X-Git-Url: https://git.liburcu.org/?a=blobdiff_plain;f=rculfhash.c;h=ce1cbb79a55f4cc69ed729dddb170773a46d17d8;hb=7a9dcf9bf15667662e61b44c611a7f4c39b10dcf;hp=40f6ff59be9b5427a64d9d4a9b07524a10fe7c6c;hpb=c9edd44acae5da757017788908936d248a062199;p=urcu.git

diff --git a/rculfhash.c b/rculfhash.c
index 40f6ff5..ce1cbb7 100644
--- a/rculfhash.c
+++ b/rculfhash.c
@@ -169,7 +169,6 @@
 /*
  * Define the minimum table size.
  */
-//#define MIN_TABLE_SIZE			128
 #define MIN_TABLE_SIZE			1
 
 #if (CAA_BITS_PER_LONG == 32)
@@ -196,7 +195,7 @@
 #define FLAGS_MASK		((1UL << 2) - 1)
 
 /* Value of the end pointer. Should not interact with flags. */
-#define END_VALUE		0x4
+#define END_VALUE		NULL
 
 struct ht_items_count {
 	unsigned long add, remove;
@@ -695,7 +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;
 
@@ -707,15 +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)
@@ -725,11 +715,10 @@ int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
 		 */
 		assert(dummy != node);
 		for (;;) {
-			assert(iter != NULL);
 			if (unlikely(is_end(iter)))
-				return 0;
+				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;
@@ -741,10 +730,9 @@ int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
 			new_next = flag_dummy(clear_flag(next));
 		else
 			new_next = clear_flag(next);
-		assert(new_next != NULL);
 		(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
 	}
-	return 0;
+	return;
 }
 
 static
@@ -757,7 +745,6 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 			*dummy_node;
 	struct _cds_lfht_node *lookup;
 	unsigned long hash, index, order;
-	int force_dummy = 0;
 
 	assert(!is_dummy(node));
 	assert(!is_removed(node));
@@ -780,26 +767,8 @@ 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 (;;) {
-			assert(iter != NULL);
-			/*
-			 * 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(is_end(iter)))
 				goto insert;
 			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
@@ -825,31 +794,14 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 		assert(!is_removed(iter_prev));
 		assert(!is_removed(iter));
 		assert(iter_prev != node);
-		if (!dummy) {
+		if (!dummy)
 			node->p.next = clear_flag(iter);
-		} else {
-			/*
-			 * Dummy node insertion is performed concurrently (help
-			 * scheme). We try to link its next node, and if this
-			 * succeeds, it _means_ it's us who link this dummy node
-			 * into the table. force_dummy is set as soon as we
-			 * succeed this cmpxchg within this function.
-			 */
-			if (!force_dummy) {
-				if (uatomic_cmpxchg(&node->p.next, NULL,
-						flag_dummy(clear_flag(iter))) != NULL) {
-					return NULL;
-				}
-				force_dummy = 1;
-			} else {
-				node->p.next = flag_dummy(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;
-		assert(new_node != NULL);
 		if (uatomic_cmpxchg(&iter_prev->p.next, iter,
 				    new_node) != iter)
 			continue;	/* retry */
@@ -861,7 +813,6 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 			new_next = flag_dummy(clear_flag(next));
 		else
 			new_next = clear_flag(next);
-		assert(new_next != NULL);
 		(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
 		/* retry */
 	}
@@ -871,11 +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;
 }
 
@@ -901,7 +848,6 @@ int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size,
 			assert(is_dummy(next));
 		else
 			assert(!is_dummy(next));
-		assert(next != NULL);
 		old = uatomic_cmpxchg(&node->p.next, next,
 				      flag_removed(next));
 	} while (old != next);
@@ -914,18 +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)
@@ -938,32 +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;
 
@@ -973,8 +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))
@@ -1006,23 +936,17 @@ void init_table(struct cds_lfht *ht,
 		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))
@@ -1049,6 +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)
@@ -1121,7 +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,
@@ -1173,7 +1102,6 @@ struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key
 	hash = ht->hash_fct(key, key_len, ht->hash_seed);
 	reverse_hash = bit_reverse_ulong(hash);
 
-restart:
 	size = rcu_dereference(ht->t.size);
 	index = hash & (size - 1);
 	order = get_count_order_ulong(index + 1);
@@ -1183,16 +1111,6 @@ restart:
 	dummy_node = (struct cds_lfht_node *) lookup;
 	/* We can always skip the dummy node initially */
 	node = rcu_dereference(dummy_node->p.next);
-	if (unlikely(node == 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, dummy_node, 0, 1);
-		goto restart;	/* retry */
-	}
 	node = clear_flag(node);
 	for (;;) {
 		if (unlikely(is_end(node))) {
@@ -1273,7 +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;
 }