2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
5 * Implementation of the lock-free skiplist data-structure created by Maurice Herlihy, Yossi Lev,
6 * and Nir Shavit. See Herlihy's and Shivit's book "The Art of Multiprocessor Programming".
7 * http://www.amazon.com/Art-Multiprocessor-Programming-Maurice-Herlihy/dp/0123705916/
9 * See also Kir Fraser's dissertation "Practical Lock Freedom".
10 * www.cl.cam.ac.uk/techreports/UCAM-CL-TR-579.pdf
12 * I've generalized the data structure to support update operations like set() and CAS() in addition to
13 * the normal add() and remove() operations.
15 * Warning: This code is written for the x86 memory-model. The algorithim depends on certain stores
16 * and loads being ordered. This code won't work correctly on platforms with weaker memory models if
17 * you don't add memory barriers in the right places.
28 // Setting MAX_LEVEL to 0 essentially makes this data structure the Harris-Michael lock-free list (in list.c).
44 const datatype_t *key_type;
47 static int random_level (void) {
48 unsigned r = nbd_rand();
52 r |= 1 << (MAX_LEVEL+1);
54 int n = __builtin_ctz(r)-1;
55 assert(n <= MAX_LEVEL);
59 static node_t *node_alloc (int level, map_key_t key, map_val_t val) {
60 assert(level >= 0 && level <= MAX_LEVEL);
61 size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *);
62 node_t *item = (node_t *)nbd_malloc(sz);
66 item->top_level = level;
70 skiplist_t *sl_alloc (const datatype_t *key_type) {
71 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
72 sl->key_type = key_type;
73 sl->head = node_alloc(MAX_LEVEL, 0, 0);
74 memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
78 void sl_free (skiplist_t *sl) {
79 node_t *item = (node_t *)(size_t)sl->head->next[0];
81 node_t *next = (node_t *)(size_t)STRIP_TAG(item->next[0], TAG1);
82 if (sl->key_type != NULL) {
83 nbd_free((void *)(size_t)item->key);
90 uint64_t sl_count (skiplist_t *sl) {
92 node_t *item = (node_t *)(size_t)sl->head->next[0];
94 if (!IS_TAGGED(item->next[0], TAG1)) {
97 item = (node_t *)(size_t)STRIP_TAG(item->next[0], TAG1);
102 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, int help_remove) {
103 node_t *pred = sl->head;
105 TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
107 int start_level = MAX_LEVEL;
109 // Optimization for small lists. No need to traverse empty higher levels.
111 while (pred->next[start_level+1] != DOES_NOT_EXIST) {
112 start_level += start_level - 1;
113 if (EXPECT_FALSE(start_level >= MAX_LEVEL)) {
114 start_level = MAX_LEVEL;
118 if (EXPECT_FALSE(start_level < n)) {
123 // Traverse the levels of <sl> from the top level to the bottom
124 for (int level = start_level; level >= 0; --level) {
125 TRACE("s3", "find_preds: level %llu", level, 0);
126 uint64_t next = pred->next[level];
127 if (EXPECT_FALSE(IS_TAGGED(next, TAG1))) {
128 TRACE("s2", "find_preds: pred %p is marked for removal (next %p); retry", pred, next);
129 return find_preds(preds, succs, n, sl, key, help_remove); // retry
131 while (next != DOES_NOT_EXIST) {
132 item = (node_t *)(size_t)next;
133 next = item->next[level];
135 // A tag means an item is logically removed but not physically unlinked yet.
136 while (EXPECT_FALSE(IS_TAGGED(next, TAG1))) {
138 // Skip over logically removed items.
140 item = (node_t *)(size_t)STRIP_TAG(next, TAG1);
141 if (EXPECT_FALSE(item == NULL))
143 next = item->next[level];
144 TRACE("s3", "find_preds: skipping marked item %p (next is 0x%llx)", item, next);
148 // Unlink logically removed items.
149 TRACE("s3", "find_preds: unlinking marked item %p; next is 0x%llx", item, next);
150 uint64_t other = SYNC_CAS(&pred->next[level], (uint64_t)(size_t)item, STRIP_TAG(next, TAG1));
151 if (other == (uint64_t)(size_t)item) {
152 item = (node_t *)(size_t)STRIP_TAG(next, TAG1);
153 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
154 TRACE("s3", "find_preds: now the current item is %p next is 0x%llx", item, next);
156 // The thread that completes the unlink should free the memory.
158 if (sl->key_type != NULL) {
159 nbd_defer_free((void *)(size_t)((node_t *)(size_t)other)->key);
161 nbd_defer_free(((node_t *)(size_t)other));
164 TRACE("s3", "find_preds: lost race to unlink item %p from pred %p", item, pred);
165 TRACE("s3", "find_preds: pred's link changed to %p", other, 0);
166 if (IS_TAGGED(other, TAG1))
167 return find_preds(preds, succs, n, sl, key, help_remove); // retry
168 item = (node_t *)(size_t)other;
169 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
173 if (EXPECT_FALSE(item == NULL))
176 TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next);
177 TRACE("s4", "find_preds: key %p val %p", STRIP_TAG(item->key, TAG1), item->val);
179 if (EXPECT_TRUE(sl->key_type == NULL)) {
180 d = (uint64_t)item->key - (uint64_t)key;
182 d = sl->key_type->cmp((void *)(size_t)item->key, (void *)(size_t)key);
186 TRACE("s4", "find_preds: found pred %p item %p", pred, item);
191 item = (node_t *)(size_t)next;
194 // The cast to unsigned is for the case when n is -1.
195 if ((unsigned)level <= (unsigned)n) {
205 // fill in empty levels
206 if (n == -1 && item != NULL) {
207 assert(item->top_level <= MAX_LEVEL);
208 for (int level = start_level + 1; level <= item->top_level; ++level) {
209 preds[level] = sl->head;
214 TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred);
217 TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key, pred);
221 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
222 map_val_t sl_lookup (skiplist_t *sl, map_key_t key) {
223 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl);
224 node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE);
226 // If we found an <item> matching the <key> return its value.
228 map_val_t val = item->val;
229 if (val != DOES_NOT_EXIST) {
230 TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val);
235 TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0);
236 return DOES_NOT_EXIST;
239 map_key_t sl_min_key (skiplist_t *sl) {
240 node_t *item = (node_t *)(size_t)sl->head->next[0];
241 while (item != NULL) {
242 uint64_t next = item->next[0];
243 if (!IS_TAGGED(next, TAG1))
245 item = (node_t *)(size_t)STRIP_TAG(next, TAG1);
247 return DOES_NOT_EXIST;
250 map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) {
251 TRACE("s1", "sl_cas: key %p skiplist %p", key, sl);
252 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
253 ASSERT((int64_t)new_val > 0);
255 node_t *preds[MAX_LEVEL+1];
256 node_t *nexts[MAX_LEVEL+1];
257 node_t *new_item = NULL;
258 int n = random_level();
260 node_t *old_item = find_preds(preds, nexts, n, sl, key, TRUE);
261 if (old_item == NULL) {
263 // There was not an item in the skiplist that matches the key.
264 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
265 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
266 return DOES_NOT_EXIST; // failure
269 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
271 // First insert <new_item> into the bottom level.
272 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]);
273 map_key_t new_key = (sl->key_type == NULL)
275 : (map_key_t)(size_t)sl->key_type->clone((void *)(size_t)key);
276 new_item = node_alloc(n, new_key, new_val);
277 node_t *pred = preds[0];
278 uint64_t next = new_item->next[0] = (uint64_t)(size_t)nexts[0];
279 for (int level = 1; level <= new_item->top_level; ++level) {
280 new_item->next[level] = (uint64_t)(size_t)nexts[level];
282 uint64_t other = SYNC_CAS(&pred->next[0], next, (uint64_t)(size_t)new_item);
284 TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0);
287 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
288 if (sl->key_type != NULL) {
289 nbd_free((void *)(size_t)new_key);
295 // Found an item in the skiplist that matches the key.
296 map_val_t old_item_val = old_item->val;
298 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
299 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
300 TRACE("s2", "sl_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
304 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
305 TRACE("s1", "sl_cas: found an item %p in the skiplist that matched the key. the expectation was "
306 "not met, the skiplist was not changed", old_item, old_item_val);
307 return old_item_val; // failure
310 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
311 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
312 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
313 // fails it means another thread either removed the node or updated its value.
314 map_val_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
315 if (ret_val == old_item_val) {
316 TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0);
317 return ret_val; // success
319 TRACE("s2", "sl_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
321 old_item_val = ret_val;
325 // Link <new_item> into <sl> from the bottom up.
326 for (int level = 1; level <= new_item->top_level; ++level) {
327 node_t *pred = preds[level];
328 uint64_t next = (uint64_t)(size_t)nexts[level];
330 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next);
331 uint64_t other = SYNC_CAS(&pred->next[level], next, (uint64_t)(size_t)new_item);
333 TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level);
336 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
337 find_preds(preds, nexts, new_item->top_level, sl, key, TRUE);
339 next = (uint64_t)(size_t)nexts[level];
341 // Update <new_item>'s next pointer
343 // There in no need to continue linking in the item if another thread removed it.
344 uint64_t old_next = ((volatile node_t *)new_item)->next[level];
345 if (IS_TAGGED(old_next, TAG1))
346 return DOES_NOT_EXIST; // success
348 // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
349 if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
354 return DOES_NOT_EXIST; // success
357 map_val_t sl_remove (skiplist_t *sl, map_key_t key) {
358 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl);
359 node_t *preds[MAX_LEVEL+1];
360 node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE);
362 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
363 return DOES_NOT_EXIST;
366 // Mark and unlink <item> at each level of <sl> from the top down. If multiple threads try to concurrently remove
367 // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds.
368 for (int level = item->top_level; level > 0; --level) {
370 uint64_t old_next = item->next[level];
373 old_next = SYNC_CAS(&item->next[level], next, TAG_VALUE(next, TAG1));
374 if (IS_TAGGED(old_next, TAG1)) {
375 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level %llu", item, level);
378 } while (next != old_next);
380 node_t *pred = preds[level];
381 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next, TAG1));
382 uint64_t other = SYNC_CAS(&pred->next[level], (uint64_t)(size_t)item, STRIP_TAG(next, TAG1));
383 if (other != (uint64_t)(size_t)item) {
384 TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
385 // If our former predecessor now points past us we know another thread unlinked us. Otherwise, we need
386 // to search for a new set of preds.
387 if (other == DOES_NOT_EXIST)
388 continue; // <pred> points past <item> to the end of the list; go on to the next level.
391 if (!IS_TAGGED(other, TAG1)) {
392 map_key_t other_key = ((node_t *)(size_t)other)->key;
393 if (EXPECT_TRUE(sl->key_type == NULL)) {
394 d = (uint64_t)item->key - (uint64_t)other_key;
396 d = sl->key_type->cmp((void *)(size_t)item->key, (void *)(size_t)other_key);
400 node_t *temp = find_preds(preds, NULL, level, sl, key, TRUE);
402 return DOES_NOT_EXIST; // Another thread removed the item we were targeting.
403 level++; // Redo this level.
409 uint64_t old_next = item->next[0];
412 old_next = SYNC_CAS(&item->next[0], next, TAG_VALUE(next, TAG1));
413 if (IS_TAGGED(old_next, TAG1)) {
414 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level 0", item, 0);
415 return DOES_NOT_EXIST;
417 } while (next != old_next);
418 TRACE("s1", "sl_remove: marked item %p removed at level 0", item, 0);
420 // Atomically swap out the item's value in case another thread is updating the item while we are
421 // removing it. This establishes which operation occurs first logically, the update or the remove.
422 map_val_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
423 TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0);
425 node_t *pred = preds[0];
426 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next, TAG1));
427 if (SYNC_CAS(&pred->next[0], item, STRIP_TAG(next, TAG1))) {
428 TRACE("s2", "sl_remove: unlinked item %p from the skiplist at level 0", item, 0);
429 // The thread that completes the unlink should free the memory.
430 if (sl->key_type != NULL) {
431 nbd_defer_free((void *)(size_t)item->key);
433 nbd_defer_free(item);
438 void sl_print (skiplist_t *sl) {
439 for (int level = MAX_LEVEL; level >= 0; --level) {
440 node_t *item = sl->head;
441 if (item->next[level] == DOES_NOT_EXIST)
443 printf("(%d) ", level);
446 uint64_t next = item->next[level];
447 printf("%s%p ", IS_TAGGED(next, TAG1) ? "*" : "", item);
448 item = (node_t *)(size_t)STRIP_TAG(next, TAG1);
457 node_t *item = sl->head;
460 int is_marked = IS_TAGGED(item->next[0], TAG1);
461 printf("%s%p:0x%llx ", is_marked ? "*" : "", item, (uint64_t)(size_t)item->key);
462 if (item != sl->head) {
463 printf("[%d]", item->top_level);
467 for (int level = 1; level <= item->top_level; ++level) {
468 node_t *next = (node_t *)(size_t)STRIP_TAG(item->next[level], TAG1);
469 is_marked = IS_TAGGED(item->next[0], TAG1);
470 printf(" %p%s", next, is_marked ? "*" : "");
471 if (item == sl->head && item->next[level] == DOES_NOT_EXIST)
476 item = (node_t *)(size_t)STRIP_TAG(item->next[0], TAG1);
484 sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) {
485 sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t));
486 find_preds(NULL, &iter->next, 0, sl, key, FALSE);
490 map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) {
492 node_t *item = iter->next;
493 while (item != NULL && IS_TAGGED(item->next[0], TAG1)) {
494 item = (node_t *)(size_t)STRIP_TAG(item->next[0], TAG1);
498 return DOES_NOT_EXIST;
500 iter->next = (node_t *)(size_t)STRIP_TAG(item->next[0], TAG1);
501 if (key_ptr != NULL) {
502 *key_ptr = item->key;
507 void sl_iter_free (sl_iter_t *iter) {