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.
14 * 11FebO9 - Bug fix in ht_iter_next() from Rui Ueyama
22 #include "hashtable.h"
25 #define GET_PTR(x) ((void *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t
27 #define GET_PTR(x) ((void *)(x))
30 typedef struct entry {
36 volatile entry_t *table;
37 hashtable_t *ht; // parent ht;
39 #ifdef USE_SYSTEM_MALLOC
40 void *unaligned_table_ptr; // system malloc doesn't guarentee cache-line alignment
45 int count; // TODO: make these counters distributed
46 int num_entries_copied;
57 const datatype_t *key_type;
60 static const map_val_t COPIED_VALUE = TAG_VALUE(DOES_NOT_EXIST, TAG1);
61 static const map_val_t TOMBSTONE = STRIP_TAG(-1, TAG1);
63 static const unsigned ENTRIES_PER_BUCKET = CACHE_LINE_SIZE/sizeof(entry_t);
64 static const unsigned ENTRIES_PER_COPY_CHUNK = CACHE_LINE_SIZE/sizeof(entry_t)*2;
65 static const unsigned MIN_SCALE = 4; // min 16 entries (4 buckets)
66 static const unsigned MAX_BUCKETS_TO_PROBE = 250;
68 static int hti_copy_entry (hti_t *ht1, volatile entry_t *ent, uint32_t ent_key_hash, hti_t *ht2);
70 // Choose the next bucket to probe using the high-order bits of <key_hash>.
71 static inline int get_next_ndx(int old_ndx, uint32_t key_hash, int ht_scale) {
72 int incr = (key_hash >> (32 - ht_scale));
73 incr += !incr; // If the increment is 0, make it 1.
74 return (old_ndx + incr) & MASK(ht_scale);
77 // Lookup <key> in <hti>.
79 // Return the entry that <key> is in, or if <key> isn't in <hti> return the entry that it would be
80 // in if it were inserted into <hti>. If there is no room for <key> in <hti> then return NULL, to
81 // indicate that the caller should look in <hti->next>.
83 // Record if the entry being returned is empty. Otherwise the caller will have to waste time
84 // re-comparing the keys to confirm that it did not lose a race to fill an empty entry.
85 static volatile entry_t *hti_lookup (hti_t *hti, map_key_t key, uint32_t key_hash, int *is_empty) {
86 TRACE("h2", "hti_lookup(key %p in hti %p)", key, hti);
89 // Probe one cache line at a time
90 int ndx = key_hash & MASK(hti->scale); // the first entry to search
91 for (int i = 0; i < hti->max_probe; ++i) {
93 // The start of the bucket is the first entry in the cache line.
94 volatile entry_t *bucket = hti->table + (ndx & ~(ENTRIES_PER_BUCKET-1));
96 // Start searching at the indexed entry. Then loop around to the begining of the cache line.
97 for (int j = 0; j < ENTRIES_PER_BUCKET; ++j) {
98 volatile entry_t *ent = bucket + ((ndx + j) & (ENTRIES_PER_BUCKET-1));
100 map_key_t ent_key = ent->key;
101 if (ent_key == DOES_NOT_EXIST) {
102 TRACE("h1", "hti_lookup: entry %p for key %p is empty", ent,
103 (hti->ht->key_type == NULL) ? (void *)key : GET_PTR(key));
104 *is_empty = 1; // indicate an empty so the caller avoids an expensive key compare
108 // Compare <key> with the key in the entry.
109 if (EXPECT_TRUE(hti->ht->key_type == NULL)) {
110 // fast path for integer keys
111 if (ent_key == key) {
112 TRACE("h1", "hti_lookup: found entry %p with key %p", ent, ent_key);
117 // The key in <ent> is made up of two parts. The 48 low-order bits are a pointer. The
118 // high-order 16 bits are taken from the hash. The bits from the hash are used as a
119 // quick check to rule out non-equal keys without doing a complete compare.
120 if ((key_hash >> 16) == (ent_key >> 48)) {
122 if (hti->ht->key_type->cmp(GET_PTR(ent_key), (void *)key) == 0) {
123 TRACE("h1", "hti_lookup: found entry %p with key %p", ent, GET_PTR(ent_key));
132 ndx = get_next_ndx(ndx, key_hash, hti->scale);
135 // maximum number of probes exceeded
136 TRACE("h1", "hti_lookup: maximum number of probes exceeded returning 0x0", 0, 0);
140 // Allocate and initialize a hti_t with 2^<scale> entries.
141 static hti_t *hti_alloc (hashtable_t *parent, int scale) {
142 hti_t *hti = (hti_t *)nbd_malloc(sizeof(hti_t));
143 memset(hti, 0, sizeof(hti_t));
146 size_t sz = sizeof(entry_t) * (1 << scale);
147 #ifdef USE_SYSTEM_MALLOC
148 hti->unaligned_table_ptr = nbd_malloc(sz + CACHE_LINE_SIZE - 1);
149 hti->table = (void *)(((size_t)hti->unaligned_table_ptr + CACHE_LINE_SIZE - 1) & ~(CACHE_LINE_SIZE - 1));
151 hti->table = nbd_malloc(sz);
153 memset((void *)hti->table, 0, sz);
155 // When searching for a key probe a maximum of 1/4 of the buckets up to 1000 buckets.
156 hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4;
157 if (hti->max_probe > MAX_BUCKETS_TO_PROBE) {
158 hti->max_probe = MAX_BUCKETS_TO_PROBE;
162 hti->ref_count = 1; // one for the parent
164 assert(hti->scale >= MIN_SCALE && hti->scale < 63); // size must be a power of 2
165 assert(sizeof(entry_t) * ENTRIES_PER_BUCKET % CACHE_LINE_SIZE == 0); // divisible into cache
166 assert((size_t)hti->table % CACHE_LINE_SIZE == 0); // cache aligned
171 // Called when <hti> runs out of room for new keys.
173 // Initiates a copy by creating a larger hti_t and installing it in <hti->next>.
174 static void hti_start_copy (hti_t *hti) {
175 TRACE("h0", "hti_start_copy(hti %p scale %llu)", hti, hti->scale);
177 // heuristics to determine the size of the new table
178 size_t count = ht_count(hti->ht);
179 unsigned int new_scale = hti->scale;
180 new_scale += (count > (1 << (new_scale - 2))); // double size if more than 1/4 full
181 new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full
183 // Allocate the new table and attempt to install it.
184 hti_t *next = hti_alloc(hti->ht, new_scale);
185 hti_t *old_next = SYNC_CAS(&hti->next, NULL, next);
186 if (old_next != NULL) {
187 // Another thread beat us to it.
188 TRACE("h0", "hti_start_copy: lost race to install new hti; found %p", old_next, 0);
189 #ifdef USE_SYSTEM_MALLOC
190 nbd_free(next->unaligned_table_ptr);
192 nbd_free((void *)next->table);
196 TRACE("h0", "hti_start_copy: new hti %p scale %llu", next, next->scale);
199 // Copy the key and value stored in <ht1_ent> (which must be an entry in <ht1>) to <ht2>.
201 // Return 1 unless <ht1_ent> is already copied (then return 0), so the caller can account for the total
202 // number of entries left to copy.
203 static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_hash, hti_t *ht2) {
204 TRACE("h2", "hti_copy_entry: entry %p to table %p", ht1_ent, ht2);
208 assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1 << ht1->scale));
210 assert(key_hash == 0 || ht1->ht->key_type == NULL || (key_hash >> 16) == (ht1_ent->key >> 48));
213 map_val_t ht1_ent_val = ht1_ent->val;
214 if (EXPECT_FALSE(ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1))) {
215 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
216 return FALSE; // already copied
219 // Kill empty entries.
220 if (EXPECT_FALSE(ht1_ent_val == DOES_NOT_EXIST)) {
221 map_val_t ht1_ent_val = SYNC_CAS(&ht1_ent->val, DOES_NOT_EXIST, COPIED_VALUE);
222 if (ht1_ent_val == DOES_NOT_EXIST) {
223 TRACE("h1", "hti_copy_entry: empty entry %p killed", ht1_ent, 0);
226 TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is not empty", ht1_ent, 0);
229 // Tag the value in the old entry to indicate a copy is in progress.
230 ht1_ent_val = SYNC_FETCH_AND_OR(&ht1_ent->val, TAG_VALUE(0, TAG1));
231 TRACE("h2", "hti_copy_entry: tagged the value %p in old entry %p", ht1_ent_val, ht1_ent);
232 if (ht1_ent_val == COPIED_VALUE || ht1_ent_val == TAG_VALUE(TOMBSTONE, TAG1)) {
233 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
234 return FALSE; // <value> was already copied by another thread.
237 // The old table's dead entries don't need to be copied to the new table
238 if (ht1_ent_val == TOMBSTONE)
241 // Install the key in the new table.
242 map_key_t ht1_ent_key = ht1_ent->key;
243 map_key_t key = (ht1->ht->key_type == NULL) ? (map_key_t)ht1_ent_key : (map_key_t)GET_PTR(ht1_ent_key);
245 // We use 0 to indicate that <key_hash> is uninitiallized. Occasionally the key's hash will really be 0 and we
246 // waste time recomputing it every time. It is rare enough that it won't hurt performance.
249 key_hash = (ht1->ht->key_type == NULL) ? murmur32_4b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key);
251 key_hash = (ht1->ht->key_type == NULL) ? murmur32_8b(ht1_ent_key) : ht1->ht->key_type->hash((void *)key);
255 int ht2_ent_is_empty;
256 volatile entry_t *ht2_ent = hti_lookup(ht2, key, key_hash, &ht2_ent_is_empty);
257 TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_ent, ht2_ent);
259 // It is possible that there isn't any room in the new table either.
260 if (EXPECT_FALSE(ht2_ent == NULL)) {
261 TRACE("h0", "hti_copy_entry: no room in table %p copy to next table %p", ht2, ht2->next);
262 if (ht2->next == NULL) {
263 hti_start_copy(ht2); // initiate nested copy, if not already started
265 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call
268 if (ht2_ent_is_empty) {
269 map_key_t old_ht2_ent_key = SYNC_CAS(&ht2_ent->key, DOES_NOT_EXIST, ht1_ent_key);
270 if (old_ht2_ent_key != DOES_NOT_EXIST) {
271 TRACE("h0", "hti_copy_entry: lost race to CAS key %p into new entry; found %p",
272 ht1_ent_key, old_ht2_ent_key);
273 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2); // recursive tail-call
277 // Copy the value to the entry in the new table.
278 ht1_ent_val = STRIP_TAG(ht1_ent_val, TAG1);
279 map_val_t old_ht2_ent_val = SYNC_CAS(&ht2_ent->val, DOES_NOT_EXIST, ht1_ent_val);
281 // If there is a nested copy in progress, we might have installed the key into a dead entry.
282 if (old_ht2_ent_val == COPIED_VALUE) {
283 TRACE("h0", "hti_copy_entry: nested copy in progress; copy %p to next table %p", ht2_ent, ht2->next);
284 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call
287 // Mark the old entry as dead.
288 ht1_ent->val = COPIED_VALUE;
290 // Update the count if we were the one that completed the copy.
291 if (old_ht2_ent_val == DOES_NOT_EXIST) {
292 TRACE("h0", "hti_copy_entry: key %p value %p copied to new entry", key, ht1_ent_val);
293 (void)SYNC_ADD(&ht1->count, -1);
294 (void)SYNC_ADD(&ht2->count, 1);
298 TRACE("h0", "hti_copy_entry: lost race to install value %p in new entry; found value %p",
299 ht1_ent_val, old_ht2_ent_val);
300 return FALSE; // another thread completed the copy
303 // Compare <expected> with the existing value associated with <key>. If the values match then
304 // replace the existing value with <new>. If <new> is DOES_NOT_EXIST, delete the value associated with
305 // the key by replacing it with a TOMBSTONE.
307 // Return the previous value associated with <key>, or DOES_NOT_EXIST if <key> is not in the table
308 // or associated with a TOMBSTONE. If a copy is in progress and <key> has been copied to the next
309 // table then return COPIED_VALUE.
311 // NOTE: the returned value matches <expected> iff the set succeeds
313 // Certain values of <expected> have special meaning. If <expected> is CAS_EXPECT_EXISTS then any
314 // real value matches (i.ent. not a TOMBSTONE or DOES_NOT_EXIST) as long as <key> is in the table. If
315 // <expected> is CAS_EXPECT_WHATEVER then skip the test entirely.
317 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) {
318 TRACE("h1", "hti_cas: hti %p key %p", hti, key);
319 TRACE("h1", "hti_cas: value %p expect %p", new, expected);
321 assert(!IS_TAGGED(new, TAG1));
325 volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty);
327 // There is no room for <key>, grow the table and try again.
329 if (hti->next == NULL) {
335 // Install <key> in the table if it doesn't exist.
337 TRACE("h0", "hti_cas: entry %p is empty", ent, 0);
338 if (expected != CAS_EXPECT_WHATEVER && expected != CAS_EXPECT_DOES_NOT_EXIST)
339 return DOES_NOT_EXIST;
341 // No need to do anything, <key> is already deleted.
342 if (new == DOES_NOT_EXIST)
343 return DOES_NOT_EXIST;
345 // Allocate <new_key>.
346 map_key_t new_key = (hti->ht->key_type == NULL)
348 : (map_key_t)hti->ht->key_type->clone((void *)key);
350 if (EXPECT_FALSE(hti->ht->key_type != NULL)) {
351 // Combine <new_key> pointer with bits from its hash
352 new_key = ((uint64_t)(key_hash >> 16) << 48) | new_key;
356 // CAS the key into the table.
357 map_key_t old_ent_key = SYNC_CAS(&ent->key, DOES_NOT_EXIST, new_key);
359 // Retry if another thread stole the entry out from under us.
360 if (old_ent_key != DOES_NOT_EXIST) {
361 TRACE("h0", "hti_cas: lost race to install key %p in entry %p", new_key, ent);
362 TRACE("h0", "hti_cas: found %p instead of NULL",
363 (hti->ht->key_type == NULL) ? (void *)old_ent_key : GET_PTR(old_ent_key), 0);
364 if (hti->ht->key_type != NULL) {
365 nbd_free(GET_PTR(new_key));
367 return hti_cas(hti, key, key_hash, expected, new); // tail-call
369 TRACE("h2", "hti_cas: installed key %p in entry %p", new_key, ent);
372 TRACE("h0", "hti_cas: entry for key %p is %p",
373 (hti->ht->key_type == NULL) ? (void *)ent->key : GET_PTR(ent->key), ent);
375 // If the entry is in the middle of a copy, the copy must be completed first.
376 map_val_t ent_val = ent->val;
377 if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
378 if (ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1)) {
379 int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next);
381 (void)SYNC_ADD(&hti->num_entries_copied, 1);
383 TRACE("h0", "hti_cas: value in the middle of a copy, copy completed by %s",
384 (did_copy ? "self" : "other"), 0);
386 TRACE("h0", "hti_cas: value copied to next table, retry on next table", 0, 0);
390 // Fail if the old value is not consistent with the caller's expectation.
391 int old_existed = (ent_val != TOMBSTONE && ent_val != DOES_NOT_EXIST);
392 if (EXPECT_FALSE(expected != CAS_EXPECT_WHATEVER && expected != ent_val)) {
393 if (EXPECT_FALSE(expected != (old_existed ? CAS_EXPECT_EXISTS : CAS_EXPECT_DOES_NOT_EXIST))) {
394 TRACE("h1", "hti_cas: value %p expected by caller not found; found value %p",
400 // No need to update if value is unchanged.
401 if ((new == DOES_NOT_EXIST && !old_existed) || ent_val == new) {
402 TRACE("h1", "hti_cas: old value and new value were the same", 0, 0);
406 // CAS the value into the entry. Retry if it fails.
407 map_val_t v = SYNC_CAS(&ent->val, ent_val, new == DOES_NOT_EXIST ? TOMBSTONE : new);
408 if (EXPECT_FALSE(v != ent_val)) {
409 TRACE("h0", "hti_cas: value CAS failed; expected %p found %p", ent_val, v);
410 return hti_cas(hti, key, key_hash, expected, new); // recursive tail-call
413 // The set succeeded. Adjust the value count.
414 if (old_existed && new == DOES_NOT_EXIST) {
415 (void)SYNC_ADD(&hti->count, -1);
416 } else if (!old_existed && new != DOES_NOT_EXIST) {
417 (void)SYNC_ADD(&hti->count, 1);
420 // Return the previous value.
421 TRACE("h0", "hti_cas: CAS succeeded; old value %p new value %p", ent_val, new);
426 static map_val_t hti_get (hti_t *hti, map_key_t key, uint32_t key_hash) {
428 volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty);
430 // When hti_lookup() returns NULL it means we hit the reprobe limit while
431 // searching the table. In that case, if a copy is in progress the key
432 // might exist in the copy.
433 if (EXPECT_FALSE(ent == NULL)) {
434 if (VOLATILE_DEREF(hti).next != NULL)
435 return hti_get(hti->next, key, key_hash); // recursive tail-call
436 return DOES_NOT_EXIST;
440 return DOES_NOT_EXIST;
442 // If the entry is being copied, finish the copy and retry on the next table.
443 map_val_t ent_val = ent->val;
444 if (EXPECT_FALSE(IS_TAGGED(ent_val, TAG1))) {
445 if (EXPECT_FALSE(ent_val != COPIED_VALUE && ent_val != TAG_VALUE(TOMBSTONE, TAG1))) {
446 int did_copy = hti_copy_entry(hti, ent, key_hash, VOLATILE_DEREF(hti).next);
448 (void)SYNC_ADD(&hti->num_entries_copied, 1);
451 return hti_get(VOLATILE_DEREF(hti).next, key, key_hash); // tail-call
454 return (ent_val == TOMBSTONE) ? DOES_NOT_EXIST : ent_val;
458 map_val_t ht_get (hashtable_t *ht, map_key_t key) {
460 uint32_t hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
462 uint32_t hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
464 return hti_get(ht->hti, key, hash);
467 // returns TRUE if copy is done
468 static int hti_help_copy (hti_t *hti) {
469 volatile entry_t *ent;
471 size_t total_copied = hti->num_entries_copied;
472 size_t num_copied = 0;
473 size_t x = hti->copy_scan;
475 TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<<hti->scale);
476 if (total_copied != (1 << hti->scale)) {
477 // Panic if we've been around the array twice and still haven't finished the copy.
478 int panic = (x >= (1 << (hti->scale + 1)));
480 limit = ENTRIES_PER_COPY_CHUNK;
482 // Reserve some entries for this thread to copy. There is a race condition here because the
483 // fetch and add isn't atomic, but that is ok.
484 hti->copy_scan = x + ENTRIES_PER_COPY_CHUNK;
486 // <copy_scan> might be larger than the size of the table, if some thread stalls while
487 // copying. In that case we just wrap around to the begining and make another pass through
489 ent = hti->table + (x & MASK(hti->scale));
491 TRACE("h1", "ht_cas: help copy panic", 0, 0);
492 // scan the whole table
494 limit = (1 << hti->scale);
498 for (int i = 0; i < limit; ++i) {
499 num_copied += hti_copy_entry(hti, ent++, 0, hti->next);
500 assert(ent <= hti->table + (1 << hti->scale));
502 if (num_copied != 0) {
503 total_copied = SYNC_ADD(&hti->num_entries_copied, num_copied);
507 return (total_copied == (1 << hti->scale));
510 static void hti_defer_free (hti_t *hti) {
511 assert(hti->ref_count == 0);
513 for (uint32_t i = 0; i < (1 << hti->scale); ++i) {
514 map_key_t key = hti->table[i].key;
515 map_val_t val = hti->table[i].val;
516 if (val == COPIED_VALUE)
518 assert(!IS_TAGGED(val, TAG1) || val == TAG_VALUE(TOMBSTONE, TAG1)); // copy not in progress
519 if (hti->ht->key_type != NULL && key != DOES_NOT_EXIST) {
520 rcu_defer_free(GET_PTR(key));
523 #ifdef USE_SYSTEM_MALLOC
524 rcu_defer_free(hti->unaligned_table_ptr);
526 rcu_defer_free((void *)hti->table);
531 static void hti_release (hti_t *hti) {
532 assert(hti->ref_count > 0);
533 int ref_count = SYNC_ADD(&hti->ref_count, -1);
534 if (ref_count == 0) {
540 map_val_t ht_cas (hashtable_t *ht, map_key_t key, map_val_t expected_val, map_val_t new_val) {
542 TRACE("h2", "ht_cas: key %p ht %p", key, ht);
543 TRACE("h2", "ht_cas: expected val %p new val %p", expected_val, new_val);
544 assert(key != DOES_NOT_EXIST);
545 assert(!IS_TAGGED(new_val, TAG1) && new_val != DOES_NOT_EXIST && new_val != TOMBSTONE);
547 hti_t *hti = ht->hti;
549 // Help with an ongoing copy.
550 if (EXPECT_FALSE(hti->next != NULL)) {
551 int done = hti_help_copy(hti);
553 // Unlink fully copied tables.
556 if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) {
564 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
566 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
568 while ((old_val = hti_cas(hti, key, key_hash, expected_val, new_val)) == COPIED_VALUE) {
573 return old_val == TOMBSTONE ? DOES_NOT_EXIST : old_val;
576 // Remove the value in <ht> associated with <key>. Returns the value removed, or DOES_NOT_EXIST if there was
577 // no value for that key.
578 map_val_t ht_remove (hashtable_t *ht, map_key_t key) {
579 hti_t *hti = ht->hti;
582 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_4b((uint64_t)key) : ht->key_type->hash((void *)key);
584 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash((void *)key);
587 val = hti_cas(hti, key, key_hash, CAS_EXPECT_WHATEVER, DOES_NOT_EXIST);
588 if (val != COPIED_VALUE)
589 return val == TOMBSTONE ? DOES_NOT_EXIST : val;
596 // Returns the number of key-values pairs in <ht>
597 size_t ht_count (hashtable_t *ht) {
598 hti_t *hti = ht->hti;
607 // Allocate and initialize a new hash table.
608 hashtable_t *ht_alloc (const datatype_t *key_type) {
609 hashtable_t *ht = nbd_malloc(sizeof(hashtable_t));
610 ht->key_type = key_type;
611 ht->hti = (hti_t *)hti_alloc(ht, MIN_SCALE);
615 // Free <ht> and its internal structures.
616 void ht_free (hashtable_t *ht) {
617 hti_t *hti = ht->hti;
619 hti_t *next = hti->next;
620 assert(hti->ref_count == 1);
627 void ht_print (hashtable_t *ht) {
628 hti_t *hti = ht->hti;
630 printf("hti:%p scale:%u count:%d copied:%d\n", hti, hti->scale, hti->count, hti->num_entries_copied);
631 for (int i = 0; i < (1 << hti->scale); ++i) {
632 volatile entry_t *ent = hti->table + i;
633 printf("[0x%x] 0x%llx:0x%llx\n", i, (uint64_t)ent->key, (uint64_t)ent->val);
643 ht_iter_t *ht_iter_begin (hashtable_t *ht, map_key_t key) {
648 while (hti->next != NULL) {
649 do { } while (hti_help_copy(hti) != TRUE);
653 ref_count = hti->ref_count;
656 } while (ref_count != SYNC_CAS(&hti->ref_count, ref_count, ref_count + 1));
657 } while (ref_count == 0);
659 ht_iter_t *iter = nbd_malloc(sizeof(ht_iter_t));
666 map_val_t ht_iter_next (ht_iter_t *iter, map_key_t *key_ptr) {
667 volatile entry_t *ent;
670 size_t table_size = (1 << iter->hti->scale);
673 if (iter->idx == table_size) {
674 return DOES_NOT_EXIST;
676 ent = &iter->hti->table[iter->idx];
677 key = (iter->hti->ht->key_type == NULL) ? (map_key_t)ent->key : (map_key_t)GET_PTR(ent->key);
680 } while (key == DOES_NOT_EXIST || val == DOES_NOT_EXIST || val == TOMBSTONE);
682 if (val == COPIED_VALUE) {
683 const datatype_t *key_type = iter->hti->ht->key_type;
685 uint32_t hash = (key_type == NULL) ? murmur32_4b((uint64_t)key) : key_type->hash((void *)key);
687 uint32_t hash = (key_type == NULL) ? murmur32_8b((uint64_t)key) : key_type->hash((void *)key);
689 val = hti_get(iter->hti->next, (map_key_t)ent->key, hash);
691 // Go to the next entry if key is already deleted.
692 if (val == DOES_NOT_EXIST)
693 return ht_iter_next(iter, key_ptr); // recursive tail-call
702 void ht_iter_free (ht_iter_t *iter) {
703 hti_release(iter->hti);