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
19 markable_t next; // next node
28 const datatype_t *key_type;
31 // Marking the <next> field of a node logically removes it from the list
32 #define MARK_NODE(x) TAG_VALUE((markable_t)(x), TAG1)
33 #define HAS_MARK(x) (IS_TAGGED((x), TAG1) == TAG1)
34 #define GET_NODE(x) ((node_t *)(x))
35 #define STRIP_MARK(x) ((node_t *)STRIP_TAG((x), TAG1))
37 static node_t *node_alloc (map_key_t key, map_val_t val) {
38 node_t *item = (node_t *)nbd_malloc(sizeof(node_t));
44 list_t *ll_alloc (const datatype_t *key_type) {
45 list_t *ll = (list_t *)nbd_malloc(sizeof(list_t));
46 ll->key_type = key_type;
47 ll->head = node_alloc(0, 0);
48 ll->head->next = DOES_NOT_EXIST;
52 void ll_free (list_t *ll) {
53 node_t *item = STRIP_MARK(ll->head->next);
54 while (item != NULL) {
55 node_t *next = STRIP_MARK(item->next);
61 size_t ll_count (list_t *ll) {
63 node_t *item = STRIP_MARK(ll->head->next);
65 if (!HAS_MARK(item->next)) {
68 item = STRIP_MARK(item->next);
73 static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, map_key_t key, int help_remove) {
74 node_t *pred = ll->head;
75 node_t *item = GET_NODE(pred->next);
76 TRACE("l2", "find_pred: searching for key %p in list (head is %p)", key, pred);
78 while (item != NULL) {
79 markable_t next = item->next;
81 // A mark means the node is logically removed but not physically unlinked yet.
82 while (EXPECT_FALSE(HAS_MARK(next))) {
84 // Skip over logically removed items.
86 item = STRIP_MARK(item->next);
87 if (EXPECT_FALSE(item == NULL))
89 TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
94 // Unlink logically removed items.
95 TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
97 markable_t other = SYNC_CAS(&pred->next, item, STRIP_MARK(next));
98 if (other == (markable_t)item) {
99 TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
100 item = STRIP_MARK(next);
101 next = (item != NULL) ? item->next : DOES_NOT_EXIST;
102 TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
104 // The thread that completes the unlink should free the memory.
105 node_t *unlinked = GET_NODE(other);
106 if (ll->key_type != NULL) {
107 nbd_defer_free((void *)unlinked->key);
109 nbd_defer_free(unlinked);
111 TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
112 TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
114 return find_pred(pred_ptr, item_ptr, ll, key, help_remove); // retry
115 item = GET_NODE(other);
116 next = (item != NULL) ? item->next : DOES_NOT_EXIST;
120 if (EXPECT_FALSE(item == NULL))
123 TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
124 TRACE("l4", "find_pred: key %p val %p", item->key, item->val);
127 if (EXPECT_TRUE(ll->key_type == NULL)) {
130 d = ll->key_type->cmp((void *)item->key, (void *)key);
133 // If we reached the key (or passed where it should be), we found the right predesssor
135 if (pred_ptr != NULL) {
140 TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
143 TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key, pred);
148 item = GET_NODE(next);
151 // <key> is not in <ll>.
152 if (pred_ptr != NULL) {
156 TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
160 // Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
161 map_val_t ll_lookup (list_t *ll, map_key_t key) {
162 TRACE("l1", "ll_lookup: searching for key %p in list %p", key, ll);
164 int found = find_pred(NULL, &item, ll, key, FALSE);
166 // If we found an <item> matching the key return its value.
168 map_val_t val = item->val;
169 if (val != DOES_NOT_EXIST) {
170 TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
175 TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
176 return DOES_NOT_EXIST;
179 map_val_t ll_cas (list_t *ll, map_key_t key, map_val_t expectation, map_val_t new_val) {
180 TRACE("l1", "ll_cas: key %p list %p", key, ll);
181 TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
182 ASSERT((int64_t)new_val > 0);
185 node_t *pred, *old_item;
186 int found = find_pred(&pred, &old_item, ll, key, TRUE);
189 // There was not an item in the list that matches the key.
190 if (EXPECT_FALSE(expectation != CAS_EXPECT_DOES_NOT_EXIST && expectation != CAS_EXPECT_WHATEVER)) {
191 TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
192 return DOES_NOT_EXIST; // failure
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 map_key_t new_key = ll->key_type == NULL ? key : (map_key_t)ll->key_type->clone((void *)key);
198 node_t *new_item = node_alloc(new_key, new_val);
199 markable_t next = new_item->next = (markable_t)old_item;
200 markable_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) {
209 nbd_free((void *)new_key);
215 // Found an item in the list that matches the key.
216 map_val_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 map_val_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 map_val_t ll_remove (list_t *ll, map_key_t 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 markable_t old_next = item->next;
261 old_next = SYNC_CAS(&item->next, next, MARK_NODE(STRIP_MARK(next)));
262 if (HAS_MARK(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(HAS_MARK(((volatile node_t *)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 map_val_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)) != (markable_t)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((void *)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) {
295 markable_t next = ll->head->next;
297 while (next != DOES_NOT_EXIST) {
298 if (HAS_MARK(next)) {
301 node_t *item = STRIP_MARK(next);
304 printf("%p:0x%llx ", item, (uint64_t)item->key);
315 ll_iter_t *ll_iter_begin (list_t *ll, map_key_t key) {
316 ll_iter_t *iter = (ll_iter_t *)nbd_malloc(sizeof(ll_iter_t));
317 if (key != DOES_NOT_EXIST) {
318 find_pred(NULL, &iter->next, ll, key, FALSE);
320 iter->next = GET_NODE(ll->head->next);
325 map_val_t ll_iter_next (ll_iter_t *iter, map_key_t *key_ptr) {
327 node_t *item = iter->next;
328 while (item != NULL && HAS_MARK(item->next)) {
329 item = STRIP_MARK(item->next);
333 return DOES_NOT_EXIST;
335 iter->next = STRIP_MARK(item->next);
336 if (key_ptr != NULL) {
337 *key_ptr = item->key;
342 void ll_iter_free (ll_iter_t *iter) {