Loading...
Note: File does not exist in v4.6.
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 { \
48 if (verbose) \
49 pr_alert("%s" SCALE_FLAG s "\n", scale_type, ## x); \
50 } while (0)
51
52static atomic_t verbose_batch_ctr;
53
54#define VERBOSE_SCALEOUT_BATCH(s, x...) \
55do { \
56 if (verbose && \
57 (verbose_batched <= 0 || \
58 !(atomic_inc_return(&verbose_batch_ctr) % verbose_batched))) { \
59 schedule_timeout_uninterruptible(1); \
60 pr_alert("%s" SCALE_FLAG s "\n", scale_type, ## x); \
61 } \
62} while (0)
63
64#define SCALEOUT_ERRSTRING(s, x...) pr_alert("%s" SCALE_FLAG "!!! " s "\n", scale_type, ## x)
65
66MODULE_LICENSE("GPL");
67MODULE_AUTHOR("Joel Fernandes (Google) <joel@joelfernandes.org>");
68
69static char *scale_type = "rcu";
70module_param(scale_type, charp, 0444);
71MODULE_PARM_DESC(scale_type, "Type of test (rcu, srcu, refcnt, rwsem, rwlock.");
72
73torture_param(int, verbose, 0, "Enable verbose debugging printk()s");
74torture_param(int, verbose_batched, 0, "Batch verbose debugging printk()s");
75
76// Wait until there are multiple CPUs before starting test.
77torture_param(int, holdoff, IS_BUILTIN(CONFIG_RCU_REF_SCALE_TEST) ? 10 : 0,
78 "Holdoff time before test start (s)");
79// Number of loops per experiment, all readers execute operations concurrently.
80torture_param(long, loops, 10000, "Number of loops per experiment.");
81// Number of readers, with -1 defaulting to about 75% of the CPUs.
82torture_param(int, nreaders, -1, "Number of readers, -1 for 75% of CPUs.");
83// Number of runs.
84torture_param(int, nruns, 30, "Number of experiments to run.");
85// Reader delay in nanoseconds, 0 for no delay.
86torture_param(int, readdelay, 0, "Read-side delay in nanoseconds.");
87
88#ifdef MODULE
89# define REFSCALE_SHUTDOWN 0
90#else
91# define REFSCALE_SHUTDOWN 1
92#endif
93
94torture_param(bool, shutdown, REFSCALE_SHUTDOWN,
95 "Shutdown at end of scalability tests.");
96
97struct reader_task {
98 struct task_struct *task;
99 int start_reader;
100 wait_queue_head_t wq;
101 u64 last_duration_ns;
102};
103
104static struct task_struct *shutdown_task;
105static wait_queue_head_t shutdown_wq;
106
107static struct task_struct *main_task;
108static wait_queue_head_t main_wq;
109static int shutdown_start;
110
111static struct reader_task *reader_tasks;
112
113// Number of readers that are part of the current experiment.
114static atomic_t nreaders_exp;
115
116// Use to wait for all threads to start.
117static atomic_t n_init;
118static atomic_t n_started;
119static atomic_t n_warmedup;
120static atomic_t n_cooleddown;
121
122// Track which experiment is currently running.
123static int exp_idx;
124
125// Operations vector for selecting different types of tests.
126struct ref_scale_ops {
127 void (*init)(void);
128 void (*cleanup)(void);
129 void (*readsection)(const int nloops);
130 void (*delaysection)(const int nloops, const int udl, const int ndl);
131 const char *name;
132};
133
134static struct ref_scale_ops *cur_ops;
135
136static void un_delay(const int udl, const int ndl)
137{
138 if (udl)
139 udelay(udl);
140 if (ndl)
141 ndelay(ndl);
142}
143
144static void ref_rcu_read_section(const int nloops)
145{
146 int i;
147
148 for (i = nloops; i >= 0; i--) {
149 rcu_read_lock();
150 rcu_read_unlock();
151 }
152}
153
154static void ref_rcu_delay_section(const int nloops, const int udl, const int ndl)
155{
156 int i;
157
158 for (i = nloops; i >= 0; i--) {
159 rcu_read_lock();
160 un_delay(udl, ndl);
161 rcu_read_unlock();
162 }
163}
164
165static void rcu_sync_scale_init(void)
166{
167}
168
169static struct ref_scale_ops rcu_ops = {
170 .init = rcu_sync_scale_init,
171 .readsection = ref_rcu_read_section,
172 .delaysection = ref_rcu_delay_section,
173 .name = "rcu"
174};
175
176// Definitions for SRCU ref scale testing.
177DEFINE_STATIC_SRCU(srcu_refctl_scale);
178static struct srcu_struct *srcu_ctlp = &srcu_refctl_scale;
179
180static void srcu_ref_scale_read_section(const int nloops)
181{
182 int i;
183 int idx;
184
185 for (i = nloops; i >= 0; i--) {
186 idx = srcu_read_lock(srcu_ctlp);
187 srcu_read_unlock(srcu_ctlp, idx);
188 }
189}
190
191static void srcu_ref_scale_delay_section(const int nloops, const int udl, const int ndl)
192{
193 int i;
194 int idx;
195
196 for (i = nloops; i >= 0; i--) {
197 idx = srcu_read_lock(srcu_ctlp);
198 un_delay(udl, ndl);
199 srcu_read_unlock(srcu_ctlp, idx);
200 }
201}
202
203static struct ref_scale_ops srcu_ops = {
204 .init = rcu_sync_scale_init,
205 .readsection = srcu_ref_scale_read_section,
206 .delaysection = srcu_ref_scale_delay_section,
207 .name = "srcu"
208};
209
210#ifdef CONFIG_TASKS_RCU
211
212// Definitions for RCU Tasks ref scale testing: Empty read markers.
213// These definitions also work for RCU Rude readers.
214static void rcu_tasks_ref_scale_read_section(const int nloops)
215{
216 int i;
217
218 for (i = nloops; i >= 0; i--)
219 continue;
220}
221
222static void rcu_tasks_ref_scale_delay_section(const int nloops, const int udl, const int ndl)
223{
224 int i;
225
226 for (i = nloops; i >= 0; i--)
227 un_delay(udl, ndl);
228}
229
230static struct ref_scale_ops rcu_tasks_ops = {
231 .init = rcu_sync_scale_init,
232 .readsection = rcu_tasks_ref_scale_read_section,
233 .delaysection = rcu_tasks_ref_scale_delay_section,
234 .name = "rcu-tasks"
235};
236
237#define RCU_TASKS_OPS &rcu_tasks_ops,
238
239#else // #ifdef CONFIG_TASKS_RCU
240
241#define RCU_TASKS_OPS
242
243#endif // #else // #ifdef CONFIG_TASKS_RCU
244
245#ifdef CONFIG_TASKS_TRACE_RCU
246
247// Definitions for RCU Tasks Trace ref scale testing.
248static void rcu_trace_ref_scale_read_section(const int nloops)
249{
250 int i;
251
252 for (i = nloops; i >= 0; i--) {
253 rcu_read_lock_trace();
254 rcu_read_unlock_trace();
255 }
256}
257
258static void rcu_trace_ref_scale_delay_section(const int nloops, const int udl, const int ndl)
259{
260 int i;
261
262 for (i = nloops; i >= 0; i--) {
263 rcu_read_lock_trace();
264 un_delay(udl, ndl);
265 rcu_read_unlock_trace();
266 }
267}
268
269static struct ref_scale_ops rcu_trace_ops = {
270 .init = rcu_sync_scale_init,
271 .readsection = rcu_trace_ref_scale_read_section,
272 .delaysection = rcu_trace_ref_scale_delay_section,
273 .name = "rcu-trace"
274};
275
276#define RCU_TRACE_OPS &rcu_trace_ops,
277
278#else // #ifdef CONFIG_TASKS_TRACE_RCU
279
280#define RCU_TRACE_OPS
281
282#endif // #else // #ifdef CONFIG_TASKS_TRACE_RCU
283
284// Definitions for reference count
285static atomic_t refcnt;
286
287static void ref_refcnt_section(const int nloops)
288{
289 int i;
290
291 for (i = nloops; i >= 0; i--) {
292 atomic_inc(&refcnt);
293 atomic_dec(&refcnt);
294 }
295}
296
297static void ref_refcnt_delay_section(const int nloops, const int udl, const int ndl)
298{
299 int i;
300
301 for (i = nloops; i >= 0; i--) {
302 atomic_inc(&refcnt);
303 un_delay(udl, ndl);
304 atomic_dec(&refcnt);
305 }
306}
307
308static struct ref_scale_ops refcnt_ops = {
309 .init = rcu_sync_scale_init,
310 .readsection = ref_refcnt_section,
311 .delaysection = ref_refcnt_delay_section,
312 .name = "refcnt"
313};
314
315// Definitions for rwlock
316static rwlock_t test_rwlock;
317
318static void ref_rwlock_init(void)
319{
320 rwlock_init(&test_rwlock);
321}
322
323static void ref_rwlock_section(const int nloops)
324{
325 int i;
326
327 for (i = nloops; i >= 0; i--) {
328 read_lock(&test_rwlock);
329 read_unlock(&test_rwlock);
330 }
331}
332
333static void ref_rwlock_delay_section(const int nloops, const int udl, const int ndl)
334{
335 int i;
336
337 for (i = nloops; i >= 0; i--) {
338 read_lock(&test_rwlock);
339 un_delay(udl, ndl);
340 read_unlock(&test_rwlock);
341 }
342}
343
344static struct ref_scale_ops rwlock_ops = {
345 .init = ref_rwlock_init,
346 .readsection = ref_rwlock_section,
347 .delaysection = ref_rwlock_delay_section,
348 .name = "rwlock"
349};
350
351// Definitions for rwsem
352static struct rw_semaphore test_rwsem;
353
354static void ref_rwsem_init(void)
355{
356 init_rwsem(&test_rwsem);
357}
358
359static void ref_rwsem_section(const int nloops)
360{
361 int i;
362
363 for (i = nloops; i >= 0; i--) {
364 down_read(&test_rwsem);
365 up_read(&test_rwsem);
366 }
367}
368
369static void ref_rwsem_delay_section(const int nloops, const int udl, const int ndl)
370{
371 int i;
372
373 for (i = nloops; i >= 0; i--) {
374 down_read(&test_rwsem);
375 un_delay(udl, ndl);
376 up_read(&test_rwsem);
377 }
378}
379
380static struct ref_scale_ops rwsem_ops = {
381 .init = ref_rwsem_init,
382 .readsection = ref_rwsem_section,
383 .delaysection = ref_rwsem_delay_section,
384 .name = "rwsem"
385};
386
387// Definitions for global spinlock
388static DEFINE_RAW_SPINLOCK(test_lock);
389
390static void ref_lock_section(const int nloops)
391{
392 int i;
393
394 preempt_disable();
395 for (i = nloops; i >= 0; i--) {
396 raw_spin_lock(&test_lock);
397 raw_spin_unlock(&test_lock);
398 }
399 preempt_enable();
400}
401
402static void ref_lock_delay_section(const int nloops, const int udl, const int ndl)
403{
404 int i;
405
406 preempt_disable();
407 for (i = nloops; i >= 0; i--) {
408 raw_spin_lock(&test_lock);
409 un_delay(udl, ndl);
410 raw_spin_unlock(&test_lock);
411 }
412 preempt_enable();
413}
414
415static struct ref_scale_ops lock_ops = {
416 .readsection = ref_lock_section,
417 .delaysection = ref_lock_delay_section,
418 .name = "lock"
419};
420
421// Definitions for global irq-save spinlock
422
423static void ref_lock_irq_section(const int nloops)
424{
425 unsigned long flags;
426 int i;
427
428 preempt_disable();
429 for (i = nloops; i >= 0; i--) {
430 raw_spin_lock_irqsave(&test_lock, flags);
431 raw_spin_unlock_irqrestore(&test_lock, flags);
432 }
433 preempt_enable();
434}
435
436static void ref_lock_irq_delay_section(const int nloops, const int udl, const int ndl)
437{
438 unsigned long flags;
439 int i;
440
441 preempt_disable();
442 for (i = nloops; i >= 0; i--) {
443 raw_spin_lock_irqsave(&test_lock, flags);
444 un_delay(udl, ndl);
445 raw_spin_unlock_irqrestore(&test_lock, flags);
446 }
447 preempt_enable();
448}
449
450static struct ref_scale_ops lock_irq_ops = {
451 .readsection = ref_lock_irq_section,
452 .delaysection = ref_lock_irq_delay_section,
453 .name = "lock-irq"
454};
455
456// Definitions acquire-release.
457static DEFINE_PER_CPU(unsigned long, test_acqrel);
458
459static void ref_acqrel_section(const int nloops)
460{
461 unsigned long x;
462 int i;
463
464 preempt_disable();
465 for (i = nloops; i >= 0; i--) {
466 x = smp_load_acquire(this_cpu_ptr(&test_acqrel));
467 smp_store_release(this_cpu_ptr(&test_acqrel), x + 1);
468 }
469 preempt_enable();
470}
471
472static void ref_acqrel_delay_section(const int nloops, const int udl, const int ndl)
473{
474 unsigned long x;
475 int i;
476
477 preempt_disable();
478 for (i = nloops; i >= 0; i--) {
479 x = smp_load_acquire(this_cpu_ptr(&test_acqrel));
480 un_delay(udl, ndl);
481 smp_store_release(this_cpu_ptr(&test_acqrel), x + 1);
482 }
483 preempt_enable();
484}
485
486static struct ref_scale_ops acqrel_ops = {
487 .readsection = ref_acqrel_section,
488 .delaysection = ref_acqrel_delay_section,
489 .name = "acqrel"
490};
491
492static volatile u64 stopopts;
493
494static void ref_clock_section(const int nloops)
495{
496 u64 x = 0;
497 int i;
498
499 preempt_disable();
500 for (i = nloops; i >= 0; i--)
501 x += ktime_get_real_fast_ns();
502 preempt_enable();
503 stopopts = x;
504}
505
506static void ref_clock_delay_section(const int nloops, const int udl, const int ndl)
507{
508 u64 x = 0;
509 int i;
510
511 preempt_disable();
512 for (i = nloops; i >= 0; i--) {
513 x += ktime_get_real_fast_ns();
514 un_delay(udl, ndl);
515 }
516 preempt_enable();
517 stopopts = x;
518}
519
520static struct ref_scale_ops clock_ops = {
521 .readsection = ref_clock_section,
522 .delaysection = ref_clock_delay_section,
523 .name = "clock"
524};
525
526static void rcu_scale_one_reader(void)
527{
528 if (readdelay <= 0)
529 cur_ops->readsection(loops);
530 else
531 cur_ops->delaysection(loops, readdelay / 1000, readdelay % 1000);
532}
533
534// Reader kthread. Repeatedly does empty RCU read-side
535// critical section, minimizing update-side interference.
536static int
537ref_scale_reader(void *arg)
538{
539 unsigned long flags;
540 long me = (long)arg;
541 struct reader_task *rt = &(reader_tasks[me]);
542 u64 start;
543 s64 duration;
544
545 VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: task started", me);
546 WARN_ON_ONCE(set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids)));
547 set_user_nice(current, MAX_NICE);
548 atomic_inc(&n_init);
549 if (holdoff)
550 schedule_timeout_interruptible(holdoff * HZ);
551repeat:
552 VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: waiting to start next experiment on cpu %d", me, raw_smp_processor_id());
553
554 // Wait for signal that this reader can start.
555 wait_event(rt->wq, (atomic_read(&nreaders_exp) && smp_load_acquire(&rt->start_reader)) ||
556 torture_must_stop());
557
558 if (torture_must_stop())
559 goto end;
560
561 // Make sure that the CPU is affinitized appropriately during testing.
562 WARN_ON_ONCE(raw_smp_processor_id() != me);
563
564 WRITE_ONCE(rt->start_reader, 0);
565 if (!atomic_dec_return(&n_started))
566 while (atomic_read_acquire(&n_started))
567 cpu_relax();
568
569 VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: experiment %d started", me, exp_idx);
570
571
572 // To reduce noise, do an initial cache-warming invocation, check
573 // in, and then keep warming until everyone has checked in.
574 rcu_scale_one_reader();
575 if (!atomic_dec_return(&n_warmedup))
576 while (atomic_read_acquire(&n_warmedup))
577 rcu_scale_one_reader();
578 // Also keep interrupts disabled. This also has the effect
579 // of preventing entries into slow path for rcu_read_unlock().
580 local_irq_save(flags);
581 start = ktime_get_mono_fast_ns();
582
583 rcu_scale_one_reader();
584
585 duration = ktime_get_mono_fast_ns() - start;
586 local_irq_restore(flags);
587
588 rt->last_duration_ns = WARN_ON_ONCE(duration < 0) ? 0 : duration;
589 // To reduce runtime-skew noise, do maintain-load invocations until
590 // everyone is done.
591 if (!atomic_dec_return(&n_cooleddown))
592 while (atomic_read_acquire(&n_cooleddown))
593 rcu_scale_one_reader();
594
595 if (atomic_dec_and_test(&nreaders_exp))
596 wake_up(&main_wq);
597
598 VERBOSE_SCALEOUT_BATCH("ref_scale_reader %ld: experiment %d ended, (readers remaining=%d)",
599 me, exp_idx, atomic_read(&nreaders_exp));
600
601 if (!torture_must_stop())
602 goto repeat;
603end:
604 torture_kthread_stopping("ref_scale_reader");
605 return 0;
606}
607
608static void reset_readers(void)
609{
610 int i;
611 struct reader_task *rt;
612
613 for (i = 0; i < nreaders; i++) {
614 rt = &(reader_tasks[i]);
615
616 rt->last_duration_ns = 0;
617 }
618}
619
620// Print the results of each reader and return the sum of all their durations.
621static u64 process_durations(int n)
622{
623 int i;
624 struct reader_task *rt;
625 char buf1[64];
626 char *buf;
627 u64 sum = 0;
628
629 buf = kmalloc(800 + 64, GFP_KERNEL);
630 if (!buf)
631 return 0;
632 buf[0] = 0;
633 sprintf(buf, "Experiment #%d (Format: <THREAD-NUM>:<Total loop time in ns>)",
634 exp_idx);
635
636 for (i = 0; i < n && !torture_must_stop(); i++) {
637 rt = &(reader_tasks[i]);
638 sprintf(buf1, "%d: %llu\t", i, rt->last_duration_ns);
639
640 if (i % 5 == 0)
641 strcat(buf, "\n");
642 if (strlen(buf) >= 800) {
643 pr_alert("%s", buf);
644 buf[0] = 0;
645 }
646 strcat(buf, buf1);
647
648 sum += rt->last_duration_ns;
649 }
650 pr_alert("%s\n", buf);
651
652 kfree(buf);
653 return sum;
654}
655
656// The main_func is the main orchestrator, it performs a bunch of
657// experiments. For every experiment, it orders all the readers
658// involved to start and waits for them to finish the experiment. It
659// then reads their timestamps and starts the next experiment. Each
660// experiment progresses from 1 concurrent reader to N of them at which
661// point all the timestamps are printed.
662static int main_func(void *arg)
663{
664 int exp, r;
665 char buf1[64];
666 char *buf;
667 u64 *result_avg;
668
669 set_cpus_allowed_ptr(current, cpumask_of(nreaders % nr_cpu_ids));
670 set_user_nice(current, MAX_NICE);
671
672 VERBOSE_SCALEOUT("main_func task started");
673 result_avg = kzalloc(nruns * sizeof(*result_avg), GFP_KERNEL);
674 buf = kzalloc(800 + 64, GFP_KERNEL);
675 if (!result_avg || !buf) {
676 SCALEOUT_ERRSTRING("out of memory");
677 goto oom_exit;
678 }
679 if (holdoff)
680 schedule_timeout_interruptible(holdoff * HZ);
681
682 // Wait for all threads to start.
683 atomic_inc(&n_init);
684 while (atomic_read(&n_init) < nreaders + 1)
685 schedule_timeout_uninterruptible(1);
686
687 // Start exp readers up per experiment
688 for (exp = 0; exp < nruns && !torture_must_stop(); exp++) {
689 if (torture_must_stop())
690 goto end;
691
692 reset_readers();
693 atomic_set(&nreaders_exp, nreaders);
694 atomic_set(&n_started, nreaders);
695 atomic_set(&n_warmedup, nreaders);
696 atomic_set(&n_cooleddown, nreaders);
697
698 exp_idx = exp;
699
700 for (r = 0; r < nreaders; r++) {
701 smp_store_release(&reader_tasks[r].start_reader, 1);
702 wake_up(&reader_tasks[r].wq);
703 }
704
705 VERBOSE_SCALEOUT("main_func: experiment started, waiting for %d readers",
706 nreaders);
707
708 wait_event(main_wq,
709 !atomic_read(&nreaders_exp) || torture_must_stop());
710
711 VERBOSE_SCALEOUT("main_func: experiment ended");
712
713 if (torture_must_stop())
714 goto end;
715
716 result_avg[exp] = div_u64(1000 * process_durations(nreaders), nreaders * loops);
717 }
718
719 // Print the average of all experiments
720 SCALEOUT("END OF TEST. Calculating average duration per loop (nanoseconds)...\n");
721
722 pr_alert("Runs\tTime(ns)\n");
723 for (exp = 0; exp < nruns; exp++) {
724 u64 avg;
725 u32 rem;
726
727 avg = div_u64_rem(result_avg[exp], 1000, &rem);
728 sprintf(buf1, "%d\t%llu.%03u\n", exp + 1, avg, rem);
729 strcat(buf, buf1);
730 if (strlen(buf) >= 800) {
731 pr_alert("%s", buf);
732 buf[0] = 0;
733 }
734 }
735
736 pr_alert("%s", buf);
737
738oom_exit:
739 // This will shutdown everything including us.
740 if (shutdown) {
741 shutdown_start = 1;
742 wake_up(&shutdown_wq);
743 }
744
745 // Wait for torture to stop us
746 while (!torture_must_stop())
747 schedule_timeout_uninterruptible(1);
748
749end:
750 torture_kthread_stopping("main_func");
751 kfree(result_avg);
752 kfree(buf);
753 return 0;
754}
755
756static void
757ref_scale_print_module_parms(struct ref_scale_ops *cur_ops, const char *tag)
758{
759 pr_alert("%s" SCALE_FLAG
760 "--- %s: verbose=%d shutdown=%d holdoff=%d loops=%ld nreaders=%d nruns=%d readdelay=%d\n", scale_type, tag,
761 verbose, shutdown, holdoff, loops, nreaders, nruns, readdelay);
762}
763
764static void
765ref_scale_cleanup(void)
766{
767 int i;
768
769 if (torture_cleanup_begin())
770 return;
771
772 if (!cur_ops) {
773 torture_cleanup_end();
774 return;
775 }
776
777 if (reader_tasks) {
778 for (i = 0; i < nreaders; i++)
779 torture_stop_kthread("ref_scale_reader",
780 reader_tasks[i].task);
781 }
782 kfree(reader_tasks);
783
784 torture_stop_kthread("main_task", main_task);
785 kfree(main_task);
786
787 // Do scale-type-specific cleanup operations.
788 if (cur_ops->cleanup != NULL)
789 cur_ops->cleanup();
790
791 torture_cleanup_end();
792}
793
794// Shutdown kthread. Just waits to be awakened, then shuts down system.
795static int
796ref_scale_shutdown(void *arg)
797{
798 wait_event(shutdown_wq, shutdown_start);
799
800 smp_mb(); // Wake before output.
801 ref_scale_cleanup();
802 kernel_power_off();
803
804 return -EINVAL;
805}
806
807static int __init
808ref_scale_init(void)
809{
810 long i;
811 int firsterr = 0;
812 static struct ref_scale_ops *scale_ops[] = {
813 &rcu_ops, &srcu_ops, RCU_TRACE_OPS RCU_TASKS_OPS &refcnt_ops, &rwlock_ops,
814 &rwsem_ops, &lock_ops, &lock_irq_ops, &acqrel_ops, &clock_ops,
815 };
816
817 if (!torture_init_begin(scale_type, verbose))
818 return -EBUSY;
819
820 for (i = 0; i < ARRAY_SIZE(scale_ops); i++) {
821 cur_ops = scale_ops[i];
822 if (strcmp(scale_type, cur_ops->name) == 0)
823 break;
824 }
825 if (i == ARRAY_SIZE(scale_ops)) {
826 pr_alert("rcu-scale: invalid scale type: \"%s\"\n", scale_type);
827 pr_alert("rcu-scale types:");
828 for (i = 0; i < ARRAY_SIZE(scale_ops); i++)
829 pr_cont(" %s", scale_ops[i]->name);
830 pr_cont("\n");
831 firsterr = -EINVAL;
832 cur_ops = NULL;
833 goto unwind;
834 }
835 if (cur_ops->init)
836 cur_ops->init();
837
838 ref_scale_print_module_parms(cur_ops, "Start of test");
839
840 // Shutdown task
841 if (shutdown) {
842 init_waitqueue_head(&shutdown_wq);
843 firsterr = torture_create_kthread(ref_scale_shutdown, NULL,
844 shutdown_task);
845 if (torture_init_error(firsterr))
846 goto unwind;
847 schedule_timeout_uninterruptible(1);
848 }
849
850 // Reader tasks (default to ~75% of online CPUs).
851 if (nreaders < 0)
852 nreaders = (num_online_cpus() >> 1) + (num_online_cpus() >> 2);
853 if (WARN_ONCE(loops <= 0, "%s: loops = %ld, adjusted to 1\n", __func__, loops))
854 loops = 1;
855 if (WARN_ONCE(nreaders <= 0, "%s: nreaders = %d, adjusted to 1\n", __func__, nreaders))
856 nreaders = 1;
857 if (WARN_ONCE(nruns <= 0, "%s: nruns = %d, adjusted to 1\n", __func__, nruns))
858 nruns = 1;
859 reader_tasks = kcalloc(nreaders, sizeof(reader_tasks[0]),
860 GFP_KERNEL);
861 if (!reader_tasks) {
862 SCALEOUT_ERRSTRING("out of memory");
863 firsterr = -ENOMEM;
864 goto unwind;
865 }
866
867 VERBOSE_SCALEOUT("Starting %d reader threads", nreaders);
868
869 for (i = 0; i < nreaders; i++) {
870 firsterr = torture_create_kthread(ref_scale_reader, (void *)i,
871 reader_tasks[i].task);
872 if (torture_init_error(firsterr))
873 goto unwind;
874
875 init_waitqueue_head(&(reader_tasks[i].wq));
876 }
877
878 // Main Task
879 init_waitqueue_head(&main_wq);
880 firsterr = torture_create_kthread(main_func, NULL, main_task);
881 if (torture_init_error(firsterr))
882 goto unwind;
883
884 torture_init_end();
885 return 0;
886
887unwind:
888 torture_init_end();
889 ref_scale_cleanup();
890 if (shutdown) {
891 WARN_ON(!IS_MODULE(CONFIG_RCU_REF_SCALE_TEST));
892 kernel_power_off();
893 }
894 return firsterr;
895}
896
897module_init(ref_scale_init);
898module_exit(ref_scale_cleanup);