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.
29 // Setting MAX_LEVEL to 0 essentially makes this data structure the Harris-Michael lock-free list (in list.c).
45 const datatype_t *key_type;
48 // Marking the <next> field of a node logically removes it from the list
50 static inline markable_t MARK_NODE(node_t * x) { return TAG_VALUE((markable_t)x, TAG1); }
51 static inline int HAS_MARK(markable_t x) { return (IS_TAGGED(x, TAG1) == TAG1); }
52 static inline node_t * GET_NODE(markable_t x) { assert(!HAS_MARK(x)); return (node_t *)x; }
53 static inline node_t * STRIP_MARK(markable_t x) { return ((node_t *)STRIP_TAG(x, TAG1)); }
55 #define MARK_NODE(x) TAG_VALUE((markable_t)(x), TAG1)
56 #define HAS_MARK(x) (IS_TAGGED((x), TAG1) == TAG1)
57 #define GET_NODE(x) ((node_t *)(x))
58 #define STRIP_MARK(x) ((node_t *)STRIP_TAG((x), TAG1))
61 static int random_level (void) {
62 unsigned r = nbd_rand();
66 r |= 1 << (MAX_LEVEL+1);
68 int n = __builtin_ctz(r)-1;
69 assert(n <= MAX_LEVEL);
73 static node_t *node_alloc (int level, map_key_t key, map_val_t val) {
74 assert(level >= 0 && level <= MAX_LEVEL);
75 size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *);
76 node_t *item = (node_t *)nbd_malloc(sz);
80 item->top_level = level;
84 skiplist_t *sl_alloc (const datatype_t *key_type) {
85 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
86 sl->key_type = key_type;
87 sl->head = node_alloc(MAX_LEVEL, 0, 0);
88 memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
92 void sl_free (skiplist_t *sl) {
93 node_t *item = GET_NODE(sl->head->next[0]);
95 node_t *next = STRIP_MARK(item->next[0]);
96 if (sl->key_type != NULL) {
97 nbd_free((void *)item->key);
104 size_t sl_count (skiplist_t *sl) {
106 node_t *item = GET_NODE(sl->head->next[0]);
108 if (!HAS_MARK(item->next[0])) {
111 item = STRIP_MARK(item->next[0]);
116 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, int help_remove) {
117 node_t *pred = sl->head;
119 TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
121 int start_level = MAX_LEVEL;
123 // Optimization for small lists. No need to traverse empty higher levels.
125 while (pred->next[start_level+1] != DOES_NOT_EXIST) {
126 start_level += start_level - 1;
127 if (EXPECT_FALSE(start_level >= MAX_LEVEL)) {
128 start_level = MAX_LEVEL;
132 if (EXPECT_FALSE(start_level < n)) {
137 // Traverse the levels of <sl> from the top level to the bottom
138 for (int level = start_level; level >= 0; --level) {
139 TRACE("s3", "find_preds: level %llu", level, 0);
140 markable_t next = pred->next[level];
141 if (EXPECT_FALSE(HAS_MARK(next))) {
142 TRACE("s2", "find_preds: pred %p is marked for removal (next %p); retry", pred, next);
143 return find_preds(preds, succs, n, sl, key, help_remove); // retry
145 item = GET_NODE(next);
146 while (item != NULL) {
147 next = item->next[level];
149 // A tag means an item is logically removed but not physically unlinked yet.
150 while (EXPECT_FALSE(HAS_MARK(next))) {
152 // Skip over logically removed items.
154 item = STRIP_MARK(next);
155 if (EXPECT_FALSE(item == NULL))
157 next = item->next[level];
158 TRACE("s3", "find_preds: skipping marked item %p (next is 0x%llx)", item, next);
162 // Unlink logically removed items.
163 TRACE("s3", "find_preds: unlinking marked item %p; next is 0x%llx", item, next);
164 markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next));
165 if (other == (markable_t)item) {
166 item = STRIP_MARK(next);
167 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
168 TRACE("s3", "find_preds: now the current item is %p next is 0x%llx", item, next);
170 // The thread that completes the unlink should free the memory.
172 node_t *unlinked = GET_NODE(other);
173 if (sl->key_type != NULL) {
174 rcu_defer_free((void *)unlinked->key);
176 rcu_defer_free(unlinked);
179 TRACE("s3", "find_preds: lost race to unlink item %p from pred %p", item, pred);
180 TRACE("s3", "find_preds: pred's link changed to %p", other, 0);
182 return find_preds(preds, succs, n, sl, key, help_remove); // retry
183 item = GET_NODE(other);
184 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
188 if (EXPECT_FALSE(item == NULL))
191 TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next);
192 TRACE("s4", "find_preds: key %p val %p", STRIP_MARK(item->key), item->val);
194 if (EXPECT_TRUE(sl->key_type == NULL)) {
197 d = sl->key_type->cmp((void *)item->key, (void *)key);
201 TRACE("s4", "find_preds: found pred %p item %p", pred, item);
206 item = GET_NODE(next);
209 // The cast to unsigned is for the case when n is -1.
210 if ((unsigned)level <= (unsigned)n) {
220 // fill in empty levels
221 if (n == -1 && item != NULL) {
222 assert(item->top_level <= MAX_LEVEL);
223 for (int level = start_level + 1; level <= item->top_level; ++level) {
224 preds[level] = sl->head;
229 TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred);
232 TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key, pred);
236 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
237 map_val_t sl_lookup (skiplist_t *sl, map_key_t key) {
238 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl);
239 node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE);
241 // If we found an <item> matching the <key> return its value.
243 map_val_t val = item->val;
244 if (val != DOES_NOT_EXIST) {
245 TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val);
250 TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0);
251 return DOES_NOT_EXIST;
254 map_key_t sl_min_key (skiplist_t *sl) {
255 node_t *item = GET_NODE(sl->head->next[0]);
256 while (item != NULL) {
257 markable_t next = item->next[0];
260 item = STRIP_MARK(next);
262 return DOES_NOT_EXIST;
265 map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) {
266 TRACE("s1", "sl_cas: key %p skiplist %p", key, sl);
267 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
268 ASSERT((int64_t)new_val > 0);
270 node_t *preds[MAX_LEVEL+1];
271 node_t *nexts[MAX_LEVEL+1];
272 node_t *new_item = NULL;
273 int n = random_level();
275 node_t *old_item = find_preds(preds, nexts, n, sl, key, TRUE);
276 if (old_item == NULL) {
278 // There was not an item in the skiplist that matches the key.
279 if (EXPECT_FALSE(expectation != CAS_EXPECT_DOES_NOT_EXIST && expectation != CAS_EXPECT_WHATEVER)) {
280 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
281 return DOES_NOT_EXIST; // failure
284 // First insert <new_item> into the bottom level.
285 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]);
286 map_key_t new_key = sl->key_type == NULL ? key : (map_key_t)sl->key_type->clone((void *)key);
287 new_item = node_alloc(n, new_key, new_val);
288 node_t *pred = preds[0];
289 markable_t next = new_item->next[0] = (markable_t)nexts[0];
290 for (int level = 1; level <= new_item->top_level; ++level) {
291 new_item->next[level] = (markable_t)nexts[level];
293 markable_t other = SYNC_CAS(&pred->next[0], next, new_item);
295 TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0);
298 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
299 if (sl->key_type != NULL) {
300 nbd_free((void *)new_key);
306 // Found an item in the skiplist that matches the key.
307 map_val_t old_item_val = old_item->val;
309 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
310 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
311 TRACE("s2", "sl_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
315 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
316 TRACE("s1", "sl_cas: found an item %p in the skiplist that matched the key. the expectation was "
317 "not met, the skiplist was not changed", old_item, old_item_val);
318 return old_item_val; // failure
321 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
322 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
323 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
324 // fails it means another thread either removed the node or updated its value.
325 map_val_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
326 if (ret_val == old_item_val) {
327 TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0);
328 return ret_val; // success
330 TRACE("s2", "sl_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
332 old_item_val = ret_val;
336 // Link <new_item> into <sl> from the bottom up.
337 for (int level = 1; level <= new_item->top_level; ++level) {
338 node_t *pred = preds[level];
339 markable_t next = (markable_t)nexts[level];
341 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next);
342 markable_t other = SYNC_CAS(&pred->next[level], next, (markable_t)new_item);
344 TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level);
347 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
348 find_preds(preds, nexts, new_item->top_level, sl, key, TRUE);
350 next = (markable_t)nexts[level];
352 // Update <new_item>'s next pointer
354 // There in no need to continue linking in the item if another thread removed it.
355 markable_t old_next = ((volatile node_t *)new_item)->next[level];
356 if (HAS_MARK(old_next))
357 return DOES_NOT_EXIST; // success
359 // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
360 if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
365 return DOES_NOT_EXIST; // success
368 map_val_t sl_remove (skiplist_t *sl, map_key_t key) {
369 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl);
370 node_t *preds[MAX_LEVEL+1];
371 node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE);
373 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
374 return DOES_NOT_EXIST;
377 // Mark and unlink <item> at each level of <sl> from the top down. If multiple threads try to concurrently remove
378 // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds.
379 for (int level = item->top_level; level > 0; --level) {
381 markable_t old_next = item->next[level];
384 old_next = SYNC_CAS(&item->next[level], next, MARK_NODE((node_t *)next));
385 if (HAS_MARK(old_next)) {
386 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level %llu", item, level);
389 } while (next != old_next);
391 node_t *pred = preds[level];
392 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_MARK(next));
393 markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next));
394 if (other != (markable_t)item) {
395 TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
396 // If our former predecessor now points past us we know another thread unlinked us. Otherwise, we need
397 // to search for a new set of preds.
398 if (other == DOES_NOT_EXIST)
399 continue; // <pred> points past <item> to the end of the list; go on to the next level.
402 if (!HAS_MARK(other)) {
403 map_key_t other_key = GET_NODE(other)->key;
404 if (EXPECT_TRUE(sl->key_type == NULL)) {
405 d = item->key - other_key;
407 d = sl->key_type->cmp((void *)item->key, (void *)other_key);
411 node_t *temp = find_preds(preds, NULL, level, sl, key, TRUE);
413 return DOES_NOT_EXIST; // Another thread removed the item we were targeting.
414 level++; // Redo this level.
420 markable_t old_next = item->next[0];
423 old_next = SYNC_CAS(&item->next[0], next, MARK_NODE((node_t *)next));
424 if (HAS_MARK(old_next)) {
425 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level 0", item, 0);
426 return DOES_NOT_EXIST;
428 } while (next != old_next);
429 TRACE("s1", "sl_remove: marked item %p removed at level 0", item, 0);
431 // Atomically swap out the item's value in case another thread is updating the item while we are
432 // removing it. This establishes which operation occurs first logically, the update or the remove.
433 map_val_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
434 TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0);
436 node_t *pred = preds[0];
437 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_MARK(next));
438 if (SYNC_CAS(&pred->next[0], item, STRIP_MARK(next))) {
439 TRACE("s2", "sl_remove: unlinked item %p from the skiplist at level 0", item, 0);
440 // The thread that completes the unlink should free the memory.
441 if (sl->key_type != NULL) {
442 rcu_defer_free((void *)item->key);
444 rcu_defer_free(item);
449 void sl_print (skiplist_t *sl) {
450 for (int level = MAX_LEVEL; level >= 0; --level) {
451 node_t *item = sl->head;
452 if (item->next[level] == DOES_NOT_EXIST)
454 printf("(%d) ", level);
457 markable_t next = item->next[level];
458 printf("%s%p ", HAS_MARK(next) ? "*" : "", item);
459 item = STRIP_MARK(next);
468 node_t *item = sl->head;
471 int is_marked = HAS_MARK(item->next[0]);
472 printf("%s%p:0x%llx ", is_marked ? "*" : "", item, (uint64_t)item->key);
473 if (item != sl->head) {
474 printf("[%d]", item->top_level);
478 for (int level = 1; level <= item->top_level; ++level) {
479 node_t *next = STRIP_MARK(item->next[level]);
480 is_marked = HAS_MARK(item->next[0]);
481 printf(" %p%s", next, is_marked ? "*" : "");
482 if (item == sl->head && item->next[level] == DOES_NOT_EXIST)
487 item = STRIP_MARK(item->next[0]);
495 sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) {
496 sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t));
497 if (key != DOES_NOT_EXIST) {
498 find_preds(NULL, &iter->next, 0, sl, key, FALSE);
500 iter->next = GET_NODE(sl->head->next[0]);
505 map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) {
507 node_t *item = iter->next;
508 while (item != NULL && HAS_MARK(item->next[0])) {
509 item = STRIP_MARK(item->next[0]);
513 return DOES_NOT_EXIST;
515 iter->next = STRIP_MARK(item->next[0]);
516 if (key_ptr != NULL) {
517 *key_ptr = item->key;
522 void sl_iter_free (sl_iter_t *iter) {