X-Git-Url: https://pd.if.org/git/?p=nbds;a=blobdiff_plain;f=map%2Fskiplist.c;h=6e02e12600673267a80291b2005ae1e90662d4b3;hp=49ef21b3d32bcfe43c0e38cfdad883be1994eda5;hb=a1d0b3ca99552878b1becf561d8f3291992aaa67;hpb=f7a1c10d18dcc2654d0c9b1f5ffc9f4ec9b23776 diff --git a/map/skiplist.c b/map/skiplist.c index 49ef21b..6e02e12 100644 --- a/map/skiplist.c +++ b/map/skiplist.c @@ -21,25 +21,43 @@ #include #include "common.h" -#include "runtime.h" #include "skiplist.h" +#include "runtime.h" #include "mem.h" +#include "rcu.h" // Setting MAX_LEVEL to 0 essentially makes this data structure the Harris-Michael lock-free list (in list.c). #define MAX_LEVEL 31 typedef struct node { - void *key; - uint64_t val; + map_key_t key; + map_val_t val; int top_level; - struct node *next[]; + markable_t next[]; } node_t; +struct sl_iter { + node_t *next; +}; + struct sl { node_t *head; const datatype_t *key_type; }; +// Marking the field of a node logically removes it from the list +#if 0 +static inline markable_t MARK_NODE(node_t * x) { return TAG_VALUE((markable_t)x, 0x1); } +static inline int HAS_MARK(markable_t x) { return (IS_TAGGED(x, 0x1) == 0x1); } +static inline node_t * GET_NODE(markable_t x) { assert(!HAS_MARK(x)); return (node_t *)x; } +static inline node_t * STRIP_MARK(markable_t x) { return ((node_t *)STRIP_TAG(x, 0x1)); } +#else +#define MARK_NODE(x) TAG_VALUE((markable_t)(x), 0x1) +#define HAS_MARK(x) (IS_TAGGED((x), 0x1) == 0x1) +#define GET_NODE(x) ((node_t *)(x)) +#define STRIP_MARK(x) ((node_t *)STRIP_TAG((x), 0x1)) +#endif + static int random_level (void) { unsigned r = nbd_rand(); if (r & 1) @@ -52,7 +70,7 @@ static int random_level (void) { return n; } -static node_t *node_alloc (int level, void *key, uint64_t val) { +static node_t *node_alloc (int level, map_key_t key, map_val_t val) { assert(level >= 0 && level <= MAX_LEVEL); size_t sz = sizeof(node_t) + (level + 1) * sizeof(node_t *); node_t *item = (node_t *)nbd_malloc(sz); @@ -66,42 +84,45 @@ static node_t *node_alloc (int level, void *key, uint64_t val) { skiplist_t *sl_alloc (const datatype_t *key_type) { skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t)); sl->key_type = key_type; - sl->head = node_alloc(MAX_LEVEL, NULL, 0); + sl->head = node_alloc(MAX_LEVEL, 0, 0); memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *)); return sl; } void sl_free (skiplist_t *sl) { - node_t *item = sl->head->next[0]; + node_t *item = GET_NODE(sl->head->next[0]); while (item) { - node_t *next = (node_t *)STRIP_TAG(item->next[0]); + node_t *next = STRIP_MARK(item->next[0]); + if (sl->key_type != NULL) { + nbd_free((void *)item->key); + } nbd_free(item); item = next; } } -uint64_t sl_count (skiplist_t *sl) { - uint64_t count = 0; - node_t *item = sl->head->next[0]; +size_t sl_count (skiplist_t *sl) { + size_t count = 0; + node_t *item = GET_NODE(sl->head->next[0]); while (item) { - if (!IS_TAGGED(item->next[0])) { + if (!HAS_MARK(item->next[0])) { count++; } - item = (node_t *)STRIP_TAG(item->next[0]); + item = STRIP_MARK(item->next[0]); } return count; } -static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, void *key, int help_remove) { +static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, map_key_t key, int help_remove) { node_t *pred = sl->head; node_t *item = NULL; TRACE("s2", "find_preds: searching for key %p in skiplist (head is %p)", key, pred); - int d; + int d = 0; int start_level = MAX_LEVEL; #if MAX_LEVEL > 2 // Optimization for small lists. No need to traverse empty higher levels. start_level = 2; - while (pred->next[start_level+1] != NULL) { + while (pred->next[start_level+1] != DOES_NOT_EXIST) { start_level += start_level - 1; if (EXPECT_FALSE(start_level >= MAX_LEVEL)) { start_level = MAX_LEVEL; @@ -116,53 +137,51 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl // Traverse the levels of from the top level to the bottom for (int level = start_level; level >= 0; --level) { TRACE("s3", "find_preds: level %llu", level, 0); - item = pred->next[level]; - if (EXPECT_FALSE(IS_TAGGED(item))) { - TRACE("s2", "find_preds: pred %p is marked for removal (item %p); retry", pred, item); + markable_t next = pred->next[level]; + if (EXPECT_FALSE(HAS_MARK(next))) { + TRACE("s2", "find_preds: pred %p is marked for removal (next %p); retry", pred, next); return find_preds(preds, succs, n, sl, key, help_remove); // retry } + item = GET_NODE(next); while (item != NULL) { - node_t *next = item->next[level]; + next = item->next[level]; // A tag means an item is logically removed but not physically unlinked yet. - while (EXPECT_FALSE(IS_TAGGED(next))) { + while (EXPECT_FALSE(HAS_MARK(next))) { // Skip over logically removed items. if (!help_remove) { - item = (node_t *)STRIP_TAG(item->next); + item = STRIP_MARK(next); if (EXPECT_FALSE(item == NULL)) break; - TRACE("s3", "find_preds: skipping marked item %p (next is %p)", item, next); next = item->next[level]; + TRACE("s3", "find_preds: skipping marked item %p (next is 0x%llx)", item, next); continue; } // Unlink logically removed items. - node_t *other; - TRACE("s3", "find_preds: unlinking marked item %p; next is %p", item, next); - if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) == item) { - item = (node_t *)STRIP_TAG(next); - if (EXPECT_FALSE(item == NULL)) - break; - next = item->next[level]; - TRACE("s3", "find_preds: now the current item is %p next is %p", item, next); + TRACE("s3", "find_preds: unlinking marked item %p; next is 0x%llx", item, next); + markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next)); + if (other == (markable_t)item) { + item = STRIP_MARK(next); + next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST; + TRACE("s3", "find_preds: now the current item is %p next is 0x%llx", item, next); // The thread that completes the unlink should free the memory. if (level == 0) { + node_t *unlinked = GET_NODE(other); if (sl->key_type != NULL) { - nbd_defer_free((void*)other->key); + rcu_defer_free((void *)unlinked->key); } - nbd_defer_free(other); + rcu_defer_free(unlinked); } } else { TRACE("s3", "find_preds: lost race to unlink item %p from pred %p", item, pred); TRACE("s3", "find_preds: pred's link changed to %p", other, 0); - if (IS_TAGGED(other)) + if (HAS_MARK(other)) return find_preds(preds, succs, n, sl, key, help_remove); // retry - item = other; - if (EXPECT_FALSE(item == NULL)) - break; - next = item->next[level]; + item = GET_NODE(other); + next = (item != NULL) ? item->next[level] : DOES_NOT_EXIST; } } @@ -170,12 +189,12 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl break; TRACE("s4", "find_preds: visiting item %p (next is %p)", item, next); - TRACE("s4", "find_preds: key %p val %p", STRIP_TAG(item->key), item->val); + TRACE("s4", "find_preds: key %p val %p", STRIP_MARK(item->key), item->val); if (EXPECT_TRUE(sl->key_type == NULL)) { - d = (uint64_t)item->key - (uint64_t)key; + d = item->key - key; } else { - d = sl->key_type->cmp(item->key, key); + d = sl->key_type->cmp((void *)item->key, (void *)key); } if (d >= 0) { @@ -184,7 +203,7 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl } pred = item; - item = next; + item = GET_NODE(next); } // The cast to unsigned is for the case when n is -1. @@ -200,6 +219,7 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl // fill in empty levels if (n == -1 && item != NULL) { + assert(item->top_level <= MAX_LEVEL); for (int level = start_level + 1; level <= item->top_level; ++level) { preds[level] = sl->head; } @@ -214,13 +234,13 @@ static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl } // Fast find that does not help unlink partially removed nodes and does not return the node's predecessors. -uint64_t sl_lookup (skiplist_t *sl, void *key) { +map_val_t sl_lookup (skiplist_t *sl, map_key_t key) { TRACE("s1", "sl_lookup: searching for key %p in skiplist %p", key, sl); node_t *item = find_preds(NULL, NULL, 0, sl, key, FALSE); // If we found an matching the return its value. if (item != NULL) { - uint64_t val = item->val; + map_val_t val = item->val; if (val != DOES_NOT_EXIST) { TRACE("s1", "sl_lookup: found item %p. val %p. returning item", item, item->val); return val; @@ -231,18 +251,18 @@ uint64_t sl_lookup (skiplist_t *sl, void *key) { return DOES_NOT_EXIST; } -void *sl_min_key (skiplist_t *sl) { - node_t *item = sl->head->next[0]; +map_key_t sl_min_key (skiplist_t *sl) { + node_t *item = GET_NODE(sl->head->next[0]); while (item != NULL) { - node_t *next = item->next[0]; - if (!IS_TAGGED(next)) + markable_t next = item->next[0]; + if (!HAS_MARK(next)) return item->key; - item = (node_t *)STRIP_TAG(next); + item = STRIP_MARK(next); } return DOES_NOT_EXIST; } -uint64_t sl_cas (skiplist_t *sl, void *key, uint64_t expectation, uint64_t new_val) { +map_val_t sl_cas (skiplist_t *sl, map_key_t key, map_val_t expectation, map_val_t new_val) { TRACE("s1", "sl_cas: key %p skiplist %p", key, sl); TRACE("s1", "sl_cas: expectation %p new value %p", expectation, new_val); ASSERT((int64_t)new_val > 0); @@ -256,37 +276,35 @@ uint64_t sl_cas (skiplist_t *sl, void *key, uint64_t expectation, uint64_t new_v if (old_item == NULL) { // There was not an item in the skiplist that matches the key. - if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == CAS_EXPECT_EXISTS)) { + if (EXPECT_FALSE(expectation != CAS_EXPECT_DOES_NOT_EXIST && expectation != CAS_EXPECT_WHATEVER)) { TRACE("l1", "sl_cas: the expectation was not met, the skiplist was not changed", 0, 0); return DOES_NOT_EXIST; // failure } - ASSERT(expectation == CAS_EXPECT_DOES_NOT_EXIST || expectation == CAS_EXPECT_WHATEVER); - // First insert into the bottom level. TRACE("s3", "sl_cas: attempting to insert item between %p and %p", preds[0], nexts[0]); - void *new_key = (sl->key_type == NULL) ? key : sl->key_type->clone(key); + map_key_t new_key = sl->key_type == NULL ? key : (map_key_t)sl->key_type->clone((void *)key); new_item = node_alloc(n, new_key, new_val); node_t *pred = preds[0]; - node_t *next = new_item->next[0] = nexts[0]; + markable_t next = new_item->next[0] = (markable_t)nexts[0]; for (int level = 1; level <= new_item->top_level; ++level) { - new_item->next[level] = nexts[level]; + new_item->next[level] = (markable_t)nexts[level]; } - node_t *other = SYNC_CAS(&pred->next[0], next, new_item); + markable_t other = SYNC_CAS(&pred->next[0], next, new_item); if (other == next) { TRACE("s3", "sl_cas: successfully inserted item %p at level 0", new_item, 0); break; // success } TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other); if (sl->key_type != NULL) { - nbd_free(new_key); + nbd_free((void *)new_key); } nbd_free(new_item); continue; } // Found an item in the skiplist that matches the key. - uint64_t old_item_val = old_item->val; + map_val_t old_item_val = old_item->val; do { // If the item's value is DOES_NOT_EXIST it means another thread removed the node out from under us. if (EXPECT_FALSE(old_item_val == DOES_NOT_EXIST)) { @@ -304,7 +322,7 @@ uint64_t sl_cas (skiplist_t *sl, void *key, uint64_t expectation, uint64_t new_v // replace DOES_NOT_EXIST with our value. Then another thread that is updating the value could think it // succeeded and return our value even though we indicated that the node has been removed. If the CAS // fails it means another thread either removed the node or updated its value. - uint64_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val); + map_val_t ret_val = SYNC_CAS(&old_item->val, old_item_val, new_val); if (ret_val == old_item_val) { TRACE("s1", "sl_cas: the CAS succeeded. updated the value of the item", 0, 0); return ret_val; // success @@ -318,10 +336,10 @@ uint64_t sl_cas (skiplist_t *sl, void *key, uint64_t expectation, uint64_t new_v // Link into from the bottom up. for (int level = 1; level <= new_item->top_level; ++level) { node_t *pred = preds[level]; - node_t *next = nexts[level]; + markable_t next = (markable_t)nexts[level]; do { TRACE("s3", "sl_cas: attempting to insert item between %p and %p", pred, next); - node_t *other = SYNC_CAS(&pred->next[level], next, new_item); + markable_t other = SYNC_CAS(&pred->next[level], next, (markable_t)new_item); if (other == next) { TRACE("s3", "sl_cas: successfully inserted item %p at level %llu", new_item, level); break; // success @@ -329,13 +347,13 @@ uint64_t sl_cas (skiplist_t *sl, void *key, uint64_t expectation, uint64_t new_v TRACE("s3", "sl_cas: failed to change pred's link: expected %p found %p", next, other); find_preds(preds, nexts, new_item->top_level, sl, key, TRUE); pred = preds[level]; - next = nexts[level]; + next = (markable_t)nexts[level]; // Update 's next pointer do { // There in no need to continue linking in the item if another thread removed it. - node_t *old_next = ((volatile node_t *)new_item)->next[level]; - if (IS_TAGGED(old_next)) + markable_t old_next = ((volatile node_t *)new_item)->next[level]; + if (HAS_MARK(old_next)) return DOES_NOT_EXIST; // success // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item. @@ -347,7 +365,7 @@ uint64_t sl_cas (skiplist_t *sl, void *key, uint64_t expectation, uint64_t new_v return DOES_NOT_EXIST; // success } -uint64_t sl_remove (skiplist_t *sl, void *key) { +map_val_t sl_remove (skiplist_t *sl, map_key_t key) { TRACE("s1", "sl_remove: removing item with key %p from skiplist %p", key, sl); node_t *preds[MAX_LEVEL+1]; node_t *item = find_preds(preds, NULL, -1, sl, key, TRUE); @@ -359,33 +377,34 @@ uint64_t sl_remove (skiplist_t *sl, void *key) { // Mark and unlink at each level of from the top down. If multiple threads try to concurrently remove // the same item only one of them should succeed. Marking the bottom level establishes which of them succeeds. for (int level = item->top_level; level > 0; --level) { - node_t *next; - node_t *old_next = item->next[level]; + markable_t next; + markable_t old_next = item->next[level]; do { next = old_next; - old_next = SYNC_CAS(&item->next[level], next, TAG_VALUE(next)); - if (IS_TAGGED(old_next)) { + old_next = SYNC_CAS(&item->next[level], next, MARK_NODE((node_t *)next)); + if (HAS_MARK(old_next)) { TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level %llu", item, level); break; } } while (next != old_next); node_t *pred = preds[level]; - TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next)); - node_t *other = NULL; - if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) != item) { + TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_MARK(next)); + markable_t other = SYNC_CAS(&pred->next[level], item, STRIP_MARK(next)); + if (other != (markable_t)item) { TRACE("s1", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other); // If our former predecessor now points past us we know another thread unlinked us. Otherwise, we need // to search for a new set of preds. - if (other == NULL) + if (other == DOES_NOT_EXIST) continue; // points past to the end of the list; go on to the next level. int d = -1; - if (!IS_TAGGED(other)) { + if (!HAS_MARK(other)) { + map_key_t other_key = GET_NODE(other)->key; if (EXPECT_TRUE(sl->key_type == NULL)) { - d = (uint64_t)item->key - (uint64_t)other->key; + d = item->key - other_key; } else { - d = sl->key_type->cmp(item->key, other->key); + d = sl->key_type->cmp((void *)item->key, (void *)other_key); } } if (d > 0) { @@ -397,12 +416,12 @@ uint64_t sl_remove (skiplist_t *sl, void *key) { } } - node_t *next; - node_t *old_next = item->next[0]; + markable_t next; + markable_t old_next = item->next[0]; do { next = old_next; - old_next = SYNC_CAS(&item->next[0], next, TAG_VALUE(next)); - if (IS_TAGGED(old_next)) { + old_next = SYNC_CAS(&item->next[0], next, MARK_NODE((node_t *)next)); + if (HAS_MARK(old_next)) { TRACE("s2", "sl_remove: %p is already marked for removal by another thread at level 0", item, 0); return DOES_NOT_EXIST; } @@ -411,18 +430,18 @@ uint64_t sl_remove (skiplist_t *sl, void *key) { // Atomically swap out the item's value in case another thread is updating the item while we are // removing it. This establishes which operation occurs first logically, the update or the remove. - uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST); + map_val_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST); TRACE("s2", "sl_remove: replaced item %p's value with DOES_NOT_EXIT", item, 0); node_t *pred = preds[0]; - TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_TAG(next)); - if (SYNC_CAS(&pred->next[0], item, STRIP_TAG(next))) { + TRACE("s2", "sl_remove: linking the item's pred %p to the item's successor %p", pred, STRIP_MARK(next)); + if (SYNC_CAS(&pred->next[0], item, STRIP_MARK(next))) { TRACE("s2", "sl_remove: unlinked item %p from the skiplist at level 0", item, 0); // The thread that completes the unlink should free the memory. if (sl->key_type != NULL) { - nbd_defer_free(item->key); + rcu_defer_free((void *)item->key); } - nbd_defer_free(item); + rcu_defer_free(item); } return val; } @@ -430,14 +449,14 @@ uint64_t sl_remove (skiplist_t *sl, void *key) { void sl_print (skiplist_t *sl) { for (int level = MAX_LEVEL; level >= 0; --level) { node_t *item = sl->head; - if (item->next[level] == NULL) + if (item->next[level] == DOES_NOT_EXIST) continue; printf("(%d) ", level); int i = 0; while (item) { - node_t *next = item->next[level]; - printf("%s%p ", IS_TAGGED(next) ? "*" : "", item); - item = (node_t *)STRIP_TAG(next); + markable_t next = item->next[level]; + printf("%s%p ", HAS_MARK(next) ? "*" : "", item); + item = STRIP_MARK(next); if (i++ > 30) { printf("..."); break; @@ -449,26 +468,57 @@ void sl_print (skiplist_t *sl) { node_t *item = sl->head; int i = 0; while (item) { - int is_marked = IS_TAGGED(item->next[0]); - printf("%s%p:%p ", is_marked ? "*" : "", item, item->key); + int is_marked = HAS_MARK(item->next[0]); + printf("%s%p:0x%llx ", is_marked ? "*" : "", item, (uint64_t)item->key); if (item != sl->head) { printf("[%d]", item->top_level); } else { printf("[HEAD]"); } for (int level = 1; level <= item->top_level; ++level) { - node_t *next = (node_t *)STRIP_TAG(item->next[level]); - is_marked = IS_TAGGED(item->next[0]); + node_t *next = STRIP_MARK(item->next[level]); + is_marked = HAS_MARK(item->next[0]); printf(" %p%s", next, is_marked ? "*" : ""); - if (item == sl->head && item->next[level] == NULL) + if (item == sl->head && item->next[level] == DOES_NOT_EXIST) break; } printf("\n"); fflush(stdout); - item = (node_t *)STRIP_TAG(item->next[0]); + item = STRIP_MARK(item->next[0]); if (i++ > 30) { printf("...\n"); break; } } } + +sl_iter_t *sl_iter_begin (skiplist_t *sl, map_key_t key) { + sl_iter_t *iter = (sl_iter_t *)nbd_malloc(sizeof(sl_iter_t)); + if (key != DOES_NOT_EXIST) { + find_preds(NULL, &iter->next, 0, sl, key, FALSE); + } else { + iter->next = GET_NODE(sl->head->next[0]); + } + return iter; +} + +map_val_t sl_iter_next (sl_iter_t *iter, map_key_t *key_ptr) { + assert(iter); + node_t *item = iter->next; + while (item != NULL && HAS_MARK(item->next[0])) { + item = STRIP_MARK(item->next[0]); + } + if (item == NULL) { + iter->next = NULL; + return DOES_NOT_EXIST; + } + iter->next = STRIP_MARK(item->next[0]); + if (key_ptr != NULL) { + *key_ptr = item->key; + } + return item->val; +} + +void sl_iter_free (sl_iter_t *iter) { + nbd_free(iter); +}