Merge tag 'ide-5.11-2021-02-28' of git://git.kernel.dk/linux-block
[linux-2.6-microblaze.git] / kernel / rcu / refscale.c
1 // SPDX-License-Identifier: GPL-2.0+
2 //
3 // Scalability test comparing RCU vs other mechanisms
4 // for acquiring references on objects.
5 //
6 // Copyright (C) Google, 2020.
7 //
8 // Author: Joel Fernandes <joel@joelfernandes.org>
9
10 #define pr_fmt(fmt) fmt
11
12 #include <linux/atomic.h>
13 #include <linux/bitops.h>
14 #include <linux/completion.h>
15 #include <linux/cpu.h>
16 #include <linux/delay.h>
17 #include <linux/err.h>
18 #include <linux/init.h>
19 #include <linux/interrupt.h>
20 #include <linux/kthread.h>
21 #include <linux/kernel.h>
22 #include <linux/mm.h>
23 #include <linux/module.h>
24 #include <linux/moduleparam.h>
25 #include <linux/notifier.h>
26 #include <linux/percpu.h>
27 #include <linux/rcupdate.h>
28 #include <linux/rcupdate_trace.h>
29 #include <linux/reboot.h>
30 #include <linux/sched.h>
31 #include <linux/spinlock.h>
32 #include <linux/smp.h>
33 #include <linux/stat.h>
34 #include <linux/srcu.h>
35 #include <linux/slab.h>
36 #include <linux/torture.h>
37 #include <linux/types.h>
38
39 #include "rcu.h"
40
41 #define SCALE_FLAG "-ref-scale: "
42
43 #define SCALEOUT(s, x...) \
44         pr_alert("%s" SCALE_FLAG s, scale_type, ## x)
45
46 #define VERBOSE_SCALEOUT(s, x...) \
47         do { if (verbose) pr_alert("%s" SCALE_FLAG s, scale_type, ## x); } while (0)
48
49 static atomic_t verbose_batch_ctr;
50
51 #define VERBOSE_SCALEOUT_BATCH(s, x...)                                                 \
52 do {                                                                                    \
53         if (verbose &&                                                                  \
54             (verbose_batched <= 0 ||                                                    \
55              !(atomic_inc_return(&verbose_batch_ctr) % verbose_batched))) {             \
56                 schedule_timeout_uninterruptible(1);                                    \
57                 pr_alert("%s" SCALE_FLAG s, scale_type, ## x);                          \
58         }                                                                               \
59 } while (0)
60
61 #define VERBOSE_SCALEOUT_ERRSTRING(s, x...) \
62         do { if (verbose) pr_alert("%s" SCALE_FLAG "!!! " s, scale_type, ## x); } while (0)
63
64 MODULE_LICENSE("GPL");
65 MODULE_AUTHOR("Joel Fernandes (Google) <joel@joelfernandes.org>");
66
67 static char *scale_type = "rcu";
68 module_param(scale_type, charp, 0444);
69 MODULE_PARM_DESC(scale_type, "Type of test (rcu, srcu, refcnt, rwsem, rwlock.");
70
71 torture_param(int, verbose, 0, "Enable verbose debugging printk()s");
72 torture_param(int, verbose_batched, 0, "Batch verbose debugging printk()s");
73
74 // Wait until there are multiple CPUs before starting test.
75 torture_param(int, holdoff, IS_BUILTIN(CONFIG_RCU_REF_SCALE_TEST) ? 10 : 0,
76               "Holdoff time before test start (s)");
77 // Number of loops per experiment, all readers execute operations concurrently.
78 torture_param(long, loops, 10000, "Number of loops per experiment.");
79 // Number of readers, with -1 defaulting to about 75% of the CPUs.
80 torture_param(int, nreaders, -1, "Number of readers, -1 for 75% of CPUs.");
81 // Number of runs.
82 torture_param(int, nruns, 30, "Number of experiments to run.");
83 // Reader delay in nanoseconds, 0 for no delay.
84 torture_param(int, readdelay, 0, "Read-side delay in nanoseconds.");
85
86 #ifdef MODULE
87 # define REFSCALE_SHUTDOWN 0
88 #else
89 # define REFSCALE_SHUTDOWN 1
90 #endif
91
92 torture_param(bool, shutdown, REFSCALE_SHUTDOWN,
93               "Shutdown at end of scalability tests.");
94
95 struct reader_task {
96         struct task_struct *task;
97         int start_reader;
98         wait_queue_head_t wq;
99         u64 last_duration_ns;
100 };
101
102 static struct task_struct *shutdown_task;
103 static wait_queue_head_t shutdown_wq;
104
105 static struct task_struct *main_task;
106 static wait_queue_head_t main_wq;
107 static int shutdown_start;
108
109 static struct reader_task *reader_tasks;
110
111 // Number of readers that are part of the current experiment.
112 static atomic_t nreaders_exp;
113
114 // Use to wait for all threads to start.
115 static atomic_t n_init;
116 static atomic_t n_started;
117 static atomic_t n_warmedup;
118 static atomic_t n_cooleddown;
119
120 // Track which experiment is currently running.
121 static int exp_idx;
122
123 // Operations vector for selecting different types of tests.
124 struct ref_scale_ops {
125         void (*init)(void);
126         void (*cleanup)(void);
127         void (*readsection)(const int nloops);
128         void (*delaysection)(const int nloops, const int udl, const int ndl);
129         const char *name;
130 };
131
132 static struct ref_scale_ops *cur_ops;
133
134 static void un_delay(const int udl, const int ndl)
135 {
136         if (udl)
137                 udelay(udl);
138         if (ndl)
139                 ndelay(ndl);
140 }
141
142 static void ref_rcu_read_section(const int nloops)
143 {
144         int i;
145
146         for (i = nloops; i >= 0; i--) {
147                 rcu_read_lock();
148                 rcu_read_unlock();
149         }
150 }
151
152 static void ref_rcu_delay_section(const int nloops, const int udl, const int ndl)
153 {
154         int i;
155
156         for (i = nloops; i >= 0; i--) {
157                 rcu_read_lock();
158                 un_delay(udl, ndl);
159                 rcu_read_unlock();
160         }
161 }
162
163 static void rcu_sync_scale_init(void)
164 {
165 }
166
167 static struct ref_scale_ops rcu_ops = {
168         .init           = rcu_sync_scale_init,
169         .readsection    = ref_rcu_read_section,
170         .delaysection   = ref_rcu_delay_section,
171         .name           = "rcu"
172 };
173
174 // Definitions for SRCU ref scale testing.
175 DEFINE_STATIC_SRCU(srcu_refctl_scale);
176 static struct srcu_struct *srcu_ctlp = &srcu_refctl_scale;
177
178 static void srcu_ref_scale_read_section(const int nloops)
179 {
180         int i;
181         int idx;
182
183         for (i = nloops; i >= 0; i--) {
184                 idx = srcu_read_lock(srcu_ctlp);
185                 srcu_read_unlock(srcu_ctlp, idx);
186         }
187 }
188
189 static void srcu_ref_scale_delay_section(const int nloops, const int udl, const int ndl)
190 {
191         int i;
192         int idx;
193
194         for (i = nloops; i >= 0; i--) {
195                 idx = srcu_read_lock(srcu_ctlp);
196                 un_delay(udl, ndl);
197                 srcu_read_unlock(srcu_ctlp, idx);
198         }
199 }
200
201 static struct ref_scale_ops srcu_ops = {
202         .init           = rcu_sync_scale_init,
203         .readsection    = srcu_ref_scale_read_section,
204         .delaysection   = srcu_ref_scale_delay_section,
205         .name           = "srcu"
206 };
207
208 // Definitions for RCU Tasks ref scale testing: Empty read markers.
209 // These definitions also work for RCU Rude readers.
210 static void rcu_tasks_ref_scale_read_section(const int nloops)
211 {
212         int i;
213
214         for (i = nloops; i >= 0; i--)
215                 continue;
216 }
217
218 static void rcu_tasks_ref_scale_delay_section(const int nloops, const int udl, const int ndl)
219 {
220         int i;
221
222         for (i = nloops; i >= 0; i--)
223                 un_delay(udl, ndl);
224 }
225
226 static struct ref_scale_ops rcu_tasks_ops = {
227         .init           = rcu_sync_scale_init,
228         .readsection    = rcu_tasks_ref_scale_read_section,
229         .delaysection   = rcu_tasks_ref_scale_delay_section,
230         .name           = "rcu-tasks"
231 };
232
233 // Definitions for RCU Tasks Trace ref scale testing.
234 static void rcu_trace_ref_scale_read_section(const int nloops)
235 {
236         int i;
237
238         for (i = nloops; i >= 0; i--) {
239                 rcu_read_lock_trace();
240                 rcu_read_unlock_trace();
241         }
242 }
243
244 static void rcu_trace_ref_scale_delay_section(const int nloops, const int udl, const int ndl)
245 {
246         int i;
247
248         for (i = nloops; i >= 0; i--) {
249                 rcu_read_lock_trace();
250                 un_delay(udl, ndl);
251                 rcu_read_unlock_trace();
252         }
253 }
254
255 static struct ref_scale_ops rcu_trace_ops = {
256         .init           = rcu_sync_scale_init,
257         .readsection    = rcu_trace_ref_scale_read_section,
258         .delaysection   = rcu_trace_ref_scale_delay_section,
259         .name           = "rcu-trace"
260 };
261
262 // Definitions for reference count
263 static atomic_t refcnt;
264
265 static void ref_refcnt_section(const int nloops)
266 {
267         int i;
268
269         for (i = nloops; i >= 0; i--) {
270                 atomic_inc(&refcnt);
271                 atomic_dec(&refcnt);
272         }
273 }
274
275 static void ref_refcnt_delay_section(const int nloops, const int udl, const int ndl)
276 {
277         int i;
278
279         for (i = nloops; i >= 0; i--) {
280                 atomic_inc(&refcnt);
281                 un_delay(udl, ndl);
282                 atomic_dec(&refcnt);
283         }
284 }
285
286 static struct ref_scale_ops refcnt_ops = {
287         .init           = rcu_sync_scale_init,
288         .readsection    = ref_refcnt_section,
289         .delaysection   = ref_refcnt_delay_section,
290         .name           = "refcnt"
291 };
292
293 // Definitions for rwlock
294 static rwlock_t test_rwlock;
295
296 static void ref_rwlock_init(void)
297 {
298         rwlock_init(&test_rwlock);
299 }
300
301 static void ref_rwlock_section(const int nloops)
302 {
303         int i;
304
305         for (i = nloops; i >= 0; i--) {
306                 read_lock(&test_rwlock);
307                 read_unlock(&test_rwlock);
308         }
309 }
310
311 static void ref_rwlock_delay_section(const int nloops, const int udl, const int ndl)
312 {
313         int i;
314
315         for (i = nloops; i >= 0; i--) {
316                 read_lock(&test_rwlock);
317                 un_delay(udl, ndl);
318                 read_unlock(&test_rwlock);
319         }
320 }
321
322 static struct ref_scale_ops rwlock_ops = {
323         .init           = ref_rwlock_init,
324         .readsection    = ref_rwlock_section,
325         .delaysection   = ref_rwlock_delay_section,
326         .name           = "rwlock"
327 };
328
329 // Definitions for rwsem
330 static struct rw_semaphore test_rwsem;
331
332 static void ref_rwsem_init(void)
333 {
334         init_rwsem(&test_rwsem);
335 }
336
337 static void ref_rwsem_section(const int nloops)
338 {
339         int i;
340
341         for (i = nloops; i >= 0; i--) {
342                 down_read(&test_rwsem);
343                 up_read(&test_rwsem);
344         }
345 }
346
347 static void ref_rwsem_delay_section(const int nloops, const int udl, const int ndl)
348 {
349         int i;
350
351         for (i = nloops; i >= 0; i--) {
352                 down_read(&test_rwsem);
353                 un_delay(udl, ndl);
354                 up_read(&test_rwsem);
355         }
356 }
357
358 static struct ref_scale_ops rwsem_ops = {
359         .init           = ref_rwsem_init,
360         .readsection    = ref_rwsem_section,
361         .delaysection   = ref_rwsem_delay_section,
362         .name           = "rwsem"
363 };
364
365 static void rcu_scale_one_reader(void)
366 {
367         if (readdelay <= 0)
368                 cur_ops->readsection(loops);
369         else
370                 cur_ops->delaysection(loops, readdelay / 1000, readdelay % 1000);
371 }
372
373 // Reader kthread.  Repeatedly does empty RCU read-side
374 // critical section, minimizing update-side interference.
375 static int
376 ref_scale_reader(void *arg)
377 {
378         unsigned long flags;
379         long me = (long)arg;
380         struct reader_task *rt = &(reader_tasks[me]);
381         u64 start;
382         s64 duration;
383
384         VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: task started", me);
385         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
386         set_user_nice(current, MAX_NICE);
387         atomic_inc(&n_init);
388         if (holdoff)
389                 schedule_timeout_interruptible(holdoff * HZ);
390 repeat:
391         VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: waiting to start next experiment on cpu %d", me, smp_processor_id());
392
393         // Wait for signal that this reader can start.
394         wait_event(rt->wq, (atomic_read(&nreaders_exp) && smp_load_acquire(&rt->start_reader)) ||
395                            torture_must_stop());
396
397         if (torture_must_stop())
398                 goto end;
399
400         // Make sure that the CPU is affinitized appropriately during testing.
401         WARN_ON_ONCE(smp_processor_id() != me);
402
403         WRITE_ONCE(rt->start_reader, 0);
404         if (!atomic_dec_return(&n_started))
405                 while (atomic_read_acquire(&n_started))
406                         cpu_relax();
407
408         VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: experiment %d started", me, exp_idx);
409
410
411         // To reduce noise, do an initial cache-warming invocation, check
412         // in, and then keep warming until everyone has checked in.
413         rcu_scale_one_reader();
414         if (!atomic_dec_return(&n_warmedup))
415                 while (atomic_read_acquire(&n_warmedup))
416                         rcu_scale_one_reader();
417         // Also keep interrupts disabled.  This also has the effect
418         // of preventing entries into slow path for rcu_read_unlock().
419         local_irq_save(flags);
420         start = ktime_get_mono_fast_ns();
421
422         rcu_scale_one_reader();
423
424         duration = ktime_get_mono_fast_ns() - start;
425         local_irq_restore(flags);
426
427         rt->last_duration_ns = WARN_ON_ONCE(duration < 0) ? 0 : duration;
428         // To reduce runtime-skew noise, do maintain-load invocations until
429         // everyone is done.
430         if (!atomic_dec_return(&n_cooleddown))
431                 while (atomic_read_acquire(&n_cooleddown))
432                         rcu_scale_one_reader();
433
434         if (atomic_dec_and_test(&nreaders_exp))
435                 wake_up(&main_wq);
436
437         VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: experiment %d ended, (readers remaining=%d)",
438                                 me, exp_idx, atomic_read(&nreaders_exp));
439
440         if (!torture_must_stop())
441                 goto repeat;
442 end:
443         torture_kthread_stopping("ref_scale_reader");
444         return 0;
445 }
446
447 static void reset_readers(void)
448 {
449         int i;
450         struct reader_task *rt;
451
452         for (i = 0; i < nreaders; i++) {
453                 rt = &(reader_tasks[i]);
454
455                 rt->last_duration_ns = 0;
456         }
457 }
458
459 // Print the results of each reader and return the sum of all their durations.
460 static u64 process_durations(int n)
461 {
462         int i;
463         struct reader_task *rt;
464         char buf1[64];
465         char *buf;
466         u64 sum = 0;
467
468         buf = kmalloc(128 + nreaders * 32, GFP_KERNEL);
469         if (!buf)
470                 return 0;
471         buf[0] = 0;
472         sprintf(buf, "Experiment #%d (Format: <THREAD-NUM>:<Total loop time in ns>)",
473                 exp_idx);
474
475         for (i = 0; i < n && !torture_must_stop(); i++) {
476                 rt = &(reader_tasks[i]);
477                 sprintf(buf1, "%d: %llu\t", i, rt->last_duration_ns);
478
479                 if (i % 5 == 0)
480                         strcat(buf, "\n");
481                 strcat(buf, buf1);
482
483                 sum += rt->last_duration_ns;
484         }
485         strcat(buf, "\n");
486
487         SCALEOUT("%s\n", buf);
488
489         kfree(buf);
490         return sum;
491 }
492
493 // The main_func is the main orchestrator, it performs a bunch of
494 // experiments.  For every experiment, it orders all the readers
495 // involved to start and waits for them to finish the experiment. It
496 // then reads their timestamps and starts the next experiment. Each
497 // experiment progresses from 1 concurrent reader to N of them at which
498 // point all the timestamps are printed.
499 static int main_func(void *arg)
500 {
501         bool errexit = false;
502         int exp, r;
503         char buf1[64];
504         char *buf;
505         u64 *result_avg;
506
507         set_cpus_allowed_ptr(current, cpumask_of(nreaders % nr_cpu_ids));
508         set_user_nice(current, MAX_NICE);
509
510         VERBOSE_SCALEOUT("main_func task started");
511         result_avg = kzalloc(nruns * sizeof(*result_avg), GFP_KERNEL);
512         buf = kzalloc(64 + nruns * 32, GFP_KERNEL);
513         if (!result_avg || !buf) {
514                 VERBOSE_SCALEOUT_ERRSTRING("out of memory");
515                 errexit = true;
516         }
517         if (holdoff)
518                 schedule_timeout_interruptible(holdoff * HZ);
519
520         // Wait for all threads to start.
521         atomic_inc(&n_init);
522         while (atomic_read(&n_init) < nreaders + 1)
523                 schedule_timeout_uninterruptible(1);
524
525         // Start exp readers up per experiment
526         for (exp = 0; exp < nruns && !torture_must_stop(); exp++) {
527                 if (errexit)
528                         break;
529                 if (torture_must_stop())
530                         goto end;
531
532                 reset_readers();
533                 atomic_set(&nreaders_exp, nreaders);
534                 atomic_set(&n_started, nreaders);
535                 atomic_set(&n_warmedup, nreaders);
536                 atomic_set(&n_cooleddown, nreaders);
537
538                 exp_idx = exp;
539
540                 for (r = 0; r < nreaders; r++) {
541                         smp_store_release(&reader_tasks[r].start_reader, 1);
542                         wake_up(&reader_tasks[r].wq);
543                 }
544
545                 VERBOSE_SCALEOUT("main_func: experiment started, waiting for %d readers",
546                                 nreaders);
547
548                 wait_event(main_wq,
549                            !atomic_read(&nreaders_exp) || torture_must_stop());
550
551                 VERBOSE_SCALEOUT("main_func: experiment ended");
552
553                 if (torture_must_stop())
554                         goto end;
555
556                 result_avg[exp] = div_u64(1000 * process_durations(nreaders), nreaders * loops);
557         }
558
559         // Print the average of all experiments
560         SCALEOUT("END OF TEST. Calculating average duration per loop (nanoseconds)...\n");
561
562         if (!errexit) {
563                 buf[0] = 0;
564                 strcat(buf, "\n");
565                 strcat(buf, "Runs\tTime(ns)\n");
566         }
567
568         for (exp = 0; exp < nruns; exp++) {
569                 u64 avg;
570                 u32 rem;
571
572                 if (errexit)
573                         break;
574                 avg = div_u64_rem(result_avg[exp], 1000, &rem);
575                 sprintf(buf1, "%d\t%llu.%03u\n", exp + 1, avg, rem);
576                 strcat(buf, buf1);
577         }
578
579         if (!errexit)
580                 SCALEOUT("%s", buf);
581
582         // This will shutdown everything including us.
583         if (shutdown) {
584                 shutdown_start = 1;
585                 wake_up(&shutdown_wq);
586         }
587
588         // Wait for torture to stop us
589         while (!torture_must_stop())
590                 schedule_timeout_uninterruptible(1);
591
592 end:
593         torture_kthread_stopping("main_func");
594         kfree(result_avg);
595         kfree(buf);
596         return 0;
597 }
598
599 static void
600 ref_scale_print_module_parms(struct ref_scale_ops *cur_ops, const char *tag)
601 {
602         pr_alert("%s" SCALE_FLAG
603                  "--- %s:  verbose=%d shutdown=%d holdoff=%d loops=%ld nreaders=%d nruns=%d readdelay=%d\n", scale_type, tag,
604                  verbose, shutdown, holdoff, loops, nreaders, nruns, readdelay);
605 }
606
607 static void
608 ref_scale_cleanup(void)
609 {
610         int i;
611
612         if (torture_cleanup_begin())
613                 return;
614
615         if (!cur_ops) {
616                 torture_cleanup_end();
617                 return;
618         }
619
620         if (reader_tasks) {
621                 for (i = 0; i < nreaders; i++)
622                         torture_stop_kthread("ref_scale_reader",
623                                              reader_tasks[i].task);
624         }
625         kfree(reader_tasks);
626
627         torture_stop_kthread("main_task", main_task);
628         kfree(main_task);
629
630         // Do scale-type-specific cleanup operations.
631         if (cur_ops->cleanup != NULL)
632                 cur_ops->cleanup();
633
634         torture_cleanup_end();
635 }
636
637 // Shutdown kthread.  Just waits to be awakened, then shuts down system.
638 static int
639 ref_scale_shutdown(void *arg)
640 {
641         wait_event(shutdown_wq, shutdown_start);
642
643         smp_mb(); // Wake before output.
644         ref_scale_cleanup();
645         kernel_power_off();
646
647         return -EINVAL;
648 }
649
650 static int __init
651 ref_scale_init(void)
652 {
653         long i;
654         int firsterr = 0;
655         static struct ref_scale_ops *scale_ops[] = {
656                 &rcu_ops, &srcu_ops, &rcu_trace_ops, &rcu_tasks_ops,
657                 &refcnt_ops, &rwlock_ops, &rwsem_ops,
658         };
659
660         if (!torture_init_begin(scale_type, verbose))
661                 return -EBUSY;
662
663         for (i = 0; i < ARRAY_SIZE(scale_ops); i++) {
664                 cur_ops = scale_ops[i];
665                 if (strcmp(scale_type, cur_ops->name) == 0)
666                         break;
667         }
668         if (i == ARRAY_SIZE(scale_ops)) {
669                 pr_alert("rcu-scale: invalid scale type: \"%s\"\n", scale_type);
670                 pr_alert("rcu-scale types:");
671                 for (i = 0; i < ARRAY_SIZE(scale_ops); i++)
672                         pr_cont(" %s", scale_ops[i]->name);
673                 pr_cont("\n");
674                 firsterr = -EINVAL;
675                 cur_ops = NULL;
676                 goto unwind;
677         }
678         if (cur_ops->init)
679                 cur_ops->init();
680
681         ref_scale_print_module_parms(cur_ops, "Start of test");
682
683         // Shutdown task
684         if (shutdown) {
685                 init_waitqueue_head(&shutdown_wq);
686                 firsterr = torture_create_kthread(ref_scale_shutdown, NULL,
687                                                   shutdown_task);
688                 if (firsterr)
689                         goto unwind;
690                 schedule_timeout_uninterruptible(1);
691         }
692
693         // Reader tasks (default to ~75% of online CPUs).
694         if (nreaders < 0)
695                 nreaders = (num_online_cpus() >> 1) + (num_online_cpus() >> 2);
696         if (WARN_ONCE(loops <= 0, "%s: loops = %ld, adjusted to 1\n", __func__, loops))
697                 loops = 1;
698         if (WARN_ONCE(nreaders <= 0, "%s: nreaders = %d, adjusted to 1\n", __func__, nreaders))
699                 nreaders = 1;
700         if (WARN_ONCE(nruns <= 0, "%s: nruns = %d, adjusted to 1\n", __func__, nruns))
701                 nruns = 1;
702         reader_tasks = kcalloc(nreaders, sizeof(reader_tasks[0]),
703                                GFP_KERNEL);
704         if (!reader_tasks) {
705                 VERBOSE_SCALEOUT_ERRSTRING("out of memory");
706                 firsterr = -ENOMEM;
707                 goto unwind;
708         }
709
710         VERBOSE_SCALEOUT("Starting %d reader threads\n", nreaders);
711
712         for (i = 0; i < nreaders; i++) {
713                 firsterr = torture_create_kthread(ref_scale_reader, (void *)i,
714                                                   reader_tasks[i].task);
715                 if (firsterr)
716                         goto unwind;
717
718                 init_waitqueue_head(&(reader_tasks[i].wq));
719         }
720
721         // Main Task
722         init_waitqueue_head(&main_wq);
723         firsterr = torture_create_kthread(main_func, NULL, main_task);
724         if (firsterr)
725                 goto unwind;
726
727         torture_init_end();
728         return 0;
729
730 unwind:
731         torture_init_end();
732         ref_scale_cleanup();
733         if (shutdown) {
734                 WARN_ON(!IS_MODULE(CONFIG_RCU_REF_SCALE_TEST));
735                 kernel_power_off();
736         }
737         return firsterr;
738 }
739
740 module_init(ref_scale_init);
741 module_exit(ref_scale_cleanup);