#include "urcu/list.h"
#include "urcu/futex.h"
#include "urcu/tls-compat.h"
+#include "urcu/ref.h"
#include "urcu-die.h"
+#define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */
+#define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1)
+
/* Data structure that identifies a call_rcu thread. */
struct call_rcu_data {
unsigned long qlen; /* maintained for debugging. */
pthread_t tid;
int cpu_affinity;
+ unsigned long gp_count;
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 {
urcu_die(ret);
}
+/*
+ * Periodically retry setting CPU affinity if we migrate.
+ * Losing affinity can be caused by CPU hotunplug/hotplug, or by
+ * cpuset(7).
+ */
#if HAVE_SCHED_SETAFFINITY
static
int set_thread_cpu_affinity(struct call_rcu_data *crdp)
{
cpu_set_t mask;
+ int ret;
if (crdp->cpu_affinity < 0)
return 0;
+ if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
+ return 0;
+ if (urcu_sched_getcpu() == crdp->cpu_affinity)
+ return 0;
CPU_ZERO(&mask);
CPU_SET(crdp->cpu_affinity, &mask);
#if SCHED_SETAFFINITY_ARGS == 2
- return sched_setaffinity(0, &mask);
+ ret = sched_setaffinity(0, &mask);
#else
- return sched_setaffinity(0, sizeof(mask), &mask);
+ ret = sched_setaffinity(0, sizeof(mask), &mask);
#endif
+ /*
+ * EINVAL is fine: can be caused by hotunplugged CPUs, or by
+ * cpuset(7). This is why we should always retry if we detect
+ * migration.
+ */
+ if (ret && errno == EINVAL) {
+ ret = 0;
+ errno = 0;
+ }
+ return ret;
}
#else
static
{
/* Read call_rcu list before read futex */
cmm_smp_mb();
- if (uatomic_read(&crdp->futex) == -1)
- futex_async(&crdp->futex, FUTEX_WAIT, -1,
- NULL, NULL, 0);
+ if (uatomic_read(&crdp->futex) != -1)
+ return;
+ while (futex_async(&crdp->futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0)) {
+ switch (errno) {
+ case EWOULDBLOCK:
+ /* Value already changed. */
+ return;
+ case EINTR:
+ /* Retry if interrupted by signal. */
+ break; /* Get out of switch. */
+ default:
+ /* Unexpected error. */
+ urcu_die(errno);
+ }
+ }
}
static void call_rcu_wake_up(struct call_rcu_data *crdp)
cmm_smp_mb();
if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
uatomic_set(&crdp->futex, 0);
- futex_async(&crdp->futex, FUTEX_WAKE, 1,
- NULL, NULL, 0);
+ if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
+ NULL, NULL, 0) < 0)
+ urcu_die(errno);
}
}
{
/* 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);
+ if (uatomic_read(&completion->futex) != -1)
+ return;
+ while (futex_async(&completion->futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0)) {
+ switch (errno) {
+ case EWOULDBLOCK:
+ /* Value already changed. */
+ return;
+ case EINTR:
+ /* Retry if interrupted by signal. */
+ break; /* Get out of switch. */
+ default:
+ /* Unexpected error. */
+ urcu_die(errno);
+ }
+ }
}
static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
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);
+ if (futex_async(&completion->futex, FUTEX_WAKE, 1,
+ NULL, NULL, 0) < 0)
+ urcu_die(errno);
}
}
unsigned long cbcount;
struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
- int ret;
- ret = set_thread_cpu_affinity(crdp);
- if (ret)
+ if (set_thread_cpu_affinity(crdp))
urcu_die(errno);
/*
struct cds_wfcq_node *cbs, *cbs_tmp_n;
enum cds_wfcq_ret splice_ret;
+ if (set_thread_cpu_affinity(crdp))
+ urcu_die(errno);
+
if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
/*
* Pause requested. Become quiescent: remove
cmm_smp_mb__before_uatomic_or();
uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
- poll(NULL, 0, 1);
+ (void) poll(NULL, 0, 1);
+ uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
+ cmm_smp_mb__after_uatomic_and();
rcu_register_thread();
}
if (cds_wfcq_empty(&crdp->cbs_head,
&crdp->cbs_tail)) {
call_rcu_wait(crdp);
- poll(NULL, 0, 10);
+ (void) poll(NULL, 0, 10);
uatomic_dec(&crdp->futex);
/*
* Decrement futex before reading
*/
cmm_smp_mb();
} else {
- poll(NULL, 0, 10);
+ (void) poll(NULL, 0, 10);
}
} else {
- poll(NULL, 0, 10);
+ (void) poll(NULL, 0, 10);
}
rcu_thread_online();
}
crdp->flags = flags;
cds_list_add(&crdp->list, &call_rcu_data_list);
crdp->cpu_affinity = cpu_affinity;
+ crdp->gp_count = 0;
cmm_smp_mb(); /* Structure initialized before pointer is planted. */
*crdpp = crdp;
ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
struct call_rcu_data *crdp;
/* Holding rcu read-side lock across use of per-cpu crdp */
- rcu_read_lock();
+ _rcu_read_lock();
crdp = get_call_rcu_data();
_call_rcu(head, func, crdp);
- rcu_read_unlock();
+ _rcu_read_unlock();
}
/*
uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
wake_call_rcu_thread(crdp);
while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
- poll(NULL, 0, 1);
+ (void) poll(NULL, 0, 1);
}
if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
/* Create default call rcu data if need be */
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)
{
work = caa_container_of(head, struct call_rcu_completion_work, head);
completion = work->completion;
- uatomic_dec(&completion->barrier_count);
- call_rcu_completion_wake_up(completion);
+ if (!uatomic_sub_return(&completion->barrier_count, 1))
+ call_rcu_completion_wake_up(completion);
+ urcu_ref_put(&completion->ref, free_completion);
free(work);
}
void rcu_barrier(void)
{
struct call_rcu_data *crdp;
- struct call_rcu_completion completion;
- int count = 0, work_count = 0;
+ struct call_rcu_completion *completion;
+ int count = 0;
int was_online;
/* Put in offline state in QSBR. */
- was_online = rcu_read_ongoing();
+ 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()) {
+ if (_rcu_read_ongoing()) {
static int warned = 0;
if (!warned) {
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++;
- completion.barrier_count = 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) {
- static int warned = 0;
-
- if (!warned) {
- fprintf(stderr, "[error] liburcu: unable to allocate memory for rcu_barrier()\n");
- }
- warned = 1;
- break;
- }
- work->completion = &completion;
+ if (!work)
+ urcu_die(errno);
+ work->completion = completion;
_call_rcu(&work->head, _rcu_barrier_complete, crdp);
- work_count++;
}
call_rcu_unlock(&call_rcu_mutex);
- if (work_count != count)
- uatomic_sub(&completion.barrier_count, count - work_count);
-
/* Wait for them */
for (;;) {
- uatomic_dec(&completion.futex);
+ uatomic_dec(&completion->futex);
/* Decrement futex before reading barrier_count */
cmm_smp_mb();
- if (!uatomic_read(&completion.barrier_count))
+ if (!uatomic_read(&completion->barrier_count))
break;
- call_rcu_completion_wait(&completion);
+ call_rcu_completion_wait(completion);
}
+
+ urcu_ref_put(&completion->ref, free_completion);
+
online:
if (was_online)
rcu_thread_online();
}
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);
+ (void) poll(NULL, 0, 1);
}
}
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)
+ (void) poll(NULL, 0, 1);
+ }
call_rcu_unlock(&call_rcu_mutex);
}