2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
23 typedef struct skiplist {
29 static int random_level (int r) {
32 int n = __builtin_ctz(r);
38 node_t *node_alloc (int top_level, uint64_t key, uint64_t value) {
39 assert(top_level >= 0 && top_level <= MAX_LEVEL);
40 size_t sz = sizeof(node_t) + (top_level + 1) * sizeof(node_t *);
41 node_t *item = (node_t *)nbd_malloc(sz);
45 item->top_level = top_level;
49 skiplist_t *skiplist_alloc (void) {
50 skiplist_t *skiplist = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
51 skiplist->head = node_alloc(MAX_LEVEL, 0, 0);
52 skiplist->last = node_alloc(MAX_LEVEL, (uint64_t)-1, 0);
53 for (int level = 0; level <= MAX_LEVEL; ++level) {
54 skiplist->head->next[level] = skiplist->last;
59 static node_t *find_preds (node_t **preds, skiplist_t *skiplist, uint64_t key, int help_remove) {
60 node_t *pred = skiplist->head;
62 TRACE("s3", "find_preds: searching for key %p in skiplist (head is %p)", key, pred);
67 // Traverse the levels of the skiplist from the top level to the bottom
68 for (int level = MAX_LEVEL; level >= 0; --level) {
69 TRACE("s3", "find_preds: level %llu", level, 0);
70 item = pred->next[level];
71 if (IS_TAGGED(item)) {
72 TRACE("s3", "find_preds: pred %p is marked for removal (item %p); retry", pred, item);
73 return find_preds(preds, skiplist, key, help_remove); // retry
76 node_t *next = item->next[level];
77 TRACE("s3", "find_preds: visiting item %p (next %p)", item, next);
78 TRACE("s3", "find_preds: key %p", item->key, 0);
80 // Marked items are logically removed, but not fully unlinked yet.
81 while (EXPECT_FALSE(IS_TAGGED(next))) {
83 // Skip over partially removed items.
85 item = (node_t *)STRIP_TAG(item->next);
86 next = item->next[level];
90 // Unlink partially removed items.
92 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) == item) {
93 item = (node_t *)STRIP_TAG(next);
94 next = item->next[level];
95 TRACE("s3", "find_preds: unlinked item %p from pred %p", item, pred);
96 TRACE("s3", "find_preds: now item is %p next is %p", item, next);
98 // The thread that completes the unlink should free the memory.
99 if (level == 0) { nbd_defer_free(other); }
101 TRACE("s3", "find_preds: lost race to unlink item from pred %p; its link changed to %p",pred,other);
102 if (IS_TAGGED(other))
103 return find_preds(preds, skiplist, key, help_remove); // retry
105 next = item->next[level];
109 // If we reached the key (or passed where it should be), we found a pred. Save it and continue down.
110 if (item->key >= key) {
111 TRACE("s3", "find_preds: found pred %p item %p", pred, item);
118 assert(count++ < 18);
127 // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors.
128 uint64_t skiplist_lookup (skiplist_t *skiplist, uint64_t key) {
129 TRACE("s3", "skiplist_lookup: searching for key %p in skiplist %p", key, skiplist);
130 node_t *item = find_preds(NULL, skiplist, key, FALSE);
132 // If we found an <item> matching the <key> return its value.
133 return (item->key == key) ? item->value : DOES_NOT_EXIST;
136 // Insert the <key> if it doesn't already exist in the <skiplist>
137 uint64_t skiplist_add_i (int n, skiplist_t *skiplist, uint64_t key, uint64_t value) {
138 TRACE("s3", "skiplist_add: inserting key %p value %p", key, value);
139 node_t *preds[MAX_LEVEL+1];
142 node_t *next = find_preds(preds, skiplist, key, TRUE);
144 // If a node matching <key> already exists in the skiplist, return its value.
145 if (next->key == key) {
146 TRACE("s3", "skiplist_add: there is already an item %p (value %p) with the same key", next, next->value);
147 if (EXPECT_FALSE(item != NULL)) { nbd_free(item); }
151 // First insert <item> into the bottom level.
152 if (EXPECT_TRUE(item == NULL)) { item = node_alloc(random_level(n), key, value); }
153 TRACE("s3", "skiplist_add: attempting to insert item between %p and %p", preds[0], next);
154 item->next[0] = next;
155 for (int level = 1; level <= item->top_level; ++level) {
156 item->next[level] = preds[level]->next[level];
158 node_t *other = SYNC_CAS(&preds[0]->next[0], next, item);
160 TRACE("s3", "skiplist_add: successfully inserted item %p at level 0", item, 0);
163 TRACE("s3", "skiplist_add: failed to change pred's link: expected %p found %p", next, other);
167 // Insert <item> into the skiplist from the bottom level up.
168 for (int level = 1; level <= item->top_level; ++level) {
170 node_t *pred = preds[level];
171 node_t *next = pred->next[level];
172 while (EXPECT_FALSE(IS_TAGGED(next) || next->key < key)) {
173 find_preds(preds, skiplist, key, TRUE);
175 next = pred->next[level];
179 // There in no need to continue linking in the item if another thread removed it.
180 node_t *old_next = ((volatile node_t *)item)->next[level];
181 if (IS_TAGGED(old_next))
182 return DOES_NOT_EXIST; // success
184 // Use a CAS so we to not inadvertantly remove a mark another thread placed on the item.
185 if (next == old_next || SYNC_CAS(&item->next[level], old_next, next) == old_next)
189 TRACE("s3", "skiplist_add: attempting to insert item between %p and %p", pred, next);
190 node_t *other = SYNC_CAS(&pred->next[level], next, item);
192 TRACE("s3", "skiplist_add: successfully inserted item %p at level %llu", item, level);
195 TRACE("s3", "skiplist_add: failed to change pred's link: expected %p found %p", next, other);
202 uint64_t skiplist_remove (skiplist_t *skiplist, uint64_t key) {
203 TRACE("s3", "skiplist_remove: removing item with key %p from skiplist %p", key, skiplist);
204 node_t *preds[MAX_LEVEL+1];
205 node_t *item = find_preds(preds, skiplist, key, TRUE);
206 if (item->key != key) {
207 TRACE("s3", "skiplist_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
208 return DOES_NOT_EXIST;
211 // Mark <item> removed at each level of the skiplist from the top down. This must be atomic. If multiple threads
212 // try to remove the same item only one of them should succeed. Marking the bottom level establishes which of
214 for (int level = item->top_level; level >= 0; --level) {
215 if (EXPECT_FALSE(IS_TAGGED(item->next[level]))) {
216 TRACE("s3", "skiplist_remove: %p is already marked for removal by another thread", item, 0);
218 return DOES_NOT_EXIST;
221 node_t *next = SYNC_FETCH_AND_OR(&item->next[level], TAG);
222 if (EXPECT_FALSE(IS_TAGGED(next))) {
223 TRACE("s3", "skiplist_remove: lost race -- %p is already marked for removal by another thread", item, 0);
225 return DOES_NOT_EXIST;
230 uint64_t value = item->value;
232 // Unlink <item> from the top down.
233 int level = item->top_level;
235 node_t *pred = preds[level];
236 node_t *next = item->next[level];
237 TRACE("s3", "skiplist_remove: link item's pred %p to it's successor %p", pred, STRIP_TAG(next));
238 node_t *other = NULL;
239 if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) != item) {
240 TRACE("s3", "skiplist_remove: unlink failed; pred's link changed from %p to %p", item, other);
241 // By marking the item earlier, we logically removed it. It is safe to leave the item partially
242 // unlinked. Another thread will finish physically removing it from the skiplist.
248 // The thread that completes the unlink should free the memory.
249 nbd_defer_free(item);
253 void skiplist_print (skiplist_t *skiplist) {
254 for (int level = MAX_LEVEL; level >= 0; --level) {
255 printf("(%d) ", level);
256 node_t *item = skiplist->head;
258 node_t *next = item->next[level];
259 printf("%s%p:0x%llx ", IS_TAGGED(next) ? "*" : "", item, item->key);
260 item = (node_t *)STRIP_TAG(next);
266 node_t *item = skiplist->head;
268 assert(item->top_level <= MAX_LEVEL);
269 int is_marked = IS_TAGGED(item->next[0]);
270 printf("%s%p:0x%llx (%d", is_marked ? "*" : "", item, item->key, item->top_level);
271 for (int i = 1; i <= item->top_level; ++i) {
272 node_t *next = (node_t *)STRIP_TAG(item->next[i]);
273 printf(" %p:0x%llx", item->next[i], next ? next->key : 0);
277 item = (node_t *)STRIP_TAG(item->next[0]);
281 #ifdef MAKE_skiplist_test
284 #include <sys/time.h>
288 #define NUM_ITERATIONS 10000000
290 static volatile int wait_;
291 static long num_threads_;
292 static skiplist_t *skiplist_;
294 void *worker (void *arg) {
295 int id = (int)(size_t)arg;
297 unsigned int rand_seed = id+1;//rdtsc_l();
299 // Wait for all the worker threads to be ready.
300 SYNC_ADD(&wait_, -1);
303 for (int i = 0; i < NUM_ITERATIONS/num_threads_; ++i) {
304 int n = rand_r(&rand_seed);
305 int key = (n & 0xF) + 1;
307 skiplist_add_i(n, skiplist_, key, 1);
309 skiplist_remove(skiplist_, key);
318 int main (int argc, char **argv) {
320 lwt_set_trace_level("s3");
322 char* program_name = argv[0];
323 pthread_t thread[MAX_NUM_THREADS];
326 fprintf(stderr, "Usage: %s num_threads\n", program_name);
334 num_threads_ = strtol(argv[1], NULL, 10);
336 fprintf(stderr, "%s: Invalid argument for number of threads\n", program_name);
339 if (num_threads_ <= 0) {
340 fprintf(stderr, "%s: Number of threads must be at least 1\n", program_name);
343 if (num_threads_ > MAX_NUM_THREADS) {
344 fprintf(stderr, "%s: Number of threads cannot be more than %d\n", program_name, MAX_NUM_THREADS);
349 skiplist_ = skiplist_alloc();
351 struct timeval tv1, tv2;
352 gettimeofday(&tv1, NULL);
354 wait_ = num_threads_;
356 for (int i = 0; i < num_threads_; ++i) {
357 int rc = nbd_thread_create(thread + i, i, worker, (void*)(size_t)i);
358 if (rc != 0) { perror("pthread_create"); return rc; }
361 for (int i = 0; i < num_threads_; ++i) {
362 pthread_join(thread[i], NULL);
365 gettimeofday(&tv2, NULL);
366 int ms = (int)(1000000*(tv2.tv_sec - tv1.tv_sec) + tv2.tv_usec - tv1.tv_usec) / 1000;
367 printf("Th:%ld Time:%dms\n", num_threads_, ms);
368 skiplist_print(skiplist_);
373 #endif//skiplist_test