2 * Written by Josh Dybnis and released to the public domain, as explained at
3 * http://creativecommons.org/licenses/publicdomain
5 * Harris-Michael lock-free list-based set
6 * http://www.research.ibm.com/people/m/michael/spaa-2002.pdf
25 clone_fun_t clone_fun;
28 static const map_impl_t ll_map_impl = {
29 (map_alloc_t)ll_alloc, (map_cas_t)ll_cas, (map_get_t)ll_lookup, (map_remove_t)ll_remove,
30 (map_count_t)ll_count, (map_print_t)ll_print, (map_free_t)ll_free
33 const map_impl_t *MAP_TYPE_LIST = &ll_map_impl;
35 static node_t *node_alloc (void *key, uint64_t val) {
36 node_t *item = (node_t *)nbd_malloc(sizeof(node_t));
42 list_t *ll_alloc (cmp_fun_t cmp_fun, hash_fun_t hash_fun, clone_fun_t clone_fun) {
43 list_t *ll = (list_t *)nbd_malloc(sizeof(list_t));
44 ll->cmp_fun = cmp_fun;
45 ll->clone_fun = clone_fun;
46 ll->head = node_alloc(NULL, 0);
47 ll->head->next = NULL;
51 void ll_free (list_t *ll) {
52 node_t *item = ll->head->next;
54 node_t *next = (node_t *)STRIP_TAG(item->next);
60 uint64_t ll_count (list_t *ll) {
62 node_t *item = ll->head->next;
65 item = (node_t *)STRIP_TAG(item->next);
70 static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, void *key, int help_remove) {
71 node_t *pred = ll->head;
72 node_t *item = pred->next;
73 TRACE("l2", "find_pred: searching for key %p in list (head is %p)", key, pred);
75 while (item != NULL) {
76 node_t *next = item->next;
78 // A tag means an item is logically removed but not physically unlinked yet.
79 while (EXPECT_FALSE(IS_TAGGED(next))) {
81 // Skip over logically removed items.
83 item = (node_t *)STRIP_TAG(item->next);
84 if (EXPECT_FALSE(item == NULL))
86 TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
91 // Unlink logically removed items.
93 TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
94 if ((other = SYNC_CAS(&pred->next, item, STRIP_TAG(next))) == item) {
95 TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
96 item = (node_t *)STRIP_TAG(next);
97 if (EXPECT_FALSE(item == NULL))
100 TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
102 // The thread that completes the unlink should free the memory.
103 if (ll->clone_fun != NULL) {
104 nbd_defer_free((void*)other->key);
106 nbd_defer_free(other);
108 TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
109 TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
110 if (IS_TAGGED(other))
111 return find_pred(pred_ptr, item_ptr, ll, key, help_remove); // retry
113 if (EXPECT_FALSE(item == NULL))
119 if (EXPECT_FALSE(item == NULL))
122 TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
123 TRACE("l4", "find_pred: key %p val %p", item->key, item->val);
126 if (EXPECT_TRUE(ll->cmp_fun == NULL)) {
127 d = (uint64_t)item->key - (uint64_t)key;
129 d = ll->cmp_fun(item->key, key);
132 // If we reached the key (or passed where it should be), we found the right predesssor
134 if (pred_ptr != NULL) {
139 TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
142 TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key, pred);
150 // <key> is not in <ll>.
151 if (pred_ptr != NULL) {
155 TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
159 // Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
160 uint64_t ll_lookup (list_t *ll, void *key) {
161 TRACE("l1", "ll_lookup: searching for key %p in list %p", key, ll);
163 int found = find_pred(NULL, &item, ll, key, FALSE);
165 // If we found an <item> matching the key return its value.
167 uint64_t val = item->val;
168 if (val != DOES_NOT_EXIST) {
169 TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
174 TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
175 return DOES_NOT_EXIST;
178 uint64_t ll_cas (list_t *ll, void *key, uint64_t expectation, uint64_t new_val) {
179 TRACE("l1", "ll_cas: key %p list %p", key, ll);
180 TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
181 ASSERT((int64_t)new_val > 0);
184 node_t *pred, *old_item;
185 int found = find_pred(&pred, &old_item, ll, key, TRUE);
188 // There was not an item in the list that matches the key.
189 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
190 TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
191 return DOES_NOT_EXIST; // failure
194 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
196 // Create a new item and insert it into the list.
197 TRACE("l2", "ll_cas: attempting to insert item between %p and %p", pred, pred->next);
198 void *new_key = (ll->clone_fun == NULL) ? key : ll->clone_fun(key);
199 node_t *new_item = node_alloc(new_key, new_val);
200 node_t *next = new_item->next = old_item;
201 node_t *other = SYNC_CAS(&pred->next, next, new_item);
203 TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0);
204 return DOES_NOT_EXIST; // success
207 // Lost a race. Failed to insert the new item into the list.
208 TRACE("l1", "ll_cas: lost a race. CAS failed. expected pred's link to be %p but found %p", next, other);
209 if (ll->clone_fun != NULL) {
216 // Found an item in the list that matches the key.
217 uint64_t old_item_val = old_item->val;
219 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
220 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
221 TRACE("l2", "ll_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
225 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
226 TRACE("l1", "ll_cas: found an item %p in the list that matched the key. the expectation was "
227 "not met, the list was not changed", old_item, old_item_val);
228 return old_item_val; // failure
231 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
232 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
233 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
234 // fails it means another thread either removed the node or updated its value.
235 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
236 if (ret_val == old_item_val) {
237 TRACE("l1", "ll_cas: the CAS succeeded. updated the value of the item", 0, 0);
238 return ret_val; // success
240 TRACE("l2", "ll_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
242 old_item_val = ret_val;
247 uint64_t ll_remove (list_t *ll, void *key) {
248 TRACE("l1", "ll_remove: removing item with key %p from list %p", key, ll);
251 int found = find_pred(&pred, &item, ll, key, TRUE);
253 TRACE("l1", "ll_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
254 return DOES_NOT_EXIST;
257 // Mark <item> removed. This must be atomic. If multiple threads try to remove the same item
258 // only one of them should succeed.
260 node_t *old_next = item->next;
263 old_next = SYNC_CAS(&item->next, next, TAG_VALUE(next));
264 if (IS_TAGGED(old_next)) {
265 TRACE("l1", "ll_remove: lost a race -- %p is already marked for removal by another thread", item, 0);
266 return DOES_NOT_EXIST;
268 } while (next != old_next);
269 TRACE("l2", "ll_remove: logically removed item %p", item, 0);
270 ASSERT(IS_TAGGED(item->next));
272 // This has to be an atomic swap in case another thread is updating the item while we are removing it.
273 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
274 TRACE("l2", "ll_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
276 // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
277 // item logically removed for a later call (or some other thread) to physically unlink. By marking the
278 // item earlier, we logically removed it.
279 TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next);
281 if ((other = SYNC_CAS(&pred->next, item, next)) != item) {
282 TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other);
286 // The thread that completes the unlink should free the memory.
287 if (ll->clone_fun != NULL) {
288 nbd_defer_free(item->key);
290 nbd_defer_free(item);
291 TRACE("l1", "ll_remove: successfully unlinked item %p from the list", item, 0);
295 void ll_print (list_t *ll) {
297 item = ll->head->next;
300 node_t *next = item->next;
301 if (IS_TAGGED(item)) {
304 printf("%p:%p ", item, item->key);
306 item = (node_t *)STRIP_TAG(next);