+ 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 };