X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-defer.c;h=c28e8488354e4138f24114200258eb3abdf7f3d3;hp=5e4650ff003b4c8d97ff0fb240a0d14fa4d8660a;hb=06f22bdbb0c4c4d5db42a2e2dc35818aa61415be;hpb=2c22932bff02d0124579d70c67695ce2a439006a

diff --git a/urcu-defer.c b/urcu-defer.c
index 5e4650f..c28e848 100644
--- a/urcu-defer.c
+++ b/urcu-defer.c
@@ -3,7 +3,7 @@
  *
  * Userspace RCU library - batch memory reclamation
  *
- * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  *
  * This library is free software; you can redistribute it and/or
  * modify it under the terms of the GNU Lesser General Public
@@ -28,72 +28,36 @@
 #include <string.h>
 #include <errno.h>
 #include <poll.h>
-#include <linux/futex.h>
 #include <sys/time.h>
 #include <syscall.h>
 #include <unistd.h>
 
+#include "urcu/urcu-futex.h"
 #include "urcu-defer-static.h"
 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 #include "urcu-defer.h"
 
-#define futex(...)	syscall(__NR_futex, __VA_ARGS__)
-
-void __attribute__((destructor)) urcu_defer_exit(void);
+void __attribute__((destructor)) rcu_defer_exit(void);
 
 extern void synchronize_rcu(void);
 
 /*
- * urcu_defer_mutex nests inside defer_thread_mutex.
+ * rcu_defer_mutex nests inside defer_thread_mutex.
  */
-static pthread_mutex_t urcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
+static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
 static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
 
-int defer_thread_futex;
+static int defer_thread_futex;
 
 /*
  * Written to only by each individual deferer. Read by both the deferer and
  * the reclamation tread.
  */
-struct defer_queue __thread defer_queue;
-
-/* Thread IDs of registered deferers */
-#define INIT_NUM_THREADS 4
-
-struct deferer_registry {
-	pthread_t tid;
-	struct defer_queue *defer_queue;
-	unsigned long last_head;
-};
-
-static struct deferer_registry *registry;
-static int num_deferers, alloc_deferers;
-
+static struct defer_queue __thread defer_queue;
+static LIST_HEAD(registry);
 static pthread_t tid_defer;
 
-/*
- * Wake-up any waiting defer thread. Called from many concurrent threads.
- */
-static void wake_up_defer(void)
-{
-	if (unlikely(atomic_read(&defer_thread_futex) == -1))
-		atomic_set(&defer_thread_futex, 0);
-		futex(&defer_thread_futex, FUTEX_WAKE,
-		      0, NULL, NULL, 0);
-}
-
-/*
- * Defer thread waiting. Single thread.
- */
-static void wait_defer(void)
-{
-	atomic_dec(&defer_thread_futex);
-	if (atomic_read(&defer_thread_futex) == -1)
-		futex(&defer_thread_futex, FUTEX_WAIT, -1,
-		      NULL, NULL, 0);
-}
-
-static void internal_urcu_lock(pthread_mutex_t *mutex)
+static void mutex_lock(pthread_mutex_t *mutex)
 {
 	int ret;
 
@@ -116,7 +80,7 @@ static void internal_urcu_lock(pthread_mutex_t *mutex)
 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 }
 
-static void internal_urcu_unlock(pthread_mutex_t *mutex)
+static void mutex_unlock(pthread_mutex_t *mutex)
 {
 	int ret;
 
@@ -127,6 +91,51 @@ static void internal_urcu_unlock(pthread_mutex_t *mutex)
 	}
 }
 
+/*
+ * Wake-up any waiting defer thread. Called from many concurrent threads.
+ */
+static void wake_up_defer(void)
+{
+	if (unlikely(uatomic_read(&defer_thread_futex) == -1)) {
+		uatomic_set(&defer_thread_futex, 0);
+		futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
+		      NULL, NULL, 0);
+	}
+}
+
+static unsigned long rcu_defer_num_callbacks(void)
+{
+	unsigned long num_items = 0, head;
+	struct defer_queue *index;
+
+	mutex_lock(&rcu_defer_mutex);
+	list_for_each_entry(index, &registry, list) {
+		head = CAA_LOAD_SHARED(index->head);
+		num_items += head - index->tail;
+	}
+	mutex_unlock(&rcu_defer_mutex);
+	return num_items;
+}
+
+/*
+ * Defer thread waiting. Single thread.
+ */
+static void wait_defer(void)
+{
+	uatomic_dec(&defer_thread_futex);
+	cmm_smp_mb();	/* Write futex before read queue */
+	if (rcu_defer_num_callbacks()) {
+		cmm_smp_mb();	/* Read queue before write futex */
+		/* Callbacks are queued, don't wait. */
+		uatomic_set(&defer_thread_futex, 0);
+	} else {
+		cmm_smp_rmb();	/* Read queue before read futex */
+		if (uatomic_read(&defer_thread_futex) == -1)
+			futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
+			      NULL, NULL, 0);
+	}
+}
+
 /*
  * Must be called after Q.S. is reached.
  */
@@ -143,22 +152,22 @@ static void rcu_defer_barrier_queue(struct defer_queue *queue,
 	 */
 
 	for (i = queue->tail; i != head;) {
-		smp_rmb();       /* read head before q[]. */
-		p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
+		cmm_smp_rmb();       /* read head before q[]. */
+		p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 		if (unlikely(DQ_IS_FCT_BIT(p))) {
 			DQ_CLEAR_FCT_BIT(p);
 			queue->last_fct_out = p;
-			p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
+			p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 		} else if (unlikely(p == DQ_FCT_MARK)) {
-			p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
+			p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 			queue->last_fct_out = p;
-			p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
+			p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 		}
 		fct = queue->last_fct_out;
 		fct(p);
 	}
-	smp_mb();	/* push tail after having used q[] */
-	STORE_SHARED(queue->tail, i);
+	cmm_smp_mb();	/* push tail after having used q[] */
+	CAA_STORE_SHARED(queue->tail, i);
 }
 
 static void _rcu_defer_barrier_thread(void)
@@ -173,12 +182,11 @@ static void _rcu_defer_barrier_thread(void)
 	rcu_defer_barrier_queue(&defer_queue, head);
 }
 
-
 void rcu_defer_barrier_thread(void)
 {
-	internal_urcu_lock(&urcu_defer_mutex);
+	mutex_lock(&rcu_defer_mutex);
 	_rcu_defer_barrier_thread();
-	internal_urcu_unlock(&urcu_defer_mutex);
+	mutex_unlock(&rcu_defer_mutex);
 }
 
 /*
@@ -196,16 +204,16 @@ void rcu_defer_barrier_thread(void)
 
 void rcu_defer_barrier(void)
 {
-	struct deferer_registry *index;
+	struct defer_queue *index;
 	unsigned long num_items = 0;
 
-	if (!registry)
+	if (list_empty(&registry))
 		return;
 
-	internal_urcu_lock(&urcu_defer_mutex);
-	for (index = registry; index < registry + num_deferers; index++) {
-		index->last_head = LOAD_SHARED(index->defer_queue->head);
-		num_items += index->last_head - index->defer_queue->tail;
+	mutex_lock(&rcu_defer_mutex);
+	list_for_each_entry(index, &registry, list) {
+		index->last_head = CAA_LOAD_SHARED(index->head);
+		num_items += index->last_head - index->tail;
 	}
 	if (likely(!num_items)) {
 		/*
@@ -215,17 +223,16 @@ void rcu_defer_barrier(void)
 		goto end;
 	}
 	synchronize_rcu();
-	for (index = registry; index < registry + num_deferers; index++)
-		rcu_defer_barrier_queue(index->defer_queue,
-					  index->last_head);
+	list_for_each_entry(index, &registry, list)
+		rcu_defer_barrier_queue(index, index->last_head);
 end:
-	internal_urcu_unlock(&urcu_defer_mutex);
+	mutex_unlock(&rcu_defer_mutex);
 }
 
 /*
- * _rcu_defer_queue - Queue a RCU callback.
+ * _defer_rcu - Queue a RCU callback.
  */
-void _rcu_defer_queue(void (*fct)(void *p), void *p)
+void _defer_rcu(void (*fct)(void *p), void *p)
 {
 	unsigned long head, tail;
 
@@ -234,16 +241,16 @@ void _rcu_defer_queue(void (*fct)(void *p), void *p)
 	 * thread.
 	 */
 	head = defer_queue.head;
-	tail = LOAD_SHARED(defer_queue.tail);
+	tail = CAA_LOAD_SHARED(defer_queue.tail);
 
 	/*
-	 * If queue is full, empty it ourself.
+	 * If queue is full, or reached threshold. Empty queue ourself.
 	 * Worse-case: must allow 2 supplementary entries for fct pointer.
 	 */
 	if (unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) {
 		assert(head - tail <= DEFER_QUEUE_SIZE);
 		rcu_defer_barrier_thread();
-		assert(head - LOAD_SHARED(defer_queue.tail) == 0);
+		assert(head - CAA_LOAD_SHARED(defer_queue.tail) == 0);
 	}
 
 	if (unlikely(defer_queue.last_fct_in != fct)) {
@@ -254,13 +261,13 @@ void _rcu_defer_queue(void (*fct)(void *p), void *p)
 			 * marker, write DQ_FCT_MARK followed by the function
 			 * pointer.
 			 */
-			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+			_CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 				      DQ_FCT_MARK);
-			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+			_CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 				      fct);
 		} else {
 			DQ_SET_FCT_BIT(fct);
-			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+			_CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 				      fct);
 		}
 	} else {
@@ -269,16 +276,17 @@ void _rcu_defer_queue(void (*fct)(void *p), void *p)
 			 * If the data to encode is not aligned or the marker,
 			 * write DQ_FCT_MARK followed by the function pointer.
 			 */
-			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+			_CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 				      DQ_FCT_MARK);
-			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+			_CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 				      fct);
 		}
 	}
-	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
-	smp_wmb();	/* Publish new pointer before head */
+	_CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
+	cmm_smp_wmb();	/* Publish new pointer before head */
 			/* Write q[] before head. */
-	STORE_SHARED(defer_queue.head, head);
+	CAA_STORE_SHARED(defer_queue.head, head);
+	cmm_smp_mb();	/* Write queue head before read futex */
 	/*
 	 * Wake-up any waiting defer thread.
 	 */
@@ -289,21 +297,15 @@ void *thr_defer(void *args)
 {
 	for (;;) {
 		pthread_testcancel();
-		printf("a\n");
 		/*
 		 * "Be green". Don't wake up the CPU if there is no RCU work
 		 * to perform whatsoever. Aims at saving laptop battery life by
 		 * leaving the processor in sleep state when idle.
 		 */
-		printf("b\n");
 		wait_defer();
-		printf("e\n");
 		/* Sleeping after wait_defer to let many callbacks enqueue */
-		//TEST poll(NULL,0,100);	/* wait for 100ms */
-		printf("f\n");
+		poll(NULL,0,100);	/* wait for 100ms */
 		rcu_defer_barrier();
-		printf("perform deferred call_rcu() from worker thread %lu.\n",
-			time(NULL));
 	}
 
 	return NULL;
@@ -313,67 +315,16 @@ void *thr_defer(void *args)
  * library wrappers to be used by non-LGPL compatible source code.
  */
 
-void rcu_defer_queue(void (*fct)(void *p), void *p)
+void defer_rcu(void (*fct)(void *p), void *p)
 {
-	_rcu_defer_queue(fct, p);
-}
-
-static void rcu_add_deferer(pthread_t id)
-{
-	struct deferer_registry *oldarray;
-
-	if (!registry) {
-		alloc_deferers = INIT_NUM_THREADS;
-		num_deferers = 0;
-		registry =
-			malloc(sizeof(struct deferer_registry) * alloc_deferers);
-	}
-	if (alloc_deferers < num_deferers + 1) {
-		oldarray = registry;
-		registry = malloc(sizeof(struct deferer_registry)
-				* (alloc_deferers << 1));
-		memcpy(registry, oldarray,
-			sizeof(struct deferer_registry) * alloc_deferers);
-		alloc_deferers <<= 1;
-		free(oldarray);
-	}
-	registry[num_deferers].tid = id;
-	/* reference to the TLS of _this_ deferer thread. */
-	registry[num_deferers].defer_queue = &defer_queue;
-	registry[num_deferers].last_head = 0;
-	num_deferers++;
-}
-
-/*
- * Never shrink (implementation limitation).
- * This is O(nb threads). Eventually use a hash table.
- */
-static void rcu_remove_deferer(pthread_t id)
-{
-	struct deferer_registry *index;
-
-	assert(registry != NULL);
-	for (index = registry; index < registry + num_deferers; index++) {
-		if (pthread_equal(index->tid, id)) {
-			memcpy(index, &registry[num_deferers - 1],
-				sizeof(struct deferer_registry));
-			registry[num_deferers - 1].tid = 0;
-			registry[num_deferers - 1].defer_queue = NULL;
-			registry[num_deferers - 1].last_head = 0;
-			num_deferers--;
-			return;
-		}
-	}
-	/* Hrm not found, forgot to register ? */
-	assert(0);
+	_defer_rcu(fct, p);
 }
 
 static void start_defer_thread(void)
 {
 	int ret;
 
-	ret = pthread_create(&tid_defer, NULL, thr_defer,
-		NULL);
+	ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
 	assert(!ret);
 }
 
@@ -390,39 +341,42 @@ static void stop_defer_thread(void)
 
 void rcu_defer_register_thread(void)
 {
-	int deferers;
+	int was_empty;
 
-	internal_urcu_lock(&defer_thread_mutex);
-	internal_urcu_lock(&urcu_defer_mutex);
+	assert(defer_queue.last_head == 0);
+	assert(defer_queue.q == NULL);
 	defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
-	rcu_add_deferer(pthread_self());
-	deferers = num_deferers;
-	internal_urcu_unlock(&urcu_defer_mutex);
 
-	if (deferers == 1)
+	mutex_lock(&defer_thread_mutex);
+	mutex_lock(&rcu_defer_mutex);
+	was_empty = list_empty(&registry);
+	list_add(&defer_queue.list, &registry);
+	mutex_unlock(&rcu_defer_mutex);
+
+	if (was_empty)
 		start_defer_thread();
-	internal_urcu_unlock(&defer_thread_mutex);
+	mutex_unlock(&defer_thread_mutex);
 }
 
 void rcu_defer_unregister_thread(void)
 {
-	int deferers;
+	int is_empty;
 
-	internal_urcu_lock(&defer_thread_mutex);
-	internal_urcu_lock(&urcu_defer_mutex);
-	rcu_remove_deferer(pthread_self());
+	mutex_lock(&defer_thread_mutex);
+	mutex_lock(&rcu_defer_mutex);
+	list_del(&defer_queue.list);
 	_rcu_defer_barrier_thread();
 	free(defer_queue.q);
 	defer_queue.q = NULL;
-	deferers = num_deferers;
-	internal_urcu_unlock(&urcu_defer_mutex);
+	is_empty = list_empty(&registry);
+	mutex_unlock(&rcu_defer_mutex);
 
-	if (deferers == 0)
+	if (is_empty)
 		stop_defer_thread();
-	internal_urcu_unlock(&defer_thread_mutex);
+	mutex_unlock(&defer_thread_mutex);
 }
 
-void urcu_defer_exit(void)
+void rcu_defer_exit(void)
 {
-	free(registry);
+	assert(list_empty(&registry));
 }