bloom: use independent hashes

Signed-off-by: Jens Axboe <axboe@fb.com>

diff --git a/Makefile b/Makefile
index d93ce7809af6f7d8fc477c8ac4e49317541556fc..8c424e3df0b25afba4bfdfde8983af3d54cbedb4 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -192,7 +192,8 @@ endif
 ifeq ($(CONFIG_TARGET_OS), Linux)
 T_DEDUPE_OBJS = t/dedupe.o
 T_DEDUPE_OBJS += lib/rbtree.o t/log.o mutex.o smalloc.o gettime.o crc/md5.o \
-               memalign.o lib/bloom.o t/debug.o
+               memalign.o lib/bloom.o t/debug.o crc/xxhash.o crc/crc32c.o \
+               crc/crc32c-intel.o
 T_DEDUPE_PROGS = t/dedupe
 endif
 

diff --git a/crc/xxhash.c b/crc/xxhash.c
index eedaecb4912db91625af09d5f5a617e0dc9bb2ca..4736c528fc07750dd5641721756e1ac860a8f9d2 100644 (file)
--- a/crc/xxhash.c
+++ b/crc/xxhash.c
@@ -221,7 +221,7 @@ static uint32_t XXH32_endian_align(const void* input, int len, uint32_t seed, XX
 }
 
 
-uint32_t XXH32(const void* input, int len, uint32_t seed)
+uint32_t XXH32(const void* input, uint32_t len, uint32_t seed)
 {
 #if 0
     // Simple version, good for code maintenance, but unfortunately slow for small inputs

diff --git a/crc/xxhash.h b/crc/xxhash.h
index e80a91d094b2598c161ad0b475c52a76e891a07a..8850d208a00fbfbf08d7f89ffe617f2f49bdd1c3 100644 (file)
--- a/crc/xxhash.h
+++ b/crc/xxhash.h
@@ -88,7 +88,7 @@ typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode;
 // Simple Hash Functions
 //****************************
 
-unsigned int XXH32 (const void* input, int len, unsigned int seed);
+uint32_t XXH32 (const void* input, uint32_t len, uint32_t seed);
 
 /*
 XXH32() :

diff --git a/lib/bloom.c b/lib/bloom.c
index 9ab89e4b6f29af482f210f98bd2a088a64d9c416..b469fdedac0d57cc133bbbdc4067032a4950e13f 100644 (file)
--- a/lib/bloom.c
+++ b/lib/bloom.c
@@ -4,6 +4,8 @@
 #include "bloom.h"
 #include "../hash.h"
 #include "../minmax.h"
+#include "../crc/xxhash.h"
+#include "../crc/crc32c.h"
 
 struct bloom {
        uint64_t nentries;
@@ -14,7 +16,31 @@ struct bloom {
 #define BITS_PER_INDEX (sizeof(uint32_t) * 8)
 #define BITS_INDEX_MASK        (BITS_PER_INDEX - 1)
 
-static unsigned int jhash_init[] = { 0, 0x12db635, 0x2a4a53 };
+struct bloom_hash {
+       unsigned int seed;
+       uint32_t (*fn)(const void *, uint32_t, uint32_t);
+};
+
+static uint32_t b_crc32c(const void *buf, uint32_t len, uint32_t seed)
+{
+       return fio_crc32c(buf, len);
+}
+
+struct bloom_hash hashes[] = {
+       {
+               .seed = 0x8989,
+               .fn = jhash,
+       },
+       {
+               .seed = 0x8989,
+               .fn = XXH32,
+       },
+       {
+               .seed = 0,
+               .fn = b_crc32c,
+       },
+};
+
 #define N_HASHES       3
 
 #define MIN_ENTRIES    1073741824UL
@@ -24,6 +50,8 @@ struct bloom *bloom_new(uint64_t entries)
        struct bloom *b;
        size_t no_uints;
 
+       crc32c_intel_probe();
+
        b = malloc(sizeof(*b));
        b->nentries = entries;
        no_uints = (entries + BITS_PER_INDEX - 1) / BITS_PER_INDEX;
@@ -46,16 +74,18 @@ void bloom_free(struct bloom *b)
 static int __bloom_check(struct bloom *b, uint32_t *data, unsigned int nwords,
                         int set)
 {
-       uint32_t hashes[N_HASHES];
+       uint32_t hash[N_HASHES];
        int i, was_set;
 
-       for (i = 0; i < N_HASHES; i++)
-               hashes[i] = jhash(data, nwords, jhash_init[i]) % b->nentries;
+       for (i = 0; i < N_HASHES; i++) {
+               hash[i] = hashes[i].fn(data, nwords, hashes[i].seed);
+               hash[i] = hash[i] % b->nentries;
+       }
 
        was_set = 0;
        for (i = 0; i < N_HASHES; i++) {
-               const unsigned int index = hashes[i] / BITS_PER_INDEX;
-               const unsigned int bit = hashes[i] & BITS_INDEX_MASK;
+               const unsigned int index = hash[i] / BITS_PER_INDEX;
+               const unsigned int bit = hash[i] & BITS_INDEX_MASK;
 
                if (b->map[index] & (1U << bit))
                        was_set++;