+/*
+ * Note: there is no need to lookup the pointer address associated with
+ * each node's nth item after taking the lock: it's already been done by
+ * cds_ja_del while holding the rcu read-side lock, and our node rules
+ * ensure that when a match value -> pointer is found in a node, it is
+ * _NEVER_ changed for that node without recompaction, and recompaction
+ * reallocates the node.
+ */
+static
+int ja_detach_node(struct cds_ja *ja,
+ struct cds_ja_inode_flag **snapshot,
+ struct cds_ja_inode_flag ***snapshot_ptr,
+ uint8_t *snapshot_n,
+ int nr_snapshot,
+ uint64_t key,
+ struct cds_ja_node *node)
+{
+ struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH];
+ struct cds_ja_inode_flag **node_flag_ptr = NULL,
+ *parent_node_flag = NULL,
+ **parent_node_flag_ptr = NULL;
+ struct cds_ja_inode_flag *iter_node_flag;
+ int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0;
+ uint8_t n = 0;
+
+ assert(nr_snapshot == ja->tree_depth + 1);
+
+ /*
+ * From the last internal level node going up, get the node
+ * lock, check if the node has only one child left. If it is the
+ * case, we continue iterating upward. When we reach a node
+ * which has more that one child left, we lock the parent, and
+ * proceed to the node deletion (removing its children too).
+ */
+ for (i = nr_snapshot - 2; i >= 1; i--) {
+ struct cds_ja_shadow_node *shadow_node;
+
+ shadow_node = rcuja_shadow_lookup_lock(ja->ht,
+ snapshot[i]);
+ if (!shadow_node) {
+ ret = -EAGAIN;
+ goto end;
+ }
+ assert(shadow_node->nr_child > 0);
+ shadow_nodes[nr_shadow++] = shadow_node;
+ if (shadow_node->nr_child == 1)
+ nr_clear++;
+ nr_branch++;
+ if (shadow_node->nr_child > 1 || i == 1) {
+ /* Lock parent and break */
+ shadow_node = rcuja_shadow_lookup_lock(ja->ht,
+ snapshot[i - 1]);
+ if (!shadow_node) {
+ ret = -EAGAIN;
+ goto end;
+ }
+ shadow_nodes[nr_shadow++] = shadow_node;
+ node_flag_ptr = snapshot_ptr[i + 1];
+ n = snapshot_n[i + 1];
+ parent_node_flag_ptr = snapshot_ptr[i];
+ parent_node_flag = snapshot[i];
+ if (i > 1) {
+ /*
+ * Lock parent's parent, in case we need
+ * to recompact parent.
+ */
+ shadow_node = rcuja_shadow_lookup_lock(ja->ht,
+ snapshot[i - 2]);
+ if (!shadow_node) {
+ ret = -EAGAIN;
+ goto end;
+ }
+ shadow_nodes[nr_shadow++] = shadow_node;
+ }
+ break;
+ }
+ }
+
+ /*
+ * At this point, we want to delete all nodes that are about to
+ * be removed from shadow_nodes (except the last one, which is
+ * either the root or the parent of the upmost node with 1
+ * child). OK to as to free lock here, because RCU read lock is
+ * held, and free only performed in call_rcu.
+ */
+
+ for (i = 0; i < nr_clear; i++) {
+ ret = rcuja_shadow_clear(ja->ht,
+ shadow_nodes[i]->node_flag,
+ shadow_nodes[i],
+ RCUJA_SHADOW_CLEAR_FREE_NODE
+ | RCUJA_SHADOW_CLEAR_FREE_LOCK);
+ assert(!ret);
+ }
+
+ iter_node_flag = parent_node_flag;
+ /* Remove from parent */
+ ret = ja_node_clear_ptr(ja,
+ node_flag_ptr, /* Pointer to location to nullify */
+ &iter_node_flag, /* Old new parent ptr in its parent */
+ shadow_nodes[nr_branch - 1], /* of parent */
+ n);
+
+ dbg_printf("ja_detach_node: publish %p instead of %p\n",
+ iter_node_flag, *parent_node_flag_ptr);
+ /* Update address of parent ptr in its parent */
+ rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag);
+
+end:
+ for (i = 0; i < nr_shadow; i++)
+ rcuja_shadow_unlock(shadow_nodes[i]);
+ return ret;
+}
+
+static
+int ja_unchain_node(struct cds_ja *ja,
+ struct cds_ja_inode_flag *parent_node_flag,
+ struct cds_hlist_head *head,
+ struct cds_ja_node *node)
+{
+ struct cds_ja_shadow_node *shadow_node;
+ struct cds_hlist_node *hlist_node;
+ int ret = 0, count = 0;
+
+ shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag);
+ if (!shadow_node)
+ return -EAGAIN;
+ /*
+ * Retry if another thread removed all but one of duplicates
+ * since check (that was performed without lock).
+ */
+ cds_hlist_for_each_rcu(hlist_node, head, list) {
+ count++;
+ }
+
+ if (count == 1) {
+ ret = -EAGAIN;
+ goto end;
+ }
+ cds_hlist_del_rcu(&node->list);
+end:
+ rcuja_shadow_unlock(shadow_node);
+ return ret;
+}
+
+/*
+ * Called with RCU read lock held.
+ */
+int cds_ja_del(struct cds_ja *ja, uint64_t key,
+ struct cds_ja_node *node)
+{
+ unsigned int tree_depth, i;
+ struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH];
+ struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH];
+ uint8_t snapshot_n[JA_MAX_DEPTH];
+ struct cds_ja_inode_flag *node_flag;
+ struct cds_ja_inode_flag **prev_node_flag_ptr;
+ int nr_snapshot;
+ int ret;
+
+ if (caa_unlikely(key > ja->key_max))
+ return -EINVAL;
+ tree_depth = ja->tree_depth;
+
+retry:
+ nr_snapshot = 0;
+ dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n",
+ key, node);
+
+ /* snapshot for level 0 is only for shadow node lookup */
+ snapshot_n[0] = 0;
+ snapshot_n[1] = 0;
+ snapshot_ptr[nr_snapshot] = NULL;
+ snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root;
+ node_flag = rcu_dereference(ja->root);
+ prev_node_flag_ptr = &ja->root;
+
+ /* Iterate on all internal levels */
+ for (i = 1; i < tree_depth; i++) {
+ uint8_t iter_key;
+
+ dbg_printf("cds_ja_del iter node_flag %p\n",
+ node_flag);
+ if (!ja_node_ptr(node_flag)) {
+ return -ENOENT;
+ }
+ iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1)));
+ snapshot_n[nr_snapshot + 1] = iter_key;
+ snapshot_ptr[nr_snapshot] = prev_node_flag_ptr;
+ snapshot[nr_snapshot++] = node_flag;
+ node_flag = ja_node_get_nth(node_flag,
+ &prev_node_flag_ptr,
+ iter_key);
+ dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n",
+ (unsigned int) iter_key, node_flag,
+ prev_node_flag_ptr);
+ }
+
+ /*
+ * We reached bottom of tree, try to find the node we are trying
+ * to remove. Fail if we cannot find it.
+ */
+ if (!ja_node_ptr(node_flag)) {
+ dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n",
+ key);
+ return -ENOENT;
+ } else {
+ struct cds_hlist_head hlist_head;
+ struct cds_hlist_node *hlist_node;
+ struct cds_ja_node *entry, *match = NULL;
+ int count = 0;
+
+ hlist_head.next =
+ (struct cds_hlist_node *) ja_node_ptr(node_flag);
+ cds_hlist_for_each_entry_rcu(entry,
+ hlist_node,
+ &hlist_head,
+ list) {
+ dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry);
+ if (entry == node)
+ match = entry;
+ count++;
+ }
+ if (!match) {
+ dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key);
+ return -ENOENT;
+ }
+ assert(count > 0);
+ if (count == 1) {
+ /*
+ * Removing last of duplicates. Last snapshot
+ * does not have a shadow node (external leafs).
+ */
+ snapshot_ptr[nr_snapshot] = prev_node_flag_ptr;
+ snapshot[nr_snapshot++] = node_flag;
+ ret = ja_detach_node(ja, snapshot, snapshot_ptr,
+ snapshot_n, nr_snapshot, key, node);
+ } else {
+ ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1],
+ &hlist_head, match);
+ }
+ }
+ if (ret == -EAGAIN)
+ goto retry;
+ return ret;
+}
+