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 like SPARC that provide
12 * weaker operations which would still do the job.
21 #define GET_PTR(x) ((void *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t
23 typedef struct ht_entry {
29 volatile entry_t *table;
30 hashtable_t *ht; // parent ht;
34 int count; // TODO: make these counters distributed
35 int num_entries_copied;
41 const datatype_t *key_type;
44 static const map_impl_t ht_map_impl = {
45 (map_alloc_t)ht_alloc, (map_cas_t)ht_cas, (map_get_t)ht_get, (map_remove_t)ht_remove,
46 (map_count_t)ht_count, (map_print_t)ht_print, (map_free_t)ht_free
49 const map_impl_t *MAP_TYPE_HASHTABLE = &ht_map_impl;
51 static const uint64_t COPIED_VALUE = -1;
52 static const uint64_t TOMBSTONE = STRIP_TAG(-1);
54 static const unsigned ENTRIES_PER_BUCKET = CACHE_LINE_SIZE/sizeof(entry_t);
55 static const unsigned ENTRIES_PER_COPY_CHUNK = CACHE_LINE_SIZE/sizeof(entry_t)*2;
56 static const unsigned MIN_SCALE = 4; // min 16 entries (4 buckets)
57 static const unsigned MAX_BUCKETS_TO_PROBE = 250;
59 static int hti_copy_entry (hti_t *ht1, volatile entry_t *ent, uint32_t ent_key_hash, hti_t *ht2);
61 // Choose the next bucket to probe using the high-order bits of <key_hash>.
62 static inline int get_next_ndx(int old_ndx, uint32_t key_hash, int ht_scale) {
63 int incr = (key_hash >> (32 - ht_scale));
64 incr += !incr; // If the increment is 0, make it 1.
65 return (old_ndx + incr) & MASK(ht_scale);
68 // Lookup <key> in <hti>.
70 // Return the entry that <key> is in, or if <key> isn't in <hti> return the entry that it would be
71 // in if it were inserted into <hti>. If there is no room for <key> in <hti> then return NULL, to
72 // indicate that the caller should look in <hti->next>.
74 // Record if the entry being returned is empty. Otherwise the caller will have to waste time
75 // re-comparing the keys to confirm that it did not lose a race to fill an empty entry.
76 static volatile entry_t *hti_lookup (hti_t *hti, void *key, uint32_t key_hash, int *is_empty) {
77 TRACE("h2", "hti_lookup(key %p in hti %p)", key, hti);
80 // Probe one cache line at a time
81 int ndx = key_hash & MASK(hti->scale); // the first entry to search
82 for (int i = 0; i < hti->max_probe; ++i) {
84 // The start of the bucket is the first entry in the cache line.
85 volatile entry_t *bucket = hti->table + (ndx & ~(ENTRIES_PER_BUCKET-1));
87 // Start searching at the indexed entry. Then loop around to the begining of the cache line.
88 for (int j = 0; j < ENTRIES_PER_BUCKET; ++j) {
89 volatile entry_t *ent = bucket + ((ndx + j) & (ENTRIES_PER_BUCKET-1));
91 uint64_t ent_key = ent->key;
92 if (ent_key == DOES_NOT_EXIST) {
93 TRACE("h1", "hti_lookup: entry %p for key %p is empty", ent,
94 (hti->ht->key_type == NULL) ? (void *)ent_key : GET_PTR(ent_key));
95 *is_empty = 1; // indicate an empty so the caller avoids an expensive key compare
99 // Compare <key> with the key in the entry.
100 if (EXPECT_TRUE(hti->ht->key_type == NULL)) {
101 // fast path for integer keys
102 if (ent_key == (uint64_t)key) {
103 TRACE("h1", "hti_lookup: found entry %p with key %p", ent, ent_key);
107 // The key in <ent> is made up of two parts. The 48 low-order bits are a pointer. The
108 // high-order 16 bits are taken from the hash. The bits from the hash are used as a
109 // quick check to rule out non-equal keys without doing a complete compare.
110 if ((key_hash >> 16) == (ent_key >> 48) && hti->ht->key_type->cmp(GET_PTR(ent_key), key) == 0) {
111 TRACE("h1", "hti_lookup: found entry %p with key %p", ent, GET_PTR(ent_key));
117 ndx = get_next_ndx(ndx, key_hash, hti->scale);
120 // maximum number of probes exceeded
121 TRACE("h1", "hti_lookup: maximum number of probes exceeded returning 0x0", 0, 0);
125 // Allocate and initialize a hti_t with 2^<scale> entries.
126 static hti_t *hti_alloc (hashtable_t *parent, int scale) {
127 // Include enough slop to align the actual table on a cache line boundry
128 size_t n = sizeof(hti_t)
129 + sizeof(entry_t) * (1 << scale)
130 + (CACHE_LINE_SIZE - 1);
131 hti_t *hti = (hti_t *)calloc(n, 1);
133 // Align the table of hash entries on a cache line boundry.
134 hti->table = (entry_t *)(((uint64_t)hti + sizeof(hti_t) + (CACHE_LINE_SIZE-1))
135 & ~(CACHE_LINE_SIZE-1));
139 // When searching for a key probe a maximum of 1/4 of the buckets up to 1000 buckets.
140 hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 4;
141 if (hti->max_probe > MAX_BUCKETS_TO_PROBE) {
142 hti->max_probe = MAX_BUCKETS_TO_PROBE;
147 assert(hti->scale >= MIN_SCALE && hti->scale < 63); // size must be a power of 2
148 assert(sizeof(entry_t) * ENTRIES_PER_BUCKET % CACHE_LINE_SIZE == 0); // divisible into cache
149 assert((size_t)hti->table % CACHE_LINE_SIZE == 0); // cache aligned
154 // Called when <hti> runs out of room for new keys.
156 // Initiates a copy by creating a larger hti_t and installing it in <hti->next>.
157 static void hti_start_copy (hti_t *hti) {
158 TRACE("h0", "hti_start_copy(hti %p scale %llu)", hti, hti->scale);
160 // heuristics to determine the size of the new table
161 uint64_t count = ht_count(hti->ht);
162 unsigned int new_scale = hti->scale;
163 new_scale += (count > (1 << (new_scale - 2))); // double size if more than 1/4 full
164 new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full
166 // Allocate the new table and attempt to install it.
167 hti_t *next = hti_alloc(hti->ht, new_scale);
168 hti_t *old_next = SYNC_CAS(&hti->next, NULL, next);
169 if (old_next != NULL) {
170 // Another thread beat us to it.
171 TRACE("h0", "hti_start_copy: lost race to install new hti; found %p", old_next, 0);
175 TRACE("h0", "hti_start_copy: new hti %p scale %llu", next, next->scale);
178 // Copy the key and value stored in <ht1_ent> (which must be an entry in <ht1>) to <ht2>.
180 // Return 1 unless <ht1_ent> is already copied (then return 0), so the caller can account for the total
181 // number of entries left to copy.
182 static int hti_copy_entry (hti_t *ht1, volatile entry_t *ht1_ent, uint32_t key_hash, hti_t *ht2) {
183 TRACE("h2", "hti_copy_entry: entry %p to table %p", ht1_ent, ht2);
187 assert(ht1_ent >= ht1->table && ht1_ent < ht1->table + (1 << ht1->scale));
188 assert(key_hash == 0 || ht1->ht->key_type == NULL || (key_hash >> 16) == (ht1_ent->key >> 48));
190 uint64_t ht1_ent_val = ht1_ent->val;
191 if (EXPECT_FALSE(ht1_ent_val == COPIED_VALUE)) {
192 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
193 return FALSE; // already copied
196 // Kill empty entries.
197 if (EXPECT_FALSE(ht1_ent_val == DOES_NOT_EXIST)) {
198 uint64_t ht1_ent_val = SYNC_CAS(&ht1_ent->val, DOES_NOT_EXIST, COPIED_VALUE);
199 if (ht1_ent_val == DOES_NOT_EXIST) {
200 TRACE("h1", "hti_copy_entry: empty entry %p killed", ht1_ent, 0);
203 if (ht1_ent_val == COPIED_VALUE) {
204 TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is already killed", ht1_ent, 0);
205 return FALSE; // another thread beat us to it
207 TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p; the entry is not empty", ht1_ent, 0);
210 // Tag the value in the old entry to indicate a copy is in progress.
211 ht1_ent_val = SYNC_FETCH_AND_OR(&ht1_ent->val, TAG_VALUE(0));
212 TRACE("h2", "hti_copy_entry: tagged the value %p in old entry %p", ht1_ent_val, ht1_ent);
213 if (ht1_ent_val == COPIED_VALUE) {
214 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_ent, ht2);
215 return FALSE; // <value> was already copied by another thread.
218 // Install the key in the new table.
219 uint64_t ht1_ent_key = ht1_ent->key;
220 void *key = (ht1->ht->key_type == NULL) ? (void *)ht1_ent_key : GET_PTR(ht1_ent_key);
222 // The old table's dead entries don't need to be copied to the new table, but their keys need to be freed.
223 assert(COPIED_VALUE == TAG_VALUE(TOMBSTONE));
224 if (ht1_ent_val == TOMBSTONE) {
225 TRACE("h1", "hti_copy_entry: entry %p old value was deleted, now freeing key %p", ht1_ent, key);
226 if (EXPECT_FALSE(ht1->ht->key_type != NULL)) {
232 // We use 0 to indicate that <key_hash> is uninitiallized. Occasionally the key's hash will really be 0 and we
233 // waste time recomputing it every time. It is rare enough (1 in 65k) that it won't hurt performance.
235 key_hash = (ht1->ht->key_type == NULL) ? murmur32_8b(ht1_ent_key) : ht1->ht->key_type->hash(key);
238 int ht2_ent_is_empty;
239 volatile entry_t *ht2_ent = hti_lookup(ht2, key, key_hash, &ht2_ent_is_empty);
240 TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_ent, ht2_ent);
242 // It is possible that there isn't any room in the new table either.
243 if (EXPECT_FALSE(ht2_ent == NULL)) {
244 TRACE("h0", "hti_copy_entry: no room in table %p copy to next table %p", ht2, ht2->next);
245 if (ht2->next == NULL) {
246 hti_start_copy(ht2); // initiate nested copy, if not already started
248 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call
251 if (ht2_ent_is_empty) {
252 uint64_t old_ht2_ent_key = SYNC_CAS(&ht2_ent->key, DOES_NOT_EXIST, ht1_ent_key);
253 if (old_ht2_ent_key != DOES_NOT_EXIST) {
254 TRACE("h0", "hti_copy_entry: lost race to CAS key %p into new entry; found %p",
255 ht1_ent_key, old_ht2_ent_key);
256 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2); // recursive tail-call
260 // Copy the value to the entry in the new table.
261 ht1_ent_val = STRIP_TAG(ht1_ent_val);
262 uint64_t old_ht2_ent_val = SYNC_CAS(&ht2_ent->val, DOES_NOT_EXIST, ht1_ent_val);
264 // If there is a nested copy in progress, we might have installed the key into a dead entry.
265 if (old_ht2_ent_val == COPIED_VALUE) {
266 TRACE("h0", "hti_copy_entry: nested copy in progress; copy %p to next table %p", ht2_ent, ht2->next);
267 return hti_copy_entry(ht1, ht1_ent, key_hash, ht2->next); // recursive tail-call
270 // Mark the old entry as dead.
271 ht1_ent->val = COPIED_VALUE;
273 // Update the count if we were the one that completed the copy.
274 if (old_ht2_ent_val == DOES_NOT_EXIST) {
275 TRACE("h0", "hti_copy_entry: key %p value %p copied to new entry", key, ht1_ent_val);
276 SYNC_ADD(&ht1->count, -1);
277 SYNC_ADD(&ht2->count, 1);
281 TRACE("h0", "hti_copy_entry: lost race to install value %p in new entry; found value %p",
282 ht1_ent_val, old_ht2_ent_val);
283 return FALSE; // another thread completed the copy
286 // Compare <expected> with the existing value associated with <key>. If the values match then
287 // replace the existing value with <new>. If <new> is TOMBSTONE, delete the value associated with
288 // the key by replacing it with a TOMBSTONE.
290 // Return the previous value associated with <key>, or DOES_NOT_EXIST if <key> is not in the table
291 // or associated with a TOMBSTONE. If a copy is in progress and <key> has been copied to the next
292 // table then return COPIED_VALUE.
294 // NOTE: the returned value matches <expected> iff the set succeeds
296 // Certain values of <expected> have special meaning. If <expected> is CAS_EXPECT_EXISTS then any
297 // real value matches (i.ent. not a TOMBSTONE or DOES_NOT_EXIST) as long as <key> is in the table. If
298 // <expected> is CAS_EXPECT_WHATEVER then skip the test entirely.
300 static uint64_t hti_cas (hti_t *hti, void *key, uint32_t key_hash, uint64_t expected, uint64_t new) {
301 TRACE("h1", "hti_cas: hti %p key %p", hti, key);
302 TRACE("h1", "hti_cas: value %p expect %p", new, expected);
304 assert(new != DOES_NOT_EXIST && !IS_TAGGED(new));
308 volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty);
310 // There is no room for <key>, grow the table and try again.
312 if (hti->next == NULL) {
318 // Install <key> in the table if it doesn't exist.
320 TRACE("h0", "hti_cas: entry %p is empty", ent, 0);
321 if (expected != CAS_EXPECT_WHATEVER && expected != CAS_EXPECT_DOES_NOT_EXIST)
322 return DOES_NOT_EXIST;
324 // No need to do anything, <key> is already deleted.
325 if (new == TOMBSTONE)
326 return DOES_NOT_EXIST;
328 // Allocate <new_key>.
329 uint64_t new_key = (uint64_t)((hti->ht->key_type == NULL) ? key : hti->ht->key_type->clone(key));
330 if (EXPECT_FALSE(hti->ht->key_type != NULL)) {
331 // Combine <new_key> pointer with bits from its hash
332 new_key = ((uint64_t)(key_hash >> 16) << 48) | new_key;
335 // CAS the key into the table.
336 uint64_t old_ent_key = SYNC_CAS(&ent->key, DOES_NOT_EXIST, new_key);
338 // Retry if another thread stole the entry out from under us.
339 if (old_ent_key != DOES_NOT_EXIST) {
340 TRACE("h0", "hti_cas: lost race to install key %p in entry %p", new_key, ent);
341 TRACE("h0", "hti_cas: found %p instead of NULL",
342 (hti->ht->key_type == NULL) ? (void *)old_ent_key : GET_PTR(old_ent_key), 0);
343 if (hti->ht->key_type != NULL) {
344 nbd_free(GET_PTR(new_key));
346 return hti_cas(hti, key, key_hash, expected, new); // tail-call
348 TRACE("h2", "hti_cas: installed key %p in entry %p", new_key, ent);
351 TRACE("h0", "hti_cas: entry for key %p is %p",
352 (hti->ht->key_type == NULL) ? (void *)ent->key : GET_PTR(ent->key), ent);
354 // If the entry is in the middle of a copy, the copy must be completed first.
355 uint64_t ent_val = ent->val;
356 if (EXPECT_FALSE(IS_TAGGED(ent_val))) {
357 if (ent_val != COPIED_VALUE) {
358 int did_copy = hti_copy_entry(hti, ent, key_hash, ((volatile hti_t *)hti)->next);
360 SYNC_ADD(&hti->num_entries_copied, 1);
362 TRACE("h0", "hti_cas: value in the middle of a copy, copy completed by %s",
363 (did_copy ? "self" : "other"), 0);
365 TRACE("h0", "hti_cas: value copied to next table, retry on next table", 0, 0);
369 // Fail if the old value is not consistent with the caller's expectation.
370 int old_existed = (ent_val != TOMBSTONE && ent_val != DOES_NOT_EXIST);
371 if (EXPECT_FALSE(expected != CAS_EXPECT_WHATEVER && expected != ent_val)) {
372 if (EXPECT_FALSE(expected != (old_existed ? CAS_EXPECT_EXISTS : CAS_EXPECT_DOES_NOT_EXIST))) {
373 TRACE("h1", "hti_cas: value %p expected by caller not found; found value %p",
379 // No need to update if value is unchanged.
380 if ((new == TOMBSTONE && !old_existed) || ent_val == new) {
381 TRACE("h1", "hti_cas: old value and new value were the same", 0, 0);
385 // CAS the value into the entry. Retry if it fails.
386 uint64_t v = SYNC_CAS(&ent->val, ent_val, new);
387 if (EXPECT_FALSE(v != ent_val)) {
388 TRACE("h0", "hti_cas: value CAS failed; expected %p found %p", ent_val, v);
389 return hti_cas(hti, key, key_hash, expected, new); // recursive tail-call
392 // The set succeeded. Adjust the value count.
393 if (old_existed && new == TOMBSTONE) {
394 SYNC_ADD(&hti->count, -1);
395 } else if (!old_existed && new != TOMBSTONE) {
396 SYNC_ADD(&hti->count, 1);
399 // Return the previous value.
400 TRACE("h0", "hti_cas: CAS succeeded; old value %p new value %p", ent_val, new);
405 static uint64_t hti_get (hti_t *hti, void *key, uint32_t key_hash) {
407 volatile entry_t *ent = hti_lookup(hti, key, key_hash, &is_empty);
409 // When hti_lookup() returns NULL it means we hit the reprobe limit while
410 // searching the table. In that case, if a copy is in progress the key
411 // might exist in the copy.
412 if (EXPECT_FALSE(ent == NULL)) {
413 if (((volatile hti_t *)hti)->next != NULL)
414 return hti_get(hti->next, key, key_hash); // recursive tail-call
415 return DOES_NOT_EXIST;
419 return DOES_NOT_EXIST;
421 // If the entry is being copied, finish the copy and retry on the next table.
422 uint64_t ent_val = ent->val;
423 if (EXPECT_FALSE(IS_TAGGED(ent_val))) {
424 if (EXPECT_FALSE(ent_val != COPIED_VALUE)) {
425 int did_copy = hti_copy_entry(hti, ent, key_hash, ((volatile hti_t *)hti)->next);
427 SYNC_ADD(&hti->num_entries_copied, 1);
430 return hti_get(((volatile hti_t *)hti)->next, key, key_hash); // tail-call
433 return (ent_val == TOMBSTONE) ? DOES_NOT_EXIST : ent_val;
437 uint64_t ht_get (hashtable_t *ht, void *key) {
438 uint32_t hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash(key);
439 return hti_get(ht->hti, key, hash);
443 uint64_t ht_cas (hashtable_t *ht, void *key, uint64_t expected_val, uint64_t new_val) {
445 TRACE("h2", "ht_cas: key %p ht %p", key, ht);
446 TRACE("h2", "ht_cas: expected val %p new val %p", expected_val, new_val);
447 assert(key != DOES_NOT_EXIST);
448 assert(!IS_TAGGED(new_val) && new_val != DOES_NOT_EXIST && new_val != TOMBSTONE);
450 hti_t *hti = ht->hti;
452 // Help with an ongoing copy.
453 if (EXPECT_FALSE(hti->next != NULL)) {
454 volatile entry_t *ent;
459 TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<<hti->scale);
460 // Panic if we've been around the array twice and still haven't finished the copy.
461 int panic = (x >= (1 << (hti->scale + 1)));
463 limit = ENTRIES_PER_COPY_CHUNK;
465 // Reserve some entries for this thread to copy. There is a race condition here because the
466 // fetch and add isn't atomic, but that is ok.
467 hti->scan = x + ENTRIES_PER_COPY_CHUNK;
469 // <hti->scan> might be larger than the size of the table, if some thread stalls while
470 // copying. In that case we just wrap around to the begining and make another pass through
472 ent = hti->table + (x & MASK(hti->scale));
474 TRACE("h1", "ht_cas: help copy panic", 0, 0);
475 // scan the whole table
476 limit = (1 << hti->scale);
481 for (int i = 0; i < limit; ++i) {
482 num_copied += hti_copy_entry(hti, ent++, 0, hti->next);
483 assert(ent <= hti->table + (1 << hti->scale));
485 if (num_copied != 0) {
486 SYNC_ADD(&hti->num_entries_copied, num_copied);
489 // Dispose of fully copied tables.
490 if (hti->num_entries_copied == (1 << hti->scale) || panic) {
492 if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) {
499 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash(key);
500 while ((old_val = hti_cas(hti, key, key_hash, expected_val, new_val)) == COPIED_VALUE) {
505 return old_val == TOMBSTONE ? DOES_NOT_EXIST : old_val;
508 // Remove the value in <ht> associated with <key>. Returns the value removed, or DOES_NOT_EXIST if there was
509 // no value for that key.
510 uint64_t ht_remove (hashtable_t *ht, void *key) {
511 hti_t *hti = ht->hti;
513 uint32_t key_hash = (ht->key_type == NULL) ? murmur32_8b((uint64_t)key) : ht->key_type->hash(key);
515 val = hti_cas(hti, key, key_hash, CAS_EXPECT_WHATEVER, TOMBSTONE);
516 if (val != COPIED_VALUE)
517 return val == TOMBSTONE ? DOES_NOT_EXIST : val;
524 // Returns the number of key-values pairs in <ht>
525 uint64_t ht_count (hashtable_t *ht) {
526 hti_t *hti = ht->hti;
535 // Allocate and initialize a new hash table.
536 hashtable_t *ht_alloc (const datatype_t *key_type) {
537 hashtable_t *ht = nbd_malloc(sizeof(hashtable_t));
538 ht->key_type = key_type;
539 ht->hti = (hti_t *)hti_alloc(ht, MIN_SCALE);
543 // Free <ht> and its internal structures.
544 void ht_free (hashtable_t *ht) {
545 hti_t *hti = ht->hti;
547 for (uint32_t i = 0; i < (1 << hti->scale); ++i) {
548 assert(hti->table[i].val == COPIED_VALUE || !IS_TAGGED(hti->table[i].val));
549 if (ht->key_type != NULL && hti->table[i].key != DOES_NOT_EXIST) {
550 nbd_free(GET_PTR(hti->table[i].key));
553 hti_t *next = hti->next;
560 void ht_print (hashtable_t *ht) {
561 hti_t *hti = ht->hti;
563 printf("hti:%p scale:%u count:%d copied:%d\n", hti, hti->scale, hti->count, hti->num_entries_copied);
564 for (int i = 0; i < (1 << hti->scale); ++i) {
565 volatile entry_t *ent = hti->table + i;
566 printf("[0x%x] %p:%p\n", i, (void *)ent->key, (void *)ent->val);