X-Git-Url: https://pd.if.org/git/?p=nbds;a=blobdiff_plain;f=struct%2Fhashtable.c;h=cea12326cae2b7dea34e1cb8b3b0543d50ce48d5;hp=953200d33d8e8da1a005249602de17942a814d28;hb=d791fc64b23a9c0a3ed90aa60571344d7735aee9;hpb=598ef6123ccf44d64347ab4c871c6baae7d29bc4 diff --git a/struct/hashtable.c b/struct/hashtable.c index 953200d..cea1232 100644 --- a/struct/hashtable.c +++ b/struct/hashtable.c @@ -13,34 +13,29 @@ */ #include "common.h" -#include "ht.h" #include "murmur.h" #include "mem.h" #include "struct.h" +#include "nstring.h" -#define GET_PTR(x) ((string_t *)((x) & MASK(48))) // low-order 48 bits is a pointer to a string_t +#define GET_PTR(x) ((nstring_t *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t typedef struct ht_entry { - uint64_t key; + uint64_t key; // ptr to nstring_t uint64_t value; } entry_t; -typedef struct string { - uint32_t len; - char val[]; -} string_t; - -typedef struct hash_table_i { +typedef struct hti { volatile entry_t *table; - hash_table_t *ht; // parent ht; - struct hash_table_i *next; - struct hash_table_i *next_free; + hashtable_t *ht; // parent ht; + struct hti *next; + struct hti *next_free; unsigned int scale; int max_probe; int count; // TODO: make these counters distributed int num_entries_copied; int scan; -} hash_table_i_t; +} hashtable_i_t; static const uint64_t COPIED_VALUE = -1; static const uint64_t TOMBSTONE = STRIP_TAG(-1); @@ -50,8 +45,7 @@ static const unsigned ENTRIES_PER_COPY_CHUNK = CACHE_LINE_SIZE/sizeof(entry_t)*2 static const unsigned MIN_SCALE = 4; // min 16 entries (4 buckets) static const unsigned MAX_BUCKETS_TO_PROBE = 250; -static int hti_copy_entry - (hash_table_i_t *ht1, volatile entry_t *e, uint32_t e_key_hash, hash_table_i_t *ht2); +static int hti_copy_entry (hashtable_i_t *ht1, volatile entry_t *e, uint32_t e_key_hash, hashtable_i_t *ht2); // Choose the next bucket to probe using the high-order bits of . static inline int get_next_ndx(int old_ndx, uint32_t key_hash, int ht_scale) { @@ -68,9 +62,7 @@ static inline int get_next_ndx(int old_ndx, uint32_t key_hash, int ht_scale) { static inline int ht_key_equals (uint64_t a, uint32_t b_hash, const char *b_value, uint32_t b_len) { if ((b_hash >> 16) != (a >> 48)) // high-order 16 bits are from the hash value return FALSE; - const string_t *a_key = GET_PTR(a); - assert(a_key); - return a_key->len == b_len && memcmp(a_key->val, b_value, b_len) == 0; + return ns_equalsc(GET_PTR(a), b_value, b_len); } // Lookup in . @@ -81,7 +73,7 @@ static inline int ht_key_equals (uint64_t a, uint32_t b_hash, const char *b_valu // // Record if the entry being returned is empty. Otherwise the caller will have to waste time with // ht_key_equals() to confirm that it did not lose a race to fill an empty entry. -static volatile entry_t *hti_lookup (hash_table_i_t *hti, uint32_t key_hash, const char *key_val, uint32_t key_len, int *is_empty) { +static volatile entry_t *hti_lookup (hashtable_i_t *hti, uint32_t key_hash, const char *key_val, uint32_t key_len, int *is_empty) { TRACE("h2", "hti_lookup(key %p in hti %p)", key_val, hti); *is_empty = 0; @@ -98,14 +90,14 @@ static volatile entry_t *hti_lookup (hash_table_i_t *hti, uint32_t key_hash, con uint64_t e_key = e->key; if (e_key == DOES_NOT_EXIST) { - TRACE("h1", "hti_lookup: entry %p for key \"%s\" is empty", e, GET_PTR(e_key)->val); + TRACE("h1", "hti_lookup: entry %p for key \"%s\" is empty", e, ns_val(GET_PTR(e_key))); *is_empty = 1; // indicate an empty so the caller avoids an expensive ht_key_equals return e; } if (ht_key_equals(e_key, key_hash, key_val, key_len)) { - TRACE("h1", "hti_lookup: entry %p key \"%s\"", e, GET_PTR(e_key)->val); - TRACE("h2", "hti_lookup: entry key len %llu, value %p", GET_PTR(e_key)->len, e->value); + TRACE("h1", "hti_lookup: entry %p key \"%s\"", e, ns_val(GET_PTR(e_key))); + TRACE("h2", "hti_lookup: entry key len %llu, value %p", ns_len(GET_PTR(e_key)), e->value); return e; } } @@ -118,16 +110,16 @@ static volatile entry_t *hti_lookup (hash_table_i_t *hti, uint32_t key_hash, con return NULL; } -// Allocate and initialize a hash_table_i_t with 2^ entries. -static hash_table_i_t *hti_alloc (hash_table_t *parent, int scale) { +// Allocate and initialize a hashtable_i_t with 2^ entries. +static hashtable_i_t *hti_alloc (hashtable_t *parent, int scale) { // Include enough slop to align the actual table on a cache line boundry - size_t n = sizeof(hash_table_i_t) + size_t n = sizeof(hashtable_i_t) + sizeof(entry_t) * (1 << scale) + (CACHE_LINE_SIZE - 1); - hash_table_i_t *hti = (hash_table_i_t *)calloc(n, 1); + hashtable_i_t *hti = (hashtable_i_t *)calloc(n, 1); // Align the table of hash entries on a cache line boundry. - hti->table = (entry_t *)(((uint64_t)hti + sizeof(hash_table_i_t) + (CACHE_LINE_SIZE-1)) + hti->table = (entry_t *)(((uint64_t)hti + sizeof(hashtable_i_t) + (CACHE_LINE_SIZE-1)) & ~(CACHE_LINE_SIZE-1)); hti->scale = scale; @@ -149,8 +141,8 @@ static hash_table_i_t *hti_alloc (hash_table_t *parent, int scale) { // Called when runs out of room for new keys. // -// Initiates a copy by creating a larger hash_table_i_t and installing it in next>. -static void hti_start_copy (hash_table_i_t *hti) { +// Initiates a copy by creating a larger hashtable_i_t and installing it in next>. +static void hti_start_copy (hashtable_i_t *hti) { TRACE("h0", "hti_start_copy(hti %p scale %llu)", hti, hti->scale); // heuristics to determine the size of the new table @@ -160,8 +152,8 @@ static void hti_start_copy (hash_table_i_t *hti) { new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full // Allocate the new table and attempt to install it. - hash_table_i_t *next = hti_alloc(hti->ht, new_scale); - hash_table_i_t *old_next = SYNC_CAS(&hti->next, NULL, next); + hashtable_i_t *next = hti_alloc(hti->ht, new_scale); + hashtable_i_t *old_next = SYNC_CAS(&hti->next, NULL, next); if (old_next != NULL) { // Another thread beat us to it. TRACE("h0", "hti_start_copy: lost race to install new hti; found %p", old_next, 0); @@ -175,8 +167,8 @@ static void hti_start_copy (hash_table_i_t *hti) { // // Return 1 unless is already copied (then return 0), so the caller can account for the total // number of entries left to copy. -static int hti_copy_entry (hash_table_i_t *ht1, volatile entry_t *ht1_e, uint32_t key_hash, - hash_table_i_t *ht2) { +static int hti_copy_entry (hashtable_i_t *ht1, volatile entry_t *ht1_e, uint32_t key_hash, + hashtable_i_t *ht2) { TRACE("h2", "hti_copy_entry: entry %p to table %p", ht1_e, ht2); assert(ht1); assert(ht1->next); @@ -217,25 +209,24 @@ static int hti_copy_entry (hash_table_i_t *ht1, volatile entry_t *ht1_e, uint32_ // to be freed. assert(COPIED_VALUE == TAG_VALUE(TOMBSTONE)); if (ht1_e_value == TOMBSTONE) { - TRACE("h1", "hti_copy_entry: entry %p old value was deleted, now freeing key %p", ht1_e, - GET_PTR(ht1_e->key)); + TRACE("h1", "hti_copy_entry: entry %p old value was deleted, now freeing key %p", ht1_e, GET_PTR(ht1_e->key)); nbd_defer_free(GET_PTR(ht1_e->key)); return TRUE; } // Install the key in the new table. uint64_t key = ht1_e->key; - string_t *key_string = GET_PTR(key); + nstring_t *key_string = GET_PTR(key); uint64_t value = STRIP_TAG(ht1_e_value); // We use 0 to indicate that isn't initiallized. Occasionally the will // really be 0 and we will waste time recomputing it. That is rare enough that it is OK. if (key_hash == 0) { - key_hash = murmur32(key_string->val, key_string->len); + key_hash = murmur32(ns_val(key_string), ns_len(key_string)); } int is_empty; - volatile entry_t *ht2_e = hti_lookup(ht2, key_hash, key_string->val, key_string->len, &is_empty); + volatile entry_t *ht2_e = hti_lookup(ht2, key_hash, ns_val(key_string), ns_len(key_string), &is_empty); TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_e, ht2_e); // it is possible that there is not any room in the new table either @@ -271,7 +262,7 @@ static int hti_copy_entry (hash_table_i_t *ht1, volatile entry_t *ht1_e, uint32_ // Update the count if we were the one that completed the copy. if (old_ht2_e_value == DOES_NOT_EXIST) { - TRACE("h0", "hti_copy_entry: key \"%s\" value %p copied to new entry", key_string->val, value); + TRACE("h0", "hti_copy_entry: key \"%s\" value %p copied to new entry", ns_val(key_string), value); SYNC_ADD(&ht1->count, -1); SYNC_ADD(&ht2->count, 1); return TRUE; @@ -292,11 +283,11 @@ static int hti_copy_entry (hash_table_i_t *ht1, volatile entry_t *ht1_e, uint32_ // // NOTE: the returned value matches iff the set succeeds // -// Certain values of have special meaning. If is HT_EXPECT_EXISTS then any +// Certain values of have special meaning. If is EXPECT_EXISTS then any // real value matches (i.e. not a TOMBSTONE or DOES_NOT_EXIST) as long as is in the table. If -// is HT_EXPECT_WHATEVER then skip the test entirely. +// is EXPECT_WHATEVER then skip the test entirely. // -static uint64_t hti_compare_and_set (hash_table_i_t *hti, uint32_t key_hash, const char *key_val, +static uint64_t hti_compare_and_set (hashtable_i_t *hti, uint32_t key_hash, const char *key_val, uint32_t key_len, uint64_t expected, uint64_t new) { TRACE("h1", "hti_compare_and_set: hti %p key %p", hti, key_val); TRACE("h1", "hti_compare_and_set: value %p expect %p", new, expected); @@ -318,7 +309,7 @@ static uint64_t hti_compare_and_set (hash_table_i_t *hti, uint32_t key_hash, con // Install in the table if it doesn't exist. if (is_empty) { TRACE("h0", "hti_compare_and_set: entry %p is empty", e, 0); - if (expected != HT_EXPECT_WHATEVER && expected != HT_EXPECT_NOT_EXISTS) + if (expected != EXPECT_WHATEVER && expected != EXPECT_DOES_NOT_EXIST) return DOES_NOT_EXIST; // No need to do anything, is already deleted. @@ -326,9 +317,7 @@ static uint64_t hti_compare_and_set (hash_table_i_t *hti, uint32_t key_hash, con return DOES_NOT_EXIST; // Allocate . - string_t *key = nbd_malloc(sizeof(uint32_t) + key_len); - key->len = key_len; - memcpy(key->val, key_val, key_len); + nstring_t *key = ns_alloc(key_val, key_len); // Combine pointer with bits from its hash, CAS it into the table. uint64_t temp = ((uint64_t)(key_hash >> 16) << 48) | (uint64_t)key; @@ -344,13 +333,13 @@ static uint64_t hti_compare_and_set (hash_table_i_t *hti, uint32_t key_hash, con TRACE("h2", "hti_compare_and_set: installed key %p in entry %p", key, e); } - TRACE("h0", "hti_compare_and_set: entry for key \"%s\" is %p", GET_PTR(e->key)->val, e); + TRACE("h0", "hti_compare_and_set: entry for key \"%s\" is %p", ns_val(GET_PTR(e->key)), e); // If the entry is in the middle of a copy, the copy must be completed first. uint64_t e_value = e->value; if (EXPECT_FALSE(IS_TAGGED(e_value))) { if (e_value != COPIED_VALUE) { - int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hash_table_i_t *)hti)->next); + int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hashtable_i_t *)hti)->next); if (did_copy) { SYNC_ADD(&hti->num_entries_copied, 1); } @@ -363,8 +352,8 @@ static uint64_t hti_compare_and_set (hash_table_i_t *hti, uint32_t key_hash, con // Fail if the old value is not consistent with the caller's expectation. int old_existed = (e_value != TOMBSTONE && e_value != DOES_NOT_EXIST); - if (EXPECT_FALSE(expected != HT_EXPECT_WHATEVER && expected != e_value)) { - if (EXPECT_FALSE(expected != (old_existed ? HT_EXPECT_EXISTS : HT_EXPECT_NOT_EXISTS))) { + if (EXPECT_FALSE(expected != EXPECT_WHATEVER && expected != e_value)) { + if (EXPECT_FALSE(expected != (old_existed ? EXPECT_EXISTS : EXPECT_DOES_NOT_EXIST))) { TRACE("h1", "hti_compare_and_set: value %p expected by caller not found; found value %p", expected, e_value); return e_value; @@ -397,7 +386,7 @@ static uint64_t hti_compare_and_set (hash_table_i_t *hti, uint32_t key_hash, con } // -static uint64_t hti_get (hash_table_i_t *hti, uint32_t key_hash, const char *key_val, uint32_t key_len) { +static uint64_t hti_get (hashtable_i_t *hti, uint32_t key_hash, const char *key_val, uint32_t key_len) { assert(key_val); int is_empty; @@ -407,7 +396,7 @@ static uint64_t hti_get (hash_table_i_t *hti, uint32_t key_hash, const char *key // searching the table. In that case, if a copy is in progress the key // might exist in the copy. if (EXPECT_FALSE(e == NULL)) { - if (((volatile hash_table_i_t *)hti)->next != NULL) + if (((volatile hashtable_i_t *)hti)->next != NULL) return hti_get(hti->next, key_hash, key_val, key_len); // recursive tail-call return DOES_NOT_EXIST; } @@ -419,24 +408,24 @@ static uint64_t hti_get (hash_table_i_t *hti, uint32_t key_hash, const char *key uint64_t e_value = e->value; if (EXPECT_FALSE(IS_TAGGED(e_value))) { if (EXPECT_FALSE(e_value != COPIED_VALUE)) { - int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hash_table_i_t *)hti)->next); + int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hashtable_i_t *)hti)->next); if (did_copy) { SYNC_ADD(&hti->num_entries_copied, 1); } } - return hti_get(((volatile hash_table_i_t *)hti)->next, key_hash, key_val, key_len); // tail-call + return hti_get(((volatile hashtable_i_t *)hti)->next, key_hash, key_val, key_len); // tail-call } return (e_value == TOMBSTONE) ? DOES_NOT_EXIST : e_value; } // -uint64_t ht_get (hash_table_t *ht, const char *key_val, uint32_t key_len) { +uint64_t ht_get (hashtable_t *ht, const char *key_val, uint32_t key_len) { return hti_get(*ht, murmur32(key_val, key_len), key_val, key_len); } // -uint64_t ht_compare_and_set (hash_table_t *ht, const char *key_val, uint32_t key_len, +uint64_t ht_compare_and_set (hashtable_t *ht, const char *key_val, uint32_t key_len, uint64_t expected_val, uint64_t new_val) { TRACE("h2", "ht_compare_and_set: key %p len %u", key_val, key_len); @@ -444,7 +433,7 @@ uint64_t ht_compare_and_set (hash_table_t *ht, const char *key_val, uint32_t key assert(key_val); assert(!IS_TAGGED(new_val) && new_val != DOES_NOT_EXIST); - hash_table_i_t *hti = *ht; + hashtable_i_t *hti = *ht; // Help with an ongoing copy. if (EXPECT_FALSE(hti->next != NULL)) { @@ -505,12 +494,12 @@ uint64_t ht_compare_and_set (hash_table_t *ht, const char *key_val, uint32_t key // Remove the value in associated with . Returns the value removed, or // DOES_NOT_EXIST if there was no value for that key. -uint64_t ht_remove (hash_table_t *ht, const char *key_val, uint32_t key_len) { - hash_table_i_t *hti = *ht; +uint64_t ht_remove (hashtable_t *ht, const char *key_val, uint32_t key_len) { + hashtable_i_t *hti = *ht; uint64_t val; uint32_t key_hash = murmur32(key_val, key_len); do { - val = hti_compare_and_set(hti, key_hash, key_val, key_len, HT_EXPECT_WHATEVER, TOMBSTONE); + val = hti_compare_and_set(hti, key_hash, key_val, key_len, EXPECT_WHATEVER, TOMBSTONE); if (val != COPIED_VALUE) return val == TOMBSTONE ? DOES_NOT_EXIST : val; assert(hti->next); @@ -520,8 +509,8 @@ uint64_t ht_remove (hash_table_t *ht, const char *key_val, uint32_t key_len) { } // Returns the number of key-values pairs in -uint64_t ht_count (hash_table_t *ht) { - hash_table_i_t *hti = *ht; +uint64_t ht_count (hashtable_t *ht) { + hashtable_i_t *hti = *ht; uint64_t count = 0; while (hti) { count += hti->count; @@ -531,15 +520,15 @@ uint64_t ht_count (hash_table_t *ht) { } // Allocate and initialize a new hash table. -hash_table_t *ht_alloc (void) { - hash_table_t *ht = nbd_malloc(sizeof(hash_table_t)); - *ht = (hash_table_i_t *)hti_alloc(ht, MIN_SCALE); +hashtable_t *ht_alloc (void) { + hashtable_t *ht = nbd_malloc(sizeof(hashtable_t)); + *ht = (hashtable_i_t *)hti_alloc(ht, MIN_SCALE); return ht; } // Free and its internal structures. -void ht_free (hash_table_t *ht) { - hash_table_i_t *hti = *ht; +void ht_free (hashtable_t *ht) { + hashtable_i_t *hti = *ht; do { for (uint32_t i = 0; i < (1 << hti->scale); ++i) { assert(hti->table[i].value == COPIED_VALUE || !IS_TAGGED(hti->table[i].value)); @@ -547,7 +536,7 @@ void ht_free (hash_table_t *ht) { nbd_free(GET_PTR(hti->table[i].key)); } } - hash_table_i_t *next = hti->next; + hashtable_i_t *next = hti->next; nbd_free(hti); hti = next; } while (hti);