Loading...
1/*
2 * Read-Copy Update module-based performance-test facility
3 *
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.
8 *
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.
13 *
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.
17 *
18 * Copyright (C) IBM Corporation, 2015
19 *
20 * Authors: Paul E. McKenney <paulmck@us.ibm.com>
21 */
22#include <linux/types.h>
23#include <linux/kernel.h>
24#include <linux/init.h>
25#include <linux/module.h>
26#include <linux/kthread.h>
27#include <linux/err.h>
28#include <linux/spinlock.h>
29#include <linux/smp.h>
30#include <linux/rcupdate.h>
31#include <linux/interrupt.h>
32#include <linux/sched.h>
33#include <linux/atomic.h>
34#include <linux/bitops.h>
35#include <linux/completion.h>
36#include <linux/moduleparam.h>
37#include <linux/percpu.h>
38#include <linux/notifier.h>
39#include <linux/reboot.h>
40#include <linux/freezer.h>
41#include <linux/cpu.h>
42#include <linux/delay.h>
43#include <linux/stat.h>
44#include <linux/srcu.h>
45#include <linux/slab.h>
46#include <asm/byteorder.h>
47#include <linux/torture.h>
48#include <linux/vmalloc.h>
49
50MODULE_LICENSE("GPL");
51MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.vnet.ibm.com>");
52
53#define PERF_FLAG "-perf:"
54#define PERFOUT_STRING(s) \
55 pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
56#define VERBOSE_PERFOUT_STRING(s) \
57 do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
58#define VERBOSE_PERFOUT_ERRSTRING(s) \
59 do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
60
61torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
62torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
63torture_param(int, nreaders, -1, "Number of RCU reader threads");
64torture_param(int, nwriters, -1, "Number of RCU updater threads");
65torture_param(bool, shutdown, false, "Shutdown at end of performance tests.");
66torture_param(bool, verbose, true, "Enable verbose debugging printk()s");
67
68static char *perf_type = "rcu";
69module_param(perf_type, charp, 0444);
70MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, rcu_bh, ...)");
71
72static int nrealreaders;
73static int nrealwriters;
74static struct task_struct **writer_tasks;
75static struct task_struct **reader_tasks;
76static struct task_struct *shutdown_task;
77
78static u64 **writer_durations;
79static int *writer_n_durations;
80static atomic_t n_rcu_perf_reader_started;
81static atomic_t n_rcu_perf_writer_started;
82static atomic_t n_rcu_perf_writer_finished;
83static wait_queue_head_t shutdown_wq;
84static u64 t_rcu_perf_writer_started;
85static u64 t_rcu_perf_writer_finished;
86static unsigned long b_rcu_perf_writer_started;
87static unsigned long b_rcu_perf_writer_finished;
88
89static int rcu_perf_writer_state;
90#define RTWS_INIT 0
91#define RTWS_EXP_SYNC 1
92#define RTWS_SYNC 2
93#define RTWS_IDLE 2
94#define RTWS_STOPPING 3
95
96#define MAX_MEAS 10000
97#define MIN_MEAS 100
98
99static int perf_runnable = IS_ENABLED(MODULE);
100module_param(perf_runnable, int, 0444);
101MODULE_PARM_DESC(perf_runnable, "Start rcuperf at boot");
102
103/*
104 * Operations vector for selecting different types of tests.
105 */
106
107struct rcu_perf_ops {
108 int ptype;
109 void (*init)(void);
110 void (*cleanup)(void);
111 int (*readlock)(void);
112 void (*readunlock)(int idx);
113 unsigned long (*started)(void);
114 unsigned long (*completed)(void);
115 unsigned long (*exp_completed)(void);
116 void (*sync)(void);
117 void (*exp_sync)(void);
118 const char *name;
119};
120
121static struct rcu_perf_ops *cur_ops;
122
123/*
124 * Definitions for rcu perf testing.
125 */
126
127static int rcu_perf_read_lock(void) __acquires(RCU)
128{
129 rcu_read_lock();
130 return 0;
131}
132
133static void rcu_perf_read_unlock(int idx) __releases(RCU)
134{
135 rcu_read_unlock();
136}
137
138static unsigned long __maybe_unused rcu_no_completed(void)
139{
140 return 0;
141}
142
143static void rcu_sync_perf_init(void)
144{
145}
146
147static struct rcu_perf_ops rcu_ops = {
148 .ptype = RCU_FLAVOR,
149 .init = rcu_sync_perf_init,
150 .readlock = rcu_perf_read_lock,
151 .readunlock = rcu_perf_read_unlock,
152 .started = rcu_batches_started,
153 .completed = rcu_batches_completed,
154 .exp_completed = rcu_exp_batches_completed,
155 .sync = synchronize_rcu,
156 .exp_sync = synchronize_rcu_expedited,
157 .name = "rcu"
158};
159
160/*
161 * Definitions for rcu_bh perf testing.
162 */
163
164static int rcu_bh_perf_read_lock(void) __acquires(RCU_BH)
165{
166 rcu_read_lock_bh();
167 return 0;
168}
169
170static void rcu_bh_perf_read_unlock(int idx) __releases(RCU_BH)
171{
172 rcu_read_unlock_bh();
173}
174
175static struct rcu_perf_ops rcu_bh_ops = {
176 .ptype = RCU_BH_FLAVOR,
177 .init = rcu_sync_perf_init,
178 .readlock = rcu_bh_perf_read_lock,
179 .readunlock = rcu_bh_perf_read_unlock,
180 .started = rcu_batches_started_bh,
181 .completed = rcu_batches_completed_bh,
182 .exp_completed = rcu_exp_batches_completed_sched,
183 .sync = synchronize_rcu_bh,
184 .exp_sync = synchronize_rcu_bh_expedited,
185 .name = "rcu_bh"
186};
187
188/*
189 * Definitions for srcu perf testing.
190 */
191
192DEFINE_STATIC_SRCU(srcu_ctl_perf);
193static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
194
195static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
196{
197 return srcu_read_lock(srcu_ctlp);
198}
199
200static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
201{
202 srcu_read_unlock(srcu_ctlp, idx);
203}
204
205static unsigned long srcu_perf_completed(void)
206{
207 return srcu_batches_completed(srcu_ctlp);
208}
209
210static void srcu_perf_synchronize(void)
211{
212 synchronize_srcu(srcu_ctlp);
213}
214
215static void srcu_perf_synchronize_expedited(void)
216{
217 synchronize_srcu_expedited(srcu_ctlp);
218}
219
220static struct rcu_perf_ops srcu_ops = {
221 .ptype = SRCU_FLAVOR,
222 .init = rcu_sync_perf_init,
223 .readlock = srcu_perf_read_lock,
224 .readunlock = srcu_perf_read_unlock,
225 .started = NULL,
226 .completed = srcu_perf_completed,
227 .exp_completed = srcu_perf_completed,
228 .sync = srcu_perf_synchronize,
229 .exp_sync = srcu_perf_synchronize_expedited,
230 .name = "srcu"
231};
232
233/*
234 * Definitions for sched perf testing.
235 */
236
237static int sched_perf_read_lock(void)
238{
239 preempt_disable();
240 return 0;
241}
242
243static void sched_perf_read_unlock(int idx)
244{
245 preempt_enable();
246}
247
248static struct rcu_perf_ops sched_ops = {
249 .ptype = RCU_SCHED_FLAVOR,
250 .init = rcu_sync_perf_init,
251 .readlock = sched_perf_read_lock,
252 .readunlock = sched_perf_read_unlock,
253 .started = rcu_batches_started_sched,
254 .completed = rcu_batches_completed_sched,
255 .exp_completed = rcu_exp_batches_completed_sched,
256 .sync = synchronize_sched,
257 .exp_sync = synchronize_sched_expedited,
258 .name = "sched"
259};
260
261#ifdef CONFIG_TASKS_RCU
262
263/*
264 * Definitions for RCU-tasks perf testing.
265 */
266
267static int tasks_perf_read_lock(void)
268{
269 return 0;
270}
271
272static void tasks_perf_read_unlock(int idx)
273{
274}
275
276static struct rcu_perf_ops tasks_ops = {
277 .ptype = RCU_TASKS_FLAVOR,
278 .init = rcu_sync_perf_init,
279 .readlock = tasks_perf_read_lock,
280 .readunlock = tasks_perf_read_unlock,
281 .started = rcu_no_completed,
282 .completed = rcu_no_completed,
283 .sync = synchronize_rcu_tasks,
284 .exp_sync = synchronize_rcu_tasks,
285 .name = "tasks"
286};
287
288#define RCUPERF_TASKS_OPS &tasks_ops,
289
290static bool __maybe_unused torturing_tasks(void)
291{
292 return cur_ops == &tasks_ops;
293}
294
295#else /* #ifdef CONFIG_TASKS_RCU */
296
297#define RCUPERF_TASKS_OPS
298
299static bool __maybe_unused torturing_tasks(void)
300{
301 return false;
302}
303
304#endif /* #else #ifdef CONFIG_TASKS_RCU */
305
306/*
307 * If performance tests complete, wait for shutdown to commence.
308 */
309static void rcu_perf_wait_shutdown(void)
310{
311 cond_resched_rcu_qs();
312 if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
313 return;
314 while (!torture_must_stop())
315 schedule_timeout_uninterruptible(1);
316}
317
318/*
319 * RCU perf reader kthread. Repeatedly does empty RCU read-side
320 * critical section, minimizing update-side interference.
321 */
322static int
323rcu_perf_reader(void *arg)
324{
325 unsigned long flags;
326 int idx;
327 long me = (long)arg;
328
329 VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
330 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
331 set_user_nice(current, MAX_NICE);
332 atomic_inc(&n_rcu_perf_reader_started);
333
334 do {
335 local_irq_save(flags);
336 idx = cur_ops->readlock();
337 cur_ops->readunlock(idx);
338 local_irq_restore(flags);
339 rcu_perf_wait_shutdown();
340 } while (!torture_must_stop());
341 torture_kthread_stopping("rcu_perf_reader");
342 return 0;
343}
344
345/*
346 * RCU perf writer kthread. Repeatedly does a grace period.
347 */
348static int
349rcu_perf_writer(void *arg)
350{
351 int i = 0;
352 int i_max;
353 long me = (long)arg;
354 struct sched_param sp;
355 bool started = false, done = false, alldone = false;
356 u64 t;
357 u64 *wdp;
358 u64 *wdpp = writer_durations[me];
359
360 VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
361 WARN_ON(!wdpp);
362 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
363 sp.sched_priority = 1;
364 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
365
366 if (holdoff)
367 schedule_timeout_uninterruptible(holdoff * HZ);
368
369 t = ktime_get_mono_fast_ns();
370 if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
371 t_rcu_perf_writer_started = t;
372 if (gp_exp) {
373 b_rcu_perf_writer_started =
374 cur_ops->exp_completed() / 2;
375 } else {
376 b_rcu_perf_writer_started =
377 cur_ops->completed();
378 }
379 }
380
381 do {
382 wdp = &wdpp[i];
383 *wdp = ktime_get_mono_fast_ns();
384 if (gp_exp) {
385 rcu_perf_writer_state = RTWS_EXP_SYNC;
386 cur_ops->exp_sync();
387 } else {
388 rcu_perf_writer_state = RTWS_SYNC;
389 cur_ops->sync();
390 }
391 rcu_perf_writer_state = RTWS_IDLE;
392 t = ktime_get_mono_fast_ns();
393 *wdp = t - *wdp;
394 i_max = i;
395 if (!started &&
396 atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
397 started = true;
398 if (!done && i >= MIN_MEAS) {
399 done = true;
400 sp.sched_priority = 0;
401 sched_setscheduler_nocheck(current,
402 SCHED_NORMAL, &sp);
403 pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
404 perf_type, PERF_FLAG, me, MIN_MEAS);
405 if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
406 nrealwriters) {
407 schedule_timeout_interruptible(10);
408 rcu_ftrace_dump(DUMP_ALL);
409 PERFOUT_STRING("Test complete");
410 t_rcu_perf_writer_finished = t;
411 if (gp_exp) {
412 b_rcu_perf_writer_finished =
413 cur_ops->exp_completed() / 2;
414 } else {
415 b_rcu_perf_writer_finished =
416 cur_ops->completed();
417 }
418 if (shutdown) {
419 smp_mb(); /* Assign before wake. */
420 wake_up(&shutdown_wq);
421 }
422 }
423 }
424 if (done && !alldone &&
425 atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
426 alldone = true;
427 if (started && !alldone && i < MAX_MEAS - 1)
428 i++;
429 rcu_perf_wait_shutdown();
430 } while (!torture_must_stop());
431 rcu_perf_writer_state = RTWS_STOPPING;
432 writer_n_durations[me] = i_max;
433 torture_kthread_stopping("rcu_perf_writer");
434 return 0;
435}
436
437static inline void
438rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
439{
440 pr_alert("%s" PERF_FLAG
441 "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
442 perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
443}
444
445static void
446rcu_perf_cleanup(void)
447{
448 int i;
449 int j;
450 int ngps = 0;
451 u64 *wdp;
452 u64 *wdpp;
453
454 if (torture_cleanup_begin())
455 return;
456
457 if (reader_tasks) {
458 for (i = 0; i < nrealreaders; i++)
459 torture_stop_kthread(rcu_perf_reader,
460 reader_tasks[i]);
461 kfree(reader_tasks);
462 }
463
464 if (writer_tasks) {
465 for (i = 0; i < nrealwriters; i++) {
466 torture_stop_kthread(rcu_perf_writer,
467 writer_tasks[i]);
468 if (!writer_n_durations)
469 continue;
470 j = writer_n_durations[i];
471 pr_alert("%s%s writer %d gps: %d\n",
472 perf_type, PERF_FLAG, i, j);
473 ngps += j;
474 }
475 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
476 perf_type, PERF_FLAG,
477 t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
478 t_rcu_perf_writer_finished -
479 t_rcu_perf_writer_started,
480 ngps,
481 b_rcu_perf_writer_finished -
482 b_rcu_perf_writer_started);
483 for (i = 0; i < nrealwriters; i++) {
484 if (!writer_durations)
485 break;
486 if (!writer_n_durations)
487 continue;
488 wdpp = writer_durations[i];
489 if (!wdpp)
490 continue;
491 for (j = 0; j <= writer_n_durations[i]; j++) {
492 wdp = &wdpp[j];
493 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
494 perf_type, PERF_FLAG,
495 i, j, *wdp);
496 if (j % 100 == 0)
497 schedule_timeout_uninterruptible(1);
498 }
499 kfree(writer_durations[i]);
500 }
501 kfree(writer_tasks);
502 kfree(writer_durations);
503 kfree(writer_n_durations);
504 }
505
506 /* Do flavor-specific cleanup operations. */
507 if (cur_ops->cleanup != NULL)
508 cur_ops->cleanup();
509
510 torture_cleanup_end();
511}
512
513/*
514 * Return the number if non-negative. If -1, the number of CPUs.
515 * If less than -1, that much less than the number of CPUs, but
516 * at least one.
517 */
518static int compute_real(int n)
519{
520 int nr;
521
522 if (n >= 0) {
523 nr = n;
524 } else {
525 nr = num_online_cpus() + 1 + n;
526 if (nr <= 0)
527 nr = 1;
528 }
529 return nr;
530}
531
532/*
533 * RCU perf shutdown kthread. Just waits to be awakened, then shuts
534 * down system.
535 */
536static int
537rcu_perf_shutdown(void *arg)
538{
539 do {
540 wait_event(shutdown_wq,
541 atomic_read(&n_rcu_perf_writer_finished) >=
542 nrealwriters);
543 } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters);
544 smp_mb(); /* Wake before output. */
545 rcu_perf_cleanup();
546 kernel_power_off();
547 return -EINVAL;
548}
549
550static int __init
551rcu_perf_init(void)
552{
553 long i;
554 int firsterr = 0;
555 static struct rcu_perf_ops *perf_ops[] = {
556 &rcu_ops, &rcu_bh_ops, &srcu_ops, &sched_ops,
557 RCUPERF_TASKS_OPS
558 };
559
560 if (!torture_init_begin(perf_type, verbose, &perf_runnable))
561 return -EBUSY;
562
563 /* Process args and tell the world that the perf'er is on the job. */
564 for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
565 cur_ops = perf_ops[i];
566 if (strcmp(perf_type, cur_ops->name) == 0)
567 break;
568 }
569 if (i == ARRAY_SIZE(perf_ops)) {
570 pr_alert("rcu-perf: invalid perf type: \"%s\"\n",
571 perf_type);
572 pr_alert("rcu-perf types:");
573 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
574 pr_alert(" %s", perf_ops[i]->name);
575 pr_alert("\n");
576 firsterr = -EINVAL;
577 goto unwind;
578 }
579 if (cur_ops->init)
580 cur_ops->init();
581
582 nrealwriters = compute_real(nwriters);
583 nrealreaders = compute_real(nreaders);
584 atomic_set(&n_rcu_perf_reader_started, 0);
585 atomic_set(&n_rcu_perf_writer_started, 0);
586 atomic_set(&n_rcu_perf_writer_finished, 0);
587 rcu_perf_print_module_parms(cur_ops, "Start of test");
588
589 /* Start up the kthreads. */
590
591 if (shutdown) {
592 init_waitqueue_head(&shutdown_wq);
593 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
594 shutdown_task);
595 if (firsterr)
596 goto unwind;
597 schedule_timeout_uninterruptible(1);
598 }
599 reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
600 GFP_KERNEL);
601 if (reader_tasks == NULL) {
602 VERBOSE_PERFOUT_ERRSTRING("out of memory");
603 firsterr = -ENOMEM;
604 goto unwind;
605 }
606 for (i = 0; i < nrealreaders; i++) {
607 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
608 reader_tasks[i]);
609 if (firsterr)
610 goto unwind;
611 }
612 while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
613 schedule_timeout_uninterruptible(1);
614 writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
615 GFP_KERNEL);
616 writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
617 GFP_KERNEL);
618 writer_n_durations =
619 kcalloc(nrealwriters, sizeof(*writer_n_durations),
620 GFP_KERNEL);
621 if (!writer_tasks || !writer_durations || !writer_n_durations) {
622 VERBOSE_PERFOUT_ERRSTRING("out of memory");
623 firsterr = -ENOMEM;
624 goto unwind;
625 }
626 if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp) {
627 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
628 firsterr = -EINVAL;
629 goto unwind;
630 }
631 if (rcu_gp_is_normal() && gp_exp) {
632 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
633 firsterr = -EINVAL;
634 goto unwind;
635 }
636 for (i = 0; i < nrealwriters; i++) {
637 writer_durations[i] =
638 kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
639 GFP_KERNEL);
640 if (!writer_durations[i]) {
641 firsterr = -ENOMEM;
642 goto unwind;
643 }
644 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
645 writer_tasks[i]);
646 if (firsterr)
647 goto unwind;
648 }
649 torture_init_end();
650 return 0;
651
652unwind:
653 torture_init_end();
654 rcu_perf_cleanup();
655 return firsterr;
656}
657
658module_init(rcu_perf_init);
659module_exit(rcu_perf_cleanup);
1// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update module-based performance-test facility
4 *
5 * Copyright (C) IBM Corporation, 2015
6 *
7 * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
8 */
9
10#define pr_fmt(fmt) fmt
11
12#include <linux/types.h>
13#include <linux/kernel.h>
14#include <linux/init.h>
15#include <linux/module.h>
16#include <linux/kthread.h>
17#include <linux/err.h>
18#include <linux/spinlock.h>
19#include <linux/smp.h>
20#include <linux/rcupdate.h>
21#include <linux/interrupt.h>
22#include <linux/sched.h>
23#include <uapi/linux/sched/types.h>
24#include <linux/atomic.h>
25#include <linux/bitops.h>
26#include <linux/completion.h>
27#include <linux/moduleparam.h>
28#include <linux/percpu.h>
29#include <linux/notifier.h>
30#include <linux/reboot.h>
31#include <linux/freezer.h>
32#include <linux/cpu.h>
33#include <linux/delay.h>
34#include <linux/stat.h>
35#include <linux/srcu.h>
36#include <linux/slab.h>
37#include <asm/byteorder.h>
38#include <linux/torture.h>
39#include <linux/vmalloc.h>
40
41#include "rcu.h"
42
43MODULE_LICENSE("GPL");
44MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
45
46#define PERF_FLAG "-perf:"
47#define PERFOUT_STRING(s) \
48 pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
49#define VERBOSE_PERFOUT_STRING(s) \
50 do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
51#define VERBOSE_PERFOUT_ERRSTRING(s) \
52 do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
53
54/*
55 * The intended use cases for the nreaders and nwriters module parameters
56 * are as follows:
57 *
58 * 1. Specify only the nr_cpus kernel boot parameter. This will
59 * set both nreaders and nwriters to the value specified by
60 * nr_cpus for a mixed reader/writer test.
61 *
62 * 2. Specify the nr_cpus kernel boot parameter, but set
63 * rcuperf.nreaders to zero. This will set nwriters to the
64 * value specified by nr_cpus for an update-only test.
65 *
66 * 3. Specify the nr_cpus kernel boot parameter, but set
67 * rcuperf.nwriters to zero. This will set nreaders to the
68 * value specified by nr_cpus for a read-only test.
69 *
70 * Various other use cases may of course be specified.
71 */
72
73#ifdef MODULE
74# define RCUPERF_SHUTDOWN 0
75#else
76# define RCUPERF_SHUTDOWN 1
77#endif
78
79torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
80torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
81torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
82torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
83torture_param(int, nreaders, -1, "Number of RCU reader threads");
84torture_param(int, nwriters, -1, "Number of RCU updater threads");
85torture_param(bool, shutdown, RCUPERF_SHUTDOWN,
86 "Shutdown at end of performance tests.");
87torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
88torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
89
90static char *perf_type = "rcu";
91module_param(perf_type, charp, 0444);
92MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, srcu, ...)");
93
94static int nrealreaders;
95static int nrealwriters;
96static struct task_struct **writer_tasks;
97static struct task_struct **reader_tasks;
98static struct task_struct *shutdown_task;
99
100static u64 **writer_durations;
101static int *writer_n_durations;
102static atomic_t n_rcu_perf_reader_started;
103static atomic_t n_rcu_perf_writer_started;
104static atomic_t n_rcu_perf_writer_finished;
105static wait_queue_head_t shutdown_wq;
106static u64 t_rcu_perf_writer_started;
107static u64 t_rcu_perf_writer_finished;
108static unsigned long b_rcu_perf_writer_started;
109static unsigned long b_rcu_perf_writer_finished;
110static DEFINE_PER_CPU(atomic_t, n_async_inflight);
111
112static int rcu_perf_writer_state;
113#define RTWS_INIT 0
114#define RTWS_ASYNC 1
115#define RTWS_BARRIER 2
116#define RTWS_EXP_SYNC 3
117#define RTWS_SYNC 4
118#define RTWS_IDLE 5
119#define RTWS_STOPPING 6
120
121#define MAX_MEAS 10000
122#define MIN_MEAS 100
123
124/*
125 * Operations vector for selecting different types of tests.
126 */
127
128struct rcu_perf_ops {
129 int ptype;
130 void (*init)(void);
131 void (*cleanup)(void);
132 int (*readlock)(void);
133 void (*readunlock)(int idx);
134 unsigned long (*get_gp_seq)(void);
135 unsigned long (*gp_diff)(unsigned long new, unsigned long old);
136 unsigned long (*exp_completed)(void);
137 void (*async)(struct rcu_head *head, rcu_callback_t func);
138 void (*gp_barrier)(void);
139 void (*sync)(void);
140 void (*exp_sync)(void);
141 const char *name;
142};
143
144static struct rcu_perf_ops *cur_ops;
145
146/*
147 * Definitions for rcu perf testing.
148 */
149
150static int rcu_perf_read_lock(void) __acquires(RCU)
151{
152 rcu_read_lock();
153 return 0;
154}
155
156static void rcu_perf_read_unlock(int idx) __releases(RCU)
157{
158 rcu_read_unlock();
159}
160
161static unsigned long __maybe_unused rcu_no_completed(void)
162{
163 return 0;
164}
165
166static void rcu_sync_perf_init(void)
167{
168}
169
170static struct rcu_perf_ops rcu_ops = {
171 .ptype = RCU_FLAVOR,
172 .init = rcu_sync_perf_init,
173 .readlock = rcu_perf_read_lock,
174 .readunlock = rcu_perf_read_unlock,
175 .get_gp_seq = rcu_get_gp_seq,
176 .gp_diff = rcu_seq_diff,
177 .exp_completed = rcu_exp_batches_completed,
178 .async = call_rcu,
179 .gp_barrier = rcu_barrier,
180 .sync = synchronize_rcu,
181 .exp_sync = synchronize_rcu_expedited,
182 .name = "rcu"
183};
184
185/*
186 * Definitions for srcu perf testing.
187 */
188
189DEFINE_STATIC_SRCU(srcu_ctl_perf);
190static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
191
192static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
193{
194 return srcu_read_lock(srcu_ctlp);
195}
196
197static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
198{
199 srcu_read_unlock(srcu_ctlp, idx);
200}
201
202static unsigned long srcu_perf_completed(void)
203{
204 return srcu_batches_completed(srcu_ctlp);
205}
206
207static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
208{
209 call_srcu(srcu_ctlp, head, func);
210}
211
212static void srcu_rcu_barrier(void)
213{
214 srcu_barrier(srcu_ctlp);
215}
216
217static void srcu_perf_synchronize(void)
218{
219 synchronize_srcu(srcu_ctlp);
220}
221
222static void srcu_perf_synchronize_expedited(void)
223{
224 synchronize_srcu_expedited(srcu_ctlp);
225}
226
227static struct rcu_perf_ops srcu_ops = {
228 .ptype = SRCU_FLAVOR,
229 .init = rcu_sync_perf_init,
230 .readlock = srcu_perf_read_lock,
231 .readunlock = srcu_perf_read_unlock,
232 .get_gp_seq = srcu_perf_completed,
233 .gp_diff = rcu_seq_diff,
234 .exp_completed = srcu_perf_completed,
235 .async = srcu_call_rcu,
236 .gp_barrier = srcu_rcu_barrier,
237 .sync = srcu_perf_synchronize,
238 .exp_sync = srcu_perf_synchronize_expedited,
239 .name = "srcu"
240};
241
242static struct srcu_struct srcud;
243
244static void srcu_sync_perf_init(void)
245{
246 srcu_ctlp = &srcud;
247 init_srcu_struct(srcu_ctlp);
248}
249
250static void srcu_sync_perf_cleanup(void)
251{
252 cleanup_srcu_struct(srcu_ctlp);
253}
254
255static struct rcu_perf_ops srcud_ops = {
256 .ptype = SRCU_FLAVOR,
257 .init = srcu_sync_perf_init,
258 .cleanup = srcu_sync_perf_cleanup,
259 .readlock = srcu_perf_read_lock,
260 .readunlock = srcu_perf_read_unlock,
261 .get_gp_seq = srcu_perf_completed,
262 .gp_diff = rcu_seq_diff,
263 .exp_completed = srcu_perf_completed,
264 .async = srcu_call_rcu,
265 .gp_barrier = srcu_rcu_barrier,
266 .sync = srcu_perf_synchronize,
267 .exp_sync = srcu_perf_synchronize_expedited,
268 .name = "srcud"
269};
270
271/*
272 * Definitions for RCU-tasks perf testing.
273 */
274
275static int tasks_perf_read_lock(void)
276{
277 return 0;
278}
279
280static void tasks_perf_read_unlock(int idx)
281{
282}
283
284static struct rcu_perf_ops tasks_ops = {
285 .ptype = RCU_TASKS_FLAVOR,
286 .init = rcu_sync_perf_init,
287 .readlock = tasks_perf_read_lock,
288 .readunlock = tasks_perf_read_unlock,
289 .get_gp_seq = rcu_no_completed,
290 .gp_diff = rcu_seq_diff,
291 .async = call_rcu_tasks,
292 .gp_barrier = rcu_barrier_tasks,
293 .sync = synchronize_rcu_tasks,
294 .exp_sync = synchronize_rcu_tasks,
295 .name = "tasks"
296};
297
298static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old)
299{
300 if (!cur_ops->gp_diff)
301 return new - old;
302 return cur_ops->gp_diff(new, old);
303}
304
305/*
306 * If performance tests complete, wait for shutdown to commence.
307 */
308static void rcu_perf_wait_shutdown(void)
309{
310 cond_resched_tasks_rcu_qs();
311 if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
312 return;
313 while (!torture_must_stop())
314 schedule_timeout_uninterruptible(1);
315}
316
317/*
318 * RCU perf reader kthread. Repeatedly does empty RCU read-side
319 * critical section, minimizing update-side interference.
320 */
321static int
322rcu_perf_reader(void *arg)
323{
324 unsigned long flags;
325 int idx;
326 long me = (long)arg;
327
328 VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
329 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
330 set_user_nice(current, MAX_NICE);
331 atomic_inc(&n_rcu_perf_reader_started);
332
333 do {
334 local_irq_save(flags);
335 idx = cur_ops->readlock();
336 cur_ops->readunlock(idx);
337 local_irq_restore(flags);
338 rcu_perf_wait_shutdown();
339 } while (!torture_must_stop());
340 torture_kthread_stopping("rcu_perf_reader");
341 return 0;
342}
343
344/*
345 * Callback function for asynchronous grace periods from rcu_perf_writer().
346 */
347static void rcu_perf_async_cb(struct rcu_head *rhp)
348{
349 atomic_dec(this_cpu_ptr(&n_async_inflight));
350 kfree(rhp);
351}
352
353/*
354 * RCU perf writer kthread. Repeatedly does a grace period.
355 */
356static int
357rcu_perf_writer(void *arg)
358{
359 int i = 0;
360 int i_max;
361 long me = (long)arg;
362 struct rcu_head *rhp = NULL;
363 struct sched_param sp;
364 bool started = false, done = false, alldone = false;
365 u64 t;
366 u64 *wdp;
367 u64 *wdpp = writer_durations[me];
368
369 VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
370 WARN_ON(!wdpp);
371 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
372 sp.sched_priority = 1;
373 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
374
375 if (holdoff)
376 schedule_timeout_uninterruptible(holdoff * HZ);
377
378 /*
379 * Wait until rcu_end_inkernel_boot() is called for normal GP tests
380 * so that RCU is not always expedited for normal GP tests.
381 * The system_state test is approximate, but works well in practice.
382 */
383 while (!gp_exp && system_state != SYSTEM_RUNNING)
384 schedule_timeout_uninterruptible(1);
385
386 t = ktime_get_mono_fast_ns();
387 if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
388 t_rcu_perf_writer_started = t;
389 if (gp_exp) {
390 b_rcu_perf_writer_started =
391 cur_ops->exp_completed() / 2;
392 } else {
393 b_rcu_perf_writer_started = cur_ops->get_gp_seq();
394 }
395 }
396
397 do {
398 if (writer_holdoff)
399 udelay(writer_holdoff);
400 wdp = &wdpp[i];
401 *wdp = ktime_get_mono_fast_ns();
402 if (gp_async) {
403retry:
404 if (!rhp)
405 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
406 if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
407 rcu_perf_writer_state = RTWS_ASYNC;
408 atomic_inc(this_cpu_ptr(&n_async_inflight));
409 cur_ops->async(rhp, rcu_perf_async_cb);
410 rhp = NULL;
411 } else if (!kthread_should_stop()) {
412 rcu_perf_writer_state = RTWS_BARRIER;
413 cur_ops->gp_barrier();
414 goto retry;
415 } else {
416 kfree(rhp); /* Because we are stopping. */
417 }
418 } else if (gp_exp) {
419 rcu_perf_writer_state = RTWS_EXP_SYNC;
420 cur_ops->exp_sync();
421 } else {
422 rcu_perf_writer_state = RTWS_SYNC;
423 cur_ops->sync();
424 }
425 rcu_perf_writer_state = RTWS_IDLE;
426 t = ktime_get_mono_fast_ns();
427 *wdp = t - *wdp;
428 i_max = i;
429 if (!started &&
430 atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
431 started = true;
432 if (!done && i >= MIN_MEAS) {
433 done = true;
434 sp.sched_priority = 0;
435 sched_setscheduler_nocheck(current,
436 SCHED_NORMAL, &sp);
437 pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
438 perf_type, PERF_FLAG, me, MIN_MEAS);
439 if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
440 nrealwriters) {
441 schedule_timeout_interruptible(10);
442 rcu_ftrace_dump(DUMP_ALL);
443 PERFOUT_STRING("Test complete");
444 t_rcu_perf_writer_finished = t;
445 if (gp_exp) {
446 b_rcu_perf_writer_finished =
447 cur_ops->exp_completed() / 2;
448 } else {
449 b_rcu_perf_writer_finished =
450 cur_ops->get_gp_seq();
451 }
452 if (shutdown) {
453 smp_mb(); /* Assign before wake. */
454 wake_up(&shutdown_wq);
455 }
456 }
457 }
458 if (done && !alldone &&
459 atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
460 alldone = true;
461 if (started && !alldone && i < MAX_MEAS - 1)
462 i++;
463 rcu_perf_wait_shutdown();
464 } while (!torture_must_stop());
465 if (gp_async) {
466 rcu_perf_writer_state = RTWS_BARRIER;
467 cur_ops->gp_barrier();
468 }
469 rcu_perf_writer_state = RTWS_STOPPING;
470 writer_n_durations[me] = i_max;
471 torture_kthread_stopping("rcu_perf_writer");
472 return 0;
473}
474
475static void
476rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
477{
478 pr_alert("%s" PERF_FLAG
479 "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
480 perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
481}
482
483static void
484rcu_perf_cleanup(void)
485{
486 int i;
487 int j;
488 int ngps = 0;
489 u64 *wdp;
490 u64 *wdpp;
491
492 /*
493 * Would like warning at start, but everything is expedited
494 * during the mid-boot phase, so have to wait till the end.
495 */
496 if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
497 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
498 if (rcu_gp_is_normal() && gp_exp)
499 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
500 if (gp_exp && gp_async)
501 VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
502
503 if (torture_cleanup_begin())
504 return;
505 if (!cur_ops) {
506 torture_cleanup_end();
507 return;
508 }
509
510 if (reader_tasks) {
511 for (i = 0; i < nrealreaders; i++)
512 torture_stop_kthread(rcu_perf_reader,
513 reader_tasks[i]);
514 kfree(reader_tasks);
515 }
516
517 if (writer_tasks) {
518 for (i = 0; i < nrealwriters; i++) {
519 torture_stop_kthread(rcu_perf_writer,
520 writer_tasks[i]);
521 if (!writer_n_durations)
522 continue;
523 j = writer_n_durations[i];
524 pr_alert("%s%s writer %d gps: %d\n",
525 perf_type, PERF_FLAG, i, j);
526 ngps += j;
527 }
528 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
529 perf_type, PERF_FLAG,
530 t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
531 t_rcu_perf_writer_finished -
532 t_rcu_perf_writer_started,
533 ngps,
534 rcuperf_seq_diff(b_rcu_perf_writer_finished,
535 b_rcu_perf_writer_started));
536 for (i = 0; i < nrealwriters; i++) {
537 if (!writer_durations)
538 break;
539 if (!writer_n_durations)
540 continue;
541 wdpp = writer_durations[i];
542 if (!wdpp)
543 continue;
544 for (j = 0; j <= writer_n_durations[i]; j++) {
545 wdp = &wdpp[j];
546 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
547 perf_type, PERF_FLAG,
548 i, j, *wdp);
549 if (j % 100 == 0)
550 schedule_timeout_uninterruptible(1);
551 }
552 kfree(writer_durations[i]);
553 }
554 kfree(writer_tasks);
555 kfree(writer_durations);
556 kfree(writer_n_durations);
557 }
558
559 /* Do torture-type-specific cleanup operations. */
560 if (cur_ops->cleanup != NULL)
561 cur_ops->cleanup();
562
563 torture_cleanup_end();
564}
565
566/*
567 * Return the number if non-negative. If -1, the number of CPUs.
568 * If less than -1, that much less than the number of CPUs, but
569 * at least one.
570 */
571static int compute_real(int n)
572{
573 int nr;
574
575 if (n >= 0) {
576 nr = n;
577 } else {
578 nr = num_online_cpus() + 1 + n;
579 if (nr <= 0)
580 nr = 1;
581 }
582 return nr;
583}
584
585/*
586 * RCU perf shutdown kthread. Just waits to be awakened, then shuts
587 * down system.
588 */
589static int
590rcu_perf_shutdown(void *arg)
591{
592 do {
593 wait_event(shutdown_wq,
594 atomic_read(&n_rcu_perf_writer_finished) >=
595 nrealwriters);
596 } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters);
597 smp_mb(); /* Wake before output. */
598 rcu_perf_cleanup();
599 kernel_power_off();
600 return -EINVAL;
601}
602
603static int __init
604rcu_perf_init(void)
605{
606 long i;
607 int firsterr = 0;
608 static struct rcu_perf_ops *perf_ops[] = {
609 &rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops,
610 };
611
612 if (!torture_init_begin(perf_type, verbose))
613 return -EBUSY;
614
615 /* Process args and tell the world that the perf'er is on the job. */
616 for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
617 cur_ops = perf_ops[i];
618 if (strcmp(perf_type, cur_ops->name) == 0)
619 break;
620 }
621 if (i == ARRAY_SIZE(perf_ops)) {
622 pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type);
623 pr_alert("rcu-perf types:");
624 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
625 pr_cont(" %s", perf_ops[i]->name);
626 pr_cont("\n");
627 WARN_ON(!IS_MODULE(CONFIG_RCU_PERF_TEST));
628 firsterr = -EINVAL;
629 cur_ops = NULL;
630 goto unwind;
631 }
632 if (cur_ops->init)
633 cur_ops->init();
634
635 nrealwriters = compute_real(nwriters);
636 nrealreaders = compute_real(nreaders);
637 atomic_set(&n_rcu_perf_reader_started, 0);
638 atomic_set(&n_rcu_perf_writer_started, 0);
639 atomic_set(&n_rcu_perf_writer_finished, 0);
640 rcu_perf_print_module_parms(cur_ops, "Start of test");
641
642 /* Start up the kthreads. */
643
644 if (shutdown) {
645 init_waitqueue_head(&shutdown_wq);
646 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
647 shutdown_task);
648 if (firsterr)
649 goto unwind;
650 schedule_timeout_uninterruptible(1);
651 }
652 reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
653 GFP_KERNEL);
654 if (reader_tasks == NULL) {
655 VERBOSE_PERFOUT_ERRSTRING("out of memory");
656 firsterr = -ENOMEM;
657 goto unwind;
658 }
659 for (i = 0; i < nrealreaders; i++) {
660 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
661 reader_tasks[i]);
662 if (firsterr)
663 goto unwind;
664 }
665 while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
666 schedule_timeout_uninterruptible(1);
667 writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
668 GFP_KERNEL);
669 writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
670 GFP_KERNEL);
671 writer_n_durations =
672 kcalloc(nrealwriters, sizeof(*writer_n_durations),
673 GFP_KERNEL);
674 if (!writer_tasks || !writer_durations || !writer_n_durations) {
675 VERBOSE_PERFOUT_ERRSTRING("out of memory");
676 firsterr = -ENOMEM;
677 goto unwind;
678 }
679 for (i = 0; i < nrealwriters; i++) {
680 writer_durations[i] =
681 kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
682 GFP_KERNEL);
683 if (!writer_durations[i]) {
684 firsterr = -ENOMEM;
685 goto unwind;
686 }
687 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
688 writer_tasks[i]);
689 if (firsterr)
690 goto unwind;
691 }
692 torture_init_end();
693 return 0;
694
695unwind:
696 torture_init_end();
697 rcu_perf_cleanup();
698 return firsterr;
699}
700
701module_init(rcu_perf_init);
702module_exit(rcu_perf_cleanup);