/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
#include "urcu-qsbr.h"
-pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+void __attribute__((destructor)) rcu_exit(void);
+
+static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+int gp_futex;
/*
* Global grace period counter.
* Written to only by each individual reader. Read by both the reader and the
* writers.
*/
-unsigned long __thread rcu_reader_qs_gp;
-
-/* Thread IDs of registered readers */
-#define INIT_NUM_THREADS 4
-
-struct reader_registry {
- pthread_t tid;
- unsigned long *rcu_reader_qs_gp;
-};
+struct urcu_reader __thread urcu_reader;
#ifdef DEBUG_YIELD
unsigned int yield_active;
unsigned int __thread rand_yield;
#endif
-static struct reader_registry *registry;
-static int num_readers, alloc_readers;
+static LIST_HEAD(registry);
static void internal_urcu_lock(void)
{
}
}
-#ifdef HAS_INCOHERENT_CACHES
-static void force_mb_single_thread(struct reader_registry *index)
+/*
+ * synchronize_rcu() waiting. Single thread.
+ */
+static void wait_gp(void)
{
- smp_mb();
+ /* Read reader_gp before read futex */
+ smp_rmb();
+ if (uatomic_read(&gp_futex) == -1)
+ futex_noasync(&gp_futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0);
}
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
static void wait_for_quiescent_state(void)
{
- struct reader_registry *index;
+ LIST_HEAD(qsreaders);
+ int wait_loops = 0;
+ struct urcu_reader *index, *tmp;
- if (!registry)
+ if (list_empty(®istry))
return;
/*
* Wait for each thread rcu_reader_qs_gp count to become 0.
*/
- for (index = registry; index < registry + num_readers; index++) {
-#ifndef HAS_INCOHERENT_CACHES
- while (rcu_gp_ongoing(index->rcu_reader_qs_gp))
- cpu_relax();
-#else /* #ifndef HAS_INCOHERENT_CACHES */
- int wait_loops = 0;
- /*
- * BUSY-LOOP. Force the reader thread to commit its
- * rcu_reader_qs_gp update to memory if we wait for too long.
- */
- while (rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
- if (wait_loops++ == KICK_READER_LOOPS) {
- force_mb_single_thread(index);
- wait_loops = 0;
+ for (;;) {
+ wait_loops++;
+ if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ uatomic_dec(&gp_futex);
+ /* Write futex before read reader_gp */
+ smp_mb();
+ }
+
+ list_for_each_entry_safe(index, tmp, ®istry, head) {
+ if (!rcu_gp_ongoing(&index->ctr))
+ list_move(&index->head, &qsreaders);
+ }
+
+ if (list_empty(®istry)) {
+ if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ /* Read reader_gp before write futex */
+ smp_mb();
+ uatomic_set(&gp_futex, 0);
+ }
+ break;
+ } else {
+ if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ wait_gp();
} else {
+#ifndef HAS_INCOHERENT_CACHES
cpu_relax();
+#else /* #ifndef HAS_INCOHERENT_CACHES */
+ smp_mb();
+#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
}
}
-#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
}
+ /* put back the reader list in the registry */
+ list_splice(&qsreaders, ®istry);
+}
+
+/*
+ * 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)
+{
+ 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;
+ was_online = urcu_reader.ctr;
+
+ /* All threads should read qparity before accessing data structure
+ * where new ptr points to.
+ */
+ /* Write new ptr before changing the qparity */
+ smp_mb();
+
+ /*
+ * 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.
+ */
+ if (was_online)
+ STORE_SHARED(urcu_reader.ctr, 0);
+
+ internal_urcu_lock();
+
+ 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.
+ */
+
+ /*
+ * 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.
+ */
+
+ 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.
+ */
+
+ /*
+ * 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(urcu_reader.ctr, LOAD_SHARED(urcu_gp_ctr));
+ smp_mb();
+}
+#else /* !(BITS_PER_LONG < 64) */
+void synchronize_rcu(void)
+{
+ unsigned long was_online;
+
+ was_online = urcu_reader.ctr;
/*
* Mark the writer thread offline to make sure we don't wait for
*/
smp_mb();
if (was_online)
- STORE_SHARED(rcu_reader_qs_gp, 0);
+ STORE_SHARED(urcu_reader.ctr, 0);
internal_urcu_lock();
- STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_COUNT);
+ STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR);
wait_for_quiescent_state();
internal_urcu_unlock();
if (was_online)
- _STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
+ _STORE_SHARED(urcu_reader.ctr, LOAD_SHARED(urcu_gp_ctr));
smp_mb();
}
+#endif /* !(BITS_PER_LONG < 64) */
/*
* library wrappers to be used by non-LGPL compatible source code.
_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 rcu_quiescent_state(void)
{
_rcu_quiescent_state();
_rcu_thread_online();
}
-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].rcu_reader_qs_gp = &rcu_reader_qs_gp;
- 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, ®istry[num_readers - 1],
- sizeof(struct reader_registry));
- registry[num_readers - 1].tid = 0;
- registry[num_readers - 1].rcu_reader_qs_gp = NULL;
- num_readers--;
- return;
- }
- }
- /* Hrm not found, forgot to register ? */
- assert(0);
-}
-
void rcu_register_thread(void)
{
+ urcu_reader.tid = pthread_self();
+ assert(urcu_reader.ctr == 0);
+
internal_urcu_lock();
- rcu_add_reader(pthread_self());
+ list_add(&urcu_reader.head, ®istry);
internal_urcu_unlock();
_rcu_thread_online();
}
*/
_rcu_thread_offline();
internal_urcu_lock();
- rcu_remove_reader(pthread_self());
+ list_del(&urcu_reader.head);
internal_urcu_unlock();
}
+
+void rcu_exit(void)
+{
+ assert(list_empty(®istry));
+}