#include <poll.h>
#include <stdlib.h>
#include <unistd.h>
+#include <stdio.h>
#include <sys/syscall.h>
#include <linux/membarrier.h>
#include "rcu.h"
#include "smp.h"
+/*
+ * If both rseq (with glibc support) and membarrier system calls are
+ * available, use them to replace barriers and atomics on the fast-path.
+ */
+unsigned int side_rcu_rseq_membarrier_available;
+
static int
membarrier(int cmd, unsigned int flags, int cpu_id)
{
return syscall(__NR_membarrier, cmd, flags, cpu_id);
}
+/*
+ * Wait/wakeup scheme with single waiter/many wakers.
+ */
+static
+void wait_gp_prepare(struct side_rcu_gp_state *gp_state)
+{
+ __atomic_store_n(&gp_state->futex, -1, __ATOMIC_RELAXED);
+ /*
+ * This memory barrier (H) pairs with memory barrier (F). It
+ * orders store to futex before load of RCU reader's counter
+ * state, thus ensuring that load of RCU reader's counters does
+ * not leak outside of futex state=-1.
+ */
+ if (side_rcu_rseq_membarrier_available) {
+ if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) {
+ perror("membarrier");
+ abort();
+ }
+ } else {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ }
+}
+
+static
+void wait_gp_end(struct side_rcu_gp_state *gp_state)
+{
+ /*
+ * This memory barrier (G) pairs with memory barrier (F). It
+ * orders load of RCU reader's counter state before storing the
+ * futex value, thus ensuring that load of RCU reader's counters
+ * does not leak outside of futex state=-1.
+ */
+ if (side_rcu_rseq_membarrier_available) {
+ if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) {
+ perror("membarrier");
+ abort();
+ }
+ } else {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ }
+ __atomic_store_n(&gp_state->futex, 0, __ATOMIC_RELAXED);
+}
+
+static
+void wait_gp(struct side_rcu_gp_state *gp_state)
+{
+ /*
+ * This memory barrier (G) pairs with memory barrier (F). It
+ * orders load of RCU reader's counter state before loading the
+ * futex value.
+ */
+ if (side_rcu_rseq_membarrier_available) {
+ if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) {
+ perror("membarrier");
+ abort();
+ }
+ } else {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ }
+ while (__atomic_load_n(&gp_state->futex, __ATOMIC_RELAXED) == -1) {
+ if (!futex(&gp_state->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
+ /*
+ * May be awakened by either spurious wake up or
+ * because the state is now as expected.
+ */
+ continue;
+ }
+ switch (errno) {
+ case EWOULDBLOCK:
+ /* Value already changed. */
+ return;
+ case EINTR:
+ /* Retry if interrupted by signal. */
+ break; /* Get out of switch. */
+ default:
+ /* Unexpected error. */
+ abort();
+ }
+ }
+ return;
+}
+
/* active_readers is an input/output parameter. */
static
void check_active_readers(struct side_rcu_gp_state *gp_state, bool *active_readers)
* incremented before "end", as guaranteed by memory barriers
* (A) or (B).
*/
- if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0))
- abort();
+ if (side_rcu_rseq_membarrier_available) {
+ if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) {
+ perror("membarrier");
+ abort();
+ }
+ } else {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ }
for (i = 0; i < gp_state->nr_cpus; i++) {
struct side_rcu_cpu_gp_state *cpu_state = &gp_state->percpu_state[i];
* previous period.
*/
for (;;) {
+ wait_gp_prepare(gp_state);
check_active_readers(gp_state, active_readers);
- if (!active_readers[prev_period])
+ if (!active_readers[prev_period]) {
+ wait_gp_end(gp_state);
break;
- /* Retry after 10ms. */
- poll(NULL, 0, 10);
+ }
+ wait_gp(gp_state);
}
}
* exist after the grace period completes are ordered after
* loads and stores performed before the grace period.
*/
- if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0))
- abort();
+ if (side_rcu_rseq_membarrier_available) {
+ if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) {
+ perror("membarrier");
+ abort();
+ }
+ } else {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ }
/*
* First scan through all cpus, for both period. If no readers
* are ordered before loads and stores performed after the grace
* period.
*/
- if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0))
- abort();
+ if (side_rcu_rseq_membarrier_available) {
+ if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0, 0)) {
+ perror("membarrier");
+ abort();
+ }
+ } else {
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+ }
}
void side_rcu_gp_init(struct side_rcu_gp_state *rcu_gp)
{
+ bool has_membarrier = false, has_rseq = false;
+
memset(rcu_gp, 0, sizeof(*rcu_gp));
rcu_gp->nr_cpus = get_possible_cpus_array_len();
if (!rcu_gp->nr_cpus)
abort();
pthread_mutex_init(&rcu_gp->gp_lock, NULL);
- rcu_gp->percpu_state = calloc(rcu_gp->nr_cpus, sizeof(struct side_rcu_cpu_gp_state));
+ rcu_gp->percpu_state = (struct side_rcu_cpu_gp_state *)
+ calloc(rcu_gp->nr_cpus, sizeof(struct side_rcu_cpu_gp_state));
if (!rcu_gp->percpu_state)
abort();
- if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0, 0))
- abort();
+ if (!membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0, 0))
+ has_membarrier = true;
+ if (rseq_available(RSEQ_AVAILABLE_QUERY_LIBC))
+ has_rseq = true;
+ if (has_membarrier && has_rseq)
+ side_rcu_rseq_membarrier_available = 1;
}
void side_rcu_gp_exit(struct side_rcu_gp_state *rcu_gp)