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 * This code is written for the x86 memory-model. The algorithim depends on certain stores and
13 * loads being ordered. Be careful, this code won't work correctly on platforms with weaker memory
14 * models if you don't add memory barriers in the right places.
27 // Setting MAX_LEVEL to 0 essentially makes this data structure the Harris-Michael lock-free list
42 static const map_impl_t sl_map_impl = {
43 (map_alloc_t)sl_alloc, (map_cas_t)sl_cas, (map_get_t)sl_lookup, (map_remove_t)sl_remove,
44 (map_count_t)sl_count, (map_print_t)sl_print, (map_free_t)sl_free
47 const map_impl_t *MAP_TYPE_SKIPLIST = &sl_map_impl;
49 static int random_level (void) {
50 unsigned r = nbd_rand();
54 r |= 1 << (MAX_LEVEL+1);
56 int n = __builtin_ctz(r)-1;
57 assert(n <= MAX_LEVEL);
61 static node_t *node_alloc (int level, const void *key_data, uint32_t key_len, uint64_t val) {
62 assert(level >= 0 && level <= MAX_LEVEL);
63 size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *);
64 node_t *item = (node_t *)nbd_malloc(sz);
66 // If <key_len> is -1 it indicates <key_data> is an integer and not a pointer
67 item->key = (key_len == (unsigned)-1)
68 ? (void *)TAG_VALUE(key_data)
69 : ns_alloc(key_data, key_len);
71 item->top_level = level;
75 static void node_free (node_t *item) {
76 if (!IS_TAGGED(item->key)) {
82 static void node_defer_free (node_t *item) {
83 if (!IS_TAGGED(item->key)) {
84 nbd_defer_free(item->key);
89 skiplist_t *sl_alloc (void) {
90 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
91 sl->head = node_alloc(MAX_LEVEL, " ", 0, 0);
92 memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
96 void sl_free (skiplist_t *sl) {
97 node_t *item = sl->head->next[0];
99 node_t *next = (node_t *)STRIP_TAG(item->next[0]);
105 uint64_t sl_count (skiplist_t *sl) {
107 node_t *item = sl->head->next[0];
110 item = (node_t *)STRIP_TAG(item->next[0]);
115 static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, const void *key_data, uint32_t key_len, int help_remove) {
116 node_t *pred = sl->head;
118 TRACE("s2", "find_preds: searching for key %p in sl (head is %p)", key_data, 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] != NULL) {
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 item = pred->next[level];
140 if (EXPECT_FALSE(IS_TAGGED(item))) {
141 TRACE("s2", "find_preds: pred %p is marked for removal (item %p); retry", pred, item);
142 return find_preds(preds, succs, n, sl, key_data, key_len, help_remove); // retry
144 while (item != NULL) {
145 node_t *next = item->next[level];
147 // A tag means an item is logically removed but not physically unlinked yet.
148 while (EXPECT_FALSE(IS_TAGGED(next))) {
150 // Skip over logically removed items.
152 item = (node_t *)STRIP_TAG(item->next);
153 if (EXPECT_FALSE(item == NULL))
155 TRACE("s3", "find_preds: skipping marked item %p (next is %p)", item, next);
156 next = item->next[level];
160 // Unlink logically removed items.
162 TRACE("s3", "find_preds: unlinking marked item %p; next is %p", item, next);
163 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) == item) {
164 item = (node_t *)STRIP_TAG(next);
165 if (EXPECT_FALSE(item == NULL))
167 next = item->next[level];
168 TRACE("s3", "find_preds: now the current item is %p next is %p", item, next);
170 // The thread that completes the unlink should free the memory.
171 if (level == 0) { node_defer_free(other); }
173 TRACE("s2", "find_preds: lost race to unlink item %p from pred %p", item, pred);
174 TRACE("s2", "find_preds: pred's link changed to %p", other, 0);
175 if (IS_TAGGED(other))
176 return find_preds(preds, succs, n, sl, key_data, key_len, help_remove); // retry
178 if (EXPECT_FALSE(item == NULL))
180 next = item->next[level];
184 if (EXPECT_FALSE(item == NULL))
187 TRACE("s3", "find_preds: visiting item %p (next is %p)", item, next);
188 TRACE("s4", "find_preds: key %p val %p", STRIP_TAG(item->key), item->val);
190 // A tagged key is an integer, otherwise it is a pointer to a string
191 if (IS_TAGGED(item->key)) {
192 d = (STRIP_TAG(item->key) - (uint64_t)key_data);
194 int item_key_len = item->key->len;
195 int len = (key_len < item_key_len) ? key_len : item_key_len;
196 d = memcmp(item->key->data, key_data, len);
197 if (d == 0) { d = item_key_len - key_len; }
201 TRACE("s2", "find_preds: found pred %p item %p", pred, item);
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 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_data, pred);
235 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
236 uint64_t sl_lookup (skiplist_t *sl, const void *key_data, uint32_t key_len) {
237 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key_data, sl);
238 node_t *item = find_preds(NULL, NULL, 0, sl, key_data, key_len, FALSE);
240 // If we found an <item> matching the <key> return its value.
242 uint64_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 uint64_t sl_cas (skiplist_t *sl, const void *key_data, uint32_t key_len, uint64_t expectation, uint64_t new_val) {
254 TRACE("s1", "sl_cas: key %p skiplist %p", key_data, sl);
255 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
256 ASSERT((int64_t)new_val > 0);
258 node_t *preds[MAX_LEVEL+1];
259 node_t *nexts[MAX_LEVEL+1];
260 node_t *new_item = NULL;
261 int n = random_level();
263 node_t *old_item = find_preds(preds, nexts, n, sl, key_data, key_len, TRUE);
264 if (old_item == NULL) {
266 // There was not an item in the skiplist that matches the key.
267 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
268 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
269 return DOES_NOT_EXIST; // failure
272 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
274 // First insert <new_item> into the bottom level.
275 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]);
276 new_item = node_alloc(n, key_data, key_len, new_val);
277 node_t *pred = preds[0];
278 node_t *next = new_item->next[0] = nexts[0];
279 for (int level = 1; level <= new_item->top_level; ++level) {
280 new_item->next[level] = nexts[level];
282 node_t *other = SYNC_CAS(&pred->next[0], next, new_item);
284 TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0);
287 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
292 // Found an item in the skiplist that matches the key.
293 uint64_t old_item_val = old_item->val;
295 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
296 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
297 TRACE("s2", "sl_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
301 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
302 TRACE("s1", "sl_cas: found an item %p in the skiplist that matched the key. the expectation was "
303 "not met, the skiplist was not changed", old_item, old_item_val);
304 return old_item_val; // failure
307 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
308 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
309 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
310 // fails it means another thread either removed the node or updated its value.
311 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
312 if (ret_val == old_item_val) {
313 TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0);
314 return ret_val; // success
316 TRACE("s2", "sl_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
318 old_item_val = ret_val;
322 // Link <new_item> into <sl> from the bottom up.
323 for (int level = 1; level <= new_item->top_level; ++level) {
324 node_t *pred = preds[level];
325 node_t *next = nexts[level];
327 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next);
328 node_t *other = SYNC_CAS(&pred->next[level], next, new_item);
330 TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level);
333 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
334 find_preds(preds, nexts, new_item->top_level, sl, key_data, key_len, TRUE);
338 // Update <new_item>'s next pointer
340 // There in no need to continue linking in the item if another thread removed it.
341 node_t *old_next = ((volatile node_t *)new_item)->next[level];
342 if (IS_TAGGED(old_next))
345 // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
346 if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
354 uint64_t sl_remove (skiplist_t *sl, const void *key_data, uint32_t key_len) {
355 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key_data, sl);
356 node_t *preds[MAX_LEVEL+1];
357 node_t *item = find_preds(preds, NULL, -1, sl, key_data, key_len, TRUE);
359 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
360 return DOES_NOT_EXIST;
363 // Mark <item> removed at each level of <sl> from the top down. This must be atomic. If multiple threads
364 // try to remove the same item only one of them should succeed. Marking the bottom level establishes which of
366 for (int level = item->top_level; level >= 0; --level) {
367 if (EXPECT_FALSE(IS_TAGGED(item->next[level]))) {
368 TRACE("s3", "sl_remove: %p is already marked for removal by another thread", item, 0);
370 return DOES_NOT_EXIST;
374 node_t *old_next = item->next[level];
377 old_next = SYNC_CAS(&item->next[level], next, TAG_VALUE(next));
378 if (IS_TAGGED(old_next)) {
379 TRACE("s2", "sl_remove: lost race -- %p is already marked for removal by another thread", item, 0);
381 return DOES_NOT_EXIST;
383 } while (!IS_TAGGED(old_next) || next != old_next);
386 // This has to be an atomic swap in case another thread is updating the item while we are removing it.
387 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
388 TRACE("s2", "sl_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
390 // Unlink <item> from <sl>. If we lose a race to another thread just back off. It is safe to leave the
391 // item partially unlinked for a later call (or some other thread) to physically unlink. By marking the
392 // item earlier, we logically removed it.
393 int level = item->top_level;
395 node_t *pred = preds[level];
396 node_t *next = item->next[level];
397 TRACE("s2", "sl_remove: unlink the item by linking its pred %p to it's successor %p", pred, STRIP_TAG(next));
398 node_t *other = NULL;
399 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) != item) {
400 TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
406 // The thread that completes the unlink should free the memory.
407 TRACE("s1", "sl_remove: successfully unlinked item %p from the skiplist", item, 0);
408 node_defer_free(item);
412 void sl_print (skiplist_t *sl) {
413 for (int level = MAX_LEVEL; level >= 0; --level) {
414 node_t *item = sl->head;
415 if (item->next[level] == NULL)
417 printf("(%d) ", level);
419 node_t *next = item->next[level];
420 printf("%s%p ", IS_TAGGED(next) ? "*" : "", item);
421 item = (node_t *)STRIP_TAG(next);
428 node_t *item = sl->head;
430 int is_marked = IS_TAGGED(item->next[0]);
432 if (IS_TAGGED(item->key)) {
433 printf("%s%p:%llx ", is_marked ? "*" : "", item, STRIP_TAG(item->key));
435 printf("%s%p:%s ", is_marked ? "*" : "", item, (char *)item->key->data);
437 if (item != sl->head) {
438 printf("[%d]", item->top_level);
442 for (int level = 1; level <= item->top_level; ++level) {
443 node_t *next = (node_t *)STRIP_TAG(item->next[level]);
444 is_marked = IS_TAGGED(item->next[0]);
445 printf(" %p%s", next, is_marked ? "*" : "");
446 if (item == sl->head && item->next[level] == NULL)
451 item = (node_t *)STRIP_TAG(item->next[0]);