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).
31 typedef struct sl_iter node_t;
42 const datatype_t *key_type;
45 static int random_level (void) {
46 unsigned r = nbd_rand();
50 r |= 1 << (MAX_LEVEL+1);
52 int n = __builtin_ctz(r)-1;
53 assert(n <= MAX_LEVEL);
57 static node_t *node_alloc (int level, map_key_t key, map_val_t val) {
58 assert(level >= 0 && level <= MAX_LEVEL);
59 size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *);
60 node_t *item = (node_t *)nbd_malloc(sz);
64 item->top_level = level;
68 skiplist_t *sl_alloc (const datatype_t *key_type) {
69 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
70 sl->key_type = key_type;
71 sl->head = node_alloc(MAX_LEVEL, NULL, 0);
72 memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
76 void sl_free (skiplist_t *sl) {
77 node_t *item = sl->head->next[0];
79 node_t *next = (node_t *)STRIP_TAG(item->next[0], TAG1);
85 uint64_t sl_count (skiplist_t *sl) {
87 node_t *item = sl->head->next[0];
89 if (!IS_TAGGED(item->next[0], TAG1)) {
92 item = (node_t *)STRIP_TAG(item->next[0], TAG1);
97 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, int help_remove) {
98 node_t *pred = sl->head;
100 TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
102 int start_level = MAX_LEVEL;
104 // Optimization for small lists. No need to traverse empty higher levels.
106 while (pred->next[start_level+1] != NULL) {
107 start_level += start_level - 1;
108 if (EXPECT_FALSE(start_level >= MAX_LEVEL)) {
109 start_level = MAX_LEVEL;
113 if (EXPECT_FALSE(start_level < n)) {
118 // Traverse the levels of <sl> from the top level to the bottom
119 for (int level = start_level; level >= 0; --level) {
120 TRACE("s3", "find_preds: level %llu", level, 0);
121 item = pred->next[level];
122 if (EXPECT_FALSE(IS_TAGGED(item, TAG1))) {
123 TRACE("s2", "find_preds: pred %p is marked for removal (item %p); retry", pred, item);
124 return find_preds(preds, succs, n, sl, key, help_remove); // retry
126 while (item != NULL) {
127 node_t *next = item->next[level];
129 // A tag means an item is logically removed but not physically unlinked yet.
130 while (EXPECT_FALSE(IS_TAGGED(next, TAG1))) {
132 // Skip over logically removed items.
134 item = (node_t *)STRIP_TAG(item->next, TAG1);
135 if (EXPECT_FALSE(item == NULL))
137 TRACE("s3", "find_preds: skipping marked item %p (next is %p)", item, next);
138 next = item->next[level];
142 // Unlink logically removed items.
144 TRACE("s3", "find_preds: unlinking marked item %p; next is %p", item, next);
145 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next, TAG1))) == item) {
146 item = (node_t *)STRIP_TAG(next, TAG1);
147 if (EXPECT_FALSE(item == NULL))
149 next = item->next[level];
150 TRACE("s3", "find_preds: now the current item is %p next is %p", item, next);
152 // The thread that completes the unlink should free the memory.
154 if (sl->key_type != NULL) {
155 nbd_defer_free((void*)other->key);
157 nbd_defer_free(other);
160 TRACE("s3", "find_preds: lost race to unlink item %p from pred %p", item, pred);
161 TRACE("s3", "find_preds: pred's link changed to %p", other, 0);
162 if (IS_TAGGED(other, TAG1))
163 return find_preds(preds, succs, n, sl, key, help_remove); // retry
165 if (EXPECT_FALSE(item == NULL))
167 next = item->next[level];
171 if (EXPECT_FALSE(item == NULL))
174 TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next);
175 TRACE("s4", "find_preds: key %p val %p", STRIP_TAG(item->key, TAG1), item->val);
177 if (EXPECT_TRUE(sl->key_type == NULL)) {
178 d = (uint64_t)item->key - (uint64_t)key;
180 d = sl->key_type->cmp(item->key, key);
184 TRACE("s4", "find_preds: found pred %p item %p", pred, item);
192 // The cast to unsigned is for the case when n is -1.
193 if ((unsigned)level <= (unsigned)n) {
203 // fill in empty levels
204 if (n == -1 && item != NULL) {
205 for (int level = start_level + 1; level <= item->top_level; ++level) {
206 preds[level] = sl->head;
211 TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred);
214 TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key, pred);
218 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
219 map_val_t sl_lookup (skiplist_t *sl, map_key_t key) {
220 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl);
221 node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE);
223 // If we found an <item> matching the <key> return its value.
225 map_val_t val = item->val;
226 if (val != DOES_NOT_EXIST) {
227 TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val);
232 TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0);
233 return DOES_NOT_EXIST;
236 map_key_t sl_min_key (skiplist_t *sl) {
237 node_t *item = sl->head->next[0];
238 while (item != NULL) {
239 node_t *next = item->next[0];
240 if (!IS_TAGGED(next, TAG1))
242 item = (node_t *)STRIP_TAG(next, TAG1);
244 return DOES_NOT_EXIST;
247 map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) {
248 TRACE("s1", "sl_cas: key %p skiplist %p", key, sl);
249 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
250 ASSERT((int64_t)new_val > 0);
252 node_t *preds[MAX_LEVEL+1];
253 node_t *nexts[MAX_LEVEL+1];
254 node_t *new_item = NULL;
255 int n = random_level();
257 node_t *old_item = find_preds(preds, nexts, n, sl, key, TRUE);
258 if (old_item == NULL) {
260 // There was not an item in the skiplist that matches the key.
261 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
262 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
263 return DOES_NOT_EXIST; // failure
266 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
268 // First insert <new_item> into the bottom level.
269 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]);
270 map_key_t new_key = (sl->key_type == NULL) ? key : sl->key_type->clone(key);
271 new_item = node_alloc(n, new_key, new_val);
272 node_t *pred = preds[0];
273 node_t *next = new_item->next[0] = nexts[0];
274 for (int level = 1; level <= new_item->top_level; ++level) {
275 new_item->next[level] = nexts[level];
277 node_t *other = SYNC_CAS(&pred->next[0], next, new_item);
279 TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0);
282 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
283 if (sl->key_type != NULL) {
290 // Found an item in the skiplist that matches the key.
291 map_val_t old_item_val = old_item->val;
293 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
294 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
295 TRACE("s2", "sl_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
299 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
300 TRACE("s1", "sl_cas: found an item %p in the skiplist that matched the key. the expectation was "
301 "not met, the skiplist was not changed", old_item, old_item_val);
302 return old_item_val; // failure
305 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
306 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
307 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
308 // fails it means another thread either removed the node or updated its value.
309 map_val_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
310 if (ret_val == old_item_val) {
311 TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0);
312 return ret_val; // success
314 TRACE("s2", "sl_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
316 old_item_val = ret_val;
320 // Link <new_item> into <sl> from the bottom up.
321 for (int level = 1; level <= new_item->top_level; ++level) {
322 node_t *pred = preds[level];
323 node_t *next = nexts[level];
325 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next);
326 node_t *other = SYNC_CAS(&pred->next[level], next, new_item);
328 TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level);
331 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
332 find_preds(preds, nexts, new_item->top_level, sl, key, TRUE);
336 // Update <new_item>'s next pointer
338 // There in no need to continue linking in the item if another thread removed it.
339 node_t *old_next = ((volatile node_t *)new_item)->next[level];
340 if (IS_TAGGED(old_next, TAG1))
341 return DOES_NOT_EXIST; // success
343 // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
344 if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
349 return DOES_NOT_EXIST; // success
352 map_val_t sl_remove (skiplist_t *sl, map_key_t key) {
353 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl);
354 node_t *preds[MAX_LEVEL+1];
355 node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE);
357 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
358 return DOES_NOT_EXIST;
361 // Mark and unlink <item> at each level of <sl> from the top down. If multiple threads try to concurrently remove
362 // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds.
363 for (int level = item->top_level; level > 0; --level) {
365 node_t *old_next = item->next[level];
368 old_next = SYNC_CAS(&item->next[level], next, TAG_VALUE(next, TAG1));
369 if (IS_TAGGED(old_next, TAG1)) {
370 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level %llu", item, level);
373 } while (next != old_next);
375 node_t *pred = preds[level];
376 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next, TAG1));
377 node_t *other = NULL;
378 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next, TAG1))) != item) {
379 TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
380 // If our former predecessor now points past us we know another thread unlinked us. Otherwise, we need
381 // to search for a new set of preds.
383 continue; // <pred> points past <item> to the end of the list; go on to the next level.
386 if (!IS_TAGGED(other, TAG1)) {
387 if (EXPECT_TRUE(sl->key_type == NULL)) {
388 d = (uint64_t)item->key - (uint64_t)other->key;
390 d = sl->key_type->cmp(item->key, other->key);
394 node_t *temp = find_preds(preds, NULL, level, sl, key, TRUE);
396 return DOES_NOT_EXIST; // Another thread removed the item we were targeting.
397 level++; // Redo this level.
403 node_t *old_next = item->next[0];
406 old_next = SYNC_CAS(&item->next[0], next, TAG_VALUE(next, TAG1));
407 if (IS_TAGGED(old_next, TAG1)) {
408 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level 0", item, 0);
409 return DOES_NOT_EXIST;
411 } while (next != old_next);
412 TRACE("s1", "sl_remove: marked item %p removed at level 0", item, 0);
414 // Atomically swap out the item's value in case another thread is updating the item while we are
415 // removing it. This establishes which operation occurs first logically, the update or the remove.
416 map_val_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
417 TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0);
419 node_t *pred = preds[0];
420 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next, TAG1));
421 if (SYNC_CAS(&pred->next[0], item, STRIP_TAG(next, TAG1))) {
422 TRACE("s2", "sl_remove: unlinked item %p from the skiplist at level 0", item, 0);
423 // The thread that completes the unlink should free the memory.
424 if (sl->key_type != NULL) {
425 nbd_defer_free(item->key);
427 nbd_defer_free(item);
432 void sl_print (skiplist_t *sl) {
433 for (int level = MAX_LEVEL; level >= 0; --level) {
434 node_t *item = sl->head;
435 if (item->next[level] == NULL)
437 printf("(%d) ", level);
440 node_t *next = item->next[level];
441 printf("%s%p ", IS_TAGGED(next, TAG1) ? "*" : "", item);
442 item = (node_t *)STRIP_TAG(next, TAG1);
451 node_t *item = sl->head;
454 int is_marked = IS_TAGGED(item->next[0], TAG1);
455 printf("%s%p:%p ", is_marked ? "*" : "", item, item->key);
456 if (item != sl->head) {
457 printf("[%d]", item->top_level);
461 for (int level = 1; level <= item->top_level; ++level) {
462 node_t *next = (node_t *)STRIP_TAG(item->next[level], TAG1);
463 is_marked = IS_TAGGED(item->next[0], TAG1);
464 printf(" %p%s", next, is_marked ? "*" : "");
465 if (item == sl->head && item->next[level] == NULL)
470 item = (node_t *)STRIP_TAG(item->next[0], TAG1);
478 sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) {
479 node_t *iter = node_alloc(0, 0, 0);
480 find_preds(NULL, &iter->next[0], 0, sl, key, FALSE);
484 map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) {
486 node_t *item = iter->next[0];
487 while (item != NULL && IS_TAGGED(item->next[0], TAG1)) {
488 item = (node_t *)STRIP_TAG(item->next[0], TAG1);
491 iter->next[0] = NULL;
492 return DOES_NOT_EXIST;
494 iter->next[0] = item->next[0];
495 if (key_ptr != NULL) {
496 *key_ptr = item->key;
501 void sl_iter_free (sl_iter_t *iter) {