2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
5 * C implementation of Cliff Click's lock-free hash table from
6 * http://www.azulsystems.com/events/javaone_2008/2008_CodingNonBlock.pdf
7 * http://sourceforge.net/projects/high-scale-lib
9 * Note: This is code uses synchronous atomic operations because that is all that x86 provides.
10 * Every atomic operation is also an implicit full memory barrier. The upshot is that it simplifies
11 * the code a bit, but it won't be as fast as it could be on platforms that provide weaker
12 * operations like and unfenced CAS which would still do the job.
20 #include "hashtable.h"
23 #define GET_PTR(x) ((void *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t
25 #define GET_PTR(x) ((void *)(x))
28 typedef struct entry {
34 volatile entry_t *table;
35 hashtable_t *ht; // parent ht;
37 #ifdef USE_SYSTEM_MALLOC
38 void *unaligned_table_ptr; // system malloc doesn't guarentee cache-line alignment
43 int count; // TODO: make these counters distributed
44 int num_entries_copied;
55 const datatype_t *key_type;
58 static const map_val_t COPIED_VALUE = TAG_VALUE(DOES_NOT_EXIST, TAG1);
59 static const map_val_t TOMBSTONE = STRIP_TAG(-1, TAG1);
61 static const unsigned ENTRIES_PER_BUCKET = CACHE_LINE_SIZE/sizeof(entry_t);
62 static const unsigned ENTRIES_PER_COPY_CHUNK = CACHE_LINE_SIZE/sizeof(entry_t)*2;
63 static const unsigned MIN_SCALE = 4; // min 16 entries (4 buckets)
64 static const unsigned MAX_BUCKETS_TO_PROBE = 250;
66 static int hti_copy_entry (hti_t *ht1, volatile entry_t *ent, uint32_t ent_key_hash, hti_t *ht2);
68 // Choose the next bucket to probe using the high-order bits of <key_hash>.
69 static inline int get_next_ndx(int old_ndx, uint32_t key_hash, int ht_scale) {
70 int incr = (key_hash >> (32 - ht_scale));
71 incr += !incr; // If the increment is 0, make it 1.
72 return (old_ndx + incr) & MASK(ht_scale);
75 // Lookup <key> in <hti>.
77 // Return the entry that <key> is in, or if <key> isn't in <hti> return the entry that it would be
78 // in if it were inserted into <hti>. If there is no room for <key> in <hti> then return NULL, to
79 // indicate that the caller should look in <hti->next>.
81 // Record if the entry being returned is empty. Otherwise the caller will have to waste time
82 // re-comparing the keys to confirm that it did not lose a race to fill an empty entry.
83 static volatile entry_t *hti_lookup (hti_t *hti, map_key_t key, uint32_t key_hash, int *is_empty) {
84 TRACE("h2", "hti_lookup(key %p in hti %p)", key, hti);
87 // Probe one cache line at a time
88 int ndx = key_hash & MASK(hti->scale); // the first entry to search
89 for (int i = 0; i < hti->max_probe; ++i) {
91 // The start of the bucket is the first entry in the cache line.
92 volatile entry_t *bucket = hti->table + (ndx & ~(ENTRIES_PER_BUCKET-1));
94 // Start searching at the indexed entry. Then loop around to the begining of the cache line.
95 for (int j = 0; j < ENTRIES_PER_BUCKET; ++j) {
96 volatile entry_t *ent = bucket + ((ndx + j) & (ENTRIES_PER_BUCKET-1));
98 map_key_t ent_key = ent->key;
99 if (ent_key == DOES_NOT_EXIST) {
100 TRACE("h1", "hti_lookup: entry %p for key %p is empty", ent,
101 (hti->ht->key_type == NULL) ? (void *)key : GET_PTR(key));
102 *is_empty = 1; // indicate an empty so the caller avoids an expensive key compare
106 // Compare <key> with the key in the entry.
107 if (EXPECT_TRUE(hti->ht->key_type == NULL)) {
108 // fast path for integer keys
109 if (ent_key == key) {
110 TRACE("h1", "hti_lookup: found entry %p with key %p", ent, ent_key);
115 // The key in <ent> is made up of two parts. The 48 low-order bits are a pointer. The
116 // high-order 16 bits are taken from the hash. The bits from the hash are used as a
117 // quick check to rule out non-equal keys without doing a complete compare.
118 if ((key_hash >> 16) == (ent_key >> 48)) {
120 if (hti->ht->key_type->cmp(GET_PTR(ent_key), (void *)key) == 0) {
121 TRACE("h1", "hti_lookup: found entry %p with key %p", ent, GET_PTR(ent_key));
130 ndx = get_next_ndx(ndx, key_hash, hti->scale);
133 // maximum number of probes exceeded
134 TRACE("h1", "hti_lookup: maximum number of probes exceeded returning 0x0", 0, 0);
138 // Allocate and initialize a hti_t with 2^<scale> entries.
139 static hti_t *hti_alloc (hashtable_t *parent, int scale) {
140 hti_t *hti = (hti_t *)nbd_malloc(sizeof(hti_t));
141 memset(hti, 0, sizeof(hti_t));
144 size_t sz = sizeof(entry_t) * (1 << scale);
145 #ifdef USE_SYSTEM_MALLOC
146 hti->unaligned_table_ptr = nbd_malloc(sz + CACHE_LINE_SIZE - 1);
147 hti->table = (void *)(((size_t)hti->unaligned_table_ptr + CACHE_LINE_SIZE - 1) & ~(CACHE_LINE_SIZE - 1));
149 hti->table = nbd_malloc(sz);
151 memset((void *)hti->table, 0, sz);
153 // When searching for a key probe a maximum of 1/4 of the buckets up to 1000 buckets.
154 hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4;
155 if (hti->max_probe > MAX_BUCKETS_TO_PROBE) {
156 hti->max_probe = MAX_BUCKETS_TO_PROBE;
160 hti->ref_count = 1; // one for the parent
162 assert(hti->scale >= MIN_SCALE && hti->scale < 63); // size must be a power of 2
163 assert(sizeof(entry_t) * ENTRIES_PER_BUCKET % CACHE_LINE_SIZE == 0); // divisible into cache
164 assert((size_t)hti->table % CACHE_LINE_SIZE == 0); // cache aligned
169 // Called when <hti> runs out of room for new keys.
171 // Initiates a copy by creating a larger hti_t and installing it in <hti->next>.
172 static void hti_start_copy (hti_t *hti) {
173 TRACE("h0", "hti_start_copy(hti %p scale %llu)", hti, hti->scale);
175 // heuristics to determine the size of the new table
176 size_t count = ht_count(hti->ht);
177 unsigned int new_scale = hti->scale;
178 new_scale += (count > (1 << (new_scale - 2))); // double size if more than 1/4 full
179 new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full
181 // Allocate the new table and attempt to install it.
182 hti_t *next = hti_alloc(hti->ht, new_scale);
183 hti_t *old_next = SYNC_CAS(&hti->next, NULL, next);
184 if (old_next != NULL) {
185 // Another thread beat us to it.
186 TRACE("h0", "hti_start_copy: lost race to install new hti; found %p", old_next, 0);
187 #ifdef USE_SYSTEM_MALLOC
188 nbd_free(next->unaligned_table_ptr);
190 nbd_free((void *)next->table);
194 TRACE("h0", "hti_start_copy: new hti %p scale %llu", next, next->scale);
197 // Copy the key and value stored in <ht1_ent> (which must be an entry in <ht1>) to <ht2>.
199 // Return 1 unless <ht1_ent> is already copied (then return 0), so the caller can account for the total
200 // number of entries left to copy.
201 static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_hash, hti_t *ht2) {
202 TRACE("h2", "hti_copy_entry: entry %p to table %p", ht1_ent, ht2);
206 assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1 << ht1->scale));
208 assert(key_hash == 0 || ht1->ht->key_type == NULL || (key_hash >> 16) == (ht1_ent->key >> 48));
211 map_val_t ht1_ent_val = ht1_ent->val;
212 if (EXPECT_FALSE(ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1))) {
213 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
214 return FALSE; // already copied
217 // Kill empty entries.
218 if (EXPECT_FALSE(ht1_ent_val == DOES_NOT_EXIST)) {
219 map_val_t ht1_ent_val = SYNC_CAS(&ht1_ent->val, DOES_NOT_EXIST, COPIED_VALUE);
220 if (ht1_ent_val == DOES_NOT_EXIST) {
221 TRACE("h1", "hti_copy_entry: empty entry %p killed", ht1_ent, 0);
224 TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is not empty", ht1_ent, 0);
227 // Tag the value in the old entry to indicate a copy is in progress.
228 ht1_ent_val = SYNC_FETCH_AND_OR(&ht1_ent->val, TAG_VALUE(0, TAG1));
229 TRACE("h2", "hti_copy_entry: tagged the value %p in old entry %p", ht1_ent_val, ht1_ent);
230 if (ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1)) {
231 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
232 return FALSE; // <value> was already copied by another thread.
235 // The old table's dead entries don't need to be copied to the new table
236 if (ht1_ent_val == TOMBSTONE)
239 // Install the key in the new table.
240 map_key_t ht1_ent_key = ht1_ent->key;
241 map_key_t key = (ht1->ht->key_type == NULL) ? (map_key_t)ht1_ent_key : (map_key_t)GET_PTR(ht1_ent_key);
243 // We use 0 to indicate that <key_hash> is uninitiallized. Occasionally the key's hash will really be 0 and we
244 // waste time recomputing it every time. It is rare enough that it won't hurt performance.
247 key_hash = (ht1->ht->key_type == NULL) ? murmur32_4b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key);
249 key_hash = (ht1->ht->key_type == NULL) ? murmur32_8b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key);
253 int ht2_ent_is_empty;
254 volatile entry_t *ht2_ent = hti_lookup(ht2, key, key_hash, &ht2_ent_is_empty);
255 TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_ent, ht2_ent);
257 // It is possible that there isn't any room in the new table either.
258 if (EXPECT_FALSE(ht2_ent == NULL)) {
259 TRACE("h0", "hti_copy_entry: no room in table %p copy to next table %p", ht2, ht2->next);
260 if (ht2->next == NULL) {
261 hti_start_copy(ht2); // initiate nested copy, if not already started
263 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call
266 if (ht2_ent_is_empty) {
267 map_key_t old_ht2_ent_key = SYNC_CAS(&ht2_ent->key, DOES_NOT_EXIST, ht1_ent_key);
268 if (old_ht2_ent_key != DOES_NOT_EXIST) {
269 TRACE("h0", "hti_copy_entry: lost race to CAS key %p into new entry; found %p",
270 ht1_ent_key, old_ht2_ent_key);
271 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2); // recursive tail-call
275 // Copy the value to the entry in the new table.
276 ht1_ent_val = STRIP_TAG(ht1_ent_val, TAG1);
277 map_val_t old_ht2_ent_val = SYNC_CAS(&ht2_ent->val, DOES_NOT_EXIST, ht1_ent_val);
279 // If there is a nested copy in progress, we might have installed the key into a dead entry.
280 if (old_ht2_ent_val == COPIED_VALUE) {
281 TRACE("h0", "hti_copy_entry: nested copy in progress; copy %p to next table %p", ht2_ent, ht2->next);
282 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call
285 // Mark the old entry as dead.
286 ht1_ent->val = COPIED_VALUE;
288 // Update the count if we were the one that completed the copy.
289 if (old_ht2_ent_val == DOES_NOT_EXIST) {
290 TRACE("h0", "hti_copy_entry: key %p value %p copied to new entry", key, ht1_ent_val);
291 (void)SYNC_ADD(&ht1->count, -1);
292 (void)SYNC_ADD(&ht2->count, 1);
296 TRACE("h0", "hti_copy_entry: lost race to install value %p in new entry; found value %p",
297 ht1_ent_val, old_ht2_ent_val);
298 return FALSE; // another thread completed the copy
301 // Compare <expected> with the existing value associated with <key>. If the values match then
302 // replace the existing value with <new>. If <new> is DOES_NOT_EXIST, delete the value associated with
303 // the key by replacing it with a TOMBSTONE.
305 // Return the previous value associated with <key>, or DOES_NOT_EXIST if <key> is not in the table
306 // or associated with a TOMBSTONE. If a copy is in progress and <key> has been copied to the next
307 // table then return COPIED_VALUE.
309 // NOTE: the returned value matches <expected> iff the set succeeds
311 // Certain values of <expected> have special meaning. If <expected> is CAS_EXPECT_EXISTS then any
312 // real value matches (i.ent. not a TOMBSTONE or DOES_NOT_EXIST) as long as <key> is in the table. If
313 // <expected> is CAS_EXPECT_WHATEVER then skip the test entirely.
315 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) {
316 TRACE("h1", "hti_cas: hti %p key %p", hti, key);
317 TRACE("h1", "hti_cas: value %p expect %p", new, expected);
319 assert(!IS_TAGGED(new, TAG1));
323 volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty);
325 // There is no room for <key>, grow the table and try again.
327 if (hti->next == NULL) {
333 // Install <key> in the table if it doesn't exist.
335 TRACE("h0", "hti_cas: entry %p is empty", ent, 0);
336 if (expected != CAS_EXPECT_WHATEVER && expected != CAS_EXPECT_DOES_NOT_EXIST)
337 return DOES_NOT_EXIST;
339 // No need to do anything, <key> is already deleted.
340 if (new == DOES_NOT_EXIST)
341 return DOES_NOT_EXIST;
343 // Allocate <new_key>.
344 map_key_t new_key = (hti->ht->key_type == NULL)
346 : (map_key_t)hti->ht->key_type->clone((void *)key);
348 if (EXPECT_FALSE(hti->ht->key_type != NULL)) {
349 // Combine <new_key> pointer with bits from its hash
350 new_key = ((uint64_t)(key_hash >> 16) << 48) | new_key;
354 // CAS the key into the table.
355 map_key_t old_ent_key = SYNC_CAS(&ent->key, DOES_NOT_EXIST, new_key);
357 // Retry if another thread stole the entry out from under us.
358 if (old_ent_key != DOES_NOT_EXIST) {
359 TRACE("h0", "hti_cas: lost race to install key %p in entry %p", new_key, ent);
360 TRACE("h0", "hti_cas: found %p instead of NULL",
361 (hti->ht->key_type == NULL) ? (void *)old_ent_key : GET_PTR(old_ent_key), 0);
362 if (hti->ht->key_type != NULL) {
363 nbd_free(GET_PTR(new_key));
365 return hti_cas(hti, key, key_hash, expected, new); // tail-call
367 TRACE("h2", "hti_cas: installed key %p in entry %p", new_key, ent);
370 TRACE("h0", "hti_cas: entry for key %p is %p",
371 (hti->ht->key_type == NULL) ? (void *)ent->key : GET_PTR(ent->key), ent);
373 // If the entry is in the middle of a copy, the copy must be completed first.
374 map_val_t ent_val = ent->val;
375 if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
376 if (ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1)) {
377 int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next);
379 (void)SYNC_ADD(&hti->num_entries_copied, 1);
381 TRACE("h0", "hti_cas: value in the middle of a copy, copy completed by %s",
382 (did_copy ? "self" : "other"), 0);
384 TRACE("h0", "hti_cas: value copied to next table, retry on next table", 0, 0);
388 // Fail if the old value is not consistent with the caller's expectation.
389 int old_existed = (ent_val != TOMBSTONE && ent_val != DOES_NOT_EXIST);
390 if (EXPECT_FALSE(expected != CAS_EXPECT_WHATEVER && expected != ent_val)) {
391 if (EXPECT_FALSE(expected != (old_existed ? CAS_EXPECT_EXISTS : CAS_EXPECT_DOES_NOT_EXIST))) {
392 TRACE("h1", "hti_cas: value %p expected by caller not found; found value %p",
398 // No need to update if value is unchanged.
399 if ((new == DOES_NOT_EXIST && !old_existed) || ent_val == new) {
400 TRACE("h1", "hti_cas: old value and new value were the same", 0, 0);
404 // CAS the value into the entry. Retry if it fails.
405 map_val_t v = SYNC_CAS(&ent->val, ent_val, new == DOES_NOT_EXIST ? TOMBSTONE : new);
406 if (EXPECT_FALSE(v != ent_val)) {
407 TRACE("h0", "hti_cas: value CAS failed; expected %p found %p", ent_val, v);
408 return hti_cas(hti, key, key_hash, expected, new); // recursive tail-call
411 // The set succeeded. Adjust the value count.
412 if (old_existed && new == DOES_NOT_EXIST) {
413 (void)SYNC_ADD(&hti->count, -1);
414 } else if (!old_existed && new != DOES_NOT_EXIST) {
415 (void)SYNC_ADD(&hti->count, 1);
418 // Return the previous value.
419 TRACE("h0", "hti_cas: CAS succeeded; old value %p new value %p", ent_val, new);
424 static map_val_t hti_get (hti_t *hti, map_key_t key, uint32_t key_hash) {
426 volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty);
428 // When hti_lookup() returns NULL it means we hit the reprobe limit while
429 // searching the table. In that case, if a copy is in progress the key
430 // might exist in the copy.
431 if (EXPECT_FALSE(ent == NULL)) {
432 if (VOLATILE_DEREF(hti).next != NULL)
433 return hti_get(hti->next, key, key_hash); // recursive tail-call
434 return DOES_NOT_EXIST;
438 return DOES_NOT_EXIST;
440 // If the entry is being copied, finish the copy and retry on the next table.
441 map_val_t ent_val = ent->val;
442 if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
443 if (EXPECT_FALSE(ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1))) {
444 int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next);
446 (void)SYNC_ADD(&hti->num_entries_copied, 1);
449 return hti_get(VOLATILE_DEREF(hti).next, key, key_hash); // tail-call
452 return (ent_val == TOMBSTONE) ? DOES_NOT_EXIST : ent_val;
456 map_val_t ht_get (hashtable_t *ht, map_key_t key) {
458 uint32_t hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
460 uint32_t hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
462 return hti_get(ht->hti, key, hash);
465 // returns TRUE if copy is done
466 static int hti_help_copy (hti_t *hti) {
467 volatile entry_t *ent;
469 size_t total_copied = hti->num_entries_copied;
470 size_t num_copied = 0;
471 size_t x = hti->copy_scan;
473 TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<<hti->scale);
474 if (total_copied != (1 << hti->scale)) {
475 // Panic if we've been around the array twice and still haven't finished the copy.
476 int panic = (x >= (1 << (hti->scale + 1)));
478 limit = ENTRIES_PER_COPY_CHUNK;
480 // Reserve some entries for this thread to copy. There is a race condition here because the
481 // fetch and add isn't atomic, but that is ok.
482 hti->copy_scan = x + ENTRIES_PER_COPY_CHUNK;
484 // <copy_scan> might be larger than the size of the table, if some thread stalls while
485 // copying. In that case we just wrap around to the begining and make another pass through
487 ent = hti->table + (x & MASK(hti->scale));
489 TRACE("h1", "ht_cas: help copy panic", 0, 0);
490 // scan the whole table
492 limit = (1 << hti->scale);
496 for (int i = 0; i < limit; ++i) {
497 num_copied += hti_copy_entry(hti, ent++, 0, hti->next);
498 assert(ent <= hti->table + (1 << hti->scale));
500 if (num_copied != 0) {
501 total_copied = SYNC_ADD(&hti->num_entries_copied, num_copied);
505 return (total_copied == (1 << hti->scale));
508 static void hti_defer_free (hti_t *hti) {
509 assert(hti->ref_count == 0);
511 for (uint32_t i = 0; i < (1 << hti->scale); ++i) {
512 map_key_t key = hti->table[i].key;
513 map_val_t val = hti->table[i].val;
514 if (val == COPIED_VALUE)
516 assert(!IS_TAGGED(val, TAG1) || val == TAG_VALUE(TOMBSTONE, TAG1)); // copy not in progress
517 if (hti->ht->key_type != NULL && key != DOES_NOT_EXIST) {
518 rcu_defer_free(GET_PTR(key));
521 #ifdef USE_SYSTEM_MALLOC
522 rcu_defer_free(hti->unaligned_table_ptr);
524 rcu_defer_free((void *)hti->table);
529 static void hti_release (hti_t *hti) {
530 assert(hti->ref_count > 0);
531 int ref_count = SYNC_ADD(&hti->ref_count, -1);
532 if (ref_count == 0) {
538 map_val_t ht_cas (hashtable_t *ht, map_key_t key, map_val_t expected_val, map_val_t new_val) {
540 TRACE("h2", "ht_cas: key %p ht %p", key, ht);
541 TRACE("h2", "ht_cas: expected val %p new val %p", expected_val, new_val);
542 assert(key != DOES_NOT_EXIST);
543 assert(!IS_TAGGED(new_val, TAG1) && new_val != DOES_NOT_EXIST && new_val != TOMBSTONE);
545 hti_t *hti = ht->hti;
547 // Help with an ongoing copy.
548 if (EXPECT_FALSE(hti->next != NULL)) {
549 int done = hti_help_copy(hti);
551 // Unlink fully copied tables.
554 if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) {
562 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
564 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
566 while ((old_val = hti_cas(hti, key, key_hash, expected_val, new_val)) == COPIED_VALUE) {
571 return old_val == TOMBSTONE ? DOES_NOT_EXIST : old_val;
574 // Remove the value in <ht> associated with <key>. Returns the value removed, or DOES_NOT_EXIST if there was
575 // no value for that key.
576 map_val_t ht_remove (hashtable_t *ht, map_key_t key) {
577 hti_t *hti = ht->hti;
580 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
582 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
585 val = hti_cas(hti, key, key_hash, CAS_EXPECT_WHATEVER, DOES_NOT_EXIST);
586 if (val != COPIED_VALUE)
587 return val == TOMBSTONE ? DOES_NOT_EXIST : val;
594 // Returns the number of key-values pairs in <ht>
595 size_t ht_count (hashtable_t *ht) {
596 hti_t *hti = ht->hti;
605 // Allocate and initialize a new hash table.
606 hashtable_t *ht_alloc (const datatype_t *key_type) {
607 hashtable_t *ht = nbd_malloc(sizeof(hashtable_t));
608 ht->key_type = key_type;
609 ht->hti = (hti_t *)hti_alloc(ht, MIN_SCALE);
613 // Free <ht> and its internal structures.
614 void ht_free (hashtable_t *ht) {
615 hti_t *hti = ht->hti;
617 hti_t *next = hti->next;
618 assert(hti->ref_count == 1);
625 void ht_print (hashtable_t *ht) {
626 hti_t *hti = ht->hti;
628 printf("hti:%p scale:%u count:%d copied:%d\n", hti, hti->scale, hti->count, hti->num_entries_copied);
629 for (int i = 0; i < (1 << hti->scale); ++i) {
630 volatile entry_t *ent = hti->table + i;
631 printf("[0x%x] 0x%llx:0x%llx\n", i, (uint64_t)ent->key, (uint64_t)ent->val);
641 ht_iter_t *ht_iter_begin (hashtable_t *ht, map_key_t key) {
646 while (hti->next != NULL) {
647 do { } while (hti_help_copy(hti) != TRUE);
651 ref_count = hti->ref_count;
654 } while (ref_count != SYNC_CAS(&hti->ref_count, ref_count, ref_count + 1));
655 } while (ref_count == 0);
657 ht_iter_t *iter = nbd_malloc(sizeof(ht_iter_t));
664 map_val_t ht_iter_next (ht_iter_t *iter, map_key_t *key_ptr) {
665 volatile entry_t *ent;
668 size_t table_size = (1 << iter->hti->scale);
671 if (iter->idx == table_size) {
672 return DOES_NOT_EXIST;
674 ent = &iter->hti->table[iter->idx];
675 key = (iter->hti->ht->key_type == NULL) ? (map_key_t)ent->key : (map_key_t)GET_PTR(ent->key);
678 } while (key == DOES_NOT_EXIST || val == DOES_NOT_EXIST || val == TOMBSTONE);
683 if (val == COPIED_VALUE) {
684 const datatype_t *key_type = iter->hti->ht->key_type;
686 uint32_t hash = (key_type == NULL) ? murmur32_4b((uint64_t)key) : key_type->hash((void *)key);
688 uint32_t hash = (key_type == NULL) ? murmur32_8b((uint64_t)key) : key_type->hash((void *)key);
690 val = hti_get(iter->hti->next, (map_key_t)ent->key, hash);
696 void ht_iter_free (ht_iter_t *iter) {
697 hti_release(iter->hti);