X-Git-Url: https://git.lttng.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=e5c91b9730bb2ffea069dbd1f68e84a41ef811e3;hp=18a3cb89c8b273b81757fd763f4b472d8301e6d6;hb=b84f92d592010f01851c3fdcd5ac7d3666028146;hpb=93d46c395ff595372827983e9e84bd1e49eb6fab

diff --git a/rculfhash.c b/rculfhash.c
index 18a3cb8..e5c91b9 100644
--- a/rculfhash.c
+++ b/rculfhash.c
@@ -53,7 +53,7 @@
  *   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
+ * - 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.
@@ -173,12 +173,13 @@
 #endif
 
 /*
- * Per-CPU split-counters lazily update the global counter each 1024
+ * 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 DEFAULT_SPLIT_COUNT_MASK	0xFUL
 #define CHAIN_LEN_TARGET		1
 #define CHAIN_LEN_RESIZE_THRESHOLD	3
 
@@ -221,15 +222,37 @@
 /* Value of the end pointer. Should not interact with flags. */
 #define END_VALUE		NULL
 
+/*
+ * ht_items_count: Split-counters counting the number of node addition
+ * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
+ * is set at hash table creation.
+ *
+ * These are free-running counters, never reset to zero. They count the
+ * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
+ * operations to update the global counter. We choose a power-of-2 value
+ * for the trigger to deal with 32 or 64-bit overflow of the counter.
+ */
 struct ht_items_count {
 	unsigned long add, del;
 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 
+/*
+ * rcu_level: Contains the per order-index-level dummy node table. The
+ * size of each dummy node table is half the number of hashes contained
+ * in this order (except for order 0). The minimum allocation size
+ * parameter allows combining the dummy node arrays of the lowermost
+ * levels to improve cache locality for small index orders.
+ */
 struct rcu_level {
 	/* Note: manually update allocation length when adding a field */
 	struct _cds_lfht_node nodes[0];
 };
 
+/*
+ * rcu_table: Contains the size and desired new size if a resize
+ * operation is in progress, as well as the statically-sized array of
+ * rcu_level pointers.
+ */
 struct rcu_table {
 	unsigned long size;	/* always a power of 2, shared (RCU) */
 	unsigned long resize_target;
@@ -237,11 +260,15 @@ struct rcu_table {
 	struct rcu_level *tbl[MAX_TABLE_ORDER];
 };
 
+/*
+ * cds_lfht: Top-level data structure representing a lock-free hash
+ * table. Defined in the implementation file to make it be an opaque
+ * cookie to users.
+ */
 struct cds_lfht {
 	struct rcu_table t;
-	cds_lfht_hash_fct hash_fct;
-	cds_lfht_compare_fct compare_fct;
-	unsigned long hash_seed;
+	unsigned long min_alloc_order;
+	unsigned long min_alloc_size;
 	int flags;
 	/*
 	 * We need to put the work threads offline (QSBR) when taking this
@@ -263,14 +290,23 @@ struct cds_lfht {
 	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 ht_items_count *split_count;	/* split item count */
 };
 
+/*
+ * rcu_resize_work: Contains arguments passed to RCU worker thread
+ * responsible for performing lazy resize.
+ */
 struct rcu_resize_work {
 	struct rcu_head head;
 	struct cds_lfht *ht;
 };
 
+/*
+ * partition_resize_work: Contains arguments passed to worker threads
+ * executing the hash table resize on partitions of the hash table
+ * assigned to each processor's worker thread.
+ */
 struct partition_resize_work {
 	pthread_t thread_id;
 	struct cds_lfht *ht;
@@ -281,6 +317,7 @@ struct partition_resize_work {
 
 static
 void _cds_lfht_add(struct cds_lfht *ht,
+		cds_lfht_match_fct match,
 		unsigned long size,
 		struct cds_lfht_node *node,
 		struct cds_lfht_iter *unique_ret,
@@ -451,7 +488,7 @@ unsigned int fls_u32(uint32_t x)
 
 unsigned int fls_ulong(unsigned long x)
 {
-#if (CAA_BITS_PER_lONG == 32)
+#if (CAA_BITS_PER_LONG == 32)
 	return fls_u32(x);
 #else
 	return fls_u64(x);
@@ -483,97 +520,115 @@ int get_count_order_ulong(unsigned long x)
 }
 
 #ifdef POISON_FREE
-#define poison_free(ptr)				\
-	do {						\
-		memset(ptr, 0x42, sizeof(*(ptr)));	\
-		free(ptr);				\
+#define poison_free(ptr)					\
+	do {							\
+		if (ptr) {					\
+			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)
+void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
 
 static
 void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
 				unsigned long count);
 
 static long nr_cpus_mask = -1;
+static long split_count_mask = -1;
+
+#if defined(HAVE_SYSCONF)
+static void ht_init_nr_cpus_mask(void)
+{
+	long maxcpus;
+
+	maxcpus = sysconf(_SC_NPROCESSORS_CONF);
+	if (maxcpus <= 0) {
+		nr_cpus_mask = -2;
+		return;
+	}
+	/*
+	 * 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;
+}
+#else /* #if defined(HAVE_SYSCONF) */
+static void ht_init_nr_cpus_mask(void)
+{
+	nr_cpus_mask = -2;
+}
+#endif /* #else #if defined(HAVE_SYSCONF) */
 
 static
-struct ht_items_count *alloc_per_cpu_items_count(void)
+void alloc_split_items_count(struct cds_lfht *ht)
 {
 	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;
+	if (nr_cpus_mask == -1)	{
+		ht_init_nr_cpus_mask();
+		if (nr_cpus_mask < 0)
+			split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
+		else
+			split_count_mask = nr_cpus_mask;
 	}
-		/* Fall-through */
-	default:
-		return calloc(nr_cpus_mask + 1, sizeof(*count));
+
+	assert(split_count_mask >= 0);
+
+	if (ht->flags & CDS_LFHT_ACCOUNTING) {
+		ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
+		assert(ht->split_count);
+	} else {
+		ht->split_count = NULL;
 	}
 }
 
 static
-void free_per_cpu_items_count(struct ht_items_count *count)
+void free_split_items_count(struct cds_lfht *ht)
 {
-	poison_free(count);
+	poison_free(ht->split_count);
 }
 
+#if defined(HAVE_SCHED_GETCPU)
 static
-int ht_get_cpu(void)
+int ht_get_split_count_index(unsigned long hash)
 {
 	int cpu;
 
-	assert(nr_cpus_mask >= 0);
+	assert(split_count_mask >= 0);
 	cpu = sched_getcpu();
-	if (unlikely(cpu < 0))
-		return cpu;
+	if (caa_unlikely(cpu < 0))
+		return hash & split_count_mask;
 	else
-		return cpu & nr_cpus_mask;
+		return cpu & split_count_mask;
 }
+#else /* #if defined(HAVE_SCHED_GETCPU) */
+static
+int ht_get_split_count_index(unsigned long hash)
+{
+	return hash & split_count_mask;
+}
+#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
 
 static
-void ht_count_add(struct cds_lfht *ht, unsigned long size)
+void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
 {
-	unsigned long percpu_count;
-	int cpu;
+	unsigned long split_count;
+	int index;
 
-	if (unlikely(!ht->percpu_count))
-		return;
-	cpu = ht_get_cpu();
-	if (unlikely(cpu < 0))
+	if (caa_unlikely(!ht->split_count))
 		return;
-	percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
-	if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+	index = ht_get_split_count_index(hash);
+	split_count = uatomic_add_return(&ht->split_count[index].add, 1);
+	if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
 		long count;
 
-		dbg_printf("add percpu %lu\n", percpu_count);
+		dbg_printf("add split count %lu\n", split_count);
 		count = uatomic_add_return(&ht->count,
 					   1UL << COUNT_COMMIT_ORDER);
 		/* If power of 2 */
@@ -588,21 +643,19 @@ void ht_count_add(struct cds_lfht *ht, unsigned long size)
 }
 
 static
-void ht_count_del(struct cds_lfht *ht, unsigned long size)
+void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
 {
-	unsigned long percpu_count;
-	int cpu;
+	unsigned long split_count;
+	int index;
 
-	if (unlikely(!ht->percpu_count))
-		return;
-	cpu = ht_get_cpu();
-	if (unlikely(cpu < 0))
+	if (caa_unlikely(!ht->split_count))
 		return;
-	percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
-	if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+	index = ht_get_split_count_index(hash);
+	split_count = uatomic_add_return(&ht->split_count[index].del, 1);
+	if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
 		long count;
 
-		dbg_printf("del percpu %lu\n", percpu_count);
+		dbg_printf("del split count %lu\n", split_count);
 		count = uatomic_add_return(&ht->count,
 					   -(1UL << COUNT_COMMIT_ORDER));
 		/* If power of 2 */
@@ -614,7 +667,7 @@ void ht_count_del(struct cds_lfht *ht, unsigned long size)
 			 * Don't shrink table if the number of nodes is below a
 			 * certain threshold.
 			 */
-			if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
+			if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
 				return;
 			cds_lfht_resize_lazy_count(ht, size,
 				count >> (CHAIN_LEN_TARGET - 1));
@@ -622,34 +675,6 @@ void ht_count_del(struct cds_lfht *ht, unsigned long size)
 	}
 }
 
-#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, unsigned long size, uint32_t chain_len)
 {
@@ -668,7 +693,7 @@ void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
 		dbg_printf("WARNING: large chain length: %u.\n",
 			   chain_len);
 	if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
-		cds_lfht_resize_lazy(ht, size,
+		cds_lfht_resize_lazy_grow(ht, size,
 			get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
 }
 
@@ -715,7 +740,8 @@ int is_end(struct cds_lfht_node *node)
 }
 
 static
-unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
+unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
+		unsigned long v)
 {
 	unsigned long old1, old2;
 
@@ -725,7 +751,7 @@ unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
 		if (old2 >= v)
 			return old2;
 	} while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
-	return v;
+	return old2;
 }
 
 static
@@ -736,17 +762,21 @@ struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
 
 	assert(size > 0);
 	index = hash & (size - 1);
+
+	if (index < ht->min_alloc_size) {
+		dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
+			   hash, index);
+		return &ht->t.tbl[0]->nodes[index];
+	}
 	/*
 	 * 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))];
+		   hash, index, order, index & ((1UL << (order - 1)) - 1));
+	return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)];
 }
 
 /*
@@ -775,12 +805,12 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node
 		 */
 		assert(dummy != node);
 		for (;;) {
-			if (unlikely(is_end(iter)))
+			if (caa_unlikely(is_end(iter)))
 				return;
-			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
+			if (caa_likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
 				return;
 			next = rcu_dereference(clear_flag(iter)->p.next);
-			if (likely(is_removed(next)))
+			if (caa_likely(is_removed(next)))
 				break;
 			iter_prev = clear_flag(iter);
 			iter = next;
@@ -860,6 +890,7 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
  */
 static
 void _cds_lfht_add(struct cds_lfht *ht,
+		cds_lfht_match_fct match,
 		unsigned long size,
 		struct cds_lfht_node *node,
 		struct cds_lfht_iter *unique_ret,
@@ -871,12 +902,6 @@ void _cds_lfht_add(struct cds_lfht *ht,
 
 	assert(!is_dummy(node));
 	assert(!is_removed(node));
-	if (!size) {
-		assert(dummy);
-		assert(!unique_ret);
-		node->p.next = flag_dummy(get_end());
-		return;		/* Initial first add (head) */
-	}
 	lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
 	for (;;) {
 		uint32_t chain_len = 0;
@@ -890,26 +915,42 @@ void _cds_lfht_add(struct cds_lfht *ht,
 		iter = rcu_dereference(iter_prev->p.next);
 		assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
 		for (;;) {
-			if (unlikely(is_end(iter)))
+			if (caa_unlikely(is_end(iter)))
 				goto insert;
-			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
+			if (caa_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)))
+			if (caa_unlikely(is_removed(next)))
 				goto gc_node;
+
+			/* uniquely add */
 			if (unique_ret
 			    && !is_dummy(next)
-			    && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash
-			    && !ht->compare_fct(node->key, node->key_len,
-						clear_flag(iter)->key,
-						clear_flag(iter)->key_len)) {
-				unique_ret->node = clear_flag(iter);
-				unique_ret->next = next;
+			    && 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, match, &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))
@@ -974,7 +1015,7 @@ int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
 		struct cds_lfht_node *new_next;
 
 		next = old;
-		if (unlikely(is_removed(next)))
+		if (caa_unlikely(is_removed(next)))
 			return -ENOENT;
 		if (dummy_removal)
 			assert(is_dummy(next));
@@ -1068,16 +1109,17 @@ void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
 {
 	unsigned long j;
 
+	assert(i > ht->min_alloc_order);
 	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, !i ? 0 : (1UL << (i - 1)) + j);
+			   i, j, (1UL << (i - 1)) + j);
 		new_node->p.reverse_hash =
-			bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
-		_cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
+				bit_reverse_ulong((1UL << (i - 1)) + j);
+		_cds_lfht_add(ht, NULL, 1UL << (i - 1),
 				new_node, NULL, 1);
 	}
 	ht->cds_lfht_rcu_read_unlock();
@@ -1105,14 +1147,15 @@ void init_table(struct cds_lfht *ht,
 
 	dbg_printf("init table: first_order %lu last_order %lu\n",
 		   first_order, last_order);
+	assert(first_order > ht->min_alloc_order);
 	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);
 
 		/* Stop expand if the resize target changes under us */
-		if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
+		if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
 			break;
 
 		ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
@@ -1128,9 +1171,9 @@ void init_table(struct cds_lfht *ht,
 		 * Update table size.
 		 */
 		cmm_smp_wmb();	/* populate data before RCU size */
-		CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
+		CMM_STORE_SHARED(ht->t.size, 1UL << i);
 
-		dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
+		dbg_printf("init new size: %lu\n", 1UL << i);
 		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
 			break;
 	}
@@ -1167,17 +1210,17 @@ void remove_table_partition(struct cds_lfht *ht, unsigned long i,
 {
 	unsigned long j;
 
+	assert(i > ht->min_alloc_order);
 	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, !i ? 0 : (1UL << (i - 1)) + j);
+			   i, j, (1UL << (i - 1)) + j);
 		fini_node->p.reverse_hash =
-			bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
-		(void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
-				fini_node, 1);
+			bit_reverse_ulong((1UL << (i - 1)) + j);
+		(void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
 	}
 	ht->cds_lfht_rcu_read_unlock();
 }
@@ -1205,11 +1248,11 @@ void fini_table(struct cds_lfht *ht,
 
 	dbg_printf("fini table: first_order %lu last_order %lu\n",
 		   first_order, last_order);
-	assert(first_order > 0);
+	assert(first_order > ht->min_alloc_order);
 	for (i = last_order; i >= first_order; i--) {
 		unsigned long len;
 
-		len = !i ? 1 : 1UL << (i - 1);
+		len = 1UL << (i - 1);
 		dbg_printf("fini order %lu len: %lu\n", i, len);
 
 		/* Stop shrink if the resize target changes under us */
@@ -1250,10 +1293,51 @@ void fini_table(struct cds_lfht *ht,
 	}
 }
 
-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,
+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, ht->min_alloc_size * 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);
+		if (order <= ht->min_alloc_order) {
+			ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
+		} else {
+			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(unsigned long init_size,
+			unsigned long min_alloc_size,
 			int flags,
 			void (*cds_lfht_call_rcu)(struct rcu_head *head,
 					void (*func)(struct rcu_head *head)),
@@ -1269,14 +1353,17 @@ struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
 	struct cds_lfht *ht;
 	unsigned long order;
 
+	/* min_alloc_size must be power of two */
+	if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
+		return NULL;
 	/* init_size must be power of two */
-	if (init_size && (init_size & (init_size - 1)))
+	if (!init_size || (init_size & (init_size - 1)))
 		return NULL;
+	min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
+	init_size = max(init_size, min_alloc_size);
 	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->flags = flags;
 	ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
 	ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
 	ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
@@ -1286,28 +1373,25 @@ struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
 	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();
+	alloc_split_items_count(ht);
 	/* 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, MIN_TABLE_SIZE));
-	ht->flags = flags;
-	ht->cds_lfht_rcu_thread_offline();
-	pthread_mutex_lock(&ht->resize_mutex);
+	order = get_count_order_ulong(init_size);
 	ht->t.resize_target = 1UL << order;
-	init_table(ht, 0, order);
-	pthread_mutex_unlock(&ht->resize_mutex);
-	ht->cds_lfht_rcu_thread_online();
+	ht->min_alloc_size = min_alloc_size;
+	ht->min_alloc_order = get_count_order_ulong(min_alloc_size);
+	cds_lfht_create_dummy(ht, 1UL << order);
+	ht->t.size = 1UL << order;
 	return ht;
 }
 
-void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
-		struct cds_lfht_iter *iter)
+void cds_lfht_lookup(struct cds_lfht *ht, cds_lfht_match_fct match,
+		unsigned long hash, void *key, struct cds_lfht_iter *iter)
 {
 	struct cds_lfht_node *node, *next, *dummy_node;
 	struct _cds_lfht_node *lookup;
-	unsigned long hash, reverse_hash, size;
+	unsigned long reverse_hash, size;
 
-	hash = ht->hash_fct(key, key_len, ht->hash_seed);
 	reverse_hash = bit_reverse_ulong(hash);
 
 	size = rcu_dereference(ht->t.size);
@@ -1317,19 +1401,20 @@ void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
 	node = rcu_dereference(dummy_node->p.next);
 	node = clear_flag(node);
 	for (;;) {
-		if (unlikely(is_end(node))) {
+		if (caa_unlikely(is_end(node))) {
 			node = next = NULL;
 			break;
 		}
-		if (unlikely(node->p.reverse_hash > reverse_hash)) {
+		if (caa_unlikely(node->p.reverse_hash > reverse_hash)) {
 			node = next = NULL;
 			break;
 		}
 		next = rcu_dereference(node->p.next);
-		if (likely(!is_removed(next))
+		assert(node == clear_flag(node));
+		if (caa_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))) {
+		    && node->p.reverse_hash == reverse_hash
+		    && caa_likely(match(node, key))) {
 				break;
 		}
 		node = clear_flag(next);
@@ -1339,33 +1424,32 @@ void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
 	iter->next = next;
 }
 
-void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
+		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))) {
+		if (caa_unlikely(is_end(node))) {
 			node = next = NULL;
 			break;
 		}
-		if (unlikely(node->p.reverse_hash > reverse_hash)) {
+		if (caa_unlikely(node->p.reverse_hash > reverse_hash)) {
 			node = next = NULL;
 			break;
 		}
 		next = rcu_dereference(node->p.next);
-		if (likely(!is_removed(next))
+		if (caa_likely(!is_removed(next))
 		    && !is_dummy(next)
-		    && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
+		    && caa_likely(match(node->key, key))) {
 				break;
 		}
 		node = clear_flag(next);
@@ -1381,12 +1465,12 @@ void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
 
 	node = clear_flag(iter->next);
 	for (;;) {
-		if (unlikely(is_end(node))) {
+		if (caa_unlikely(is_end(node))) {
 			node = next = NULL;
 			break;
 		}
 		next = rcu_dereference(node->p.next);
-		if (likely(!is_removed(next))
+		if (caa_likely(!is_removed(next))
 		    && !is_dummy(next)) {
 				break;
 		}
@@ -1410,48 +1494,47 @@ void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
 	cds_lfht_next(ht, iter);
 }
 
-void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
+void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
+		struct cds_lfht_node *node)
 {
-	unsigned long hash, size;
+	unsigned long size;
 
-	hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
 	node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
-
 	size = rcu_dereference(ht->t.size);
-	_cds_lfht_add(ht, size, node, NULL, 0);
-	ht_count_add(ht, size);
+	_cds_lfht_add(ht, NULL, size, node, NULL, 0);
+	ht_count_add(ht, size, hash);
 }
 
 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
+				cds_lfht_match_fct match,
+				unsigned long hash,
 				struct cds_lfht_node *node)
 {
-	unsigned long hash, size;
+	unsigned long 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);
-
 	size = rcu_dereference(ht->t.size);
-	_cds_lfht_add(ht, size, node, &iter, 0);
+	_cds_lfht_add(ht, match, size, node, &iter, 0);
 	if (iter.node == node)
-		ht_count_add(ht, size);
+		ht_count_add(ht, size, hash);
 	return iter.node;
 }
 
 struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
+				cds_lfht_match_fct match,
+				unsigned long hash,
 				struct cds_lfht_node *node)
 {
-	unsigned long hash, size;
+	unsigned long 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);
-
 	size = rcu_dereference(ht->t.size);
 	for (;;) {
-		_cds_lfht_add(ht, size, node, &iter, 0);
+		_cds_lfht_add(ht, match, size, node, &iter, 0);
 		if (iter.node == node) {
-			ht_count_add(ht, size);
+			ht_count_add(ht, size, hash);
 			return NULL;
 		}
 
@@ -1472,13 +1555,15 @@ int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
 
 int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
 {
-	unsigned long size;
+	unsigned long size, hash;
 	int ret;
 
 	size = rcu_dereference(ht->t.size);
 	ret = _cds_lfht_del(ht, size, iter->node, 0);
-	if (!ret)
-		ht_count_del(ht, size);
+	if (!ret) {
+		hash = bit_reverse_ulong(iter->node->p.reverse_hash);
+		ht_count_del(ht, size, hash);
+	}
 	return ret;
 }
 
@@ -1514,7 +1599,12 @@ int cds_lfht_delete_dummy(struct cds_lfht *ht)
 				bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
 			assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
 		}
-		poison_free(ht->t.tbl[order]);
+
+		if (order == ht->min_alloc_order)
+			poison_free(ht->t.tbl[0]);
+		else if (order > ht->min_alloc_order)
+			poison_free(ht->t.tbl[order]);
+		/* Nothing to delete for order < ht->min_alloc_order */
 	}
 	return 0;
 }
@@ -1535,7 +1625,7 @@ int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
 	ret = cds_lfht_delete_dummy(ht);
 	if (ret)
 		return ret;
-	free_per_cpu_items_count(ht->percpu_count);
+	free_split_items_count(ht);
 	if (attr)
 		*attr = ht->resize_attr;
 	poison_free(ht);
@@ -1553,12 +1643,12 @@ void cds_lfht_count_nodes(struct cds_lfht *ht,
 	unsigned long nr_dummy = 0;
 
 	*approx_before = 0;
-	if (nr_cpus_mask >= 0) {
+	if (ht->split_count) {
 		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);
+		for (i = 0; i < split_count_mask + 1; i++) {
+			*approx_before += uatomic_read(&ht->split_count[i].add);
+			*approx_before -= uatomic_read(&ht->split_count[i].del);
 		}
 	}
 
@@ -1583,12 +1673,12 @@ void cds_lfht_count_nodes(struct cds_lfht *ht,
 	} while (!is_end(node));
 	dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
 	*approx_after = 0;
-	if (nr_cpus_mask >= 0) {
+	if (ht->split_count) {
 		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);
+		for (i = 0; i < split_count_mask + 1; i++) {
+			*approx_after += uatomic_read(&ht->split_count[i].add);
+			*approx_after -= uatomic_read(&ht->split_count[i].del);
 		}
 	}
 }
@@ -1615,7 +1705,7 @@ void _do_cds_lfht_shrink(struct cds_lfht *ht,
 {
 	unsigned long old_order, new_order;
 
-	new_size = max(new_size, MIN_TABLE_SIZE);
+	new_size = max(new_size, ht->min_alloc_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",
@@ -1654,18 +1744,16 @@ void _do_cds_lfht_resize(struct cds_lfht *ht)
 }
 
 static
-unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
-				   int growth_order)
+unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
 {
-	return _uatomic_max(&ht->t.resize_target,
-			    size << growth_order);
+	return _uatomic_xchg_monotonic_increase(&ht->t.resize_target, new_size);
 }
 
 static
 void resize_target_update_count(struct cds_lfht *ht,
 				unsigned long count)
 {
-	count = max(count, MIN_TABLE_SIZE);
+	count = max(count, ht->min_alloc_size);
 	uatomic_set(&ht->t.resize_target, count);
 }
 
@@ -1698,15 +1786,13 @@ void do_resize_cb(struct rcu_head *head)
 }
 
 static
-void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
+void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
 {
 	struct rcu_resize_work *work;
-	unsigned long 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) {
+	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)) {
@@ -1720,31 +1806,47 @@ void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
 	}
 }
 
-#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+static
+void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
+{
+	unsigned long target_size = size << growth;
+
+	if (resize_target_grow(ht, target_size) >= target_size)
+		return;
 
+	__cds_lfht_resize_lazy_launch(ht);
+}
+
+/*
+ * We favor grow operations over shrink. A shrink operation never occurs
+ * if a grow operation is queued for lazy execution. A grow operation
+ * cancels any pending shrink lazy execution.
+ */
 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);
+	count = max(count, ht->min_alloc_size);
+	if (count == size)
+		return;		/* Already the right size, no resize needed */
+	if (count > size) {	/* lazy grow */
+		if (resize_target_grow(ht, count) >= count)
 			return;
+	} else {		/* lazy shrink */
+		for (;;) {
+			unsigned long s;
+
+			s = uatomic_cmpxchg(&ht->t.resize_target, size, count);
+			if (s == size)
+				break;	/* no resize needed */
+			if (s > size)
+				return;	/* growing is/(was just) in progress */
+			if (s <= count)
+				return;	/* some other thread do shrink */
+			size = s;
 		}
-		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);
 	}
+	__cds_lfht_resize_lazy_launch(ht);
 }
-
-#endif