#include "common.h"
#include "runtime.h"
#include "struct.h"
+#include "nstring.h"
#include "mem.h"
#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 {
- uint64_t key;
+ nstring_t *key;
uint64_t value;
int top_level;
struct node *next[];
} node_t;
-typedef struct sl {
+struct sl {
node_t *head;
-} skiplist_t;
+};
static int random_level (void) {
unsigned r = nbd_rand();
return n;
}
-node_t *node_alloc (int level, uint64_t key, uint64_t value) {
+node_t *node_alloc (int level, const void *key_data, uint32_t key_len, uint64_t value) {
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 = key;
+ // 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->value = value;
item->top_level = level;
return item;
skiplist_t *sl_alloc (void) {
skiplist_t *sl = (skiplist_t *)nbd_malloc(sizeof(skiplist_t));
- sl->head = node_alloc(MAX_LEVEL, 0, 0);
+ sl->head = node_alloc(MAX_LEVEL, " ", 0, 0);
memset(sl->head->next, 0, (MAX_LEVEL+1) * sizeof(skiplist_t *));
return sl;
}
-static node_t *find_preds (node_t *preds[MAX_LEVEL+1], int n, skiplist_t *sl, uint64_t key, 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, pred);
-
+ TRACE("s3", "find_preds: searching for key %p in sl (head is %p)", key_data, pred);
+ int x;
int start_level = MAX_LEVEL;
#if MAX_LEVEL > 2
// Optimization for small lists. No need to traverse empty higher levels.
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(preds, n, sl, key, 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->value);
- // Marked items are logically removed, but not fully unlinked yet.
+ // A tag means an item is logically removed but not physically unlinked yet.
while (EXPECT_FALSE(IS_TAGGED(next))) {
- // Skip over partially removed items.
+ // Skip over logically removed items.
if (!help_remove) {
item = (node_t *)STRIP_TAG(item->next);
if (EXPECT_FALSE(item == NULL))
continue;
}
- // Unlink partially removed items.
+ // Unlink logically removed items.
node_t *other;
if ((other = SYNC_CAS(&pred->next[level], item, STRIP_TAG(next))) == item) {
item = (node_t *)STRIP_TAG(next);
} 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(preds, n, sl, key, 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;
break;
// If we reached the key (or passed where it should be), we found a pred. Save it and continue down.
- if (item->key >= key) {
+ 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) {
TRACE("s3", "find_preds: found pred %p item %p", pred, item);
break;
}
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, uint64_t key) {
+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);
- node_t *item = find_preds(NULL, 0, sl, key, FALSE);
+ 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 (item && item->key == key) ? item->value : DOES_NOT_EXIST;
+ return item != NULL ? item->value : DOES_NOT_EXIST;
}
// Insert the <key> if it doesn't already exist in <sl>
-uint64_t sl_add (skiplist_t *sl, uint64_t key, 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 value) {
+ TRACE("s3", "sl_add: inserting key %p value %p", key_data, value);
node_t *preds[MAX_LEVEL+1];
+ node_t *nexts[MAX_LEVEL+1];
node_t *item = NULL;
+ int n = random_level();
do {
- int n = random_level();
- node_t *next = find_preds(preds, n, sl, key, TRUE);
+ node_t *next = find_preds(preds, nexts, n, sl, key_data, key_len, TRUE);
// If a node matching <key> already exists in <sl>, return its value.
- if (next != NULL && next->key == key) {
- TRACE("s3", "sl_add: there is already an item %p (value %p) with the same key", next, next->value);
+ if (next != NULL) {
+ TRACE("s3", "sl_add: there is already an item %p (value %p) with the same key", nexts[0], nexts[0]->value);
if (EXPECT_FALSE(item != NULL)) { nbd_free(item); }
- return next->value;
+ return nexts[0]->value;
}
// First insert <item> into the bottom level.
- if (EXPECT_TRUE(item == NULL)) { item = node_alloc(n, key, value); }
- TRACE("s3", "sl_add: attempting to insert item between %p and %p", preds[0], next);
- item->next[0] = next;
+ if (EXPECT_TRUE(item == NULL)) { item = node_alloc(n, key_data, key_len, value); }
+ node_t *pred = preds[0];
+ item->next[0] = next = nexts[0];
+ TRACE("s3", "sl_add: attempting to insert item between %p and %p", pred, next);
for (int level = 1; level <= item->top_level; ++level) {
- node_t *pred = preds[level];
- item->next[level] = pred->next[level];
+ item->next[level] = nexts[level];
}
- node_t *pred = preds[0];
node_t *other = SYNC_CAS(&pred->next[0], next, item);
if (other == next) {
TRACE("s3", "sl_add: successfully inserted item %p at level 0", item, 0);
// Insert <item> into <sl> from the bottom level up.
for (int level = 1; level <= 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) && next->key > key) // pred's link changed
- break;
- find_preds(preds, item->top_level, sl, key, TRUE);
- } while (1);
-
- 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];
- if (IS_TAGGED(old_next))
- return DOES_NOT_EXIST; // success
-
- // 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)
- 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) {
break; // success
}
TRACE("s3", "sl_add: failed to change pred's link: expected %p found %p", next, other);
+ find_preds(preds, nexts, item->top_level, sl, key_data, key_len, TRUE);
+ pred = preds[level];
+ next = nexts[level];
+
+ // Update <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];
+ if (IS_TAGGED(old_next))
+ return value;
+ // Use a CAS so we to not inadvertantly stomp on a mark another thread placed on the item.
+ if (old_next == next || SYNC_CAS(&item->next[level], old_next, next) == old_next)
+ break;
+ } while (1);
} while (1);
}
return value;
}
-uint64_t sl_remove (skiplist_t *sl, uint64_t key) {
- TRACE("s3", "sl_remove: removing item with key %p from sl %p", key, sl);
+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);
node_t *preds[MAX_LEVEL+1];
- node_t *item = find_preds(preds, -1, sl, key, TRUE);
- if (item == NULL || item->key != key) {
+ 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 sl", 0, 0);
return DOES_NOT_EXIST;
}
node_t *item = sl->head;
while (item) {
int is_marked = IS_TAGGED(item->next[0]);
- printf("%s%p:0x%llx ", is_marked ? "*" : "", item, 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 {
item = (node_t *)STRIP_TAG(item->next[0]);
}
}
-
-#ifdef MAKE_skiplist_test
-#include <errno.h>
-#include <pthread.h>
-#include <sys/time.h>
-
-#include "runtime.h"
-
-#define NUM_ITERATIONS 10000000
-
-static volatile int wait_;
-static long num_threads_;
-static skiplist_t *sl_;
-
-void *worker (void *arg) {
-
- // Wait for all the worker threads to be ready.
- SYNC_ADD(&wait_, -1);
- do {} while (wait_);
-
- for (int i = 0; i < NUM_ITERATIONS/num_threads_; ++i) {
- unsigned r = nbd_rand();
- int key = r & 0xF;
- if (r & (1 << 8)) {
- sl_add(sl_, key, 1);
- } else {
- sl_remove(sl_, key);
- }
-
- rcu_update();
- }
-
- return NULL;
-}
-
-int main (int argc, char **argv) {
- nbd_init();
- lwt_set_trace_level("s3");
-
- char* program_name = argv[0];
- pthread_t thread[MAX_NUM_THREADS];
-
- if (argc > 2) {
- fprintf(stderr, "Usage: %s num_threads\n", program_name);
- return -1;
- }
-
- num_threads_ = 2;
- if (argc == 2)
- {
- errno = 0;
- num_threads_ = strtol(argv[1], NULL, 10);
- if (errno) {
- fprintf(stderr, "%s: Invalid argument for number of threads\n", program_name);
- return -1;
- }
- if (num_threads_ <= 0) {
- fprintf(stderr, "%s: Number of threads must be at least 1\n", program_name);
- return -1;
- }
- if (num_threads_ > MAX_NUM_THREADS) {
- fprintf(stderr, "%s: Number of threads cannot be more than %d\n", program_name, MAX_NUM_THREADS);
- return -1;
- }
- }
-
- sl_ = sl_alloc();
-
- struct timeval tv1, tv2;
- gettimeofday(&tv1, NULL);
-
- wait_ = num_threads_;
-
- for (int i = 0; i < num_threads_; ++i) {
- int rc = nbd_thread_create(thread + i, i, worker, (void*)(size_t)i);
- if (rc != 0) { perror("pthread_create"); return rc; }
- }
-
- for (int i = 0; i < num_threads_; ++i) {
- pthread_join(thread[i], NULL);
- }
-
- gettimeofday(&tv2, NULL);
- int ms = (int)(1000000*(tv2.tv_sec - tv1.tv_sec) + tv2.tv_usec - tv1.tv_usec) / 1000;
- sl_print(sl_);
- printf("Th:%ld Time:%dms\n", num_threads_, ms);
-
- return 0;
-}
-#endif//skiplist_test