#include "urcu/list.h"
#include "urcu/futex.h"
#include "urcu/tls-compat.h"
+#include "urcu/ref.h"
#include "urcu-die.h"
/* Data structure that identifies a call_rcu thread. */
struct cds_list_head list;
} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+struct call_rcu_completion {
+ int barrier_count;
+ int32_t futex;
+ struct urcu_ref ref;
+};
+
+struct call_rcu_completion_work {
+ struct rcu_head head;
+ struct call_rcu_completion *completion;
+};
+
/*
* List of all call_rcu_data structures to keep valgrind happy.
* Protected by call_rcu_mutex.
* CPUs rather than only to specific threads.
*/
-#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+#ifdef HAVE_SCHED_GETCPU
+
+static int urcu_sched_getcpu(void)
+{
+ return sched_getcpu();
+}
+
+#else /* #ifdef HAVE_SCHED_GETCPU */
+
+static int urcu_sched_getcpu(void)
+{
+ return -1;
+}
+
+#endif /* #else #ifdef HAVE_SCHED_GETCPU */
+
+#if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU)
/*
* Pointer to array of pointers to per-CPU call_rcu_data structures
}
}
-#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+#else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
/*
* per_cpu_call_rcu_data should be constant, but some functions below, used both
{
}
-static int sched_getcpu(void)
-{
- return -1;
-}
-
-#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+#endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
/* Acquire the specified pthread mutex. */
}
}
+static void call_rcu_completion_wait(struct call_rcu_completion *completion)
+{
+ /* Read completion barrier count before read futex */
+ cmm_smp_mb();
+ if (uatomic_read(&completion->futex) == -1)
+ futex_async(&completion->futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0);
+}
+
+static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
+{
+ /* Write to completion barrier count before reading/writing futex */
+ cmm_smp_mb();
+ if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
+ uatomic_set(&completion->futex, 0);
+ futex_async(&completion->futex, FUTEX_WAKE, 1,
+ NULL, NULL, 0);
+ }
+}
+
/* This is the code run by each call_rcu thread. */
static void *call_rcu_thread(void *arg)
uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
poll(NULL, 0, 1);
+ uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
+ cmm_smp_mb__after_uatomic_and();
rcu_register_thread();
}
* Return a pointer to the call_rcu_data structure for the specified
* CPU, returning NULL if there is none. We cannot automatically
* created it because the platform we are running on might not define
- * sched_getcpu().
+ * urcu_sched_getcpu().
*
* The call to this function and use of the returned call_rcu_data
* should be protected by RCU read-side lock.
return URCU_TLS(thread_call_rcu_data);
if (maxcpus > 0) {
- crd = get_cpu_call_rcu_data(sched_getcpu());
+ crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
if (crd)
return crd;
}
call_rcu_wake_up(crdp);
}
+static void _call_rcu(struct rcu_head *head,
+ void (*func)(struct rcu_head *head),
+ struct call_rcu_data *crdp)
+{
+ cds_wfcq_node_init(&head->next);
+ head->func = func;
+ cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
+ uatomic_inc(&crdp->qlen);
+ wake_call_rcu_thread(crdp);
+}
+
/*
* Schedule a function to be invoked after a following grace period.
* This is the only function that must be called -- the others are
*
* call_rcu must be called by registered RCU read-side threads.
*/
-
void call_rcu(struct rcu_head *head,
void (*func)(struct rcu_head *head))
{
struct call_rcu_data *crdp;
- cds_wfcq_node_init(&head->next);
- head->func = func;
/* Holding rcu read-side lock across use of per-cpu crdp */
rcu_read_lock();
crdp = get_call_rcu_data();
- cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
- uatomic_inc(&crdp->qlen);
- wake_call_rcu_thread(crdp);
+ _call_rcu(head, func, crdp);
rcu_read_unlock();
}
free(crdp);
}
+static
+void free_completion(struct urcu_ref *ref)
+{
+ struct call_rcu_completion *completion;
+
+ completion = caa_container_of(ref, struct call_rcu_completion, ref);
+ free(completion);
+}
+
+static
+void _rcu_barrier_complete(struct rcu_head *head)
+{
+ struct call_rcu_completion_work *work;
+ struct call_rcu_completion *completion;
+
+ work = caa_container_of(head, struct call_rcu_completion_work, head);
+ completion = work->completion;
+ if (!uatomic_sub_return(&completion->barrier_count, 1))
+ call_rcu_completion_wake_up(completion);
+ urcu_ref_put(&completion->ref, free_completion);
+ free(work);
+}
+
+/*
+ * Wait for all in-flight call_rcu callbacks to complete execution.
+ */
+void rcu_barrier(void)
+{
+ struct call_rcu_data *crdp;
+ struct call_rcu_completion *completion;
+ int count = 0;
+ int was_online;
+
+ /* Put in offline state in QSBR. */
+ was_online = rcu_read_ongoing();
+ if (was_online)
+ rcu_thread_offline();
+ /*
+ * Calling a rcu_barrier() within a RCU read-side critical
+ * section is an error.
+ */
+ if (rcu_read_ongoing()) {
+ static int warned = 0;
+
+ if (!warned) {
+ fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
+ }
+ warned = 1;
+ goto online;
+ }
+
+ completion = calloc(sizeof(*completion), 1);
+ if (!completion)
+ urcu_die(errno);
+
+ call_rcu_lock(&call_rcu_mutex);
+ cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
+ count++;
+
+ /* Referenced by rcu_barrier() and each call_rcu thread. */
+ urcu_ref_set(&completion->ref, count + 1);
+ completion->barrier_count = count;
+
+ cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
+ struct call_rcu_completion_work *work;
+
+ work = calloc(sizeof(*work), 1);
+ if (!work)
+ urcu_die(errno);
+ work->completion = completion;
+ _call_rcu(&work->head, _rcu_barrier_complete, crdp);
+ }
+ call_rcu_unlock(&call_rcu_mutex);
+
+ /* Wait for them */
+ for (;;) {
+ uatomic_dec(&completion->futex);
+ /* Decrement futex before reading barrier_count */
+ cmm_smp_mb();
+ if (!uatomic_read(&completion->barrier_count))
+ break;
+ call_rcu_completion_wait(completion);
+ }
+
+ urcu_ref_put(&completion->ref, free_completion);
+
+online:
+ if (was_online)
+ rcu_thread_online();
+}
+
/*
* Acquire the call_rcu_mutex in order to ensure that the child sees
* all of the call_rcu() data structures in a consistent state. Ensure
cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
+ cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
+ while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
+ poll(NULL, 0, 1);
+ }
call_rcu_unlock(&call_rcu_mutex);
}