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_LEVELS to 1 essentially makes this data structure the Harris-Michael lock-free list (see list.c).
51 const datatype_t *key_type;
52 int high_water; // max level of any item in the list
55 // Marking the <next> field of a node logically removes it from the list
57 static inline markable_t MARK_NODE(node_t * x) { return TAG_VALUE((markable_t)x, 0x1); }
58 static inline int HAS_MARK(markable_t x) { return (IS_TAGGED(x, 0x1) == 0x1); }
59 static inline node_t * GET_NODE(markable_t x) { assert(!HAS_MARK(x)); return (node_t *)x; }
60 static inline node_t * STRIP_MARK(markable_t x) { return ((node_t *)STRIP_TAG(x, 0x1)); }
62 #define MARK_NODE(x) TAG_VALUE((markable_t)(x), 0x1)
63 #define HAS_MARK(x) (IS_TAGGED((x), 0x1) == 0x1)
64 #define GET_NODE(x) ((node_t *)(x))
65 #define STRIP_MARK(x) ((node_t *)STRIP_TAG((x), 0x1))
68 static int random_levels (void) {
69 unsigned r = nbd_rand();
70 int n = __builtin_ctz(r) / 2 + 1;
71 if (n > MAX_LEVELS) { n = MAX_LEVELS; }
75 static node_t *node_alloc (int num_levels, map_key_t key, map_val_t val) {
76 assert(num_levels >= 0 && num_levels <= MAX_LEVELS);
77 size_t sz = sizeof(node_t) + (num_levels - 1) * sizeof(node_t *);
78 node_t *item = (node_t *)nbd_malloc(sz);
82 item->num_levels = num_levels;
83 TRACE("s2", "node_alloc: new node %p (%llu levels)", item, num_levels);
87 skiplist_t *sl_alloc (const datatype_t *key_type) {
88 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
89 sl->key_type = key_type;
91 sl->head = node_alloc(MAX_LEVELS, 0, 0);
92 memset(sl->head->next, 0, MAX_LEVELS * sizeof(skiplist_t *));
96 void sl_free (skiplist_t *sl) {
97 node_t *item = GET_NODE(sl->head->next[0]);
99 node_t *next = STRIP_MARK(item->next[0]);
100 if (sl->key_type != NULL) {
101 nbd_free((void *)item->key);
108 size_t sl_count (skiplist_t *sl) {
110 node_t *item = GET_NODE(sl->head->next[0]);
112 if (!HAS_MARK(item->next[0])) {
115 item = STRIP_MARK(item->next[0]);
120 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, enum unlink unlink) {
121 node_t *pred = sl->head;
123 TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
126 // Traverse the levels of <sl> from the top level to the bottom
127 for (int level = sl->high_water - 1; level >= 0; --level) {
128 markable_t next = pred->next[level];
129 if (next == DOES_NOT_EXIST && level >= n)
131 TRACE("s3", "find_preds: traversing level %p starting at %p", level, pred);
132 if (EXPECT_FALSE(HAS_MARK(next))) {
133 TRACE("s2", "find_preds: pred %p is marked for removal (next %p); retry", pred, next);
134 ASSERT(level == pred->num_levels - 1 || HAS_MARK(pred->next[level+1]));
135 return find_preds(preds, succs, n, sl, key, unlink); // retry
137 item = GET_NODE(next);
138 while (item != NULL) {
139 next = item->next[level];
141 // A tag means an item is logically removed but not physically unlinked yet.
142 while (EXPECT_FALSE(HAS_MARK(next))) {
143 TRACE("s3", "find_preds: found marked item %p (next is %p)", item, next);
144 if (unlink == DONT_UNLINK) {
146 // Skip over logically removed items.
147 item = STRIP_MARK(next);
148 if (EXPECT_FALSE(item == NULL))
150 next = item->next[level];
153 // Unlink logically removed items.
154 markable_t other = SYNC_CAS(&pred->next[level], (markable_t)item, (markable_t)STRIP_MARK(next));
155 if (other == (markable_t)item) {
156 TRACE("s3", "find_preds: unlinked item from pred %p", pred, 0);
157 item = STRIP_MARK(next);
159 TRACE("s3", "find_preds: lost race to unlink item pred %p's link changed to %p", pred, other);
161 return find_preds(preds, succs, n, sl, key, unlink); // retry
162 item = GET_NODE(other);
164 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
168 if (EXPECT_FALSE(item == NULL)) {
169 TRACE("s3", "find_preds: past the last item in the skiplist", 0, 0);
173 TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next);
174 TRACE("s4", "find_preds: key %p val %p", STRIP_MARK(item->key), item->val);
176 if (EXPECT_TRUE(sl->key_type == NULL)) {
179 d = sl->key_type->cmp((void *)item->key, (void *)key);
184 if (d == 0 && unlink != FORCE_UNLINK)
188 item = GET_NODE(next);
191 TRACE("s3", "find_preds: found pred %p next %p", pred, item);
204 TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred);
207 TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key, pred);
211 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
212 map_val_t sl_lookup (skiplist_t *sl, map_key_t key) {
213 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl);
214 node_t *item = find_preds(NULL, NULL, 0, sl, key, DONT_UNLINK);
216 // If we found an <item> matching the <key> return its value.
218 map_val_t val = item->val;
219 if (val != DOES_NOT_EXIST) {
220 TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val);
225 TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0);
226 return DOES_NOT_EXIST;
229 map_key_t sl_min_key (skiplist_t *sl) {
230 node_t *item = GET_NODE(sl->head->next[0]);
231 while (item != NULL) {
232 markable_t next = item->next[0];
235 item = STRIP_MARK(next);
237 return DOES_NOT_EXIST;
240 static map_val_t update_item (node_t *item, map_val_t expectation, map_val_t new_val) {
241 map_val_t old_val = item->val;
243 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
244 if (EXPECT_FALSE(old_val == DOES_NOT_EXIST)) {
245 TRACE("s2", "update_item: lost a race to another thread removing the item. retry", 0, 0);
246 return DOES_NOT_EXIST; // retry
249 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
250 TRACE("s1", "update_item: found an item %p in the skiplist that matched the key. the expectation was "
251 "not met, the skiplist was not changed", item, old_val);
252 return old_val; // failure
255 // Use a CAS and not a SWAP. If the CAS fails it means another thread removed the node or updated its
256 // value. If another thread removed the node but it is not unlinked yet and we used a SWAP, we could
257 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
258 // succeeded and return our value even though it should return DOES_NOT_EXIST.
259 if (old_val == SYNC_CAS(&item->val, old_val, new_val)) {
260 TRACE("s1", "update_item: the CAS succeeded. updated the value of the item", 0, 0);
261 return old_val; // success
263 TRACE("s2", "update_item: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
266 return update_item(item, expectation, new_val); // tail call
269 map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) {
270 TRACE("s1", "sl_cas: key %p skiplist %p", key, sl);
271 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
272 ASSERT((int64_t)new_val > 0);
274 node_t *preds[MAX_LEVELS];
275 node_t *nexts[MAX_LEVELS];
276 node_t *new_item = NULL;
277 int n = random_levels();
278 if (n > sl->high_water) {
279 n = SYNC_ADD(&sl->high_water, 1);
280 TRACE("s2", "sl_cas: incremented high water mark to %p", n, 0);
282 node_t *old_item = find_preds(preds, nexts, n, sl, key, ASSIST_UNLINK);
284 // If there is already an item in the skiplist that matches the key just update its value.
285 if (old_item != NULL) {
286 map_val_t ret_val = update_item(old_item, expectation, new_val);
287 if (ret_val != DOES_NOT_EXIST)
290 // If we lose a race with a thread removing the item we tried to update then we have to retry.
291 return sl_cas(sl, key, expectation, new_val); // tail call
294 if (EXPECT_FALSE(expectation != CAS_EXPECT_DOES_NOT_EXIST && expectation != CAS_EXPECT_WHATEVER)) {
295 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
296 return DOES_NOT_EXIST; // failure, the caller expected an item for the <key> to already exist
299 // Create a new node and insert it into the skiplist.
300 TRACE("s3", "sl_cas: attempting to insert a new item between %p and %p", preds[0], nexts[0]);
301 map_key_t new_key = sl->key_type == NULL ? key : (map_key_t)sl->key_type->clone((void *)key);
302 new_item = node_alloc(n, new_key, new_val);
304 // Set <new_item>'s next pointers to their proper values
305 markable_t next = new_item->next[0] = (markable_t)nexts[0];
306 for (int level = 1; level < new_item->num_levels; ++level) {
307 new_item->next[level] = (markable_t)nexts[level];
310 // Link <new_item> into <sl> from the bottom level up. After <new_item> is inserted into the bottom level
311 // it is officially part of the skiplist.
312 node_t *pred = preds[0];
313 markable_t other = SYNC_CAS(&pred->next[0], next, (markable_t)new_item);
315 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
317 // Lost a race to another thread modifying the skiplist. Free the new item we allocated and retry.
318 if (sl->key_type != NULL) {
319 nbd_free((void *)new_key);
322 return sl_cas(sl, key, expectation, new_val); // tail call
325 TRACE("s3", "sl_cas: successfully inserted a new item %p at the bottom level", new_item, 0);
327 ASSERT(new_item->num_levels <= MAX_LEVELS);
328 for (int level = 1; level < new_item->num_levels; ++level) {
329 TRACE("s3", "sl_cas: inserting the new item %p at level %p", new_item, level);
331 node_t * pred = preds[level];
332 ASSERT(new_item->next[level]==(markable_t)nexts[level] || new_item->next[level]==MARK_NODE(nexts[level]));
333 TRACE("s3", "sl_cas: attempting to to insert the new item between %p and %p", pred, nexts[level]);
335 markable_t other = SYNC_CAS(&pred->next[level], (markable_t)nexts[level], (markable_t)new_item);
336 if (other == (markable_t)nexts[level])
337 break; // successfully linked <new_item> into the skiplist at the current <level>
338 TRACE("s3", "sl_cas: lost a race. failed to change pred's link. expected %p found %p", nexts[level], other);
340 // Find <new_item>'s new preds and nexts.
341 find_preds(preds, nexts, new_item->num_levels, sl, key, ASSIST_UNLINK);
343 for (int i = level; i < new_item->num_levels; ++i) {
344 markable_t old_next = new_item->next[i];
345 if ((markable_t)nexts[i] == old_next)
348 // Update <new_item>'s inconsistent next pointer before trying again. Use a CAS so if another thread
349 // is trying to remove the new item concurrently we do not stomp on the mark it places on the item.
350 TRACE("s3", "sl_cas: attempting to update the new item's link from %p to %p", old_next, nexts[i]);
351 other = SYNC_CAS(&new_item->next[i], old_next, (markable_t)nexts[i]);
352 ASSERT(other == old_next || other == MARK_NODE(old_next));
354 // If another thread is removing this item we can stop linking it into to skiplist
355 if (HAS_MARK(other)) {
356 find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK); // see comment below
357 return DOES_NOT_EXIST;
363 // In case another thread was in the process of removing the <new_item> while we were added it, we have to
364 // make sure it is completely unlinked before we return. We might have lost a race and inserted the new item
365 // at some level after the other thread thought it was fully removed. That is a problem because once a thread
366 // thinks it completely unlinks a node it queues it to be freed
367 if (HAS_MARK(new_item->next[new_item->num_levels - 1])) {
368 find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK);
371 return DOES_NOT_EXIST; // success, inserted a new item
374 map_val_t sl_remove (skiplist_t *sl, map_key_t key) {
375 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl);
376 node_t *preds[MAX_LEVELS];
377 node_t *item = find_preds(preds, NULL, sl->high_water, sl, key, ASSIST_UNLINK);
379 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
380 return DOES_NOT_EXIST;
383 // Mark <item> at each level of <sl> from the top down. If multiple threads try to concurrently remove
384 // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds.
385 markable_t old_next = 0;
386 for (int level = item->num_levels - 1; level >= 0; --level) {
388 old_next = item->next[level];
390 TRACE("s3", "sl_remove: marking item at level %p (next %p)", level, old_next);
392 old_next = SYNC_CAS(&item->next[level], next, MARK_NODE((node_t *)next));
393 if (HAS_MARK(old_next)) {
394 TRACE("s2", "sl_remove: %p is already marked for removal by another thread (next %p)", item, old_next);
396 return DOES_NOT_EXIST;
399 } while (next != old_next);
402 // Atomically swap out the item's value in case another thread is updating the item while we are
403 // removing it. This establishes which operation occurs first logically, the update or the remove.
404 map_val_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
405 TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0);
408 find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK);
411 if (sl->key_type != NULL) {
412 rcu_defer_free((void *)item->key);
414 rcu_defer_free(item);
419 void sl_print (skiplist_t *sl) {
421 printf("high water: %d levels\n", sl->high_water);
422 for (int level = MAX_LEVELS - 1; level >= 0; --level) {
423 node_t *item = sl->head;
424 if (item->next[level] == DOES_NOT_EXIST)
426 printf("(%d) ", level);
429 markable_t next = item->next[level];
430 printf("%s%p ", HAS_MARK(next) ? "*" : "", item);
431 item = STRIP_MARK(next);
440 node_t *item = sl->head;
443 int is_marked = HAS_MARK(item->next[0]);
444 printf("%s%p:0x%llx ", is_marked ? "*" : "", item, (uint64_t)item->key);
445 if (item != sl->head) {
446 printf("[%d]", item->num_levels);
450 for (int level = 1; level < item->num_levels; ++level) {
451 node_t *next = STRIP_MARK(item->next[level]);
452 is_marked = HAS_MARK(item->next[0]);
453 printf(" %p%s", next, is_marked ? "*" : "");
454 if (item == sl->head && item->next[level] == DOES_NOT_EXIST)
459 item = STRIP_MARK(item->next[0]);
467 sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) {
468 sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t));
469 if (key != DOES_NOT_EXIST) {
470 find_preds(NULL, &iter->next, 1, sl, key, DONT_UNLINK);
472 iter->next = GET_NODE(sl->head->next[0]);
477 map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) {
479 node_t *item = iter->next;
480 while (item != NULL && HAS_MARK(item->next[0])) {
481 item = STRIP_MARK(item->next[0]);
485 return DOES_NOT_EXIST;
487 iter->next = STRIP_MARK(item->next[0]);
488 if (key_ptr != NULL) {
489 *key_ptr = item->key;
494 void sl_iter_free (sl_iter_t *iter) {