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 // Marking the <next> field of a node logically removes it from the list
49 static inline markable_t MARK_NODE(node_t * x) { return TAG_VALUE((markable_t)x, TAG1); }
50 static inline int HAS_MARK(markable_t x) { return (IS_TAGGED(x, TAG1) == TAG1); }
51 static inline node_t * GET_NODE(markable_t x) { assert(!HAS_MARK(x)); return (node_t *)x; }
52 static inline node_t * STRIP_MARK(markable_t x) { return ((node_t *)STRIP_TAG(x, TAG1)); }
54 #define MARK_NODE(x) TAG_VALUE((markable_t)(x), TAG1)
55 #define HAS_MARK(x) (IS_TAGGED((x), TAG1) == TAG1)
56 #define GET_NODE(x) ((node_t *)(x))
57 #define STRIP_MARK(x) ((node_t *)STRIP_TAG((x), TAG1))
60 static int random_level (void) {
61 unsigned r = nbd_rand();
65 r |= 1 << (MAX_LEVEL+1);
67 int n = __builtin_ctz(r)-1;
68 assert(n <= MAX_LEVEL);
72 static node_t *node_alloc (int level, map_key_t key, map_val_t val) {
73 assert(level >= 0 && level <= MAX_LEVEL);
74 size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *);
75 node_t *item = (node_t *)nbd_malloc(sz);
79 item->top_level = level;
83 skiplist_t *sl_alloc (const datatype_t *key_type) {
84 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
85 sl->key_type = key_type;
86 sl->head = node_alloc(MAX_LEVEL, 0, 0);
87 memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
91 void sl_free (skiplist_t *sl) {
92 node_t *item = GET_NODE(sl->head->next[0]);
94 node_t *next = STRIP_MARK(item->next[0]);
95 if (sl->key_type != NULL) {
96 nbd_free((void *)item->key);
103 size_t sl_count (skiplist_t *sl) {
105 node_t *item = GET_NODE(sl->head->next[0]);
107 if (!HAS_MARK(item->next[0])) {
110 item = STRIP_MARK(item->next[0]);
115 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, int help_remove) {
116 node_t *pred = sl->head;
118 TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
120 int start_level = MAX_LEVEL;
122 // Optimization for small lists. No need to traverse empty higher levels.
124 while (pred->next[start_level+1] != DOES_NOT_EXIST) {
125 start_level += start_level - 1;
126 if (EXPECT_FALSE(start_level >= MAX_LEVEL)) {
127 start_level = MAX_LEVEL;
131 if (EXPECT_FALSE(start_level < n)) {
136 // Traverse the levels of <sl> from the top level to the bottom
137 for (int level = start_level; level >= 0; --level) {
138 TRACE("s3", "find_preds: level %llu", level, 0);
139 markable_t next = pred->next[level];
140 if (EXPECT_FALSE(HAS_MARK(next))) {
141 TRACE("s2", "find_preds: pred %p is marked for removal (next %p); retry", pred, next);
142 return find_preds(preds, succs, n, sl, key, help_remove); // retry
144 item = GET_NODE(next);
145 while (item != NULL) {
146 next = item->next[level];
148 // A tag means an item is logically removed but not physically unlinked yet.
149 while (EXPECT_FALSE(HAS_MARK(next))) {
151 // Skip over logically removed items.
153 item = STRIP_MARK(next);
154 if (EXPECT_FALSE(item == NULL))
156 next = item->next[level];
157 TRACE("s3", "find_preds: skipping marked item %p (next is 0x%llx)", item, next);
161 // Unlink logically removed items.
162 TRACE("s3", "find_preds: unlinking marked item %p; next is 0x%llx", item, next);
163 markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next));
164 if (other == (markable_t)item) {
165 item = STRIP_MARK(next);
166 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
167 TRACE("s3", "find_preds: now the current item is %p next is 0x%llx", item, next);
169 // The thread that completes the unlink should free the memory.
171 node_t *unlinked = GET_NODE(other);
172 if (sl->key_type != NULL) {
173 nbd_defer_free((void *)unlinked->key);
175 nbd_defer_free(unlinked);
178 TRACE("s3", "find_preds: lost race to unlink item %p from pred %p", item, pred);
179 TRACE("s3", "find_preds: pred's link changed to %p", other, 0);
181 return find_preds(preds, succs, n, sl, key, help_remove); // retry
182 item = GET_NODE(other);
183 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
187 if (EXPECT_FALSE(item == NULL))
190 TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next);
191 TRACE("s4", "find_preds: key %p val %p", STRIP_MARK(item->key), item->val);
193 if (EXPECT_TRUE(sl->key_type == NULL)) {
196 d = sl->key_type->cmp((void *)item->key, (void *)key);
200 TRACE("s4", "find_preds: found pred %p item %p", pred, item);
205 item = GET_NODE(next);
208 // The cast to unsigned is for the case when n is -1.
209 if ((unsigned)level <= (unsigned)n) {
219 // fill in empty levels
220 if (n == -1 && item != NULL) {
221 assert(item->top_level <= MAX_LEVEL);
222 for (int level = start_level + 1; level <= item->top_level; ++level) {
223 preds[level] = sl->head;
228 TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred);
231 TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key, pred);
235 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
236 map_val_t sl_lookup (skiplist_t *sl, map_key_t key) {
237 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl);
238 node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE);
240 // If we found an <item> matching the <key> return its value.
242 map_val_t val = item->val;
243 if (val != DOES_NOT_EXIST) {
244 TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val);
249 TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0);
250 return DOES_NOT_EXIST;
253 map_key_t sl_min_key (skiplist_t *sl) {
254 node_t *item = GET_NODE(sl->head->next[0]);
255 while (item != NULL) {
256 markable_t next = item->next[0];
259 item = STRIP_MARK(next);
261 return DOES_NOT_EXIST;
264 map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) {
265 TRACE("s1", "sl_cas: key %p skiplist %p", key, sl);
266 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
267 ASSERT((int64_t)new_val > 0);
269 node_t *preds[MAX_LEVEL+1];
270 node_t *nexts[MAX_LEVEL+1];
271 node_t *new_item = NULL;
272 int n = random_level();
274 node_t *old_item = find_preds(preds, nexts, n, sl, key, TRUE);
275 if (old_item == NULL) {
277 // There was not an item in the skiplist that matches the key.
278 if (EXPECT_FALSE(expectation != CAS_EXPECT_DOES_NOT_EXIST && expectation != CAS_EXPECT_WHATEVER)) {
279 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
280 return DOES_NOT_EXIST; // failure
283 // First insert <new_item> into the bottom level.
284 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]);
285 map_key_t new_key = sl->key_type == NULL ? key : (map_key_t)sl->key_type->clone((void *)key);
286 new_item = node_alloc(n, new_key, new_val);
287 node_t *pred = preds[0];
288 markable_t next = new_item->next[0] = (markable_t)nexts[0];
289 for (int level = 1; level <= new_item->top_level; ++level) {
290 new_item->next[level] = (markable_t)nexts[level];
292 markable_t other = SYNC_CAS(&pred->next[0], next, new_item);
294 TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0);
297 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
298 if (sl->key_type != NULL) {
299 nbd_free((void *)new_key);
305 // Found an item in the skiplist that matches the key.
306 map_val_t old_item_val = old_item->val;
308 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
309 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
310 TRACE("s2", "sl_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
314 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
315 TRACE("s1", "sl_cas: found an item %p in the skiplist that matched the key. the expectation was "
316 "not met, the skiplist was not changed", old_item, old_item_val);
317 return old_item_val; // failure
320 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
321 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
322 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
323 // fails it means another thread either removed the node or updated its value.
324 map_val_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
325 if (ret_val == old_item_val) {
326 TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0);
327 return ret_val; // success
329 TRACE("s2", "sl_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
331 old_item_val = ret_val;
335 // Link <new_item> into <sl> from the bottom up.
336 for (int level = 1; level <= new_item->top_level; ++level) {
337 node_t *pred = preds[level];
338 markable_t next = (markable_t)nexts[level];
340 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next);
341 markable_t other = SYNC_CAS(&pred->next[level], next, (markable_t)new_item);
343 TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level);
346 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
347 find_preds(preds, nexts, new_item->top_level, sl, key, TRUE);
349 next = (markable_t)nexts[level];
351 // Update <new_item>'s next pointer
353 // There in no need to continue linking in the item if another thread removed it.
354 markable_t old_next = ((volatile node_t *)new_item)->next[level];
355 if (HAS_MARK(old_next))
356 return DOES_NOT_EXIST; // success
358 // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
359 if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
364 return DOES_NOT_EXIST; // success
367 map_val_t sl_remove (skiplist_t *sl, map_key_t key) {
368 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl);
369 node_t *preds[MAX_LEVEL+1];
370 node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE);
372 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
373 return DOES_NOT_EXIST;
376 // Mark and unlink <item> at each level of <sl> from the top down. If multiple threads try to concurrently remove
377 // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds.
378 for (int level = item->top_level; level > 0; --level) {
380 markable_t old_next = item->next[level];
383 old_next = SYNC_CAS(&item->next[level], next, MARK_NODE((node_t *)next));
384 if (HAS_MARK(old_next)) {
385 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level %llu", item, level);
388 } while (next != old_next);
390 node_t *pred = preds[level];
391 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_MARK(next));
392 markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next));
393 if (other != (markable_t)item) {
394 TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
395 // If our former predecessor now points past us we know another thread unlinked us. Otherwise, we need
396 // to search for a new set of preds.
397 if (other == DOES_NOT_EXIST)
398 continue; // <pred> points past <item> to the end of the list; go on to the next level.
401 if (!HAS_MARK(other)) {
402 map_key_t other_key = GET_NODE(other)->key;
403 if (EXPECT_TRUE(sl->key_type == NULL)) {
404 d = item->key - other_key;
406 d = sl->key_type->cmp((void *)item->key, (void *)other_key);
410 node_t *temp = find_preds(preds, NULL, level, sl, key, TRUE);
412 return DOES_NOT_EXIST; // Another thread removed the item we were targeting.
413 level++; // Redo this level.
419 markable_t old_next = item->next[0];
422 old_next = SYNC_CAS(&item->next[0], next, MARK_NODE((node_t *)next));
423 if (HAS_MARK(old_next)) {
424 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level 0", item, 0);
425 return DOES_NOT_EXIST;
427 } while (next != old_next);
428 TRACE("s1", "sl_remove: marked item %p removed at level 0", item, 0);
430 // Atomically swap out the item's value in case another thread is updating the item while we are
431 // removing it. This establishes which operation occurs first logically, the update or the remove.
432 map_val_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
433 TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0);
435 node_t *pred = preds[0];
436 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_MARK(next));
437 if (SYNC_CAS(&pred->next[0], item, STRIP_MARK(next))) {
438 TRACE("s2", "sl_remove: unlinked item %p from the skiplist at level 0", item, 0);
439 // The thread that completes the unlink should free the memory.
440 if (sl->key_type != NULL) {
441 nbd_defer_free((void *)item->key);
443 nbd_defer_free(item);
448 void sl_print (skiplist_t *sl) {
449 for (int level = MAX_LEVEL; level >= 0; --level) {
450 node_t *item = sl->head;
451 if (item->next[level] == DOES_NOT_EXIST)
453 printf("(%d) ", level);
456 markable_t next = item->next[level];
457 printf("%s%p ", HAS_MARK(next) ? "*" : "", item);
458 item = STRIP_MARK(next);
467 node_t *item = sl->head;
470 int is_marked = HAS_MARK(item->next[0]);
471 printf("%s%p:0x%llx ", is_marked ? "*" : "", item, (uint64_t)item->key);
472 if (item != sl->head) {
473 printf("[%d]", item->top_level);
477 for (int level = 1; level <= item->top_level; ++level) {
478 node_t *next = STRIP_MARK(item->next[level]);
479 is_marked = HAS_MARK(item->next[0]);
480 printf(" %p%s", next, is_marked ? "*" : "");
481 if (item == sl->head && item->next[level] == DOES_NOT_EXIST)
486 item = STRIP_MARK(item->next[0]);
494 sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) {
495 sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t));
496 if (key != DOES_NOT_EXIST) {
497 find_preds(NULL, &iter->next, 0, sl, key, FALSE);
499 iter->next = GET_NODE(sl->head->next[0]);
504 map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) {
506 node_t *item = iter->next;
507 while (item != NULL && HAS_MARK(item->next[0])) {
508 item = STRIP_MARK(item->next[0]);
512 return DOES_NOT_EXIST;
514 iter->next = STRIP_MARK(item->next[0]);
515 if (key_ptr != NULL) {
516 *key_ptr = item->key;
521 void sl_iter_free (sl_iter_t *iter) {