+// SPDX-FileCopyrightText: 2008 Paul E. McKenney, IBM Corporation.
+//
+// SPDX-License-Identifier: GPL-2.0-or-later
+
/*
* rcutorture.h: simple user-level performance/stress test of RCU.
*
* data. A correct RCU implementation will have all but the first two
* numbers non-zero.
*
+ * rcu_stress_count: Histogram of "ages" of structures seen by readers. If any
+ * entries past the first two are non-zero, RCU is broken. The age of a newly
+ * allocated structure is zero, it becomes one when removed from reader
+ * visibility, and is incremented once per grace period subsequently -- and is
+ * freed after passing through (RCU_STRESS_PIPE_LEN-2) grace periods. Since
+ * this tests only has one true writer (there are fake writers), only buckets at
+ * indexes 0 and 1 should be none-zero.
+ *
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
#include <stdlib.h>
#include "tap.h"
+#include <urcu/uatomic.h>
+
+#include "urcu-wait.h"
+
#define NR_TESTS 1
DEFINE_PER_THREAD(long long, n_reads_pt);
CALLRCU_PERTHREAD,
};
+enum writer_state {
+ WRITER_STATE_SYNC_RCU,
+ WRITER_STATE_CALL_RCU,
+ WRITER_STATE_POLL_RCU,
+};
+
static enum callrcu_type callrcu_type = CALLRCU_GLOBAL;
long long n_reads = 0LL;
#endif
#ifndef mark_rcu_quiescent_state
-#define mark_rcu_quiescent_state() do ; while (0)
+#define mark_rcu_quiescent_state() do {} while (0)
#endif /* #ifdef mark_rcu_quiescent_state */
#ifndef put_thread_offline
-#define put_thread_offline() do ; while (0)
-#define put_thread_online() do ; while (0)
-#define put_thread_online_delay() do ; while (0)
+#define put_thread_offline() do {} while (0)
+#define put_thread_online() do {} while (0)
+#define put_thread_online_delay() do {} while (0)
#else /* #ifndef put_thread_offline */
#define put_thread_online_delay() synchronize_rcu()
#endif /* #else #ifndef put_thread_offline */
* Performance test.
*/
+static
void *rcu_read_perf_test(void *arg)
{
int i;
run_on(me);
uatomic_inc(&nthreadsrunning);
put_thread_offline();
- while (goflag == GOFLAG_INIT)
+ while (uatomic_read(&goflag) == GOFLAG_INIT)
(void) poll(NULL, 0, 1);
put_thread_online();
- while (goflag == GOFLAG_RUN) {
+ while (uatomic_read(&goflag) == GOFLAG_RUN) {
for (i = 0; i < RCU_READ_RUN; i++) {
rcu_read_lock();
/* rcu_read_lock_nest(); */
return (NULL);
}
-void *rcu_update_perf_test(void *arg)
+static
+void *rcu_update_perf_test(void *arg __attribute__((unused)))
{
long long n_updates_local = 0;
}
}
uatomic_inc(&nthreadsrunning);
- while (goflag == GOFLAG_INIT)
+ while (uatomic_read(&goflag) == GOFLAG_INIT)
(void) poll(NULL, 0, 1);
- while (goflag == GOFLAG_RUN) {
+ while (uatomic_read(&goflag) == GOFLAG_RUN) {
synchronize_rcu();
n_updates_local++;
}
return NULL;
}
+static
void perftestinit(void)
{
init_per_thread(n_reads_pt, 0LL);
uatomic_set(&nthreadsrunning, 0);
}
+static
int perftestrun(int nthreads, int nreaders, int nupdaters)
{
int t;
int duration = 1;
- cmm_smp_mb();
while (uatomic_read(&nthreadsrunning) < nthreads)
(void) poll(NULL, 0, 1);
- goflag = GOFLAG_RUN;
- cmm_smp_mb();
+ uatomic_set(&goflag, GOFLAG_RUN);
sleep(duration);
- cmm_smp_mb();
- goflag = GOFLAG_STOP;
- cmm_smp_mb();
+ uatomic_set(&goflag, GOFLAG_STOP);
wait_all_threads();
for_each_thread(t) {
n_reads += per_thread(n_reads_pt, t);
return 0;
}
+static
int perftest(int nreaders, int cpustride)
{
int i;
return perftestrun(i + 1, nreaders, 1);
}
+static
int rperftest(int nreaders, int cpustride)
{
int i;
return perftestrun(i, nreaders, 0);
}
+static
int uperftest(int nupdaters, int cpustride)
{
int i;
int mbtest;
};
-struct rcu_stress rcu_stress_array[RCU_STRESS_PIPE_LEN] = { { 0 } };
+struct rcu_stress rcu_stress_array[RCU_STRESS_PIPE_LEN] = { { 0, 0 } };
struct rcu_stress *rcu_stress_current;
int rcu_stress_idx = 0;
+/*
+ * How many time a reader has seen something that should not be visible. It is
+ * an error if this value is different than zero at the end of the stress test.
+ *
+ * Here, the something that should not be visibile is an old pipe that has been
+ * freed (mbtest = 0).
+ */
int n_mberror = 0;
DEFINE_PER_THREAD(long long [RCU_STRESS_PIPE_LEN + 1], rcu_stress_count);
int garbage = 0;
-void *rcu_read_stress_test(void *arg)
+static
+void *rcu_read_stress_test(void *arg __attribute__((unused)))
{
int i;
int itercnt = 0;
rcu_register_thread();
put_thread_offline();
- while (goflag == GOFLAG_INIT)
+ while (uatomic_read(&goflag) == GOFLAG_INIT)
(void) poll(NULL, 0, 1);
put_thread_online();
- while (goflag == GOFLAG_RUN) {
+ while (uatomic_read(&goflag) == GOFLAG_RUN) {
rcu_read_lock();
p = rcu_dereference(rcu_stress_current);
if (p->mbtest == 0)
- n_mberror++;
+ uatomic_inc_mo(&n_mberror, CMM_RELAXED);
rcu_read_lock_nest();
+ /*
+ * The value of garbage is nothing important. This is
+ * essentially a busy loop. The atomic operation -- while not
+ * important here -- helps tools such as TSAN to not flag this
+ * as a race condition.
+ */
for (i = 0; i < 100; i++)
- garbage++;
+ uatomic_inc(&garbage);
rcu_read_unlock_nest();
- pc = p->pipe_count;
+ pc = uatomic_read(&p->pipe_count);
rcu_read_unlock();
if ((pc > RCU_STRESS_PIPE_LEN) || (pc < 0))
pc = RCU_STRESS_PIPE_LEN;
return (NULL);
}
-static pthread_mutex_t call_rcu_test_mutex = PTHREAD_MUTEX_INITIALIZER;
-static pthread_cond_t call_rcu_test_cond = PTHREAD_COND_INITIALIZER;
+static DEFINE_URCU_WAIT_QUEUE(call_rcu_waiters);
-void rcu_update_stress_test_rcu(struct rcu_head *head)
+static
+void rcu_update_stress_test_rcu(struct rcu_head *head __attribute__((unused)))
{
- int ret;
+ struct urcu_waiters waiters;
- ret = pthread_mutex_lock(&call_rcu_test_mutex);
- if (ret) {
- errno = ret;
- diag("pthread_mutex_lock: %s",
- strerror(errno));
- abort();
- }
- ret = pthread_cond_signal(&call_rcu_test_cond);
- if (ret) {
- errno = ret;
- diag("pthread_cond_signal: %s",
- strerror(errno));
- abort();
- }
- ret = pthread_mutex_unlock(&call_rcu_test_mutex);
- if (ret) {
- errno = ret;
- diag("pthread_mutex_unlock: %s",
- strerror(errno));
- abort();
+ urcu_move_waiters(&waiters, &call_rcu_waiters);
+ urcu_wake_all_waiters(&waiters);
+}
+
+static
+void advance_writer_state(enum writer_state *state)
+{
+ switch (*state) {
+ case WRITER_STATE_SYNC_RCU:
+ *state = WRITER_STATE_CALL_RCU;
+ break;
+ case WRITER_STATE_CALL_RCU:
+ *state = WRITER_STATE_POLL_RCU;
+ break;
+ case WRITER_STATE_POLL_RCU:
+ *state = WRITER_STATE_SYNC_RCU;
+ break;
}
}
-void *rcu_update_stress_test(void *arg)
+static
+void *rcu_update_stress_test(void *arg __attribute__((unused)))
{
int i;
- struct rcu_stress *p;
+ struct rcu_stress *p, *old_p;
struct rcu_head rh;
+ enum writer_state writer_state = WRITER_STATE_SYNC_RCU;
+
+ rcu_register_thread();
- while (goflag == GOFLAG_INIT)
+ /* Offline for poll. */
+ put_thread_offline();
+ while (uatomic_read(&goflag) == GOFLAG_INIT)
(void) poll(NULL, 0, 1);
- while (goflag == GOFLAG_RUN) {
+ put_thread_online();
+
+ old_p = NULL;
+ while (uatomic_read(&goflag) == GOFLAG_RUN) {
i = rcu_stress_idx + 1;
if (i >= RCU_STRESS_PIPE_LEN)
i = 0;
+
+ rcu_read_lock();
+ old_p = rcu_dereference(rcu_stress_current);
+ rcu_read_unlock();
+
+ /*
+ * Allocate a new pipe.
+ */
p = &rcu_stress_array[i];
- p->mbtest = 0;
- cmm_smp_mb();
p->pipe_count = 0;
p->mbtest = 1;
+
rcu_assign_pointer(rcu_stress_current, p);
rcu_stress_idx = i;
+
+ /*
+ * Increment every pipe except the freshly allocated one. A
+ * reader should only see either the old pipe or the new
+ * pipe. This is reflected in the rcu_stress_count histogram.
+ */
for (i = 0; i < RCU_STRESS_PIPE_LEN; i++)
if (i != rcu_stress_idx)
- rcu_stress_array[i].pipe_count++;
- if (n_updates & 0x1)
+ uatomic_inc(&rcu_stress_array[i].pipe_count);
+
+ switch (writer_state) {
+ case WRITER_STATE_SYNC_RCU:
synchronize_rcu();
- else {
- int ret;
-
- ret = pthread_mutex_lock(&call_rcu_test_mutex);
- if (ret) {
- errno = ret;
- diag("pthread_mutex_lock: %s",
- strerror(errno));
- abort();
- }
- rcu_register_thread();
+ break;
+ case WRITER_STATE_CALL_RCU:
+ {
+ DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
+
+ urcu_wait_add(&call_rcu_waiters, &wait);
+
call_rcu(&rh, rcu_update_stress_test_rcu);
- rcu_unregister_thread();
- /*
- * Our MacOS X test machine with the following
- * config:
- * 15.6.0 Darwin Kernel Version 15.6.0
- * root:xnu-3248.60.10~1/RELEASE_X86_64
- * appears to have issues with liburcu-signal
- * signal being delivered on top of
- * pthread_cond_wait. It seems to make the
- * thread continue, and therefore corrupt the
- * rcu_head. Work around this issue by
- * unregistering the RCU read-side thread
- * immediately after call_rcu (call_rcu needs
- * us to be registered RCU readers).
- */
- ret = pthread_cond_wait(&call_rcu_test_cond,
- &call_rcu_test_mutex);
- if (ret) {
- errno = ret;
- diag("pthread_cond_signal: %s",
- strerror(errno));
- abort();
- }
- ret = pthread_mutex_unlock(&call_rcu_test_mutex);
- if (ret) {
- errno = ret;
- diag("pthread_mutex_unlock: %s",
- strerror(errno));
- abort();
- }
+
+ /* Offline for busy-wait. */
+ put_thread_offline();
+ urcu_adaptative_busy_wait(&wait);
+ put_thread_online();
+ break;
}
+ case WRITER_STATE_POLL_RCU:
+ {
+ struct urcu_gp_poll_state poll_state;
+
+ poll_state = start_poll_synchronize_rcu();
+
+ /* Offline for poll. */
+ put_thread_offline();
+ while (!poll_state_synchronize_rcu(poll_state))
+ (void) poll(NULL, 0, 1); /* Wait for 1ms */
+ put_thread_online();
+ break;
+ }
+ }
+ /*
+ * No readers should see that old pipe now. Setting mbtest to 0
+ * to mark it as "freed".
+ */
+ if (old_p) {
+ old_p->mbtest = 0;
+ }
+ old_p = p;
n_updates++;
+ advance_writer_state(&writer_state);
}
+ rcu_unregister_thread();
+
return NULL;
}
-void *rcu_fake_update_stress_test(void *arg)
+static
+void *rcu_fake_update_stress_test(void *arg __attribute__((unused)))
{
if (callrcu_type == CALLRCU_PERTHREAD) {
struct call_rcu_data *crdp;
set_thread_call_rcu_data(crdp);
}
}
- while (goflag == GOFLAG_INIT)
+ while (uatomic_read(&goflag) == GOFLAG_INIT)
(void) poll(NULL, 0, 1);
- while (goflag == GOFLAG_RUN) {
+ while (uatomic_read(&goflag) == GOFLAG_RUN) {
synchronize_rcu();
(void) poll(NULL, 0, 1);
}
return NULL;
}
+static
int stresstest(int nreaders)
{
int i;
int t;
long long *p;
long long sum;
+ int ret;
init_per_thread(n_reads_pt, 0LL);
for_each_thread(t) {
create_thread(rcu_update_stress_test, NULL);
for (i = 0; i < 5; i++)
create_thread(rcu_fake_update_stress_test, NULL);
- cmm_smp_mb();
- goflag = GOFLAG_RUN;
- cmm_smp_mb();
+ uatomic_set(&goflag, GOFLAG_RUN);
sleep(10);
- cmm_smp_mb();
- goflag = GOFLAG_STOP;
- cmm_smp_mb();
+ uatomic_set(&goflag, GOFLAG_STOP);
wait_all_threads();
for_each_thread(t)
n_reads += per_thread(n_reads_pt, t);
n_reads, n_updates, n_mberror);
rdiag_start();
rdiag("rcu_stress_count:");
+ ret = 0;
for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++) {
sum = 0LL;
for_each_thread(t) {
sum += per_thread(rcu_stress_count, t)[i];
}
+ /*
+ * If any entries past the first two are non-zero, RCU is
+ * broken. See details above about rcu_stress_count.
+ */
+ if (i > 1 && sum != 0) {
+ ret = -1;
+ }
rdiag(" %lld", sum);
}
rdiag_end();
diag("Deallocating per-CPU call_rcu threads.");
free_all_cpu_call_rcu_data();
}
- if (!n_mberror)
- return 0;
- else
- return -1;
+ if (n_mberror)
+ ret = -1;
+ return ret;
}
/*
* Mainprogram.
*/
-void usage(int argc, char *argv[])
+static
+void usage(char *argv[]) __attribute__((__noreturn__));
+
+static
+void usage(char *argv[])
{
diag("Usage: %s nreaders [ perf | rperf | uperf | stress ] [ stride ] [ callrcu_global | callrcu_percpu | callrcu_perthread ]\n", argv[0]);
exit(-1);
} else if (strcmp(callrcu_str, "callrcu_perthread") == 0) {
callrcu_type = CALLRCU_PERTHREAD;
} else {
- usage(argc, argv);
+ usage(argv);
goto end;
}
}
if (argc > 1) {
if (strcmp(argv[1], "-h") == 0
|| strcmp(argv[1], "--help") == 0) {
- usage(argc, argv);
+ usage(argv);
goto end;
}
nreaders = strtoul(argv[1], NULL, 0);
"stresstest readers: %d, stride: %d",
nreaders, cpustride);
else
- usage(argc, argv);
+ usage(argv);
} else {
- usage(argc, argv);
+ usage(argv);
}
end:
return exit_status();