X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=0bcbb1e832a5ab41c4d4357c0dafa4ff87832db7;hp=9aeb3e76598071c0b450a4287c2fe8c615bc9a29;hb=4f6e90b73cb103d7d49b8b4838c5f48c3d15b02e;hpb=14044b37bb7ba51676f1c3f139facae9b06dbc51

diff --git a/rculfhash.c b/rculfhash.c
index 9aeb3e7..0bcbb1e 100644
--- a/rculfhash.c
+++ b/rculfhash.c
@@ -1,7 +1,7 @@
 /*
  * rculfhash.c
  *
- * Userspace RCU library - Lock-Free Expandable RCU Hash Table
+ * Userspace RCU library - Lock-Free Resizable RCU Hash Table
  *
  * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  *
@@ -29,7 +29,7 @@
  *   symposium on Parallel algorithms and architectures, ACM Press,
  *   (2002), 73-82.
  *
- * Some specificities of this Lock-Free Expandable RCU Hash Table
+ * Some specificities of this Lock-Free Resizable RCU Hash Table
  * implementation:
  *
  * - RCU read-side critical section allows readers to perform hash
@@ -48,9 +48,13 @@
  * - An index of dummy nodes is kept. These dummy nodes are the hash
  *   table "buckets", and they are also chained together in the
  *   split-ordered list, which allows recursive expansion.
- * - The resize operation only allows expanding the hash table.
- *   It is triggered either through an API call or automatically by
- *   detecting long chains in the add operation.
+ * - The resize operation for small tables only allows expanding the hash table.
+ *   It is triggered automatically by detecting long chains in the add
+ *   operation.
+ * - The resize operation for larger tables (and available through an
+ *   API) allows both expanding and shrinking the hash table.
+ * - Per-CPU Split-counters are used to keep track of the number of
+ *   nodes within the hash table for automatic resize triggering.
  * - Resize operation initiated by long chain detection is executed by a
  *   call_rcu thread, which keeps lock-freedom of add and remove.
  * - Resize operations are protected by a mutex.
@@ -87,11 +91,61 @@
  *   the "dummy node" tables.
  * - There is one dummy node table per hash index order. The size of
  *   each dummy node table is half the number of hashes contained in
- *   this order.
- * - call_rcu is used to garbage-collect the old order table.
+ *   this order (except for order 0).
+ * - synchronzie_rcu is used to garbage-collect the old dummy node table.
  * - The per-order dummy node tables contain a compact version of the
  *   hash table nodes. These tables are invariant after they are
  *   populated into the hash table.
+ *
+ * Dummy node tables:
+ *
+ * hash table	hash table	the last	all dummy node tables
+ * order	size		dummy node	0   1   2   3   4   5   6(index)
+ * 				table size
+ * 0		1		1		1
+ * 1		2		1		1   1
+ * 2		4		2		1   1   2
+ * 3		8		4		1   1   2   4
+ * 4		16		8		1   1   2   4   8
+ * 5		32		16		1   1   2   4   8  16
+ * 6		64		32		1   1   2   4   8  16  32
+ *
+ * When growing/shrinking, we only focus on the last dummy node table
+ * which size is (!order ? 1 : (1 << (order -1))).
+ *
+ * Example for growing/shrinking:
+ * grow hash table from order 5 to 6: init the index=6 dummy node table
+ * shrink hash table from order 6 to 5: fini the index=6 dummy node table
+ *
+ * A bit of ascii art explanation:
+ * 
+ * Order index is the off-by-one compare to the actual power of 2 because 
+ * we use index 0 to deal with the 0 special-case.
+ * 
+ * This shows the nodes for a small table ordered by reversed bits:
+ * 
+ *    bits   reverse
+ * 0  000        000
+ * 4  100        001
+ * 2  010        010
+ * 6  110        011
+ * 1  001        100
+ * 5  101        101
+ * 3  011        110
+ * 7  111        111
+ * 
+ * This shows the nodes in order of non-reversed bits, linked by 
+ * reversed-bit order.
+ * 
+ * order              bits       reverse
+ * 0               0  000        000
+ * 1               |  1  001        100             <-
+ * 2               |  |  2  010        010    <-     |
+ *                 |  |  |  3  011        110  | <-  |
+ * 3               -> |  |  |  4  100        001  |  |
+ *                    -> |  |     5  101        101  |
+ *                       -> |        6  110        011
+ *                          ->          7  111        111
  */
 
 #define _LGPL_SOURCE
@@ -102,11 +156,11 @@
 #include <stdint.h>
 #include <string.h>
 
+#include "config.h"
 #include <urcu.h>
 #include <urcu-call-rcu.h>
 #include <urcu/arch.h>
 #include <urcu/uatomic.h>
-#include <urcu/jhash.h>
 #include <urcu/compiler.h>
 #include <urcu/rculfhash.h>
 #include <stdio.h>
@@ -118,8 +172,36 @@
 #define dbg_printf(fmt, args...)
 #endif
 
-#define CHAIN_LEN_TARGET		4
-#define CHAIN_LEN_RESIZE_THRESHOLD	8
+/*
+ * Per-CPU split-counters lazily update the global counter each 1024
+ * addition/removal. It automatically keeps track of resize required.
+ * We use the bucket length as indicator for need to expand for small
+ * tables and machines lacking per-cpu data suppport.
+ */
+#define COUNT_COMMIT_ORDER		10
+#define CHAIN_LEN_TARGET		1
+#define CHAIN_LEN_RESIZE_THRESHOLD	3
+
+/*
+ * Define the minimum table size.
+ */
+#define MIN_TABLE_SIZE			1
+
+#if (CAA_BITS_PER_LONG == 32)
+#define MAX_TABLE_ORDER			32
+#else
+#define MAX_TABLE_ORDER			64
+#endif
+
+/*
+ * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
+ */
+#define MIN_PARTITION_PER_THREAD_ORDER	12
+#define MIN_PARTITION_PER_THREAD	(1UL << MIN_PARTITION_PER_THREAD_ORDER)
+
+#ifndef min
+#define min(a, b)	((a) < (b) ? (a) : (b))
+#endif
 
 #ifndef max
 #define max(a, b)	((a) > (b) ? (a) : (b))
@@ -127,6 +209,8 @@
 
 /*
  * The removed flag needs to be updated atomically with the pointer.
+ * It indicates that no node must attach to the node scheduled for
+ * removal, and that node garbage collection must be performed.
  * The dummy flag does not require to be updated atomically with the
  * pointer, but it is added as a pointer low bit flag to save space.
  */
@@ -134,23 +218,52 @@
 #define DUMMY_FLAG		(1UL << 1)
 #define FLAGS_MASK		((1UL << 2) - 1)
 
+/* Value of the end pointer. Should not interact with flags. */
+#define END_VALUE		NULL
+
+struct ht_items_count {
+	unsigned long add, del;
+} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+
+struct rcu_level {
+	/* Note: manually update allocation length when adding a field */
+	struct _cds_lfht_node nodes[0];
+};
+
 struct rcu_table {
-	unsigned long size;	/* always a power of 2 */
+	unsigned long size;	/* always a power of 2, shared (RCU) */
 	unsigned long resize_target;
 	int resize_initiated;
-	struct rcu_head head;
-	struct _cds_lfht_node *tbl[0];
+	struct rcu_level *tbl[MAX_TABLE_ORDER];
 };
 
 struct cds_lfht {
-	struct rcu_table *t;		/* shared */
+	struct rcu_table t;
 	cds_lfht_hash_fct hash_fct;
 	cds_lfht_compare_fct compare_fct;
 	unsigned long hash_seed;
+	int flags;
+	/*
+	 * We need to put the work threads offline (QSBR) when taking this
+	 * mutex, because we use synchronize_rcu within this mutex critical
+	 * section, which waits on read-side critical sections, and could
+	 * therefore cause grace-period deadlock if we hold off RCU G.P.
+	 * completion.
+	 */
 	pthread_mutex_t resize_mutex;	/* resize mutex: add/del mutex */
 	unsigned int in_progress_resize, in_progress_destroy;
 	void (*cds_lfht_call_rcu)(struct rcu_head *head,
 		      void (*func)(struct rcu_head *head));
+	void (*cds_lfht_synchronize_rcu)(void);
+	void (*cds_lfht_rcu_read_lock)(void);
+	void (*cds_lfht_rcu_read_unlock)(void);
+	void (*cds_lfht_rcu_thread_offline)(void);
+	void (*cds_lfht_rcu_thread_online)(void);
+	void (*cds_lfht_rcu_register_thread)(void);
+	void (*cds_lfht_rcu_unregister_thread)(void);
+	pthread_attr_t *resize_attr;	/* Resize threads attributes */
+	long count;			/* global approximate item count */
+	struct ht_items_count *percpu_count;	/* per-cpu item count */
 };
 
 struct rcu_resize_work {
@@ -158,6 +271,21 @@ struct rcu_resize_work {
 	struct cds_lfht *ht;
 };
 
+struct partition_resize_work {
+	pthread_t thread_id;
+	struct cds_lfht *ht;
+	unsigned long i, start, len;
+	void (*fct)(struct cds_lfht *ht, unsigned long i,
+		    unsigned long start, unsigned long len);
+};
+
+static
+void _cds_lfht_add(struct cds_lfht *ht,
+		unsigned long size,
+		struct cds_lfht_node *node,
+		struct cds_lfht_iter *unique_ret,
+		int dummy);
+
 /*
  * Algorithm to reverse bits in a word by lookup table, extended to
  * 64-bit words.
@@ -330,38 +458,217 @@ unsigned int fls_ulong(unsigned long x)
 #endif
 }
 
+/*
+ * Return the minimum order for which x <= (1UL << order).
+ * Return -1 if x is 0.
+ */
 int get_count_order_u32(uint32_t x)
 {
-	int order;
+	if (!x)
+		return -1;
 
-	order = fls_u32(x) - 1;
-	if (x & (x - 1))
-		order++;
-	return order;
+	return fls_u32(x - 1);
 }
 
+/*
+ * Return the minimum order for which x <= (1UL << order).
+ * Return -1 if x is 0.
+ */
 int get_count_order_ulong(unsigned long x)
 {
-	int order;
+	if (!x)
+		return -1;
 
-	order = fls_ulong(x) - 1;
-	if (x & (x - 1))
-		order++;
-	return order;
+	return fls_ulong(x - 1);
 }
 
+#ifdef POISON_FREE
+#define poison_free(ptr)				\
+	do {						\
+		memset(ptr, 0x42, sizeof(*(ptr)));	\
+		free(ptr);				\
+	} while (0)
+#else
+#define poison_free(ptr)	free(ptr)
+#endif
+
+static
+void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
+
+/*
+ * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
+ * available, then we support hash table item accounting.
+ * In the unfortunate event the number of CPUs reported would be
+ * inaccurate, we use modulo arithmetic on the number of CPUs we got.
+ */
+#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+
 static
-void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth);
+void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
+				unsigned long count);
+
+static long nr_cpus_mask = -1;
+
+static
+struct ht_items_count *alloc_per_cpu_items_count(void)
+{
+	struct ht_items_count *count;
+
+	switch (nr_cpus_mask) {
+	case -2:
+		return NULL;
+	case -1:
+	{
+		long maxcpus;
+
+		maxcpus = sysconf(_SC_NPROCESSORS_CONF);
+		if (maxcpus <= 0) {
+			nr_cpus_mask = -2;
+			return NULL;
+		}
+		/*
+		 * round up number of CPUs to next power of two, so we
+		 * can use & for modulo.
+		 */
+		maxcpus = 1UL << get_count_order_ulong(maxcpus);
+		nr_cpus_mask = maxcpus - 1;
+	}
+		/* Fall-through */
+	default:
+		return calloc(nr_cpus_mask + 1, sizeof(*count));
+	}
+}
+
+static
+void free_per_cpu_items_count(struct ht_items_count *count)
+{
+	poison_free(count);
+}
+
+static
+int ht_get_cpu(void)
+{
+	int cpu;
+
+	assert(nr_cpus_mask >= 0);
+	cpu = sched_getcpu();
+	if (unlikely(cpu < 0))
+		return cpu;
+	else
+		return cpu & nr_cpus_mask;
+}
+
+static
+void ht_count_add(struct cds_lfht *ht, unsigned long size)
+{
+	unsigned long percpu_count;
+	int cpu;
+
+	if (unlikely(!ht->percpu_count))
+		return;
+	cpu = ht_get_cpu();
+	if (unlikely(cpu < 0))
+		return;
+	percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
+	if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+		long count;
+
+		dbg_printf("add percpu %lu\n", percpu_count);
+		count = uatomic_add_return(&ht->count,
+					   1UL << COUNT_COMMIT_ORDER);
+		/* If power of 2 */
+		if (!(count & (count - 1))) {
+			if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
+				return;
+			dbg_printf("add set global %ld\n", count);
+			cds_lfht_resize_lazy_count(ht, size,
+				count >> (CHAIN_LEN_TARGET - 1));
+		}
+	}
+}
+
+static
+void ht_count_del(struct cds_lfht *ht, unsigned long size)
+{
+	unsigned long percpu_count;
+	int cpu;
+
+	if (unlikely(!ht->percpu_count))
+		return;
+	cpu = ht_get_cpu();
+	if (unlikely(cpu < 0))
+		return;
+	percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
+	if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+		long count;
+
+		dbg_printf("del percpu %lu\n", percpu_count);
+		count = uatomic_add_return(&ht->count,
+					   -(1UL << COUNT_COMMIT_ORDER));
+		/* If power of 2 */
+		if (!(count & (count - 1))) {
+			if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
+				return;
+			dbg_printf("del set global %ld\n", count);
+			/*
+			 * Don't shrink table if the number of nodes is below a
+			 * certain threshold.
+			 */
+			if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
+				return;
+			cds_lfht_resize_lazy_count(ht, size,
+				count >> (CHAIN_LEN_TARGET - 1));
+		}
+	}
+}
+
+#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+
+static const long nr_cpus_mask = -2;
+
+static
+struct ht_items_count *alloc_per_cpu_items_count(void)
+{
+	return NULL;
+}
+
+static
+void free_per_cpu_items_count(struct ht_items_count *count)
+{
+}
+
+static
+void ht_count_add(struct cds_lfht *ht, unsigned long size)
+{
+}
+
+static
+void ht_count_del(struct cds_lfht *ht, unsigned long size)
+{
+}
+
+#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+
 
 static
-void check_resize(struct cds_lfht *ht, struct rcu_table *t,
-		  uint32_t chain_len)
+void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
 {
+	unsigned long count;
+
+	if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
+		return;
+	count = uatomic_read(&ht->count);
+	/*
+	 * Use bucket-local length for small table expand and for
+	 * environments lacking per-cpu data support.
+	 */
+	if (count >= (1UL << COUNT_COMMIT_ORDER))
+		return;
 	if (chain_len > 100)
 		dbg_printf("WARNING: large chain length: %u.\n",
 			   chain_len);
 	if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
-		cds_lfht_resize_lazy(ht, t,
+		cds_lfht_resize_lazy(ht, size,
 			get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
 }
 
@@ -394,7 +701,19 @@ struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
 {
 	return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
 }
- 
+
+static
+struct cds_lfht_node *get_end(void)
+{
+	return (struct cds_lfht_node *) END_VALUE;
+}
+
+static
+int is_end(struct cds_lfht_node *node)
+{
+	return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
+}
+
 static
 unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
 {
@@ -409,6 +728,27 @@ unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
 	return v;
 }
 
+static
+struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
+		unsigned long hash)
+{
+	unsigned long index, order;
+
+	assert(size > 0);
+	index = hash & (size - 1);
+	/*
+	 * equivalent to get_count_order_ulong(index + 1), but optimizes
+	 * away the non-existing 0 special-case for
+	 * get_count_order_ulong.
+	 */
+	order = fls_ulong(index);
+
+	dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
+		   hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
+
+	return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
+}
+
 /*
  * Remove all logically deleted nodes from a bucket up to a certain node key.
  */
@@ -417,13 +757,25 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node
 {
 	struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
 
+	assert(!is_dummy(dummy));
+	assert(!is_removed(dummy));
+	assert(!is_dummy(node));
+	assert(!is_removed(node));
 	for (;;) {
 		iter_prev = dummy;
 		/* We can always skip the dummy node initially */
 		iter = rcu_dereference(iter_prev->p.next);
+		assert(!is_removed(iter));
 		assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
+		/*
+		 * We should never be called with dummy (start of chain)
+		 * and logically removed node (end of path compression
+		 * marker) being the actual same node. This would be a
+		 * bug in the algorithm implementation.
+		 */
+		assert(dummy != node);
 		for (;;) {
-			if (unlikely(!clear_flag(iter)))
+			if (unlikely(is_end(iter)))
 				return;
 			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
 				return;
@@ -440,23 +792,86 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node
 			new_next = clear_flag(next);
 		(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
 	}
+	return;
 }
 
 static
-struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t,
-				struct cds_lfht_node *node, int unique, int dummy)
+int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
+		struct cds_lfht_node *old_node,
+		struct cds_lfht_node *old_next,
+		struct cds_lfht_node *new_node)
 {
-	struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
-			*dummy_node;
+	struct cds_lfht_node *dummy, *ret_next;
 	struct _cds_lfht_node *lookup;
-	unsigned long hash, index, order;
 
-	if (!t->size) {
-		assert(dummy);
-		node->p.next = flag_dummy(NULL);
-		return node;	/* Initial first add (head) */
+	if (!old_node)	/* Return -ENOENT if asked to replace NULL node */
+		return -ENOENT;
+
+	assert(!is_removed(old_node));
+	assert(!is_dummy(old_node));
+	assert(!is_removed(new_node));
+	assert(!is_dummy(new_node));
+	assert(new_node != old_node);
+	for (;;) {
+		/* Insert after node to be replaced */
+		if (is_removed(old_next)) {
+			/*
+			 * Too late, the old node has been removed under us
+			 * between lookup and replace. Fail.
+			 */
+			return -ENOENT;
+		}
+		assert(!is_dummy(old_next));
+		assert(new_node != clear_flag(old_next));
+		new_node->p.next = clear_flag(old_next);
+		/*
+		 * Here is the whole trick for lock-free replace: we add
+		 * the replacement node _after_ the node we want to
+		 * replace by atomically setting its next pointer at the
+		 * same time we set its removal flag. Given that
+		 * the lookups/get next use an iterator aware of the
+		 * next pointer, they will either skip the old node due
+		 * to the removal flag and see the new node, or use
+		 * the old node, but will not see the new one.
+		 */
+		ret_next = uatomic_cmpxchg(&old_node->p.next,
+			      old_next, flag_removed(new_node));
+		if (ret_next == old_next)
+			break;		/* We performed the replacement. */
+		old_next = ret_next;
 	}
-	hash = bit_reverse_ulong(node->p.reverse_hash);
+
+	/*
+	 * Ensure that the old node is not visible to readers anymore:
+	 * lookup for the node, and remove it (along with any other
+	 * logically removed node) if found.
+	 */
+	lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
+	dummy = (struct cds_lfht_node *) lookup;
+	_cds_lfht_gc_bucket(dummy, new_node);
+
+	assert(is_removed(rcu_dereference(old_node->p.next)));
+	return 0;
+}
+
+/*
+ * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
+ * mode. A NULL unique_ret allows creation of duplicate keys.
+ */
+static
+void _cds_lfht_add(struct cds_lfht *ht,
+		unsigned long size,
+		struct cds_lfht_node *node,
+		struct cds_lfht_iter *unique_ret,
+		int dummy)
+{
+	struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
+			*return_node;
+	struct _cds_lfht_node *lookup;
+
+	assert(!is_dummy(node));
+	assert(!is_removed(node));
+	lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
 	for (;;) {
 		uint32_t chain_len = 0;
 
@@ -464,37 +879,59 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t,
 		 * iter_prev points to the non-removed node prior to the
 		 * insert location.
 		 */
-		index = hash & (t->size - 1);
-		order = get_count_order_ulong(index + 1);
-		lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
 		iter_prev = (struct cds_lfht_node *) lookup;
 		/* We can always skip the dummy node initially */
 		iter = rcu_dereference(iter_prev->p.next);
 		assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
 		for (;;) {
-			if (unlikely(!clear_flag(iter)))
+			if (unlikely(is_end(iter)))
 				goto insert;
 			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
 				goto insert;
+
+			/* dummy node is the first node of the identical-hash-value chain */
+			if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
+				goto insert;
+
 			next = rcu_dereference(clear_flag(iter)->p.next);
 			if (unlikely(is_removed(next)))
 				goto gc_node;
-			if (unique
+
+			/* uniquely add */
+			if (unique_ret
 			    && !is_dummy(next)
-			    && !ht->compare_fct(node->key, node->key_len,
-						clear_flag(iter)->key,
-						clear_flag(iter)->key_len))
-				return clear_flag(iter);
+			    && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) {
+				struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
+
+				/*
+				 * uniquely adding inserts the node as the first
+				 * node of the identical-hash-value node chain.
+				 *
+				 * This semantic ensures no duplicated keys
+				 * should ever be observable in the table
+				 * (including observe one node by one node
+				 * by forward iterations)
+				 */
+				cds_lfht_next_duplicate(ht, &d_iter);
+				if (!d_iter.node)
+					goto insert;
+
+				*unique_ret = d_iter;
+				return;
+			}
+
 			/* Only account for identical reverse hash once */
 			if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
 			    && !is_dummy(next))
-				check_resize(ht, t, ++chain_len);
+				check_resize(ht, size, ++chain_len);
 			iter_prev = clear_flag(iter);
 			iter = next;
 		}
+
 	insert:
 		assert(node != clear_flag(iter));
 		assert(!is_removed(iter_prev));
+		assert(!is_removed(iter));
 		assert(iter_prev != node);
 		if (!dummy)
 			node->p.next = clear_flag(iter);
@@ -505,10 +942,13 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t,
 		else
 			new_node = node;
 		if (uatomic_cmpxchg(&iter_prev->p.next, iter,
-				    new_node) != iter)
+				    new_node) != iter) {
 			continue;	/* retry */
-		else
-			goto gc_end;
+		} else {
+			return_node = node;
+			goto end;
+		}
+
 	gc_node:
 		assert(!is_removed(iter));
 		if (is_dummy(iter))
@@ -518,226 +958,611 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t,
 		(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
 		/* retry */
 	}
-gc_end:
-	/* Garbage collect logically removed nodes in the bucket */
-	index = hash & (t->size - 1);
-	order = get_count_order_ulong(index + 1);
-	lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
-	dummy_node = (struct cds_lfht_node *) lookup;
-	_cds_lfht_gc_bucket(dummy_node, node);
-	return node;
+end:
+	if (unique_ret) {
+		unique_ret->node = return_node;
+		/* unique_ret->next left unset, never used. */
+	}
 }
 
 static
-int _cds_lfht_remove(struct cds_lfht *ht, struct rcu_table *t,
-		struct cds_lfht_node *node)
+int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
+		struct cds_lfht_node *node,
+		int dummy_removal)
 {
 	struct cds_lfht_node *dummy, *next, *old;
 	struct _cds_lfht_node *lookup;
-	int flagged = 0;
-	unsigned long hash, index, order;
+
+	if (!node)	/* Return -ENOENT if asked to delete NULL node */
+		return -ENOENT;
 
 	/* logically delete the node */
+	assert(!is_dummy(node));
+	assert(!is_removed(node));
 	old = rcu_dereference(node->p.next);
 	do {
+		struct cds_lfht_node *new_next;
+
 		next = old;
 		if (unlikely(is_removed(next)))
-			goto end;
-		assert(!is_dummy(next));
-		old = uatomic_cmpxchg(&node->p.next, next,
-				      flag_removed(next));
+			return -ENOENT;
+		if (dummy_removal)
+			assert(is_dummy(next));
+		else
+			assert(!is_dummy(next));
+		new_next = flag_removed(next);
+		old = uatomic_cmpxchg(&node->p.next, next, new_next);
 	} while (old != next);
-
 	/* We performed the (logical) deletion. */
-	flagged = 1;
 
 	/*
 	 * Ensure that the node is not visible to readers anymore: lookup for
 	 * the node, and remove it (along with any other logically removed node)
 	 * if found.
 	 */
-	hash = bit_reverse_ulong(node->p.reverse_hash);
-	index = hash & (t->size - 1);
-	order = get_count_order_ulong(index + 1);
-	lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
+	lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
 	dummy = (struct cds_lfht_node *) lookup;
 	_cds_lfht_gc_bucket(dummy, node);
-end:
+
+	assert(is_removed(rcu_dereference(node->p.next)));
+	return 0;
+}
+
+static
+void *partition_resize_thread(void *arg)
+{
+	struct partition_resize_work *work = arg;
+
+	work->ht->cds_lfht_rcu_register_thread();
+	work->fct(work->ht, work->i, work->start, work->len);
+	work->ht->cds_lfht_rcu_unregister_thread();
+	return NULL;
+}
+
+static
+void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
+		unsigned long len,
+		void (*fct)(struct cds_lfht *ht, unsigned long i,
+			unsigned long start, unsigned long len))
+{
+	unsigned long partition_len;
+	struct partition_resize_work *work;
+	int thread, ret;
+	unsigned long nr_threads;
+
 	/*
-	 * Only the flagging action indicated that we (and no other)
-	 * removed the node from the hash.
+	 * Note: nr_cpus_mask + 1 is always power of 2.
+	 * We spawn just the number of threads we need to satisfy the minimum
+	 * partition size, up to the number of CPUs in the system.
 	 */
-	if (flagged) {
-		assert(is_removed(rcu_dereference(node->p.next)));
-		return 0;
-	} else
-		return -ENOENT;
+	if (nr_cpus_mask > 0) {
+		nr_threads = min(nr_cpus_mask + 1,
+				 len >> MIN_PARTITION_PER_THREAD_ORDER);
+	} else {
+		nr_threads = 1;
+	}
+	partition_len = len >> get_count_order_ulong(nr_threads);
+	work = calloc(nr_threads, sizeof(*work));
+	assert(work);
+	for (thread = 0; thread < nr_threads; thread++) {
+		work[thread].ht = ht;
+		work[thread].i = i;
+		work[thread].len = partition_len;
+		work[thread].start = thread * partition_len;
+		work[thread].fct = fct;
+		ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
+			partition_resize_thread, &work[thread]);
+		assert(!ret);
+	}
+	for (thread = 0; thread < nr_threads; thread++) {
+		ret = pthread_join(work[thread].thread_id, NULL);
+		assert(!ret);
+	}
+	free(work);
+}
+
+/*
+ * Holding RCU read lock to protect _cds_lfht_add against memory
+ * reclaim that could be performed by other call_rcu worker threads (ABA
+ * problem).
+ *
+ * When we reach a certain length, we can split this population phase over
+ * many worker threads, based on the number of CPUs available in the system.
+ * This should therefore take care of not having the expand lagging behind too
+ * many concurrent insertion threads by using the scheduler's ability to
+ * schedule dummy node population fairly with insertions.
+ */
+static
+void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
+				   unsigned long start, unsigned long len)
+{
+	unsigned long j;
+
+	assert(i > 0);
+	ht->cds_lfht_rcu_read_lock();
+	for (j = start; j < start + len; j++) {
+		struct cds_lfht_node *new_node =
+			(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
+
+		dbg_printf("init populate: i %lu j %lu hash %lu\n",
+			   i, j, (1UL << (i - 1)) + j);
+		new_node->p.reverse_hash =
+				bit_reverse_ulong((1UL << (i - 1)) + j);
+		_cds_lfht_add(ht, 1UL << (i - 1),
+				new_node, NULL, 1);
+	}
+	ht->cds_lfht_rcu_read_unlock();
 }
 
 static
-void init_table(struct cds_lfht *ht, struct rcu_table *t,
-		unsigned long first_order, unsigned long len_order)
+void init_table_populate(struct cds_lfht *ht, unsigned long i,
+			 unsigned long len)
 {
-	unsigned long i, end_order;
+	assert(nr_cpus_mask != -1);
+	if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
+		ht->cds_lfht_rcu_thread_online();
+		init_table_populate_partition(ht, i, 0, len);
+		ht->cds_lfht_rcu_thread_offline();
+		return;
+	}
+	partition_resize_helper(ht, i, len, init_table_populate_partition);
+}
 
-	dbg_printf("init table: first_order %lu end_order %lu\n",
-		   first_order, first_order + len_order);
-	end_order = first_order + len_order;
-	t->size = !first_order ? 0 : (1UL << (first_order - 1));
-	for (i = first_order; i < end_order; i++) {
-		unsigned long j, len;
+static
+void init_table(struct cds_lfht *ht,
+		unsigned long first_order, unsigned long last_order)
+{
+	unsigned long i;
+
+	dbg_printf("init table: first_order %lu last_order %lu\n",
+		   first_order, last_order);
+	assert(first_order > 0);
+	for (i = first_order; i <= last_order; i++) {
+		unsigned long len;
 
-		len = !i ? 1 : 1UL << (i - 1);
+		len = 1UL << (i - 1);
 		dbg_printf("init order %lu len: %lu\n", i, len);
-		t->tbl[i] = calloc(len, sizeof(struct _cds_lfht_node));
-		for (j = 0; j < len; j++) {
-			dbg_printf("init entry: i %lu j %lu hash %lu\n",
-				   i, j, !i ? 0 : (1UL << (i - 1)) + j);
-			struct cds_lfht_node *new_node =
-				(struct cds_lfht_node *) &t->tbl[i][j];
-			new_node->p.reverse_hash =
-				bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
-			(void) _cds_lfht_add(ht, t, new_node, 0, 1);
-			if (CMM_LOAD_SHARED(ht->in_progress_destroy))
-				break;
-		}
-		/* Update table size */
-		t->size = !i ? 1 : (1UL << i);
-		dbg_printf("init new size: %lu\n", t->size);
+
+		/* Stop expand if the resize target changes under us */
+		if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
+			break;
+
+		ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
+		assert(ht->t.tbl[i]);
+
+		/*
+		 * Set all dummy nodes reverse hash values for a level and
+		 * link all dummy nodes into the table.
+		 */
+		init_table_populate(ht, i, len);
+
+		/*
+		 * Update table size.
+		 */
+		cmm_smp_wmb();	/* populate data before RCU size */
+		CMM_STORE_SHARED(ht->t.size, 1UL << i);
+
+		dbg_printf("init new size: %lu\n", 1UL << i);
 		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
 			break;
 	}
-	t->resize_target = t->size;
-	t->resize_initiated = 0;
 }
 
-struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
+/*
+ * Holding RCU read lock to protect _cds_lfht_remove against memory
+ * reclaim that could be performed by other call_rcu worker threads (ABA
+ * problem).
+ * For a single level, we logically remove and garbage collect each node.
+ *
+ * As a design choice, we perform logical removal and garbage collection on a
+ * node-per-node basis to simplify this algorithm. We also assume keeping good
+ * cache locality of the operation would overweight possible performance gain
+ * that could be achieved by batching garbage collection for multiple levels.
+ * However, this would have to be justified by benchmarks.
+ *
+ * Concurrent removal and add operations are helping us perform garbage
+ * collection of logically removed nodes. We guarantee that all logically
+ * removed nodes have been garbage-collected (unlinked) before call_rcu is
+ * invoked to free a hole level of dummy nodes (after a grace period).
+ *
+ * Logical removal and garbage collection can therefore be done in batch or on a
+ * node-per-node basis, as long as the guarantee above holds.
+ *
+ * When we reach a certain length, we can split this removal over many worker
+ * threads, based on the number of CPUs available in the system. This should
+ * take care of not letting resize process lag behind too many concurrent
+ * updater threads actively inserting into the hash table.
+ */
+static
+void remove_table_partition(struct cds_lfht *ht, unsigned long i,
+			    unsigned long start, unsigned long len)
+{
+	unsigned long j;
+
+	assert(i > 0);
+	ht->cds_lfht_rcu_read_lock();
+	for (j = start; j < start + len; j++) {
+		struct cds_lfht_node *fini_node =
+			(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
+
+		dbg_printf("remove entry: i %lu j %lu hash %lu\n",
+			   i, j, (1UL << (i - 1)) + j);
+		fini_node->p.reverse_hash =
+			bit_reverse_ulong((1UL << (i - 1)) + j);
+		(void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
+	}
+	ht->cds_lfht_rcu_read_unlock();
+}
+
+static
+void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
+{
+
+	assert(nr_cpus_mask != -1);
+	if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
+		ht->cds_lfht_rcu_thread_online();
+		remove_table_partition(ht, i, 0, len);
+		ht->cds_lfht_rcu_thread_offline();
+		return;
+	}
+	partition_resize_helper(ht, i, len, remove_table_partition);
+}
+
+static
+void fini_table(struct cds_lfht *ht,
+		unsigned long first_order, unsigned long last_order)
+{
+	long i;
+	void *free_by_rcu = NULL;
+
+	dbg_printf("fini table: first_order %lu last_order %lu\n",
+		   first_order, last_order);
+	assert(first_order > 0);
+	for (i = last_order; i >= first_order; i--) {
+		unsigned long len;
+
+		len = 1UL << (i - 1);
+		dbg_printf("fini order %lu len: %lu\n", i, len);
+
+		/* Stop shrink if the resize target changes under us */
+		if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
+			break;
+
+		cmm_smp_wmb();	/* populate data before RCU size */
+		CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
+
+		/*
+		 * We need to wait for all add operations to reach Q.S. (and
+		 * thus use the new table for lookups) before we can start
+		 * releasing the old dummy nodes. Otherwise their lookup will
+		 * return a logically removed node as insert position.
+		 */
+		ht->cds_lfht_synchronize_rcu();
+		if (free_by_rcu)
+			free(free_by_rcu);
+
+		/*
+		 * Set "removed" flag in dummy nodes about to be removed.
+		 * Unlink all now-logically-removed dummy node pointers.
+		 * Concurrent add/remove operation are helping us doing
+		 * the gc.
+		 */
+		remove_table(ht, i, len);
+
+		free_by_rcu = ht->t.tbl[i];
+
+		dbg_printf("fini new size: %lu\n", 1UL << i);
+		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
+			break;
+	}
+
+	if (free_by_rcu) {
+		ht->cds_lfht_synchronize_rcu();
+		free(free_by_rcu);
+	}
+}
+
+static
+void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
+{
+	struct _cds_lfht_node *prev, *node;
+	unsigned long order, len, i, j;
+
+	ht->t.tbl[0] = calloc(1, sizeof(struct _cds_lfht_node));
+	assert(ht->t.tbl[0]);
+
+	dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
+	ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end());
+	ht->t.tbl[0]->nodes[0].reverse_hash = 0;
+
+	for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
+		len = 1UL << (order - 1);
+		ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
+		assert(ht->t.tbl[order]);
+
+		i = 0;
+		prev = ht->t.tbl[i]->nodes;
+		for (j = 0; j < len; j++) {
+			if (j & (j - 1)) {	/* Between power of 2 */
+				prev++;
+			} else if (j) {		/* At each power of 2 */
+				i++;
+				prev = ht->t.tbl[i]->nodes;
+			}
+
+			node = &ht->t.tbl[order]->nodes[j];
+			dbg_printf("create dummy: order %lu index %lu hash %lu\n",
+				   order, j, j + len);
+			node->next = prev->next;
+			assert(is_dummy(node->next));
+			node->reverse_hash = bit_reverse_ulong(j + len);
+			prev->next = flag_dummy((struct cds_lfht_node *)node);
+		}
+	}
+}
+
+struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
 			cds_lfht_compare_fct compare_fct,
 			unsigned long hash_seed,
 			unsigned long init_size,
+			int flags,
 			void (*cds_lfht_call_rcu)(struct rcu_head *head,
-					void (*func)(struct rcu_head *head)))
+					void (*func)(struct rcu_head *head)),
+			void (*cds_lfht_synchronize_rcu)(void),
+			void (*cds_lfht_rcu_read_lock)(void),
+			void (*cds_lfht_rcu_read_unlock)(void),
+			void (*cds_lfht_rcu_thread_offline)(void),
+			void (*cds_lfht_rcu_thread_online)(void),
+			void (*cds_lfht_rcu_register_thread)(void),
+			void (*cds_lfht_rcu_unregister_thread)(void),
+			pthread_attr_t *attr)
 {
 	struct cds_lfht *ht;
 	unsigned long order;
 
+	/* init_size must be power of two */
+	if (init_size && (init_size & (init_size - 1)))
+		return NULL;
 	ht = calloc(1, sizeof(struct cds_lfht));
+	assert(ht);
 	ht->hash_fct = hash_fct;
 	ht->compare_fct = compare_fct;
 	ht->hash_seed = hash_seed;
 	ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
-	ht->in_progress_resize = 0;
+	ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
+	ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
+	ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
+	ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
+	ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
+	ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
+	ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
+	ht->resize_attr = attr;
+	ht->percpu_count = alloc_per_cpu_items_count();
 	/* this mutex should not nest in read-side C.S. */
 	pthread_mutex_init(&ht->resize_mutex, NULL);
-	order = get_count_order_ulong(max(init_size, 1)) + 1;
-	ht->t = calloc(1, sizeof(struct cds_lfht)
-		       + (order * sizeof(struct _cds_lfht_node *)));
-	ht->t->size = 0;
-	pthread_mutex_lock(&ht->resize_mutex);
-	init_table(ht, ht->t, 0, order);
-	pthread_mutex_unlock(&ht->resize_mutex);
+	ht->flags = flags;
+	order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE));
+	ht->t.resize_target = 1UL << order;
+	cds_lfht_create_dummy(ht, 1UL << order);
+	ht->t.size = 1UL << order;
 	return ht;
 }
 
-struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
+void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
+		struct cds_lfht_iter *iter)
 {
-	struct rcu_table *t;
-	struct cds_lfht_node *node, *next;
+	struct cds_lfht_node *node, *next, *dummy_node;
 	struct _cds_lfht_node *lookup;
-	unsigned long hash, reverse_hash, index, order;
+	unsigned long hash, reverse_hash, size;
 
 	hash = ht->hash_fct(key, key_len, ht->hash_seed);
 	reverse_hash = bit_reverse_ulong(hash);
 
-	t = rcu_dereference(ht->t);
-	index = hash & (t->size - 1);
-	order = get_count_order_ulong(index + 1);
-	lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
-	dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
-		   hash, index, order, index & ((1UL << (order - 1)) - 1));
-	node = (struct cds_lfht_node *) lookup;
+	size = rcu_dereference(ht->t.size);
+	lookup = lookup_bucket(ht, size, hash);
+	dummy_node = (struct cds_lfht_node *) lookup;
+	/* We can always skip the dummy node initially */
+	node = rcu_dereference(dummy_node->p.next);
+	node = clear_flag(node);
 	for (;;) {
-		if (unlikely(!node))
+		if (unlikely(is_end(node))) {
+			node = next = NULL;
 			break;
+		}
 		if (unlikely(node->p.reverse_hash > reverse_hash)) {
-			node = NULL;
+			node = next = NULL;
 			break;
 		}
 		next = rcu_dereference(node->p.next);
 		if (likely(!is_removed(next))
 		    && !is_dummy(next)
+		    && clear_flag(node)->p.reverse_hash == reverse_hash
 		    && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
 				break;
 		}
 		node = clear_flag(next);
 	}
 	assert(!node || !is_dummy(rcu_dereference(node->p.next)));
-	return node;
+	iter->node = node;
+	iter->next = next;
+}
+
+void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+{
+	struct cds_lfht_node *node, *next;
+	unsigned long reverse_hash;
+	void *key;
+	size_t key_len;
+
+	node = iter->node;
+	reverse_hash = node->p.reverse_hash;
+	key = node->key;
+	key_len = node->key_len;
+	next = iter->next;
+	node = clear_flag(next);
+
+	for (;;) {
+		if (unlikely(is_end(node))) {
+			node = next = NULL;
+			break;
+		}
+		if (unlikely(node->p.reverse_hash > reverse_hash)) {
+			node = next = NULL;
+			break;
+		}
+		next = rcu_dereference(node->p.next);
+		if (likely(!is_removed(next))
+		    && !is_dummy(next)
+		    && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
+				break;
+		}
+		node = clear_flag(next);
+	}
+	assert(!node || !is_dummy(rcu_dereference(node->p.next)));
+	iter->node = node;
+	iter->next = next;
+}
+
+void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+{
+	struct cds_lfht_node *node, *next;
+
+	node = clear_flag(iter->next);
+	for (;;) {
+		if (unlikely(is_end(node))) {
+			node = next = NULL;
+			break;
+		}
+		next = rcu_dereference(node->p.next);
+		if (likely(!is_removed(next))
+		    && !is_dummy(next)) {
+				break;
+		}
+		node = clear_flag(next);
+	}
+	assert(!node || !is_dummy(rcu_dereference(node->p.next)));
+	iter->node = node;
+	iter->next = next;
+}
+
+void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+{
+	struct _cds_lfht_node *lookup;
+
+	/*
+	 * Get next after first dummy node. The first dummy node is the
+	 * first node of the linked list.
+	 */
+	lookup = &ht->t.tbl[0]->nodes[0];
+	iter->next = lookup->next;
+	cds_lfht_next(ht, iter);
 }
 
 void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
 {
-	struct rcu_table *t;
-	unsigned long hash;
+	unsigned long hash, size;
 
 	hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
 	node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
 
-	t = rcu_dereference(ht->t);
-	(void) _cds_lfht_add(ht, t, node, 0, 0);
+	size = rcu_dereference(ht->t.size);
+	_cds_lfht_add(ht, size, node, NULL, 0);
+	ht_count_add(ht, size);
 }
 
 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
-					struct cds_lfht_node *node)
+				struct cds_lfht_node *node)
 {
-	struct rcu_table *t;
-	unsigned long hash;
+	unsigned long hash, size;
+	struct cds_lfht_iter iter;
 
 	hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
 	node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
 
-	t = rcu_dereference(ht->t);
-	return _cds_lfht_add(ht, t, node, 1, 0);
+	size = rcu_dereference(ht->t.size);
+	_cds_lfht_add(ht, size, node, &iter, 0);
+	if (iter.node == node)
+		ht_count_add(ht, size);
+	return iter.node;
 }
 
-int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node)
+struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
+				struct cds_lfht_node *node)
 {
-	struct rcu_table *t;
+	unsigned long hash, size;
+	struct cds_lfht_iter iter;
+
+	hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
+	node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
 
-	t = rcu_dereference(ht->t);
-	return _cds_lfht_remove(ht, t, node);
+	size = rcu_dereference(ht->t.size);
+	for (;;) {
+		_cds_lfht_add(ht, size, node, &iter, 0);
+		if (iter.node == node) {
+			ht_count_add(ht, size);
+			return NULL;
+		}
+
+		if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
+			return iter.node;
+	}
+}
+
+int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
+		struct cds_lfht_node *new_node)
+{
+	unsigned long size;
+
+	size = rcu_dereference(ht->t.size);
+	return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
+			new_node);
+}
+
+int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+{
+	unsigned long size;
+	int ret;
+
+	size = rcu_dereference(ht->t.size);
+	ret = _cds_lfht_del(ht, size, iter->node, 0);
+	if (!ret)
+		ht_count_del(ht, size);
+	return ret;
 }
 
 static
 int cds_lfht_delete_dummy(struct cds_lfht *ht)
 {
-	struct rcu_table *t;
 	struct cds_lfht_node *node;
 	struct _cds_lfht_node *lookup;
-	unsigned long order, i;
+	unsigned long order, i, size;
 
-	t = ht->t;
 	/* Check that the table is empty */
-	lookup = &t->tbl[0][0];
+	lookup = &ht->t.tbl[0]->nodes[0];
 	node = (struct cds_lfht_node *) lookup;
 	do {
 		node = clear_flag(node)->p.next;
 		if (!is_dummy(node))
 			return -EPERM;
 		assert(!is_removed(node));
-	} while (clear_flag(node));
+	} while (!is_end(node));
+	/*
+	 * size accessed without rcu_dereference because hash table is
+	 * being destroyed.
+	 */
+	size = ht->t.size;
 	/* Internal sanity check: all nodes left should be dummy */
-	for (order = 0; order < get_count_order_ulong(t->size) + 1; order++) {
+	for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
 		unsigned long len;
 
 		len = !order ? 1 : 1UL << (order - 1);
 		for (i = 0; i < len; i++) {
 			dbg_printf("delete order %lu i %lu hash %lu\n",
 				order, i,
-				bit_reverse_ulong(t->tbl[order][i].reverse_hash));
-			assert(is_dummy(t->tbl[order][i].next));
+				bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
+			assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
 		}
-		free(t->tbl[order]);
+		poison_free(ht->t.tbl[order]);
 	}
 	return 0;
 }
@@ -746,116 +1571,161 @@ int cds_lfht_delete_dummy(struct cds_lfht *ht)
  * Should only be called when no more concurrent readers nor writers can
  * possibly access the table.
  */
-int cds_lfht_destroy(struct cds_lfht *ht)
+int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
 {
 	int ret;
 
 	/* Wait for in-flight resize operations to complete */
-	CMM_STORE_SHARED(ht->in_progress_destroy, 1);
+	_CMM_STORE_SHARED(ht->in_progress_destroy, 1);
+	cmm_smp_mb();	/* Store destroy before load resize */
 	while (uatomic_read(&ht->in_progress_resize))
 		poll(NULL, 0, 100);	/* wait for 100ms */
 	ret = cds_lfht_delete_dummy(ht);
 	if (ret)
 		return ret;
-	free(ht->t);
-	free(ht);
+	free_per_cpu_items_count(ht->percpu_count);
+	if (attr)
+		*attr = ht->resize_attr;
+	poison_free(ht);
 	return ret;
 }
 
 void cds_lfht_count_nodes(struct cds_lfht *ht,
+		long *approx_before,
 		unsigned long *count,
-		unsigned long *removed)
+		unsigned long *removed,
+		long *approx_after)
 {
-	struct rcu_table *t;
 	struct cds_lfht_node *node, *next;
 	struct _cds_lfht_node *lookup;
 	unsigned long nr_dummy = 0;
 
+	*approx_before = 0;
+	if (nr_cpus_mask >= 0) {
+		int i;
+
+		for (i = 0; i < nr_cpus_mask + 1; i++) {
+			*approx_before += uatomic_read(&ht->percpu_count[i].add);
+			*approx_before -= uatomic_read(&ht->percpu_count[i].del);
+		}
+	}
+
 	*count = 0;
 	*removed = 0;
 
-	t = rcu_dereference(ht->t);
 	/* Count non-dummy nodes in the table */
-	lookup = &t->tbl[0][0];
+	lookup = &ht->t.tbl[0]->nodes[0];
 	node = (struct cds_lfht_node *) lookup;
 	do {
 		next = rcu_dereference(node->p.next);
 		if (is_removed(next)) {
-			assert(!is_dummy(next));
-			(*removed)++;
+			if (!is_dummy(next))
+				(*removed)++;
+			else
+				(nr_dummy)++;
 		} else if (!is_dummy(next))
 			(*count)++;
 		else
 			(nr_dummy)++;
 		node = clear_flag(next);
-	} while (node);
+	} while (!is_end(node));
 	dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
+	*approx_after = 0;
+	if (nr_cpus_mask >= 0) {
+		int i;
+
+		for (i = 0; i < nr_cpus_mask + 1; i++) {
+			*approx_after += uatomic_read(&ht->percpu_count[i].add);
+			*approx_after -= uatomic_read(&ht->percpu_count[i].del);
+		}
+	}
 }
 
+/* called with resize mutex held */
 static
-void cds_lfht_free_table_cb(struct rcu_head *head)
+void _do_cds_lfht_grow(struct cds_lfht *ht,
+		unsigned long old_size, unsigned long new_size)
 {
-	struct rcu_table *t =
-		caa_container_of(head, struct rcu_table, head);
-	free(t);
+	unsigned long old_order, new_order;
+
+	old_order = get_count_order_ulong(old_size);
+	new_order = get_count_order_ulong(new_size);
+	dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
+		   old_size, old_order, new_size, new_order);
+	assert(new_size > old_size);
+	init_table(ht, old_order + 1, new_order);
 }
 
 /* called with resize mutex held */
 static
-void _do_cds_lfht_resize(struct cds_lfht *ht)
+void _do_cds_lfht_shrink(struct cds_lfht *ht,
+		unsigned long old_size, unsigned long new_size)
 {
-	unsigned long new_size, old_size, old_order, new_order;
-	struct rcu_table *new_t, *old_t;
+	unsigned long old_order, new_order;
 
-	old_t = ht->t;
-	old_size = old_t->size;
-	old_order = get_count_order_ulong(old_size) + 1;
-
-	new_size = CMM_LOAD_SHARED(old_t->resize_target);
-	if (old_size == new_size)
-		return;
-	new_order = get_count_order_ulong(new_size) + 1;
+	new_size = max(new_size, MIN_TABLE_SIZE);
+	old_order = get_count_order_ulong(old_size);
+	new_order = get_count_order_ulong(new_size);
 	dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
-	       old_size, old_order, new_size, new_order);
-	new_t = malloc(sizeof(struct cds_lfht)
-			+ (new_order * sizeof(struct _cds_lfht_node *)));
-	assert(new_size > old_size);
-	memcpy(&new_t->tbl, &old_t->tbl,
-	       old_order * sizeof(struct _cds_lfht_node *));
-	init_table(ht, new_t, old_order, new_order - old_order);
-	/* Changing table and size atomically wrt lookups */
-	rcu_assign_pointer(ht->t, new_t);
-	ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb);
+		   old_size, old_order, new_size, new_order);
+	assert(new_size < old_size);
+
+	/* Remove and unlink all dummy nodes to remove. */
+	fini_table(ht, new_order + 1, old_order);
 }
 
+
+/* called with resize mutex held */
 static
-unsigned long resize_target_update(struct rcu_table *t,
-				   int growth_order)
+void _do_cds_lfht_resize(struct cds_lfht *ht)
 {
-	return _uatomic_max(&t->resize_target,
-			    t->size << growth_order);
+	unsigned long new_size, old_size;
+
+	/*
+	 * Resize table, re-do if the target size has changed under us.
+	 */
+	do {
+		assert(uatomic_read(&ht->in_progress_resize));
+		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
+			break;
+		ht->t.resize_initiated = 1;
+		old_size = ht->t.size;
+		new_size = CMM_LOAD_SHARED(ht->t.resize_target);
+		if (old_size < new_size)
+			_do_cds_lfht_grow(ht, old_size, new_size);
+		else if (old_size > new_size)
+			_do_cds_lfht_shrink(ht, old_size, new_size);
+		ht->t.resize_initiated = 0;
+		/* write resize_initiated before read resize_target */
+		cmm_smp_mb();
+	} while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
 }
 
-void cds_lfht_resize(struct cds_lfht *ht, int growth)
+static
+unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
+				   int growth_order)
 {
-	struct rcu_table *t = rcu_dereference(ht->t);
-	unsigned long target_size;
+	return _uatomic_max(&ht->t.resize_target,
+			    size << growth_order);
+}
 
-	if (growth < 0) {
-		/*
-		 * Silently refuse to shrink hash table. (not supported)
-		 */
-		dbg_printf("shrinking hash table not supported.\n");
-		return;
-	}
+static
+void resize_target_update_count(struct cds_lfht *ht,
+				unsigned long count)
+{
+	count = max(count, MIN_TABLE_SIZE);
+	uatomic_set(&ht->t.resize_target, count);
+}
 
-	target_size = resize_target_update(t, growth);
-	if (t->size < target_size) {
-		CMM_STORE_SHARED(t->resize_initiated, 1);
-		pthread_mutex_lock(&ht->resize_mutex);
-		_do_cds_lfht_resize(ht);
-		pthread_mutex_unlock(&ht->resize_mutex);
-	}
+void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
+{
+	resize_target_update_count(ht, new_size);
+	CMM_STORE_SHARED(ht->t.resize_initiated, 1);
+	ht->cds_lfht_rcu_thread_offline();
+	pthread_mutex_lock(&ht->resize_mutex);
+	_do_cds_lfht_resize(ht);
+	pthread_mutex_unlock(&ht->resize_mutex);
+	ht->cds_lfht_rcu_thread_online();
 }
 
 static
@@ -865,27 +1735,64 @@ void do_resize_cb(struct rcu_head *head)
 		caa_container_of(head, struct rcu_resize_work, head);
 	struct cds_lfht *ht = work->ht;
 
+	ht->cds_lfht_rcu_thread_offline();
 	pthread_mutex_lock(&ht->resize_mutex);
 	_do_cds_lfht_resize(ht);
 	pthread_mutex_unlock(&ht->resize_mutex);
-	free(work);
+	ht->cds_lfht_rcu_thread_online();
+	poison_free(work);
 	cmm_smp_mb();	/* finish resize before decrement */
 	uatomic_dec(&ht->in_progress_resize);
 }
 
 static
-void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth)
+void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
 {
 	struct rcu_resize_work *work;
 	unsigned long target_size;
 
-	target_size = resize_target_update(t, growth);
-	if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) {
+	target_size = resize_target_update(ht, size, growth);
+	/* Store resize_target before read resize_initiated */
+	cmm_smp_mb();
+	if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
 		uatomic_inc(&ht->in_progress_resize);
-		cmm_smp_mb();	/* increment resize count before calling it */
+		cmm_smp_mb();	/* increment resize count before load destroy */
+		if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
+			uatomic_dec(&ht->in_progress_resize);
+			return;
+		}
 		work = malloc(sizeof(*work));
 		work->ht = ht;
 		ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
-		CMM_STORE_SHARED(t->resize_initiated, 1);
+		CMM_STORE_SHARED(ht->t.resize_initiated, 1);
 	}
 }
+
+#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+
+static
+void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
+				unsigned long count)
+{
+	struct rcu_resize_work *work;
+
+	if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
+		return;
+	resize_target_update_count(ht, count);
+	/* Store resize_target before read resize_initiated */
+	cmm_smp_mb();
+	if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
+		uatomic_inc(&ht->in_progress_resize);
+		cmm_smp_mb();	/* increment resize count before load destroy */
+		if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
+			uatomic_dec(&ht->in_progress_resize);
+			return;
+		}
+		work = malloc(sizeof(*work));
+		work->ht = ht;
+		ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
+		CMM_STORE_SHARED(ht->t.resize_initiated, 1);
+	}
+}
+
+#endif