+int _ja_node_set_nth(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_inode_flag *node_flag,
+ 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, node_flag, 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;
+}
+
+static
+int ja_linear_node_clear_ptr(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_shadow_node *shadow_node,
+ struct cds_ja_inode_flag **node_flag_ptr)
+{
+ uint8_t nr_child;
+ uint8_t *nr_child_ptr;
+
+ assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL);
+
+ nr_child_ptr = &node->u.data[0];
+ nr_child = *nr_child_ptr;
+ assert(nr_child <= type->max_linear_child);
+
+ if (type->type_class == RCU_JA_LINEAR) {
+ assert(!shadow_node->fallback_removal_count);
+ if (shadow_node->nr_child <= type->min_child) {
+ /* We need to try recompacting the node */
+ return -EFBIG;
+ }
+ }
+ dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr);
+ assert(*node_flag_ptr != NULL);
+ rcu_assign_pointer(*node_flag_ptr, NULL);
+ /*
+ * Value and nr_child are never changed (would cause ABA issue).
+ * Instead, we leave the pointer to NULL and recompact the node
+ * once in a while. It is allowed to set a NULL pointer to a new
+ * value without recompaction though.
+ * Only update the shadow node accounting.
+ */
+ shadow_node->nr_child--;
+ dbg_printf("linear clear ptr: %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_clear_ptr(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_inode_flag *node_flag,
+ struct cds_ja_shadow_node *shadow_node,
+ struct cds_ja_inode_flag **node_flag_ptr,
+ uint8_t n)
+{
+ struct cds_ja_inode *linear;
+
+ assert(type->type_class == RCU_JA_POOL);
+
+ if (shadow_node->fallback_removal_count) {
+ shadow_node->fallback_removal_count--;
+ } else {
+ /* We should try recompacting the node */
+ if (shadow_node->nr_child <= type->min_child)
+ return -EFBIG;
+ }
+
+ switch (type->nr_pool_order) {
+ case 1:
+ {
+ unsigned long bitsel, index;
+
+ bitsel = ja_node_pool_1d_bitsel(node_flag);
+ assert(bitsel < CHAR_BIT);
+ index = ((unsigned long) n >> bitsel) & type->nr_pool_order;
+ linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order];
+ break;
+ }
+ case 2:
+ {
+ unsigned long bitsel[2], index[2], rindex;
+
+ ja_node_pool_2d_bitsel(node_flag, bitsel);
+ assert(bitsel[0] < CHAR_BIT);
+ assert(bitsel[1] < CHAR_BIT);
+ index[0] = ((unsigned long) n >> bitsel[0]) & 0x1;
+ index[0] <<= 1;
+ index[1] = ((unsigned long) n >> bitsel[1]) & 0x1;
+ rindex = index[0] | index[1];
+ linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order];
+ break;
+ }
+ default:
+ linear = NULL;
+ assert(0);
+ }
+
+ return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr);
+}
+
+static
+int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_shadow_node *shadow_node,
+ struct cds_ja_inode_flag **node_flag_ptr)
+{
+ assert(type->type_class == RCU_JA_PIGEON);
+
+ if (shadow_node->fallback_removal_count) {
+ shadow_node->fallback_removal_count--;
+ } else {
+ /* We should try recompacting the node */
+ if (shadow_node->nr_child <= type->min_child)
+ return -EFBIG;
+ }
+ dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr);
+ rcu_assign_pointer(*node_flag_ptr, NULL);
+ shadow_node->nr_child--;
+ return 0;
+}
+
+/*
+ * _ja_node_clear_ptr: clear ptr item within a node. Return an error
+ * (negative error value) if it is not found (-ENOENT).
+ */
+static
+int _ja_node_clear_ptr(const struct cds_ja_type *type,
+ struct cds_ja_inode *node,
+ struct cds_ja_inode_flag *node_flag,
+ struct cds_ja_shadow_node *shadow_node,
+ struct cds_ja_inode_flag **node_flag_ptr,
+ uint8_t n)
+{
+ switch (type->type_class) {
+ case RCU_JA_LINEAR:
+ return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr);
+ case RCU_JA_POOL:
+ return ja_pool_node_clear_ptr(type, node, node_flag, shadow_node, node_flag_ptr, n);
+ case RCU_JA_PIGEON:
+ return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr);
+ case RCU_JA_NULL:
+ return -ENOENT;
+ default:
+ assert(0);
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+/*
+ * Calculate bit distribution. Returns the bit (0 to 7) that splits the
+ * distribution in two sub-distributions containing as much elements one
+ * compared to the other.
+ */
+static
+unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode,
+ struct cds_ja *ja,
+ unsigned int type_index,
+ 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 **nullify_node_flag_ptr)
+{
+ uint8_t nr_one[JA_BITS_PER_BYTE];
+ unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX;
+ unsigned int distrib_nr_child = 0;
+
+ memset(nr_one, 0, sizeof(nr_one));
+
+ switch (type->type_class) {
+ case RCU_JA_LINEAR:
+ {
+ uint8_t nr_child =
+ ja_linear_node_get_nr_child(type, 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(type, node, i, &v, &iter);
+ if (!iter)
+ continue;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ if (v & (1U << bit_i))
+ nr_one[bit_i]++;
+ }
+ distrib_nr_child++;
+ }
+ break;
+ }
+ case RCU_JA_POOL:
+ {
+ unsigned int pool_nr;
+
+ for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) {
+ struct cds_ja_inode *pool =
+ ja_pool_node_get_ith_pool(type,
+ node, pool_nr);
+ uint8_t nr_child =
+ ja_linear_node_get_nr_child(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(type, pool,
+ j, &v, &iter);
+ if (!iter)
+ continue;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ if (v & (1U << bit_i))
+ nr_one[bit_i]++;
+ }
+ distrib_nr_child++;
+ }
+ }
+ break;
+ }
+ case RCU_JA_PIGEON:
+ {
+ unsigned int i;
+
+ assert(mode == JA_RECOMPACT_DEL);
+ for (i = 0; i < JA_ENTRY_PER_NODE; i++) {
+ struct cds_ja_inode_flag *iter;
+
+ iter = ja_pigeon_node_get_ith_pos(type, node, i);
+ if (!iter)
+ continue;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ if (i & (1U << bit_i))
+ nr_one[bit_i]++;
+ }
+ distrib_nr_child++;
+ }
+ break;
+ }
+ case RCU_JA_NULL:
+ assert(mode == JA_RECOMPACT_ADD_NEXT);
+ break;
+ default:
+ assert(0);
+ break;
+ }
+
+ if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) {
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ if (n & (1U << bit_i))
+ nr_one[bit_i]++;
+ }
+ distrib_nr_child++;
+ }
+
+ /*
+ * The best bit selector is that for which the number of ones is
+ * closest to half of the number of children in the
+ * distribution. We calculate the distance using the double of
+ * the sub-distribution sizes to eliminate truncation error.
+ */
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ unsigned int distance_to_best;
+
+ distance_to_best = abs_int((nr_one[bit_i] << 1U) - distrib_nr_child);
+ if (distance_to_best < overall_best_distance) {
+ overall_best_distance = distance_to_best;
+ bitsel = bit_i;
+ }
+ }
+ dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel);
+ return bitsel;
+}
+
+/*
+ * Calculate bit distribution in two dimensions. Returns the two bits
+ * (each 0 to 7) that splits the distribution in four sub-distributions
+ * containing as much elements one compared to the other.
+ */
+static
+void ja_node_sum_distribution_2d(enum ja_recompact mode,
+ struct cds_ja *ja,
+ unsigned int type_index,
+ 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 **nullify_node_flag_ptr,
+ unsigned int *_bitsel)
+{
+ uint8_t nr_2d_11[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE],
+ nr_2d_10[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE],
+ nr_2d_01[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE],
+ nr_2d_00[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE];
+ unsigned int bitsel[2] = { 0, 1 };
+ unsigned int bit_i, bit_j;
+ int overall_best_distance = INT_MAX;
+ unsigned int distrib_nr_child = 0;
+
+ memset(nr_2d_11, 0, sizeof(nr_2d_11));
+ memset(nr_2d_10, 0, sizeof(nr_2d_10));
+ memset(nr_2d_01, 0, sizeof(nr_2d_01));
+ memset(nr_2d_00, 0, sizeof(nr_2d_00));
+
+ switch (type->type_class) {
+ case RCU_JA_LINEAR:
+ {
+ uint8_t nr_child =
+ ja_linear_node_get_nr_child(type, 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(type, node, i, &v, &iter);
+ if (!iter)
+ continue;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ for (bit_j = 0; bit_j < bit_i; bit_j++) {
+ if ((v & (1U << bit_i)) && (v & (1U << bit_j))) {
+ nr_2d_11[bit_i][bit_j]++;
+ }
+ if ((v & (1U << bit_i)) && !(v & (1U << bit_j))) {
+ nr_2d_10[bit_i][bit_j]++;
+ }
+ if (!(v & (1U << bit_i)) && (v & (1U << bit_j))) {
+ nr_2d_01[bit_i][bit_j]++;
+ }
+ if (!(v & (1U << bit_i)) && !(v & (1U << bit_j))) {
+ nr_2d_00[bit_i][bit_j]++;
+ }
+ }
+ }
+ distrib_nr_child++;
+ }
+ break;
+ }
+ case RCU_JA_POOL:
+ {
+ unsigned int pool_nr;
+
+ for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) {
+ struct cds_ja_inode *pool =
+ ja_pool_node_get_ith_pool(type,
+ node, pool_nr);
+ uint8_t nr_child =
+ ja_linear_node_get_nr_child(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(type, pool,
+ j, &v, &iter);
+ if (!iter)
+ continue;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ for (bit_j = 0; bit_j < bit_i; bit_j++) {
+ if ((v & (1U << bit_i)) && (v & (1U << bit_j))) {
+ nr_2d_11[bit_i][bit_j]++;
+ }
+ if ((v & (1U << bit_i)) && !(v & (1U << bit_j))) {
+ nr_2d_10[bit_i][bit_j]++;
+ }
+ if (!(v & (1U << bit_i)) && (v & (1U << bit_j))) {
+ nr_2d_01[bit_i][bit_j]++;
+ }
+ if (!(v & (1U << bit_i)) && !(v & (1U << bit_j))) {
+ nr_2d_00[bit_i][bit_j]++;
+ }
+ }
+ }
+ distrib_nr_child++;
+ }
+ }
+ break;
+ }
+ case RCU_JA_PIGEON:
+ {
+ unsigned int i;
+
+ assert(mode == JA_RECOMPACT_DEL);
+ for (i = 0; i < JA_ENTRY_PER_NODE; i++) {
+ struct cds_ja_inode_flag *iter;
+
+ iter = ja_pigeon_node_get_ith_pos(type, node, i);
+ if (!iter)
+ continue;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ for (bit_j = 0; bit_j < bit_i; bit_j++) {
+ if ((i & (1U << bit_i)) && (i & (1U << bit_j))) {
+ nr_2d_11[bit_i][bit_j]++;
+ }
+ if ((i & (1U << bit_i)) && !(i & (1U << bit_j))) {
+ nr_2d_10[bit_i][bit_j]++;
+ }
+ if (!(i & (1U << bit_i)) && (i & (1U << bit_j))) {
+ nr_2d_01[bit_i][bit_j]++;
+ }
+ if (!(i & (1U << bit_i)) && !(i & (1U << bit_j))) {
+ nr_2d_00[bit_i][bit_j]++;
+ }
+ }
+ }
+ distrib_nr_child++;
+ }
+ break;
+ }
+ case RCU_JA_NULL:
+ assert(mode == JA_RECOMPACT_ADD_NEXT);
+ break;
+ default:
+ assert(0);
+ break;
+ }
+
+ if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) {
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ for (bit_j = 0; bit_j < bit_i; bit_j++) {
+ if ((n & (1U << bit_i)) && (n & (1U << bit_j))) {
+ nr_2d_11[bit_i][bit_j]++;
+ }
+ if ((n & (1U << bit_i)) && !(n & (1U << bit_j))) {
+ nr_2d_10[bit_i][bit_j]++;
+ }
+ if (!(n & (1U << bit_i)) && (n & (1U << bit_j))) {
+ nr_2d_01[bit_i][bit_j]++;
+ }
+ if (!(n & (1U << bit_i)) && !(n & (1U << bit_j))) {
+ nr_2d_00[bit_i][bit_j]++;
+ }
+ }
+ }
+ distrib_nr_child++;
+ }
+
+ /*
+ * The best bit selector is that for which the number of nodes
+ * in each sub-class is closest to one-fourth of the number of
+ * children in the distribution. We calculate the distance using
+ * 4 times the size of the sub-distribution to eliminate
+ * truncation error.
+ */
+ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) {
+ for (bit_j = 0; bit_j < bit_i; bit_j++) {
+ int distance_to_best[4];
+
+ distance_to_best[0] = (nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child;
+ distance_to_best[1] = (nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child;
+ distance_to_best[2] = (nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child;
+ distance_to_best[3] = (nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child;
+
+ /* Consider worse distance above best */
+ if (distance_to_best[1] > 0 && distance_to_best[1] > distance_to_best[0])
+ distance_to_best[0] = distance_to_best[1];
+ if (distance_to_best[2] > 0 && distance_to_best[2] > distance_to_best[0])
+ distance_to_best[0] = distance_to_best[2];
+ if (distance_to_best[3] > 0 && distance_to_best[3] > distance_to_best[0])
+ distance_to_best[0] = distance_to_best[3];
+
+ /*
+ * If our worse distance is better than overall,
+ * we become new best candidate.
+ */
+ if (distance_to_best[0] < overall_best_distance) {
+ overall_best_distance = distance_to_best[0];
+ bitsel[0] = bit_i;
+ bitsel[1] = bit_j;
+ }
+ }
+ }
+
+ dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel[0], bitsel[1]);
+
+ /* Return our bit selection */
+ _bitsel[0] = bitsel[0];
+ _bitsel[1] = bitsel[1];
+}
+
+static
+unsigned int find_nearest_type_index(unsigned int type_index,
+ unsigned int nr_nodes)
+{
+ const struct cds_ja_type *type;
+
+ assert(type_index != NODE_INDEX_NULL);
+ if (nr_nodes == 0)
+ return NODE_INDEX_NULL;
+ for (;;) {
+ type = &ja_types[type_index];
+ if (nr_nodes < type->min_child)
+ type_index--;
+ else if (nr_nodes > type->max_child)
+ type_index++;
+ else
+ break;
+ }
+ return type_index;
+}
+
+/*
+ * ja_node_recompact_add: recompact a node, adding a new child.
+ * Return 0 on success, -EAGAIN if need to retry, or other negative
+ * error value otherwise.
+ */
+static
+int ja_node_recompact(enum ja_recompact mode,
+ 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_ptr, uint8_t n,
+ struct cds_ja_inode_flag *child_node_flag,
+ struct cds_ja_inode_flag **nullify_node_flag_ptr,
+ int level)
+{
+ unsigned int new_type_index;
+ struct cds_ja_inode *new_node;
+ struct cds_ja_shadow_node *new_shadow_node = NULL;
+ const struct cds_ja_type *new_type;
+ struct cds_ja_inode_flag *new_node_flag, *old_node_flag;
+ int ret;
+ int fallback = 0;
+
+ old_node_flag = *old_node_flag_ptr;
+
+ /*
+ * Need to find nearest type index even for ADD_SAME, because
+ * this recompaction, when applied to linear nodes, will garbage
+ * collect dummy (NULL) entries, and can therefore cause a few
+ * linear representations to be skipped.
+ */
+ switch (mode) {
+ case JA_RECOMPACT_ADD_SAME:
+ new_type_index = find_nearest_type_index(old_type_index,
+ shadow_node->nr_child + 1);
+ dbg_printf("Recompact for node with %u children\n",
+ shadow_node->nr_child + 1);
+ break;
+ case JA_RECOMPACT_ADD_NEXT:
+ if (!shadow_node || old_type_index == NODE_INDEX_NULL) {
+ new_type_index = 0;
+ dbg_printf("Recompact for NULL\n");
+ } else {
+ new_type_index = find_nearest_type_index(old_type_index,
+ shadow_node->nr_child + 1);
+ dbg_printf("Recompact for node with %u children\n",
+ shadow_node->nr_child + 1);
+ }
+ break;
+ case JA_RECOMPACT_DEL:
+ new_type_index = find_nearest_type_index(old_type_index,
+ shadow_node->nr_child - 1);
+ dbg_printf("Recompact for node with %u children\n",
+ shadow_node->nr_child - 1);
+ break;
+ default:
+ assert(0);
+ }
+
+retry: /* for fallback */
+ dbg_printf("Recompact from type %d to type %d\n",
+ old_type_index, new_type_index);
+ new_type = &ja_types[new_type_index];
+ if (new_type_index != NODE_INDEX_NULL) {
+ new_node = alloc_cds_ja_node(new_type);
+ if (!new_node)
+ return -ENOMEM;
+
+ if (new_type->type_class == RCU_JA_POOL) {
+ switch (new_type->nr_pool_order) {
+ case 1:
+ {
+ unsigned int node_distrib_bitsel;
+
+ node_distrib_bitsel =
+ ja_node_sum_distribution_1d(mode, ja,
+ old_type_index, old_type,
+ old_node, shadow_node,
+ n, child_node_flag,
+ nullify_node_flag_ptr);
+ assert(!((unsigned long) new_node & JA_POOL_1D_MASK));
+ new_node_flag = ja_node_flag_pool_1d(new_node,
+ new_type_index, node_distrib_bitsel);
+ break;
+ }
+ case 2:
+ {
+ unsigned int node_distrib_bitsel[2];
+
+ ja_node_sum_distribution_2d(mode, ja,
+ old_type_index, old_type,
+ old_node, shadow_node,
+ n, child_node_flag,
+ nullify_node_flag_ptr,
+ node_distrib_bitsel);
+ assert(!((unsigned long) new_node & JA_POOL_1D_MASK));
+ assert(!((unsigned long) new_node & JA_POOL_2D_MASK));
+ new_node_flag = ja_node_flag_pool_2d(new_node,
+ new_type_index, node_distrib_bitsel);
+ break;
+ }
+ default:
+ assert(0);
+ }
+ } else {
+ 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_flag, shadow_node, ja, level);
+ if (!new_shadow_node) {
+ free_cds_ja_node(new_node);
+ return -ENOMEM;
+ }
+ if (fallback)
+ new_shadow_node->fallback_removal_count =
+ JA_FALLBACK_REMOVAL_COUNT;
+ } else {
+ new_node = NULL;
+ new_node_flag = NULL;
+ }
+
+ assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON);
+
+ if (new_type_index == NODE_INDEX_NULL)
+ goto skip_copy;
+
+ 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;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ ret = _ja_node_set_nth(new_type, new_node, new_node_flag,
+ 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;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ ret = _ja_node_set_nth(new_type, new_node, new_node_flag,
+ new_shadow_node,
+ v, iter);
+ if (new_type->type_class == RCU_JA_POOL
+ && ret) {
+ goto fallback_toosmall;
+ }
+ assert(!ret);
+ }
+ }
+ break;
+ }
+ case RCU_JA_NULL:
+ assert(mode == JA_RECOMPACT_ADD_NEXT);
+ break;
+ case RCU_JA_PIGEON:
+ {
+ unsigned int i;
+
+ assert(mode == JA_RECOMPACT_DEL);
+ for (i = 0; i < JA_ENTRY_PER_NODE; i++) {
+ struct cds_ja_inode_flag *iter;
+
+ iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i);
+ if (!iter)
+ continue;
+ if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
+ continue;
+ ret = _ja_node_set_nth(new_type, new_node, new_node_flag,
+ new_shadow_node,
+ i, iter);
+ if (new_type->type_class == RCU_JA_POOL && ret) {
+ goto fallback_toosmall;
+ }
+ assert(!ret);
+ }
+ break;
+ }
+ default:
+ assert(0);
+ ret = -EINVAL;
+ goto end;
+ }
+skip_copy:
+
+ if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) {
+ /* add node */
+ ret = _ja_node_set_nth(new_type, new_node, new_node_flag,
+ new_shadow_node,
+ n, child_node_flag);
+ if (new_type->type_class == RCU_JA_POOL && ret) {
+ goto fallback_toosmall;
+ }
+ assert(!ret);
+ }
+
+ if (fallback) {
+ dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n",
+ new_shadow_node->nr_child, old_type_index, mode == JA_RECOMPACT_ADD_NEXT ? "add_next" :
+ (mode == JA_RECOMPACT_DEL ? "del" : "add_same"));
+ uatomic_inc(&node_fallback_count_distribution[new_shadow_node->nr_child]);
+ }
+
+ /* Return pointer to new recompacted node through old_node_flag_ptr */
+ *old_node_flag_ptr = new_node_flag;
+ if (old_node) {
+ int flags;
+
+ flags = RCUJA_SHADOW_CLEAR_FREE_NODE;
+ /*
+ * It is OK to free the lock associated with a node
+ * going to NULL, since we are holding the parent lock.
+ * This synchronizes removal with re-add of that node.
+ */
+ if (new_type_index == NODE_INDEX_NULL)
+ flags |= RCUJA_SHADOW_CLEAR_FREE_LOCK;
+ ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node,
+ flags);
+ assert(!ret);
+ }
+
+ ret = 0;
+end:
+ return ret;
+
+fallback_toosmall:
+ /* fallback if next pool is too small */
+ assert(new_shadow_node);
+ ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node,
+ RCUJA_SHADOW_CLEAR_FREE_NODE);
+ assert(!ret);
+
+ switch (mode) {
+ case JA_RECOMPACT_ADD_SAME:
+ /*
+ * JA_RECOMPACT_ADD_SAME is only triggered if a linear
+ * node within a pool has unused entries. It should
+ * therefore _never_ be too small.
+ */
+ assert(0);
+
+ /* Fall-through */
+ case JA_RECOMPACT_ADD_NEXT:
+ {
+ const struct cds_ja_type *next_type;
+
+ /*
+ * Recompaction attempt on add failed. Should only
+ * happen if target node type is pool. Caused by
+ * hard-to-split distribution. Recompact using the next
+ * distribution size.
+ */
+ assert(new_type->type_class == RCU_JA_POOL);
+ next_type = &ja_types[new_type_index + 1];
+ /*
+ * Try going to the next pool size if our population
+ * fits within its range. This is not flagged as a
+ * fallback.
+ */
+ if (shadow_node->nr_child + 1 >= next_type->min_child
+ && shadow_node->nr_child + 1 <= next_type->max_child) {
+ new_type_index++;
+ goto retry;
+ } else {
+ new_type_index++;
+ dbg_printf("Add fallback to type %d\n", new_type_index);
+ uatomic_inc(&ja->nr_fallback);
+ fallback = 1;
+ goto retry;
+ }
+ break;
+ }
+ case JA_RECOMPACT_DEL:
+ /*
+ * Recompaction attempt on delete failed. Should only
+ * happen if target node type is pool. This is caused by
+ * a hard-to-split distribution. Recompact on same node
+ * size, but flag current node as "fallback" to ensure
+ * we don't attempt recompaction before some activity
+ * has reshuffled our node.
+ */
+ assert(new_type->type_class == RCU_JA_POOL);
+ new_type_index = old_type_index;
+ dbg_printf("Delete fallback keeping type %d\n", new_type_index);
+ uatomic_inc(&ja->nr_fallback);
+ fallback = 1;
+ goto retry;
+ default:
+ assert(0);
+ return -EINVAL;
+ }
+
+ /*
+ * Last resort fallback: 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, -EAGAIN 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 level)
+{
+ 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, *node_flag, shadow_node,
+ n, child_node_flag);
+ switch (ret) {
+ case -ENOSPC:
+ /* Not enough space in node, need to recompact to next type. */
+ ret = ja_node_recompact(JA_RECOMPACT_ADD_NEXT, ja, type_index, type, node,
+ shadow_node, node_flag, n, child_node_flag, NULL, level);
+ break;
+ case -ERANGE:
+ /* Node needs to be recompacted. */
+ ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node,
+ shadow_node, node_flag, n, child_node_flag, NULL, level);
+ break;
+ }
+ return ret;
+}
+
+/*
+ * Return 0 on success, -EAGAIN if need to retry, or other negative
+ * error value otherwise.
+ */
+static
+int ja_node_clear_ptr(struct cds_ja *ja,
+ struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */
+ struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */
+ struct cds_ja_shadow_node *shadow_node, /* of parent */
+ uint8_t n, int level)
+{
+ int ret;
+ unsigned int type_index;
+ const struct cds_ja_type *type;
+ struct cds_ja_inode *node;
+
+ dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n",
+ ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr);
+
+ node = ja_node_ptr(*parent_node_flag_ptr);
+ type_index = ja_node_type(*parent_node_flag_ptr);
+ type = &ja_types[type_index];
+ ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n);
+ if (ret == -EFBIG) {
+ /* Should try recompaction. */
+ ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node,
+ shadow_node, parent_node_flag_ptr, n, NULL,
+ node_flag_ptr, level);
+ }
+ 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;
+ struct cds_hlist_head head = { NULL };
+
+ if (caa_unlikely(key > ja->key_max))
+ return head;
+ tree_depth = ja->tree_depth;
+ node_flag = rcu_dereference(ja->root);
+
+ /* level 0: root node */
+ if (!ja_node_ptr(node_flag))
+ return head;
+
+ 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);
+ dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
+ (unsigned int) iter_key, node_flag);
+ if (!ja_node_ptr(node_flag))
+ return head;
+ }
+
+ /* Last level lookup succeded. We got an actual match. */
+ head.next = (struct cds_hlist_node *) node_flag;
+ return head;
+}
+
+/*
+ * 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 **attach_node_flag_ptr,
+ struct cds_ja_inode_flag *attach_node_flag,
+ struct cds_ja_inode_flag *parent_attach_node_flag,
+ struct cds_ja_inode_flag **old_node_flag_ptr,
+ struct cds_ja_inode_flag *old_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;
+ 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 (old_node_flag %p, attach_node_flag_ptr %p attach_node_flag %p, parent_attach_node_flag %p)\n",
+ level, old_node_flag, attach_node_flag_ptr, attach_node_flag, parent_attach_node_flag);
+
+ assert(!old_node_flag);
+ if (attach_node_flag) {
+ shadow_node = rcuja_shadow_lookup_lock(ja->ht, attach_node_flag);
+ if (!shadow_node) {
+ ret = -EAGAIN;
+ goto end;
+ }
+ }
+ if (parent_attach_node_flag) {
+ parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht,
+ parent_attach_node_flag);
+ if (!parent_shadow_node) {
+ ret = -EAGAIN;
+ goto unlock_shadow;
+ }
+ }
+
+ if (old_node_flag_ptr && ja_node_ptr(*old_node_flag_ptr)) {
+ /*
+ * Target node has been updated between RCU lookup and
+ * lock acquisition. We need to re-try lookup and
+ * attach.
+ */
+ ret = -EAGAIN;
+ goto unlock_parent;
+ }
+
+ if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) !=
+ ja_node_ptr(attach_node_flag)) {
+ /*
+ * Target node has been updated between RCU lookup and
+ * lock acquisition. We need to re-try lookup and
+ * attach.
+ */
+ ret = -EAGAIN;
+ goto unlock_parent;
+ }
+
+ /* 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;
+
+ for (i = ja->tree_depth - 1; i >= (int) level; i--) {
+ uint8_t iter_key;
+
+ iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i - 1)));
+ dbg_printf("branch creation level %d, key %u\n",
+ i, (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, i);
+ if (ret)
+ goto check_error;
+ created_nodes[nr_created_nodes++] = iter_dest_node_flag;
+ iter_node_flag = iter_dest_node_flag;
+ }
+ assert(level > 0);
+
+ /* Publish branch */
+ if (level == 1) {
+ /*
+ * Attaching to root node.
+ */
+ rcu_assign_pointer(ja->root, iter_node_flag);
+ } else {
+ uint8_t iter_key;
+
+ iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level)));
+ dbg_printf("publish branch at level %d, key %u\n",
+ level - 1, (unsigned int) iter_key);
+ /* We need to use set_nth on the previous level. */
+ iter_dest_node_flag = attach_node_flag;
+ ret = ja_node_set_nth(ja, &iter_dest_node_flag,
+ iter_key,
+ iter_node_flag,
+ shadow_node, level - 1);
+ if (ret)
+ goto check_error;
+ /*
+ * Attach branch
+ */
+ rcu_assign_pointer(*attach_node_flag_ptr, iter_dest_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,
+ created_nodes[i],
+ NULL,
+ flags);
+ assert(!tmpret);
+ }
+ }
+unlock_parent:
+ 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 parent containing the hlist head pointer, and add node to list of
+ * duplicates. Failure can happen if concurrent update changes the
+ * parent before we get the lock. We return -EAGAIN in that case.
+ * Return 0 on success, negative error value on failure.
+ */
+static
+int ja_chain_node(struct cds_ja *ja,
+ struct cds_ja_inode_flag *parent_node_flag,
+ struct cds_ja_inode_flag **node_flag_ptr,
+ struct cds_ja_inode_flag *node_flag,
+ struct cds_ja_node *node)
+{
+ struct cds_ja_shadow_node *shadow_node;
+ int ret = 0;
+
+ shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag);
+ if (!shadow_node) {
+ return -EAGAIN;
+ }
+ if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) {
+ ret = -EAGAIN;
+ goto end;
+ }
+ cds_hlist_add_head_rcu(&node->list, (struct cds_hlist_head *) node_flag_ptr);
+end:
+ rcuja_shadow_unlock(shadow_node);
+ return ret;
+}
+
+static
+int _cds_ja_add(struct cds_ja *ja, uint64_t key,
+ struct cds_ja_node *new_node,
+ struct cds_ja_node **unique_node_ret)
+{
+ unsigned int tree_depth, i;
+ struct cds_ja_inode_flag *attach_node_flag,
+ *parent_node_flag,
+ *parent2_node_flag,
+ *node_flag,
+ *parent_attach_node_flag;
+ struct cds_ja_inode_flag **attach_node_flag_ptr,
+ **parent_node_flag_ptr,
+ **node_flag_ptr;
+ int ret;
+
+ if (caa_unlikely(key > ja->key_max)) {
+ return -EINVAL;
+ }
+ tree_depth = ja->tree_depth;
+
+retry:
+ dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n",
+ key, new_node);
+ parent2_node_flag = NULL;
+ parent_node_flag =
+ (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */
+ parent_node_flag_ptr = NULL;
+ node_flag = rcu_dereference(ja->root);
+ node_flag_ptr = &ja->root;
+
+ /* Iterate on all internal levels */
+ for (i = 1; i < tree_depth; i++) {
+ uint8_t iter_key;
+
+ if (!ja_node_ptr(node_flag))
+ break;
+ dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
+ parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag);
+ iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1)));
+ parent2_node_flag = parent_node_flag;
+ parent_node_flag = node_flag;
+ parent_node_flag_ptr = node_flag_ptr;
+ node_flag = ja_node_get_nth(node_flag,
+ &node_flag_ptr,
+ iter_key);
+ }
+
+ /*
+ * We reached either bottom of tree or internal NULL node,
+ * simply add node to last internal level, or chain it if key is
+ * already present.
+ */
+ if (!ja_node_ptr(node_flag)) {
+ dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
+ parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag);
+
+ attach_node_flag = parent_node_flag;
+ attach_node_flag_ptr = parent_node_flag_ptr;
+ parent_attach_node_flag = parent2_node_flag;
+
+ ret = ja_attach_node(ja, attach_node_flag_ptr,
+ attach_node_flag,
+ parent_attach_node_flag,
+ node_flag_ptr,
+ node_flag,
+ key, i, new_node);
+ } else {
+ if (unique_node_ret) {
+ *unique_node_ret = (struct cds_ja_node *) ja_node_ptr(node_flag);
+ return -EEXIST;
+ }
+
+ dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n",
+ parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag);
+
+ attach_node_flag = node_flag;
+ attach_node_flag_ptr = node_flag_ptr;
+ parent_attach_node_flag = parent_node_flag;
+
+ ret = ja_chain_node(ja,
+ parent_attach_node_flag,
+ attach_node_flag_ptr,
+ attach_node_flag,
+ new_node);
+ }
+ if (ret == -EAGAIN || ret == -EEXIST)
+ goto retry;
+
+ return ret;
+}
+
+int cds_ja_add(struct cds_ja *ja, uint64_t key,
+ struct cds_ja_node *new_node)
+{
+ return _cds_ja_add(ja, key, new_node, NULL);
+}
+
+struct cds_ja_node *cds_ja_add_unique(struct cds_ja *ja, uint64_t key,
+ struct cds_ja_node *new_node)
+{
+ int ret;
+ struct cds_ja_node *ret_node;
+
+ ret = _cds_ja_add(ja, key, new_node, &ret_node);
+ if (ret == -EEXIST)
+ return ret_node;
+ else
+ return new_node;
+}
+
+/*
+ * 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.
+ * However, when a child is removed from "linear" nodes, its pointer
+ * is set to NULL. We therefore check, while holding the locks, if this
+ * pointer is NULL, and return -ENOENT to the caller if it is the case.
+ */
+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)