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).
40 const datatype_t *key_type;
43 static int random_level (void) {
44 unsigned r = nbd_rand();
48 r |= 1 << (MAX_LEVEL+1);
50 int n = __builtin_ctz(r)-1;
51 assert(n <= MAX_LEVEL);
55 static node_t *node_alloc (int level, void *key, uint64_t val) {
56 assert(level >= 0 && level <= MAX_LEVEL);
57 size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *);
58 node_t *item = (node_t *)nbd_malloc(sz);
62 item->top_level = level;
66 skiplist_t *sl_alloc (const datatype_t *key_type) {
67 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
68 sl->key_type = key_type;
69 sl->head = node_alloc(MAX_LEVEL, NULL, 0);
70 memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
74 void sl_free (skiplist_t *sl) {
75 node_t *item = sl->head->next[0];
77 node_t *next = (node_t *)STRIP_TAG(item->next[0]);
83 uint64_t sl_count (skiplist_t *sl) {
85 node_t *item = sl->head->next[0];
87 if (!IS_TAGGED(item->next[0])) {
90 item = (node_t *)STRIP_TAG(item->next[0]);
95 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, void *key, int help_remove) {
96 node_t *pred = sl->head;
98 TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
100 int start_level = MAX_LEVEL;
102 // Optimization for small lists. No need to traverse empty higher levels.
104 while (pred->next[start_level+1] != NULL) {
105 start_level += start_level - 1;
106 if (EXPECT_FALSE(start_level >= MAX_LEVEL)) {
107 start_level = MAX_LEVEL;
111 if (EXPECT_FALSE(start_level < n)) {
116 // Traverse the levels of <sl> from the top level to the bottom
117 for (int level = start_level; level >= 0; --level) {
118 TRACE("s3", "find_preds: level %llu", level, 0);
119 item = pred->next[level];
120 if (EXPECT_FALSE(IS_TAGGED(item))) {
121 TRACE("s2", "find_preds: pred %p is marked for removal (item %p); retry", pred, item);
122 return find_preds(preds, succs, n, sl, key, help_remove); // retry
124 while (item != NULL) {
125 node_t *next = item->next[level];
127 // A tag means an item is logically removed but not physically unlinked yet.
128 while (EXPECT_FALSE(IS_TAGGED(next))) {
130 // Skip over logically removed items.
132 item = (node_t *)STRIP_TAG(item->next);
133 if (EXPECT_FALSE(item == NULL))
135 TRACE("s3", "find_preds: skipping marked item %p (next is %p)", item, next);
136 next = item->next[level];
140 // Unlink logically removed items.
142 TRACE("s3", "find_preds: unlinking marked item %p; next is %p", item, next);
143 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) == item) {
144 item = (node_t *)STRIP_TAG(next);
145 if (EXPECT_FALSE(item == NULL))
147 next = item->next[level];
148 TRACE("s3", "find_preds: now the current item is %p next is %p", item, next);
150 // The thread that completes the unlink should free the memory.
152 if (sl->key_type != NULL) {
153 nbd_defer_free((void*)other->key);
155 nbd_defer_free(other);
158 TRACE("s3", "find_preds: lost race to unlink item %p from pred %p", item, pred);
159 TRACE("s3", "find_preds: pred's link changed to %p", other, 0);
160 if (IS_TAGGED(other))
161 return find_preds(preds, succs, n, sl, key, help_remove); // retry
163 if (EXPECT_FALSE(item == NULL))
165 next = item->next[level];
169 if (EXPECT_FALSE(item == NULL))
172 TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next);
173 TRACE("s4", "find_preds: key %p val %p", STRIP_TAG(item->key), item->val);
175 if (EXPECT_TRUE(sl->key_type == NULL)) {
176 d = (uint64_t)item->key - (uint64_t)key;
178 d = sl->key_type->cmp(item->key, key);
182 TRACE("s4", "find_preds: found pred %p item %p", pred, item);
190 // The cast to unsigned is for the case when n is -1.
191 if ((unsigned)level <= (unsigned)n) {
201 // fill in empty levels
202 if (n == -1 && item != NULL) {
203 for (int level = start_level + 1; level <= item->top_level; ++level) {
204 preds[level] = sl->head;
209 TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred);
212 TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key, pred);
216 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
217 uint64_t sl_lookup (skiplist_t *sl, void *key) {
218 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl);
219 node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE);
221 // If we found an <item> matching the <key> return its value.
223 uint64_t val = item->val;
224 if (val != DOES_NOT_EXIST) {
225 TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val);
230 TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0);
231 return DOES_NOT_EXIST;
234 void *sl_min_key (skiplist_t *sl) {
235 node_t *item = sl->head->next[0];
236 while (item != NULL) {
237 node_t *next = item->next[0];
238 if (!IS_TAGGED(next))
240 item = (node_t *)STRIP_TAG(next);
242 return DOES_NOT_EXIST;
245 uint64_t sl_cas (skiplist_t *sl, void *key, uint64_t expectation, uint64_t new_val) {
246 TRACE("s1", "sl_cas: key %p skiplist %p", key, sl);
247 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
248 ASSERT((int64_t)new_val > 0);
250 node_t *preds[MAX_LEVEL+1];
251 node_t *nexts[MAX_LEVEL+1];
252 node_t *new_item = NULL;
253 int n = random_level();
255 node_t *old_item = find_preds(preds, nexts, n, sl, key, TRUE);
256 if (old_item == NULL) {
258 // There was not an item in the skiplist that matches the key.
259 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
260 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
261 return DOES_NOT_EXIST; // failure
264 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
266 // First insert <new_item> into the bottom level.
267 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]);
268 void *new_key = (sl->key_type == NULL) ? key : sl->key_type->clone(key);
269 new_item = node_alloc(n, new_key, new_val);
270 node_t *pred = preds[0];
271 node_t *next = new_item->next[0] = nexts[0];
272 for (int level = 1; level <= new_item->top_level; ++level) {
273 new_item->next[level] = nexts[level];
275 node_t *other = SYNC_CAS(&pred->next[0], next, new_item);
277 TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0);
280 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
281 if (sl->key_type != NULL) {
288 // Found an item in the skiplist that matches the key.
289 uint64_t old_item_val = old_item->val;
291 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
292 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
293 TRACE("s2", "sl_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
297 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
298 TRACE("s1", "sl_cas: found an item %p in the skiplist that matched the key. the expectation was "
299 "not met, the skiplist was not changed", old_item, old_item_val);
300 return old_item_val; // failure
303 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
304 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
305 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
306 // fails it means another thread either removed the node or updated its value.
307 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
308 if (ret_val == old_item_val) {
309 TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0);
310 return ret_val; // success
312 TRACE("s2", "sl_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
314 old_item_val = ret_val;
318 // Link <new_item> into <sl> from the bottom up.
319 for (int level = 1; level <= new_item->top_level; ++level) {
320 node_t *pred = preds[level];
321 node_t *next = nexts[level];
323 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next);
324 node_t *other = SYNC_CAS(&pred->next[level], next, new_item);
326 TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level);
329 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
330 find_preds(preds, nexts, new_item->top_level, sl, key, TRUE);
334 // Update <new_item>'s next pointer
336 // There in no need to continue linking in the item if another thread removed it.
337 node_t *old_next = ((volatile node_t *)new_item)->next[level];
338 if (IS_TAGGED(old_next))
339 return DOES_NOT_EXIST; // success
341 // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
342 if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
347 return DOES_NOT_EXIST; // success
350 uint64_t sl_remove (skiplist_t *sl, void *key) {
351 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl);
352 node_t *preds[MAX_LEVEL+1];
353 node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE);
355 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
356 return DOES_NOT_EXIST;
359 // Mark and unlink <item> at each level of <sl> from the top down. If multiple threads try to concurrently remove
360 // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds.
361 for (int level = item->top_level; level > 0; --level) {
363 node_t *old_next = item->next[level];
366 old_next = SYNC_CAS(&item->next[level], next, TAG_VALUE(next));
367 if (IS_TAGGED(old_next)) {
368 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level %llu", item, level);
371 } while (next != old_next);
373 node_t *pred = preds[level];
374 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next));
375 node_t *other = NULL;
376 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) != item) {
377 TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
378 // If our former predecessor now points past us we know another thread unlinked us. Otherwise, we need
379 // to search for a new set of preds.
381 continue; // <pred> points past <item> to the end of the list; go on to the next level.
384 if (!IS_TAGGED(other)) {
385 if (EXPECT_TRUE(sl->key_type == NULL)) {
386 d = (uint64_t)item->key - (uint64_t)other->key;
388 d = sl->key_type->cmp(item->key, other->key);
392 node_t *temp = find_preds(preds, NULL, level, sl, key, TRUE);
394 return DOES_NOT_EXIST; // Another thread removed the item we were targeting.
395 level++; // Redo this level.
401 node_t *old_next = item->next[0];
404 old_next = SYNC_CAS(&item->next[0], next, TAG_VALUE(next));
405 if (IS_TAGGED(old_next)) {
406 TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level 0", item, 0);
407 return DOES_NOT_EXIST;
409 } while (next != old_next);
410 TRACE("s1", "sl_remove: marked item %p removed at level 0", item, 0);
412 // Atomically swap out the item's value in case another thread is updating the item while we are
413 // removing it. This establishes which operation occurs first logically, the update or the remove.
414 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
415 TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0);
417 node_t *pred = preds[0];
418 TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next));
419 if (SYNC_CAS(&pred->next[0], item, STRIP_TAG(next))) {
420 TRACE("s2", "sl_remove: unlinked item %p from the skiplist at level 0", item, 0);
421 // The thread that completes the unlink should free the memory.
422 if (sl->key_type != NULL) {
423 nbd_defer_free(item->key);
425 nbd_defer_free(item);
430 void sl_print (skiplist_t *sl) {
431 for (int level = MAX_LEVEL; level >= 0; --level) {
432 node_t *item = sl->head;
433 if (item->next[level] == NULL)
435 printf("(%d) ", level);
438 node_t *next = item->next[level];
439 printf("%s%p ", IS_TAGGED(next) ? "*" : "", item);
440 item = (node_t *)STRIP_TAG(next);
449 node_t *item = sl->head;
452 int is_marked = IS_TAGGED(item->next[0]);
453 printf("%s%p:%p ", is_marked ? "*" : "", item, item->key);
454 if (item != sl->head) {
455 printf("[%d]", item->top_level);
459 for (int level = 1; level <= item->top_level; ++level) {
460 node_t *next = (node_t *)STRIP_TAG(item->next[level]);
461 is_marked = IS_TAGGED(item->next[0]);
462 printf(" %p%s", next, is_marked ? "*" : "");
463 if (item == sl->head && item->next[level] == NULL)
468 item = (node_t *)STRIP_TAG(item->next[0]);