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 const datatype_t *key_type;
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 (const datatype_t *key_type) {
43 list_t *ll = (list_t *)nbd_malloc(sizeof(list_t));
44 ll->key_type = key_type;
45 ll->head = node_alloc(NULL, 0);
46 ll->head->next = NULL;
50 void ll_free (list_t *ll) {
51 node_t *item = ll->head->next;
53 node_t *next = (node_t *)STRIP_TAG(item->next);
59 uint64_t ll_count (list_t *ll) {
61 node_t *item = ll->head->next;
64 item = (node_t *)STRIP_TAG(item->next);
69 static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, void *key, int help_remove) {
70 node_t *pred = ll->head;
71 node_t *item = pred->next;
72 TRACE("l2", "find_pred: searching for key %p in list (head is %p)", key, pred);
74 while (item != NULL) {
75 node_t *next = item->next;
77 // A tag means an item is logically removed but not physically unlinked yet.
78 while (EXPECT_FALSE(IS_TAGGED(next))) {
80 // Skip over logically removed items.
82 item = (node_t *)STRIP_TAG(item->next);
83 if (EXPECT_FALSE(item == NULL))
85 TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
90 // Unlink logically removed items.
92 TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
93 if ((other = SYNC_CAS(&pred->next, item, STRIP_TAG(next))) == item) {
94 TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
95 item = (node_t *)STRIP_TAG(next);
96 if (EXPECT_FALSE(item == NULL))
99 TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
101 // The thread that completes the unlink should free the memory.
102 if (ll->key_type != NULL) {
103 nbd_defer_free((void*)other->key);
105 nbd_defer_free(other);
107 TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
108 TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
109 if (IS_TAGGED(other))
110 return find_pred(pred_ptr, item_ptr, ll, key, help_remove); // retry
112 if (EXPECT_FALSE(item == NULL))
118 if (EXPECT_FALSE(item == NULL))
121 TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
122 TRACE("l4", "find_pred: key %p val %p", item->key, item->val);
125 if (EXPECT_TRUE(ll->key_type == NULL)) {
126 d = (uint64_t)item->key - (uint64_t)key;
128 d = ll->key_type->cmp(item->key, key);
131 // If we reached the key (or passed where it should be), we found the right predesssor
133 if (pred_ptr != NULL) {
138 TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
141 TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key, pred);
149 // <key> is not in <ll>.
150 if (pred_ptr != NULL) {
154 TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
158 // Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
159 uint64_t ll_lookup (list_t *ll, void *key) {
160 TRACE("l1", "ll_lookup: searching for key %p in list %p", key, ll);
162 int found = find_pred(NULL, &item, ll, key, FALSE);
164 // If we found an <item> matching the key return its value.
166 uint64_t val = item->val;
167 if (val != DOES_NOT_EXIST) {
168 TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
173 TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
174 return DOES_NOT_EXIST;
177 uint64_t ll_cas (list_t *ll, void *key, uint64_t expectation, uint64_t new_val) {
178 TRACE("l1", "ll_cas: key %p list %p", key, ll);
179 TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
180 ASSERT((int64_t)new_val > 0);
183 node_t *pred, *old_item;
184 int found = find_pred(&pred, &old_item, ll, key, TRUE);
187 // There was not an item in the list that matches the key.
188 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
189 TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
190 return DOES_NOT_EXIST; // failure
193 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
195 // Create a new item and insert it into the list.
196 TRACE("l2", "ll_cas: attempting to insert item between %p and %p", pred, pred->next);
197 void *new_key = (ll->key_type == NULL) ? key : ll->key_type->clone(key);
198 node_t *new_item = node_alloc(new_key, new_val);
199 node_t *next = new_item->next = old_item;
200 node_t *other = SYNC_CAS(&pred->next, next, new_item);
202 TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0);
203 return DOES_NOT_EXIST; // success
206 // Lost a race. Failed to insert the new item into the list.
207 TRACE("l1", "ll_cas: lost a race. CAS failed. expected pred's link to be %p but found %p", next, other);
208 if (ll->key_type != NULL) {
215 // Found an item in the list that matches the key.
216 uint64_t old_item_val = old_item->val;
218 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
219 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
220 TRACE("l2", "ll_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
224 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
225 TRACE("l1", "ll_cas: found an item %p in the list that matched the key. the expectation was "
226 "not met, the list was not changed", old_item, old_item_val);
227 return old_item_val; // failure
230 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
231 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
232 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
233 // fails it means another thread either removed the node or updated its value.
234 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
235 if (ret_val == old_item_val) {
236 TRACE("l1", "ll_cas: the CAS succeeded. updated the value of the item", 0, 0);
237 return ret_val; // success
239 TRACE("l2", "ll_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
241 old_item_val = ret_val;
246 uint64_t ll_remove (list_t *ll, void *key) {
247 TRACE("l1", "ll_remove: removing item with key %p from list %p", key, ll);
250 int found = find_pred(&pred, &item, ll, key, TRUE);
252 TRACE("l1", "ll_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
253 return DOES_NOT_EXIST;
256 // Mark <item> removed. If multiple threads try to remove the same item only one of them should succeed.
258 node_t *old_next = item->next;
261 old_next = SYNC_CAS(&item->next, next, TAG_VALUE(next));
262 if (IS_TAGGED(old_next)) {
263 TRACE("l1", "ll_remove: lost a race -- %p is already marked for removal by another thread", item, 0);
264 return DOES_NOT_EXIST;
266 } while (next != old_next);
267 TRACE("l2", "ll_remove: logically removed item %p", item, 0);
268 ASSERT(IS_TAGGED(item->next));
270 // Atomically swap out the item's value in case another thread is updating the item while we are
271 // removing it. This establishes which operation occurs first logically, the update or the remove.
272 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
273 TRACE("l2", "ll_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
275 // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
276 // item logically removed for a later call (or some other thread) to physically unlink. By marking the
277 // item earlier, we logically removed it.
278 TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next);
280 if ((other = SYNC_CAS(&pred->next, item, next)) != item) {
281 TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other);
285 // The thread that completes the unlink should free the memory.
286 if (ll->key_type != NULL) {
287 nbd_defer_free(item->key);
289 nbd_defer_free(item);
290 TRACE("l1", "ll_remove: successfully unlinked item %p from the list", item, 0);
294 void ll_print (list_t *ll) {
296 item = ll->head->next;
299 node_t *next = item->next;
300 if (IS_TAGGED(item)) {
303 printf("%p:%p ", item, item->key);
305 item = (node_t *)STRIP_TAG(next);