#include "common.h"
#include "struct.h"
+#include "nstring.h"
#include "mem.h"
typedef struct node {
- uint64_t key;
- uint64_t value;
+ nstring_t *key;
+ uint64_t val;
struct node *next;
} node_t;
-typedef struct list {
+struct ll {
node_t *head;
- node_t *last;
-} list_t;
+};
-node_t *node_alloc (uint64_t key, uint64_t value) {
+static node_t *node_alloc (const void *key_data, uint32_t key_len, uint64_t val) {
node_t *item = (node_t *)nbd_malloc(sizeof(node_t));
memset(item, 0, sizeof(node_t));
- item->key = key;
- 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;
return item;
}
-list_t *list_alloc (void) {
- list_t *list = (list_t *)nbd_malloc(sizeof(list_t));
- list->head = node_alloc(0, 0);
- list->last = node_alloc((uint64_t)-1, 0);
- list->head->next = list->last;
- return list;
+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);
+}
+
+list_t *ll_alloc (void) {
+ list_t *ll = (list_t *)nbd_malloc(sizeof(list_t));
+ ll->head = node_alloc(" ", 0, 0);
+ ll->head->next = NULL;
+ return ll;
}
-static node_t *find_pred (node_t **pred_ptr, list_t *list, uint64_t key, int help_remove) {
- node_t *pred = list->head;
+static int find_pred (node_t **pred_ptr, node_t **item_ptr, list_t *ll, const void *key_data, uint32_t key_len, int help_remove) {
+ node_t *pred = ll->head;
node_t *item = pred->next;
- TRACE("l3", "find_pred: searching for key %p in list (head is %p)", key, pred);
-#ifndef NDEBUG
- int count = 0;
-#endif
+ TRACE("l2", "find_pred: searching for key %p in ll (head is %p)", key_data, pred);
- do {
+ while (item != NULL) {
node_t *next = item->next;
- TRACE("l3", "find_pred: visiting item %p (next %p)", item, next);
- TRACE("l3", "find_pred: key %p", item->key, item->value);
- // Marked items are logically removed, but not 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))
+ break;
+ TRACE("l3", "find_pred: skipping marked item %p (next is %p)", item, next);
next = item->next;
continue;
}
- // Unlink partially removed items.
+ // Unlink logically removed items.
node_t *other;
+ TRACE("l3", "find_pred: unlinking marked item %p next is %p", item, next);
if ((other = SYNC_CAS(&pred->next, item, STRIP_TAG(next))) == item) {
+ TRACE("l2", "find_pred: unlinked item %p from pred %p", item, pred);
item = (node_t *)STRIP_TAG(next);
+ if (EXPECT_FALSE(item == NULL))
+ break;
next = item->next;
- TRACE("l3", "find_pred: unlinked item %p from pred %p", item, pred);
- TRACE("l3", "find_pred: now item is %p next is %p", item, next);
+ TRACE("l3", "find_pred: now current item is %p next is %p", item, next);
// The thread that completes the unlink should free the memory.
- nbd_defer_free(other);
+ node_defer_free(other);
} else {
- TRACE("l3", "find_pred: lost race to unlink item from pred %p; its link changed to %p", pred, other);
+ TRACE("l2", "find_pred: lost a race to unlink item %p from pred %p", item, pred);
+ TRACE("l2", "find_pred: pred's link changed to %p", other, 0);
if (IS_TAGGED(other))
- return find_pred(pred_ptr, list, key, help_remove); // retry
+ return find_pred(pred_ptr, item_ptr, ll, key_data, key_len, help_remove); // retry
item = other;
+ if (EXPECT_FALSE(item == NULL))
+ break;
next = item->next;
}
}
+ if (EXPECT_FALSE(item == NULL))
+ break;
+
+ TRACE("l3", "find_pred: visiting item %p (next is %p)", item, next);
+ TRACE("l4", "find_pred: key %p val %p", STRIP_TAG(item->key), item->val);
+
+ // A tagged key is an integer, otherwise it is a pointer to a string
+ int d;
+ if (IS_TAGGED(item->key)) {
+ d = (STRIP_TAG(item->key) - (uint64_t)key_data);
+ } else {
+ int item_key_len = item->key->len;
+ int len = (key_len < item_key_len) ? key_len : item_key_len;
+ d = memcmp(item->key->data, key_data, len);
+ if (d == 0) { d = item_key_len - key_len; }
+ }
+
// If we reached the key (or passed where it should be), we found the right predesssor
- if (item->key >= key) {
- TRACE("l3", "find_pred: found pred %p item %p", pred, item);
+ if (d >= 0) {
if (pred_ptr != NULL) {
*pred_ptr = pred;
}
- return item;
+ *item_ptr = item;
+ if (d == 0) {
+ TRACE("l2", "find_pred: found matching item %p in list, pred is %p", item, pred);
+ return TRUE;
+ }
+ TRACE("l2", "find_pred: found proper place for key %p in list, pred is %p", key_data, pred);
+ return FALSE;
}
- assert(count++ < 18);
pred = item;
item = next;
+ }
- } while (1);
+ // <key> is not in <ll>.
+ if (pred_ptr != NULL) {
+ *pred_ptr = pred;
+ }
+ *item_ptr = NULL;
+ TRACE("l2", "find_pred: reached end of list. last item is %p", pred, 0);
+ return FALSE;
}
// Fast find. Do not help unlink partially removed nodes and do not return the found item's predecessor.
-uint64_t list_lookup (list_t *list, uint64_t key) {
- TRACE("l3", "list_lookup: searching for key %p in list %p", key, list);
- node_t *item = find_pred(NULL, list, key, FALSE);
+uint64_t ll_lookup (list_t *ll, const void *key_data, uint32_t key_len) {
+ TRACE("l1", "ll_lookup: searching for key %p in list %p", key_data, ll);
+ node_t *item;
+ int found = find_pred(NULL, &item, ll, key_data, key_len, FALSE);
+
+ // If we found an <item> matching the key return its value.
+ if (found) {
+ uint64_t val = item->val;
+ if (val != DOES_NOT_EXIST) {
+ TRACE("l1", "ll_lookup: found item %p. val %p. returning item", item, item->val);
+ return val;
+ }
+ }
- // If we found an <item> matching the <key> return its value.
- return (item->key == key) ? item->value : DOES_NOT_EXIST;
+ TRACE("l1", "ll_lookup: no item in the list matched the key", 0, 0);
+ return DOES_NOT_EXIST;
}
-// Insert the <key>, if it doesn't already exist in the <list>
-uint64_t list_add (list_t *list, uint64_t key, uint64_t value) {
- TRACE("l3", "list_add: inserting key %p value %p", key, value);
- node_t *pred;
- node_t *item = NULL;
+uint64_t ll_cas (list_t *ll, const void *key_data, uint32_t key_len, uint64_t expectation, uint64_t new_val) {
+ TRACE("l1", "ll_cas: key %p list %p", key_data, ll);
+ TRACE("l1", "ll_cas: expectation %p new value %p", expectation, new_val);
+ ASSERT((int64_t)new_val > 0);
+
do {
- node_t *next = find_pred(&pred, list, key, TRUE);
+ node_t *pred, *old_item;
+ if (!find_pred(&pred, &old_item, ll, key_data, key_len, TRUE)) {
- // If a node matching <key> already exists in the list, return its value.
- if (next->key == key) {
- TRACE("l3", "list_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;
- }
+ // There was not an item in the list that matches the key.
+ if (EXPECT_FALSE((int64_t)expectation > 0 || expectation == EXPECT_EXISTS)) {
+ TRACE("l1", "ll_cas: the expectation was not met, the list was not changed", 0, 0);
+ return DOES_NOT_EXIST; // failure
+ }
+
+ ASSERT(expectation == EXPECT_DOES_NOT_EXIST || expectation == EXPECT_WHATEVER);
+
+ // Create a new item and insert it into the list.
+ TRACE("l2", "ll_cas: attempting to insert item between %p and %p", pred, pred->next);
+ node_t *new_item = node_alloc(key_data, key_len, new_val);
+ node_t *next = new_item->next = old_item;
+ node_t *other = SYNC_CAS(&pred->next, next, new_item);
+ if (other == next) {
+ TRACE("l1", "ll_cas: successfully inserted new item %p", new_item, 0);
+ return DOES_NOT_EXIST; // success
+ }
- TRACE("l3", "list_add: attempting to insert item between %p and %p", pred, next);
- if (EXPECT_TRUE(item == NULL)) { item = node_alloc(key, value); }
- item->next = next;
- node_t *other = SYNC_CAS(&pred->next, next, item);
- if (other == next) {
- TRACE("l3", "list_add: successfully inserted item %p", item, 0);
- return DOES_NOT_EXIST; // success
+ // Lost a race. Failed to insert the new item into the list.
+ TRACE("l1", "ll_cas: lost a race. CAS failed. expected pred's link to be %p but found %p", next, other);
+ node_free(new_item);
+ continue; // retry
}
- TRACE("l3", "list_add: failed to change pred's link: expected %p found %p", next, other);
+ // Found an item in the list that matches the key.
+ uint64_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)) {
+ TRACE("l2", "ll_cas: lost a race, found an item but another thread removed it. retry", 0, 0);
+ break; // retry
+ }
+
+ if (EXPECT_FALSE(expectation == EXPECT_DOES_NOT_EXIST)) {
+ TRACE("l1", "ll_cas: found an item %p in the list that matched the key. the expectation was "
+ "not met, the list was not changed", old_item, old_item_val);
+ return old_item_val; // failure
+ }
+
+ // Use a CAS and not a SWAP. If the node is in the process of being removed and we used a SWAP, we could
+ // 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);
+ if (ret_val == old_item_val) {
+ TRACE("l1", "ll_cas: the CAS succeeded. updated the value of the item", 0, 0);
+ return ret_val; // success
+ }
+ TRACE("l2", "ll_cas: lost a race. the CAS failed. another thread changed the item's value", 0, 0);
+
+ old_item_val = ret_val;
+ } while (1);
} while (1);
}
-uint64_t list_remove (list_t *list, uint64_t key) {
- TRACE("l3", "list_remove: removing item with key %p from list %p", key, list);
+uint64_t ll_remove (list_t *ll, const void *key_data, uint32_t key_len) {
+ TRACE("l1", "ll_remove: removing item with key %p from list %p", key_data, ll);
node_t *pred;
- node_t *item = find_pred(&pred, list, key, TRUE);
- if (item->key != key) {
- TRACE("l3", "list_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
+ node_t *item;
+ int found = find_pred(&pred, &item, ll, key_data, key_len, TRUE);
+ if (!found) {
+ TRACE("l1", "ll_remove: remove failed, an item with a matching key does not exist in the list", 0, 0);
return DOES_NOT_EXIST;
}
// Mark <item> removed. This must be atomic. If multiple threads try to remove the same item
// only one of them should succeed.
- if (EXPECT_FALSE(IS_TAGGED(item->next))) {
- TRACE("l3", "list_remove: %p is already marked for removal by another thread", item, 0);
- return DOES_NOT_EXIST;
- }
- node_t *next = SYNC_FETCH_AND_OR(&item->next, TAG);
- if (EXPECT_FALSE(IS_TAGGED(next))) {
- TRACE("l3", "list_remove: lost race -- %p is already marked for removal by another thread", item, 0);
- return DOES_NOT_EXIST;
- }
-
- uint64_t value = item->value;
-
- // Unlink <item> from the list.
- TRACE("l3", "list_remove: link item's pred %p to it's successor %p", pred, next);
+ node_t *next;
+ node_t *old_next = item->next;
+ do {
+ next = old_next;
+ old_next = SYNC_CAS(&item->next, next, TAG_VALUE(next));
+ if (IS_TAGGED(old_next)) {
+ TRACE("l1", "ll_remove: lost a race -- %p is already marked for removal by another thread", item, 0);
+ return DOES_NOT_EXIST;
+ }
+ } while (next != old_next);
+ TRACE("l2", "ll_remove: logically removed item %p", item, 0);
+ ASSERT(IS_TAGGED(item->next));
+
+ // This has to be an atomic swap in case another thread is updating the item while we are removing it.
+ uint64_t val = SYNC_SWAP(&item->val, DOES_NOT_EXIST);
+ TRACE("l2", "ll_remove: replaced item's val %p with DOES_NOT_EXIT", val, 0);
+
+ // Unlink <item> from <ll>. If we lose a race to another thread just back off. It is safe to leave the
+ // item logically removed for a later call (or some other thread) to physically unlink. By marking the
+ // item earlier, we logically removed it.
+ TRACE("l2", "ll_remove: unlink the item by linking its pred %p to its successor %p", pred, next);
node_t *other;
if ((other = SYNC_CAS(&pred->next, item, next)) != item) {
- TRACE("l3", "list_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.
- // Another thread will finish physically removing it from the list.
- return value;
+ TRACE("l1", "ll_remove: unlink failed; pred's link changed from %p to %p", item, other);
+ return val;
}
// The thread that completes the unlink should free the memory.
- nbd_defer_free(item);
- return value;
+ node_defer_free(item);
+ TRACE("l1", "ll_remove: successfully unlinked item %p from the list", item, 0);
+ return val;
}
-void list_print (list_t *list) {
+void ll_print (list_t *ll) {
node_t *item;
- item = list->head;
+ item = ll->head->next;
while (item) {
- printf("0x%llx ", item->key);
- fflush(stdout);
- item = item->next;
- }
- printf("\n");
-}
-
-#ifdef MAKE_list_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 list_t *list_;
-
-void *worker (void *arg) {
- int id = (int)(size_t)arg;
-
- unsigned int rand_seed = id+1;//rdtsc_l();
-
- // 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) {
- int n = rand_r(&rand_seed);
- int key = (n & 0xF) + 1;
- if (n & (1 << 8)) {
- list_add(list_, key, 1);
- } else {
- list_remove(list_, key);
- }
-
- rcu_update();
- }
-
- return NULL;
-}
-
-int main (int argc, char **argv) {
- nbd_init();
- //lwt_set_trace_level("m0l0");
-
- 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;
+ node_t *next = item->next;
+ if (IS_TAGGED(item)) {
+ printf("*");
}
- 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;
+ printf("%p:", item);
+ if (IS_TAGGED(item->key)) {
+ printf("0x%llx ", STRIP_TAG(item->key));
+ } else {
+ printf("%s ", (char *)item->key->data);
}
+ fflush(stdout);
+ item = (node_t *)STRIP_TAG(next);
}
-
- list_ = list_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;
- printf("Th:%ld Time:%dms\n", num_threads_, ms);
- list_print(list_);
- lwt_dump("lwt.out");
-
- return 0;
+ printf("\n");
}
-#endif//list_test
projects / nbds / blobdiff