X-Git-Url: https://pd.if.org/git/?p=nbds;a=blobdiff_plain;f=map%2Fhashtable.c;h=bff782f84afdf46867dcbc27275402f1e3ac18dd;hp=69583acf0c11e15b8534262d940d878276d3624b;hb=5aa9223647fbb52fa8941d92c8896ebaf148b41c;hpb=025017478bb385da88a6b185849c8bcffeb2e2aa diff --git a/map/hashtable.c b/map/hashtable.c index 69583ac..bff782f 100644 --- a/map/hashtable.c +++ b/map/hashtable.c @@ -8,48 +8,64 @@ * * Note: This is code uses synchronous atomic operations because that is all that x86 provides. * Every atomic operation is also an implicit full memory barrier. The upshot is that it simplifies - * the code a bit, but it won't be as fast as it could be on platforms like SPARC that provide - * weaker operations which would still do the job. + * the code a bit, but it won't be as fast as it could be on platforms that provide weaker + * operations like and unfenced CAS which would still do the job. + * + * 11FebO9 - Bug fix in ht_iter_next() from Rui Ueyama */ +#include #include "common.h" #include "murmur.h" #include "mem.h" -#include "mlocal.h" -#include "nstring.h" - -#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; // ptr to nstring_t - uint64_t value; +#include "rcu.h" +#include "hashtable.h" + +#ifndef NBD32 +#define GET_PTR(x) ((void *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t +#else +#define GET_PTR(x) ((void *)(x)) +#endif + +typedef struct entry { + map_key_t key; + map_val_t val; } entry_t; typedef struct hti { volatile entry_t *table; hashtable_t *ht; // parent ht; struct hti *next; - struct hti *next_free; - unsigned int scale; +#ifdef USE_SYSTEM_MALLOC + void *unaligned_table_ptr; // system malloc doesn't guarentee cache-line alignment +#endif + unsigned scale; int max_probe; + int ref_count; int count; // TODO: make these counters distributed int num_entries_copied; - int scan; -} hashtable_i_t; + int copy_scan; +} hti_t; + +struct ht_iter { + hti_t * hti; + int64_t idx; +}; struct ht { - hashtable_i_t *hti; + hti_t *hti; + const datatype_t *key_type; }; -static const uint64_t COPIED_VALUE = -1; -static const uint64_t TOMBSTONE = STRIP_TAG(-1); +static const map_val_t COPIED_VALUE = TAG_VALUE(DOES_NOT_EXIST, TAG1); +static const map_val_t TOMBSTONE = STRIP_TAG(-1, TAG1); static const unsigned ENTRIES_PER_BUCKET = CACHE_LINE_SIZE/sizeof(entry_t); 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 (hashtable_i_t *ht1, volatile entry_t *e, uint32_t e_key_hash, hashtable_i_t *ht2); +static int hti_copy_entry (hti_t *ht1, volatile entry_t *ent, uint32_t ent_key_hash, hti_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) { @@ -58,27 +74,16 @@ static inline int get_next_ndx(int old_ndx, uint32_t key_hash, int ht_scale) { return (old_ndx + incr) & MASK(ht_scale); } -// Compare two keys. -// -// A key is made up of two parts. The 48 low-order bits are a pointer to a null terminated string. -// The high-order 16 bits are taken from the hash of that string. The bits from the hash are used -// as a quick check to rule out non-equal keys without doing a complete string compare. -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; - return ns_cmp_raw(GET_PTR(a), b_value, b_len) == 0; -} - // Lookup in . // // Return the entry that is in, or if isn't in return the entry that it would be // in if it were inserted into . If there is no room for in then return NULL, to // indicate that the caller should look in next>. // -// 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 (hashtable_i_t *hti, uint32_t key_hash, const char *key_data, uint32_t key_len, int *is_empty) { - TRACE("h2", "hti_lookup(key %p in hti %p)", key_data, hti); +// Record if the entry being returned is empty. Otherwise the caller will have to waste time +// re-comparing the keys to confirm that it did not lose a race to fill an empty entry. +static volatile entry_t *hti_lookup (hti_t *hti, map_key_t key, uint32_t key_hash, int *is_empty) { + TRACE("h2", "hti_lookup(key %p in hti %p)", key, hti); *is_empty = 0; // Probe one cache line at a time @@ -90,19 +95,37 @@ static volatile entry_t *hti_lookup (hashtable_i_t *hti, uint32_t key_hash, cons // Start searching at the indexed entry. Then loop around to the begining of the cache line. for (int j = 0; j < ENTRIES_PER_BUCKET; ++j) { - volatile entry_t *e = bucket + ((ndx + j) & (ENTRIES_PER_BUCKET-1)); - - 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)->data); - *is_empty = 1; // indicate an empty so the caller avoids an expensive ht_key_equals - return e; + volatile entry_t *ent = bucket + ((ndx + j) & (ENTRIES_PER_BUCKET-1)); + + map_key_t ent_key = ent->key; + if (ent_key == DOES_NOT_EXIST) { + TRACE("h1", "hti_lookup: entry %p for key %p is empty", ent, + (hti->ht->key_type == NULL) ? (void *)key : GET_PTR(key)); + *is_empty = 1; // indicate an empty so the caller avoids an expensive key compare + return ent; } - if (ht_key_equals(e_key, key_hash, key_data, key_len)) { - TRACE("h1", "hti_lookup: entry %p key \"%s\"", e, GET_PTR(e_key)->data); - TRACE("h2", "hti_lookup: entry key len %llu, value %p", GET_PTR(e_key)->len, e->value); - return e; + // Compare with the key in the entry. + if (EXPECT_TRUE(hti->ht->key_type == NULL)) { + // fast path for integer keys + if (ent_key == key) { + TRACE("h1", "hti_lookup: found entry %p with key %p", ent, ent_key); + return ent; + } + } else { +#ifndef NBD32 + // The key in is made up of two parts. The 48 low-order bits are a pointer. The + // high-order 16 bits are taken from the hash. The bits from the hash are used as a + // quick check to rule out non-equal keys without doing a complete compare. + if ((key_hash >> 16) == (ent_key >> 48)) { +#endif + if (hti->ht->key_type->cmp(GET_PTR(ent_key), (void *)key) == 0) { + TRACE("h1", "hti_lookup: found entry %p with key %p", ent, GET_PTR(ent_key)); + return ent; +#ifndef NBD32 + } +#endif + } } } @@ -114,27 +137,29 @@ static volatile entry_t *hti_lookup (hashtable_i_t *hti, uint32_t key_hash, cons return NULL; } -// 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(hashtable_i_t) - + sizeof(entry_t) * (1 << scale) - + (CACHE_LINE_SIZE - 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(hashtable_i_t) + (CACHE_LINE_SIZE-1)) - & ~(CACHE_LINE_SIZE-1)); - +// Allocate and initialize a hti_t with 2^ entries. +static hti_t *hti_alloc (hashtable_t *parent, int scale) { + hti_t *hti = (hti_t *)nbd_malloc(sizeof(hti_t)); + memset(hti, 0, sizeof(hti_t)); hti->scale = scale; + size_t sz = sizeof(entry_t) * (1 << scale); +#ifdef USE_SYSTEM_MALLOC + hti->unaligned_table_ptr = nbd_malloc(sz + CACHE_LINE_SIZE - 1); + hti->table = (void *)(((size_t)hti->unaligned_table_ptr + CACHE_LINE_SIZE - 1) & ~(CACHE_LINE_SIZE - 1)); +#else + hti->table = nbd_malloc(sz); +#endif + memset((void *)hti->table, 0, sz); + // When searching for a key probe a maximum of 1/4 of the buckets up to 1000 buckets. - hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 2; + hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4; if (hti->max_probe > MAX_BUCKETS_TO_PROBE) { hti->max_probe = MAX_BUCKETS_TO_PROBE; } hti->ht = parent; + hti->ref_count = 1; // one for the parent assert(hti->scale >= MIN_SCALE && hti->scale < 63); // size must be a power of 2 assert(sizeof(entry_t) * ENTRIES_PER_BUCKET % CACHE_LINE_SIZE == 0); // divisible into cache @@ -145,140 +170,138 @@ static hashtable_i_t *hti_alloc (hashtable_t *parent, int scale) { // Called when runs out of room for new keys. // -// Initiates a copy by creating a larger hashtable_i_t and installing it in next>. -static void hti_start_copy (hashtable_i_t *hti) { +// Initiates a copy by creating a larger hti_t and installing it in next>. +static void hti_start_copy (hti_t *hti) { TRACE("h0", "hti_start_copy(hti %p scale %llu)", hti, hti->scale); // heuristics to determine the size of the new table - uint64_t count = ht_count(hti->ht); + size_t count = ht_count(hti->ht); unsigned int new_scale = hti->scale; new_scale += (count > (1 << (new_scale - 2))); // double size if more than 1/4 full 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. - hashtable_i_t *next = hti_alloc(hti->ht, new_scale); - hashtable_i_t *old_next = SYNC_CAS(&hti->next, NULL, next); + hti_t *next = hti_alloc(hti->ht, new_scale); + hti_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); - nbd_free(next); +#ifdef USE_SYSTEM_MALLOC + nbd_free(next->unaligned_table_ptr); +#else + nbd_free((void *)next->table); +#endif return; } TRACE("h0", "hti_start_copy: new hti %p scale %llu", next, next->scale); } -// Copy the key and value stored in (which must be an entry in ) to . +// Copy the key and value stored in (which must be an entry in ) to . // -// Return 1 unless is already copied (then return 0), so the caller can account for the total +// 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 (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); +static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_hash, hti_t *ht2) { + TRACE("h2", "hti_copy_entry: entry %p to table %p", ht1_ent, ht2); assert(ht1); assert(ht1->next); assert(ht2); - assert(ht1_e >= ht1->table && ht1_e < ht1->table + (1 << ht1->scale)); - assert(key_hash == 0 || (key_hash >> 16) == (ht1_e->key >> 48)); - - uint64_t ht1_e_value = ht1_e->value; - if (EXPECT_FALSE(ht1_e_value == COPIED_VALUE)) { - TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_e, ht2); + assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1 << ht1->scale)); +#ifndef NBD32 + assert(key_hash == 0 || ht1->ht->key_type == NULL || (key_hash >> 16) == (ht1_ent->key >> 48)); +#endif + + map_val_t ht1_ent_val = ht1_ent->val; + if (EXPECT_FALSE(ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1))) { + TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2); return FALSE; // already copied } // Kill empty entries. - if (EXPECT_FALSE(ht1_e_value == DOES_NOT_EXIST)) { - uint64_t ht1_e_value = SYNC_CAS(&ht1_e->value, DOES_NOT_EXIST, COPIED_VALUE); - if (ht1_e_value == DOES_NOT_EXIST) { - TRACE("h1", "hti_copy_entry: empty entry %p killed", ht1_e, 0); + if (EXPECT_FALSE(ht1_ent_val == DOES_NOT_EXIST)) { + map_val_t ht1_ent_val = SYNC_CAS(&ht1_ent->val, DOES_NOT_EXIST, COPIED_VALUE); + if (ht1_ent_val == DOES_NOT_EXIST) { + TRACE("h1", "hti_copy_entry: empty entry %p killed", ht1_ent, 0); return TRUE; } - if (ht1_e_value == COPIED_VALUE) { - TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p", ht1_e, 0); - return FALSE; // another thread beat us to it - } - TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is now" - "in use and should be copied", ht1_e, 0); + TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is not empty", ht1_ent, 0); } // Tag the value in the old entry to indicate a copy is in progress. - ht1_e_value = SYNC_FETCH_AND_OR(&ht1_e->value, TAG_VALUE(0)); - TRACE("h2", "hti_copy_entry: tagged the value %p in old entry %p", ht1_e_value, ht1_e); - if (ht1_e_value == COPIED_VALUE) { - TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_e, ht2); + ht1_ent_val = SYNC_FETCH_AND_OR(&ht1_ent->val, TAG_VALUE(0, TAG1)); + TRACE("h2", "hti_copy_entry: tagged the value %p in old entry %p", ht1_ent_val, ht1_ent); + if (ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1)) { + TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2); return FALSE; // was already copied by another thread. } - // The old table's deleted entries don't need to be copied to the new table, but their keys need - // 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)); - nbd_defer_free(GET_PTR(ht1_e->key)); + // The old table's dead entries don't need to be copied to the new table + if (ht1_ent_val == TOMBSTONE) return TRUE; - } // Install the key in the new table. - uint64_t key = ht1_e->key; - nstring_t *key_string = GET_PTR(key); - uint64_t value = STRIP_TAG(ht1_e_value); + map_key_t ht1_ent_key = ht1_ent->key; + map_key_t key = (ht1->ht->key_type == NULL) ? (map_key_t)ht1_ent_key : (map_key_t)GET_PTR(ht1_ent_key); - // 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. + // We use 0 to indicate that is uninitiallized. Occasionally the key's hash will really be 0 and we + // waste time recomputing it every time. It is rare enough that it won't hurt performance. if (key_hash == 0) { - key_hash = murmur32(key_string->data, key_string->len); +#ifdef NBD32 + key_hash = (ht1->ht->key_type == NULL) ? murmur32_4b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key); +#else + key_hash = (ht1->ht->key_type == NULL) ? murmur32_8b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key); +#endif } - int is_empty; - volatile entry_t *ht2_e = hti_lookup(ht2, key_hash, key_string->data, key_string->len, &is_empty); - TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_e, ht2_e); + int ht2_ent_is_empty; + volatile entry_t *ht2_ent = hti_lookup(ht2, key, key_hash, &ht2_ent_is_empty); + TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_ent, ht2_ent); - // it is possible that there is not any room in the new table either - if (EXPECT_FALSE(ht2_e == NULL)) { + // It is possible that there isn't any room in the new table either. + if (EXPECT_FALSE(ht2_ent == NULL)) { TRACE("h0", "hti_copy_entry: no room in table %p copy to next table %p", ht2, ht2->next); if (ht2->next == NULL) { hti_start_copy(ht2); // initiate nested copy, if not already started } - return hti_copy_entry(ht1, ht1_e, key_hash, ht2->next); // recursive tail-call + return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call } - // a tagged entry returned from hti_lookup() means it is either empty or has a new key - if (is_empty) { - uint64_t old_ht2_e_key = SYNC_CAS(&ht2_e->key, DOES_NOT_EXIST, key); - if (old_ht2_e_key != DOES_NOT_EXIST) { + if (ht2_ent_is_empty) { + map_key_t old_ht2_ent_key = SYNC_CAS(&ht2_ent->key, DOES_NOT_EXIST, ht1_ent_key); + if (old_ht2_ent_key != DOES_NOT_EXIST) { TRACE("h0", "hti_copy_entry: lost race to CAS key %p into new entry; found %p", - key, old_ht2_e_key); - return hti_copy_entry(ht1, ht1_e, key_hash, ht2); // recursive tail-call + ht1_ent_key, old_ht2_ent_key); + return hti_copy_entry(ht1, ht1_ent, key_hash, ht2); // recursive tail-call } } // Copy the value to the entry in the new table. - uint64_t old_ht2_e_value = SYNC_CAS(&ht2_e->value, DOES_NOT_EXIST, value); + ht1_ent_val = STRIP_TAG(ht1_ent_val, TAG1); + map_val_t old_ht2_ent_val = SYNC_CAS(&ht2_ent->val, DOES_NOT_EXIST, ht1_ent_val); // If there is a nested copy in progress, we might have installed the key into a dead entry. - if (old_ht2_e_value == COPIED_VALUE) { - TRACE("h0", "hti_copy_entry: nested copy in progress; copy %p to next table %p", ht2_e, ht2->next); - return hti_copy_entry(ht1, ht1_e, key_hash, ht2->next); // recursive tail-call + if (old_ht2_ent_val == COPIED_VALUE) { + TRACE("h0", "hti_copy_entry: nested copy in progress; copy %p to next table %p", ht2_ent, ht2->next); + return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call } // Mark the old entry as dead. - ht1_e->value = COPIED_VALUE; + ht1_ent->val = COPIED_VALUE; // 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->data, value); - SYNC_ADD(&ht1->count, -1); - SYNC_ADD(&ht2->count, 1); + if (old_ht2_ent_val == DOES_NOT_EXIST) { + TRACE("h0", "hti_copy_entry: key %p value %p copied to new entry", key, ht1_ent_val); + (void)SYNC_ADD(&ht1->count, -1); + (void)SYNC_ADD(&ht2->count, 1); return TRUE; } TRACE("h0", "hti_copy_entry: lost race to install value %p in new entry; found value %p", - value, old_ht2_e_value); + ht1_ent_val, old_ht2_ent_val); return FALSE; // another thread completed the copy } // Compare with the existing value associated with . If the values match then -// replace the existing value with . If is TOMBSTONE, delete the value associated with +// replace the existing value with . If is DOES_NOT_EXIST, delete the value associated with // the key by replacing it with a TOMBSTONE. // // Return the previous value associated with , or DOES_NOT_EXIST if is not in the table @@ -288,22 +311,21 @@ static int hti_copy_entry (hashtable_i_t *ht1, volatile entry_t *ht1_e, uint32_t // NOTE: the returned value matches iff the set succeeds // // Certain values of have special meaning. If is CAS_EXPECT_EXISTS then any -// real value matches (i.e. not a TOMBSTONE or DOES_NOT_EXIST) as long as is in the table. If +// real value matches (i.ent. not a TOMBSTONE or DOES_NOT_EXIST) as long as is in the table. If // is CAS_EXPECT_WHATEVER then skip the test entirely. // -static uint64_t hti_cas (hashtable_i_t *hti, uint32_t key_hash, const char *key_data, uint32_t key_len, - uint64_t expected, uint64_t new) { - TRACE("h1", "hti_cas: hti %p key %p", hti, key_data); +static map_val_t hti_cas (hti_t *hti, map_key_t key, uint32_t key_hash, map_val_t expected, map_val_t new) { + TRACE("h1", "hti_cas: hti %p key %p", hti, key); TRACE("h1", "hti_cas: value %p expect %p", new, expected); assert(hti); - assert(new != DOES_NOT_EXIST && !IS_TAGGED(new)); - assert(key_data); + assert(!IS_TAGGED(new, TAG1)); + assert(key); int is_empty; - volatile entry_t *e = hti_lookup(hti, key_hash, key_data, key_len, &is_empty); + volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty); // There is no room for , grow the table and try again. - if (e == NULL) { + if (ent == NULL) { if (hti->next == NULL) { hti_start_copy(hti); } @@ -312,40 +334,51 @@ static uint64_t hti_cas (hashtable_i_t *hti, uint32_t key_hash, const char *key_ // Install in the table if it doesn't exist. if (is_empty) { - TRACE("h0", "hti_cas: entry %p is empty", e, 0); + TRACE("h0", "hti_cas: entry %p is empty", ent, 0); if (expected != CAS_EXPECT_WHATEVER && expected != CAS_EXPECT_DOES_NOT_EXIST) return DOES_NOT_EXIST; // No need to do anything, is already deleted. - if (new == TOMBSTONE) + if (new == DOES_NOT_EXIST) return DOES_NOT_EXIST; - // Allocate . - nstring_t *key = ns_alloc(key_data, key_len); + // Allocate . + map_key_t new_key = (hti->ht->key_type == NULL) + ? (map_key_t)key + : (map_key_t)hti->ht->key_type->clone((void *)key); +#ifndef NBD32 + if (EXPECT_FALSE(hti->ht->key_type != NULL)) { + // Combine pointer with bits from its hash + new_key = ((uint64_t)(key_hash >> 16) << 48) | new_key; + } +#endif - // Combine pointer with bits from its hash, CAS it into the table. - uint64_t temp = ((uint64_t)(key_hash >> 16) << 48) | (uint64_t)key; - uint64_t e_key = SYNC_CAS(&e->key, DOES_NOT_EXIST, temp); + // CAS the key into the table. + map_key_t old_ent_key = SYNC_CAS(&ent->key, DOES_NOT_EXIST, new_key); // Retry if another thread stole the entry out from under us. - if (e_key != DOES_NOT_EXIST) { - TRACE("h0", "hti_cas: lost race to install key %p in entry %p", key, e); - TRACE("h0", "hti_cas: found %p instead of NULL", GET_PTR(e_key), 0); - nbd_free(key); - return hti_cas(hti, key_hash, key_data, key_len, expected, new); // tail-call + if (old_ent_key != DOES_NOT_EXIST) { + TRACE("h0", "hti_cas: lost race to install key %p in entry %p", new_key, ent); + TRACE("h0", "hti_cas: found %p instead of NULL", + (hti->ht->key_type == NULL) ? (void *)old_ent_key : GET_PTR(old_ent_key), 0); + if (hti->ht->key_type != NULL) { + nbd_free(GET_PTR(new_key)); + } + return hti_cas(hti, key, key_hash, expected, new); // tail-call } - TRACE("h2", "hti_cas: installed key %p in entry %p", key, e); + TRACE("h2", "hti_cas: installed key %p in entry %p", new_key, ent); } - TRACE("h0", "hti_cas: entry for key \"%s\" is %p", GET_PTR(e->key)->data, e); + TRACE("h0", "hti_cas: entry for key %p is %p", + (hti->ht->key_type == NULL) ? (void *)ent->key : GET_PTR(ent->key), ent); // 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 hashtable_i_t *)hti)->next); + map_val_t ent_val = ent->val; + if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) { + if (ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1)) { + int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next); if (did_copy) { - SYNC_ADD(&hti->num_entries_copied, 1); + (void)SYNC_ADD(&hti->num_entries_copied, 1); } TRACE("h0", "hti_cas: value in the middle of a copy, copy completed by %s", (did_copy ? "self" : "other"), 0); @@ -355,53 +388,51 @@ static uint64_t hti_cas (hashtable_i_t *hti, uint32_t key_hash, const char *key_ } // 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 != CAS_EXPECT_WHATEVER && expected != e_value)) { + int old_existed = (ent_val != TOMBSTONE && ent_val != DOES_NOT_EXIST); + if (EXPECT_FALSE(expected != CAS_EXPECT_WHATEVER && expected != ent_val)) { if (EXPECT_FALSE(expected != (old_existed ? CAS_EXPECT_EXISTS : CAS_EXPECT_DOES_NOT_EXIST))) { TRACE("h1", "hti_cas: value %p expected by caller not found; found value %p", - expected, e_value); - return e_value; + expected, ent_val); + return ent_val; } } // No need to update if value is unchanged. - if ((new == TOMBSTONE && !old_existed) || e_value == new) { + if ((new == DOES_NOT_EXIST && !old_existed) || ent_val == new) { TRACE("h1", "hti_cas: old value and new value were the same", 0, 0); - return e_value; + return ent_val; } // CAS the value into the entry. Retry if it fails. - uint64_t v = SYNC_CAS(&e->value, e_value, new); - if (EXPECT_FALSE(v != e_value)) { - TRACE("h0", "hti_cas: value CAS failed; expected %p found %p", e_value, v); - return hti_cas(hti, key_hash, key_data, key_len, expected, new); // recursive tail-call + map_val_t v = SYNC_CAS(&ent->val, ent_val, new == DOES_NOT_EXIST ? TOMBSTONE : new); + if (EXPECT_FALSE(v != ent_val)) { + TRACE("h0", "hti_cas: value CAS failed; expected %p found %p", ent_val, v); + return hti_cas(hti, key, key_hash, expected, new); // recursive tail-call } // The set succeeded. Adjust the value count. - if (old_existed && new == TOMBSTONE) { - SYNC_ADD(&hti->count, -1); - } else if (!old_existed && new != TOMBSTONE) { - SYNC_ADD(&hti->count, 1); + if (old_existed && new == DOES_NOT_EXIST) { + (void)SYNC_ADD(&hti->count, -1); + } else if (!old_existed && new != DOES_NOT_EXIST) { + (void)SYNC_ADD(&hti->count, 1); } // Return the previous value. - TRACE("h0", "hti_cas: CAS succeeded; old value %p new value %p", e_value, new); - return e_value; + TRACE("h0", "hti_cas: CAS succeeded; old value %p new value %p", ent_val, new); + return ent_val; } // -static uint64_t hti_get (hashtable_i_t *hti, uint32_t key_hash, const char *key_data, uint32_t key_len) { - assert(key_data); - +static map_val_t hti_get (hti_t *hti, map_key_t key, uint32_t key_hash) { int is_empty; - volatile entry_t *e = hti_lookup(hti, key_hash, key_data, key_len, &is_empty); + volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty); // When hti_lookup() returns NULL it means we hit the reprobe limit while // 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 hashtable_i_t *)hti)->next != NULL) - return hti_get(hti->next, key_hash, key_data, key_len); // recursive tail-call + if (EXPECT_FALSE(ent == NULL)) { + if (VOLATILE_DEREF(hti).next != NULL) + return hti_get(hti->next, key, key_hash); // recursive tail-call return DOES_NOT_EXIST; } @@ -409,43 +440,40 @@ static uint64_t hti_get (hashtable_i_t *hti, uint32_t key_hash, const char *key_ return DOES_NOT_EXIST; // If the entry is being copied, finish the copy and retry on the next table. - 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 hashtable_i_t *)hti)->next); + map_val_t ent_val = ent->val; + if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) { + if (EXPECT_FALSE(ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1))) { + int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next); if (did_copy) { - SYNC_ADD(&hti->num_entries_copied, 1); + (void)SYNC_ADD(&hti->num_entries_copied, 1); } } - return hti_get(((volatile hashtable_i_t *)hti)->next, key_hash, key_data, key_len); // tail-call + return hti_get(VOLATILE_DEREF(hti).next, key, key_hash); // tail-call } - return (e_value == TOMBSTONE) ? DOES_NOT_EXIST : e_value; + return (ent_val == TOMBSTONE) ? DOES_NOT_EXIST : ent_val; } // -uint64_t ht_get (hashtable_t *ht, const char *key_data, uint32_t key_len) { - return hti_get(ht->hti, murmur32(key_data, key_len), key_data, key_len); +map_val_t ht_get (hashtable_t *ht, map_key_t key) { +#ifdef NBD32 + uint32_t hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key); +#else + uint32_t hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key); +#endif + return hti_get(ht->hti, key, hash); } -// -uint64_t ht_cas (hashtable_t *ht, const char *key_data, uint32_t key_len, uint64_t expected_val, uint64_t new_val) { - - TRACE("h2", "ht_cas: key %p len %u", key_data, key_len); - TRACE("h2", "ht_cas: expected val %p new val %p", expected_val, new_val); - assert(key_data); - assert(!IS_TAGGED(new_val) && new_val != DOES_NOT_EXIST); - - hashtable_i_t *hti = ht->hti; +// returns TRUE if copy is done +static int hti_help_copy (hti_t *hti) { + volatile entry_t *ent; + size_t limit; + size_t total_copied = hti->num_entries_copied; + size_t num_copied = 0; + size_t x = hti->copy_scan; - // Help with an ongoing copy. - if (EXPECT_FALSE(hti->next != NULL)) { - volatile entry_t *e; - uint64_t limit; - int num_copied = 0; - int x = hti->scan; - - TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<scale); + TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<scale); + if (total_copied != (1 << hti->scale)) { // Panic if we've been around the array twice and still haven't finished the copy. int panic = (x >= (1 << (hti->scale + 1))); if (!panic) { @@ -453,41 +481,91 @@ uint64_t ht_cas (hashtable_t *ht, const char *key_data, uint32_t key_len, uint64 // Reserve some entries for this thread to copy. There is a race condition here because the // fetch and add isn't atomic, but that is ok. - hti->scan = x + ENTRIES_PER_COPY_CHUNK; + hti->copy_scan = x + ENTRIES_PER_COPY_CHUNK; - // scan> might be larger than the size of the table, if some thread stalls while + // might be larger than the size of the table, if some thread stalls while // copying. In that case we just wrap around to the begining and make another pass through // the table. - e = hti->table + (x & MASK(hti->scale)); + ent = hti->table + (x & MASK(hti->scale)); } else { TRACE("h1", "ht_cas: help copy panic", 0, 0); // scan the whole table + ent = hti->table; limit = (1 << hti->scale); - e = hti->table; } // Copy the entries for (int i = 0; i < limit; ++i) { - num_copied += hti_copy_entry(hti, e++, 0, hti->next); - assert(e <= hti->table + (1 << hti->scale)); + num_copied += hti_copy_entry(hti, ent++, 0, hti->next); + assert(ent <= hti->table + (1 << hti->scale)); } if (num_copied != 0) { - SYNC_ADD(&hti->num_entries_copied, num_copied); + total_copied = SYNC_ADD(&hti->num_entries_copied, num_copied); + } + } + + return (total_copied == (1 << hti->scale)); +} + +static void hti_defer_free (hti_t *hti) { + assert(hti->ref_count == 0); + + for (uint32_t i = 0; i < (1 << hti->scale); ++i) { + map_key_t key = hti->table[i].key; + map_val_t val = hti->table[i].val; + if (val == COPIED_VALUE) + continue; + assert(!IS_TAGGED(val, TAG1) || val == TAG_VALUE(TOMBSTONE, TAG1)); // copy not in progress + if (hti->ht->key_type != NULL && key != DOES_NOT_EXIST) { + rcu_defer_free(GET_PTR(key)); } + } +#ifdef USE_SYSTEM_MALLOC + rcu_defer_free(hti->unaligned_table_ptr); +#else + rcu_defer_free((void *)hti->table); +#endif + rcu_defer_free(hti); +} + +static void hti_release (hti_t *hti) { + assert(hti->ref_count > 0); + int ref_count = SYNC_ADD(&hti->ref_count, -1); + if (ref_count == 0) { + hti_defer_free(hti); + } +} + +// +map_val_t ht_cas (hashtable_t *ht, map_key_t key, map_val_t expected_val, map_val_t new_val) { + + TRACE("h2", "ht_cas: key %p ht %p", key, ht); + TRACE("h2", "ht_cas: expected val %p new val %p", expected_val, new_val); + assert(key != DOES_NOT_EXIST); + assert(!IS_TAGGED(new_val, TAG1) && new_val != DOES_NOT_EXIST && new_val != TOMBSTONE); + + hti_t *hti = ht->hti; + + // Help with an ongoing copy. + if (EXPECT_FALSE(hti->next != NULL)) { + int done = hti_help_copy(hti); - // Dispose of fully copied tables. - if (hti->num_entries_copied == (1 << hti->scale) || panic) { + // Unlink fully copied tables. + if (done) { assert(hti->next); if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) { - nbd_defer_free(hti); + hti_release(hti); } } } - uint64_t old_val; - uint32_t key_hash = murmur32(key_data, key_len); - while ((old_val = hti_cas(hti, key_hash, key_data, key_len, expected_val, new_val)) - == COPIED_VALUE) { + map_val_t old_val; +#ifdef NBD32 + uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key); +#else + uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key); +#endif + while ((old_val = hti_cas(hti, key, key_hash, expected_val, new_val)) == COPIED_VALUE) { assert(hti->next); hti = hti->next; } @@ -495,14 +573,18 @@ uint64_t ht_cas (hashtable_t *ht, const char *key_data, uint32_t key_len, uint64 return old_val == TOMBSTONE ? DOES_NOT_EXIST : old_val; } -// 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 (hashtable_t *ht, const char *key_data, uint32_t key_len) { - hashtable_i_t *hti = ht->hti; - uint64_t val; - uint32_t key_hash = murmur32(key_data, key_len); +// Remove the value in associated with . Returns the value removed, or DOES_NOT_EXIST if there was +// no value for that key. +map_val_t ht_remove (hashtable_t *ht, map_key_t key) { + hti_t *hti = ht->hti; + map_val_t val; +#ifdef NBD32 + uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key); +#else + uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key); +#endif do { - val = hti_cas(hti, key_hash, key_data, key_len, CAS_EXPECT_WHATEVER, TOMBSTONE); + val = hti_cas(hti, key, key_hash, CAS_EXPECT_WHATEVER, DOES_NOT_EXIST); if (val != COPIED_VALUE) return val == TOMBSTONE ? DOES_NOT_EXIST : val; assert(hti->next); @@ -512,9 +594,9 @@ uint64_t ht_remove (hashtable_t *ht, const char *key_data, uint32_t key_len) { } // Returns the number of key-values pairs in -uint64_t ht_count (hashtable_t *ht) { - hashtable_i_t *hti = ht->hti; - uint64_t count = 0; +size_t ht_count (hashtable_t *ht) { + hti_t *hti = ht->hti; + size_t count = 0; while (hti) { count += hti->count; hti = hti->next; @@ -523,28 +605,101 @@ uint64_t ht_count (hashtable_t *ht) { } // Allocate and initialize a new hash table. -hashtable_t *ht_alloc (void) { +hashtable_t *ht_alloc (const datatype_t *key_type) { hashtable_t *ht = nbd_malloc(sizeof(hashtable_t)); - ht->hti = (hashtable_i_t *)hti_alloc(ht, MIN_SCALE); + ht->key_type = key_type; + ht->hti = (hti_t *)hti_alloc(ht, MIN_SCALE); return ht; } // Free and its internal structures. void ht_free (hashtable_t *ht) { - hashtable_i_t *hti = ht->hti; + hti_t *hti = ht->hti; do { - for (uint32_t i = 0; i < (1 << hti->scale); ++i) { - assert(hti->table[i].value == COPIED_VALUE || !IS_TAGGED(hti->table[i].value)); - if (hti->table[i].key != DOES_NOT_EXIST) { - nbd_free(GET_PTR(hti->table[i].key)); - } - } - hashtable_i_t *next = hti->next; - nbd_free(hti); + hti_t *next = hti->next; + assert(hti->ref_count == 1); + hti_release(hti); hti = next; } while (hti); nbd_free(ht); } void ht_print (hashtable_t *ht) { + hti_t *hti = ht->hti; + while (hti) { + printf("hti:%p scale:%u count:%d copied:%d\n", hti, hti->scale, hti->count, hti->num_entries_copied); + for (int i = 0; i < (1 << hti->scale); ++i) { + volatile entry_t *ent = hti->table + i; + printf("[0x%x] 0x%llx:0x%llx\n", i, (uint64_t)ent->key, (uint64_t)ent->val); + if (i > 30) { + printf("...\n"); + break; + } + } + hti = hti->next; + } +} + +ht_iter_t *ht_iter_begin (hashtable_t *ht, map_key_t key) { + hti_t *hti; + int ref_count; + do { + hti = ht->hti; + while (hti->next != NULL) { + do { } while (hti_help_copy(hti) != TRUE); + hti = hti->next; + } + do { + ref_count = hti->ref_count; + if(ref_count == 0) + break; + } while (ref_count != SYNC_CAS(&hti->ref_count, ref_count, ref_count + 1)); + } while (ref_count == 0); + + ht_iter_t *iter = nbd_malloc(sizeof(ht_iter_t)); + iter->hti = hti; + iter->idx = -1; + + return iter; +} + +map_val_t ht_iter_next (ht_iter_t *iter, map_key_t *key_ptr) { + volatile entry_t *ent; + map_key_t key; + map_val_t val; + size_t table_size = (1 << iter->hti->scale); + do { + iter->idx++; + if (iter->idx == table_size) { + return DOES_NOT_EXIST; + } + ent = &iter->hti->table[iter->idx]; + key = (iter->hti->ht->key_type == NULL) ? (map_key_t)ent->key : (map_key_t)GET_PTR(ent->key); + val = ent->val; + + } while (key == DOES_NOT_EXIST || val == DOES_NOT_EXIST || val == TOMBSTONE); + + if (val == COPIED_VALUE) { + const datatype_t *key_type = iter->hti->ht->key_type; +#ifdef NBD32 + uint32_t hash = (key_type == NULL) ? murmur32_4b((uint64_t)key) : key_type->hash((void *)key); +#else + uint32_t hash = (key_type == NULL) ? murmur32_8b((uint64_t)key) : key_type->hash((void *)key); +#endif + val = hti_get(iter->hti->next, (map_key_t)ent->key, hash); + + // Go to the next entry if key is already deleted. + if (val == DOES_NOT_EXIST) + return ht_iter_next(iter, key_ptr); // recursive tail-call + } + + if (key_ptr) { + *key_ptr = key; + } + return val; +} + +void ht_iter_free (ht_iter_t *iter) { + hti_release(iter->hti); + nbd_free(iter); }