Add speculative execution (prefetch) to model

[urcu.git] / urcu.c

/*
 * urcu.c
 *
 * Userspace RCU library
 *
 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
 *
 * 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>
#include <pthread.h>
#include <signal.h>
#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"

void __attribute__((constructor)) urcu_init(void);
void __attribute__((destructor)) urcu_exit(void);

int init_done;

pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;

/*
 * 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;

/*
 * 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_registry {
	pthread_t tid;
	long *urcu_active_readers;
	char *need_mb;
};

#ifdef DEBUG_YIELD
unsigned int yield_active;
unsigned int __thread rand_yield;
#endif

static struct reader_registry *registry;
static char __thread need_mb;
static int num_readers, alloc_readers;

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)
{
	int ret;

	ret = pthread_mutex_unlock(&urcu_mutex);
	if (ret) {
		perror("Error in pthread mutex unlock");
		exit(-1);
	}
}

/*
 * called with urcu_mutex held.
 */
static void switch_next_urcu_qparity(void)
{
	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)
{
	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();
	/*
	 * 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 (!registry)
		return;
	/*
	 * 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);
	}
	/*
	 * Wait for sighandler (and thus mb()) to execute on every thread.
	 *
	 * 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).
	 */
	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(void)
{
	struct reader_registry *index;

	if (!registry)
		return;
	/*
	 * Wait for each thread urcu_active_readers count to become 0.
	 */
	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. Force the reader thread to commit its
		 * urcu_active_readers update to memory if we wait for too long.
		 */
		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 */
	}
}

void synchronize_rcu(void)
{
	internal_urcu_lock();

	/* 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();

	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.
	 */

	/*
	 * Adding a smp_mb() which is _not_ formally required, but makes the
	 * model easier to understand. It does not have a big performance impact
	 * anyway, given this is the write-side.
	 */
	smp_mb();

	/*
	 * Wait for previous parity to be empty of readers.
	 */
	wait_for_quiescent_state();	/* Wait readers in parity 0 */

	/*
	 * 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.
	 */

	/*
	 * Adding a smp_mb() which is _not_ formally required, but makes the
	 * model easier to understand. It does not have a big performance impact
	 * anyway, given this is the write-side.
	 */
	smp_mb();

	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.
	 */

	/*
	 * Adding a smp_mb() which is _not_ formally required, but makes the
	 * model easier to understand. It does not have a big performance impact
	 * anyway, given this is the write-side.
	 */
	smp_mb();

	/*
	 * 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();
}

/*
 * library wrappers to be used by non-LGPL compatible source code.
 */

void rcu_read_lock(void)
{
	_rcu_read_lock();
}

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;
}

static void rcu_add_reader(pthread_t id)
{
	struct reader_registry *oldarray;

	if (!registry) {
		alloc_readers = INIT_NUM_THREADS;
		num_readers = 0;
		registry =
			malloc(sizeof(struct reader_registry) * alloc_readers);
	}
	if (alloc_readers < num_readers + 1) {
		oldarray = registry;
		registry = malloc(sizeof(struct reader_registry)
				* (alloc_readers << 1));
		memcpy(registry, oldarray,
			sizeof(struct reader_registry) * alloc_readers);
		alloc_readers <<= 1;
		free(oldarray);
	}
	registry[num_readers].tid = id;
	/* reference to the TLS of _this_ reader thread. */
	registry[num_readers].urcu_active_readers = &urcu_active_readers;
	registry[num_readers].need_mb = &need_mb;
	num_readers++;
}

/*
 * Never shrink (implementation limitation).
 * This is O(nb threads). Eventually use a hash table.
 */
static void rcu_remove_reader(pthread_t id)
{
	struct reader_registry *index;

	assert(registry != NULL);
	for (index = registry; index < registry + num_readers; index++) {
		if (pthread_equal(index->tid, id)) {
			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;
		}
	}
	/* Hrm not found, forgot to register ? */
	assert(0);
}

void rcu_register_thread(void)
{
	internal_urcu_lock();
	urcu_init();	/* In case gcc does not support constructor attribute */
	rcu_add_reader(pthread_self());
	internal_urcu_unlock();
}

void rcu_unregister_thread(void)
{
	internal_urcu_lock();
	rcu_remove_reader(pthread_self());
	internal_urcu_unlock();
}

#ifndef DEBUG_FULL_MB
static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
{
	/*
	 * 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();
}

/*
 * urcu_init constructor. Called when the library is linked, but also when
 * reader threads are calling rcu_register_thread().
 * Should only be called by a single thread at a given time. This is ensured by
 * holing the internal_urcu_lock() from rcu_register_thread() or by running at
 * library load time, which should not be executed by multiple threads nor
 * concurrently with rcu_register_thread() anyway.
 */
void urcu_init(void)
{
	struct sigaction act;
	int ret;

	if (init_done)
		return;
	init_done = 1;

	act.sa_sigaction = sigurcu_handler;
	ret = sigaction(SIGURCU, &act, NULL);
	if (ret) {
		perror("Error in sigaction");
		exit(-1);
	}
}

void urcu_exit(void)
{
	struct sigaction act;
	int ret;

	ret = sigaction(SIGURCU, NULL, &act);
	if (ret) {
		perror("Error in sigaction");
		exit(-1);
	}
	assert(act.sa_sigaction == sigurcu_handler);
	free(registry);
}
#endif /* #ifndef DEBUG_FULL_MB */