X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=197a00f14984d09d8d3e08068a8f7bc89318b636;hp=f95ef5dc083aa4e22618512708beb37cee22e1e3;hb=bb7b2f262ba9a888271c31f48d8c4331d33f2376;hpb=5f51139190536f948f9571fdf3c97cf0198356d0

diff --git a/rculfhash.c b/rculfhash.c
index f95ef5d..197a00f 100644
--- a/rculfhash.c
+++ b/rculfhash.c
@@ -167,12 +167,7 @@
 #define CHAIN_LEN_RESIZE_THRESHOLD	3
 
 /*
- * Define the minimum table size. Protects against hash table resize overload
- * when too many entries are added quickly before the resize can complete.
- * This is especially the case if the table could be shrinked to a size of 1.
- * TODO: we might want to make the add/remove operations help the resize to
- * add or remove dummy nodes when a resize is ongoing to ensure upper-bound on
- * chain length.
+ * Define the minimum table size.
  */
 #define MIN_TABLE_SIZE			128
 
@@ -199,6 +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		0x4
+
 struct ht_items_count {
 	unsigned long add, remove;
 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
@@ -246,6 +244,12 @@ struct rcu_resize_work {
 	struct cds_lfht *ht;
 };
 
+static
+struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
+				unsigned long size,
+				struct cds_lfht_node *node,
+				int unique, int dummy);
+
 /*
  * Algorithm to reverse bits in a word by lookup table, extended to
  * 64-bit words.
@@ -651,7 +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)
 {
@@ -678,7 +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
-void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
+int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
 {
 	struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
 
@@ -690,6 +706,15 @@ void _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)
@@ -699,10 +724,11 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node
 		 */
 		assert(dummy != node);
 		for (;;) {
-			if (unlikely(!clear_flag(iter)))
-				return;
+			assert(iter != NULL);
+			if (unlikely(is_end(iter)))
+				return 0;
 			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
-				return;
+				return 0;
 			next = rcu_dereference(clear_flag(iter)->p.next);
 			if (likely(is_removed(next)))
 				break;
@@ -714,8 +740,10 @@ void _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;
 }
 
 static
@@ -728,12 +756,13 @@ 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));
 	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);
@@ -750,12 +779,28 @@ 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 (;;) {
-			/* TODO: check if removed */
-			if (unlikely(!clear_flag(iter)))
+			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;
-			/* TODO: check if removed */
 			if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
 				goto insert;
 			next = rcu_dereference(clear_flag(iter)->p.next);
@@ -779,14 +824,31 @@ 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
-			node->p.next = flag_dummy(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));
+			}
+		}
 		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 */
@@ -798,6 +860,7 @@ 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 */
 	}
@@ -807,7 +870,11 @@ 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;
-	_cds_lfht_gc_bucket(dummy_node, node);
+	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 */
+	}
 	return node;
 }
 
@@ -833,6 +900,7 @@ 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);
@@ -845,13 +913,18 @@ 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;
-	_cds_lfht_gc_bucket(dummy, node);
+	if (_cds_lfht_gc_bucket(dummy, node)) {
+		/* Help expand */
+		(void) _cds_lfht_add(ht, size >> 1, dummy, 0, 1);
+		goto gc_retry;	/* retry */
+	}
 end:
 	/*
 	 * Only the flagging action indicated that we (and no other)
@@ -883,6 +956,11 @@ void init_table_hash(struct cds_lfht *ht, unsigned long i,
 	}
 }
 
+/*
+ * Holding RCU read lock to protect _cds_lfht_add against memory
+ * reclaim that could be performed by other call_rcu worker threads (ABA
+ * problem).
+ */
 static
 void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
 {
@@ -905,11 +983,6 @@ void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
 	ht->cds_lfht_rcu_thread_offline();
 }
 
-/*
- * Holding RCU read lock to protect _cds_lfht_add against memory
- * reclaim that could be performed by other call_rcu worker threads (ABA
- * problem).
- */
 static
 void init_table(struct cds_lfht *ht,
 		unsigned long first_order, unsigned long len_order)
@@ -924,30 +997,58 @@ 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.
+		 */
+		cmm_smp_wmb();	/* populate data before RCU size */
+		CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
+
 		/*
 		 * Link all dummy nodes into the table. Concurrent
 		 * add/remove are helping us.
 		 */
 		init_table_link(ht, i, len);
 
-		/*
-		 * Update table size (after init for now, because no
-		 * concurrent updater help (TODO)).
-		 */
-		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))
 			break;
 	}
 }
 
+/*
+ * Holding RCU read lock to protect _cds_lfht_remove against memory
+ * reclaim that could be performed by other call_rcu worker threads (ABA
+ * problem).
+ * For a single level, we logically remove and garbage collect each node.
+ *
+ * As a design choice, we perform logical removal and garbage collection on a
+ * node-per-node basis to simplify this algorithm. We also assume keeping good
+ * cache locality of the operation would overweight possible performance gain
+ * that could be achieved by batching garbage collection for multiple levels.
+ * However, this would have to be justified by benchmarks.
+ *
+ * Concurrent removal and add operations are helping us perform garbage
+ * collection of logically removed nodes. We guarantee that all logically
+ * removed nodes have been garbage-collected (unlinked) before call_rcu is
+ * invoked to free a hole level of dummy nodes (after a grace period).
+ *
+ * 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.
+ */
 static
 void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
 {
@@ -972,11 +1073,6 @@ void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
 	ht->cds_lfht_rcu_thread_offline();
 }
 
-/*
- * Holding RCU read lock to protect _cds_lfht_remove against memory
- * reclaim that could be performed by other call_rcu worker threads (ABA
- * problem).
- */
 static
 void fini_table(struct cds_lfht *ht,
 		unsigned long first_order, unsigned long len_order)
@@ -987,13 +1083,27 @@ 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.
@@ -1046,6 +1156,7 @@ 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();
@@ -1054,23 +1165,39 @@ 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;
 
 	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);
 	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);
+	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(!node))
+		if (unlikely(is_end(node))) {
+			node = NULL;
 			break;
+		}
 		if (unlikely(node->p.reverse_hash > reverse_hash)) {
 			node = NULL;
 			break;
@@ -1102,8 +1229,10 @@ 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;
@@ -1175,7 +1304,7 @@ 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.
@@ -1241,7 +1370,7 @@ 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);
 }
 
@@ -1274,17 +1403,6 @@ 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);
 }
@@ -1310,7 +1428,7 @@ 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