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
24 const datatype_t *key_type;
27 static node_t *node_alloc (void *key, uint64_t val) {
28 node_t *item = (node_t *)nbd_malloc(sizeof(node_t));
34 list_t *ll_alloc (const datatype_t *key_type) {
35 list_t *ll = (list_t *)nbd_malloc(sizeof(list_t));
36 ll->key_type = key_type;
37 ll->head = node_alloc(NULL, 0);
38 ll->head->next = NULL;
42 void ll_free (list_t *ll) {
43 node_t *item = ll->head->next;
45 node_t *next = (node_t *)STRIP_TAG(item->next, TAG1);
51 uint64_t ll_count (list_t *ll) {
53 node_t *item = ll->head->next;
55 if (!IS_TAGGED(item->next, TAG1)) {
58 item = (node_t *)STRIP_TAG(item->next, TAG1);
63 static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, void *key, int help_remove) {
64 node_t *pred = ll->head;
65 node_t *item = pred->next;
66 TRACE("l2", "find_pred: searching for key %p in list (head is %p)", key, pred);
68 while (item != NULL) {
69 node_t *next = item->next;
71 // A tag means an item is logically removed but not physically unlinked yet.
72 while (EXPECT_FALSE(IS_TAGGED(next, TAG1))) {
74 // Skip over logically removed items.
76 item = (node_t *)STRIP_TAG(item->next, TAG1);
77 if (EXPECT_FALSE(item == NULL))
79 TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
84 // Unlink logically removed items.
86 TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
87 if ((other = SYNC_CAS(&pred->next, item, STRIP_TAG(next, TAG1))) == item) {
88 TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
89 item = (node_t *)STRIP_TAG(next, TAG1);
90 if (EXPECT_FALSE(item == NULL))
93 TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
95 // The thread that completes the unlink should free the memory.
96 if (ll->key_type != NULL) {
97 nbd_defer_free((void*)other->key);
99 nbd_defer_free(other);
101 TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
102 TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
103 if (IS_TAGGED(other, TAG1))
104 return find_pred(pred_ptr, item_ptr, ll, key, help_remove); // retry
106 if (EXPECT_FALSE(item == NULL))
112 if (EXPECT_FALSE(item == NULL))
115 TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
116 TRACE("l4", "find_pred: key %p val %p", item->key, item->val);
119 if (EXPECT_TRUE(ll->key_type == NULL)) {
120 d = (uint64_t)item->key - (uint64_t)key;
122 d = ll->key_type->cmp(item->key, key);
125 // If we reached the key (or passed where it should be), we found the right predesssor
127 if (pred_ptr != NULL) {
132 TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
135 TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key, pred);
143 // <key> is not in <ll>.
144 if (pred_ptr != NULL) {
148 TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
152 // Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
153 uint64_t ll_lookup (list_t *ll, void *key) {
154 TRACE("l1", "ll_lookup: searching for key %p in list %p", key, ll);
156 int found = find_pred(NULL, &item, ll, key, FALSE);
158 // If we found an <item> matching the key return its value.
160 uint64_t val = item->val;
161 if (val != DOES_NOT_EXIST) {
162 TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
167 TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
168 return DOES_NOT_EXIST;
171 uint64_t ll_cas (list_t *ll, void *key, uint64_t expectation, uint64_t new_val) {
172 TRACE("l1", "ll_cas: key %p list %p", key, ll);
173 TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
174 ASSERT((int64_t)new_val > 0);
177 node_t *pred, *old_item;
178 int found = find_pred(&pred, &old_item, ll, key, TRUE);
181 // There was not an item in the list that matches the key.
182 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
183 TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
184 return DOES_NOT_EXIST; // failure
187 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
189 // Create a new item and insert it into the list.
190 TRACE("l2", "ll_cas: attempting to insert item between %p and %p", pred, pred->next);
191 void *new_key = (ll->key_type == NULL) ? key : ll->key_type->clone(key);
192 node_t *new_item = node_alloc(new_key, new_val);
193 node_t *next = new_item->next = old_item;
194 node_t *other = SYNC_CAS(&pred->next, next, new_item);
196 TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0);
197 return DOES_NOT_EXIST; // success
200 // Lost a race. Failed to insert the new item into the list.
201 TRACE("l1", "ll_cas: lost a race. CAS failed. expected pred's link to be %p but found %p", next, other);
202 if (ll->key_type != NULL) {
209 // Found an item in the list that matches the key.
210 uint64_t old_item_val = old_item->val;
212 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
213 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
214 TRACE("l2", "ll_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
218 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
219 TRACE("l1", "ll_cas: found an item %p in the list that matched the key. the expectation was "
220 "not met, the list was not changed", old_item, old_item_val);
221 return old_item_val; // failure
224 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
225 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
226 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
227 // fails it means another thread either removed the node or updated its value.
228 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
229 if (ret_val == old_item_val) {
230 TRACE("l1", "ll_cas: the CAS succeeded. updated the value of the item", 0, 0);
231 return ret_val; // success
233 TRACE("l2", "ll_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
235 old_item_val = ret_val;
240 uint64_t ll_remove (list_t *ll, void *key) {
241 TRACE("l1", "ll_remove: removing item with key %p from list %p", key, ll);
244 int found = find_pred(&pred, &item, ll, key, TRUE);
246 TRACE("l1", "ll_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
247 return DOES_NOT_EXIST;
250 // Mark <item> removed. If multiple threads try to remove the same item only one of them should succeed.
252 node_t *old_next = item->next;
255 old_next = SYNC_CAS(&item->next, next, TAG_VALUE(next, TAG1));
256 if (IS_TAGGED(old_next, TAG1)) {
257 TRACE("l1", "ll_remove: lost a race -- %p is already marked for removal by another thread", item, 0);
258 return DOES_NOT_EXIST;
260 } while (next != old_next);
261 TRACE("l2", "ll_remove: logically removed item %p", item, 0);
262 ASSERT(IS_TAGGED(item->next, TAG1));
264 // Atomically swap out the item's value in case another thread is updating the item while we are
265 // removing it. This establishes which operation occurs first logically, the update or the remove.
266 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
267 TRACE("l2", "ll_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
269 // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
270 // item logically removed for a later call (or some other thread) to physically unlink. By marking the
271 // item earlier, we logically removed it.
272 TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next);
274 if ((other = SYNC_CAS(&pred->next, item, next)) != item) {
275 TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other);
279 // The thread that completes the unlink should free the memory.
280 if (ll->key_type != NULL) {
281 nbd_defer_free(item->key);
283 nbd_defer_free(item);
284 TRACE("l1", "ll_remove: successfully unlinked item %p from the list", item, 0);
288 void ll_print (list_t *ll) {
290 item = ll->head->next;
293 node_t *next = item->next;
294 if (IS_TAGGED(item, TAG1)) {
297 printf("%p:%p ", item, item->key);
299 item = (node_t *)STRIP_TAG(next, TAG1);