasthash64(fusepath_,
return wyhash(fusepath_,
fusepath_len_, fusepath_len_,
fs::inode::MAGIC);
fs::inode::MAGIC,
_wyp);
} }
static static
@ -98,9 +99,10 @@ devino_hash(const char *fusepath_,
buf[0] = dev_; buf[0] = dev_;
buf[1] = ino_; buf[1] = ino_;
return fasthash64((void*)&buf[0],
return wyhash((void*)&buf[0],
sizeof(buf), sizeof(buf),
fs::inode::MAGIC);
fs::inode::MAGIC,
_wyp);
} }
static static

4
src/hashset.hpp
View File

@ -19,7 +19,7 @@
#pragma once #pragma once
#include "khash.h" #include "khash.h"
#include "fasthash.h"
#include "wyhash.h"
KHASH_SET_INIT_INT64(hashset); KHASH_SET_INIT_INT64(hashset);
@ -45,7 +45,7 @@ public:
uint64_t h; uint64_t h;
khint_t key; khint_t key;
h = fasthash64(str_,len_,0x7472617065786974);
h = wyhash(str_,len_,0x7472617065786974,_wyp);
key = kh_put(hashset,_set,h,&rv); key = kh_put(hashset,_set,h,&rv);
if(rv == 0) if(rv == 0)

140
src/wyhash.h
View File

@ -0,0 +1,140 @@
//Author: Wang Yi <godspeed_china@yeah.net>
#ifndef wyhash_final_version
#define wyhash_final_version
//defines that change behavior
#ifndef WYHASH_CONDOM
#define WYHASH_CONDOM 0 //0,1,2
#endif
#define WYHASH_32BIT_MUM 0 //faster on 32 bit system
//includes
#include <stdint.h>
#include <string.h>
#if defined(_MSC_VER) && defined(_M_X64)
#include <intrin.h>
#pragma intrinsic(_umul128)
#endif
#if defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
#define _likely_(x) __builtin_expect(x,1)
#define _unlikely_(x) __builtin_expect(x,0)
#else
#define _likely_(x) (x)
#define _unlikely_(x) (x)
#endif
//mum function
static inline uint64_t _wyrot(uint64_t x) { return (x>>32)|(x<<32); }
static inline void _wymum(uint64_t *A, uint64_t *B){
#if(WYHASH_32BIT_MUM)
uint64_t hh=(*A>>32)*(*B>>32), hl=(*A>>32)*(unsigned)*B, lh=(unsigned)*A*(*B>>32), ll=(uint64_t)(unsigned)*A*(unsigned)*B;
#if(WYHASH_CONDOM>1)
*A^=_wyrot(hl)^hh; *B^=_wyrot(lh)^ll;
#else
*A=_wyrot(hl)^hh; *B=_wyrot(lh)^ll;
#endif
#elif defined(__SIZEOF_INT128__)
__uint128_t r=*A; r*=*B;
#if(WYHASH_CONDOM>1)
*A^=(uint64_t)r; *B^=(uint64_t)(r>>64);
#else
*A=(uint64_t)r; *B=(uint64_t)(r>>64);
#endif
#elif defined(_MSC_VER) && defined(_M_X64)
#if(WYHASH_CONDOM>1)
uint64_t a, b;
a=_umul128(*A,*B,&b);
*A^=a; *B^=b;
#else
*A=_umul128(*A,*B,B);
#endif
#else
uint64_t ha=*A>>32, hb=*B>>32, la=(uint32_t)*A, lb=(uint32_t)*B, hi, lo;
uint64_t rh=ha*hb, rm0=ha*lb, rm1=hb*la, rl=la*lb, t=rl+(rm0<<32), c=t<rl;
lo=t+(rm1<<32); c+=lo<t; hi=rh+(rm0>>32)+(rm1>>32)+c;
#if(WYHASH_CONDOM>1)
*A^=lo; *B^=hi;
#else
*A=lo; *B=hi;
#endif
#endif
}
static inline uint64_t _wymix(uint64_t A, uint64_t B){ _wymum(&A,&B); return A^B; }
//read functions
#ifndef WYHASH_LITTLE_ENDIAN
#if defined(_WIN32) || defined(__LITTLE_ENDIAN__) || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
#define WYHASH_LITTLE_ENDIAN 1
#elif defined(__BIG_ENDIAN__) || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
#define WYHASH_LITTLE_ENDIAN 0
#endif
#endif
#if (WYHASH_LITTLE_ENDIAN)
static inline uint64_t _wyr8(const uint8_t *p) { uint64_t v; memcpy(&v, p, 8); return v;}
static inline uint64_t _wyr4(const uint8_t *p) { unsigned v; memcpy(&v, p, 4); return v;}
#elif defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
static inline uint64_t _wyr8(const uint8_t *p) { uint64_t v; memcpy(&v, p, 8); return __builtin_bswap64(v);}
static inline uint64_t _wyr4(const uint8_t *p) { unsigned v; memcpy(&v, p, 4); return __builtin_bswap32(v);}
#elif defined(_MSC_VER)
static inline uint64_t _wyr8(const uint8_t *p) { uint64_t v; memcpy(&v, p, 8); return _byteswap_uint64(v);}
static inline uint64_t _wyr4(const uint8_t *p) { unsigned v; memcpy(&v, p, 4); return _byteswap_ulong(v);}
#endif
static inline uint64_t _wyr3(const uint8_t *p, unsigned k) { return (((uint64_t)p[0])<<16)|(((uint64_t)p[k>>1])<<8)|p[k-1];}
//wyhash function
static inline uint64_t _wyfinish16(const uint8_t *p, uint64_t len, uint64_t seed, const uint64_t *secret, uint64_t i){
#if(WYHASH_CONDOM>0)
uint64_t a, b;
if(_likely_(i<=8)){
if(_likely_(i>=4)){ a=_wyr4(p); b=_wyr4(p+i-4); }
else if (_likely_(i)){ a=_wyr3(p,i); b=0; }
else a=b=0;
}
else{ a=_wyr8(p); b=_wyr8(p+i-8); }
return _wymix(secret[1]^len,_wymix(a^secret[1], b^seed));
#else
#define oneshot_shift ((i<8)*((8-i)<<3))
return _wymix(secret[1]^len,_wymix((_wyr8(p)<<oneshot_shift)^secret[1],(_wyr8(p+i-8)>>oneshot_shift)^seed));
#endif
}
static inline uint64_t _wyfinish(const uint8_t *p, uint64_t len, uint64_t seed, const uint64_t *secret, uint64_t i){
if(_likely_(i<=16)) return _wyfinish16(p,len,seed,secret,i);
return _wyfinish(p+16,len,_wymix(_wyr8(p)^secret[1],_wyr8(p+8)^seed),secret,i-16);
}
static inline uint64_t wyhash(const void *key, uint64_t len, uint64_t seed, const uint64_t *secret){
const uint8_t *p=(const uint8_t *)key;
uint64_t i=len; seed^=*secret;
if(_unlikely_(i>64)){
uint64_t see1=seed;
do{
seed=_wymix(_wyr8(p)^secret[1],_wyr8(p+8)^seed)^_wymix(_wyr8(p+16)^secret[2],_wyr8(p+24)^seed);
see1=_wymix(_wyr8(p+32)^secret[3],_wyr8(p+40)^see1)^_wymix(_wyr8(p+48)^secret[4],_wyr8(p+56)^see1);
p+=64; i-=64;
}while(i>64);
seed^=see1;
}
return _wyfinish(p,len,seed,secret,i);
}
//utility functions
const uint64_t _wyp[5] = {0xa0761d6478bd642full, 0xe7037ed1a0b428dbull, 0x8ebc6af09c88c6e3ull, 0x589965cc75374cc3ull, 0x1d8e4e27c47d124full};
static inline uint64_t wyhash64(uint64_t A, uint64_t B){ A^=_wyp[0]; B^=_wyp[1]; _wymum(&A,&B); return _wymix(A^_wyp[0],B^_wyp[1]);}
static inline uint64_t wyrand(uint64_t *seed){ *seed+=_wyp[0]; return _wymix(*seed,*seed^_wyp[1]);}
static inline double wy2u01(uint64_t r){ const double _wynorm=1.0/(1ull<<52); return (r>>12)*_wynorm;}
static inline double wy2gau(uint64_t r){ const double _wynorm=1.0/(1ull<<20); return ((r&0x1fffff)+((r>>21)&0x1fffff)+((r>>42)&0x1fffff))*_wynorm-3.0;}
static inline void make_secret(uint64_t seed, uint64_t *secret){
uint8_t c[] = {15, 23, 27, 29, 30, 39, 43, 45, 46, 51, 53, 54, 57, 58, 60, 71, 75, 77, 78, 83, 85, 86, 89, 90, 92, 99, 101, 102, 105, 106, 108, 113, 114, 116, 120, 135, 139, 141, 142, 147, 149, 150, 153, 154, 156, 163, 165, 166, 169, 170, 172, 177, 178, 180, 184, 195, 197, 198, 201, 202, 204, 209, 210, 212, 216, 225, 226, 228, 232, 240 };
for(size_t i=0;i<5;i++){
uint8_t ok;
do{
ok=1; secret[i]=0;
for(size_t j=0;j<64;j+=8) secret[i]|=((uint64_t)c[wyrand(&seed)%sizeof(c)])<<j;
if(secret[i]%2==0){ ok=0; continue; }
for(size_t j=0;j<i;j++)
#if defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
if(__builtin_popcountll(secret[j]^secret[i])!=32){ ok=0; break; }
#elif defined(_MSC_VER)
if(_mm_popcnt_u64(secret[j]^secret[i])!=32){ ok=0; break; }
#endif
if(!ok)continue;
for(uint64_t j=3;j<0x100000000ull;j+=2) if(secret[i]%j==0){ ok=0; break; }
}while(!ok);
}
}
#endif
Write Preview
Loading…
Cancel
Save