X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=1487980c832c115154937cda4b185024636f9e44;hp=09840f2f75694529c17ab2ede16be0ec3b475bda;hb=b198f0fda77c816d80f63cfa2e71a7e0b4496736;hpb=dc1da8f6d538b4b56326617b391182f6bec36b59

diff --git a/rculfhash.c b/rculfhash.c
index 09840f2..1487980 100644
--- a/rculfhash.c
+++ b/rculfhash.c
@@ -177,6 +177,12 @@
 #define MAX_TABLE_ORDER			64
 #endif
 
+/*
+ * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
+ */
+#define MIN_PARTITION_PER_THREAD_ORDER	12
+#define MIN_PARTITION_PER_THREAD	(1UL << MIN_PARTITION_PER_THREAD_ORDER)
+
 #ifndef min
 #define min(a, b)	((a) < (b) ? (a) : (b))
 #endif
@@ -187,6 +193,8 @@
 
 /*
  * The removed flag needs to be updated atomically with the pointer.
+ * It indicates that no node must attach to the node scheduled for
+ * removal, and that node garbage collection must be performed.
  * 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.
  */
@@ -198,7 +206,7 @@
 #define END_VALUE		NULL
 
 struct ht_items_count {
-	unsigned long add, remove;
+	unsigned long add, del;
 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
 
 struct rcu_level {
@@ -235,6 +243,9 @@ struct cds_lfht {
 	void (*cds_lfht_rcu_read_unlock)(void);
 	void (*cds_lfht_rcu_thread_offline)(void);
 	void (*cds_lfht_rcu_thread_online)(void);
+	void (*cds_lfht_rcu_register_thread)(void);
+	void (*cds_lfht_rcu_unregister_thread)(void);
+	pthread_attr_t *resize_attr;	/* Resize threads attributes */
 	unsigned long count;		/* global approximate item count */
 	struct ht_items_count *percpu_count;	/* per-cpu item count */
 };
@@ -244,11 +255,25 @@ struct rcu_resize_work {
 	struct cds_lfht *ht;
 };
 
+struct partition_resize_work {
+	struct rcu_head head;
+	struct cds_lfht *ht;
+	unsigned long i, start, len;
+	void (*fct)(struct cds_lfht *ht, unsigned long i,
+		    unsigned long start, unsigned long len);
+};
+
+enum add_mode {
+	ADD_DEFAULT = 0,
+	ADD_UNIQUE = 1,
+	ADD_REPLACE = 2,
+};
+
 static
 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 				unsigned long size,
 				struct cds_lfht_node *node,
-				int unique, int dummy);
+				enum add_mode mode, int dummy);
 
 /*
  * Algorithm to reverse bits in a word by lookup table, extended to
@@ -548,7 +573,7 @@ void ht_count_add(struct cds_lfht *ht, unsigned long size)
 }
 
 static
-void ht_count_remove(struct cds_lfht *ht, unsigned long size)
+void ht_count_del(struct cds_lfht *ht, unsigned long size)
 {
 	unsigned long percpu_count;
 	int cpu;
@@ -558,18 +583,18 @@ void ht_count_remove(struct cds_lfht *ht, unsigned long size)
 	cpu = ht_get_cpu();
 	if (unlikely(cpu < 0))
 		return;
-	percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1);
+	percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, -1);
 	if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
 		unsigned long count;
 
-		dbg_printf("remove percpu %lu\n", percpu_count);
+		dbg_printf("del percpu %lu\n", percpu_count);
 		count = uatomic_add_return(&ht->count,
 					   -(1UL << COUNT_COMMIT_ORDER));
 		/* If power of 2 */
 		if (!(count & (count - 1))) {
 			if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
 				return;
-			dbg_printf("remove set global %lu\n", count);
+			dbg_printf("del set global %lu\n", count);
 			cds_lfht_resize_lazy_count(ht, size,
 				count >> (CHAIN_LEN_TARGET - 1));
 		}
@@ -597,7 +622,7 @@ void ht_count_add(struct cds_lfht *ht, unsigned long size)
 }
 
 static
-void ht_count_remove(struct cds_lfht *ht, unsigned long size)
+void ht_count_del(struct cds_lfht *ht, unsigned long size)
 {
 }
 
@@ -730,6 +755,8 @@ 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);
+		if (is_removed(iter))
+			new_next = flag_removed(new_next);
 		(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
 	}
 	return;
@@ -739,10 +766,10 @@ static
 struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 				unsigned long size,
 				struct cds_lfht_node *node,
-				int unique, int dummy)
+				enum add_mode mode, int dummy)
 {
 	struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
-			*dummy_node;
+			*dummy_node, *return_node;
 	struct _cds_lfht_node *lookup;
 	unsigned long hash, index, order;
 
@@ -776,12 +803,16 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 			next = rcu_dereference(clear_flag(iter)->p.next);
 			if (unlikely(is_removed(next)))
 				goto gc_node;
-			if (unique
+			if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
 			    && !is_dummy(next)
 			    && !ht->compare_fct(node->key, node->key_len,
 						clear_flag(iter)->key,
-						clear_flag(iter)->key_len))
-				return clear_flag(iter);
+						clear_flag(iter)->key_len)) {
+				if (mode == ADD_UNIQUE)
+					return clear_flag(iter);
+				else /* mode == ADD_REPLACE */
+					goto replace;
+			}
 			/* Only account for identical reverse hash once */
 			if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
 			    && !is_dummy(next))
@@ -789,6 +820,7 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 			iter_prev = clear_flag(iter);
 			iter = next;
 		}
+
 	insert:
 		assert(node != clear_flag(iter));
 		assert(!is_removed(iter_prev));
@@ -803,17 +835,55 @@ struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
 		else
 			new_node = node;
 		if (uatomic_cmpxchg(&iter_prev->p.next, iter,
-				    new_node) != iter)
+				    new_node) != iter) {
 			continue;	/* retry */
+		} else {
+			if (mode == ADD_REPLACE)
+				return_node = NULL;
+			else	/* ADD_DEFAULT and ADD_UNIQUE */
+				return_node = node;
+			goto gc_end;
+		}
+
+	replace:
+		/* Insert after node to be replaced */
+		iter_prev = clear_flag(iter);
+		iter = next;
+		assert(node != clear_flag(iter));
+		assert(!is_removed(iter_prev));
+		assert(!is_removed(iter));
+		assert(iter_prev != node);
+		assert(!dummy);
+		node->p.next = clear_flag(iter);
+		if (is_dummy(iter))
+			new_node = flag_dummy(node);
 		else
+			new_node = node;
+		/*
+		 * Here is the whole trick for lock-free replace: we add
+		 * the replacement node _after_ the node we want to
+		 * replace by atomically setting its next pointer at the
+		 * same time we set its removal flag. Given that
+		 * the lookups/get next use an iterator aware of the
+		 * next pointer, they will either skip the old node due
+		 * to the removal flag and see the new node, or use
+		 * the old node, but will not see the new one.
+		 */
+		new_node = flag_removed(new_node);
+		if (uatomic_cmpxchg(&iter_prev->p.next,
+			      iter, new_node) != iter) {
+			continue;	/* retry */
+		} else {
+			return_node = iter_prev;
 			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);
-		assert(new_next != NULL);
 		(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
 		/* retry */
 	}
@@ -824,11 +894,11 @@ gc_end:
 	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);
-	return node;
+	return return_node;
 }
 
 static
-int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size,
+int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
 		struct cds_lfht_node *node,
 		int dummy_removal)
 {
@@ -842,6 +912,8 @@ int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size,
 	assert(!is_removed(node));
 	old = rcu_dereference(node->p.next);
 	do {
+		struct cds_lfht_node *new_next;
+
 		next = old;
 		if (unlikely(is_removed(next)))
 			goto end;
@@ -849,9 +921,8 @@ 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));
+		new_next = flag_removed(next);
+		old = uatomic_cmpxchg(&node->p.next, next, new_next);
 	} while (old != next);
 
 	/* We performed the (logical) deletion. */
@@ -881,19 +952,77 @@ end:
 		return -ENOENT;
 }
 
+static
+void *partition_resize_thread(void *arg)
+{
+	struct partition_resize_work *work = arg;
+
+	work->ht->cds_lfht_rcu_register_thread();
+	work->fct(work->ht, work->i, work->start, work->len);
+	work->ht->cds_lfht_rcu_unregister_thread();
+	return NULL;
+}
+
+static
+void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
+		unsigned long len,
+		void (*fct)(struct cds_lfht *ht, unsigned long i,
+			unsigned long start, unsigned long len))
+{
+	unsigned long partition_len;
+	struct partition_resize_work *work;
+	int thread, ret;
+	unsigned long nr_threads;
+	pthread_t *thread_id;
+
+	/*
+	 * Note: nr_cpus_mask + 1 is always power of 2.
+	 * We spawn just the number of threads we need to satisfy the minimum
+	 * partition size, up to the number of CPUs in the system.
+	 */
+	nr_threads = min(nr_cpus_mask + 1,
+			 len >> MIN_PARTITION_PER_THREAD_ORDER);
+	partition_len = len >> get_count_order_ulong(nr_threads);
+	work = calloc(nr_threads, sizeof(*work));
+	thread_id = calloc(nr_threads, sizeof(*thread_id));
+	assert(work);
+	for (thread = 0; thread < nr_threads; thread++) {
+		work[thread].ht = ht;
+		work[thread].i = i;
+		work[thread].len = partition_len;
+		work[thread].start = thread * partition_len;
+		work[thread].fct = fct;
+		ret = pthread_create(&thread_id[thread], ht->resize_attr,
+			partition_resize_thread, &work[thread]);
+		assert(!ret);
+	}
+	for (thread = 0; thread < nr_threads; thread++) {
+		ret = pthread_join(thread_id[thread], NULL);
+		assert(!ret);
+	}
+	free(work);
+	free(thread_id);
+}
+
 /*
  * Holding RCU read lock to protect _cds_lfht_add against memory
  * reclaim that could be performed by other call_rcu worker threads (ABA
  * problem).
+ *
+ * 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_populate(struct cds_lfht *ht, unsigned long i, unsigned long len)
+void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
+				   unsigned long start, unsigned long len)
 {
 	unsigned long j;
 
-	ht->cds_lfht_rcu_thread_online();
 	ht->cds_lfht_rcu_read_lock();
-	for (j = 0; j < len; j++) {
+	for (j = start; j < start + len; j++) {
 		struct cds_lfht_node *new_node =
 			(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
 
@@ -902,12 +1031,25 @@ void init_table_populate(struct cds_lfht *ht, unsigned long i, unsigned long len
 		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);
+				new_node, ADD_DEFAULT, 1);
 		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
 			break;
 	}
 	ht->cds_lfht_rcu_read_unlock();
-	ht->cds_lfht_rcu_thread_offline();
+}
+
+static
+void init_table_populate(struct cds_lfht *ht, unsigned long i,
+			 unsigned long len)
+{
+	assert(nr_cpus_mask != -1);
+	if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
+		ht->cds_lfht_rcu_thread_online();
+		init_table_populate_partition(ht, i, 0, len);
+		ht->cds_lfht_rcu_thread_offline();
+		return;
+	}
+	partition_resize_helper(ht, i, len, init_table_populate_partition);
 }
 
 static
@@ -931,6 +1073,7 @@ void init_table(struct cds_lfht *ht,
 
 		ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
 				+ (len * sizeof(struct _cds_lfht_node)));
+		assert(ht->t.tbl[i]);
 
 		/*
 		 * Set all dummy nodes reverse hash values for a level and
@@ -969,15 +1112,20 @@ 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.
+ *
+ * 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)
+void remove_table_partition(struct cds_lfht *ht, unsigned long i,
+			    unsigned long start, unsigned long len)
 {
 	unsigned long j;
 
-	ht->cds_lfht_rcu_thread_online();
 	ht->cds_lfht_rcu_read_lock();
-	for (j = 0; j < len; j++) {
+	for (j = start; j < start + len; j++) {
 		struct cds_lfht_node *fini_node =
 			(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
 
@@ -985,13 +1133,26 @@ void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
 			   i, j, !i ? 0 : (1UL << (i - 1)) + j);
 		fini_node->p.reverse_hash =
 			bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
-		(void) _cds_lfht_remove(ht, !i ? 0 : (1UL << (i - 1)),
+		(void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
 				fini_node, 1);
 		if (CMM_LOAD_SHARED(ht->in_progress_destroy))
 			break;
 	}
 	ht->cds_lfht_rcu_read_unlock();
-	ht->cds_lfht_rcu_thread_offline();
+}
+
+static
+void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
+{
+
+	assert(nr_cpus_mask != -1);
+	if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
+		ht->cds_lfht_rcu_thread_online();
+		remove_table_partition(ht, i, 0, len);
+		ht->cds_lfht_rcu_thread_offline();
+		return;
+	}
+	partition_resize_helper(ht, i, len, remove_table_partition);
 }
 
 static
@@ -1041,7 +1202,7 @@ 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,
@@ -1052,7 +1213,10 @@ struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
 			void (*cds_lfht_rcu_read_lock)(void),
 			void (*cds_lfht_rcu_read_unlock)(void),
 			void (*cds_lfht_rcu_thread_offline)(void),
-			void (*cds_lfht_rcu_thread_online)(void))
+			void (*cds_lfht_rcu_thread_online)(void),
+			void (*cds_lfht_rcu_register_thread)(void),
+			void (*cds_lfht_rcu_unregister_thread)(void),
+			pthread_attr_t *attr)
 {
 	struct cds_lfht *ht;
 	unsigned long order;
@@ -1061,6 +1225,7 @@ struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
 	if (init_size && (init_size & (init_size - 1)))
 		return NULL;
 	ht = calloc(1, sizeof(struct cds_lfht));
+	assert(ht);
 	ht->hash_fct = hash_fct;
 	ht->compare_fct = compare_fct;
 	ht->hash_seed = hash_seed;
@@ -1070,6 +1235,9 @@ struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
 	ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
 	ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
 	ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
+	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();
 	/* this mutex should not nest in read-side C.S. */
 	pthread_mutex_init(&ht->resize_mutex, NULL);
@@ -1084,7 +1252,8 @@ struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
 	return ht;
 }
 
-struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
+void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
+		struct cds_lfht_iter *iter)
 {
 	struct cds_lfht_node *node, *next, *dummy_node;
 	struct _cds_lfht_node *lookup;
@@ -1121,21 +1290,22 @@ struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key
 		node = clear_flag(next);
 	}
 	assert(!node || !is_dummy(rcu_dereference(node->p.next)));
-	return node;
+	iter->node = node;
+	iter->next = next;
 }
 
-struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
-				struct cds_lfht_node *node)
+void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
 {
-	struct cds_lfht_node *next;
+	struct cds_lfht_node *node, *next;
 	unsigned long reverse_hash;
 	void *key;
 	size_t key_len;
 
+	node = iter->node;
 	reverse_hash = node->p.reverse_hash;
 	key = node->key;
 	key_len = node->key_len;
-	next = rcu_dereference(node->p.next);
+	next = iter->next;
 	node = clear_flag(next);
 
 	for (;;) {
@@ -1156,7 +1326,8 @@ struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
 		node = clear_flag(next);
 	}
 	assert(!node || !is_dummy(rcu_dereference(node->p.next)));
-	return node;
+	iter->node = node;
+	iter->next = next;
 }
 
 void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
@@ -1167,12 +1338,12 @@ void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
 	node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
 
 	size = rcu_dereference(ht->t.size);
-	(void) _cds_lfht_add(ht, size, node, 0, 0);
+	(void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
 	ht_count_add(ht, size);
 }
 
 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
-					struct cds_lfht_node *node)
+				struct cds_lfht_node *node)
 {
 	unsigned long hash, size;
 	struct cds_lfht_node *ret;
@@ -1181,21 +1352,37 @@ struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
 	node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
 
 	size = rcu_dereference(ht->t.size);
-	ret = _cds_lfht_add(ht, size, node, 1, 0);
+	ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
 	if (ret == node)
 		ht_count_add(ht, size);
 	return ret;
 }
 
-int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node)
+struct cds_lfht_node *cds_lfht_replace(struct cds_lfht *ht,
+				struct cds_lfht_node *node)
+{
+	unsigned long hash, size;
+	struct cds_lfht_node *ret;
+
+	hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
+	node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
+
+	size = rcu_dereference(ht->t.size);
+	ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
+	if (ret == NULL)
+		ht_count_add(ht, size);
+	return ret;
+}
+
+int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
 {
 	unsigned long size;
 	int ret;
 
 	size = rcu_dereference(ht->t.size);
-	ret = _cds_lfht_remove(ht, size, node, 0);
+	ret = _cds_lfht_del(ht, size, node, 0);
 	if (!ret)
-		ht_count_remove(ht, size);
+		ht_count_del(ht, size);
 	return ret;
 }
 
@@ -1240,7 +1427,7 @@ int cds_lfht_delete_dummy(struct cds_lfht *ht)
  * Should only be called when no more concurrent readers nor writers can
  * possibly access the table.
  */
-int cds_lfht_destroy(struct cds_lfht *ht)
+int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
 {
 	int ret;
 
@@ -1252,6 +1439,8 @@ int cds_lfht_destroy(struct cds_lfht *ht)
 	if (ret)
 		return ret;
 	free_per_cpu_items_count(ht->percpu_count);
+	if (attr)
+		*attr = ht->resize_attr;
 	poison_free(ht);
 	return ret;
 }