+ * TODO: use ja_get_nr_child to monitor limits triggering shrink
+ * recompaction.
+ * Also use ja_get_nr_child to make the difference between resize and
+ * pool change of compaction bit(s).
+ */
+static
+unsigned int ja_get_nr_child(struct cds_ja_shadow_node *shadow_node)
+{
+ return shadow_node->nr_child;
+}
+
+static
+int ja_linear_node_set_nth(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_shadow_node *shadow_node,
+ uint8_t n,
+ struct cds_ja_inode_flag *child_node_flag)
+{
+ uint8_t nr_child;
+ uint8_t *values, *nr_child_ptr;
+ struct cds_ja_inode_flag **pointers;
+ unsigned int i;
+
+ assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL);
+
+ nr_child_ptr = &node->u.data[0];
+ dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr);
+ nr_child = *nr_child_ptr;
+ assert(nr_child <= type->max_linear_child);
+
+ values = &node->u.data[1];
+ for (i = 0; i < nr_child; i++) {
+ if (values[i] == n)
+ return -EEXIST;
+ }
+ if (nr_child >= type->max_linear_child) {
+ /* No space left in this node type */
+ return -ENOSPC;
+ }
+ pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]);
+ assert(pointers[nr_child] == NULL);
+ rcu_assign_pointer(pointers[nr_child], child_node_flag);
+ CMM_STORE_SHARED(values[nr_child], n);
+ cmm_smp_wmb(); /* write value and pointer before nr_child */
+ CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1);
+ shadow_node->nr_child++;
+ dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
+ (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr),
+ (unsigned int) shadow_node->nr_child,
+ node, shadow_node);
+
+ return 0;
+}
+
+static
+int ja_pool_node_set_nth(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_shadow_node *shadow_node,
+ uint8_t n,
+ struct cds_ja_inode_flag *child_node_flag)
+{
+ struct cds_ja_inode *linear;
+
+ assert(type->type_class == RCU_JA_POOL);
+ linear = (struct cds_ja_inode *)
+ &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order];
+ return ja_linear_node_set_nth(type, linear, shadow_node,
+ n, child_node_flag);
+}
+
+static
+int ja_pigeon_node_set_nth(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_shadow_node *shadow_node,
+ uint8_t n,
+ struct cds_ja_inode_flag *child_node_flag)
+{
+ struct cds_ja_inode_flag **ptr;
+
+ assert(type->type_class == RCU_JA_PIGEON);
+ ptr = &((struct cds_ja_inode_flag **) node->u.data)[n];
+ if (*ptr)
+ return -EEXIST;
+ rcu_assign_pointer(*ptr, child_node_flag);
+ shadow_node->nr_child++;
+ return 0;
+}
+
+/*
+ * _ja_node_set_nth: set nth item within a node. Return an error
+ * (negative error value) if it is already there.
+ * TODO: exclusive access on node.
+ */
+static
+int _ja_node_set_nth(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_shadow_node *shadow_node,
+ uint8_t n,
+ struct cds_ja_inode_flag *child_node_flag)
+{
+ switch (type->type_class) {
+ case RCU_JA_LINEAR:
+ return ja_linear_node_set_nth(type, node, shadow_node, n,
+ child_node_flag);
+ case RCU_JA_POOL:
+ return ja_pool_node_set_nth(type, node, shadow_node, n,
+ child_node_flag);
+ case RCU_JA_PIGEON:
+ return ja_pigeon_node_set_nth(type, node, shadow_node, n,
+ child_node_flag);
+ case RCU_JA_NULL:
+ return -ENOSPC;
+ default:
+ assert(0);
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+/*
+ * ja_node_recompact_add: recompact a node, adding a new child.
+ * TODO: for pool type, take selection bit(s) into account.
+ * Return 0 on success, -ENOENT if need to retry, or other negative
+ * error value otherwise.
+ */
+static
+int ja_node_recompact_add(struct cds_ja *ja,
+ unsigned int old_type_index,
+ const struct cds_ja_type *old_type,
+ struct cds_ja_inode *old_node,
+ struct cds_ja_shadow_node *shadow_node,
+ struct cds_ja_inode_flag **old_node_flag, uint8_t n,
+ struct cds_ja_inode_flag *child_node_flag)
+{
+ unsigned int new_type_index;
+ struct cds_ja_inode *new_node;
+ struct cds_ja_shadow_node *new_shadow_node;
+ const struct cds_ja_type *new_type;
+ struct cds_ja_inode_flag *new_node_flag;
+ int ret;
+ int fallback = 0;
+
+ if (!shadow_node || old_type_index == NODE_INDEX_NULL) {
+ new_type_index = 0;
+ } else {
+ new_type_index = old_type_index + 1;
+ }
+
+retry: /* for fallback */
+ dbg_printf("Recompact to type %d\n", new_type_index);
+ new_type = &ja_types[new_type_index];
+ new_node = alloc_cds_ja_node(new_type);
+ if (!new_node)
+ return -ENOMEM;
+ new_node_flag = ja_node_flag(new_node, new_type_index);
+
+ dbg_printf("Recompact inherit lock from %p\n", shadow_node);
+ new_shadow_node = rcuja_shadow_set(ja->ht, new_node, shadow_node);
+ if (!new_shadow_node) {
+ free(new_node);
+ return -ENOMEM;
+ }
+ if (fallback)
+ new_shadow_node->fallback_removal_count =
+ JA_FALLBACK_REMOVAL_COUNT;
+
+ assert(old_type->type_class != RCU_JA_PIGEON);
+ switch (old_type->type_class) {
+ case RCU_JA_LINEAR:
+ {
+ uint8_t nr_child =
+ ja_linear_node_get_nr_child(old_type, old_node);
+ unsigned int i;
+
+ for (i = 0; i < nr_child; i++) {
+ struct cds_ja_inode_flag *iter;
+ uint8_t v;
+
+ ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter);
+ if (!iter)
+ continue;
+ ret = _ja_node_set_nth(new_type, new_node,
+ new_shadow_node,
+ v, iter);
+ if (new_type->type_class == RCU_JA_POOL && ret) {
+ goto fallback_toosmall;
+ }
+ assert(!ret);
+ }
+ break;
+ }
+ case RCU_JA_POOL:
+ {
+ unsigned int pool_nr;
+
+ for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) {
+ struct cds_ja_inode *pool =
+ ja_pool_node_get_ith_pool(old_type,
+ old_node, pool_nr);
+ uint8_t nr_child =
+ ja_linear_node_get_nr_child(old_type, pool);
+ unsigned int j;
+
+ for (j = 0; j < nr_child; j++) {
+ struct cds_ja_inode_flag *iter;
+ uint8_t v;
+
+ ja_linear_node_get_ith_pos(old_type, pool,
+ j, &v, &iter);
+ if (!iter)
+ continue;
+ ret = _ja_node_set_nth(new_type, new_node,
+ new_shadow_node,
+ v, iter);
+ if (new_type->type_class == RCU_JA_POOL
+ && ret) {
+ goto fallback_toosmall;
+ }
+ assert(!ret);
+ }
+ }
+ break;
+ }
+ case RCU_JA_NULL:
+ /* Nothing to copy */
+ break;
+ case RCU_JA_PIGEON:
+ default:
+ assert(0);
+ ret = -EINVAL;
+ goto end;
+ }
+
+ /* add node */
+ ret = _ja_node_set_nth(new_type, new_node,
+ new_shadow_node,
+ n, child_node_flag);
+ assert(!ret);
+ /* Return pointer to new recompacted new through old_node_flag */
+ *old_node_flag = new_node_flag;
+ if (old_node) {
+ ret = rcuja_shadow_clear(ja->ht, old_node, shadow_node,
+ RCUJA_SHADOW_CLEAR_FREE_NODE);
+ assert(!ret);
+ }
+
+ ret = 0;
+end:
+ return ret;
+
+fallback_toosmall:
+ /* fallback if next pool is too small */
+ ret = rcuja_shadow_clear(ja->ht, new_node, new_shadow_node,
+ RCUJA_SHADOW_CLEAR_FREE_NODE);
+ assert(!ret);
+
+ /* Last type: pigeon */
+ new_type_index = (1UL << JA_TYPE_BITS) - 1;
+ dbg_printf("Fallback to type %d\n", new_type_index);
+ uatomic_inc(&ja->nr_fallback);
+ fallback = 1;
+ goto retry;
+}
+
+/*
+ * Return 0 on success, -ENOENT if need to retry, or other negative
+ * error value otherwise.
+ */
+static
+int ja_node_set_nth(struct cds_ja *ja,
+ struct cds_ja_inode_flag **node_flag, uint8_t n,
+ struct cds_ja_inode_flag *child_node_flag,
+ struct cds_ja_shadow_node *shadow_node)
+{
+ int ret;
+ unsigned int type_index;
+ const struct cds_ja_type *type;
+ struct cds_ja_inode *node;
+
+ dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
+ (unsigned int) n, ja_node_ptr(*node_flag), shadow_node);
+
+ node = ja_node_ptr(*node_flag);
+ type_index = ja_node_type(*node_flag);
+ type = &ja_types[type_index];
+ ret = _ja_node_set_nth(type, node, shadow_node,
+ n, child_node_flag);
+ if (ret == -ENOSPC) {
+ /* Not enough space in node, need to recompact. */
+ ret = ja_node_recompact_add(ja, type_index, type, node,
+ shadow_node, node_flag, n, child_node_flag);
+ }
+ return ret;
+}
+
+struct cds_hlist_head *cds_ja_lookup(struct cds_ja *ja, uint64_t key)
+{
+ unsigned int tree_depth, i;
+ struct cds_ja_inode_flag *node_flag;
+
+ if (caa_unlikely(key > ja->key_max))
+ return NULL;
+ tree_depth = ja->tree_depth;
+ node_flag = rcu_dereference(ja->root);
+
+ /* level 0: root node */
+ if (!ja_node_ptr(node_flag))
+ return NULL;
+
+ for (i = 1; i < tree_depth; i++) {
+ uint8_t iter_key;
+
+ iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1)));
+ node_flag = ja_node_get_nth(node_flag, NULL,
+ iter_key);
+ if (!ja_node_ptr(node_flag))
+ return NULL;
+ }
+
+ /* Last level lookup succeded. We got an actual match. */
+ return (struct cds_hlist_head *) node_flag;
+}
+
+/*
+ * We reached an unpopulated node. Create it and the children we need,
+ * and then attach the entire branch to the current node. This may
+ * trigger recompaction of the current node. Locks needed: node lock
+ * (for add), and, possibly, parent node lock (to update pointer due to
+ * node recompaction).
+ *
+ * First take node lock, check if recompaction is needed, then take
+ * parent lock (if needed). Then we can proceed to create the new
+ * branch. Publish the new branch, and release locks.
+ * TODO: we currently always take the parent lock even when not needed.
+ */
+static
+int ja_attach_node(struct cds_ja *ja,
+ struct cds_ja_inode_flag **node_flag_ptr,
+ struct cds_ja_inode_flag *node_flag,
+ struct cds_ja_inode_flag *parent_node_flag,
+ uint64_t key,
+ unsigned int level,
+ struct cds_ja_node *child_node)
+{
+ struct cds_ja_shadow_node *shadow_node = NULL,
+ *parent_shadow_node = NULL,
+ *iter_shadow_node;
+ struct cds_ja_inode *node = ja_node_ptr(node_flag);
+ struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag);
+ struct cds_hlist_head head;
+ struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag;
+ int ret, i;
+ struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH];
+ int nr_created_nodes = 0;
+
+ dbg_printf("Attach node at level %u\n", level);
+
+ assert(node);
+ shadow_node = rcuja_shadow_lookup_lock(ja->ht, node);
+ if (!shadow_node) {
+ ret = -ENOENT;
+ goto end;
+ }
+ if (parent_node) {
+ parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht,
+ parent_node);
+ if (!parent_shadow_node) {
+ ret = -ENOENT;
+ goto unlock_shadow;
+ }
+ }
+
+ /* Create new branch, starting from bottom */
+ CDS_INIT_HLIST_HEAD(&head);
+ cds_hlist_add_head_rcu(&child_node->list, &head);
+ iter_node_flag = (struct cds_ja_inode_flag *) head.next;
+
+ /* Create shadow node for the leaf node */
+ dbg_printf("leaf shadow node creation\n");
+ iter_shadow_node = rcuja_shadow_set(ja->ht,
+ ja_node_ptr(iter_node_flag), NULL);
+ if (!iter_shadow_node) {
+ ret = -ENOMEM;
+ goto check_error;
+ }
+ created_nodes[nr_created_nodes++] = iter_node_flag;
+
+ for (i = ja->tree_depth; i > (int) level; i--) {
+ uint8_t iter_key;
+
+ iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i)));
+ dbg_printf("branch creation level %d, key %u\n",
+ i - 1, (unsigned int) iter_key);
+ iter_dest_node_flag = NULL;
+ ret = ja_node_set_nth(ja, &iter_dest_node_flag,
+ iter_key,
+ iter_node_flag,
+ NULL);
+ if (ret)
+ goto check_error;
+ created_nodes[nr_created_nodes++] = iter_dest_node_flag;
+ iter_node_flag = iter_dest_node_flag;
+ }
+
+ if (level > 1) {
+ uint8_t iter_key;
+
+ iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level)));
+ /* We need to use set_nth on the previous level. */
+ iter_dest_node_flag = node_flag;
+ ret = ja_node_set_nth(ja, &iter_dest_node_flag,
+ iter_key,
+ iter_node_flag,
+ shadow_node);
+ if (ret)
+ goto check_error;
+ created_nodes[nr_created_nodes++] = iter_dest_node_flag;
+ iter_node_flag = iter_dest_node_flag;
+ }
+
+ /* Publish new branch */
+ dbg_printf("Publish branch %p, replacing %p\n",
+ iter_node_flag, *node_flag_ptr);
+ rcu_assign_pointer(*node_flag_ptr, iter_node_flag);
+
+ /* Success */
+ ret = 0;
+
+check_error:
+ if (ret) {
+ for (i = 0; i < nr_created_nodes; i++) {
+ int tmpret;
+ int flags;
+
+ flags = RCUJA_SHADOW_CLEAR_FREE_LOCK;
+ if (i)
+ flags |= RCUJA_SHADOW_CLEAR_FREE_NODE;
+ tmpret = rcuja_shadow_clear(ja->ht,
+ ja_node_ptr(created_nodes[i]),
+ NULL,
+ flags);
+ assert(!tmpret);
+ }
+ }
+ if (parent_shadow_node)
+ rcuja_shadow_unlock(parent_shadow_node);
+unlock_shadow:
+ if (shadow_node)
+ rcuja_shadow_unlock(shadow_node);
+end:
+ return ret;
+}
+
+/*
+ * Lock the hlist head shadow node mutex, and add node to list of
+ * duplicates. Failure can happen if concurrent removal removes the last
+ * node with same key before we get the lock.
+ * Return 0 on success, negative error value on failure.