X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=f3c96880fc6dc12c920a204bf284be14f688c16c;hp=8058931b386cf9d75f64173d7ff0e3333d486032;hb=33c7c748c098dfc963c934fde62328169834c1ed;hpb=9d2614f07691a813a3c560a6c0bcd0a7be854ed5

diff --git a/rculfhash.c b/rculfhash.c
index 8058931..f3c9688 100644
--- a/rculfhash.c
+++ b/rculfhash.c
@@ -38,25 +38,45 @@
 #include <stdio.h>
 #include <pthread.h>
 
-#define BUCKET_SIZE_RESIZE_THRESHOLD	32
-#define MAX_NR_BUCKETS			1048576		/* 1M buckets */
+//#define DEBUG		/* Test */
+
+#ifdef DEBUG
+#define dbg_printf(args...)     printf(args)
+#else
+#define dbg_printf(args...)
+#endif
+
+#define CHAIN_LEN_TARGET		4
+#define CHAIN_LEN_RESIZE_THRESHOLD	8
 
 #ifndef max
 #define max(a, b)	((a) > (b) ? (a) : (b))
 #endif
 
+/*
+ * The removed flag needs to be updated atomically with the pointer.
+ * 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.
+ */
+#define REMOVED_FLAG		(1UL << 0)
+#define DUMMY_FLAG		(1UL << 1)
+#define FLAGS_MASK		((1UL << 2) - 1)
+
 struct rcu_table {
 	unsigned long size;	/* always a power of 2 */
+	unsigned long resize_target;
+	int resize_initiated;
 	struct rcu_head head;
-	struct rcu_ht_node *tbl[0];
+	struct _rcu_ht_node *tbl[0];
 };
 
 struct rcu_ht {
 	struct rcu_table *t;		/* shared */
 	ht_hash_fct hash_fct;
-	void *hashseed;
+	ht_compare_fct compare_fct;
+	unsigned long hash_seed;
 	pthread_mutex_t resize_mutex;	/* resize mutex: add/del mutex */
-	unsigned long target_size;
+	unsigned int in_progress_resize, in_progress_destroy;
 	void (*ht_call_rcu)(struct rcu_head *head,
 		      void (*func)(struct rcu_head *head));
 };
@@ -66,23 +86,12 @@ struct rcu_resize_work {
 	struct rcu_ht *ht;
 };
 
-static
-void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth);
-
-static
-void check_resize(struct rcu_ht *ht, struct rcu_table *t,
-		  unsigned long chain_len)
-{
-	//printf("check resize chain len %lu\n", chain_len);
-	if (chain_len >= BUCKET_SIZE_RESIZE_THRESHOLD)
-		ht_resize_lazy(ht, t, chain_len / BUCKET_SIZE_RESIZE_THRESHOLD);
-}
-
 /*
  * Algorithm to reverse bits in a word by lookup table, extended to
  * 64-bit words.
- * ref.
+ * Source:
  * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
+ * Originally from Public Domain.
  */
 
 static const uint8_t BitReverseTable256[256] = 
@@ -134,26 +143,188 @@ unsigned long bit_reverse_ulong(unsigned long v)
 #endif
 }
 
+/*
+ * fls: returns the position of the most significant bit.
+ * Returns 0 if no bit is set, else returns the position of the most
+ * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
+ */
+#if defined(__i386) || defined(__x86_64)
+static inline
+unsigned int fls_u32(uint32_t x)
+{
+	int r;
+
+	asm("bsrl %1,%0\n\t"
+	    "jnz 1f\n\t"
+	    "movl $-1,%0\n\t"
+	    "1:\n\t"
+	    : "=r" (r) : "rm" (x));
+	return r + 1;
+}
+#define HAS_FLS_U32
+#endif
+
+#if defined(__x86_64)
+static inline
+unsigned int fls_u64(uint64_t x)
+{
+	long r;
+
+	asm("bsrq %1,%0\n\t"
+	    "jnz 1f\n\t"
+	    "movq $-1,%0\n\t"
+	    "1:\n\t"
+	    : "=r" (r) : "rm" (x));
+	return r + 1;
+}
+#define HAS_FLS_U64
+#endif
+
+#ifndef HAS_FLS_U64
+static __attribute__((unused))
+unsigned int fls_u64(uint64_t x)
+{
+	unsigned int r = 64;
+
+	if (!x)
+		return 0;
+
+	if (!(x & 0xFFFFFFFF00000000ULL)) {
+		x <<= 32;
+		r -= 32;
+	}
+	if (!(x & 0xFFFF000000000000ULL)) {
+		x <<= 16;
+		r -= 16;
+	}
+	if (!(x & 0xFF00000000000000ULL)) {
+		x <<= 8;
+		r -= 8;
+	}
+	if (!(x & 0xF000000000000000ULL)) {
+		x <<= 4;
+		r -= 4;
+	}
+	if (!(x & 0xC000000000000000ULL)) {
+		x <<= 2;
+		r -= 2;
+	}
+	if (!(x & 0x8000000000000000ULL)) {
+		x <<= 1;
+		r -= 1;
+	}
+	return r;
+}
+#endif
+
+#ifndef HAS_FLS_U32
+static __attribute__((unused))
+unsigned int fls_u32(uint32_t x)
+{
+	unsigned int r = 32;
+
+	if (!x)
+		return 0;
+	if (!(x & 0xFFFF0000U)) {
+		x <<= 16;
+		r -= 16;
+	}
+	if (!(x & 0xFF000000U)) {
+		x <<= 8;
+		r -= 8;
+	}
+	if (!(x & 0xF0000000U)) {
+		x <<= 4;
+		r -= 4;
+	}
+	if (!(x & 0xC0000000U)) {
+		x <<= 2;
+		r -= 2;
+	}
+	if (!(x & 0x80000000U)) {
+		x <<= 1;
+		r -= 1;
+	}
+	return r;
+}
+#endif
+
+unsigned int fls_ulong(unsigned long x)
+{
+#if (CAA_BITS_PER_lONG == 32)
+	return fls_u32(x);
+#else
+	return fls_u64(x);
+#endif
+}
+
+int get_count_order_u32(uint32_t x)
+{
+	int order;
+
+	order = fls_u32(x) - 1;
+	if (x & (x - 1))
+		order++;
+	return order;
+}
+
+int get_count_order_ulong(unsigned long x)
+{
+	int order;
+
+	order = fls_ulong(x) - 1;
+	if (x & (x - 1))
+		order++;
+	return order;
+}
+
+static
+void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth);
+
+static
+void check_resize(struct rcu_ht *ht, struct rcu_table *t,
+		  uint32_t chain_len)
+{
+	if (chain_len > 100)
+		dbg_printf("rculfhash: WARNING: large chain length: %u.\n",
+			   chain_len);
+	if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
+		ht_resize_lazy(ht, t,
+			get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
+}
+
 static
 struct rcu_ht_node *clear_flag(struct rcu_ht_node *node)
 {
-	return (struct rcu_ht_node *) (((unsigned long) node) & ~0x1);
+	return (struct rcu_ht_node *) (((unsigned long) node) & ~FLAGS_MASK);
 }
 
 static
 int is_removed(struct rcu_ht_node *node)
 {
-	return ((unsigned long) node) & 0x1;
+	return ((unsigned long) node) & REMOVED_FLAG;
 }
 
 static
 struct rcu_ht_node *flag_removed(struct rcu_ht_node *node)
 {
-	return (struct rcu_ht_node *) (((unsigned long) node) | 0x1);
+	return (struct rcu_ht_node *) (((unsigned long) node) | REMOVED_FLAG);
+}
+
+static
+int is_dummy(struct rcu_ht_node *node)
+{
+	return ((unsigned long) node) & DUMMY_FLAG;
 }
 
 static
-void _uatomic_max(unsigned long *ptr, unsigned long v)
+struct rcu_ht_node *flag_dummy(struct rcu_ht_node *node)
+{
+	return (struct rcu_ht_node *) (((unsigned long) node) | DUMMY_FLAG);
+}
+ 
+static
+unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
 {
 	unsigned long old1, old2;
 
@@ -161,186 +332,298 @@ void _uatomic_max(unsigned long *ptr, unsigned long v)
 	do {
 		old2 = old1;
 		if (old2 >= v)
-			break;
+			return old2;
 	} while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
+	return v;
 }
 
+/*
+ * Remove all logically deleted nodes from a bucket up to a certain node key.
+ */
 static
-void _ht_add(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node)
+void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node)
 {
-	struct rcu_ht_node *iter_prev = NULL, *iter = NULL;
+	struct rcu_ht_node *iter_prev, *iter, *next, *new_next;
 
-	if (!t->size)
-		return;
 	for (;;) {
-		unsigned long chain_len = 0;
-
-		iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]);
-		//printf("iter prev %p hash %lu bucket %lu\n", iter_prev,
-		//	node->hash, node->hash & (t->size - 1));
-		assert(iter_prev);
-		assert(iter_prev->reverse_hash <= node->reverse_hash);
+		iter_prev = dummy;
+		/* 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 (;;) {
-			iter = clear_flag(rcu_dereference(iter_prev->next));
-			if (unlikely(!iter))
+			if (unlikely(!clear_flag(iter)))
+				return;
+			if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)
+				return;
+			next = rcu_dereference(clear_flag(iter)->p.next);
+			if (is_removed(next))
 				break;
-			if (iter->reverse_hash < node->reverse_hash)
-				break;
-			iter_prev = iter;
-			check_resize(ht, t, ++chain_len);
+			iter_prev = clear_flag(iter);
+			iter = next;
 		}
-		/* add in iter_prev->next */
-		if (is_removed(iter))
-			continue;
-		assert(node != iter);
-		node->next = iter;
+		assert(!is_removed(iter));
+		if (is_dummy(iter))
+			new_next = flag_dummy(clear_flag(next));
+		else
+			new_next = clear_flag(next);
+		(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
+	}
+}
+
+static
+struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t,
+			    struct rcu_ht_node *node, int unique, int dummy)
+{
+	struct rcu_ht_node *iter_prev, *iter, *next, *new_node, *new_next,
+			*dummy_node;
+	struct _rcu_ht_node *lookup;
+	unsigned long hash, index, order;
+
+	if (!t->size) {
+		assert(dummy);
+		node->p.next = flag_dummy(NULL);
+		return node;	/* Initial first add (head) */
+	}
+	hash = bit_reverse_ulong(node->p.reverse_hash);
+	for (;;) {
+		uint32_t chain_len = 0;
+
+		/*
+		 * 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 rcu_ht_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)))
+				goto insert;
+			if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)
+				goto insert;
+			next = rcu_dereference(clear_flag(iter)->p.next);
+			if (is_removed(next))
+				goto gc_node;
+			if (unique
+			    && !is_dummy(next)
+			    && !ht->compare_fct(node->key, node->key_len,
+						clear_flag(iter)->key,
+						clear_flag(iter)->key_len))
+				return clear_flag(iter);
+			/* 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);
+			iter_prev = clear_flag(iter);
+			iter = next;
+		}
+	insert:
+		assert(node != clear_flag(iter));
+		assert(!is_removed(iter_prev));
 		assert(iter_prev != node);
-		if (uatomic_cmpxchg(&iter_prev->next, iter, node) != iter)
-			continue;
-		break;
+		if (!dummy)
+			node->p.next = clear_flag(iter);
+		else
+			node->p.next = flag_dummy(clear_flag(iter));
+		if (is_dummy(iter))
+			new_node = flag_dummy(node);
+		else
+			new_node = node;
+		if (uatomic_cmpxchg(&iter_prev->p.next, iter,
+				    new_node) != iter)
+			continue;	/* retry */
+		else
+			goto gc_end;
+	gc_node:
+		assert(!is_removed(iter));
+		if (is_dummy(iter))
+			new_next = flag_dummy(clear_flag(next));
+		else
+			new_next = clear_flag(next);
+		(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 rcu_ht_node *) lookup;
+	_ht_gc_bucket(dummy_node, node);
+	return node;
 }
 
 static
 int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node)
 {
-	struct rcu_ht_node *iter_prev, *iter, *next, *old;
-	unsigned long chain_len;
-	int found, ret = 0;
+	struct rcu_ht_node *dummy, *next, *old;
+	struct _rcu_ht_node *lookup;
 	int flagged = 0;
+	unsigned long hash, index, order;
 
-retry:
-	chain_len = 0;
-	found = 0;
-	iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]);
-	assert(iter_prev);
-	assert(iter_prev->reverse_hash <= node->reverse_hash);
-	for (;;) {
-		iter = clear_flag(rcu_dereference(iter_prev->next));
-		if (unlikely(!iter))
-			break;
-		if (iter->reverse_hash < node->reverse_hash)
-			break;
-		if (iter == node) {
-			found = 1;
-			break;
-		}
-		iter_prev = iter;
-	} 
-	if (!found) {
-		ret = -ENOENT;
-		goto end;
-	}
-	next = rcu_dereference(iter->next);
-	if (!flagged) {
-		if (is_removed(next)) {
-			ret = -ENOENT;
+	/* logically delete the node */
+	old = rcu_dereference(node->p.next);
+	do {
+		next = old;
+		if (is_removed(next))
 			goto end;
-		}
-		/* set deletion flag */
-		if ((old = uatomic_cmpxchg(&iter->next, next,
-					   flag_removed(next))) != next) {
-			if (old == flag_removed(next)) {
-				ret = -ENOENT;
-				goto end;
-			} else {
-				goto retry;
-			}
-		}
-		flagged = 1;
-	}
+		assert(!is_dummy(next));
+		old = uatomic_cmpxchg(&node->p.next, next,
+				      flag_removed(next));
+	} while (old != next);
+
+	/* We performed the (logical) deletion. */
+	flagged = 1;
+
 	/*
-	 * Remove the element from the list. Retry if there has been a
-	 * concurrent add (there cannot be a concurrent delete, because
-	 * we won the deletion flag cmpxchg).
+	 * 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.
 	 */
-	if (uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next)) != iter)
-		goto retry;
+	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)];
+	dummy = (struct rcu_ht_node *) lookup;
+	_ht_gc_bucket(dummy, node);
 end:
-	return ret;
+	/*
+	 * Only the flagging action indicated that we (and no other)
+	 * removed the node from the hash.
+	 */
+	if (flagged) {
+		assert(is_removed(rcu_dereference(node->p.next)));
+		return 0;
+	} else
+		return -ENOENT;
 }
 
 static
 void init_table(struct rcu_ht *ht, struct rcu_table *t,
-		unsigned long first, unsigned long len)
-{
-	unsigned long i, end;
-
-	end = first + len;
-	for (i = first; i < end; i++) {
-		/* Update table size when power of two */
-		if (i != 0 && !(i & (i - 1)))
-			t->size = i;
-		t->tbl[i] = calloc(1, sizeof(struct rcu_ht_node));
-		t->tbl[i]->dummy = 1;
-		t->tbl[i]->hash = i;
-		t->tbl[i]->reverse_hash = bit_reverse_ulong(i);
-		_ht_add(ht, t, t->tbl[i]);
+		unsigned long first_order, unsigned long len_order)
+{
+	unsigned long i, end_order;
+
+	dbg_printf("rculfhash: 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;
+
+		len = !i ? 1 : 1UL << (i - 1);
+		dbg_printf("rculfhash: init order %lu len: %lu\n", i, len);
+		t->tbl[i] = calloc(len, sizeof(struct _rcu_ht_node));
+		for (j = 0; j < len; j++) {
+			dbg_printf("rculfhash: init entry: i %lu j %lu hash %lu\n",
+				   i, j, !i ? 0 : (1UL << (i - 1)) + j);
+			struct rcu_ht_node *new_node =
+				(struct rcu_ht_node *) &t->tbl[i][j];
+			new_node->p.reverse_hash =
+				bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
+			(void) _ht_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("rculfhash: init new size: %lu\n", t->size);
+		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
+			break;
 	}
-	t->size = end;
+	t->resize_target = t->size;
+	t->resize_initiated = 0;
 }
 
 struct rcu_ht *ht_new(ht_hash_fct hash_fct,
-		      void *hashseed,
+		      ht_compare_fct compare_fct,
+		      unsigned long hash_seed,
 		      unsigned long init_size,
 		      void (*ht_call_rcu)(struct rcu_head *head,
 				void (*func)(struct rcu_head *head)))
 {
 	struct rcu_ht *ht;
+	unsigned long order;
 
 	ht = calloc(1, sizeof(struct rcu_ht));
 	ht->hash_fct = hash_fct;
-	ht->hashseed = hashseed;
+	ht->compare_fct = compare_fct;
+	ht->hash_seed = hash_seed;
 	ht->ht_call_rcu = ht_call_rcu;
+	ht->in_progress_resize = 0;
 	/* 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 rcu_table)
-		       + (max(init_size, 1) * sizeof(struct rcu_ht_node *)));
+		       + (order * sizeof(struct _rcu_ht_node *)));
 	ht->t->size = 0;
 	pthread_mutex_lock(&ht->resize_mutex);
-	init_table(ht, ht->t, 0, max(init_size, 1));
+	init_table(ht, ht->t, 0, order);
 	pthread_mutex_unlock(&ht->resize_mutex);
-	ht->target_size = ht->t->size;
 	return ht;
 }
 
-struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key)
+struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len)
 {
 	struct rcu_table *t;
-	struct rcu_ht_node *node;
-	unsigned long hash, reverse_hash;
+	struct rcu_ht_node *node, *next;
+	struct _rcu_ht_node *lookup;
+	unsigned long hash, reverse_hash, index, order;
 
-	hash = ht->hash_fct(ht->hashseed, key);
+	hash = ht->hash_fct(key, key_len, ht->hash_seed);
 	reverse_hash = bit_reverse_ulong(hash);
 
 	t = rcu_dereference(ht->t);
-	node = rcu_dereference(t->tbl[hash & (t->size - 1)]);
+	index = hash & (t->size - 1);
+	order = get_count_order_ulong(index + 1);
+	lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
+	dbg_printf("rculfhash: lookup hash %lu index %lu order %lu aridx %lu\n",
+		   hash, index, order, index & ((1UL << (order - 1)) - 1));
+	node = (struct rcu_ht_node *) lookup;
 	for (;;) {
 		if (unlikely(!node))
 			break;
-		if (node->reverse_hash > reverse_hash) {
+		if (unlikely(node->p.reverse_hash > reverse_hash)) {
 			node = NULL;
 			break;
 		}
-		if (node->key == key) {
-			if (is_removed(rcu_dereference(node->next)))
-				node = 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(rcu_dereference(node->next));
+		node = clear_flag(next);
 	}
+	assert(!node || !is_dummy(rcu_dereference(node->p.next)));
 	return node;
 }
 
 void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node)
 {
 	struct rcu_table *t;
+	unsigned long hash;
+
+	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) _ht_add(ht, t, node, 0, 0);
+}
+
+struct rcu_ht_node *ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node)
+{
+	struct rcu_table *t;
+	unsigned long hash;
 
-	node->hash = ht->hash_fct(ht->hashseed, node->key);
-	node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash);
+	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);
-	_ht_add(ht, t, node);
+	return _ht_add(ht, t, node, 1, 0);
 }
 
 int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node)
@@ -356,20 +639,31 @@ int ht_delete_dummy(struct rcu_ht *ht)
 {
 	struct rcu_table *t;
 	struct rcu_ht_node *node;
-	unsigned long i;
+	struct _rcu_ht_node *lookup;
+	unsigned long order, i;
 
 	t = ht->t;
 	/* Check that the table is empty */
-	node = t->tbl[0];
+	lookup = &t->tbl[0][0];
+	node = (struct rcu_ht_node *) lookup;
 	do {
-		if (!node->dummy)
+		node = clear_flag(node)->p.next;
+		if (!is_dummy(node))
 			return -EPERM;
-		node = node->next;
-	} while (node);
+		assert(!is_removed(node));
+	} while (clear_flag(node));
 	/* Internal sanity check: all nodes left should be dummy */
-	for (i = 0; i < t->size; i++) {
-		assert(t->tbl[i]->dummy);
-		free(t->tbl[i]);
+	for (order = 0; order < get_count_order_ulong(t->size) + 1; order++) {
+		unsigned long len;
+
+		len = !order ? 1 : 1UL << (order - 1);
+		for (i = 0; i < len; i++) {
+			dbg_printf("rculfhash: 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));
+		}
+		free(t->tbl[order]);
 	}
 	return 0;
 }
@@ -382,6 +676,10 @@ int ht_destroy(struct rcu_ht *ht)
 {
 	int ret;
 
+	/* Wait for in-flight resize operations to complete */
+	CMM_STORE_SHARED(ht->in_progress_destroy, 1);
+	while (uatomic_read(&ht->in_progress_resize))
+		poll(NULL, 0, 100);	/* wait for 100ms */
 	ret = ht_delete_dummy(ht);
 	if (ret)
 		return ret;
@@ -390,6 +688,36 @@ int ht_destroy(struct rcu_ht *ht)
 	return ret;
 }
 
+void ht_count_nodes(struct rcu_ht *ht,
+		unsigned long *count,
+		unsigned long *removed)
+{
+	struct rcu_table *t;
+	struct rcu_ht_node *node, *next;
+	struct _rcu_ht_node *lookup;
+	unsigned long nr_dummy = 0;
+
+	*count = 0;
+	*removed = 0;
+
+	t = rcu_dereference(ht->t);
+	/* Count non-dummy nodes in the table */
+	lookup = &t->tbl[0][0];
+	node = (struct rcu_ht_node *) lookup;
+	do {
+		next = rcu_dereference(node->p.next);
+		if (is_removed(next)) {
+			assert(!is_dummy(next));
+			(*removed)++;
+		} else if (!is_dummy(next))
+			(*count)++;
+		else
+			(nr_dummy)++;
+		node = clear_flag(next);
+	} while (node);
+	dbg_printf("rculfhash: number of dummy nodes: %lu\n", nr_dummy);
+}
+
 static
 void ht_free_table_cb(struct rcu_head *head)
 {
@@ -402,48 +730,50 @@ void ht_free_table_cb(struct rcu_head *head)
 static
 void _do_ht_resize(struct rcu_ht *ht)
 {
-	unsigned long new_size, old_size;
+	unsigned long new_size, old_size, old_order, new_order;
 	struct rcu_table *new_t, *old_t;
 
-	//return; //TEST
-
 	old_t = ht->t;
 	old_size = old_t->size;
+	old_order = get_count_order_ulong(old_size) + 1;
 
-	new_size = CMM_LOAD_SHARED(ht->target_size);
+	new_size = CMM_LOAD_SHARED(old_t->resize_target);
 	if (old_size == new_size)
 		return;
+	new_order = get_count_order_ulong(new_size) + 1;
+	printf("rculfhash: resize from %lu (order %lu) to %lu (order %lu) buckets\n",
+	       old_size, old_order, new_size, new_order);
 	new_t = malloc(sizeof(struct rcu_table)
-			+ (new_size * sizeof(struct rcu_ht_node *)));
+			+ (new_order * sizeof(struct _rcu_ht_node *)));
 	assert(new_size > old_size);
 	memcpy(&new_t->tbl, &old_t->tbl,
-	       old_size * sizeof(struct rcu_ht_node *));
-	init_table(ht, new_t, old_size, new_size - old_size);
-	new_t->size = new_size;
+	       old_order * sizeof(struct _rcu_ht_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->ht_call_rcu(&old_t->head, ht_free_table_cb);
 }
 
 static
-void resize_target_update(struct rcu_ht *ht, struct rcu_table *t,
-			  int growth_order)
+unsigned long resize_target_update(struct rcu_table *t,
+				   int growth_order)
 {
-	unsigned long new_size = t->size << growth_order;
-
-	if (new_size > MAX_NR_BUCKETS)
-		new_size = MAX_NR_BUCKETS;
-	//printf("resize update prevtarget %lu current %lu order %d\n",
-	//	ht->target_size, t->size, growth_order);
-	_uatomic_max(&ht->target_size, new_size);
+	return _uatomic_max(&t->resize_target,
+			    t->size << growth_order);
 }
 
 void ht_resize(struct rcu_ht *ht, int growth)
 {
-	resize_target_update(ht, rcu_dereference(ht->t), growth);
-	pthread_mutex_lock(&ht->resize_mutex);
-	_do_ht_resize(ht);
-	pthread_mutex_unlock(&ht->resize_mutex);
+	struct rcu_table *t = rcu_dereference(ht->t);
+	unsigned long target_size;
+
+	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_ht_resize(ht);
+		pthread_mutex_unlock(&ht->resize_mutex);
+	}
 }
 
 static
@@ -457,15 +787,23 @@ void do_resize_cb(struct rcu_head *head)
 	_do_ht_resize(ht);
 	pthread_mutex_unlock(&ht->resize_mutex);
 	free(work);
+	cmm_smp_mb();	/* finish resize before decrement */
+	uatomic_dec(&ht->in_progress_resize);
 }
 
 static
 void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth)
 {
 	struct rcu_resize_work *work;
+	unsigned long target_size;
 
-	work = malloc(sizeof(*work));
-	work->ht = ht;
-	resize_target_update(ht, t, growth);
-	ht->ht_call_rcu(&work->head, do_resize_cb);
+	target_size = resize_target_update(t, growth);
+	if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) {
+		uatomic_inc(&ht->in_progress_resize);
+		cmm_smp_mb();	/* increment resize count before calling it */
+		work = malloc(sizeof(*work));
+		work->ht = ht;
+		ht->ht_call_rcu(&work->head, do_resize_cb);
+		CMM_STORE_SHARED(t->resize_initiated, 1);
+	}
 }