*
* 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.
*/
+#define _BSD_SOURCE
#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"
-pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+#ifdef RCU_MEMBARRIER
+static int init_done;
+int has_sys_membarrier;
-/* Global quiescent period parity */
-int urcu_qparity;
+void __attribute__((constructor)) rcu_init(void);
+#endif
+
+#ifdef RCU_MB
+void rcu_init(void)
+{
+}
+#endif
-int __thread urcu_active_readers[2];
+#ifdef RCU_SIGNAL
+static int init_done;
+
+void __attribute__((constructor)) rcu_init(void);
+void __attribute__((destructor)) rcu_exit(void);
+#endif
+
+static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
+
+int gp_futex;
+
+/*
+ * Global grace period counter.
+ * Contains the current RCU_GP_CTR_PHASE.
+ * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
+ * Written to only by writer with mutex taken. Read by both writer and readers.
+ */
+unsigned long rcu_gp_ctr = RCU_GP_COUNT;
-/* Thread IDs of registered readers */
-#define INIT_NUM_THREADS 4
+/*
+ * Written to only by each individual reader. Read by both the reader and the
+ * writers.
+ */
+struct rcu_reader __thread rcu_reader;
-struct reader_data {
- pthread_t tid;
- int *urcu_active_readers;
-};
+#ifdef DEBUG_YIELD
+unsigned int yield_active;
+unsigned int __thread rand_yield;
+#endif
-static struct reader_data *reader_data;
-static int num_readers, alloc_readers;
-static int sig_done;
+static LIST_HEAD(registry);
-void internal_urcu_lock(void)
+static void mutex_lock(pthread_mutex_t *mutex)
{
int ret;
- ret = pthread_mutex_lock(&urcu_mutex);
+
+#ifndef DISTRUST_SIGNALS_EXTREME
+ ret = pthread_mutex_lock(mutex);
if (ret) {
perror("Error in pthread mutex lock");
exit(-1);
}
+#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
+ while ((ret = pthread_mutex_trylock(mutex)) != 0) {
+ if (ret != EBUSY && ret != EINTR) {
+ printf("ret = %d, errno = %d\n", ret, errno);
+ perror("Error in pthread mutex lock");
+ exit(-1);
+ }
+ if (LOAD_SHARED(rcu_reader.need_mb)) {
+ smp_mb();
+ _STORE_SHARED(rcu_reader.need_mb, 0);
+ smp_mb();
+ }
+ poll(NULL,0,10);
+ }
+#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
}
-void internal_urcu_unlock(void)
+static void mutex_unlock(pthread_mutex_t *mutex)
{
int ret;
- ret = pthread_mutex_unlock(&urcu_mutex);
+ ret = pthread_mutex_unlock(mutex);
if (ret) {
perror("Error in pthread mutex unlock");
exit(-1);
}
}
-/*
- * called with urcu_mutex held.
- */
-static int switch_next_urcu_qparity(void)
+#ifdef RCU_MEMBARRIER
+static void smp_mb_master(int group)
{
- int old_parity = urcu_qparity;
- urcu_qparity = 1 - old_parity;
- return old_parity;
+ if (likely(has_sys_membarrier))
+ membarrier(MEMBARRIER_EXPEDITED);
+ else
+ smp_mb();
}
+#endif
-static void force_mb_all_threads(void)
+#ifdef RCU_MB
+static void smp_mb_master(int group)
{
- struct reader_data *index;
+ smp_mb();
+}
+#endif
+
+#ifdef RCU_SIGNAL
+static void force_mb_all_readers(void)
+{
+ struct rcu_reader *index;
+
/*
- * Ask for each threads to execute a mb() so we can consider the
+ * 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 (list_empty(®istry))
return;
- sig_done = 0;
- mb(); /* write sig_done before sending the signals */
- for (index = reader_data; index < reader_data + num_readers; index++)
- pthread_kill(index->tid, SIGURCU);
+ /*
+ * 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.
+ */
+ list_for_each_entry(index, ®istry, node) {
+ STORE_SHARED(index->need_mb, 1);
+ pthread_kill(index->tid, SIGRCU);
+ }
/*
* 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();
- mb(); /* read sig_done before ending the barrier */
+ list_for_each_entry(index, ®istry, node) {
+ while (LOAD_SHARED(index->need_mb)) {
+ pthread_kill(index->tid, SIGRCU);
+ poll(NULL, 0, 1);
+ }
+ }
+ smp_mb(); /* read ->need_mb before ending the barrier */
}
-void wait_for_quiescent_state(int parity)
+static void smp_mb_master(int group)
{
- struct reader_data *index;
+ force_mb_all_readers();
+}
+#endif /* #ifdef RCU_SIGNAL */
- if (!reader_data)
- return;
- /* Wait for each thread urcu_active_readers count to become 0.
+/*
+ * synchronize_rcu() waiting. Single thread.
+ */
+static void wait_gp(void)
+{
+ /* Read reader_gp before read futex */
+ smp_mb_master(RCU_MB_GROUP);
+ if (uatomic_read(&gp_futex) == -1)
+ futex_async(&gp_futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0);
+}
+
+void update_counter_and_wait(void)
+{
+ LIST_HEAD(qsreaders);
+ int wait_loops = 0;
+ struct rcu_reader *index, *tmp;
+
+ /* Switch parity: 0 -> 1, 1 -> 0 */
+ STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
+
+ /*
+ * Must commit rcu_gp_ctr update to memory before waiting for quiescent
+ * state. Failure to do so could result in the writer waiting forever
+ * while new readers are always accessing data (no progress). Enforce
+ * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr.
+ */
+ barrier();
+
+ /*
+ *
+ * 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.
*/
- for (index = reader_data; index < reader_data + num_readers; index++) {
+ smp_mb();
+
+ /*
+ * Wait for each thread rcu_reader.ctr count to become 0.
+ */
+ for (;;) {
+ wait_loops++;
+ if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ uatomic_dec(&gp_futex);
+ /* Write futex before read reader_gp */
+ smp_mb_master(RCU_MB_GROUP);
+ }
+
+ list_for_each_entry_safe(index, tmp, ®istry, node) {
+ if (!rcu_gp_ongoing(&index->ctr))
+ list_move(&index->node, &qsreaders);
+ }
+
+#ifndef HAS_INCOHERENT_CACHES
+ if (list_empty(®istry)) {
+ if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ /* Read reader_gp before write futex */
+ smp_mb_master(RCU_MB_GROUP);
+ uatomic_set(&gp_futex, 0);
+ }
+ break;
+ } else {
+ if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
+ wait_gp();
+ else
+ cpu_relax();
+ }
+#else /* #ifndef HAS_INCOHERENT_CACHES */
/*
- * BUSY-LOOP.
+ * BUSY-LOOP. Force the reader thread to commit its
+ * rcu_reader.ctr update to memory if we wait for too long.
*/
- while (index->urcu_active_readers[parity] != 0)
- barrier();
+ if (list_empty(®istry)) {
+ if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ /* Read reader_gp before write futex */
+ smp_mb_master(RCU_MB_GROUP);
+ uatomic_set(&gp_futex, 0);
+ }
+ break;
+ } else {
+ switch (wait_loops) {
+ case RCU_QS_ACTIVE_ATTEMPTS:
+ wait_gp();
+ break; /* only escape switch */
+ case KICK_READER_LOOPS:
+ smp_mb_master(RCU_MB_GROUP);
+ wait_loops = 0;
+ break; /* only escape switch */
+ default:
+ 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();
+ /* put back the reader list in the registry */
+ list_splice(&qsreaders, ®istry);
}
-static void switch_qparity(void)
+void synchronize_rcu(void)
{
- int prev_parity;
+ mutex_lock(&rcu_gp_lock);
+
+ if (list_empty(®istry))
+ goto out;
- /* All threads should read qparity before accessing data structure. */
- /* Write ptr before changing the qparity */
- force_mb_all_threads();
- prev_parity = switch_next_urcu_qparity();
+ /* All threads should read qparity before accessing data structure
+ * where new ptr points to. Must be done within rcu_gp_lock because it
+ * iterates on reader threads.*/
+ /* Write new ptr before changing the qparity */
+ smp_mb_master(RCU_MB_GROUP);
/*
* Wait for previous parity to be empty of readers.
*/
- wait_for_quiescent_state(prev_parity);
-}
+ update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
-void synchronize_rcu(void)
-{
- internal_urcu_lock();
- switch_qparity();
- switch_qparity();
- internal_urcu_lock();
-}
+ /*
+ * Must finish waiting for quiescent state for parity 0 before
+ * committing next rcu_gp_ctr update to memory. Failure to do so could
+ * result in the writer waiting forever while new readers are always
+ * accessing data (no progress). Enforce compiler-order of load
+ * rcu_reader ctr before store to rcu_gp_ctr.
+ */
+ barrier();
-/*
- * Return old pointer, OK to free, no more reference exist.
- * Called under rcu_write_lock.
- */
-void *urcu_publish_content(void **ptr, void *new)
-{
- void *oldptr;
+ /*
+ * 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();
- internal_urcu_lock();
/*
- * 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.
+ * Wait for previous parity to be empty of readers.
*/
- oldptr = *ptr;
- *ptr = new;
+ update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
+
+ /* Finish waiting for reader threads before letting the old ptr being
+ * freed. Must be done within rcu_gp_lock because it iterates on reader
+ * threads. */
+ smp_mb_master(RCU_MB_GROUP);
+out:
+ mutex_unlock(&rcu_gp_lock);
+}
- switch_qparity();
- switch_qparity();
- internal_urcu_unlock();
+/*
+ * library wrappers to be used by non-LGPL compatible source code.
+ */
- return oldptr;
+void rcu_read_lock(void)
+{
+ _rcu_read_lock();
}
-void urcu_add_reader(pthread_t id)
+void rcu_read_unlock(void)
{
- struct reader_data *oldarray;
-
- if (!reader_data) {
- alloc_readers = INIT_NUM_THREADS;
- num_readers = 0;
- reader_data =
- malloc(sizeof(struct reader_data) * alloc_readers);
- }
- if (alloc_readers < num_readers + 1) {
- oldarray = reader_data;
- reader_data = malloc(sizeof(struct reader_data)
- * (alloc_readers << 1));
- memcpy(reader_data, oldarray,
- sizeof(struct reader_data) * alloc_readers);
- alloc_readers <<= 1;
- free(oldarray);
- }
- reader_data[num_readers].tid = id;
- /* reference to the TLS of _this_ reader thread. */
- reader_data[num_readers].urcu_active_readers = urcu_active_readers;
- num_readers++;
+ _rcu_read_unlock();
}
-/*
- * Never shrink (implementation limitation).
- * This is O(nb threads). Eventually use a hash table.
- */
-void urcu_remove_reader(pthread_t id)
+void rcu_register_thread(void)
{
- struct reader_data *index;
-
- assert(reader_data != NULL);
- for (index = reader_data; index < reader_data + 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;
- num_readers--;
- return;
- }
- }
- /* Hrm not found, forgot to register ? */
- assert(0);
+ rcu_reader.tid = pthread_self();
+ assert(rcu_reader.need_mb == 0);
+ assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK));
+
+ mutex_lock(&rcu_gp_lock);
+ rcu_init(); /* In case gcc does not support constructor attribute */
+ list_add(&rcu_reader.node, ®istry);
+ mutex_unlock(&rcu_gp_lock);
}
-void urcu_register_thread(void)
+void rcu_unregister_thread(void)
{
- internal_urcu_lock();
- urcu_add_reader(pthread_self());
- internal_urcu_unlock();
+ mutex_lock(&rcu_gp_lock);
+ list_del(&rcu_reader.node);
+ mutex_unlock(&rcu_gp_lock);
}
-void urcu_unregister_thread(void)
+#ifdef RCU_MEMBARRIER
+void rcu_init(void)
{
- internal_urcu_lock();
- urcu_remove_reader(pthread_self());
- internal_urcu_unlock();
+ if (init_done)
+ return;
+ init_done = 1;
+ if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY))
+ has_sys_membarrier = 1;
}
+#endif
-void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
+#ifdef RCU_SIGNAL
+static void sigrcu_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();
+ _STORE_SHARED(rcu_reader.need_mb, 0);
+ smp_mb();
}
-void __attribute__((constructor)) urcu_init(void)
+/*
+ * rcu_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 rcu_gp_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 rcu_init(void)
{
struct sigaction act;
int ret;
- act.sa_sigaction = sigurcu_handler;
- ret = sigaction(SIGURCU, &act, NULL);
+ if (init_done)
+ return;
+ init_done = 1;
+
+ act.sa_sigaction = sigrcu_handler;
+ act.sa_flags = SA_SIGINFO | SA_RESTART;
+ sigemptyset(&act.sa_mask);
+ ret = sigaction(SIGRCU, &act, NULL);
if (ret) {
perror("Error in sigaction");
exit(-1);
}
}
-void __attribute__((destructor)) urcu_exit(void)
+void rcu_exit(void)
{
struct sigaction act;
int ret;
- ret = sigaction(SIGURCU, NULL, &act);
+ ret = sigaction(SIGRCU, NULL, &act);
if (ret) {
perror("Error in sigaction");
exit(-1);
}
- assert(act.sa_sigaction == sigurcu_handler);
- free(reader_data);
+ assert(act.sa_sigaction == sigrcu_handler);
+ assert(list_empty(®istry));
}
+#endif /* #ifdef RCU_SIGNAL */