X-Git-Url: http://git.liburcu.org/?a=blobdiff_plain;f=rcuja%2Frcuja.c;h=56a6e566261b955065e785c1c36fa29ce44ba89e;hb=4a073c5361c1ac6dfcf01cf3f9ec8a2380486842;hp=12cce1d1d42e8af6fc3efe2a20e5cf0b2652e373;hpb=3d8fe30760fd4fb67ce24b5aa011ab75f680a904;p=userspace-rcu.git

diff --git a/rcuja/rcuja.c b/rcuja/rcuja.c
index 12cce1d..56a6e56 100644
--- a/rcuja/rcuja.c
+++ b/rcuja/rcuja.c
@@ -24,6 +24,7 @@
 #include <stdint.h>
 #include <errno.h>
 #include <limits.h>
+#include <string.h>
 #include <urcu/rcuja.h>
 #include <urcu/compiler.h>
 #include <urcu/arch.h>
@@ -35,6 +36,10 @@
 #include "rcuja-internal.h"
 #include "bitfield.h"
 
+#ifndef abs
+#define abs_int(a)	((int) (a) > 0 ? (int) (a) : -((int) (a)))
+#endif
+
 enum cds_ja_type_class {
 	RCU_JA_LINEAR = 0,	/* Type A */
 			/* 32-bit: 1 to 25 children, 8 to 128 bytes */
@@ -120,8 +125,8 @@ const struct cds_ja_type ja_types[] = {
 	{ .type_class = RCU_JA_POOL, .min_child = 45, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 9, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 7, },
 
 	/*
-	 * TODO: Upon node removal below min_child, if child pool is
-	 * filled beyond capacity, we need to roll back to pigeon.
+	 * Upon node removal below min_child, if child pool is filled
+	 * beyond capacity, we roll back to pigeon.
 	 */
 	{ .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, },
 
@@ -168,8 +173,8 @@ const struct cds_ja_type ja_types[] = {
 	{ .type_class = RCU_JA_POOL, .min_child = 51, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 10, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 8, },
 
 	/*
-	 * TODO: Upon node removal below min_child, if child pool is
-	 * filled beyond capacity, we need to roll back to pigeon.
+	 * Upon node removal below min_child, if child pool is filled
+	 * beyond capacity, we roll back to pigeon.
 	 */
 	{ .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, },
 
@@ -234,14 +239,68 @@ struct cds_ja_inode {
 };
 
 enum ja_recompact {
-	JA_RECOMPACT,
-	JA_RECOMPACT_ADD,
+	JA_RECOMPACT_ADD_SAME,
+	JA_RECOMPACT_ADD_NEXT,
 	JA_RECOMPACT_DEL,
 };
 
+static
+unsigned long node_fallback_count_distribution[JA_ENTRY_PER_NODE];
+
+static
+struct cds_ja_inode *_ja_node_mask_ptr(struct cds_ja_inode_flag *node)
+{
+	return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK);
+}
+
+unsigned long ja_node_type(struct cds_ja_inode_flag *node)
+{
+	unsigned long type;
+
+	if (_ja_node_mask_ptr(node) == NULL) {
+		return NODE_INDEX_NULL;
+	}
+	type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK);
+	assert(type < (1UL << JA_TYPE_BITS));
+	return type;
+}
+
+struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node)
+{
+	unsigned long type_index = ja_node_type(node);
+	const struct cds_ja_type *type;
+
+	type = &ja_types[type_index];
+	switch (type->type_class) {
+	case RCU_JA_LINEAR:
+	case RCU_JA_PIGEON:	/* fall-through */
+	case RCU_JA_NULL:	/* fall-through */
+	default:		/* fall-through */
+		return _ja_node_mask_ptr(node);
+	case RCU_JA_POOL:
+		switch (type->nr_pool_order) {
+		case 1:
+			return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_1D_MASK | JA_TYPE_MASK));
+		case 2:
+			return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_2D_MASK | JA_POOL_1D_MASK | JA_TYPE_MASK));
+		default:
+			assert(0);
+		}
+	}
+}
+
 struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type)
 {
-	return calloc(1U << ja_type->order, sizeof(char));
+	size_t len = 1U << ja_type->order;
+	void *p;
+	int ret;
+
+	ret = posix_memalign(&p, len, len);
+	if (ret || !p) {
+		return NULL;
+	}
+	memset(p, 0, len);
+	return p;
 }
 
 void free_cds_ja_node(struct cds_ja_inode *node)
@@ -277,6 +336,8 @@ static
 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,
+		struct cds_ja_inode_flag **child_node_flag_v,
+		struct cds_ja_inode_flag ***node_flag_ptr,
 		uint8_t n)
 {
 	uint8_t nr_child;
@@ -297,12 +358,19 @@ struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type,
 		if (CMM_LOAD_SHARED(values[i]) == n)
 			break;
 	}
-	if (i >= nr_child)
+	if (i >= nr_child) {
+		if (caa_unlikely(node_flag_ptr))
+			*node_flag_ptr = NULL;
 		return NULL;
+	}
 	pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]);
 	ptr = rcu_dereference(pointers[i]);
 	if (caa_unlikely(child_node_flag_ptr) && ptr)
 		*child_node_flag_ptr = &pointers[i];
+	if (caa_unlikely(child_node_flag_v) && ptr)
+		*child_node_flag_v = ptr;
+	if (caa_unlikely(node_flag_ptr))
+		*node_flag_ptr = &pointers[i];
 	return ptr;
 }
 
@@ -328,19 +396,47 @@ void ja_linear_node_get_ith_pos(const struct cds_ja_type *type,
 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 *node_flag,
 		struct cds_ja_inode_flag ***child_node_flag_ptr,
+		struct cds_ja_inode_flag **child_node_flag_v,
+		struct cds_ja_inode_flag ***node_flag_ptr,
 		uint8_t n)
 {
 	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 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, child_node_flag_ptr, n);
+
+	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) & 0x1;
+		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_get_nth(type, linear, child_node_flag_ptr,
+		child_node_flag_v, node_flag_ptr, n);
 }
 
 static
@@ -357,17 +453,25 @@ static
 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,
+		struct cds_ja_inode_flag **child_node_flag_v,
+		struct cds_ja_inode_flag ***node_flag_ptr,
 		uint8_t n)
 {
 	struct cds_ja_inode_flag **child_node_flag;
+	struct cds_ja_inode_flag *child_node_flag_read;
 
 	assert(type->type_class == RCU_JA_PIGEON);
 	child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n];
+	child_node_flag_read = rcu_dereference(*child_node_flag);
 	dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n",
 		child_node_flag);
-	if (caa_unlikely(child_node_flag_ptr) && *child_node_flag)
+	if (caa_unlikely(child_node_flag_ptr) && child_node_flag_read)
 		*child_node_flag_ptr = child_node_flag;
-	return rcu_dereference(*child_node_flag);
+	if (caa_unlikely(child_node_flag_v) && child_node_flag_read)
+		*child_node_flag_v = child_node_flag_read;
+	if (caa_unlikely(node_flag_ptr))
+		*node_flag_ptr = child_node_flag;
+	return child_node_flag_read;
 }
 
 static
@@ -375,7 +479,7 @@ struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *t
 		struct cds_ja_inode *node,
 		uint8_t i)
 {
-	return ja_pigeon_node_get_nth(type, node, NULL, i);
+	return ja_pigeon_node_get_nth(type, node, NULL, NULL, NULL, i);
 }
 
 /*
@@ -383,8 +487,10 @@ struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *t
  * node_flag is already rcu_dereference'd.
  */
 static
-struct cds_ja_inode_flag * ja_node_get_nth(struct cds_ja_inode_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,
+		struct cds_ja_inode_flag **child_node_flag,
+		struct cds_ja_inode_flag ***node_flag_ptr,
 		uint8_t n)
 {
 	unsigned int type_index;
@@ -399,13 +505,16 @@ struct cds_ja_inode_flag * ja_node_get_nth(struct cds_ja_inode_flag *node_flag,
 	switch (type->type_class) {
 	case RCU_JA_LINEAR:
 		return ja_linear_node_get_nth(type, node,
-				child_node_flag_ptr, n);
+				child_node_flag_ptr, child_node_flag,
+				node_flag_ptr, n);
 	case RCU_JA_POOL:
-		return ja_pool_node_get_nth(type, node,
-				child_node_flag_ptr, n);
+		return ja_pool_node_get_nth(type, node, node_flag,
+				child_node_flag_ptr, child_node_flag,
+				node_flag_ptr, n);
 	case RCU_JA_PIGEON:
 		return ja_pigeon_node_get_nth(type, node,
-				child_node_flag_ptr, n);
+				child_node_flag_ptr, child_node_flag,
+				node_flag_ptr, n);
 	default:
 		assert(0);
 		return (void *) -1UL;
@@ -473,6 +582,7 @@ int ja_linear_node_set_nth(const struct cds_ja_type *type,
 static
 int ja_pool_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)
@@ -480,8 +590,37 @@ int ja_pool_node_set_nth(const struct cds_ja_type *type,
 	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];
+
+	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) & 0x1;
+		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_set_nth(type, linear, shadow_node,
 			n, child_node_flag);
 }
@@ -511,6 +650,7 @@ int ja_pigeon_node_set_nth(const struct cds_ja_type *type,
 static
 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)
@@ -520,7 +660,7 @@ int _ja_node_set_nth(const struct cds_ja_type *type,
 		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,
+		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,
@@ -547,18 +687,19 @@ int ja_linear_node_clear_ptr(const struct cds_ja_type *type,
 	assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL);
 
 	nr_child_ptr = &node->u.data[0];
-	dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr);
 	nr_child = *nr_child_ptr;
 	assert(nr_child <= type->max_linear_child);
 
 	if (shadow_node->fallback_removal_count) {
 		shadow_node->fallback_removal_count--;
 	} else {
-		if (shadow_node->nr_child <= type->min_child) {
+		if (type->type_class == RCU_JA_LINEAR
+				&& 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);
 	/*
@@ -573,13 +714,13 @@ int ja_linear_node_clear_ptr(const struct cds_ja_type *type,
 		(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)
@@ -587,8 +728,45 @@ int ja_pool_node_clear_ptr(const struct cds_ja_type *type,
 	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];
+
+	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);
 }
 
@@ -599,6 +777,15 @@ int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type,
 		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;
@@ -611,6 +798,7 @@ int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type,
 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)
@@ -619,7 +807,7 @@ int _ja_node_clear_ptr(const struct cds_ja_type *type,
 	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, shadow_node, node_flag_ptr, n);
+		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:
@@ -632,9 +820,359 @@ int _ja_node_clear_ptr(const struct cds_ja_type *type,
 	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:
+	{
+		uint8_t nr_child;
+		unsigned int i;
+
+		assert(mode == JA_RECOMPACT_DEL);
+		nr_child = shadow_node->nr_child;
+		for (i = 0; i < nr_child; 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:
+	{
+		uint8_t nr_child;
+		unsigned int i;
+
+		assert(mode == JA_RECOMPACT_DEL);
+		nr_child = shadow_node->nr_child;
+		for (i = 0; i < nr_child; 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];
+}
+
 /*
  * 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, -EAGAIN if need to retry, or other negative
  * error value otherwise.
  */
@@ -660,21 +1198,62 @@ int ja_node_recompact(enum ja_recompact mode,
 	old_node_flag = *old_node_flag_ptr;
 
 	switch (mode) {
-	case JA_RECOMPACT:
-		new_type_index = old_type_index;
+	case JA_RECOMPACT_ADD_SAME:
+		if (old_type->type_class == RCU_JA_POOL) {
+			/*
+			 * For pool type, try redistributing
+			 * into a different distribution of same
+			 * size if we have not reached limits.
+			 */
+			if (shadow_node->nr_child + 1 > old_type->max_child) {
+				new_type_index = old_type_index + 1;
+			} else if (shadow_node->nr_child + 1 < old_type->min_child) {
+				new_type_index = old_type_index - 1;
+			} else {
+				new_type_index = old_type_index;
+			}
+		} else {
+			new_type_index = old_type_index;
+		}
 		break;
-	case JA_RECOMPACT_ADD:
+	case JA_RECOMPACT_ADD_NEXT:
 		if (!shadow_node || old_type_index == NODE_INDEX_NULL) {
 			new_type_index = 0;
 		} else {
-			new_type_index = old_type_index + 1;
+			if (old_type->type_class == RCU_JA_POOL) {
+				/*
+				 * For pool type, try redistributing
+				 * into a different distribution of same
+				 * size if we have not reached limits.
+				 */
+				if (shadow_node->nr_child + 1 > old_type->max_child) {
+					new_type_index = old_type_index + 1;
+				} else {
+					new_type_index = old_type_index;
+				}
+			} else {
+				new_type_index = old_type_index + 1;
+			}
 		}
 		break;
 	case JA_RECOMPACT_DEL:
 		if (old_type_index == 0) {
 			new_type_index = NODE_INDEX_NULL;
 		} else {
-			new_type_index = old_type_index - 1;
+			if (old_type->type_class == RCU_JA_POOL) {
+				/*
+				 * For pool type, try redistributing
+				 * into a different distribution of same
+				 * size if we have not reached limits.
+				 */
+				if (shadow_node->nr_child - 1 < old_type->min_child) {
+					new_type_index = old_type_index - 1;
+				} else {
+					new_type_index = old_type_index;
+				}
+			} else {
+				new_type_index = old_type_index - 1;
+			}
 		}
 		break;
 	default:
@@ -689,7 +1268,47 @@ retry:		/* for fallback */
 		new_node = alloc_cds_ja_node(new_type);
 		if (!new_node)
 			return -ENOMEM;
-		new_node_flag = ja_node_flag(new_node, new_type_index);
+
+		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);
 		if (!new_shadow_node) {
@@ -704,7 +1323,7 @@ retry:		/* for fallback */
 		new_node_flag = NULL;
 	}
 
-	assert(mode != JA_RECOMPACT_ADD || old_type->type_class != RCU_JA_PIGEON);
+	assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON);
 
 	if (new_type_index == NODE_INDEX_NULL)
 		goto skip_copy;
@@ -725,7 +1344,7 @@ retry:		/* for fallback */
 				continue;
 			if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
 				continue;
-			ret = _ja_node_set_nth(new_type, new_node,
+			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) {
@@ -757,7 +1376,7 @@ retry:		/* for fallback */
 					continue;
 				if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
 					continue;
-				ret = _ja_node_set_nth(new_type, new_node,
+				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
@@ -770,7 +1389,7 @@ retry:		/* for fallback */
 		break;
 	}
 	case RCU_JA_NULL:
-		assert(mode == JA_RECOMPACT_ADD);
+		assert(mode == JA_RECOMPACT_ADD_NEXT);
 		break;
 	case RCU_JA_PIGEON:
 	{
@@ -787,7 +1406,7 @@ retry:		/* for fallback */
 				continue;
 			if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter)
 				continue;
-			ret = _ja_node_set_nth(new_type, new_node,
+			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) {
@@ -804,13 +1423,24 @@ retry:		/* for fallback */
 	}
 skip_copy:
 
-	if (JA_RECOMPACT_ADD) {
+	if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) {
 		/* add node */
-		ret = _ja_node_set_nth(new_type, new_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) {
@@ -840,7 +1470,69 @@ fallback_toosmall:
 			RCUJA_SHADOW_CLEAR_FREE_NODE);
 	assert(!ret);
 
-	/* Choose fallback type: pigeon */
+	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);
@@ -869,17 +1561,17 @@ int ja_node_set_nth(struct cds_ja *ja,
 	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,
+	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. */
-		ret = ja_node_recompact(JA_RECOMPACT_ADD, ja, type_index, type, node,
+		/* 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);
 		break;
 	case -ERANGE:
 		/* Node needs to be recompacted. */
-		ret = ja_node_recompact(JA_RECOMPACT, ja, type_index, type, node,
+		ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node,
 				shadow_node, node_flag, n, child_node_flag, NULL);
 		break;
 	}
@@ -908,9 +1600,9 @@ int ja_node_clear_ptr(struct cds_ja *ja,
 	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, shadow_node, node_flag_ptr, n);
+	ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n);
 	if (ret == -EFBIG) {
-		/* Should to try recompaction. */
+		/* 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);
@@ -937,7 +1629,7 @@ struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key)
 		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,
+		node_flag = ja_node_get_nth(node_flag, NULL, NULL, NULL,
 			iter_key);
 		dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n",
 				(unsigned int) iter_key, node_flag);
@@ -964,6 +1656,8 @@ struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key)
  */
 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 **node_flag_ptr,
 		struct cds_ja_inode_flag *node_flag,
 		struct cds_ja_inode_flag *parent_node_flag,
@@ -999,6 +1693,27 @@ int ja_attach_node(struct cds_ja *ja,
 		}
 	}
 
+	if (node_flag_ptr && ja_node_ptr(*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);
@@ -1039,8 +1754,8 @@ int ja_attach_node(struct cds_ja *ja,
 
 	/* 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);
+		iter_node_flag, *attach_node_flag_ptr);
+	rcu_assign_pointer(*attach_node_flag_ptr, iter_node_flag);
 
 	/* Success */
 	ret = 0;
@@ -1061,6 +1776,7 @@ check_error:
 			assert(!tmpret);
 		}
 	}
+unlock_parent:
 	if (parent_shadow_node)
 		rcuja_shadow_unlock(parent_shadow_node);
 unlock_shadow:
@@ -1079,31 +1795,43 @@ end:
 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_hlist_head *head,
 		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)
+	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, head);
+end:
 	rcuja_shadow_unlock(shadow_node);
-	return 0;
+	return ret;
 }
 
 int cds_ja_add(struct cds_ja *ja, uint64_t key,
 		struct cds_ja_node *new_node)
 {
 	unsigned int tree_depth, i;
-	struct cds_ja_inode_flag **node_flag_ptr;	/* in parent */
+	struct cds_ja_inode_flag **attach_node_flag_ptr,
+		**node_flag_ptr;
 	struct cds_ja_inode_flag *node_flag,
 		*parent_node_flag,
-		*parent2_node_flag;
+		*parent2_node_flag,
+		*attach_node_flag;
 	int ret;
 
-	if (caa_unlikely(key > ja->key_max))
+	if (caa_unlikely(key > ja->key_max)) {
 		return -EINVAL;
+	}
 	tree_depth = ja->tree_depth;
 
 retry:
@@ -1112,6 +1840,8 @@ retry:
 	parent2_node_flag = NULL;
 	parent_node_flag =
 		(struct cds_ja_inode_flag *) &ja->root;	/* Use root ptr address as key for mutex */
+	attach_node_flag_ptr = &ja->root;
+	attach_node_flag = rcu_dereference(ja->root);
 	node_flag_ptr = &ja->root;
 	node_flag = rcu_dereference(ja->root);
 
@@ -1119,11 +1849,14 @@ retry:
 	for (i = 1; i < tree_depth; i++) {
 		uint8_t iter_key;
 
-		dbg_printf("cds_ja_add iter node_flag_ptr %p node_flag %p\n",
-				*node_flag_ptr, node_flag);
+		dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
+				attach_node_flag_ptr, node_flag_ptr, node_flag);
 		if (!ja_node_ptr(node_flag)) {
-			ret = ja_attach_node(ja, node_flag_ptr,
-					parent_node_flag, parent2_node_flag,
+			ret = ja_attach_node(ja, attach_node_flag_ptr,
+					attach_node_flag,
+					node_flag_ptr,
+					parent_node_flag,
+					parent2_node_flag,
 					key, i, new_node);
 			if (ret == -EAGAIN || ret == -EEXIST)
 				goto retry;
@@ -1134,10 +1867,14 @@ retry:
 		parent2_node_flag = parent_node_flag;
 		parent_node_flag = node_flag;
 		node_flag = ja_node_get_nth(node_flag,
+			&attach_node_flag_ptr,
+			&attach_node_flag,
 			&node_flag_ptr,
 			iter_key);
-		dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p node_flag_ptr %p\n",
-				(unsigned int) iter_key, node_flag, *node_flag_ptr);
+		dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n",
+				(unsigned int) iter_key, node_flag,
+				attach_node_flag_ptr,
+				node_flag_ptr);
 	}
 
 	/*
@@ -1145,17 +1882,21 @@ retry:
 	 * level, or chain it if key is already present.
 	 */
 	if (!ja_node_ptr(node_flag)) {
-		dbg_printf("cds_ja_add last node_flag_ptr %p node_flag %p\n",
-				*node_flag_ptr, node_flag);
-		ret = ja_attach_node(ja, node_flag_ptr, parent_node_flag,
+		dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n",
+				attach_node_flag_ptr, node_flag_ptr, node_flag);
+		ret = ja_attach_node(ja, attach_node_flag_ptr,
+				attach_node_flag,
+				node_flag_ptr, parent_node_flag,
 				parent2_node_flag, key, i, new_node);
 	} else {
 		ret = ja_chain_node(ja,
 			parent_node_flag,
-			(struct cds_hlist_head *) node_flag_ptr,
+			node_flag_ptr,
+			node_flag,
+			(struct cds_hlist_head *) attach_node_flag_ptr,
 			new_node);
 	}
-	if (ret == -EAGAIN)
+	if (ret == -EAGAIN || ret == -EEXIST)
 		goto retry;
 end:
 	return ret;
@@ -1168,6 +1909,9 @@ end:
  * 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,
@@ -1183,10 +1927,10 @@ int ja_detach_node(struct cds_ja *ja,
 			*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;
-	uint8_t n;
+	int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0;
+	uint8_t n = 0;
 
-	assert(nr_snapshot == ja->tree_depth - 1);
+	assert(nr_snapshot == ja->tree_depth + 1);
 
 	/*
 	 * From the last internal level node going up, get the node
@@ -1195,7 +1939,7 @@ int ja_detach_node(struct cds_ja *ja,
 	 * 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 - 1; i >= 1; i--) {
+	for (i = nr_snapshot - 2; i >= 1; i--) {
 		struct cds_ja_shadow_node *shadow_node;
 
 		shadow_node = rcuja_shadow_lookup_lock(ja->ht,
@@ -1204,19 +1948,23 @@ int ja_detach_node(struct cds_ja *ja,
 			ret = -EAGAIN;
 			goto end;
 		}
-		assert(shadow_node->nr_child > 0);
 		shadow_nodes[nr_shadow++] = shadow_node;
-		nr_clear++;
-		if (i == nr_snapshot - 1) {
-			/*
-			 * Re-check that last internal node level has
-			 * only one child, else trigger a retry.
-			 */
-			if (shadow_node->nr_child != 1) {
-				ret = -EAGAIN;
-				goto end;
-			}
+
+		/*
+		 * Check if node has been removed between RCU
+		 * lookup and lock acquisition.
+		 */
+		assert(snapshot_ptr[i + 1]);
+		if (ja_node_ptr(*snapshot_ptr[i + 1])
+				!= ja_node_ptr(snapshot[i + 1])) {
+			ret = -ENOENT;
+			goto end;
 		}
+
+		assert(shadow_node->nr_child > 0);
+		if (shadow_node->nr_child == 1 && i > 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,
@@ -1226,10 +1974,23 @@ int ja_detach_node(struct cds_ja *ja,
 				goto end;
 			}
 			shadow_nodes[nr_shadow++] = shadow_node;
-			node_flag_ptr = snapshot_ptr[i];
-			n = snapshot_n[i];
-			parent_node_flag_ptr = snapshot_ptr[i - 1];
-			parent_node_flag = snapshot[i - 1];
+
+			/*
+			 * Check if node has been removed between RCU
+			 * lookup and lock acquisition.
+			 */
+			assert(snapshot_ptr[i]);
+			if (ja_node_ptr(*snapshot_ptr[i])
+					!= ja_node_ptr(snapshot[i])) {
+				ret = -ENOENT;
+				goto end;
+			}
+
+			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
@@ -1242,17 +2003,29 @@ int ja_detach_node(struct cds_ja *ja,
 					goto end;
 				}
 				shadow_nodes[nr_shadow++] = shadow_node;
+
+				/*
+				 * Check if node has been removed between RCU
+				 * lookup and lock acquisition.
+				 */
+				assert(snapshot_ptr[i - 1]);
+				if (ja_node_ptr(*snapshot_ptr[i - 1])
+						!= ja_node_ptr(snapshot[i - 1])) {
+					ret = -ENOENT;
+					goto end;
+				}
 			}
+
 			break;
 		}
 	}
 
 	/*
-	 * At this point, we want to delete all nodes in 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.
+	 * 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 free lock here, because RCU read lock is held,
+	 * and free only performed in call_rcu.
 	 */
 
 	for (i = 0; i < nr_clear; i++) {
@@ -1269,9 +2042,13 @@ int ja_detach_node(struct cds_ja *ja,
 	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_clear],	/* of parent */
+		shadow_nodes[nr_branch - 1],	/* of parent */
 		n);
+	if (ret)
+		goto end;
 
+	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);
 
@@ -1284,23 +2061,48 @@ end:
 static
 int ja_unchain_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;
+	struct cds_hlist_node *hlist_node;
+	struct cds_hlist_head hlist_head;
+	int ret = 0, count = 0, found = 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;
+	}
+	hlist_head.next = (struct cds_hlist_node *) ja_node_ptr(node_flag);
 	/*
 	 * Retry if another thread removed all but one of duplicates
-	 * since check.
+	 * since check (this check was performed without lock).
+	 * Ensure that the node we are about to remove is still in the
+	 * list (while holding lock).
 	 */
-	if (shadow_node->nr_child == 1) {
+	cds_hlist_for_each_rcu(hlist_node, &hlist_head) {
+		if (count == 0) {
+			/* FIXME: currently a work-around */
+			hlist_node->prev = (struct cds_hlist_node *) node_flag_ptr;
+		}
+		count++;
+		if (hlist_node == &node->list)
+			found++;
+	}
+	assert(found <= 1);
+	if (!found || count == 1) {
 		ret = -EAGAIN;
 		goto end;
 	}
 	cds_hlist_del_rcu(&node->list);
+	/*
+	 * Validate that we indeed removed the node from linked list.
+	 */
+	assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node);
 end:
 	rcuja_shadow_unlock(shadow_node);
 	return ret;
@@ -1317,8 +2119,9 @@ int cds_ja_del(struct cds_ja *ja, uint64_t key,
 	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 = 0;
+	struct cds_ja_inode_flag **prev_node_flag_ptr,
+		**node_flag_ptr;
+	int nr_snapshot;
 	int ret;
 
 	if (caa_unlikely(key > ja->key_max))
@@ -1326,15 +2129,18 @@ int cds_ja_del(struct cds_ja *ja, uint64_t key,
 	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[nr_snapshot] = 0;
+	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;
+	node_flag_ptr = &ja->root;
 
 	/* Iterate on all internal levels */
 	for (i = 1; i < tree_depth; i++) {
@@ -1346,58 +2152,68 @@ retry:
 			return -ENOENT;
 		}
 		iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1)));
-		if (nr_snapshot <= 1)
-			snapshot_n[nr_snapshot] = 0;
-		else
-			snapshot_n[nr_snapshot - 1] = iter_key;
-
+		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,
+			NULL,
+			&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_head hlist_head;
 		struct cds_hlist_node *hlist_node;
 		struct cds_ja_node *entry, *match = NULL;
 		int count = 0;
 
-		hlist_head = (struct cds_hlist_head *) ja_node_ptr(node_flag);
+		hlist_head.next =
+			(struct cds_hlist_node *) ja_node_ptr(node_flag);
 		cds_hlist_for_each_entry_rcu(entry,
 				hlist_node,
-				hlist_head,
+				&hlist_head,
 				list) {
+			dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry);
 			if (entry == node)
 				match = entry;
 			count++;
 		}
-		if (!match)
+		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.
+			 * 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, node_flag, entry);
+			ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1],
+				node_flag_ptr, node_flag, match);
 		}
 	}
-	if (ret == -EAGAIN)
+	/*
+	 * Explanation of -ENOENT handling: caused by concurrent delete
+	 * between RCU lookup and actual removal. Need to re-do the
+	 * lookup and removal attempt.
+	 */
+	if (ret == -EAGAIN || ret == -ENOENT)
 		goto retry;
 	return ret;
 }
@@ -1415,13 +2231,13 @@ struct cds_ja *_cds_ja_new(unsigned int key_bits,
 
 	switch (key_bits) {
 	case 8:
-		ja->key_max = UINT8_MAX;
-		break;
 	case 16:
-		ja->key_max = UINT16_MAX;
-		break;
+	case 24:
 	case 32:
-		ja->key_max = UINT32_MAX;
+	case 40:
+	case 48:
+	case 56:
+		ja->key_max = (1ULL << key_bits) - 1;
 		break;
 	case 64:
 		ja->key_max = UINT64_MAX;
@@ -1563,6 +2379,20 @@ void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node,
 	}
 }
 
+static
+void print_debug_fallback_distribution(void)
+{
+	int i;
+
+	fprintf(stderr, "Fallback node distribution:\n");
+	for (i = 0; i < JA_ENTRY_PER_NODE; i++) {
+		if (!node_fallback_count_distribution[i])
+			continue;
+		fprintf(stderr, "	%3u: %4lu\n",
+			i, node_fallback_count_distribution[i]);
+	}
+}
+
 /*
  * There should be no more concurrent add to the judy array while it is
  * being destroyed (ensured by the caller).
@@ -1582,6 +2412,7 @@ int cds_ja_destroy(struct cds_ja *ja,
 		fprintf(stderr,
 			"[warning] RCU Judy Array used %lu fallback node(s)\n",
 			uatomic_read(&ja->nr_fallback));
+	print_debug_fallback_distribution();
 	free(ja);
 	return 0;
 }