*
* Userspace RCU library - batch memory reclamation
*
- * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
*
* This library is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
#include <string.h>
#include <errno.h>
#include <poll.h>
+#include <sys/time.h>
+#include <syscall.h>
+#include <unistd.h>
+#include "urcu/urcu-futex.h"
#include "urcu-defer-static.h"
/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
#include "urcu-defer.h"
-void __attribute__((destructor)) urcu_defer_exit(void);
+void __attribute__((destructor)) rcu_defer_exit(void);
extern void synchronize_rcu(void);
/*
- * urcu_defer_mutex nests inside defer_thread_mutex.
+ * rcu_defer_mutex nests inside defer_thread_mutex.
*/
-static pthread_mutex_t urcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
+static pthread_mutex_t rcu_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;
-
-/* Thread IDs of registered deferers */
-#define INIT_NUM_THREADS 4
-
-struct deferer_registry {
- pthread_t tid;
- struct defer_queue *defer_queue;
- unsigned long last_head;
-};
-
-static struct deferer_registry *registry;
-static int num_deferers, alloc_deferers;
-
+static struct defer_queue __thread defer_queue;
+static CDS_LIST_HEAD(registry);
static pthread_t tid_defer;
-static int exit_defer;
-static void internal_urcu_lock(pthread_mutex_t *mutex)
+static void mutex_lock(pthread_mutex_t *mutex)
{
int ret;
perror("Error in pthread mutex lock");
exit(-1);
}
+ pthread_testcancel();
poll(NULL,0,10);
}
#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
}
-static void internal_urcu_unlock(pthread_mutex_t *mutex)
+static void mutex_unlock(pthread_mutex_t *mutex)
{
int ret;
}
}
+/*
+ * Wake-up any waiting defer thread. Called from many concurrent threads.
+ */
+static void wake_up_defer(void)
+{
+ if (unlikely(uatomic_read(&defer_thread_futex) == -1)) {
+ uatomic_set(&defer_thread_futex, 0);
+ futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
+ NULL, NULL, 0);
+ }
+}
+
+static unsigned long rcu_defer_num_callbacks(void)
+{
+ unsigned long num_items = 0, head;
+ struct defer_queue *index;
+
+ mutex_lock(&rcu_defer_mutex);
+ cds_list_for_each_entry(index, ®istry, list) {
+ head = CMM_LOAD_SHARED(index->head);
+ num_items += head - index->tail;
+ }
+ mutex_unlock(&rcu_defer_mutex);
+ return num_items;
+}
+
+/*
+ * Defer thread waiting. Single thread.
+ */
+static void wait_defer(void)
+{
+ uatomic_dec(&defer_thread_futex);
+ cmm_smp_mb(); /* Write futex before read queue */
+ if (rcu_defer_num_callbacks()) {
+ cmm_smp_mb(); /* Read queue before write futex */
+ /* Callbacks are queued, don't wait. */
+ uatomic_set(&defer_thread_futex, 0);
+ } else {
+ cmm_smp_rmb(); /* Read queue before read futex */
+ if (uatomic_read(&defer_thread_futex) == -1)
+ futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0);
+ }
+}
+
/*
* Must be called after Q.S. is reached.
*/
*/
for (i = queue->tail; i != head;) {
- smp_rmb(); /* read head before q[]. */
- p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
+ cmm_smp_rmb(); /* read head before q[]. */
+ p = CMM_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]);
+ p = CMM_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]);
+ p = CMM_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);
+ p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
+ }
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[] */
- STORE_SHARED(queue->tail, i);
+ cmm_smp_mb(); /* push tail after having used q[] */
+ CMM_STORE_SHARED(queue->tail, i);
}
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);
}
-
void rcu_defer_barrier_thread(void)
{
- internal_urcu_lock(&urcu_defer_mutex);
+ mutex_lock(&rcu_defer_mutex);
_rcu_defer_barrier_thread();
- internal_urcu_unlock(&urcu_defer_mutex);
+ mutex_unlock(&rcu_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;
+ struct defer_queue *index;
+ unsigned long num_items = 0;
- if (!registry)
+ if (cds_list_empty(®istry))
return;
- internal_urcu_lock(&urcu_defer_mutex);
- for (index = registry; index < registry + num_deferers; index++)
- index->last_head = LOAD_SHARED(index->defer_queue->head);
+ mutex_lock(&rcu_defer_mutex);
+ cds_list_for_each_entry(index, ®istry, list) {
+ index->last_head = CMM_LOAD_SHARED(index->head);
+ num_items += index->last_head - index->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);
- internal_urcu_unlock(&urcu_defer_mutex);
+ cds_list_for_each_entry(index, ®istry, list)
+ rcu_defer_barrier_queue(index, index->last_head);
+end:
+ mutex_unlock(&rcu_defer_mutex);
+}
+
+/*
+ * _defer_rcu - Queue a RCU callback.
+ */
+void _defer_rcu(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 = CMM_LOAD_SHARED(defer_queue.tail);
+
+ /*
+ * If queue is full, or reached threshold. Empty queue 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 - CMM_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.
+ */
+ _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ DQ_FCT_MARK);
+ _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ fct);
+ } else {
+ DQ_SET_FCT_BIT(fct);
+ _CMM_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.
+ */
+ _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ DQ_FCT_MARK);
+ _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
+ fct);
+ }
+ }
+ _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
+ cmm_smp_wmb(); /* Publish new pointer before head */
+ /* Write q[] before head. */
+ CMM_STORE_SHARED(defer_queue.head, head);
+ cmm_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();
}
* library wrappers to be used by non-LGPL compatible source code.
*/
-void rcu_defer_queue(void (*fct)(void *p), void *p)
-{
- _rcu_defer_queue(fct, p);
-}
-
-static void rcu_add_deferer(pthread_t id)
-{
- struct deferer_registry *oldarray;
-
- if (!registry) {
- alloc_deferers = INIT_NUM_THREADS;
- num_deferers = 0;
- registry =
- malloc(sizeof(struct deferer_registry) * alloc_deferers);
- }
- if (alloc_deferers < num_deferers + 1) {
- oldarray = registry;
- registry = malloc(sizeof(struct deferer_registry)
- * (alloc_deferers << 1));
- memcpy(registry, oldarray,
- sizeof(struct deferer_registry) * alloc_deferers);
- alloc_deferers <<= 1;
- free(oldarray);
- }
- registry[num_deferers].tid = id;
- /* reference to the TLS of _this_ deferer thread. */
- registry[num_deferers].defer_queue = &defer_queue;
- registry[num_deferers].last_head = 0;
- num_deferers++;
-}
-
-/*
- * Never shrink (implementation limitation).
- * This is O(nb threads). Eventually use a hash table.
- */
-static void rcu_remove_deferer(pthread_t id)
+void defer_rcu(void (*fct)(void *p), void *p)
{
- struct deferer_registry *index;
-
- assert(registry != NULL);
- for (index = registry; index < registry + num_deferers; index++) {
- if (pthread_equal(index->tid, id)) {
- memcpy(index, ®istry[num_deferers - 1],
- sizeof(struct deferer_registry));
- registry[num_deferers - 1].tid = 0;
- registry[num_deferers - 1].defer_queue = NULL;
- registry[num_deferers - 1].last_head = 0;
- num_deferers--;
- return;
- }
- }
- /* Hrm not found, forgot to register ? */
- assert(0);
+ _defer_rcu(fct, p);
}
static void start_defer_thread(void)
{
int ret;
- ret = pthread_create(&tid_defer, NULL, thr_defer,
- NULL);
+ ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
assert(!ret);
}
int ret;
void *tret;
- STORE_SHARED(exit_defer, 1);
+ pthread_cancel(tid_defer);
+ wake_up_defer();
ret = pthread_join(tid_defer, &tret);
assert(!ret);
}
void rcu_defer_register_thread(void)
{
- int deferers;
+ int was_empty;
- internal_urcu_lock(&defer_thread_mutex);
- internal_urcu_lock(&urcu_defer_mutex);
+ assert(defer_queue.last_head == 0);
+ assert(defer_queue.q == NULL);
defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
- rcu_add_deferer(pthread_self());
- deferers = num_deferers;
- internal_urcu_unlock(&urcu_defer_mutex);
- if (deferers == 1)
+ mutex_lock(&defer_thread_mutex);
+ mutex_lock(&rcu_defer_mutex);
+ was_empty = cds_list_empty(®istry);
+ cds_list_add(&defer_queue.list, ®istry);
+ mutex_unlock(&rcu_defer_mutex);
+
+ if (was_empty)
start_defer_thread();
- internal_urcu_unlock(&defer_thread_mutex);
+ mutex_unlock(&defer_thread_mutex);
}
void rcu_defer_unregister_thread(void)
{
- int deferers;
+ int is_empty;
- internal_urcu_lock(&defer_thread_mutex);
- internal_urcu_lock(&urcu_defer_mutex);
- rcu_remove_deferer(pthread_self());
+ mutex_lock(&defer_thread_mutex);
+ mutex_lock(&rcu_defer_mutex);
+ cds_list_del(&defer_queue.list);
_rcu_defer_barrier_thread();
free(defer_queue.q);
defer_queue.q = NULL;
- deferers = num_deferers;
- internal_urcu_unlock(&urcu_defer_mutex);
+ is_empty = cds_list_empty(®istry);
+ mutex_unlock(&rcu_defer_mutex);
- if (deferers == 0)
+ if (is_empty)
stop_defer_thread();
- internal_urcu_unlock(&defer_thread_mutex);
+ mutex_unlock(&defer_thread_mutex);
}
-void urcu_defer_exit(void)
+void rcu_defer_exit(void)
{
- free(registry);
+ assert(cds_list_empty(®istry));
}