X-Git-Url: https://pd.if.org/git/?p=nbds;a=blobdiff_plain;f=struct%2Fskiplist.c;h=a3fab396f8a16cee038c17d7841b279e5267b71c;hp=0b49fd4937b9c2170ed4c7d20816d51c691bbcea;hb=9b3e566281f7e2ac0683205042796958bfd8939f;hpb=53d171373819e921da8f8648eea236a08ff6a702

diff --git a/struct/skiplist.c b/struct/skiplist.c
index 0b49fd4..a3fab39 100644
--- a/struct/skiplist.c
+++ b/struct/skiplist.c
@@ -24,12 +24,12 @@
 #include "tls.h"
 
 // Setting MAX_LEVEL to 0 essentially makes this data structure the Harris-Michael lock-free list
-// in list.c
+// (in list.c).
 #define MAX_LEVEL 31
 
 typedef struct node {
     nstring_t *key;
-    uint64_t value;
+    uint64_t val;
     int top_level;
     struct node *next[];
 } node_t;
@@ -40,28 +40,44 @@ struct sl {
 
 static int random_level (void) {
     unsigned r = nbd_rand();
-    if (r&1)
+    if (r & 1)
         return 0;
-    int n = __builtin_ctz(r)-1;
 #if MAX_LEVEL < 31
-    if (n > MAX_LEVEL)
-        return MAX_LEVEL;
+    r |= 1 << (MAX_LEVEL+1);
 #endif
+    int n = __builtin_ctz(r)-1;
     assert(n <= MAX_LEVEL);
     return n;
 }
 
-node_t *node_alloc (int level, const void *key_data, uint32_t key_len, uint64_t value) {
+node_t *node_alloc (int level, const void *key_data, uint32_t key_len, uint64_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);
     memset(item, 0, sz);
-    item->key   = ns_alloc(key_data, key_len);
-    item->value = value;
+    // If <key_len> is -1 it indicates <key_data> is an integer and not a pointer
+    item->key = (key_len == (unsigned)-1) 
+              ? (void *)TAG_VALUE(key_data) 
+              : ns_alloc(key_data, key_len); 
+    item->val = val;
     item->top_level = level;
     return item;
 }
 
+static void node_free (node_t *item) {
+    if (!IS_TAGGED(item->key)) {
+        nbd_free(item->key);
+    }
+    nbd_free(item);
+}
+
+static void node_defer_free (node_t *item) {
+    if (!IS_TAGGED(item->key)) {
+        nbd_defer_free(item->key);
+    }
+    nbd_defer_free(item);
+}
+
 skiplist_t *sl_alloc (void) {
     skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
     sl->head = node_alloc(MAX_LEVEL, " ", 0, 0);
@@ -69,12 +85,11 @@ skiplist_t *sl_alloc (void) {
     return sl;
 }
 
-static node_t *find_preds (int *found, node_t *preds[MAX_LEVEL+1], int n, skiplist_t *sl, const void *key_data, uint32_t key_len, int help_remove) {
+static node_t *find_preds (node_t **preds, node_t **succs, int n, skiplist_t *sl, const void *key_data, uint32_t key_len, int help_remove) {
     node_t *pred = sl->head;
     node_t *item = NULL;
     TRACE("s3", "find_preds: searching for key %p in sl (head is %p)", key_data, pred);
-    *found = 0;
-
+    int x;
     int start_level = MAX_LEVEL;
 #if MAX_LEVEL > 2
     // Optimization for small lists. No need to traverse empty higher levels.
@@ -97,12 +112,12 @@ static node_t *find_preds (int *found, node_t *preds[MAX_LEVEL+1], int n, skipli
         item = pred->next[level];
         if (EXPECT_FALSE(IS_TAGGED(item))) {
             TRACE("s3", "find_preds: pred %p is marked for removal (item %p); retry", pred, item);
-            return find_preds(found, preds, n, sl, key_data, key_len, help_remove); // retry
+            return find_preds(preds, succs, n, sl, key_data, key_len, help_remove); // retry
         }
         while (item != NULL) {
             node_t *next = item->next[level];
             TRACE("s3", "find_preds: visiting item %p (next %p)", item, next);
-            TRACE("s3", "find_preds: key %p", item->key, 0);
+            TRACE("s3", "find_preds: key %p", STRIP_TAG(item->key), item->val);
 
             // A tag means an item is logically removed but not physically unlinked yet.
             while (EXPECT_FALSE(IS_TAGGED(next))) {
@@ -127,11 +142,11 @@ static node_t *find_preds (int *found, node_t *preds[MAX_LEVEL+1], int n, skipli
                     TRACE("s3", "find_preds: now item is %p next is %p", item, next);
 
                     // The thread that completes the unlink should free the memory.
-                    if (level == 0) { nbd_defer_free(other); }
+                    if (level == 0) { node_defer_free(other); }
                 } else {
                     TRACE("s3", "find_preds: lost race to unlink from pred %p; its link changed to %p", pred, other);
                     if (IS_TAGGED(other))
-                        return find_preds(found, preds, n, sl, key_data, key_len, help_remove); // retry
+                        return find_preds(preds, succs, n, sl, key_data, key_len, help_remove); // retry
                     item = other;
                     if (EXPECT_FALSE(item == NULL))
                         break;
@@ -143,11 +158,10 @@ static node_t *find_preds (int *found, node_t *preds[MAX_LEVEL+1], int n, skipli
                 break;
 
             // If we reached the key (or passed where it should be), we found a pred. Save it and continue down.
-            int x = ns_cmp_raw(item->key, key_data, key_len);
+            x = (IS_TAGGED(item->key))
+              ? (STRIP_TAG(item->key) - (uint64_t)key_data)
+              : ns_cmp_raw(item->key, key_data, key_len);
             if (x >= 0) {
-                if (level == 0 && x == 0) {
-                    *found = 1;
-                }
                 TRACE("s3", "find_preds: found pred %p item %p", pred, item);
                 break;
             }
@@ -155,111 +169,106 @@ static node_t *find_preds (int *found, node_t *preds[MAX_LEVEL+1], int n, skipli
             pred = item;
             item = next;
         }
-        if (preds != NULL) {
-            preds[level] = pred;
+
+        // The cast to unsigned is for the case when n is -1.
+        if ((unsigned)level <= (unsigned)n) { 
+            if (preds != NULL) {
+                preds[level] = pred;
+            }
+            if (succs != NULL) {
+                succs[level] = item;
+            }
         }
     }
     if (n == -1 && item != NULL) {
-        assert(preds != NULL);
         for (int level = start_level + 1; level <= item->top_level; ++level) {
             preds[level] = sl->head;
         }
     }
-    return item;
+    return x == 0 ? item : NULL;
 }
 
 // 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, const void *key_data, uint32_t key_len) {
-    TRACE("s3", "sl_lookup: searching for key %p in sl %p", key, sl);
-    int found;
-    node_t *item = find_preds(&found, NULL, 0, sl, key_data, key_len, FALSE);
+    TRACE("s3", "sl_lookup: searching for key %p in skiplist %p", key, sl);
+    node_t *item = find_preds(NULL, NULL, 0, sl, key_data, key_len, FALSE);
 
     // If we found an <item> matching the <key> return its value.
-    return found ? item->value : DOES_NOT_EXIST;
+    return item != NULL ? item->val : DOES_NOT_EXIST;
 }
 
 // Insert the <key> if it doesn't already exist in <sl>
-uint64_t sl_add (skiplist_t *sl, const void *key_data, uint32_t key_len, uint64_t value) {
-    TRACE("s3", "sl_add: inserting key %p value %p", key, value);
+uint64_t sl_add (skiplist_t *sl, const void *key_data, uint32_t key_len, uint64_t val) {
+    TRACE("s3", "sl_add: inserting key %p val %p", key_data, val);
     node_t *preds[MAX_LEVEL+1];
-    node_t *item = NULL;
+    node_t *nexts[MAX_LEVEL+1];
+    node_t *new_item = NULL;
     int n = random_level();
     do {
-        int found;
-        node_t *next = find_preds(&found, preds, n, sl, key_data, key_len, TRUE);
+        node_t *old_item = find_preds(preds, nexts, n, sl, key_data, key_len, TRUE);
 
         // If a node matching <key> already exists in <sl>, return its value.
-        if (found) {
-            TRACE("s3", "sl_add: there is already an item %p (value %p) with the same key", next, next->value);
-            if (EXPECT_FALSE(item != NULL)) { nbd_free(item); }
-            return next->value;
+        if (old_item != NULL) {
+            TRACE("s3", "sl_add: there is already an item %p (value %p) with the same key", nexts[0], nexts[0]->val);
+            return nexts[0]->val;
         }
 
-        // First insert <item> into the bottom level.
-        if (EXPECT_TRUE(item == NULL)) { item = node_alloc(n, key_data, key_len, value); }
-        TRACE("s3", "sl_add: attempting to insert item between %p and %p", preds[0], next);
-        item->next[0] = next;
-        for (int level = 1; level <= item->top_level; ++level) {
-            node_t *pred = preds[level];
-            item->next[level] = pred->next[level];
-        }
+        // First insert <new_item> into the bottom level.
+        TRACE("s3", "sl_add: attempting to insert item between %p and %p", preds[0], nexts[0]);
+        new_item = node_alloc(n, key_data, key_len, val);
         node_t *pred = preds[0];
-        node_t *other = SYNC_CAS(&pred->next[0], next, item);
+        node_t *next = new_item->next[0] = nexts[0];
+        for (int level = 1; level <= new_item->top_level; ++level) {
+            new_item->next[level] = nexts[level];
+        }
+        node_t *other = SYNC_CAS(&pred->next[0], next, new_item);
         if (other == next) {
-            TRACE("s3", "sl_add: successfully inserted item %p at level 0", item, 0);
+            TRACE("s3", "sl_add: successfully inserted item %p at level 0", new_item, 0);
             break; // success
         }
         TRACE("s3", "sl_add: failed to change pred's link: expected %p found %p", next, other);
+        node_free(new_item);
 
     } while (1);
 
-    // Insert <item> into <sl> from the bottom level up.
-    for (int level = 1; level <= item->top_level; ++level) {
+    // Insert <new_item> into <sl> from the bottom level up.
+    for (int level = 1; level <= new_item->top_level; ++level) {
+        node_t *pred = preds[level];
+        node_t *next = nexts[level];
         do {
-            node_t *pred;
-            node_t *next;
-            do {
-                pred = preds[level];
-                next = pred->next[level];
-                if (next == NULL) // item goes at the end of the list
-                    break;
-                if (!IS_TAGGED(next) && ns_cmp_raw(next->key, key_data, key_len) > 0) // pred's link changed
-                    break;
-                int found;
-                find_preds(&found, preds, item->top_level, sl, key_data, key_len, TRUE);
-            } while (1);
+            TRACE("s3", "sl_add: attempting to insert item between %p and %p", pred, next);
+            node_t *other = SYNC_CAS(&pred->next[level], next, new_item);
+            if (other == next) {
+                TRACE("s3", "sl_add: successfully inserted item %p at level %llu", new_item, level);
+                break; // success
+            }
+            TRACE("s3", "sl_add: failed to change pred's link: expected %p found %p", next, other);
+            find_preds(preds, nexts, new_item->top_level, sl, key_data, key_len, TRUE);
+            pred = preds[level];
+            next = nexts[level];
 
+            // Update <new_item>'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 *)item)->next[level];
+                node_t *old_next = ((volatile node_t *)new_item)->next[level];
                 if (IS_TAGGED(old_next))
-                    return DOES_NOT_EXIST; // success
+                    return val;
 
-                // Use a CAS so we to not inadvertantly remove a mark another thread placed on the item.
-                if (next == old_next || SYNC_CAS(&item->next[level], old_next, next) == old_next)
+                // Use a CAS so we do not inadvertantly stomp on a mark another thread placed on the item.
+                if (old_next == next || SYNC_CAS(&new_item->next[level], old_next, next) == old_next)
                     break;
             } while (1);
-
-            TRACE("s3", "sl_add: attempting to insert item between %p and %p", pred, next);
-            node_t *other = SYNC_CAS(&pred->next[level], next, item);
-            if (other == next) {
-                TRACE("s3", "sl_add: successfully inserted item %p at level %llu", item, level);
-                break; // success
-            }
-            TRACE("s3", "sl_add: failed to change pred's link: expected %p found %p", next, other);
-
         } while (1);
     }
-    return value;
+    return val;
 }
 
 uint64_t sl_remove (skiplist_t *sl, const void *key_data, uint32_t key_len) {
-    TRACE("s3", "sl_remove: removing item with key %p from sl %p", key_data, sl);
-    int found;
+    TRACE("s3", "sl_remove: removing item with key %p from skiplist %p", key_data, sl);
     node_t *preds[MAX_LEVEL+1];
-    node_t *item = find_preds(&found, preds, -1, sl, key_data, key_len, TRUE);
-    if (!found) {
-        TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the sl", 0, 0);
+    node_t *item = find_preds(preds, NULL, -1, sl, key_data, key_len, TRUE);
+    if (item == NULL) {
+        TRACE("s3", "sl_remove: remove failed, an item with a matching key does not exist in the skiplist", 0, 0);
         return DOES_NOT_EXIST;
     }
 
@@ -282,9 +291,11 @@ uint64_t sl_remove (skiplist_t *sl, const void *key_data, uint32_t key_len) {
         }
     }
 
-    uint64_t value = item->value;
+    uint64_t val = item->val;
 
-    // Unlink <item> from the top down.
+    // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
+    // item partially unlinked for a later call (or some other thread) to physically unlink. By marking the
+    // item earlier, we logically removed it. 
     int level = item->top_level;
     while (level >= 0) {
         node_t *pred = preds[level];
@@ -293,16 +304,14 @@ uint64_t sl_remove (skiplist_t *sl, const void *key_data, uint32_t key_len) {
         node_t *other = NULL;
         if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) != item) {
             TRACE("s3", "sl_remove: unlink failed; pred's link changed from %p to %p", item, other);
-            // By marking the item earlier, we logically removed it. It is safe to leave the item partially
-            // unlinked. Another thread will finish physically removing it from <sl>.
-            return value;
+            return val;
         }
         --level; 
     }
 
     // The thread that completes the unlink should free the memory.
-    nbd_defer_free(item); 
-    return value;
+    node_defer_free(item); 
+    return val;
 }
 
 void sl_print (skiplist_t *sl) {
@@ -324,7 +333,12 @@ void sl_print (skiplist_t *sl) {
     node_t *item = sl->head;
     while (item) {
         int is_marked = IS_TAGGED(item->next[0]);
-        printf("%s%p:%s ", is_marked ? "*" : "", item, (char *)ns_data(item->key));
+
+        if (IS_TAGGED(item->key)) {
+            printf("%s%p:%llx ", is_marked ? "*" : "", item, STRIP_TAG(item->key));
+        } else {
+            printf("%s%p:%s ", is_marked ? "*" : "", item, (char *)ns_data(item->key));
+        }
         if (item != sl->head) {
             printf("[%d]", item->top_level);
         } else {