#include <string.h>
#include <errno.h>
#include <poll.h>
+#include <sys/time.h>
+#include <syscall.h>
+#include <unistd.h>
#include "urcu-defer-static.h"
/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
#include "urcu-defer.h"
+#define futex(...) syscall(__NR_futex, __VA_ARGS__)
+#define FUTEX_WAIT 0
+#define FUTEX_WAKE 1
+
void __attribute__((destructor)) urcu_defer_exit(void);
extern void synchronize_rcu(void);
static pthread_mutex_t urcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
+static int defer_thread_futex;
+
/*
* Written to only by each individual deferer. Read by both the deferer and
* the reclamation tread.
*/
-struct defer_queue __thread defer_queue;
+static struct defer_queue __thread defer_queue;
/* Thread IDs of registered deferers */
#define INIT_NUM_THREADS 4
static int num_deferers, alloc_deferers;
static pthread_t tid_defer;
-static int exit_defer;
static void internal_urcu_lock(pthread_mutex_t *mutex)
{
perror("Error in pthread mutex lock");
exit(-1);
}
+ pthread_testcancel();
poll(NULL,0,10);
}
#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
}
}
+/*
+ * Wake-up any waiting defer thread. Called from many concurrent threads.
+ */
+static void wake_up_defer(void)
+{
+ if (unlikely(atomic_read(&defer_thread_futex) == -1)) {
+ atomic_set(&defer_thread_futex, 0);
+ futex(&defer_thread_futex, FUTEX_WAKE, 1,
+ NULL, NULL, 0);
+ }
+}
+
+static unsigned long rcu_defer_num_callbacks(void)
+{
+ unsigned long num_items = 0, head;
+ struct deferer_registry *index;
+
+ internal_urcu_lock(&urcu_defer_mutex);
+ for (index = registry; index < registry + num_deferers; index++) {
+ head = LOAD_SHARED(index->defer_queue->head);
+ num_items += head - index->defer_queue->tail;
+ }
+ internal_urcu_unlock(&urcu_defer_mutex);
+ return num_items;
+}
+
+/*
+ * Defer thread waiting. Single thread.
+ */
+static void wait_defer(void)
+{
+ atomic_dec(&defer_thread_futex);
+ smp_mb(); /* Write futex before read queue */
+ if (rcu_defer_num_callbacks()) {
+ smp_mb(); /* Read queue before write futex */
+ /* Callbacks are queued, don't wait. */
+ atomic_set(&defer_thread_futex, 0);
+ } else {
+ smp_rmb(); /* Read queue before read futex */
+ if (atomic_read(&defer_thread_futex) == -1)
+ futex(&defer_thread_futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0);
+ }
+}
+
/*
* Must be called after Q.S. is reached.
*/
smp_rmb(); /* read head before q[]. */
p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
if (unlikely(DQ_IS_FCT_BIT(p))) {
- //printf("%lu fct bit %p\n", i-1, p);
DQ_CLEAR_FCT_BIT(p);
queue->last_fct_out = p;
p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
} else if (unlikely(p == DQ_FCT_MARK)) {
- //printf("%lu fct mark %p\n", i-1, p);
p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
queue->last_fct_out = p;
p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
- }// else
- //printf("%lu data %p\n", i-1, p);
+ }
fct = queue->last_fct_out;
- //printf("tid %lu %lu last_fct %p data %p\n", pthread_self(), i-1, fct, p);
fct(p);
}
smp_mb(); /* push tail after having used q[] */
static void _rcu_defer_barrier_thread(void)
{
- unsigned long head;
+ unsigned long head, num_items;
head = defer_queue.head;
+ num_items = head - defer_queue.tail;
+ if (unlikely(!num_items))
+ return;
synchronize_rcu();
rcu_defer_barrier_queue(&defer_queue, head);
}
internal_urcu_unlock(&urcu_defer_mutex);
}
+/*
+ * rcu_defer_barrier - Execute all queued rcu callbacks.
+ *
+ * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
+ * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
+ * are guaranteed to be executed.
+ * Callbacks queued by other threads concurrently with rcu_defer_barrier()
+ * execution are not guaranteed to be executed in the current batch (could
+ * be left for the next batch). These callbacks queued by other threads are only
+ * guaranteed to be executed if there is explicit synchronization between
+ * the thread adding to the queue and the thread issuing the defer_barrier call.
+ */
+
void rcu_defer_barrier(void)
{
struct deferer_registry *index;
+ unsigned long num_items = 0;
if (!registry)
return;
internal_urcu_lock(&urcu_defer_mutex);
- for (index = registry; index < registry + num_deferers; index++)
+ for (index = registry; index < registry + num_deferers; index++) {
index->last_head = LOAD_SHARED(index->defer_queue->head);
+ num_items += index->last_head - index->defer_queue->tail;
+ }
+ if (likely(!num_items)) {
+ /*
+ * We skip the grace period because there are no queued
+ * callbacks to execute.
+ */
+ goto end;
+ }
synchronize_rcu();
for (index = registry; index < registry + num_deferers; index++)
rcu_defer_barrier_queue(index->defer_queue,
index->last_head);
+end:
internal_urcu_unlock(&urcu_defer_mutex);
}
+/*
+ * _rcu_defer_queue - Queue a RCU callback.
+ */
+void _rcu_defer_queue(void (*fct)(void *p), void *p)
+{
+ unsigned long head, tail;
+
+ /*
+ * Head is only modified by ourself. Tail can be modified by reclamation
+ * thread.
+ */
+ head = defer_queue.head;
+ tail = LOAD_SHARED(defer_queue.tail);
+
+ /*
+ * If queue is full, empty it ourself.
+ * Worse-case: must allow 2 supplementary entries for fct pointer.
+ */
+ if (unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) {
+ assert(head - tail <= DEFER_QUEUE_SIZE);
+ rcu_defer_barrier_thread();
+ assert(head - LOAD_SHARED(defer_queue.tail) == 0);
+ }
+
+ if (unlikely(defer_queue.last_fct_in != fct)) {
+ defer_queue.last_fct_in = fct;
+ if (unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) {
+ /*
+ * If the function to encode is not aligned or the
+ * marker, write DQ_FCT_MARK followed by the function
+ * pointer.
+ */
+ _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ DQ_FCT_MARK);
+ _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ fct);
+ } else {
+ DQ_SET_FCT_BIT(fct);
+ _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ fct);
+ }
+ } else {
+ if (unlikely(DQ_IS_FCT_BIT(p) || p == DQ_FCT_MARK)) {
+ /*
+ * If the data to encode is not aligned or the marker,
+ * write DQ_FCT_MARK followed by the function pointer.
+ */
+ _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ DQ_FCT_MARK);
+ _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ fct);
+ }
+ }
+ _STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
+ smp_wmb(); /* Publish new pointer before head */
+ /* Write q[] before head. */
+ STORE_SHARED(defer_queue.head, head);
+ smp_mb(); /* Write queue head before read futex */
+ /*
+ * Wake-up any waiting defer thread.
+ */
+ wake_up_defer();
+}
+
void *thr_defer(void *args)
{
for (;;) {
- if (LOAD_SHARED(exit_defer))
- break;
+ pthread_testcancel();
+ /*
+ * "Be green". Don't wake up the CPU if there is no RCU work
+ * to perform whatsoever. Aims at saving laptop battery life by
+ * leaving the processor in sleep state when idle.
+ */
+ wait_defer();
+ /* Sleeping after wait_defer to let many callbacks enqueue */
poll(NULL,0,100); /* wait for 100ms */
rcu_defer_barrier();
}
int ret;
void *tret;
- STORE_SHARED(exit_defer, 1);
+ pthread_cancel(tid_defer);
+ wake_up_defer();
ret = pthread_join(tid_defer, &tret);
assert(!ret);
}
projects / urcu.git / blobdiff