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) ((nstring_t *)((x) & MASK(48))) // low-order 48 bits is a pointer to a nstring_t
23 typedef struct ht_entry {
24 uint64_t key; // ptr to nstring_t
29 volatile entry_t *table;
30 hashtable_t *ht; // parent ht;
32 struct hti *next_free;
35 int count; // TODO: make these counters distributed
36 int num_entries_copied;
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 (hashtable_i_t *ht1, volatile entry_t *e, uint32_t e_key_hash, hashtable_i_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);
70 // A key is made up of two parts. The 48 low-order bits are a pointer to a null terminated string.
71 // The high-order 16 bits are taken from the hash of that string. The bits from the hash are used
72 // as a quick check to rule out non-equal keys without doing a complete string compare.
73 static inline int ht_key_equals (uint64_t a, uint32_t b_hash, const char *b_value, uint32_t b_len) {
74 if ((b_hash >> 16) != (a >> 48)) // high-order 16 bits are from the hash value
76 return ns_cmp_raw(GET_PTR(a), b_value, b_len) == 0;
79 // Lookup <key> in <hti>.
81 // Return the entry that <key> is in, or if <key> isn't in <hti> return the entry that it would be
82 // in if it were inserted into <hti>. If there is no room for <key> in <hti> then return NULL, to
83 // indicate that the caller should look in <hti->next>.
85 // Record if the entry being returned is empty. Otherwise the caller will have to waste time with
86 // ht_key_equals() to confirm that it did not lose a race to fill an empty entry.
87 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) {
88 TRACE("h2", "hti_lookup(key %p in hti %p)", key_data, hti);
91 // Probe one cache line at a time
92 int ndx = key_hash & MASK(hti->scale); // the first entry to search
93 for (int i = 0; i < hti->max_probe; ++i) {
95 // The start of the bucket is the first entry in the cache line.
96 volatile entry_t *bucket = hti->table + (ndx & ~(ENTRIES_PER_BUCKET-1));
98 // Start searching at the indexed entry. Then loop around to the begining of the cache line.
99 for (int j = 0; j < ENTRIES_PER_BUCKET; ++j) {
100 volatile entry_t *e = bucket + ((ndx + j) & (ENTRIES_PER_BUCKET-1));
102 uint64_t e_key = e->key;
103 if (e_key == DOES_NOT_EXIST) {
104 TRACE("h1", "hti_lookup: entry %p for key \"%s\" is empty", e, GET_PTR(e_key)->data);
105 *is_empty = 1; // indicate an empty so the caller avoids an expensive ht_key_equals
109 if (ht_key_equals(e_key, key_hash, key_data, key_len)) {
110 TRACE("h1", "hti_lookup: entry %p key \"%s\"", e, GET_PTR(e_key)->data);
111 TRACE("h2", "hti_lookup: entry key len %llu, value %p", GET_PTR(e_key)->len, e->value);
116 ndx = get_next_ndx(ndx, key_hash, hti->scale);
119 // maximum number of probes exceeded
120 TRACE("h1", "hti_lookup: maximum number of probes exceeded returning 0x0", 0, 0);
124 // Allocate and initialize a hashtable_i_t with 2^<scale> entries.
125 static hashtable_i_t *hti_alloc (hashtable_t *parent, int scale) {
126 // Include enough slop to align the actual table on a cache line boundry
127 size_t n = sizeof(hashtable_i_t)
128 + sizeof(entry_t) * (1 << scale)
129 + (CACHE_LINE_SIZE - 1);
130 hashtable_i_t *hti = (hashtable_i_t *)calloc(n, 1);
132 // Align the table of hash entries on a cache line boundry.
133 hti->table = (entry_t *)(((uint64_t)hti + sizeof(hashtable_i_t) + (CACHE_LINE_SIZE-1))
134 & ~(CACHE_LINE_SIZE-1));
138 // When searching for a key probe a maximum of 1/4 of the buckets up to 1000 buckets.
139 hti->max_probe = ((1 << (hti->scale - 2)) / ENTRIES_PER_BUCKET) + 2;
140 if (hti->max_probe > MAX_BUCKETS_TO_PROBE) {
141 hti->max_probe = MAX_BUCKETS_TO_PROBE;
146 assert(hti->scale >= MIN_SCALE && hti->scale < 63); // size must be a power of 2
147 assert(sizeof(entry_t) * ENTRIES_PER_BUCKET % CACHE_LINE_SIZE == 0); // divisible into cache
148 assert((size_t)hti->table % CACHE_LINE_SIZE == 0); // cache aligned
153 // Called when <hti> runs out of room for new keys.
155 // Initiates a copy by creating a larger hashtable_i_t and installing it in <hti->next>.
156 static void hti_start_copy (hashtable_i_t *hti) {
157 TRACE("h0", "hti_start_copy(hti %p scale %llu)", hti, hti->scale);
159 // heuristics to determine the size of the new table
160 uint64_t count = ht_count(hti->ht);
161 unsigned int new_scale = hti->scale;
162 new_scale += (count > (1 << (new_scale - 2))); // double size if more than 1/4 full
163 new_scale += (count > (1 << (new_scale - 2))); // double size again if more than 1/2 full
165 // Allocate the new table and attempt to install it.
166 hashtable_i_t *next = hti_alloc(hti->ht, new_scale);
167 hashtable_i_t *old_next = SYNC_CAS(&hti->next, NULL, next);
168 if (old_next != NULL) {
169 // Another thread beat us to it.
170 TRACE("h0", "hti_start_copy: lost race to install new hti; found %p", old_next, 0);
174 TRACE("h0", "hti_start_copy: new hti %p scale %llu", next, next->scale);
177 // Copy the key and value stored in <ht1_e> (which must be an entry in <ht1>) to <ht2>.
179 // Return 1 unless <ht1_e> is already copied (then return 0), so the caller can account for the total
180 // number of entries left to copy.
181 static int hti_copy_entry (hashtable_i_t *ht1, volatile entry_t *ht1_e, uint32_t key_hash,
182 hashtable_i_t *ht2) {
183 TRACE("h2", "hti_copy_entry: entry %p to table %p", ht1_e, ht2);
187 assert(ht1_e >= ht1->table && ht1_e < ht1->table + (1 << ht1->scale));
188 assert(key_hash == 0 || (key_hash >> 16) == (ht1_e->key >> 48));
190 uint64_t ht1_e_value = ht1_e->value;
191 if (EXPECT_FALSE(ht1_e_value == COPIED_VALUE)) {
192 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_e, ht2);
193 return FALSE; // already copied
196 // Kill empty entries.
197 if (EXPECT_FALSE(ht1_e_value == DOES_NOT_EXIST)) {
198 uint64_t ht1_e_value = SYNC_CAS(&ht1_e->value, DOES_NOT_EXIST, COPIED_VALUE);
199 if (ht1_e_value == DOES_NOT_EXIST) {
200 TRACE("h1", "hti_copy_entry: empty entry %p killed", ht1_e, 0);
203 if (ht1_e_value == COPIED_VALUE) {
204 TRACE("h0", "hti_copy_entry: lost race to kill empty entry %p", ht1_e, 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 now"
208 "in use and should be copied", ht1_e, 0);
211 // Tag the value in the old entry to indicate a copy is in progress.
212 ht1_e_value = SYNC_FETCH_AND_OR(&ht1_e->value, TAG_VALUE(0));
213 TRACE("h2", "hti_copy_entry: tagged the value %p in old entry %p", ht1_e_value, ht1_e);
214 if (ht1_e_value == COPIED_VALUE) {
215 TRACE("h1", "hti_copy_entry: entry %p already copied to table %p", ht1_e, ht2);
216 return FALSE; // <value> was already copied by another thread.
219 // The old table's deleted entries don't need to be copied to the new table, but their keys need
221 assert(COPIED_VALUE == TAG_VALUE(TOMBSTONE));
222 if (ht1_e_value == TOMBSTONE) {
223 TRACE("h1", "hti_copy_entry: entry %p old value was deleted, now freeing key %p", ht1_e, GET_PTR(ht1_e->key));
224 nbd_defer_free(GET_PTR(ht1_e->key));
228 // Install the key in the new table.
229 uint64_t key = ht1_e->key;
230 nstring_t *key_string = GET_PTR(key);
231 uint64_t value = STRIP_TAG(ht1_e_value);
233 // We use 0 to indicate that <key_hash> isn't initiallized. Occasionally the <key_hash> will
234 // really be 0 and we will waste time recomputing it. That is rare enough that it is OK.
236 key_hash = murmur32(key_string->data, key_string->len);
240 volatile entry_t *ht2_e = hti_lookup(ht2, key_hash, key_string->data, key_string->len, &is_empty);
241 TRACE("h0", "hti_copy_entry: copy entry %p to entry %p", ht1_e, ht2_e);
243 // it is possible that there is not any room in the new table either
244 if (EXPECT_FALSE(ht2_e == NULL)) {
245 TRACE("h0", "hti_copy_entry: no room in table %p copy to next table %p", ht2, ht2->next);
246 if (ht2->next == NULL) {
247 hti_start_copy(ht2); // initiate nested copy, if not already started
249 return hti_copy_entry(ht1, ht1_e, key_hash, ht2->next); // recursive tail-call
252 // a tagged entry returned from hti_lookup() means it is either empty or has a new key
254 uint64_t old_ht2_e_key = SYNC_CAS(&ht2_e->key, DOES_NOT_EXIST, key);
255 if (old_ht2_e_key != DOES_NOT_EXIST) {
256 TRACE("h0", "hti_copy_entry: lost race to CAS key %p into new entry; found %p",
258 return hti_copy_entry(ht1, ht1_e, key_hash, ht2); // recursive tail-call
262 // Copy the value to the entry in the new table.
263 uint64_t old_ht2_e_value = SYNC_CAS(&ht2_e->value, DOES_NOT_EXIST, value);
265 // If there is a nested copy in progress, we might have installed the key into a dead entry.
266 if (old_ht2_e_value == COPIED_VALUE) {
267 TRACE("h0", "hti_copy_entry: nested copy in progress; copy %p to next table %p", ht2_e, ht2->next);
268 return hti_copy_entry(ht1, ht1_e, key_hash, ht2->next); // recursive tail-call
271 // Mark the old entry as dead.
272 ht1_e->value = COPIED_VALUE;
274 // Update the count if we were the one that completed the copy.
275 if (old_ht2_e_value == DOES_NOT_EXIST) {
276 TRACE("h0", "hti_copy_entry: key \"%s\" value %p copied to new entry", key_string->data, value);
277 SYNC_ADD(&ht1->count, -1);
278 SYNC_ADD(&ht2->count, 1);
282 TRACE("h0", "hti_copy_entry: lost race to install value %p in new entry; found value %p",
283 value, old_ht2_e_value);
284 return FALSE; // another thread completed the copy
287 // Compare <expected> with the existing value associated with <key>. If the values match then
288 // replace the existing value with <new>. If <new> is TOMBSTONE, delete the value associated with
289 // the key by replacing it with a TOMBSTONE.
291 // Return the previous value associated with <key>, or DOES_NOT_EXIST if <key> is not in the table
292 // or associated with a TOMBSTONE. If a copy is in progress and <key> has been copied to the next
293 // table then return COPIED_VALUE.
295 // NOTE: the returned value matches <expected> iff the set succeeds
297 // Certain values of <expected> have special meaning. If <expected> is CAS_EXPECT_EXISTS then any
298 // real value matches (i.e. not a TOMBSTONE or DOES_NOT_EXIST) as long as <key> is in the table. If
299 // <expected> is CAS_EXPECT_WHATEVER then skip the test entirely.
301 static uint64_t hti_cas (hashtable_i_t *hti, uint32_t key_hash, const char *key_data, uint32_t key_len,
302 uint64_t expected, uint64_t new) {
303 TRACE("h1", "hti_cas: hti %p key %p", hti, key_data);
304 TRACE("h1", "hti_cas: value %p expect %p", new, expected);
306 assert(new != DOES_NOT_EXIST && !IS_TAGGED(new));
310 volatile entry_t *e = hti_lookup(hti, key_hash, key_data, key_len, &is_empty);
312 // There is no room for <key>, grow the table and try again.
314 if (hti->next == NULL) {
320 // Install <key> in the table if it doesn't exist.
322 TRACE("h0", "hti_cas: entry %p is empty", e, 0);
323 if (expected != CAS_EXPECT_WHATEVER && expected != CAS_EXPECT_DOES_NOT_EXIST)
324 return DOES_NOT_EXIST;
326 // No need to do anything, <key> is already deleted.
327 if (new == TOMBSTONE)
328 return DOES_NOT_EXIST;
331 nstring_t *key = ns_alloc(key_data, key_len);
333 // Combine <key> pointer with bits from its hash, CAS it into the table.
334 uint64_t temp = ((uint64_t)(key_hash >> 16) << 48) | (uint64_t)key;
335 uint64_t e_key = SYNC_CAS(&e->key, DOES_NOT_EXIST, temp);
337 // Retry if another thread stole the entry out from under us.
338 if (e_key != DOES_NOT_EXIST) {
339 TRACE("h0", "hti_cas: lost race to install key %p in entry %p", key, e);
340 TRACE("h0", "hti_cas: found %p instead of NULL", GET_PTR(e_key), 0);
342 return hti_cas(hti, key_hash, key_data, key_len, expected, new); // tail-call
344 TRACE("h2", "hti_cas: installed key %p in entry %p", key, e);
347 TRACE("h0", "hti_cas: entry for key \"%s\" is %p", GET_PTR(e->key)->data, e);
349 // If the entry is in the middle of a copy, the copy must be completed first.
350 uint64_t e_value = e->value;
351 if (EXPECT_FALSE(IS_TAGGED(e_value))) {
352 if (e_value != COPIED_VALUE) {
353 int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hashtable_i_t *)hti)->next);
355 SYNC_ADD(&hti->num_entries_copied, 1);
357 TRACE("h0", "hti_cas: value in the middle of a copy, copy completed by %s",
358 (did_copy ? "self" : "other"), 0);
360 TRACE("h0", "hti_cas: value copied to next table, retry on next table", 0, 0);
364 // Fail if the old value is not consistent with the caller's expectation.
365 int old_existed = (e_value != TOMBSTONE && e_value != DOES_NOT_EXIST);
366 if (EXPECT_FALSE(expected != CAS_EXPECT_WHATEVER && expected != e_value)) {
367 if (EXPECT_FALSE(expected != (old_existed ? CAS_EXPECT_EXISTS : CAS_EXPECT_DOES_NOT_EXIST))) {
368 TRACE("h1", "hti_cas: value %p expected by caller not found; found value %p",
374 // No need to update if value is unchanged.
375 if ((new == TOMBSTONE && !old_existed) || e_value == new) {
376 TRACE("h1", "hti_cas: old value and new value were the same", 0, 0);
380 // CAS the value into the entry. Retry if it fails.
381 uint64_t v = SYNC_CAS(&e->value, e_value, new);
382 if (EXPECT_FALSE(v != e_value)) {
383 TRACE("h0", "hti_cas: value CAS failed; expected %p found %p", e_value, v);
384 return hti_cas(hti, key_hash, key_data, key_len, expected, new); // recursive tail-call
387 // The set succeeded. Adjust the value count.
388 if (old_existed && new == TOMBSTONE) {
389 SYNC_ADD(&hti->count, -1);
390 } else if (!old_existed && new != TOMBSTONE) {
391 SYNC_ADD(&hti->count, 1);
394 // Return the previous value.
395 TRACE("h0", "hti_cas: CAS succeeded; old value %p new value %p", e_value, new);
400 static uint64_t hti_get (hashtable_i_t *hti, uint32_t key_hash, const char *key_data, uint32_t key_len) {
404 volatile entry_t *e = hti_lookup(hti, key_hash, key_data, key_len, &is_empty);
406 // When hti_lookup() returns NULL it means we hit the reprobe limit while
407 // searching the table. In that case, if a copy is in progress the key
408 // might exist in the copy.
409 if (EXPECT_FALSE(e == NULL)) {
410 if (((volatile hashtable_i_t *)hti)->next != NULL)
411 return hti_get(hti->next, key_hash, key_data, key_len); // recursive tail-call
412 return DOES_NOT_EXIST;
416 return DOES_NOT_EXIST;
418 // If the entry is being copied, finish the copy and retry on the next table.
419 uint64_t e_value = e->value;
420 if (EXPECT_FALSE(IS_TAGGED(e_value))) {
421 if (EXPECT_FALSE(e_value != COPIED_VALUE)) {
422 int did_copy = hti_copy_entry(hti, e, key_hash, ((volatile hashtable_i_t *)hti)->next);
424 SYNC_ADD(&hti->num_entries_copied, 1);
427 return hti_get(((volatile hashtable_i_t *)hti)->next, key_hash, key_data, key_len); // tail-call
430 return (e_value == TOMBSTONE) ? DOES_NOT_EXIST : e_value;
434 uint64_t ht_get (hashtable_t *ht, const char *key_data, uint32_t key_len) {
435 return hti_get(ht->hti, murmur32(key_data, key_len), key_data, key_len);
439 uint64_t ht_cas (hashtable_t *ht, const char *key_data, uint32_t key_len, uint64_t expected_val, uint64_t new_val) {
441 TRACE("h2", "ht_cas: key %p len %u", key_data, key_len);
442 TRACE("h2", "ht_cas: expected val %p new val %p", expected_val, new_val);
444 assert(!IS_TAGGED(new_val) && new_val != DOES_NOT_EXIST);
446 hashtable_i_t *hti = ht->hti;
448 // Help with an ongoing copy.
449 if (EXPECT_FALSE(hti->next != NULL)) {
455 TRACE("h1", "ht_cas: help copy. scan is %llu, size is %llu", x, 1<<hti->scale);
456 // Panic if we've been around the array twice and still haven't finished the copy.
457 int panic = (x >= (1 << (hti->scale + 1)));
459 limit = ENTRIES_PER_COPY_CHUNK;
461 // Reserve some entries for this thread to copy. There is a race condition here because the
462 // fetch and add isn't atomic, but that is ok.
463 hti->scan = x + ENTRIES_PER_COPY_CHUNK;
465 // <hti->scan> might be larger than the size of the table, if some thread stalls while
466 // copying. In that case we just wrap around to the begining and make another pass through
468 e = hti->table + (x & MASK(hti->scale));
470 TRACE("h1", "ht_cas: help copy panic", 0, 0);
471 // scan the whole table
472 limit = (1 << hti->scale);
477 for (int i = 0; i < limit; ++i) {
478 num_copied += hti_copy_entry(hti, e++, 0, hti->next);
479 assert(e <= hti->table + (1 << hti->scale));
481 if (num_copied != 0) {
482 SYNC_ADD(&hti->num_entries_copied, num_copied);
485 // Dispose of fully copied tables.
486 if (hti->num_entries_copied == (1 << hti->scale) || panic) {
488 if (SYNC_CAS(&ht->hti, hti, hti->next) == hti) {
495 uint32_t key_hash = murmur32(key_data, key_len);
496 while ((old_val = hti_cas(hti, key_hash, key_data, key_len, expected_val, new_val))
502 return old_val == TOMBSTONE ? DOES_NOT_EXIST : old_val;
505 // Remove the value in <ht> associated with <key_data>. Returns the value removed, or
506 // DOES_NOT_EXIST if there was no value for that key.
507 uint64_t ht_remove (hashtable_t *ht, const char *key_data, uint32_t key_len) {
508 hashtable_i_t *hti = ht->hti;
510 uint32_t key_hash = murmur32(key_data, key_len);
512 val = hti_cas(hti, key_hash, key_data, key_len, CAS_EXPECT_WHATEVER, TOMBSTONE);
513 if (val != COPIED_VALUE)
514 return val == TOMBSTONE ? DOES_NOT_EXIST : val;
521 // Returns the number of key-values pairs in <ht>
522 uint64_t ht_count (hashtable_t *ht) {
523 hashtable_i_t *hti = ht->hti;
532 // Allocate and initialize a new hash table.
533 hashtable_t *ht_alloc (void) {
534 hashtable_t *ht = nbd_malloc(sizeof(hashtable_t));
535 ht->hti = (hashtable_i_t *)hti_alloc(ht, MIN_SCALE);
539 // Free <ht> and its internal structures.
540 void ht_free (hashtable_t *ht) {
541 hashtable_i_t *hti = ht->hti;
543 for (uint32_t i = 0; i < (1 << hti->scale); ++i) {
544 assert(hti->table[i].value == COPIED_VALUE || !IS_TAGGED(hti->table[i].value));
545 if (hti->table[i].key != DOES_NOT_EXIST) {
546 nbd_free(GET_PTR(hti->table[i].key));
549 hashtable_i_t *next = hti->next;
556 void ht_print (hashtable_t *ht) {