+ dbg_printf("init new size: %lu\n", 1UL << i);
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy))
+ break;
+ }
+}
+
+/*
+ * Holding RCU read lock to protect _cds_lfht_remove against memory
+ * reclaim that could be performed by other call_rcu worker threads (ABA
+ * problem).
+ * For a single level, we logically remove and garbage collect each node.
+ *
+ * As a design choice, we perform logical removal and garbage collection on a
+ * node-per-node basis to simplify this algorithm. We also assume keeping good
+ * cache locality of the operation would overweight possible performance gain
+ * that could be achieved by batching garbage collection for multiple levels.
+ * However, this would have to be justified by benchmarks.
+ *
+ * Concurrent removal and add operations are helping us perform garbage
+ * collection of logically removed nodes. We guarantee that all logically
+ * removed nodes have been garbage-collected (unlinked) before call_rcu is
+ * invoked to free a hole level of dummy nodes (after a grace period).
+ *
+ * Logical removal and garbage collection can therefore be done in batch or on a
+ * node-per-node basis, as long as the guarantee above holds.
+ *
+ * When we reach a certain length, we can split this removal over many worker
+ * threads, based on the number of CPUs available in the system. This should
+ * take care of not letting resize process lag behind too many concurrent
+ * updater threads actively inserting into the hash table.
+ */
+static
+void remove_table_partition(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len)
+{
+ unsigned long j;
+
+ assert(i > ht->min_alloc_order);
+ ht->cds_lfht_rcu_read_lock();
+ for (j = start; j < start + len; j++) {
+ struct cds_lfht_node *fini_node =
+ (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
+
+ dbg_printf("remove entry: i %lu j %lu hash %lu\n",
+ i, j, (1UL << (i - 1)) + j);
+ fini_node->p.reverse_hash =
+ bit_reverse_ulong((1UL << (i - 1)) + j);
+ (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
+ }
+ ht->cds_lfht_rcu_read_unlock();
+}
+
+static
+void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
+{
+
+ assert(nr_cpus_mask != -1);
+ if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
+ ht->cds_lfht_rcu_thread_online();
+ remove_table_partition(ht, i, 0, len);
+ ht->cds_lfht_rcu_thread_offline();
+ return;
+ }
+ partition_resize_helper(ht, i, len, remove_table_partition);
+}
+
+static
+void fini_table(struct cds_lfht *ht,
+ unsigned long first_order, unsigned long last_order)
+{
+ long i;
+ void *free_by_rcu = NULL;
+
+ dbg_printf("fini table: first_order %lu last_order %lu\n",
+ first_order, last_order);
+ assert(first_order > ht->min_alloc_order);
+ for (i = last_order; i >= first_order; i--) {
+ unsigned long len;
+
+ len = 1UL << (i - 1);
+ dbg_printf("fini order %lu len: %lu\n", i, len);
+
+ /* Stop shrink if the resize target changes under us */
+ if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
+ break;
+
+ cmm_smp_wmb(); /* populate data before RCU size */
+ CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
+
+ /*
+ * We need to wait for all add operations to reach Q.S. (and
+ * thus use the new table for lookups) before we can start
+ * releasing the old dummy nodes. Otherwise their lookup will
+ * return a logically removed node as insert position.
+ */
+ ht->cds_lfht_synchronize_rcu();
+ if (free_by_rcu)
+ free(free_by_rcu);
+
+ /*
+ * Set "removed" flag in dummy nodes about to be removed.
+ * Unlink all now-logically-removed dummy node pointers.
+ * Concurrent add/remove operation are helping us doing
+ * the gc.
+ */
+ remove_table(ht, i, len);
+
+ free_by_rcu = ht->t.tbl[i];
+
+ dbg_printf("fini new size: %lu\n", 1UL << i);
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy))
+ break;
+ }
+
+ if (free_by_rcu) {
+ ht->cds_lfht_synchronize_rcu();
+ free(free_by_rcu);
+ }
+}
+
+static
+void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
+{
+ struct _cds_lfht_node *prev, *node;
+ unsigned long order, len, i, j;
+
+ ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct _cds_lfht_node));
+ assert(ht->t.tbl[0]);
+
+ dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
+ ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end());
+ ht->t.tbl[0]->nodes[0].reverse_hash = 0;
+
+ for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
+ len = 1UL << (order - 1);
+ if (order <= ht->min_alloc_order) {
+ ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
+ } else {
+ ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
+ assert(ht->t.tbl[order]);
+ }
+
+ i = 0;
+ prev = ht->t.tbl[i]->nodes;
+ for (j = 0; j < len; j++) {
+ if (j & (j - 1)) { /* Between power of 2 */
+ prev++;
+ } else if (j) { /* At each power of 2 */
+ i++;
+ prev = ht->t.tbl[i]->nodes;
+ }
+
+ node = &ht->t.tbl[order]->nodes[j];
+ dbg_printf("create dummy: order %lu index %lu hash %lu\n",
+ order, j, j + len);
+ node->next = prev->next;
+ assert(is_dummy(node->next));
+ node->reverse_hash = bit_reverse_ulong(j + len);
+ prev->next = flag_dummy((struct cds_lfht_node *)node);
+ }
+ }
+}
+
+struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
+ cds_lfht_compare_fct compare_fct,
+ unsigned long hash_seed,
+ unsigned long init_size,
+ unsigned long min_alloc_size,
+ int flags,
+ void (*cds_lfht_call_rcu)(struct rcu_head *head,
+ void (*func)(struct rcu_head *head)),
+ void (*cds_lfht_synchronize_rcu)(void),
+ void (*cds_lfht_rcu_read_lock)(void),
+ void (*cds_lfht_rcu_read_unlock)(void),
+ void (*cds_lfht_rcu_thread_offline)(void),
+ void (*cds_lfht_rcu_thread_online)(void),
+ void (*cds_lfht_rcu_register_thread)(void),
+ void (*cds_lfht_rcu_unregister_thread)(void),
+ pthread_attr_t *attr)
+{
+ struct cds_lfht *ht;
+ unsigned long order;
+
+ /* min_alloc_size must be power of two */
+ if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
+ return NULL;
+ /* init_size must be power of two */
+ if (!init_size || (init_size & (init_size - 1)))
+ return NULL;
+ min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
+ init_size = max(init_size, min_alloc_size);
+ ht = calloc(1, sizeof(struct cds_lfht));
+ assert(ht);