/*
- * urcu.c
+ * urcu-qsbr.c
*
- * Userspace RCU library
+ * Userspace RCU QSBR library
*
* Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
* Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
#include <errno.h>
#include <poll.h>
-#include "urcu-static.h"
+#define BUILD_QSBR_LIB
+#include "urcu-qsbr-static.h"
/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
-#include "urcu.h"
+#include "urcu-qsbr.h"
-void __attribute__((constructor)) urcu_init(void);
-void __attribute__((destructor)) urcu_exit(void);
+static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
-int init_done;
-
-pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+int gp_futex;
/*
* Global grace period counter.
*/
-long urcu_gp_ctr = 0;
+unsigned long urcu_gp_ctr = RCU_GP_ONLINE;
/*
* Written to only by each individual reader. Read by both the reader and the
* writers.
*/
-long __thread urcu_active_readers;
+unsigned long __thread rcu_reader_qs_gp;
/* Thread IDs of registered readers */
#define INIT_NUM_THREADS 4
struct reader_registry {
pthread_t tid;
- long *urcu_active_readers;
- char *need_mb;
+ unsigned long *rcu_reader_qs_gp;
};
#ifdef DEBUG_YIELD
#endif
static struct reader_registry *registry;
-static char __thread need_mb;
static int num_readers, alloc_readers;
-void internal_urcu_lock(void)
+static void internal_urcu_lock(void)
{
int ret;
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)
+static void internal_urcu_unlock(void)
{
int ret;
}
}
-#ifdef DEBUG_FULL_MB
-#ifdef HAS_INCOHERENT_CACHES
-static void force_mb_single_thread(struct reader_registry *index)
+/*
+ * synchronize_rcu() waiting. Single thread.
+ */
+static void wait_gp(struct reader_registry *index)
{
- smp_mb();
+ atomic_dec(&gp_futex);
+ smp_mb(); /* Write futex before read reader_gp */
+ if (!rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
+ /* Read reader_gp before write futex */
+ smp_mb();
+ /* Callbacks are queued, don't wait. */
+ atomic_set(&gp_futex, 0);
+ } else {
+ /* Read reader_gp before read futex */
+ smp_rmb();
+ if (atomic_read(&gp_futex) == -1)
+ futex(&gp_futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0);
+ }
}
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
-static void force_mb_all_threads(void)
+static void wait_for_quiescent_state(void)
{
- smp_mb();
+ struct reader_registry *index;
+
+ if (!registry)
+ return;
+ /*
+ * Wait for each thread rcu_reader_qs_gp count to become 0.
+ */
+ for (index = registry; index < registry + num_readers; index++) {
+ int wait_loops = 0;
+
+ while (rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
+ if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
+ wait_gp(index);
+ } else {
+#ifndef HAS_INCOHERENT_CACHES
+ cpu_relax();
+#else /* #ifndef HAS_INCOHERENT_CACHES */
+ smp_mb();
+#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
+ }
+ }
+ }
}
-#else /* #ifdef DEBUG_FULL_MB */
-#ifdef HAS_INCOHERENT_CACHES
-static void force_mb_single_thread(struct reader_registry *index)
+
+/*
+ * Using a two-subphases algorithm for architectures with smaller than 64-bit
+ * long-size to ensure we do not encounter an overflow bug.
+ */
+
+#if (BITS_PER_LONG < 64)
+/*
+ * called with urcu_mutex held.
+ */
+static void switch_next_urcu_qparity(void)
{
- 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.
+ STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR);
+}
+
+void synchronize_rcu(void)
+{
+ unsigned long was_online;
+
+ was_online = rcu_reader_qs_gp;
+
+ /* All threads should read qparity before accessing data structure
+ * where new ptr points to.
*/
- *index->need_mb = 1;
- smp_mc(); /* write ->need_mb before sending the signals */
- pthread_kill(index->tid, SIGURCU);
+ /* Write new ptr before changing the qparity */
smp_mb();
+
/*
- * Wait for sighandler (and thus mb()) to execute on every thread.
- * BUSY-LOOP.
+ * Mark the writer thread offline to make sure we don't wait for
+ * our own quiescent state. This allows using synchronize_rcu() in
+ * threads registered as readers.
*/
- while (*index->need_mb) {
- poll(NULL, 0, 1);
- }
- smp_mb(); /* read ->need_mb before ending the barrier */
-}
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
+ if (was_online)
+ STORE_SHARED(rcu_reader_qs_gp, 0);
+
+ internal_urcu_lock();
+
+ switch_next_urcu_qparity(); /* 0 -> 1 */
-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.
+ * 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.
*/
- 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.
+ * Wait for previous parity to be empty of readers.
*/
- 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_quiescent_state(); /* Wait readers in parity 0 */
+
/*
- * 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).
+ * 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.
*/
- 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;
+ switch_next_urcu_qparity(); /* 1 -> 0 */
- if (!registry)
- return;
/*
- * Wait for each thread urcu_active_readers count to become 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.
*/
- 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 */
- }
-}
+ /*
+ * Wait for previous parity to be empty of readers.
+ */
+ wait_for_quiescent_state(); /* Wait readers in parity 1 */
+
+ internal_urcu_unlock();
+
+ /*
+ * Finish waiting for reader threads before letting the old ptr being
+ * freed.
+ */
+ if (was_online)
+ _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
+ smp_mb();
+}
+#else /* !(BITS_PER_LONG < 64) */
void synchronize_rcu(void)
{
+ unsigned long was_online;
+
+ was_online = rcu_reader_qs_gp;
+
+ /*
+ * Mark the writer thread offline to make sure we don't wait for
+ * our own quiescent state. This allows using synchronize_rcu() in
+ * threads registered as readers.
+ */
+ smp_mb();
+ if (was_online)
+ STORE_SHARED(rcu_reader_qs_gp, 0);
+
internal_urcu_lock();
- force_mb_all_threads();
- urcu_gp_ctr += 2;
+ STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR);
wait_for_quiescent_state();
- force_mb_all_threads();
internal_urcu_unlock();
+
+ if (was_online)
+ _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
+ smp_mb();
}
+#endif /* !(BITS_PER_LONG < 64) */
/*
* library wrappers to be used by non-LGPL compatible source code.
return STORE_SHARED(p, v);
}
+void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
+{
+ wmb();
+ return cmpxchg(p, old, _new);
+}
+
void *rcu_xchg_pointer_sym(void **p, void *v)
{
wmb();
return oldptr;
}
+void rcu_quiescent_state(void)
+{
+ _rcu_quiescent_state();
+}
+
+void rcu_thread_offline(void)
+{
+ _rcu_thread_offline();
+}
+
+void rcu_thread_online(void)
+{
+ _rcu_thread_online();
+}
+
static void rcu_add_reader(pthread_t id)
{
struct reader_registry *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;
+ registry[num_readers].rcu_reader_qs_gp = &rcu_reader_qs_gp;
num_readers++;
}
memcpy(index, ®istry[num_readers - 1],
sizeof(struct reader_registry));
registry[num_readers - 1].tid = 0;
- registry[num_readers - 1].urcu_active_readers = NULL;
+ registry[num_readers - 1].rcu_reader_qs_gp = NULL;
num_readers--;
return;
}
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();
+ _rcu_thread_online();
}
void rcu_unregister_thread(void)
{
+ /*
+ * We have to make the thread offline otherwise we end up dealocking
+ * with a waiting writer.
+ */
+ _rcu_thread_offline();
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 */
projects / urcu.git / blobdiff