2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
5 * C implementation of the lock-free skiplist data-structure created by Maurice Herlihy,
6 * Yossi Lev, and Nir Shavit. See "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 depends on certain stores and loads being ordered. Be careful on non-x86 platforms
13 * with weaker memory-models. This code probably won't work without adding some memory barriers.
31 typedef struct skiplist {
37 static int random_level (int r) {
40 int n = __builtin_ctz(r);
46 node_t *node_alloc (int top_level, uint64_t key, uint64_t value) {
47 assert(top_level >= 0 && top_level <= MAX_LEVEL);
48 size_t sz = sizeof(node_t) + (top_level + 1) * sizeof(node_t *);
49 node_t *item = (node_t *)nbd_malloc(sz);
53 item->top_level = top_level;
57 skiplist_t *sl_alloc (void) {
58 skiplist_t *skiplist = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
59 skiplist->head = node_alloc(MAX_LEVEL, 0, 0);
60 skiplist->last = node_alloc(MAX_LEVEL, (uint64_t)-1, 0);
61 for (int level = 0; level <= MAX_LEVEL; ++level) {
62 skiplist->head->next[level] = skiplist->last;
67 static node_t *find_preds (node_t **preds, skiplist_t *skiplist, uint64_t key, int help_remove) {
68 node_t *pred = skiplist->head;
70 TRACE("s3", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
75 // Traverse the levels of the skiplist from the top level to the bottom
76 for (int level = MAX_LEVEL; level >= 0; --level) {
77 TRACE("s3", "find_preds: level %llu", level, 0);
78 item = pred->next[level];
79 if (IS_TAGGED(item)) {
80 TRACE("s3", "find_preds: pred %p is marked for removal (item %p); retry", pred, item);
81 return find_preds(preds, skiplist, key, help_remove); // retry
84 node_t *next = item->next[level];
85 TRACE("s3", "find_preds: visiting item %p (next %p)", item, next);
86 TRACE("s3", "find_preds: key %p", item->key, 0);
88 // Marked items are logically removed, but not fully unlinked yet.
89 while (EXPECT_FALSE(IS_TAGGED(next))) {
91 // Skip over partially removed items.
93 item = (node_t *)STRIP_TAG(item->next);
94 next = item->next[level];
98 // Unlink partially removed items.
100 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) == item) {
101 item = (node_t *)STRIP_TAG(next);
102 next = item->next[level];
103 TRACE("s3", "find_preds: unlinked item %p from pred %p", item, pred);
104 TRACE("s3", "find_preds: now item is %p next is %p", item, next);
106 // The thread that completes the unlink should free the memory.
107 if (level == 0) { nbd_defer_free(other); }
109 TRACE("s3", "find_preds: lost race to unlink item from pred %p; its link changed to %p",pred,other);
110 if (IS_TAGGED(other))
111 return find_preds(preds, skiplist, key, help_remove); // retry
113 next = item->next[level];
117 // If we reached the key (or passed where it should be), we found a pred. Save it and continue down.
118 if (item->key >= key) {
119 TRACE("s3", "find_preds: found pred %p item %p", pred, item);
126 assert(count++ < 18);
135 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
136 uint64_t sl_lookup (skiplist_t *skiplist, uint64_t key) {
137 TRACE("s3", "sl_lookup: searching for key %p in skiplist %p", key, skiplist);
138 node_t *item = find_preds(NULL, skiplist, key, FALSE);
140 // If we found an <item> matching the <key> return its value.
141 return (item->key == key) ? item->value : DOES_NOT_EXIST;
144 // Insert the <key> if it doesn't already exist in the <skiplist>
145 uint64_t sl_add_r (int r, skiplist_t *skiplist, uint64_t key, uint64_t value) {
146 TRACE("s3", "sl_add: inserting key %p value %p", key, value);
147 node_t *preds[MAX_LEVEL+1];
150 node_t *next = find_preds(preds, skiplist, key, TRUE);
152 // If a node matching <key> already exists in the skiplist, return its value.
153 if (next->key == key) {
154 TRACE("s3", "sl_add: there is already an item %p (value %p) with the same key", next, next->value);
155 if (EXPECT_FALSE(item != NULL)) { nbd_free(item); }
159 // First insert <item> into the bottom level.
160 if (EXPECT_TRUE(item == NULL)) { item = node_alloc(random_level(r), key, value); }
161 TRACE("s3", "sl_add: attempting to insert item between %p and %p", preds[0], next);
162 item->next[0] = next;
163 for (int level = 1; level <= item->top_level; ++level) {
164 item->next[level] = preds[level]->next[level];
166 node_t *other = SYNC_CAS(&preds[0]->next[0], next, item);
168 TRACE("s3", "sl_add: successfully inserted item %p at level 0", item, 0);
171 TRACE("s3", "sl_add: failed to change pred's link: expected %p found %p", next, other);
175 // Insert <item> into the skiplist from the bottom level up.
176 for (int level = 1; level <= item->top_level; ++level) {
178 node_t *pred = preds[level];
179 node_t *next = pred->next[level];
180 while (EXPECT_FALSE(IS_TAGGED(next) || next->key < key)) {
181 find_preds(preds, skiplist, key, TRUE);
183 next = pred->next[level];
187 // There in no need to continue linking in the item if another thread removed it.
188 node_t *old_next = ((volatile node_t *)item)->next[level];
189 if (IS_TAGGED(old_next))
190 return DOES_NOT_EXIST; // success
192 // Use a CAS so we to not inadvertantly remove a mark another thread placed on the item.
193 if (next == old_next || SYNC_CAS(&item->next[level], old_next, next) == old_next)
197 TRACE("s3", "sl_add: attempting to insert item between %p and %p", pred, next);
198 node_t *other = SYNC_CAS(&pred->next[level], next, item);
200 TRACE("s3", "sl_add: successfully inserted item %p at level %llu", item, level);
203 TRACE("s3", "sl_add: failed to change pred's link: expected %p found %p", next, other);
210 uint64_t sl_remove (skiplist_t *skiplist, uint64_t key) {
211 TRACE("s3", "sl_remove: removing item with key %p from skiplist %p", key, skiplist);
212 node_t *preds[MAX_LEVEL+1];
213 node_t *item = find_preds(preds, skiplist, key, TRUE);
214 if (item->key != key) {
215 TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
216 return DOES_NOT_EXIST;
219 // Mark <item> removed at each level of the skiplist from the top down. This must be atomic. If multiple threads
220 // try to remove the same item only one of them should succeed. Marking the bottom level establishes which of
222 for (int level = item->top_level; level >= 0; --level) {
223 if (EXPECT_FALSE(IS_TAGGED(item->next[level]))) {
224 TRACE("s3", "sl_remove: %p is already marked for removal by another thread", item, 0);
226 return DOES_NOT_EXIST;
229 node_t *next = SYNC_FETCH_AND_OR(&item->next[level], TAG);
230 if (EXPECT_FALSE(IS_TAGGED(next))) {
231 TRACE("s3", "sl_remove: lost race -- %p is already marked for removal by another thread", item, 0);
233 return DOES_NOT_EXIST;
238 uint64_t value = item->value;
240 // Unlink <item> from the top down.
241 int level = item->top_level;
243 node_t *pred = preds[level];
244 node_t *next = item->next[level];
245 TRACE("s3", "sl_remove: link item's pred %p to it's successor %p", pred, STRIP_TAG(next));
246 node_t *other = NULL;
247 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) != item) {
248 TRACE("s3", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
249 // By marking the item earlier, we logically removed it. It is safe to leave the item partially
250 // unlinked. Another thread will finish physically removing it from the skiplist.
256 // The thread that completes the unlink should free the memory.
257 nbd_defer_free(item);
261 void sl_print (skiplist_t *skiplist) {
262 for (int level = MAX_LEVEL; level >= 0; --level) {
263 node_t *item = skiplist->head;
264 if (item->next[level] == skiplist->last)
266 printf("(%d) ", level);
268 node_t *next = item->next[level];
269 printf("%s%p:0x%llx ", IS_TAGGED(next) ? "*" : "", item, item->key);
270 item = (node_t *)STRIP_TAG(next);
277 node_t *item = skiplist->head;
279 assert(item->top_level <= MAX_LEVEL);
280 int is_marked = IS_TAGGED(item->next[0]);
281 printf("%s%p:0x%llx (%d", is_marked ? "*" : "", item, item->key, item->top_level);
282 for (int level = 1; level <= item->top_level; ++level) {
283 node_t *next = (node_t *)STRIP_TAG(item->next[level]);
284 printf(" %p:0x%llx", item->next[level], next ? next->key : 0);
285 if (item == skiplist->head && item->next[level] == skiplist->last)
287 if (item == skiplist->last && item->next[level] == NULL)
292 item = (node_t *)STRIP_TAG(item->next[0]);
296 #ifdef MAKE_skiplist_test
299 #include <sys/time.h>
303 #define NUM_ITERATIONS 10000000
305 static volatile int wait_;
306 static long num_threads_;
307 static skiplist_t *sl_;
309 void *worker (void *arg) {
310 int id = (int)(size_t)arg;
312 unsigned int rand_seed = id+1;//rdtsc_l();
314 // Wait for all the worker threads to be ready.
315 SYNC_ADD(&wait_, -1);
318 for (int i = 0; i < NUM_ITERATIONS/num_threads_; ++i) {
319 int n = rand_r(&rand_seed);
320 int key = (n & 0xF) + 1;
322 sl_add_r(n, sl_, key, 1);
333 int main (int argc, char **argv) {
335 lwt_set_trace_level("s3");
337 char* program_name = argv[0];
338 pthread_t thread[MAX_NUM_THREADS];
341 fprintf(stderr, "Usage: %s num_threads\n", program_name);
349 num_threads_ = strtol(argv[1], NULL, 10);
351 fprintf(stderr, "%s: Invalid argument for number of threads\n", program_name);
354 if (num_threads_ <= 0) {
355 fprintf(stderr, "%s: Number of threads must be at least 1\n", program_name);
358 if (num_threads_ > MAX_NUM_THREADS) {
359 fprintf(stderr, "%s: Number of threads cannot be more than %d\n", program_name, MAX_NUM_THREADS);
366 struct timeval tv1, tv2;
367 gettimeofday(&tv1, NULL);
369 wait_ = num_threads_;
371 for (int i = 0; i < num_threads_; ++i) {
372 int rc = nbd_thread_create(thread + i, i, worker, (void*)(size_t)i);
373 if (rc != 0) { perror("pthread_create"); return rc; }
376 for (int i = 0; i < num_threads_; ++i) {
377 pthread_join(thread[i], NULL);
380 gettimeofday(&tv2, NULL);
381 int ms = (int)(1000000*(tv2.tv_sec - tv1.tv_sec) + tv2.tv_usec - tv1.tv_usec) / 1000;
382 printf("Th:%ld Time:%dms\n", num_threads_, ms);
388 #endif//skiplist_test