This source file includes following definitions.
- rcu_perf_read_lock
- rcu_perf_read_unlock
- rcu_no_completed
- rcu_sync_perf_init
- srcu_perf_read_lock
- srcu_perf_read_unlock
- srcu_perf_completed
- srcu_call_rcu
- srcu_rcu_barrier
- srcu_perf_synchronize
- srcu_perf_synchronize_expedited
- srcu_sync_perf_init
- srcu_sync_perf_cleanup
- tasks_perf_read_lock
- tasks_perf_read_unlock
- rcuperf_seq_diff
- rcu_perf_wait_shutdown
- rcu_perf_reader
- rcu_perf_async_cb
- rcu_perf_writer
- rcu_perf_print_module_parms
- rcu_perf_cleanup
- compute_real
- rcu_perf_shutdown
- rcu_perf_init
1
2
3
4
5
6
7
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
43 MODULE_LICENSE("GPL");
44 MODULE_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
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73 #ifdef MODULE
74 # define RCUPERF_SHUTDOWN 0
75 #else
76 # define RCUPERF_SHUTDOWN 1
77 #endif
78
79 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
80 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
81 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
82 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
83 torture_param(int, nreaders, -1, "Number of RCU reader threads");
84 torture_param(int, nwriters, -1, "Number of RCU updater threads");
85 torture_param(bool, shutdown, RCUPERF_SHUTDOWN,
86 "Shutdown at end of performance tests.");
87 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
88 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
89
90 static char *perf_type = "rcu";
91 module_param(perf_type, charp, 0444);
92 MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, srcu, ...)");
93
94 static int nrealreaders;
95 static int nrealwriters;
96 static struct task_struct **writer_tasks;
97 static struct task_struct **reader_tasks;
98 static struct task_struct *shutdown_task;
99
100 static u64 **writer_durations;
101 static int *writer_n_durations;
102 static atomic_t n_rcu_perf_reader_started;
103 static atomic_t n_rcu_perf_writer_started;
104 static atomic_t n_rcu_perf_writer_finished;
105 static wait_queue_head_t shutdown_wq;
106 static u64 t_rcu_perf_writer_started;
107 static u64 t_rcu_perf_writer_finished;
108 static unsigned long b_rcu_perf_writer_started;
109 static unsigned long b_rcu_perf_writer_finished;
110 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
111
112 static 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
126
127
128 struct 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
144 static struct rcu_perf_ops *cur_ops;
145
146
147
148
149
150 static int rcu_perf_read_lock(void) __acquires(RCU)
151 {
152 rcu_read_lock();
153 return 0;
154 }
155
156 static void rcu_perf_read_unlock(int idx) __releases(RCU)
157 {
158 rcu_read_unlock();
159 }
160
161 static unsigned long __maybe_unused rcu_no_completed(void)
162 {
163 return 0;
164 }
165
166 static void rcu_sync_perf_init(void)
167 {
168 }
169
170 static 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
187
188
189 DEFINE_STATIC_SRCU(srcu_ctl_perf);
190 static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
191
192 static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
193 {
194 return srcu_read_lock(srcu_ctlp);
195 }
196
197 static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
198 {
199 srcu_read_unlock(srcu_ctlp, idx);
200 }
201
202 static unsigned long srcu_perf_completed(void)
203 {
204 return srcu_batches_completed(srcu_ctlp);
205 }
206
207 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
208 {
209 call_srcu(srcu_ctlp, head, func);
210 }
211
212 static void srcu_rcu_barrier(void)
213 {
214 srcu_barrier(srcu_ctlp);
215 }
216
217 static void srcu_perf_synchronize(void)
218 {
219 synchronize_srcu(srcu_ctlp);
220 }
221
222 static void srcu_perf_synchronize_expedited(void)
223 {
224 synchronize_srcu_expedited(srcu_ctlp);
225 }
226
227 static 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
242 static struct srcu_struct srcud;
243
244 static void srcu_sync_perf_init(void)
245 {
246 srcu_ctlp = &srcud;
247 init_srcu_struct(srcu_ctlp);
248 }
249
250 static void srcu_sync_perf_cleanup(void)
251 {
252 cleanup_srcu_struct(srcu_ctlp);
253 }
254
255 static 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
273
274
275 static int tasks_perf_read_lock(void)
276 {
277 return 0;
278 }
279
280 static void tasks_perf_read_unlock(int idx)
281 {
282 }
283
284 static 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
298 static 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
307
308 static 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
319
320
321 static int
322 rcu_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
346
347 static 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
355
356 static int
357 rcu_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
380
381
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) {
403 retry:
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);
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();
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
475 static void
476 rcu_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
483 static void
484 rcu_perf_cleanup(void)
485 {
486 int i;
487 int j;
488 int ngps = 0;
489 u64 *wdp;
490 u64 *wdpp;
491
492
493
494
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
560 if (cur_ops->cleanup != NULL)
561 cur_ops->cleanup();
562
563 torture_cleanup_end();
564 }
565
566
567
568
569
570
571 static 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
587
588
589 static int
590 rcu_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();
598 rcu_perf_cleanup();
599 kernel_power_off();
600 return -EINVAL;
601 }
602
603 static int __init
604 rcu_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
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
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
695 unwind:
696 torture_init_end();
697 rcu_perf_cleanup();
698 return firsterr;
699 }
700
701 module_init(rcu_perf_init);
702 module_exit(rcu_perf_cleanup);