X-Git-Url: http://git.lttng.org./?a=blobdiff_plain;f=rcuja%2Frcuja.c;h=ff018e682cf4b6265bcfb35e8b1898bcc276acd2;hb=f07b240f5d4b2c086abedcdd4f6516753fd934e4;hp=45ce8f70a816e4cca4188ab5d9c9a54bceb02655;hpb=e1db2db589f05c48ae9e8b0199989d76bb2f7b14;p=userspace-rcu.git

diff --git a/rcuja/rcuja.c b/rcuja/rcuja.c
index 45ce8f7..ff018e6 100644
--- a/rcuja/rcuja.c
+++ b/rcuja/rcuja.c
@@ -29,11 +29,13 @@
 #include <urcu/arch.h>
 #include <assert.h>
 #include <urcu-pointer.h>
+#include <urcu/uatomic.h>
+#include <stdint.h>
 
 #include "rcuja-internal.h"
 #include "bitfield.h"
 
-enum rcu_ja_type_class {
+enum cds_ja_type_class {
 	RCU_JA_LINEAR = 0,	/* Type A */
 			/* 32-bit: 1 to 25 children, 8 to 128 bytes */
 			/* 64-bit: 1 to 28 children, 16 to 256 bytes */
@@ -49,8 +51,8 @@ enum rcu_ja_type_class {
 	RCU_JA_NULL,	/* not an encoded type, but keeps code regular */
 };
 
-struct rcu_ja_type {
-	enum rcu_ja_type_class type_class;
+struct cds_ja_type {
+	enum cds_ja_type_class type_class;
 	uint16_t min_child;		/* minimum number of children: 1 to 256 */
 	uint16_t max_child;		/* maximum number of children: 1 to 256 */
 	uint16_t max_linear_child;	/* per-pool max nr. children: 1 to 256 */
@@ -64,11 +66,14 @@ struct rcu_ja_type {
  * child type.
  */
 #define JA_TYPE_BITS	3
-#define JA_TYPE_MAX_NR	(1U << JA_TYPE_BITS)
+#define JA_TYPE_MAX_NR	(1UL << JA_TYPE_BITS)
 #define JA_TYPE_MASK	(JA_TYPE_MAX_NR - 1)
 #define JA_PTR_MASK	(~JA_TYPE_MASK)
 
 #define JA_ENTRY_PER_NODE	256UL
+#define JA_BITS_PER_BYTE	3
+
+#define JA_MAX_DEPTH	5	/* Maximum depth, including leafs */
 
 /*
  * Entry for NULL node is at index 8 of the table. It is never encoded
@@ -76,6 +81,12 @@ struct rcu_ja_type {
  */
 #define NODE_INDEX_NULL		8
 
+/*
+ * Number of removals needed on a fallback node before we try to shrink
+ * it.
+ */
+#define JA_FALLBACK_REMOVAL_COUNT	8
+
 /*
  * Iteration on the array to find the right node size for the number of
  * children stops when it reaches .max_child == 256 (this is the largest
@@ -123,7 +134,7 @@ enum {
 	ja_type_6_nr_pool_order = 2,
 };
 
-const struct rcu_ja_type ja_types[] = {
+const struct cds_ja_type ja_types[] = {
 	{ .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 3, },
 	{ .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 4, },
 	{ .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 5, },
@@ -171,7 +182,7 @@ enum {
 	ja_type_6_nr_pool_order = 2,
 };
 
-const struct rcu_ja_type ja_types[] = {
+const struct cds_ja_type ja_types[] = {
 	{ .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 4, },
 	{ .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 5, },
 	{ .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 6, },
@@ -199,7 +210,7 @@ void static_array_size_check(void)
 }
 
 /*
- * The rcu_ja_node contains the compressed node data needed for
+ * The cds_ja_node contains the compressed node data needed for
  * read-side. For linear and pool node configurations, it starts with a
  * byte counting the number of children in the node.  Then, the
  * node-specific data is placed.
@@ -214,7 +225,7 @@ void static_array_size_check(void)
 	struct {										\
 		uint8_t nr_child;								\
 		uint8_t child_value[ja_type_## index ##_max_linear_child];			\
-		struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_linear_child];	\
+		struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child];	\
 	}
 
 #define DECLARE_POOL_NODE(index)								\
@@ -222,11 +233,11 @@ void static_array_size_check(void)
 		struct {									\
 			uint8_t nr_child;							\
 			uint8_t child_value[ja_type_## index ##_max_linear_child];		\
-			struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
+			struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \
 		} linear[1U << ja_type_## index ##_nr_pool_order];				\
 	}
 
-struct rcu_ja_node {
+struct cds_ja_inode {
 	union {
 		/* Linear configuration */
 		DECLARE_LINEAR_NODE(0) conf_0;
@@ -241,46 +252,46 @@ struct rcu_ja_node {
 
 		/* Pigeon configuration */
 		struct {
-			struct rcu_ja_node_flag *child[ja_type_7_max_child];
+			struct cds_ja_inode_flag *child[ja_type_7_max_child];
 		} conf_7;
 		/* data aliasing nodes for computed accesses */
-		uint8_t data[sizeof(struct rcu_ja_node_flag *) * ja_type_7_max_child];
+		uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child];
 	} u;
 };
 
 static
-struct rcu_ja_node_flag *ja_node_flag(struct rcu_ja_node *node,
-		unsigned int type)
+struct cds_ja_inode_flag *ja_node_flag(struct cds_ja_inode *node,
+		unsigned long type)
 {
-	assert(type < RCU_JA_NR_TYPES);
-	return (struct rcu_ja_node_flag *) (((unsigned long) node) | type);
+	assert(type < (1UL << JA_TYPE_BITS));
+	return (struct cds_ja_inode_flag *) (((unsigned long) node) | type);
 }
 
 static
-struct rcu_ja_node *ja_node_ptr(struct rcu_ja_node_flag *node)
+struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node)
 {
-	return (struct rcu_ja_node *) (((unsigned long) node) | JA_PTR_MASK);
+	return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK);
 }
 
 static
-unsigned int ja_node_type(struct rcu_ja_node_flag *node)
+unsigned long ja_node_type(struct cds_ja_inode_flag *node)
 {
-	unsigned int type;
+	unsigned long type;
 
 	if (ja_node_ptr(node) == NULL) {
 		return NODE_INDEX_NULL;
 	}
 	type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK);
-	assert(type < RCU_JA_NR_TYPES);
+	assert(type < (1UL << JA_TYPE_BITS));
 	return type;
 }
 
-struct rcu_ja_node *alloc_rcu_ja_node(const struct rcu_ja_type *ja_type)
+struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type)
 {
 	return calloc(1U << ja_type->order, sizeof(char));
 }
 
-void free_rcu_ja_node(struct rcu_ja_node *node)
+void free_cds_ja_node(struct cds_ja_inode *node)
 {
 	free(node);
 }
@@ -296,25 +307,34 @@ uint8_t *align_ptr_size(uint8_t *ptr)
 	return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *));
 }
 
+static
+uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type,
+		struct cds_ja_inode *node)
+{
+	assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL);
+	return CMM_LOAD_SHARED(node->u.data[0]);
+}
+
 /*
  * The order in which values and pointers are does does not matter: if
  * a value is missing, we return NULL. If a value is there, but its
  * associated pointers is still NULL, we return NULL too.
  */
 static
-struct rcu_ja_node_flag *ja_linear_node_get_nth(const struct rcu_ja_type *type,
-		struct rcu_ja_node *node,
+struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type,
+		struct cds_ja_inode *node,
+		struct cds_ja_inode_flag ***child_node_flag_ptr,
 		uint8_t n)
 {
 	uint8_t nr_child;
 	uint8_t *values;
-	struct rcu_ja_node_flag **pointers;
-	struct rcu_ja_node_flag *ptr;
+	struct cds_ja_inode_flag **pointers;
+	struct cds_ja_inode_flag *ptr;
 	unsigned int i;
 
 	assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL);
 
-	nr_child = CMM_LOAD_SHARED(node->u.data[0]);
+	nr_child = ja_linear_node_get_nr_child(type, node);
 	cmm_smp_rmb();	/* read nr_child before values and pointers */
 	assert(nr_child <= type->max_linear_child);
 	assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child);
@@ -326,36 +346,74 @@ struct rcu_ja_node_flag *ja_linear_node_get_nth(const struct rcu_ja_type *type,
 	}
 	if (i >= nr_child)
 		return NULL;
-	pointers = (struct rcu_ja_node_flag **) align_ptr_size(&values[type->max_linear_child]);
+	pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]);
+	if (caa_unlikely(child_node_flag_ptr))
+		*child_node_flag_ptr = &pointers[i];
 	ptr = rcu_dereference(pointers[i]);
 	assert(ja_node_ptr(ptr) != NULL);
 	return ptr;
 }
 
 static
-struct rcu_ja_node_flag *ja_pool_node_get_nth(const struct rcu_ja_type *type,
-		struct rcu_ja_node *node,
+void ja_linear_node_get_ith_pos(const struct cds_ja_type *type,
+		struct cds_ja_inode *node,
+		uint8_t i,
+		uint8_t *v,
+		struct cds_ja_inode_flag **iter)
+{
+	uint8_t *values;
+	struct cds_ja_inode_flag **pointers;
+
+	assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL);
+	assert(i < ja_linear_node_get_nr_child(type, node));
+
+	values = &node->u.data[1];
+	*v = values[i];
+	pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]);
+	*iter = pointers[i];
+}
+
+static
+struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type,
+		struct cds_ja_inode *node,
+		struct cds_ja_inode_flag ***child_node_flag_ptr,
 		uint8_t n)
 {
-	struct rcu_ja_node *linear;
+	struct cds_ja_inode *linear;
 
 	assert(type->type_class == RCU_JA_POOL);
 	/*
 	 * TODO: currently, we select the pool by highest bits. We
 	 * should support various encodings.
 	 */
-	linear = (struct rcu_ja_node *)
+	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_get_nth(type, linear, n);
+	return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, n);
+}
+
+static
+struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type,
+		struct cds_ja_inode *node,
+		uint8_t i)
+{
+	assert(type->type_class == RCU_JA_POOL);
+	return (struct cds_ja_inode *)
+		&node->u.data[(unsigned int) i << type->pool_size_order];
 }
 
 static
-struct rcu_ja_node_flag *ja_pigeon_node_get_nth(const struct rcu_ja_type *type,
-		struct rcu_ja_node *node,
+struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type,
+		struct cds_ja_inode *node,
+		struct cds_ja_inode_flag ***child_node_flag_ptr,
 		uint8_t n)
 {
+	struct cds_ja_inode_flag **child_node_flag;
+
 	assert(type->type_class == RCU_JA_PIGEON);
-	return rcu_dereference(((struct rcu_ja_node_flag **) node->u.data)[n]);
+	child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n];
+	if (caa_unlikely(child_node_flag_ptr))
+		*child_node_flag_ptr = child_node_flag;
+	return rcu_dereference(*child_node_flag);
 }
 
 /*
@@ -363,12 +421,13 @@ struct rcu_ja_node_flag *ja_pigeon_node_get_nth(const struct rcu_ja_type *type,
  * node_flag is already rcu_dereference'd.
  */
 static
-struct rcu_ja_node_flag *ja_node_get_nth(struct rcu_ja_node_flag *node_flag,
+struct cds_ja_inode_flag * ja_node_get_nth(struct cds_ja_inode_flag *node_flag,
+		struct cds_ja_inode_flag ***child_node_flag_ptr,
 		uint8_t n)
 {
 	unsigned int type_index;
-	struct rcu_ja_node *node;
-	const struct rcu_ja_type *type;
+	struct cds_ja_inode *node;
+	const struct cds_ja_type *type;
 
 	node = ja_node_ptr(node_flag);
 	assert(node != NULL);
@@ -377,41 +436,50 @@ struct rcu_ja_node_flag *ja_node_get_nth(struct rcu_ja_node_flag *node_flag,
 
 	switch (type->type_class) {
 	case RCU_JA_LINEAR:
-		return ja_linear_node_get_nth(type, node, n);
+		return ja_linear_node_get_nth(type, node,
+				child_node_flag_ptr, n);
 	case RCU_JA_POOL:
-		return ja_pool_node_get_nth(type, node, n);
+		return ja_pool_node_get_nth(type, node,
+				child_node_flag_ptr, n);
 	case RCU_JA_PIGEON:
-		return ja_pigeon_node_get_nth(type, node, n);
+		return ja_pigeon_node_get_nth(type, node,
+				child_node_flag_ptr, n);
 	default:
 		assert(0);
 		return (void *) -1UL;
 	}
 }
 
+/*
+ * 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 rcu_ja_shadow_node *shadow_node)
+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 rcu_ja_type *type,
-		struct rcu_ja_node *node,
-		struct rcu_ja_shadow_node *shadow_node,
+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 rcu_ja_node_flag *child_node_flag)
+		struct cds_ja_inode_flag *child_node_flag)
 {
 	uint8_t nr_child;
 	uint8_t *values, *nr_child_ptr;
-	struct rcu_ja_node_flag **pointers;
+	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);
-	assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child);
 
 	values = &node->u.data[1];
 	for (i = 0; i < nr_child; i++) {
@@ -422,44 +490,49 @@ int ja_linear_node_set_nth(const struct rcu_ja_type *type,
 		/* No space left in this node type */
 		return -ENOSPC;
 	}
-	pointers = (struct rcu_ja_node_flag **) align_ptr_size(&values[type->max_linear_child]);
+	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 rcu_ja_type *type,
-		struct rcu_ja_node *node,
-		struct rcu_ja_shadow_node *shadow_node,
+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 rcu_ja_node_flag *child_node_flag)
+		struct cds_ja_inode_flag *child_node_flag)
 {
-	struct rcu_ja_node *linear;
+	struct cds_ja_inode *linear;
 
 	assert(type->type_class == RCU_JA_POOL);
-	linear = (struct rcu_ja_node *)
+	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 rcu_ja_type *type,
-		struct rcu_ja_node *node,
-		struct rcu_ja_shadow_node *shadow_node,
+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 rcu_ja_node_flag *child_node_flag)
+		struct cds_ja_inode_flag *child_node_flag)
 {
-	struct rcu_ja_node_flag **ptr;
+	struct cds_ja_inode_flag **ptr;
 
 	assert(type->type_class == RCU_JA_PIGEON);
-	ptr = &((struct rcu_ja_node_flag **) node->u.data)[n];
-	if (*ptr != NULL)
+	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++;
@@ -472,11 +545,11 @@ int ja_pigeon_node_set_nth(const struct rcu_ja_type *type,
  * TODO: exclusive access on node.
  */
 static
-int _ja_node_set_nth(const struct rcu_ja_type *type,
-		struct rcu_ja_node *node,
-		struct rcu_ja_shadow_node *shadow_node,
+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 rcu_ja_node_flag *child_node_flag)
+		struct cds_ja_inode_flag *child_node_flag)
 {
 	switch (type->type_class) {
 	case RCU_JA_LINEAR:
@@ -501,92 +574,489 @@ int _ja_node_set_nth(const struct rcu_ja_type *type,
 /*
  * 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 rcu_ja *ja,
+int ja_node_recompact_add(struct cds_ja *ja,
 		unsigned int old_type_index,
-		const struct rcu_ja_type *old_type,
-		struct rcu_ja_node *old_node,
-		struct rcu_ja_shadow_node **shadow_node,
-		struct rcu_ja_node_flag **old_node_flag, uint8_t n,
-		struct rcu_ja_node_flag *child_node_flag)
+		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 rcu_ja_node *new_node;
-	const struct rcu_ja_type *new_type;
-	struct rcu_ja_node_flag *new_node_flag;
-	unsigned int i;
+	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 == NULL) {
+	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_rcu_ja_node(new_type);
+	new_node = alloc_cds_ja_node(new_type);
 	if (!new_node)
 		return -ENOMEM;
 	new_node_flag = ja_node_flag(new_node, new_type_index);
 
-	ret = rcuja_shadow_set(ja->ht, new_node, *shadow_node);
-	if (ret)
-		return ret;
-
-	if (*shadow_node == NULL) {
-		*shadow_node = rcuja_shadow_lookup_lock(ja->ht, new_node);
-		assert(*shadow_node);
+	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;
 
-	(*shadow_node)->nr_child = 0;
-	for (i = 0; i < old_type->max_child; i++) {
-		struct rcu_ja_node_flag *iter;
-
-		iter = ja_node_get_nth(*old_node_flag, i);
-		if (!iter)
-			continue;
-		ret = _ja_node_set_nth(new_type, new_node, *shadow_node,
-				i, iter);
-		assert(!ret);
+	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, *shadow_node,
+	ret = _ja_node_set_nth(new_type, new_node,
+			new_shadow_node,
 			n, child_node_flag);
 	assert(!ret);
-	/* Replace the old node with the new recompacted one */
-	rcu_assign_pointer(*old_node_flag, new_node_flag);
-	ret = rcuja_shadow_clear(ja->ht, old_node,
+	/* 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);
-	return 0;
+
+	/* 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 rcu_ja *ja,
-		struct rcu_ja_node_flag **node_flag, uint8_t n,
-		struct rcu_ja_node_flag *child_node_flag)
+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 rcu_ja_type *type;
-	struct rcu_ja_node *node;
-	struct rcu_ja_shadow_node *shadow_node = NULL;
+	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];
-	if (node != NULL) {
-		shadow_node = rcuja_shadow_lookup_lock(ja->ht, node);
-		assert(shadow_node);
-	}
 	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);
-		/* recompact always leave shadow_node locked */
+				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++) {
+		node_flag = ja_node_get_nth(node_flag, NULL,
+			(unsigned char) key);
+		if (!ja_node_ptr(node_flag))
+			return NULL;
+		key >>= JA_BITS_PER_BYTE;
+	}
+
+	/* 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 depth,
+		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 depth %u\n", depth);
+
+	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 - 1; i >= (int) depth; i--) {
+		dbg_printf("branch creation level %d, key %" PRIu64 "\n",
+				i, key >> (JA_BITS_PER_BYTE * (i - 2)));
+		iter_dest_node_flag = NULL;
+		ret = ja_node_set_nth(ja, &iter_dest_node_flag,
+			key >> (JA_BITS_PER_BYTE * (i - 2)),
+			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 (depth > 1) {
+		/* 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,
+			key >> (JA_BITS_PER_BYTE * (depth - 2)),
+			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.
+ */
+static
+int ja_chain_node(struct cds_ja *ja,
+		struct cds_hlist_head *head,
+		struct cds_ja_node *node)
+{
+	struct cds_ja_shadow_node *shadow_node;
+
+	shadow_node = rcuja_shadow_lookup_lock(ja->ht,
+		(struct cds_ja_inode *) head);
+	if (!shadow_node)
+		return -ENOENT;
+	cds_hlist_add_head_rcu(&node->list, head);
 	rcuja_shadow_unlock(shadow_node);
+	return 0;
+}
+
+int cds_ja_add(struct cds_ja *ja, uint64_t key,
+		struct cds_ja_node *new_node)
+{
+	unsigned int tree_depth, i;
+	uint64_t iter_key;
+	struct cds_ja_inode_flag **node_flag_ptr;	/* in parent */
+	struct cds_ja_inode_flag *node_flag,
+		*parent_node_flag,
+		*parent2_node_flag;
+	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);
+	iter_key = key;
+	parent2_node_flag = NULL;
+	parent_node_flag =
+		(struct cds_ja_inode_flag *) &ja->root;	/* Use root ptr address as key for mutex */
+	node_flag_ptr = &ja->root;
+	node_flag = rcu_dereference(*node_flag_ptr);
+
+	/* Iterate on all internal levels */
+	for (i = 1; i < tree_depth; i++) {
+		if (!ja_node_ptr(node_flag)) {
+			ret = ja_attach_node(ja, node_flag_ptr,
+					parent_node_flag, parent2_node_flag,
+					key, i, new_node);
+			if (ret == -ENOENT || ret == -EEXIST)
+				goto retry;
+			else
+				goto end;
+		}
+		parent2_node_flag = parent_node_flag;
+		parent_node_flag = node_flag;
+		node_flag = ja_node_get_nth(node_flag,
+			&node_flag_ptr,
+			(unsigned char) iter_key);
+		iter_key >>= JA_BITS_PER_BYTE;
+	}
+
+	/*
+	 * We reached bottom of tree, simply add node to last internal
+	 * level, or chain it if key is already present.
+	 */
+	if (!ja_node_ptr(node_flag)) {
+		ret = ja_attach_node(ja, node_flag_ptr, parent_node_flag,
+				parent2_node_flag, key, i, new_node);
+	} else {
+		ret = ja_chain_node(ja,
+			(struct cds_hlist_head *) ja_node_ptr(node_flag),
+			new_node);
+	}
+	if (ret == -ENOENT)
+		goto retry;
+end:
 	return ret;
 }
+
+struct cds_ja *_cds_ja_new(unsigned int key_bits,
+		const struct rcu_flavor_struct *flavor)
+{
+	struct cds_ja *ja;
+	int ret;
+	struct cds_ja_shadow_node *root_shadow_node;
+
+	ja = calloc(sizeof(*ja), 1);
+	if (!ja)
+		goto ja_error;
+
+	switch (key_bits) {
+	case 8:
+		ja->key_max = UINT8_MAX;
+		break;
+	case 16:
+		ja->key_max = UINT16_MAX;
+		break;
+	case 32:
+		ja->key_max = UINT32_MAX;
+		break;
+	case 64:
+		ja->key_max = UINT64_MAX;
+		break;
+	default:
+		goto check_error;
+	}
+
+	/* ja->root is NULL */
+	/* tree_depth 0 is for pointer to root node */
+	ja->tree_depth = (key_bits >> JA_BITS_PER_BYTE) + 1;
+	assert(ja->tree_depth <= JA_MAX_DEPTH);
+	ja->ht = rcuja_create_ht(flavor);
+	if (!ja->ht)
+		goto ht_error;
+
+	/*
+	 * Note: we should not free this node until judy array destroy.
+	 */
+	root_shadow_node = rcuja_shadow_set(ja->ht,
+			ja_node_ptr((struct cds_ja_inode_flag *) &ja->root),
+			NULL);
+	if (!root_shadow_node) {
+		ret = -ENOMEM;
+		goto ht_node_error;
+	}
+
+	return ja;
+
+ht_node_error:
+	ret = rcuja_delete_ht(ja->ht);
+	assert(!ret);
+ht_error:
+check_error:
+	free(ja);
+ja_error:
+	return NULL;
+}
+
+/*
+ * There should be no more concurrent add to the judy array while it is
+ * being destroyed (ensured by the caller).
+ */
+int cds_ja_destroy(struct cds_ja *ja)
+{
+	int ret;
+
+	rcuja_shadow_prune(ja->ht,
+		RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK);
+	ret = rcuja_delete_ht(ja->ht);
+	if (ret)
+		return ret;
+	if (uatomic_read(&ja->nr_fallback))
+		fprintf(stderr,
+			"[warning] RCU Judy Array used %lu fallback node(s)\n",
+			uatomic_read(&ja->nr_fallback));
+	free(ja);
+	return 0;
+}