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 probably won't work correctly on platforms with
14 * weaker memory models if you don't add memory barriers in the right places.
26 // Setting MAX_LEVEL to 0 essentially makes this data structure the Harris-Michael lock-free list
41 static int random_level (void) {
42 unsigned r = nbd_rand();
46 r |= 1 << (MAX_LEVEL+1);
48 int n = __builtin_ctz(r)-1;
49 assert(n <= MAX_LEVEL);
53 node_t *node_alloc (int level, const void *key_data, uint32_t key_len, uint64_t val) {
54 assert(level >= 0 && level <= MAX_LEVEL);
55 size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *);
56 node_t *item = (node_t *)nbd_malloc(sz);
58 // If <key_len> is -1 it indicates <key_data> is an integer and not a pointer
59 item->key = (key_len == (unsigned)-1)
60 ? (void *)TAG_VALUE(key_data)
61 : ns_alloc(key_data, key_len);
63 item->top_level = level;
67 static void node_free (node_t *item) {
68 if (!IS_TAGGED(item->key)) {
74 static void node_defer_free (node_t *item) {
75 if (!IS_TAGGED(item->key)) {
76 nbd_defer_free(item->key);
81 skiplist_t *sl_alloc (void) {
82 skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
83 sl->head = node_alloc(MAX_LEVEL, " ", 0, 0);
84 memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
88 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) {
89 node_t *pred = sl->head;
91 TRACE("s2", "find_preds: searching for key %p in sl (head is %p)", key_data, pred);
93 int start_level = MAX_LEVEL;
95 // Optimization for small lists. No need to traverse empty higher levels.
97 while (pred->next[start_level+1] != NULL) {
98 start_level += start_level - 1;
99 if (EXPECT_FALSE(start_level >= MAX_LEVEL)) {
100 start_level = MAX_LEVEL;
104 if (EXPECT_FALSE(start_level < n)) {
109 // Traverse the levels of <sl> from the top level to the bottom
110 for (int level = start_level; level >= 0; --level) {
111 TRACE("s3", "find_preds: level %llu", level, 0);
112 item = pred->next[level];
113 if (EXPECT_FALSE(IS_TAGGED(item))) {
114 TRACE("s2", "find_preds: pred %p is marked for removal (item %p); retry", pred, item);
115 return find_preds(preds, succs, n, sl, key_data, key_len, help_remove); // retry
117 while (item != NULL) {
118 node_t *next = item->next[level];
120 // A tag means an item is logically removed but not physically unlinked yet.
121 while (EXPECT_FALSE(IS_TAGGED(next))) {
123 // Skip over logically removed items.
125 item = (node_t *)STRIP_TAG(item->next);
126 if (EXPECT_FALSE(item == NULL))
128 TRACE("s3", "find_preds: skipping marked item %p (next is %p)", item, next);
129 next = item->next[level];
133 // Unlink logically removed items.
135 TRACE("s3", "find_preds: unlinking marked item %p; next is %p", item, next);
136 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) == item) {
137 item = (node_t *)STRIP_TAG(next);
138 if (EXPECT_FALSE(item == NULL))
140 next = item->next[level];
141 TRACE("s3", "find_preds: now the current item is %p next is %p", item, next);
143 // The thread that completes the unlink should free the memory.
144 if (level == 0) { node_defer_free(other); }
146 TRACE("s2", "find_preds: lost race to unlink item %p from pred %p", item, pred);
147 TRACE("s2", "find_preds: pred's link changed to %p", other, 0);
148 if (IS_TAGGED(other))
149 return find_preds(preds, succs, n, sl, key_data, key_len, help_remove); // retry
151 if (EXPECT_FALSE(item == NULL))
153 next = item->next[level];
157 if (EXPECT_FALSE(item == NULL))
160 TRACE("s3", "find_preds: visiting item %p (next is %p)", item, next);
161 TRACE("s4", "find_preds: key %p val %p", STRIP_TAG(item->key), item->val);
163 // A tagged key is an integer, otherwise it is a pointer to a string
164 if (IS_TAGGED(item->key)) {
165 d = (STRIP_TAG(item->key) - (uint64_t)key_data);
167 int item_key_len = item->key->len;
168 int len = (key_len < item_key_len) ? key_len : item_key_len;
169 d = memcmp(item->key->data, key_data, len);
170 if (d == 0) { d = item_key_len - key_len; }
174 TRACE("s2", "find_preds: found pred %p item %p", pred, item);
182 // The cast to unsigned is for the case when n is -1.
183 if ((unsigned)level <= (unsigned)n) {
193 // fill in empty levels
194 if (n == -1 && item != NULL) {
195 for (int level = start_level + 1; level <= item->top_level; ++level) {
196 preds[level] = sl->head;
201 TRACE("s2", "find_preds: found matching item %p in skiplist, pred is %p", item, pred);
204 TRACE("s2", "find_preds: found proper place for key %p in skiplist, pred is %p. returning null", key_data, pred);
208 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
209 uint64_t sl_lookup (skiplist_t *sl, const void *key_data, uint32_t key_len) {
210 TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key_data, sl);
211 node_t *item = find_preds(NULL, NULL, 0, sl, key_data, key_len, FALSE);
213 // If we found an <item> matching the <key> return its value.
215 uint64_t val = item->val;
216 if (val != DOES_NOT_EXIST) {
217 TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val);
222 TRACE("l1", "sl_lookup: no item in the skiplist matched the key", 0, 0);
223 return DOES_NOT_EXIST;
226 uint64_t sl_cas (skiplist_t *sl, const void *key_data, uint32_t key_len, uint64_t expectation, uint64_t new_val) {
227 TRACE("s1", "sl_cas: key %p skiplist %p", key_data, sl);
228 TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val);
229 ASSERT((int64_t)new_val > 0);
231 node_t *preds[MAX_LEVEL+1];
232 node_t *nexts[MAX_LEVEL+1];
233 node_t *new_item = NULL;
234 int n = random_level();
236 node_t *old_item = find_preds(preds, nexts, n, sl, key_data, key_len, TRUE);
237 if (old_item == NULL) {
239 // There was not an item in the skiplist that matches the key.
240 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == EXPECT_EXISTS)) {
241 TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0);
242 return DOES_NOT_EXIST; // failure
245 ASSERT(expectation == EXPECT_DOES_NOT_EXIST || expectation == EXPECT_WHATEVER);
247 // First insert <new_item> into the bottom level.
248 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]);
249 new_item = node_alloc(n, key_data, key_len, new_val);
250 node_t *pred = preds[0];
251 node_t *next = new_item->next[0] = nexts[0];
252 for (int level = 1; level <= new_item->top_level; ++level) {
253 new_item->next[level] = nexts[level];
255 node_t *other = SYNC_CAS(&pred->next[0], next, new_item);
257 TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0);
260 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
265 // Found an item in the skiplist that matches the key.
266 uint64_t old_item_val = old_item->val;
268 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
269 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
270 TRACE("s2", "sl_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
274 if (EXPECT_FALSE(expectation == EXPECT_DOES_NOT_EXIST)) {
275 TRACE("s1", "sl_cas: found an item %p in the skiplist that matched the key. the expectation was "
276 "not met, the skiplist was not changed", old_item, old_item_val);
277 return old_item_val; // failure
280 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
281 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
282 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
283 // fails it means another thread either removed the node or updated its value.
284 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
285 if (ret_val == old_item_val) {
286 TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0);
287 return ret_val; // success
289 TRACE("s2", "sl_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
291 old_item_val = ret_val;
295 // Link <new_item> into <sl> from the bottom up.
296 for (int level = 1; level <= new_item->top_level; ++level) {
297 node_t *pred = preds[level];
298 node_t *next = nexts[level];
300 TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next);
301 node_t *other = SYNC_CAS(&pred->next[level], next, new_item);
303 TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level);
306 TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other);
307 find_preds(preds, nexts, new_item->top_level, sl, key_data, key_len, TRUE);
311 // Update <new_item>'s next pointer
313 // There in no need to continue linking in the item if another thread removed it.
314 node_t *old_next = ((volatile node_t *)new_item)->next[level];
315 if (IS_TAGGED(old_next))
318 // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
319 if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
327 uint64_t sl_remove (skiplist_t *sl, const void *key_data, uint32_t key_len) {
328 TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key_data, sl);
329 node_t *preds[MAX_LEVEL+1];
330 node_t *item = find_preds(preds, NULL, -1, sl, key_data, key_len, TRUE);
332 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
333 return DOES_NOT_EXIST;
336 // Mark <item> removed at each level of <sl> from the top down. This must be atomic. If multiple threads
337 // try to remove the same item only one of them should succeed. Marking the bottom level establishes which of
339 for (int level = item->top_level; level >= 0; --level) {
340 if (EXPECT_FALSE(IS_TAGGED(item->next[level]))) {
341 TRACE("s3", "sl_remove: %p is already marked for removal by another thread", item, 0);
343 return DOES_NOT_EXIST;
347 node_t *old_next = item->next[level];
350 old_next = SYNC_CAS(&item->next[level], next, TAG_VALUE(next));
351 if (IS_TAGGED(old_next)) {
352 TRACE("s2", "sl_remove: lost race -- %p is already marked for removal by another thread", item, 0);
354 return DOES_NOT_EXIST;
356 } while (!IS_TAGGED(old_next) || next != old_next);
359 // This has to be an atomic swap in case another thread is updating the item while we are removing it.
360 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
361 TRACE("s2", "sl_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
363 // Unlink <item> from <sl>. If we lose a race to another thread just back off. It is safe to leave the
364 // item partially unlinked for a later call (or some other thread) to physically unlink. By marking the
365 // item earlier, we logically removed it.
366 int level = item->top_level;
368 node_t *pred = preds[level];
369 node_t *next = item->next[level];
370 TRACE("s2", "sl_remove: unlink the item by linking its pred %p to it's successor %p", pred, STRIP_TAG(next));
371 node_t *other = NULL;
372 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) != item) {
373 TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
379 // The thread that completes the unlink should free the memory.
380 TRACE("s1", "sl_remove: successfully unlinked item %p from the skiplist", item, 0);
381 node_defer_free(item);
385 void sl_print (skiplist_t *sl) {
386 for (int level = MAX_LEVEL; level >= 0; --level) {
387 node_t *item = sl->head;
388 if (item->next[level] == NULL)
390 printf("(%d) ", level);
392 node_t *next = item->next[level];
393 printf("%s%p ", IS_TAGGED(next) ? "*" : "", item);
394 item = (node_t *)STRIP_TAG(next);
401 node_t *item = sl->head;
403 int is_marked = IS_TAGGED(item->next[0]);
405 if (IS_TAGGED(item->key)) {
406 printf("%s%p:%llx ", is_marked ? "*" : "", item, STRIP_TAG(item->key));
408 printf("%s%p:%s ", is_marked ? "*" : "", item, (char *)item->key->data);
410 if (item != sl->head) {
411 printf("[%d]", item->top_level);
415 for (int level = 1; level <= item->top_level; ++level) {
416 node_t *next = (node_t *)STRIP_TAG(item->next[level]);
417 is_marked = IS_TAGGED(item->next[0]);
418 printf(" %p%s", next, is_marked ? "*" : "");
419 if (item == sl->head && item->next[level] == NULL)
424 item = (node_t *)STRIP_TAG(item->next[0]);