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 historic number of levels
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 (skiplist_t *sl) {
69 uint64_t r = nbd_rand();
70 int z = __builtin_ctz(r);
71 int levels = (int)(z / 1.5);
74 if (levels > sl->high_water) {
75 levels = SYNC_ADD(&sl->high_water, 1);
76 TRACE("s2", "random_levels: increased high water mark to %lld", sl->high_water, 0);
78 if (levels > MAX_LEVELS) { levels = MAX_LEVELS; }
82 static node_t *node_alloc (int num_levels, map_key_t key, map_val_t val) {
83 assert(num_levels >= 0 && num_levels <= MAX_LEVELS);
84 size_t sz = sizeof(node_t) + (num_levels - 1) * sizeof(node_t *);
85 node_t *item = (node_t *)nbd_malloc(sz);
89 item->num_levels = num_levels;
90 TRACE("s2", "node_alloc: new node %p (%llu levels)", item, num_levels);
94 skiplist_t *sl_alloc (const datatype_t *key_type) {
95 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
96 sl->key_type = key_type;
98 sl->head = node_alloc(MAX_LEVELS, 0, 0);
99 memset(sl->head->next, 0, MAX_LEVELS * sizeof(skiplist_t *));
103 void sl_free (skiplist_t *sl) {
104 node_t *item = GET_NODE(sl->head->next[0]);
106 node_t *next = STRIP_MARK(item->next[0]);
107 if (sl->key_type != NULL) {
108 nbd_free((void *)item->key);
115 size_t sl_count (skiplist_t *sl) {
117 node_t *item = GET_NODE(sl->head->next[0]);
119 if (!HAS_MARK(item->next[0])) {
122 item = STRIP_MARK(item->next[0]);
127 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, enum unlink unlink) {
128 node_t *pred = sl->head;
130 TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
133 // Traverse the levels of <sl> from the top level to the bottom
134 for (int level = sl->high_water - 1; level >= 0; --level) {
135 markable_t next = pred->next[level];
136 if (next == DOES_NOT_EXIST && level >= n)
138 TRACE("s3", "find_preds: traversing level %p starting at %p", level, pred);
139 if (EXPECT_FALSE(HAS_MARK(next))) {
140 TRACE("s2", "find_preds: pred %p is marked for removal (next %p); retry", pred, next);
141 ASSERT(level == pred->num_levels - 1 || HAS_MARK(pred->next[level+1]));
142 return find_preds(preds, succs, n, sl, key, unlink); // 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))) {
150 TRACE("s3", "find_preds: found marked item %p (next is %p)", item, next);
151 if (unlink == DONT_UNLINK) {
153 // Skip over logically removed items.
154 item = STRIP_MARK(next);
155 if (EXPECT_FALSE(item == NULL))
157 next = item->next[level];
160 // Unlink logically removed items.
161 markable_t other = SYNC_CAS(&pred->next[level], (markable_t)item, (markable_t)STRIP_MARK(next));
162 if (other == (markable_t)item) {
163 TRACE("s3", "find_preds: unlinked item from pred %p", pred, 0);
164 item = STRIP_MARK(next);
166 TRACE("s3", "find_preds: lost race to unlink item pred %p's link changed to %p", pred, other);
168 return find_preds(preds, succs, n, sl, key, unlink); // retry
169 item = GET_NODE(other);
171 next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST;
175 if (EXPECT_FALSE(item == NULL)) {
176 TRACE("s3", "find_preds: past the last item in the skiplist", 0, 0);
180 TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next);
181 TRACE("s4", "find_preds: key %p val %p", STRIP_MARK(item->key), item->val);
183 if (EXPECT_TRUE(sl->key_type == NULL)) {
186 d = sl->key_type->cmp((void *)item->key, (void *)key);
191 if (d == 0 && unlink != FORCE_UNLINK)
195 item = GET_NODE(next);
198 TRACE("s3", "find_preds: found pred %p next %p", pred, item);
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, DONT_UNLINK);
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("s1", "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 = GET_NODE(sl->head->next[0]);
238 while (item != NULL) {
239 markable_t next = item->next[0];
242 item = STRIP_MARK(next);
244 return DOES_NOT_EXIST;
247 static map_val_t update_item (node_t *item, map_val_t expectation, map_val_t new_val) {
248 map_val_t old_val = item->val;
250 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
251 if (EXPECT_FALSE(old_val == DOES_NOT_EXIST)) {
252 TRACE("s2", "update_item: lost a race to another thread removing the item. retry", 0, 0);
253 return DOES_NOT_EXIST; // retry
256 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
257 TRACE("s1", "update_item: the expectation was not met; the skiplist was not changed", 0, 0);
258 return old_val; // failure
261 // Use a CAS and not a SWAP. If the CAS fails it means another thread removed the node or updated its
262 // value. If another thread removed the node but it is not unlinked yet and we used a SWAP, we could
263 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
264 // succeeded and return our value even though it should return DOES_NOT_EXIST.
265 if (old_val == SYNC_CAS(&item->val, old_val, new_val)) {
266 TRACE("s1", "update_item: the CAS succeeded. updated the value of the item", 0, 0);
267 return old_val; // success
269 TRACE("s2", "update_item: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
272 return update_item(item, expectation, new_val); // tail call
275 map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) {
276 TRACE("s1", "sl_cas: key %p skiplist %p", key, sl);
277 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
278 ASSERT((int64_t)new_val > 0);
280 node_t *preds[MAX_LEVELS];
281 node_t *nexts[MAX_LEVELS];
282 node_t *new_item = NULL;
283 int n = random_levels(sl);
284 node_t *old_item = find_preds(preds, nexts, n, sl, key, ASSIST_UNLINK);
286 // If there is already an item in the skiplist that matches the key just update its value.
287 if (old_item != NULL) {
288 map_val_t ret_val = update_item(old_item, expectation, new_val);
289 if (ret_val != DOES_NOT_EXIST)
292 // If we lose a race with a thread removing the item we tried to update then we have to retry.
293 return sl_cas(sl, key, expectation, new_val); // tail call
296 if (EXPECT_FALSE(expectation != CAS_EXPECT_DOES_NOT_EXIST && expectation != CAS_EXPECT_WHATEVER)) {
297 TRACE("s1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
298 return DOES_NOT_EXIST; // failure, the caller expected an item for the <key> to already exist
301 // Create a new node and insert it into the skiplist.
302 TRACE("s3", "sl_cas: attempting to insert a new item between %p and %p", preds[0], nexts[0]);
303 map_key_t new_key = sl->key_type == NULL ? key : (map_key_t)sl->key_type->clone((void *)key);
304 new_item = node_alloc(n, new_key, new_val);
306 // Set <new_item>'s next pointers to their proper values
307 markable_t next = new_item->next[0] = (markable_t)nexts[0];
308 for (int level = 1; level < new_item->num_levels; ++level) {
309 new_item->next[level] = (markable_t)nexts[level];
312 // Link <new_item> into <sl> from the bottom level up. After <new_item> is inserted into the bottom level
313 // it is officially part of the skiplist.
314 node_t *pred = preds[0];
315 markable_t other = SYNC_CAS(&pred->next[0], next, (markable_t)new_item);
317 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
319 // Lost a race to another thread modifying the skiplist. Free the new item we allocated and retry.
320 if (sl->key_type != NULL) {
321 nbd_free((void *)new_key);
324 return sl_cas(sl, key, expectation, new_val); // tail call
327 TRACE("s3", "sl_cas: successfully inserted a new item %p at the bottom level", new_item, 0);
329 ASSERT(new_item->num_levels <= MAX_LEVELS);
330 for (int level = 1; level < new_item->num_levels; ++level) {
331 TRACE("s3", "sl_cas: inserting the new item %p at level %p", new_item, level);
333 node_t * pred = preds[level];
334 ASSERT(new_item->next[level]==(markable_t)nexts[level] || new_item->next[level]==MARK_NODE(nexts[level]));
335 TRACE("s3", "sl_cas: attempting to to insert the new item between %p and %p", pred, nexts[level]);
337 markable_t other = SYNC_CAS(&pred->next[level], (markable_t)nexts[level], (markable_t)new_item);
338 if (other == (markable_t)nexts[level])
339 break; // successfully linked <new_item> into the skiplist at the current <level>
340 TRACE("s3", "sl_cas: lost a race. failed to change pred's link. expected %p found %p", nexts[level], other);
342 // Find <new_item>'s new preds and nexts.
343 find_preds(preds, nexts, new_item->num_levels, sl, key, ASSIST_UNLINK);
345 for (int i = level; i < new_item->num_levels; ++i) {
346 markable_t old_next = new_item->next[i];
347 if ((markable_t)nexts[i] == old_next)
350 // Update <new_item>'s inconsistent next pointer before trying again. Use a CAS so if another thread
351 // is trying to remove the new item concurrently we do not stomp on the mark it places on the item.
352 TRACE("s3", "sl_cas: attempting to update the new item's link from %p to %p", old_next, nexts[i]);
353 other = SYNC_CAS(&new_item->next[i], old_next, (markable_t)nexts[i]);
354 ASSERT(other == old_next || other == MARK_NODE(old_next));
356 // If another thread is removing this item we can stop linking it into to skiplist
357 if (HAS_MARK(other)) {
358 find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK); // see comment below
359 return DOES_NOT_EXIST;
365 // In case another thread was in the process of removing the <new_item> while we were added it, we have to
366 // make sure it is completely unlinked before we return. We might have lost a race and inserted the new item
367 // at some level after the other thread thought it was fully removed. That is a problem because once a thread
368 // thinks it completely unlinks a node it queues it to be freed
369 if (HAS_MARK(new_item->next[new_item->num_levels - 1])) {
370 find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK);
373 return DOES_NOT_EXIST; // success, inserted a new item
376 map_val_t sl_remove (skiplist_t *sl, map_key_t key) {
377 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl);
378 node_t *preds[MAX_LEVELS];
379 node_t *item = find_preds(preds, NULL, sl->high_water, sl, key, ASSIST_UNLINK);
381 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
382 return DOES_NOT_EXIST;
385 // Mark <item> at each level of <sl> from the top down. If multiple threads try to concurrently remove
386 // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds.
387 markable_t old_next = 0;
388 for (int level = item->num_levels - 1; level >= 0; --level) {
390 old_next = item->next[level];
392 TRACE("s3", "sl_remove: marking item at level %p (next %p)", level, old_next);
394 old_next = SYNC_CAS(&item->next[level], next, MARK_NODE((node_t *)next));
395 if (HAS_MARK(old_next)) {
396 TRACE("s2", "sl_remove: %p is already marked for removal by another thread (next %p)", item, old_next);
398 return DOES_NOT_EXIST;
401 } while (next != old_next);
404 // Atomically swap out the item's value in case another thread is updating the item while we are
405 // removing it. This establishes which operation occurs first logically, the update or the remove.
406 map_val_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
407 TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0);
410 find_preds(NULL, NULL, 0, sl, key, FORCE_UNLINK);
413 if (sl->key_type != NULL) {
414 rcu_defer_free((void *)item->key);
416 rcu_defer_free(item);
421 void sl_print (skiplist_t *sl, int verbose) {
424 for (int level = MAX_LEVELS - 1; level >= 0; --level) {
425 node_t *item = sl->head;
426 if (item->next[level] == DOES_NOT_EXIST)
428 printf("(%d) ", level);
431 markable_t next = item->next[level];
432 printf("%s%p ", HAS_MARK(next) ? "*" : "", item);
433 item = STRIP_MARK(next);
442 node_t *item = sl->head;
445 int is_marked = HAS_MARK(item->next[0]);
446 printf("%s%p:0x%llx ", is_marked ? "*" : "", item, (uint64_t)item->key);
447 if (item != sl->head) {
448 printf("[%d]", item->num_levels);
452 for (int level = 1; level < item->num_levels; ++level) {
453 node_t *next = STRIP_MARK(item->next[level]);
454 is_marked = HAS_MARK(item->next[0]);
455 printf(" %p%s", next, is_marked ? "*" : "");
456 if (item == sl->head && item->next[level] == DOES_NOT_EXIST)
461 item = STRIP_MARK(item->next[0]);
468 printf("levels:%-2d count:%-6lld \n", sl->high_water, (uint64_t)sl_count(sl));
471 sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) {
472 sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t));
473 if (key != DOES_NOT_EXIST) {
474 find_preds(NULL, &iter->next, 1, sl, key, DONT_UNLINK);
476 iter->next = GET_NODE(sl->head->next[0]);
481 map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) {
483 node_t *item = iter->next;
484 while (item != NULL && HAS_MARK(item->next[0])) {
485 item = STRIP_MARK(item->next[0]);
489 return DOES_NOT_EXIST;
491 iter->next = STRIP_MARK(item->next[0]);
492 if (key_ptr != NULL) {
493 *key_ptr = item->key;
498 void sl_iter_free (sl_iter_t *iter) {