* 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 */
static struct reader_data *reader_data;
static int num_readers, alloc_readers;
+#ifndef DEBUG_FULL_MB
static int sig_done;
+#endif
void internal_urcu_lock(void)
{
*/
static void switch_next_urcu_qparity(void)
{
- urcu_gp_ctr ^= RCU_GP_CTR_BIT;
+ STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
+}
+
+#ifdef DEBUG_FULL_MB
+static void force_mb_single_thread(pthread_t tid)
+{
+ smp_mb();
+}
+
+static void force_mb_all_threads(void)
+{
+ smp_mb();
+}
+#else
+
+static void force_mb_single_thread(pthread_t tid)
+{
+ assert(reader_data);
+ sig_done = 0;
+ /*
+ * 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.
+ * smp_mb(); write sig_done before sending the signals
+ */
+ smp_mc(); /* write sig_done before sending the signals */
+ pthread_kill(tid, SIGURCU);
+ /*
+ * Wait for sighandler (and thus mb()) to execute on every thread.
+ * BUSY-LOOP.
+ */
+ while (LOAD_SHARED(sig_done) < 1)
+ cpu_relax();
+ smp_mb(); /* read sig_done before ending the barrier */
}
static void force_mb_all_threads(void)
{
struct reader_data *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)
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.
+ * smp_mb(); write sig_done before sending the signals
+ */
+ smp_mc(); /* write sig_done before sending the signals */
+ for (index = reader_data; index < reader_data + num_readers; index++)
pthread_kill(index->tid, SIGURCU);
- debug_yield_write();
- }
/*
* Wait for sighandler (and thus mb()) to execute on every thread.
* BUSY-LOOP.
*/
- while (sig_done < num_readers)
- barrier();
- debug_yield_write();
- mb(); /* read sig_done before ending the barrier */
- debug_yield_write();
+ while (LOAD_SHARED(sig_done) < num_readers)
+ cpu_relax();
+ smp_mb(); /* read sig_done before ending the barrier */
}
+#endif
void wait_for_quiescent_state(void)
{
if (!reader_data)
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++) {
+ 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 (rcu_old_gp_ongoing(index->urcu_active_readers))
- barrier();
+ while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
+ if (wait_loops++ == KICK_READER_LOOPS) {
+ force_mb_single_thread(index->tid);
+ wait_loops = 0;
+ } else {
+ cpu_relax();
+ }
+ }
}
- /*
- * 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)
{
- /* All threads should read qparity before accessing data structure. */
- /* Write ptr before changing the qparity */
+ 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();
- debug_yield_write();
- 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.
+ */
/*
* Wait for previous parity to be empty of readers.
*/
- wait_for_quiescent_state();
-}
+ 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 */
+
+ /* 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();
-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_unlock();
- debug_yield_write();
}
void urcu_add_reader(pthread_t id)
internal_urcu_unlock();
}
+#ifndef DEBUG_FULL_MB
void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
{
- mb();
+ /*
+ * 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();
atomic_inc(&sig_done);
}
assert(act.sa_sigaction == sigurcu_handler);
free(reader_data);
}
+#endif