X-Git-Url: https://git.liburcu.org/?a=blobdiff_plain;f=rculfhash.c;h=ce1cbb79a55f4cc69ed729dddb170773a46d17d8;hb=7a9dcf9bf15667662e61b44c611a7f4c39b10dcf;hp=ce09a0f93261d9a5c2a76c0a530e7cca6118fe5f;hpb=76a73da8b055a8db8b067e992b59222b653bd87d;p=urcu.git

diff --git a/rculfhash.c b/rculfhash.c
index ce09a0f..ce1cbb7 100644
--- a/rculfhash.c
+++ b/rculfhash.c
@@ -169,7 +169,7 @@
 /*
  * Define the minimum table size.
  */
-#define MIN_TABLE_SIZE			128
+#define MIN_TABLE_SIZE			1
 
 #if (CAA_BITS_PER_LONG == 32)
 #define MAX_TABLE_ORDER			32
@@ -194,6 +194,9 @@
 #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, remove;
 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
@@ -652,7 +655,19 @@ 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)
 {
@@ -679,7 +694,7 @@ void cds_lfht_free_level(struct rcu_head *head)
  * Remove all logically deleted nodes from a bucket up to a certain node key.
  */
 static
-int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
+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;
 
@@ -691,15 +706,6 @@ int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
 		iter_prev = dummy;
 		/* We can always skip the dummy node initially */
 		iter = rcu_dereference(iter_prev->p.next);
-		if (unlikely(iter == NULL)) {
-			/*
-			 * We are executing concurrently with a hash table
-			 * expand, so we see a dummy node with NULL next value.
-			 * Help expand by linking this node into the list and
-			 * retry.
-			 */
-			return 1;
-		}
 		assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
 		/*
 		 * We should never be called with dummy (start of chain)
@@ -709,10 +715,10 @@ int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
 		 */
 		assert(dummy != node);
 		for (;;) {
-			if (unlikely(!clear_flag(iter)))
-				return 0;
+			if (unlikely(is_end(iter)))
+				return;
 			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
-				return 0;
+				return;
 			next = rcu_dereference(clear_flag(iter)->p.next);
 			if (likely(is_removed(next)))
 				break;
@@ -726,7 +732,7 @@ int _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 0;
+	return;
 }
 
 static
@@ -744,7 +750,7 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 	assert(!is_removed(node));
 	if (!size) {
 		assert(dummy);
-		node->p.next = flag_dummy(NULL);
+		node->p.next = flag_dummy(get_end());
 		return node;	/* Initial first add (head) */
 	}
 	hash = bit_reverse_ulong(node->p.reverse_hash);
@@ -761,26 +767,9 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 		iter_prev = (struct cds_lfht_node *) lookup;
 		/* We can always skip the dummy node initially */
 		iter = rcu_dereference(iter_prev->p.next);
-		if (unlikely(iter == NULL)) {
-			/*
-			 * We are executing concurrently with a hash table
-			 * expand, so we see a dummy node with NULL next value.
-			 * Help expand by linking this node into the list and
-			 * retry.
-			 */
-			(void) _cds_lfht_add(ht, size >> 1, iter_prev, 0, 1);
-			continue;	/* retry */
-		}
 		assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
 		for (;;) {
-			/*
-			 * When adding a dummy node, we allow concurrent
-			 * add/removal to help. If we find the dummy node in
-			 * place, skip its insertion.
-			 */
-			if (unlikely(dummy && clear_flag(iter) == node))
-				return node;
-			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;
@@ -833,11 +822,7 @@ gc_end:
 	order = get_count_order_ulong(index + 1);
 	lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
 	dummy_node = (struct cds_lfht_node *) lookup;
-	if (_cds_lfht_gc_bucket(dummy_node, node)) {
-		/* Help expand */
-		(void) _cds_lfht_add(ht, size >> 1, dummy_node, 0, 1);
-		goto gc_end;	/* retry */
-	}
+	_cds_lfht_gc_bucket(dummy_node, node);
 	return node;
 }
 
@@ -875,18 +860,13 @@ int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size,
 	 * the node, and remove it (along with any other logically removed node)
 	 * if found.
 	 */
-gc_retry:
 	hash = bit_reverse_ulong(node->p.reverse_hash);
 	assert(size > 0);
 	index = hash & (size - 1);
 	order = get_count_order_ulong(index + 1);
 	lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
 	dummy = (struct cds_lfht_node *) lookup;
-	if (_cds_lfht_gc_bucket(dummy, node)) {
-		/* Help expand */
-		(void) _cds_lfht_add(ht, size >> 1, dummy, 0, 1);
-		goto gc_retry;	/* retry */
-	}
+	_cds_lfht_gc_bucket(dummy, node);
 end:
 	/*
 	 * Only the flagging action indicated that we (and no other)
@@ -899,32 +879,19 @@ end:
 		return -ENOENT;
 }
 
-static
-void init_table_hash(struct cds_lfht *ht, unsigned long i,
-		unsigned long len)
-{
-	unsigned long j;
-
-	for (j = 0; j < len; j++) {
-		struct cds_lfht_node *new_node =
-			(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
-
-		dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
-			   i, j, !i ? 0 : (1UL << (i - 1)) + j);
-		new_node->p.reverse_hash =
-			bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
-		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
-			break;
-	}
-}
-
 /*
  * Holding RCU read lock to protect _cds_lfht_add against memory
  * reclaim that could be performed by other call_rcu worker threads (ABA
  * problem).
+ *
+ * TODO: 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_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
+void init_table_populate(struct cds_lfht *ht, unsigned long i, unsigned long len)
 {
 	unsigned long j;
 
@@ -934,8 +901,10 @@ void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
 		struct cds_lfht_node *new_node =
 			(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
 
-		dbg_printf("init link: i %lu j %lu hash %lu\n",
+		dbg_printf("init populate: i %lu j %lu hash %lu\n",
 			   i, j, !i ? 0 : (1UL << (i - 1)) + j);
+		new_node->p.reverse_hash =
+			bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
 		(void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
 				new_node, 0, 1);
 		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
@@ -959,26 +928,25 @@ void init_table(struct cds_lfht *ht,
 
 		len = !i ? 1 : 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)))
+			break;
+
 		ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
 				+ (len * sizeof(struct _cds_lfht_node)));
 
-		/* Set all dummy nodes reverse hash values for a level */
-		init_table_hash(ht, i, len);
-
 		/*
-		 * Update table size. At this point, concurrent add/remove see
-		 * dummy nodes with correctly initialized reverse hash value,
-		 * but with NULL next pointers. If they do, they can help us
-		 * link the dummy nodes into the list and retry.
+		 * Set all dummy nodes reverse hash values for a level and
+		 * link all dummy nodes into the table.
 		 */
-		cmm_smp_wmb();	/* populate data before RCU size */
-		CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
+		init_table_populate(ht, i, len);
 
 		/*
-		 * Link all dummy nodes into the table. Concurrent
-		 * add/remove are helping us.
+		 * Update table size.
 		 */
-		init_table_link(ht, i, len);
+		cmm_smp_wmb();	/* populate data before RCU size */
+		CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
 
 		dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
 		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
@@ -1005,6 +973,11 @@ void init_table(struct cds_lfht *ht,
  *
  * 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.
+ *
+ * TODO: 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(struct cds_lfht *ht, unsigned long i, unsigned long len)
@@ -1040,13 +1013,27 @@ void fini_table(struct cds_lfht *ht,
 		   first_order, first_order + len_order);
 	end_order = first_order + len_order;
 	assert(first_order > 0);
-	assert(ht->t.size == (1UL << (first_order - 1)));
 	for (i = end_order - 1; i >= first_order; i--) {
 		unsigned long len;
 
 		len = !i ? 1 : 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();
+
 		/*
 		 * Set "removed" flag in dummy nodes about to be removed.
 		 * Unlink all now-logically-removed dummy node pointers.
@@ -1063,7 +1050,7 @@ void fini_table(struct cds_lfht *ht,
 	}
 }
 
-struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
+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,
@@ -1099,6 +1086,7 @@ struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
 	ht->flags = flags;
 	ht->cds_lfht_rcu_thread_offline();
 	pthread_mutex_lock(&ht->resize_mutex);
+	ht->t.resize_target = 1UL << (order - 1);
 	init_table(ht, 0, order);
 	pthread_mutex_unlock(&ht->resize_mutex);
 	ht->cds_lfht_rcu_thread_online();
@@ -1107,7 +1095,7 @@ struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
 
 struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
 {
-	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, size;
 
@@ -1120,10 +1108,15 @@ struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key
 	lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
 	dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
 		   hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
-	node = (struct cds_lfht_node *) lookup;
+	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 = NULL;
 			break;
+		}
 		if (unlikely(node->p.reverse_hash > reverse_hash)) {
 			node = NULL;
 			break;
@@ -1155,8 +1148,10 @@ struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
 	node = clear_flag(next);
 
 	for (;;) {
-		if (unlikely(!node))
+		if (unlikely(is_end(node))) {
+			node = NULL;
 			break;
+		}
 		if (unlikely(node->p.reverse_hash > reverse_hash)) {
 			node = NULL;
 			break;
@@ -1196,7 +1191,7 @@ struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
 
 	size = rcu_dereference(ht->t.size);
 	ret = _cds_lfht_add(ht, size, node, 1, 0);
-	if (ret != node)
+	if (ret == node)
 		ht_count_add(ht, size);
 	return ret;
 }
@@ -1228,7 +1223,7 @@ int cds_lfht_delete_dummy(struct cds_lfht *ht)
 		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.
@@ -1294,7 +1289,7 @@ void cds_lfht_count_nodes(struct cds_lfht *ht,
 		else
 			(nr_dummy)++;
 		node = clear_flag(next);
-	} while (node);
+	} while (!is_end(node));
 	dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
 }
 
@@ -1327,17 +1322,6 @@ void _do_cds_lfht_shrink(struct cds_lfht *ht,
 	       old_size, old_order, new_size, new_order);
 	assert(new_size < old_size);
 
-	cmm_smp_wmb();	/* populate data before RCU size */
-	CMM_STORE_SHARED(ht->t.size, new_size);
-
-	/*
-	 * 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();
-
 	/* Remove and unlink all dummy nodes to remove. */
 	fini_table(ht, new_order, old_order - new_order);
 }
@@ -1363,7 +1347,7 @@ void _do_cds_lfht_resize(struct cds_lfht *ht)
 		ht->t.resize_initiated = 0;
 		/* write resize_initiated before read resize_target */
 		cmm_smp_mb();
-	} while (new_size != CMM_LOAD_SHARED(ht->t.resize_target));
+	} while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
 }
 
 static