X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=efbb4b7890a71524fa8264877f3da63565291591;hp=4362217d49ab4020c0bd199c30410733f32bbd89;hb=67c2d80b2b808fe7a6d6d0660ca81c919a81010b;hpb=cf380c2fd2f85e3fba4826991fb748255bdc9b76

diff --git a/urcu.c b/urcu.c
index 4362217..efbb4b7 100644
--- a/urcu.c
+++ b/urcu.c
@@ -3,9 +3,24 @@
  *
  * Userspace RCU library
  *
- * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
  *
- * Distributed under GPLv2
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ *
+ * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  */
 
 #include <stdio.h>
@@ -14,40 +29,72 @@
 #include <assert.h>
 #include <stdlib.h>
 #include <string.h>
+#include <errno.h>
+#include <poll.h>
 
+#include "urcu-static.h"
+/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 #include "urcu.h"
 
 pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
 
-/* Global quiescent period parity */
-int urcu_qparity;
+/*
+ * Global grace period counter.
+ * Contains the current RCU_GP_CTR_BIT.
+ * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
+ * Written to only by writer with mutex taken. Read by both writer and readers.
+ */
+long urcu_gp_ctr = RCU_GP_COUNT;
 
-int __thread urcu_active_readers[2];
+/*
+ * Written to only by each individual reader. Read by both the reader and the
+ * writers.
+ */
+long __thread urcu_active_readers;
 
 /* Thread IDs of registered readers */
 #define INIT_NUM_THREADS 4
 
-struct reader_data {
+struct reader_registry {
 	pthread_t tid;
-	int *urcu_active_readers;
+	long *urcu_active_readers;
+	char *need_mb;
 };
 
 #ifdef DEBUG_YIELD
-int yield_active;
+unsigned int yield_active;
+unsigned int __thread rand_yield;
 #endif
 
-static struct reader_data *reader_data;
+static struct reader_registry *registry;
+static char __thread need_mb;
 static int num_readers, alloc_readers;
-static int sig_done;
 
 void internal_urcu_lock(void)
 {
 	int ret;
+
+#ifndef DISTRUST_SIGNALS_EXTREME
 	ret = pthread_mutex_lock(&urcu_mutex);
 	if (ret) {
 		perror("Error in pthread mutex lock");
 		exit(-1);
 	}
+#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
+	while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
+		if (ret != EBUSY && ret != EINTR) {
+			printf("ret = %d, errno = %d\n", ret, errno);
+			perror("Error in pthread mutex lock");
+			exit(-1);
+		}
+		if (need_mb) {
+			smp_mb();
+			need_mb = 0;
+			smp_mb();
+		}
+		poll(NULL,0,10);
+	}
+#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 }
 
 void internal_urcu_unlock(void)
@@ -64,154 +111,254 @@ void internal_urcu_unlock(void)
 /*
  * called with urcu_mutex held.
  */
-static int switch_next_urcu_qparity(void)
+static void switch_next_urcu_qparity(void)
 {
-	int old_parity = urcu_qparity;
-	urcu_qparity = 1 - old_parity;
-	return old_parity;
+	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 }
 
+#ifdef DEBUG_FULL_MB
+#ifdef HAS_INCOHERENT_CACHES
+static void force_mb_single_thread(struct reader_registry *index)
+{
+	smp_mb();
+}
+#endif /* #ifdef HAS_INCOHERENT_CACHES */
+
 static void force_mb_all_threads(void)
 {
-	struct reader_data *index;
+	smp_mb();
+}
+#else /* #ifdef DEBUG_FULL_MB */
+#ifdef HAS_INCOHERENT_CACHES
+static void force_mb_single_thread(struct reader_registry *index)
+{
+	assert(registry);
+	/*
+	 * pthread_kill has a smp_mb(). But beware, we assume it performs
+	 * a cache flush on architectures with non-coherent cache. Let's play
+	 * safe and don't assume anything : we use smp_mc() to make sure the
+	 * cache flush is enforced.
+	 */
+	*index->need_mb = 1;
+	smp_mc();	/* write ->need_mb before sending the signals */
+	pthread_kill(index->tid, SIGURCU);
+	smp_mb();
 	/*
-	 * Ask for each threads to execute a mb() so we can consider the
+	 * Wait for sighandler (and thus mb()) to execute on every thread.
+	 * BUSY-LOOP.
+	 */
+	while (*index->need_mb) {
+		poll(NULL, 0, 1);
+	}
+	smp_mb();	/* read ->need_mb before ending the barrier */
+}
+#endif /* #ifdef HAS_INCOHERENT_CACHES */
+
+static void force_mb_all_threads(void)
+{
+	struct reader_registry *index;
+	/*
+	 * Ask for each threads to execute a smp_mb() so we can consider the
 	 * compiler barriers around rcu read lock as real memory barriers.
 	 */
-	if (!reader_data)
+	if (!registry)
 		return;
-	debug_yield_write();
-	sig_done = 0;
-	debug_yield_write();
-	mb();	/* write sig_done before sending the signals */
-	debug_yield_write();
-	for (index = reader_data; index < reader_data + num_readers; index++) {
+	/*
+	 * pthread_kill has a smp_mb(). But beware, we assume it performs
+	 * a cache flush on architectures with non-coherent cache. Let's play
+	 * safe and don't assume anything : we use smp_mc() to make sure the
+	 * cache flush is enforced.
+	 */
+	for (index = registry; index < registry + num_readers; index++) {
+		*index->need_mb = 1;
+		smp_mc();	/* write need_mb before sending the signal */
 		pthread_kill(index->tid, SIGURCU);
-		debug_yield_write();
 	}
 	/*
 	 * Wait for sighandler (and thus mb()) to execute on every thread.
-	 * BUSY-LOOP.
+	 *
+	 * Note that the pthread_kill() will never be executed on systems
+	 * that correctly deliver signals in a timely manner.  However, it
+	 * is not uncommon for kernels to have bugs that can result in
+	 * lost or unduly delayed signals.
+	 *
+	 * If you are seeing the below pthread_kill() executing much at
+	 * all, we suggest testing the underlying kernel and filing the
+	 * relevant bug report.  For Linux kernels, we recommend getting
+	 * the Linux Test Project (LTP).
 	 */
-	while (sig_done < num_readers)
-		barrier();
-	debug_yield_write();
-	mb();	/* read sig_done before ending the barrier */
-	debug_yield_write();
+	for (index = registry; index < registry + num_readers; index++) {
+		while (*index->need_mb) {
+			pthread_kill(index->tid, SIGURCU);
+			poll(NULL, 0, 1);
+		}
+	}
+	smp_mb();	/* read ->need_mb before ending the barrier */
 }
+#endif /* #else #ifdef DEBUG_FULL_MB */
 
-void wait_for_quiescent_state(int parity)
+void wait_for_quiescent_state(void)
 {
-	struct reader_data *index;
+	struct reader_registry *index;
 
-	if (!reader_data)
+	if (!registry)
 		return;
-	/* Wait for each thread urcu_active_readers count to become 0.
+	/*
+	 * Wait for each thread urcu_active_readers count to become 0.
 	 */
-	for (index = reader_data; index < reader_data + num_readers; index++) {
+	for (index = registry; index < registry + num_readers; index++) {
+#ifndef HAS_INCOHERENT_CACHES
+		while (rcu_old_gp_ongoing(index->urcu_active_readers))
+			cpu_relax();
+#else /* #ifndef HAS_INCOHERENT_CACHES */
+		int wait_loops = 0;
 		/*
-		 * BUSY-LOOP.
+		 * BUSY-LOOP. Force the reader thread to commit its
+		 * urcu_active_readers update to memory if we wait for too long.
 		 */
-		while (index->urcu_active_readers[parity] != 0)
-			barrier();
+		while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
+			if (wait_loops++ == KICK_READER_LOOPS) {
+				force_mb_single_thread(index);
+				wait_loops = 0;
+			} else {
+				cpu_relax();
+			}
+		}
+#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 	}
-	/*
-	 * Locally : read *index->urcu_active_readers before freeing old
-	 * pointer.
-	 * Remote (reader threads) : Order urcu_qparity update and other
-	 * thread's quiescent state counter read.
-	 */
-	force_mb_all_threads();
 }
 
-static void switch_qparity(void)
+void synchronize_rcu(void)
 {
-	int prev_parity;
+	internal_urcu_lock();
 
-	/* All threads should read qparity before accessing data structure. */
-	/* Write ptr before changing the qparity */
+	/* All threads should read qparity before accessing data structure
+	 * where new ptr points to. Must be done within internal_urcu_lock
+	 * because it iterates on reader threads.*/
+	/* Write new ptr before changing the qparity */
 	force_mb_all_threads();
-	debug_yield_write();
-	prev_parity = switch_next_urcu_qparity();
-	debug_yield_write();
+
+	switch_next_urcu_qparity();	/* 0 -> 1 */
+
+	/*
+	 * Must commit qparity update to memory before waiting for parity
+	 * 0 quiescent state. Failure to do so could result in the writer
+	 * waiting forever while new readers are always accessing data (no
+	 * progress).
+	 * Ensured by STORE_SHARED and LOAD_SHARED.
+	 */
+
+	/*
+	 * Current RCU formal verification model assumes sequential execution of
+	 * the write-side. Add core synchronization instructions. Can be removed
+	 * if the formal model is extended to prove that reordering is still
+	 * correct.
+	 */
+	sync_core();	/* Formal model assumes serialized execution */
 
 	/*
 	 * Wait for previous parity to be empty of readers.
 	 */
-	wait_for_quiescent_state(prev_parity);
-}
+	wait_for_quiescent_state();	/* Wait readers in parity 0 */
 
-void synchronize_rcu(void)
-{
-	debug_yield_write();
-	internal_urcu_lock();
-	debug_yield_write();
-	switch_qparity();
-	debug_yield_write();
-	switch_qparity();
-	debug_yield_write();
-	internal_urcu_lock();
-	debug_yield_write();
+	/*
+	 * Must finish waiting for quiescent state for parity 0 before
+	 * committing qparity update to memory. Failure to do so could result in
+	 * the writer waiting forever while new readers are always accessing
+	 * data (no progress).
+	 * Ensured by STORE_SHARED and LOAD_SHARED.
+	 */
+
+	sync_core();	/* Formal model assumes serialized execution */
+
+	switch_next_urcu_qparity();	/* 1 -> 0 */
+
+	/*
+	 * Must commit qparity update to memory before waiting for parity
+	 * 1 quiescent state. Failure to do so could result in the writer
+	 * waiting forever while new readers are always accessing data (no
+	 * progress).
+	 * Ensured by STORE_SHARED and LOAD_SHARED.
+	 */
+
+	sync_core();	/* Formal model assumes serialized execution */
+
+	/*
+	 * Wait for previous parity to be empty of readers.
+	 */
+	wait_for_quiescent_state();	/* Wait readers in parity 1 */
+
+	/* Finish waiting for reader threads before letting the old ptr being
+	 * freed. Must be done within internal_urcu_lock because it iterates on
+	 * reader threads. */
+	force_mb_all_threads();
+
+	internal_urcu_unlock();
 }
 
 /*
- * Return old pointer, OK to free, no more reference exist.
- * Called under rcu_write_lock.
+ * library wrappers to be used by non-LGPL compatible source code.
  */
-void *urcu_publish_content(void **ptr, void *new)
+
+void rcu_read_lock(void)
 {
-	void *oldptr;
+	_rcu_read_lock();
+}
 
-	debug_yield_write();
-	internal_urcu_lock();
-	debug_yield_write();
-	/*
-	 * We can publish the new pointer before we change the current qparity.
-	 * Readers seeing the new pointer while being in the previous qparity
-	 * window will make us wait until the end of the quiescent state before
-	 * we release the unrelated memory area. However, given we hold the
-	 * urcu_mutex, we are making sure that no further garbage collection can
-	 * occur until we release the mutex, therefore we guarantee that this
-	 * given reader will have completed its execution using the new pointer
-	 * when the next quiescent state window will be over.
-	 */
-	oldptr = *ptr;
-	debug_yield_write();
-	*ptr = new;
-
-	debug_yield_write();
-	switch_qparity();
-	debug_yield_write();
-	switch_qparity();
-	debug_yield_write();
-	internal_urcu_unlock();
-	debug_yield_write();
+void rcu_read_unlock(void)
+{
+	_rcu_read_unlock();
+}
+
+void *rcu_dereference(void *p)
+{
+	return _rcu_dereference(p);
+}
 
+void *rcu_assign_pointer_sym(void **p, void *v)
+{
+	wmb();
+	return STORE_SHARED(p, v);
+}
+
+void *rcu_xchg_pointer_sym(void **p, void *v)
+{
+	wmb();
+	return xchg(p, v);
+}
+
+void *rcu_publish_content_sym(void **p, void *v)
+{
+	void *oldptr;
+
+	oldptr = _rcu_xchg_pointer(p, v);
+	synchronize_rcu();
 	return oldptr;
 }
 
-void urcu_add_reader(pthread_t id)
+static void rcu_add_reader(pthread_t id)
 {
-	struct reader_data *oldarray;
+	struct reader_registry *oldarray;
 
-	if (!reader_data) {
+	if (!registry) {
 		alloc_readers = INIT_NUM_THREADS;
 		num_readers = 0;
-		reader_data =
-			malloc(sizeof(struct reader_data) * alloc_readers);
+		registry =
+			malloc(sizeof(struct reader_registry) * alloc_readers);
 	}
 	if (alloc_readers < num_readers + 1) {
-		oldarray = reader_data;
-		reader_data = malloc(sizeof(struct reader_data)
+		oldarray = registry;
+		registry = malloc(sizeof(struct reader_registry)
 				* (alloc_readers << 1));
-		memcpy(reader_data, oldarray,
-			sizeof(struct reader_data) * alloc_readers);
+		memcpy(registry, oldarray,
+			sizeof(struct reader_registry) * alloc_readers);
 		alloc_readers <<= 1;
 		free(oldarray);
 	}
-	reader_data[num_readers].tid = id;
+	registry[num_readers].tid = id;
 	/* reference to the TLS of _this_ reader thread. */
-	reader_data[num_readers].urcu_active_readers = urcu_active_readers;
+	registry[num_readers].urcu_active_readers = &urcu_active_readers;
+	registry[num_readers].need_mb = &need_mb;
 	num_readers++;
 }
 
@@ -219,17 +366,17 @@ void urcu_add_reader(pthread_t id)
  * Never shrink (implementation limitation).
  * This is O(nb threads). Eventually use a hash table.
  */
-void urcu_remove_reader(pthread_t id)
+static void rcu_remove_reader(pthread_t id)
 {
-	struct reader_data *index;
+	struct reader_registry *index;
 
-	assert(reader_data != NULL);
-	for (index = reader_data; index < reader_data + num_readers; index++) {
+	assert(registry != NULL);
+	for (index = registry; index < registry + num_readers; index++) {
 		if (pthread_equal(index->tid, id)) {
-			memcpy(index, &reader_data[num_readers - 1],
-				sizeof(struct reader_data));
-			reader_data[num_readers - 1].tid = 0;
-			reader_data[num_readers - 1].urcu_active_readers = NULL;
+			memcpy(index, &registry[num_readers - 1],
+				sizeof(struct reader_registry));
+			registry[num_readers - 1].tid = 0;
+			registry[num_readers - 1].urcu_active_readers = NULL;
 			num_readers--;
 			return;
 		}
@@ -238,24 +385,31 @@ void urcu_remove_reader(pthread_t id)
 	assert(0);
 }
 
-void urcu_register_thread(void)
+void rcu_register_thread(void)
 {
 	internal_urcu_lock();
-	urcu_add_reader(pthread_self());
+	rcu_add_reader(pthread_self());
 	internal_urcu_unlock();
 }
 
-void urcu_unregister_thread(void)
+void rcu_unregister_thread(void)
 {
 	internal_urcu_lock();
-	urcu_remove_reader(pthread_self());
+	rcu_remove_reader(pthread_self());
 	internal_urcu_unlock();
 }
 
-void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
+#ifndef DEBUG_FULL_MB
+static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 {
-	mb();
-	atomic_inc(&sig_done);
+	/*
+	 * Executing this smp_mb() is the only purpose of this signal handler.
+	 * It punctually promotes barrier() into smp_mb() on every thread it is
+	 * executed on.
+	 */
+	smp_mb();
+	need_mb = 0;
+	smp_mb();
 }
 
 void __attribute__((constructor)) urcu_init(void)
@@ -282,5 +436,6 @@ void __attribute__((destructor)) urcu_exit(void)
 		exit(-1);
 	}
 	assert(act.sa_sigaction == sigurcu_handler);
-	free(reader_data);
+	free(registry);
 }
+#endif /* #ifndef DEBUG_FULL_MB */