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
16 typedef struct ll_iter node_t;
26 const datatype_t *key_type;
29 static node_t *node_alloc (void *key, uint64_t val) {
30 node_t *item = (node_t *)nbd_malloc(sizeof(node_t));
36 list_t *ll_alloc (const datatype_t *key_type) {
37 list_t *ll = (list_t *)nbd_malloc(sizeof(list_t));
38 ll->key_type = key_type;
39 ll->head = node_alloc(NULL, 0);
40 ll->head->next = NULL;
44 void ll_free (list_t *ll) {
45 node_t *item = ll->head->next;
47 node_t *next = (node_t *)STRIP_TAG(item->next, TAG1);
53 uint64_t ll_count (list_t *ll) {
55 node_t *item = ll->head->next;
57 if (!IS_TAGGED(item->next, TAG1)) {
60 item = (node_t *)STRIP_TAG(item->next, TAG1);
65 static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, void *key, int help_remove) {
66 node_t *pred = ll->head;
67 node_t *item = pred->next;
68 TRACE("l2", "find_pred: searching for key %p in list (head is %p)", key, pred);
70 while (item != NULL) {
71 node_t *next = item->next;
73 // A tag means an item is logically removed but not physically unlinked yet.
74 while (EXPECT_FALSE(IS_TAGGED(next, TAG1))) {
76 // Skip over logically removed items.
78 item = (node_t *)STRIP_TAG(item->next, TAG1);
79 if (EXPECT_FALSE(item == NULL))
81 TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
86 // Unlink logically removed items.
88 TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
89 if ((other = SYNC_CAS(&pred->next, item, STRIP_TAG(next, TAG1))) == item) {
90 TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
91 item = (node_t *)STRIP_TAG(next, TAG1);
92 if (EXPECT_FALSE(item == NULL))
95 TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
97 // The thread that completes the unlink should free the memory.
98 if (ll->key_type != NULL) {
99 nbd_defer_free((void*)other->key);
101 nbd_defer_free(other);
103 TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
104 TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
105 if (IS_TAGGED(other, TAG1))
106 return find_pred(pred_ptr, item_ptr, ll, key, help_remove); // retry
108 if (EXPECT_FALSE(item == NULL))
114 if (EXPECT_FALSE(item == NULL))
117 TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
118 TRACE("l4", "find_pred: key %p val %p", item->key, item->val);
121 if (EXPECT_TRUE(ll->key_type == NULL)) {
122 d = (uint64_t)item->key - (uint64_t)key;
124 d = ll->key_type->cmp(item->key, key);
127 // If we reached the key (or passed where it should be), we found the right predesssor
129 if (pred_ptr != NULL) {
134 TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
137 TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key, pred);
145 // <key> is not in <ll>.
146 if (pred_ptr != NULL) {
150 TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
154 // Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
155 uint64_t ll_lookup (list_t *ll, void *key) {
156 TRACE("l1", "ll_lookup: searching for key %p in list %p", key, ll);
158 int found = find_pred(NULL, &item, ll, key, FALSE);
160 // If we found an <item> matching the key return its value.
162 uint64_t val = item->val;
163 if (val != DOES_NOT_EXIST) {
164 TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
169 TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
170 return DOES_NOT_EXIST;
173 uint64_t ll_cas (list_t *ll, void *key, uint64_t expectation, uint64_t new_val) {
174 TRACE("l1", "ll_cas: key %p list %p", key, ll);
175 TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
176 ASSERT((int64_t)new_val > 0);
179 node_t *pred, *old_item;
180 int found = find_pred(&pred, &old_item, ll, key, TRUE);
183 // There was not an item in the list that matches the key.
184 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
185 TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
186 return DOES_NOT_EXIST; // failure
189 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
191 // Create a new item and insert it into the list.
192 TRACE("l2", "ll_cas: attempting to insert item between %p and %p", pred, pred->next);
193 void *new_key = (ll->key_type == NULL) ? key : ll->key_type->clone(key);
194 node_t *new_item = node_alloc(new_key, new_val);
195 node_t *next = new_item->next = old_item;
196 node_t *other = SYNC_CAS(&pred->next, next, new_item);
198 TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0);
199 return DOES_NOT_EXIST; // success
202 // Lost a race. Failed to insert the new item into the list.
203 TRACE("l1", "ll_cas: lost a race. CAS failed. expected pred's link to be %p but found %p", next, other);
204 if (ll->key_type != NULL) {
211 // Found an item in the list that matches the key.
212 uint64_t old_item_val = old_item->val;
214 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
215 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
216 TRACE("l2", "ll_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
220 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
221 TRACE("l1", "ll_cas: found an item %p in the list that matched the key. the expectation was "
222 "not met, the list was not changed", old_item, old_item_val);
223 return old_item_val; // failure
226 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
227 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
228 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
229 // fails it means another thread either removed the node or updated its value.
230 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
231 if (ret_val == old_item_val) {
232 TRACE("l1", "ll_cas: the CAS succeeded. updated the value of the item", 0, 0);
233 return ret_val; // success
235 TRACE("l2", "ll_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
237 old_item_val = ret_val;
242 uint64_t ll_remove (list_t *ll, void *key) {
243 TRACE("l1", "ll_remove: removing item with key %p from list %p", key, ll);
246 int found = find_pred(&pred, &item, ll, key, TRUE);
248 TRACE("l1", "ll_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
249 return DOES_NOT_EXIST;
252 // Mark <item> removed. If multiple threads try to remove the same item only one of them should succeed.
254 node_t *old_next = item->next;
257 old_next = SYNC_CAS(&item->next, next, TAG_VALUE(next, TAG1));
258 if (IS_TAGGED(old_next, TAG1)) {
259 TRACE("l1", "ll_remove: lost a race -- %p is already marked for removal by another thread", item, 0);
260 return DOES_NOT_EXIST;
262 } while (next != old_next);
263 TRACE("l2", "ll_remove: logically removed item %p", item, 0);
264 ASSERT(IS_TAGGED(item->next, TAG1));
266 // Atomically swap out the item's value in case another thread is updating the item while we are
267 // removing it. This establishes which operation occurs first logically, the update or the remove.
268 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
269 TRACE("l2", "ll_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
271 // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
272 // item logically removed for a later call (or some other thread) to physically unlink. By marking the
273 // item earlier, we logically removed it.
274 TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next);
276 if ((other = SYNC_CAS(&pred->next, item, next)) != item) {
277 TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other);
281 // The thread that completes the unlink should free the memory.
282 if (ll->key_type != NULL) {
283 nbd_defer_free(item->key);
285 nbd_defer_free(item);
286 TRACE("l1", "ll_remove: successfully unlinked item %p from the list", item, 0);
290 void ll_print (list_t *ll) {
292 item = ll->head->next;
295 node_t *next = item->next;
296 if (IS_TAGGED(item, TAG1)) {
299 printf("%p:%p ", item, item->key);
301 item = (node_t *)STRIP_TAG(next, TAG1);
310 ll_iter_t *ll_iter_start (list_t *ll, void *key) {
312 find_pred(NULL, &item, ll, key, FALSE);
316 ll_iter_t *ll_iter_next (ll_iter_t *iter) {
318 if (EXPECT_FALSE(!iter))
321 node_t *next = iter->next;
322 while (next != NULL && IS_TAGGED(next->next, TAG1)) {
323 next = (node_t *)STRIP_TAG(next->next, TAG1);
329 uint64_t ll_iter_val (ll_iter_t *iter) {
333 void *ll_iter_key (ll_iter_t *iter) {