2 * Read-Copy Update module-based performance-test facility
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright (C) IBM Corporation, 2015
20 * Authors: Paul E. McKenney <paulmck@us.ibm.com>
23 #define pr_fmt(fmt) fmt
25 #include <linux/types.h>
26 #include <linux/kernel.h>
27 #include <linux/init.h>
28 #include <linux/module.h>
29 #include <linux/kthread.h>
30 #include <linux/err.h>
31 #include <linux/spinlock.h>
32 #include <linux/smp.h>
33 #include <linux/rcupdate.h>
34 #include <linux/interrupt.h>
35 #include <linux/sched.h>
36 #include <uapi/linux/sched/types.h>
37 #include <linux/atomic.h>
38 #include <linux/bitops.h>
39 #include <linux/completion.h>
40 #include <linux/moduleparam.h>
41 #include <linux/percpu.h>
42 #include <linux/notifier.h>
43 #include <linux/reboot.h>
44 #include <linux/freezer.h>
45 #include <linux/cpu.h>
46 #include <linux/delay.h>
47 #include <linux/stat.h>
48 #include <linux/srcu.h>
49 #include <linux/slab.h>
50 #include <asm/byteorder.h>
51 #include <linux/torture.h>
52 #include <linux/vmalloc.h>
56 MODULE_LICENSE("GPL");
57 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.vnet.ibm.com>");
59 #define PERF_FLAG "-perf:"
60 #define PERFOUT_STRING(s) \
61 pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
62 #define VERBOSE_PERFOUT_STRING(s) \
63 do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
64 #define VERBOSE_PERFOUT_ERRSTRING(s) \
65 do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
68 * The intended use cases for the nreaders and nwriters module parameters
71 * 1. Specify only the nr_cpus kernel boot parameter. This will
72 * set both nreaders and nwriters to the value specified by
73 * nr_cpus for a mixed reader/writer test.
75 * 2. Specify the nr_cpus kernel boot parameter, but set
76 * rcuperf.nreaders to zero. This will set nwriters to the
77 * value specified by nr_cpus for an update-only test.
79 * 3. Specify the nr_cpus kernel boot parameter, but set
80 * rcuperf.nwriters to zero. This will set nreaders to the
81 * value specified by nr_cpus for a read-only test.
83 * Various other use cases may of course be specified.
86 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
87 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
88 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
89 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
90 torture_param(int, nreaders, -1, "Number of RCU reader threads");
91 torture_param(int, nwriters, -1, "Number of RCU updater threads");
92 torture_param(bool, shutdown, !IS_ENABLED(MODULE),
93 "Shutdown at end of performance tests.");
94 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
95 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
97 static char *perf_type = "rcu";
98 module_param(perf_type, charp, 0444);
99 MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, rcu_bh, ...)");
101 static int nrealreaders;
102 static int nrealwriters;
103 static struct task_struct **writer_tasks;
104 static struct task_struct **reader_tasks;
105 static struct task_struct *shutdown_task;
107 static u64 **writer_durations;
108 static int *writer_n_durations;
109 static atomic_t n_rcu_perf_reader_started;
110 static atomic_t n_rcu_perf_writer_started;
111 static atomic_t n_rcu_perf_writer_finished;
112 static wait_queue_head_t shutdown_wq;
113 static u64 t_rcu_perf_writer_started;
114 static u64 t_rcu_perf_writer_finished;
115 static unsigned long b_rcu_perf_writer_started;
116 static unsigned long b_rcu_perf_writer_finished;
117 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
119 static int rcu_perf_writer_state;
122 #define RTWS_BARRIER 2
123 #define RTWS_EXP_SYNC 3
126 #define RTWS_STOPPING 6
128 #define MAX_MEAS 10000
132 * Operations vector for selecting different types of tests.
135 struct rcu_perf_ops {
138 void (*cleanup)(void);
139 int (*readlock)(void);
140 void (*readunlock)(int idx);
141 unsigned long (*get_gp_seq)(void);
142 unsigned long (*gp_diff)(unsigned long new, unsigned long old);
143 unsigned long (*exp_completed)(void);
144 void (*async)(struct rcu_head *head, rcu_callback_t func);
145 void (*gp_barrier)(void);
147 void (*exp_sync)(void);
151 static struct rcu_perf_ops *cur_ops;
154 * Definitions for rcu perf testing.
157 static int rcu_perf_read_lock(void) __acquires(RCU)
163 static void rcu_perf_read_unlock(int idx) __releases(RCU)
168 static unsigned long __maybe_unused rcu_no_completed(void)
173 static void rcu_sync_perf_init(void)
177 static struct rcu_perf_ops rcu_ops = {
179 .init = rcu_sync_perf_init,
180 .readlock = rcu_perf_read_lock,
181 .readunlock = rcu_perf_read_unlock,
182 .get_gp_seq = rcu_get_gp_seq,
183 .gp_diff = rcu_seq_diff,
184 .exp_completed = rcu_exp_batches_completed,
186 .gp_barrier = rcu_barrier,
187 .sync = synchronize_rcu,
188 .exp_sync = synchronize_rcu_expedited,
193 * Definitions for srcu perf testing.
196 DEFINE_STATIC_SRCU(srcu_ctl_perf);
197 static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
199 static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
201 return srcu_read_lock(srcu_ctlp);
204 static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
206 srcu_read_unlock(srcu_ctlp, idx);
209 static unsigned long srcu_perf_completed(void)
211 return srcu_batches_completed(srcu_ctlp);
214 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
216 call_srcu(srcu_ctlp, head, func);
219 static void srcu_rcu_barrier(void)
221 srcu_barrier(srcu_ctlp);
224 static void srcu_perf_synchronize(void)
226 synchronize_srcu(srcu_ctlp);
229 static void srcu_perf_synchronize_expedited(void)
231 synchronize_srcu_expedited(srcu_ctlp);
234 static struct rcu_perf_ops srcu_ops = {
235 .ptype = SRCU_FLAVOR,
236 .init = rcu_sync_perf_init,
237 .readlock = srcu_perf_read_lock,
238 .readunlock = srcu_perf_read_unlock,
239 .get_gp_seq = srcu_perf_completed,
240 .gp_diff = rcu_seq_diff,
241 .exp_completed = srcu_perf_completed,
242 .async = srcu_call_rcu,
243 .gp_barrier = srcu_rcu_barrier,
244 .sync = srcu_perf_synchronize,
245 .exp_sync = srcu_perf_synchronize_expedited,
249 static struct srcu_struct srcud;
251 static void srcu_sync_perf_init(void)
254 init_srcu_struct(srcu_ctlp);
257 static void srcu_sync_perf_cleanup(void)
259 cleanup_srcu_struct(srcu_ctlp);
262 static struct rcu_perf_ops srcud_ops = {
263 .ptype = SRCU_FLAVOR,
264 .init = srcu_sync_perf_init,
265 .cleanup = srcu_sync_perf_cleanup,
266 .readlock = srcu_perf_read_lock,
267 .readunlock = srcu_perf_read_unlock,
268 .get_gp_seq = srcu_perf_completed,
269 .gp_diff = rcu_seq_diff,
270 .exp_completed = srcu_perf_completed,
271 .async = srcu_call_rcu,
272 .gp_barrier = srcu_rcu_barrier,
273 .sync = srcu_perf_synchronize,
274 .exp_sync = srcu_perf_synchronize_expedited,
279 * Definitions for RCU-tasks perf testing.
282 static int tasks_perf_read_lock(void)
287 static void tasks_perf_read_unlock(int idx)
291 static struct rcu_perf_ops tasks_ops = {
292 .ptype = RCU_TASKS_FLAVOR,
293 .init = rcu_sync_perf_init,
294 .readlock = tasks_perf_read_lock,
295 .readunlock = tasks_perf_read_unlock,
296 .get_gp_seq = rcu_no_completed,
297 .gp_diff = rcu_seq_diff,
298 .async = call_rcu_tasks,
299 .gp_barrier = rcu_barrier_tasks,
300 .sync = synchronize_rcu_tasks,
301 .exp_sync = synchronize_rcu_tasks,
305 static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old)
307 if (!cur_ops->gp_diff)
309 return cur_ops->gp_diff(new, old);
313 * If performance tests complete, wait for shutdown to commence.
315 static void rcu_perf_wait_shutdown(void)
317 cond_resched_tasks_rcu_qs();
318 if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
320 while (!torture_must_stop())
321 schedule_timeout_uninterruptible(1);
325 * RCU perf reader kthread. Repeatedly does empty RCU read-side
326 * critical section, minimizing update-side interference.
329 rcu_perf_reader(void *arg)
335 VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
336 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
337 set_user_nice(current, MAX_NICE);
338 atomic_inc(&n_rcu_perf_reader_started);
341 local_irq_save(flags);
342 idx = cur_ops->readlock();
343 cur_ops->readunlock(idx);
344 local_irq_restore(flags);
345 rcu_perf_wait_shutdown();
346 } while (!torture_must_stop());
347 torture_kthread_stopping("rcu_perf_reader");
352 * Callback function for asynchronous grace periods from rcu_perf_writer().
354 static void rcu_perf_async_cb(struct rcu_head *rhp)
356 atomic_dec(this_cpu_ptr(&n_async_inflight));
361 * RCU perf writer kthread. Repeatedly does a grace period.
364 rcu_perf_writer(void *arg)
369 struct rcu_head *rhp = NULL;
370 struct sched_param sp;
371 bool started = false, done = false, alldone = false;
374 u64 *wdpp = writer_durations[me];
376 VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
378 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
379 sp.sched_priority = 1;
380 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
383 schedule_timeout_uninterruptible(holdoff * HZ);
385 t = ktime_get_mono_fast_ns();
386 if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
387 t_rcu_perf_writer_started = t;
389 b_rcu_perf_writer_started =
390 cur_ops->exp_completed() / 2;
392 b_rcu_perf_writer_started = cur_ops->get_gp_seq();
398 udelay(writer_holdoff);
400 *wdp = ktime_get_mono_fast_ns();
404 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
405 if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
406 rcu_perf_writer_state = RTWS_ASYNC;
407 atomic_inc(this_cpu_ptr(&n_async_inflight));
408 cur_ops->async(rhp, rcu_perf_async_cb);
410 } else if (!kthread_should_stop()) {
411 rcu_perf_writer_state = RTWS_BARRIER;
412 cur_ops->gp_barrier();
415 kfree(rhp); /* Because we are stopping. */
418 rcu_perf_writer_state = RTWS_EXP_SYNC;
421 rcu_perf_writer_state = RTWS_SYNC;
424 rcu_perf_writer_state = RTWS_IDLE;
425 t = ktime_get_mono_fast_ns();
429 atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
431 if (!done && i >= MIN_MEAS) {
433 sp.sched_priority = 0;
434 sched_setscheduler_nocheck(current,
436 pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
437 perf_type, PERF_FLAG, me, MIN_MEAS);
438 if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
440 schedule_timeout_interruptible(10);
441 rcu_ftrace_dump(DUMP_ALL);
442 PERFOUT_STRING("Test complete");
443 t_rcu_perf_writer_finished = t;
445 b_rcu_perf_writer_finished =
446 cur_ops->exp_completed() / 2;
448 b_rcu_perf_writer_finished =
449 cur_ops->get_gp_seq();
452 smp_mb(); /* Assign before wake. */
453 wake_up(&shutdown_wq);
457 if (done && !alldone &&
458 atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
460 if (started && !alldone && i < MAX_MEAS - 1)
462 rcu_perf_wait_shutdown();
463 } while (!torture_must_stop());
465 rcu_perf_writer_state = RTWS_BARRIER;
466 cur_ops->gp_barrier();
468 rcu_perf_writer_state = RTWS_STOPPING;
469 writer_n_durations[me] = i_max;
470 torture_kthread_stopping("rcu_perf_writer");
475 rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
477 pr_alert("%s" PERF_FLAG
478 "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
479 perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
483 rcu_perf_cleanup(void)
492 * Would like warning at start, but everything is expedited
493 * during the mid-boot phase, so have to wait till the end.
495 if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
496 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
497 if (rcu_gp_is_normal() && gp_exp)
498 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
499 if (gp_exp && gp_async)
500 VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
502 if (torture_cleanup_begin())
506 for (i = 0; i < nrealreaders; i++)
507 torture_stop_kthread(rcu_perf_reader,
513 for (i = 0; i < nrealwriters; i++) {
514 torture_stop_kthread(rcu_perf_writer,
516 if (!writer_n_durations)
518 j = writer_n_durations[i];
519 pr_alert("%s%s writer %d gps: %d\n",
520 perf_type, PERF_FLAG, i, j);
523 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
524 perf_type, PERF_FLAG,
525 t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
526 t_rcu_perf_writer_finished -
527 t_rcu_perf_writer_started,
529 rcuperf_seq_diff(b_rcu_perf_writer_finished,
530 b_rcu_perf_writer_started));
531 for (i = 0; i < nrealwriters; i++) {
532 if (!writer_durations)
534 if (!writer_n_durations)
536 wdpp = writer_durations[i];
539 for (j = 0; j <= writer_n_durations[i]; j++) {
541 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
542 perf_type, PERF_FLAG,
545 schedule_timeout_uninterruptible(1);
547 kfree(writer_durations[i]);
550 kfree(writer_durations);
551 kfree(writer_n_durations);
554 /* Do torture-type-specific cleanup operations. */
555 if (cur_ops->cleanup != NULL)
558 torture_cleanup_end();
562 * Return the number if non-negative. If -1, the number of CPUs.
563 * If less than -1, that much less than the number of CPUs, but
566 static int compute_real(int n)
573 nr = num_online_cpus() + 1 + n;
581 * RCU perf shutdown kthread. Just waits to be awakened, then shuts
585 rcu_perf_shutdown(void *arg)
588 wait_event(shutdown_wq,
589 atomic_read(&n_rcu_perf_writer_finished) >=
591 } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters);
592 smp_mb(); /* Wake before output. */
603 static struct rcu_perf_ops *perf_ops[] = {
604 &rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops,
607 if (!torture_init_begin(perf_type, verbose))
610 /* Process args and tell the world that the perf'er is on the job. */
611 for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
612 cur_ops = perf_ops[i];
613 if (strcmp(perf_type, cur_ops->name) == 0)
616 if (i == ARRAY_SIZE(perf_ops)) {
617 pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type);
618 pr_alert("rcu-perf types:");
619 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
620 pr_cont(" %s", perf_ops[i]->name);
622 WARN_ON(!IS_MODULE(CONFIG_RCU_PERF_TEST));
629 nrealwriters = compute_real(nwriters);
630 nrealreaders = compute_real(nreaders);
631 atomic_set(&n_rcu_perf_reader_started, 0);
632 atomic_set(&n_rcu_perf_writer_started, 0);
633 atomic_set(&n_rcu_perf_writer_finished, 0);
634 rcu_perf_print_module_parms(cur_ops, "Start of test");
636 /* Start up the kthreads. */
639 init_waitqueue_head(&shutdown_wq);
640 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
644 schedule_timeout_uninterruptible(1);
646 reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
648 if (reader_tasks == NULL) {
649 VERBOSE_PERFOUT_ERRSTRING("out of memory");
653 for (i = 0; i < nrealreaders; i++) {
654 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
659 while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
660 schedule_timeout_uninterruptible(1);
661 writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
663 writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
666 kcalloc(nrealwriters, sizeof(*writer_n_durations),
668 if (!writer_tasks || !writer_durations || !writer_n_durations) {
669 VERBOSE_PERFOUT_ERRSTRING("out of memory");
673 for (i = 0; i < nrealwriters; i++) {
674 writer_durations[i] =
675 kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
677 if (!writer_durations[i]) {
681 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
695 module_init(rcu_perf_init);
696 module_exit(rcu_perf_cleanup);