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);
51 uint64_t ll_count (list_t *ll) {
53 node_t *item = ll->head->next;
56 item = (node_t *)STRIP_TAG(item->next);
61 static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, void *key, int help_remove) {
62 node_t *pred = ll->head;
63 node_t *item = pred->next;
64 TRACE("l2", "find_pred: searching for key %p in list (head is %p)", key, pred);
66 while (item != NULL) {
67 node_t *next = item->next;
69 // A tag means an item is logically removed but not physically unlinked yet.
70 while (EXPECT_FALSE(IS_TAGGED(next))) {
72 // Skip over logically removed items.
74 item = (node_t *)STRIP_TAG(item->next);
75 if (EXPECT_FALSE(item == NULL))
77 TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
82 // Unlink logically removed items.
84 TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
85 if ((other = SYNC_CAS(&pred->next, item, STRIP_TAG(next))) == item) {
86 TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
87 item = (node_t *)STRIP_TAG(next);
88 if (EXPECT_FALSE(item == NULL))
91 TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
93 // The thread that completes the unlink should free the memory.
94 if (ll->key_type != NULL) {
95 nbd_defer_free((void*)other->key);
97 nbd_defer_free(other);
99 TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
100 TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
101 if (IS_TAGGED(other))
102 return find_pred(pred_ptr, item_ptr, ll, key, help_remove); // retry
104 if (EXPECT_FALSE(item == NULL))
110 if (EXPECT_FALSE(item == NULL))
113 TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
114 TRACE("l4", "find_pred: key %p val %p", item->key, item->val);
117 if (EXPECT_TRUE(ll->key_type == NULL)) {
118 d = (uint64_t)item->key - (uint64_t)key;
120 d = ll->key_type->cmp(item->key, key);
123 // If we reached the key (or passed where it should be), we found the right predesssor
125 if (pred_ptr != NULL) {
130 TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
133 TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key, pred);
141 // <key> is not in <ll>.
142 if (pred_ptr != NULL) {
146 TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
150 // Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
151 uint64_t ll_lookup (list_t *ll, void *key) {
152 TRACE("l1", "ll_lookup: searching for key %p in list %p", key, ll);
154 int found = find_pred(NULL, &item, ll, key, FALSE);
156 // If we found an <item> matching the key return its value.
158 uint64_t val = item->val;
159 if (val != DOES_NOT_EXIST) {
160 TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
165 TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
166 return DOES_NOT_EXIST;
169 uint64_t ll_cas (list_t *ll, void *key, uint64_t expectation, uint64_t new_val) {
170 TRACE("l1", "ll_cas: key %p list %p", key, ll);
171 TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
172 ASSERT((int64_t)new_val > 0);
175 node_t *pred, *old_item;
176 int found = find_pred(&pred, &old_item, ll, key, TRUE);
179 // There was not an item in the list that matches the key.
180 if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) {
181 TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
182 return DOES_NOT_EXIST; // failure
185 ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER);
187 // Create a new item and insert it into the list.
188 TRACE("l2", "ll_cas: attempting to insert item between %p and %p", pred, pred->next);
189 void *new_key = (ll->key_type == NULL) ? key : ll->key_type->clone(key);
190 node_t *new_item = node_alloc(new_key, new_val);
191 node_t *next = new_item->next = old_item;
192 node_t *other = SYNC_CAS(&pred->next, next, new_item);
194 TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0);
195 return DOES_NOT_EXIST; // success
198 // Lost a race. Failed to insert the new item into the list.
199 TRACE("l1", "ll_cas: lost a race. CAS failed. expected pred's link to be %p but found %p", next, other);
200 if (ll->key_type != NULL) {
207 // Found an item in the list that matches the key.
208 uint64_t old_item_val = old_item->val;
210 // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us.
211 if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) {
212 TRACE("l2", "ll_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
216 if (EXPECT_FALSE(expectation == CAS_EXPECT_DOES_NOT_EXIST)) {
217 TRACE("l1", "ll_cas: found an item %p in the list that matched the key. the expectation was "
218 "not met, the list was not changed", old_item, old_item_val);
219 return old_item_val; // failure
222 // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
223 // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it
224 // succeeded and return our value even though we indicated that the node has been removed. If the CAS
225 // fails it means another thread either removed the node or updated its value.
226 uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val);
227 if (ret_val == old_item_val) {
228 TRACE("l1", "ll_cas: the CAS succeeded. updated the value of the item", 0, 0);
229 return ret_val; // success
231 TRACE("l2", "ll_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
233 old_item_val = ret_val;
238 uint64_t ll_remove (list_t *ll, void *key) {
239 TRACE("l1", "ll_remove: removing item with key %p from list %p", key, ll);
242 int found = find_pred(&pred, &item, ll, key, TRUE);
244 TRACE("l1", "ll_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
245 return DOES_NOT_EXIST;
248 // Mark <item> removed. If multiple threads try to remove the same item only one of them should succeed.
250 node_t *old_next = item->next;
253 old_next = SYNC_CAS(&item->next, next, TAG_VALUE(next));
254 if (IS_TAGGED(old_next)) {
255 TRACE("l1", "ll_remove: lost a race -- %p is already marked for removal by another thread", item, 0);
256 return DOES_NOT_EXIST;
258 } while (next != old_next);
259 TRACE("l2", "ll_remove: logically removed item %p", item, 0);
260 ASSERT(IS_TAGGED(item->next));
262 // Atomically swap out the item's value in case another thread is updating the item while we are
263 // removing it. This establishes which operation occurs first logically, the update or the remove.
264 uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
265 TRACE("l2", "ll_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
267 // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
268 // item logically removed for a later call (or some other thread) to physically unlink. By marking the
269 // item earlier, we logically removed it.
270 TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next);
272 if ((other = SYNC_CAS(&pred->next, item, next)) != item) {
273 TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other);
277 // The thread that completes the unlink should free the memory.
278 if (ll->key_type != NULL) {
279 nbd_defer_free(item->key);
281 nbd_defer_free(item);
282 TRACE("l1", "ll_remove: successfully unlinked item %p from the list", item, 0);
286 void ll_print (list_t *ll) {
288 item = ll->head->next;
291 node_t *next = item->next;
292 if (IS_TAGGED(item)) {
295 printf("%p:%p ", item, item->key);
297 item = (node_t *)STRIP_TAG(next);