X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=ebefa92ad8460c0d49501b67261b9087832167e0;hp=675ef382311beafcbb8d62ad284b2032d6db79aa;hb=6887cc5e59ed0275616771957b4158b68c5d5d9a;hpb=5488222bbdb95680d942c3fb645095b48e96620f

diff --git a/rculfhash.c b/rculfhash.c
index 675ef38..ebefa92 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
 
@@ -265,7 +266,7 @@ 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 */
 };
 
 struct rcu_resize_work {
@@ -453,7 +454,7 @@ 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);
@@ -485,10 +486,12 @@ 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)
@@ -497,85 +500,101 @@ int get_count_order_ulong(unsigned long x)
 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_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;
+		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 (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 (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 */
@@ -590,21 +609,19 @@ 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))
+	if (unlikely(!ht->split_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)))) {
+	index = ht_get_split_count_index(hash);
+	split_count = uatomic_add_return(&ht->split_count[index].del, 1);
+	if (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 */
@@ -616,7 +633,7 @@ 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));
@@ -624,34 +641,6 @@ 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)
 {
@@ -738,17 +727,21 @@ 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)];
 }
 
 /*
@@ -1280,7 +1273,7 @@ void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
 	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] = (void *)(ht->t.tbl[0]->nodes + len);
+			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]);
@@ -1337,6 +1330,7 @@ struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
 	init_size = max(init_size, min_alloc_size);
 	ht = calloc(1, sizeof(struct cds_lfht));
 	assert(ht);
+	ht->flags = flags;
 	ht->hash_fct = hash_fct;
 	ht->compare_fct = compare_fct;
 	ht->hash_seed = hash_seed;
@@ -1349,16 +1343,15 @@ 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);
-	ht->flags = flags;
 	order = get_count_order_ulong(init_size);
 	ht->t.resize_target = 1UL << order;
-	cds_lfht_create_dummy(ht, 1UL << order);
-	ht->t.size = 1UL << order;
 	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;
 }
 
@@ -1481,7 +1474,7 @@ void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
 
 	size = rcu_dereference(ht->t.size);
 	_cds_lfht_add(ht, size, node, NULL, 0);
-	ht_count_add(ht, size);
+	ht_count_add(ht, size, hash);
 }
 
 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
@@ -1496,7 +1489,7 @@ struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
 	size = rcu_dereference(ht->t.size);
 	_cds_lfht_add(ht, size, node, &iter, 0);
 	if (iter.node == node)
-		ht_count_add(ht, size);
+		ht_count_add(ht, size, hash);
 	return iter.node;
 }
 
@@ -1513,7 +1506,7 @@ struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
 	for (;;) {
 		_cds_lfht_add(ht, size, node, &iter, 0);
 		if (iter.node == node) {
-			ht_count_add(ht, size);
+			ht_count_add(ht, size, hash);
 			return NULL;
 		}
 
@@ -1534,13 +1527,15 @@ 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;
 }
 
@@ -1602,7 +1597,7 @@ 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);
@@ -1620,12 +1615,12 @@ 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);
 		}
 	}
 
@@ -1650,12 +1645,12 @@ 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);
 		}
 	}
 }
@@ -1787,8 +1782,6 @@ 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_count(struct cds_lfht *ht, unsigned long size,
 				unsigned long count)
@@ -1813,5 +1806,3 @@ void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
 		CMM_STORE_SHARED(ht->t.resize_initiated, 1);
 	}
 }
-
-#endif