Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.15.
  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/mm.h>
 16#include <linux/module.h>
 17#include <linux/kthread.h>
 18#include <linux/err.h>
 19#include <linux/spinlock.h>
 20#include <linux/smp.h>
 21#include <linux/rcupdate.h>
 22#include <linux/interrupt.h>
 23#include <linux/sched.h>
 24#include <uapi/linux/sched/types.h>
 25#include <linux/atomic.h>
 26#include <linux/bitops.h>
 27#include <linux/completion.h>
 28#include <linux/moduleparam.h>
 29#include <linux/percpu.h>
 30#include <linux/notifier.h>
 31#include <linux/reboot.h>
 32#include <linux/freezer.h>
 33#include <linux/cpu.h>
 34#include <linux/delay.h>
 35#include <linux/stat.h>
 36#include <linux/srcu.h>
 37#include <linux/slab.h>
 38#include <asm/byteorder.h>
 39#include <linux/torture.h>
 40#include <linux/vmalloc.h>
 41
 42#include "rcu.h"
 43
 44MODULE_LICENSE("GPL");
 45MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
 46
 47#define PERF_FLAG "-perf:"
 48#define PERFOUT_STRING(s) \
 49	pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
 50#define VERBOSE_PERFOUT_STRING(s) \
 51	do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
 52#define VERBOSE_PERFOUT_ERRSTRING(s) \
 53	do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
 54
 55/*
 56 * The intended use cases for the nreaders and nwriters module parameters
 57 * are as follows:
 58 *
 59 * 1.	Specify only the nr_cpus kernel boot parameter.  This will
 60 *	set both nreaders and nwriters to the value specified by
 61 *	nr_cpus for a mixed reader/writer test.
 62 *
 63 * 2.	Specify the nr_cpus kernel boot parameter, but set
 64 *	rcuperf.nreaders to zero.  This will set nwriters to the
 65 *	value specified by nr_cpus for an update-only test.
 66 *
 67 * 3.	Specify the nr_cpus kernel boot parameter, but set
 68 *	rcuperf.nwriters to zero.  This will set nreaders to the
 69 *	value specified by nr_cpus for a read-only test.
 70 *
 71 * Various other use cases may of course be specified.
 72 *
 73 * Note that this test's readers are intended only as a test load for
 74 * the writers.  The reader performance statistics will be overly
 75 * pessimistic due to the per-critical-section interrupt disabling,
 76 * test-end checks, and the pair of calls through pointers.
 77 */
 78
 79#ifdef MODULE
 80# define RCUPERF_SHUTDOWN 0
 81#else
 82# define RCUPERF_SHUTDOWN 1
 83#endif
 84
 85torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
 86torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
 87torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
 88torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
 89torture_param(int, nreaders, -1, "Number of RCU reader threads");
 90torture_param(int, nwriters, -1, "Number of RCU updater threads");
 91torture_param(bool, shutdown, RCUPERF_SHUTDOWN,
 92	      "Shutdown at end of performance tests.");
 93torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
 94torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
 95torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() perf test?");
 96torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate.");
 97
 98static char *perf_type = "rcu";
 99module_param(perf_type, charp, 0444);
100MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, srcu, ...)");
101
102static int nrealreaders;
103static int nrealwriters;
104static struct task_struct **writer_tasks;
105static struct task_struct **reader_tasks;
106static struct task_struct *shutdown_task;
107
108static u64 **writer_durations;
109static int *writer_n_durations;
110static atomic_t n_rcu_perf_reader_started;
111static atomic_t n_rcu_perf_writer_started;
112static atomic_t n_rcu_perf_writer_finished;
113static wait_queue_head_t shutdown_wq;
114static u64 t_rcu_perf_writer_started;
115static u64 t_rcu_perf_writer_finished;
116static unsigned long b_rcu_gp_test_started;
117static unsigned long b_rcu_gp_test_finished;
118static DEFINE_PER_CPU(atomic_t, n_async_inflight);
119
120#define MAX_MEAS 10000
121#define MIN_MEAS 100
122
123/*
124 * Operations vector for selecting different types of tests.
125 */
126
127struct rcu_perf_ops {
128	int ptype;
129	void (*init)(void);
130	void (*cleanup)(void);
131	int (*readlock)(void);
132	void (*readunlock)(int idx);
133	unsigned long (*get_gp_seq)(void);
134	unsigned long (*gp_diff)(unsigned long new, unsigned long old);
135	unsigned long (*exp_completed)(void);
136	void (*async)(struct rcu_head *head, rcu_callback_t func);
137	void (*gp_barrier)(void);
138	void (*sync)(void);
139	void (*exp_sync)(void);
140	const char *name;
141};
142
143static struct rcu_perf_ops *cur_ops;
144
145/*
146 * Definitions for rcu perf testing.
147 */
148
149static int rcu_perf_read_lock(void) __acquires(RCU)
150{
151	rcu_read_lock();
152	return 0;
153}
154
155static void rcu_perf_read_unlock(int idx) __releases(RCU)
156{
157	rcu_read_unlock();
158}
159
160static unsigned long __maybe_unused rcu_no_completed(void)
161{
162	return 0;
163}
164
165static void rcu_sync_perf_init(void)
166{
167}
168
169static struct rcu_perf_ops rcu_ops = {
170	.ptype		= RCU_FLAVOR,
171	.init		= rcu_sync_perf_init,
172	.readlock	= rcu_perf_read_lock,
173	.readunlock	= rcu_perf_read_unlock,
174	.get_gp_seq	= rcu_get_gp_seq,
175	.gp_diff	= rcu_seq_diff,
176	.exp_completed	= rcu_exp_batches_completed,
177	.async		= call_rcu,
178	.gp_barrier	= rcu_barrier,
179	.sync		= synchronize_rcu,
180	.exp_sync	= synchronize_rcu_expedited,
181	.name		= "rcu"
182};
183
184/*
185 * Definitions for srcu perf testing.
186 */
187
188DEFINE_STATIC_SRCU(srcu_ctl_perf);
189static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
190
191static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
192{
193	return srcu_read_lock(srcu_ctlp);
194}
195
196static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
197{
198	srcu_read_unlock(srcu_ctlp, idx);
199}
200
201static unsigned long srcu_perf_completed(void)
202{
203	return srcu_batches_completed(srcu_ctlp);
204}
205
206static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
207{
208	call_srcu(srcu_ctlp, head, func);
209}
210
211static void srcu_rcu_barrier(void)
212{
213	srcu_barrier(srcu_ctlp);
214}
215
216static void srcu_perf_synchronize(void)
217{
218	synchronize_srcu(srcu_ctlp);
219}
220
221static void srcu_perf_synchronize_expedited(void)
222{
223	synchronize_srcu_expedited(srcu_ctlp);
224}
225
226static struct rcu_perf_ops srcu_ops = {
227	.ptype		= SRCU_FLAVOR,
228	.init		= rcu_sync_perf_init,
229	.readlock	= srcu_perf_read_lock,
230	.readunlock	= srcu_perf_read_unlock,
231	.get_gp_seq	= srcu_perf_completed,
232	.gp_diff	= rcu_seq_diff,
233	.exp_completed	= srcu_perf_completed,
234	.async		= srcu_call_rcu,
235	.gp_barrier	= srcu_rcu_barrier,
236	.sync		= srcu_perf_synchronize,
237	.exp_sync	= srcu_perf_synchronize_expedited,
238	.name		= "srcu"
239};
240
241static struct srcu_struct srcud;
242
243static void srcu_sync_perf_init(void)
244{
245	srcu_ctlp = &srcud;
246	init_srcu_struct(srcu_ctlp);
247}
248
249static void srcu_sync_perf_cleanup(void)
250{
251	cleanup_srcu_struct(srcu_ctlp);
252}
253
254static struct rcu_perf_ops srcud_ops = {
255	.ptype		= SRCU_FLAVOR,
256	.init		= srcu_sync_perf_init,
257	.cleanup	= srcu_sync_perf_cleanup,
258	.readlock	= srcu_perf_read_lock,
259	.readunlock	= srcu_perf_read_unlock,
260	.get_gp_seq	= srcu_perf_completed,
261	.gp_diff	= rcu_seq_diff,
262	.exp_completed	= srcu_perf_completed,
263	.async		= srcu_call_rcu,
264	.gp_barrier	= srcu_rcu_barrier,
265	.sync		= srcu_perf_synchronize,
266	.exp_sync	= srcu_perf_synchronize_expedited,
267	.name		= "srcud"
268};
269
270/*
271 * Definitions for RCU-tasks perf testing.
272 */
273
274static int tasks_perf_read_lock(void)
275{
276	return 0;
277}
278
279static void tasks_perf_read_unlock(int idx)
280{
281}
282
283static struct rcu_perf_ops tasks_ops = {
284	.ptype		= RCU_TASKS_FLAVOR,
285	.init		= rcu_sync_perf_init,
286	.readlock	= tasks_perf_read_lock,
287	.readunlock	= tasks_perf_read_unlock,
288	.get_gp_seq	= rcu_no_completed,
289	.gp_diff	= rcu_seq_diff,
290	.async		= call_rcu_tasks,
291	.gp_barrier	= rcu_barrier_tasks,
292	.sync		= synchronize_rcu_tasks,
293	.exp_sync	= synchronize_rcu_tasks,
294	.name		= "tasks"
295};
296
297static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old)
298{
299	if (!cur_ops->gp_diff)
300		return new - old;
301	return cur_ops->gp_diff(new, old);
302}
303
304/*
305 * If performance tests complete, wait for shutdown to commence.
306 */
307static void rcu_perf_wait_shutdown(void)
308{
309	cond_resched_tasks_rcu_qs();
310	if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
311		return;
312	while (!torture_must_stop())
313		schedule_timeout_uninterruptible(1);
314}
315
316/*
317 * RCU perf reader kthread.  Repeatedly does empty RCU read-side critical
318 * section, minimizing update-side interference.  However, the point of
319 * this test is not to evaluate reader performance, but instead to serve
320 * as a test load for update-side performance testing.
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 * Callback function for asynchronous grace periods from rcu_perf_writer().
347 */
348static void rcu_perf_async_cb(struct rcu_head *rhp)
349{
350	atomic_dec(this_cpu_ptr(&n_async_inflight));
351	kfree(rhp);
352}
353
354/*
355 * RCU perf writer kthread.  Repeatedly does a grace period.
356 */
357static int
358rcu_perf_writer(void *arg)
359{
360	int i = 0;
361	int i_max;
362	long me = (long)arg;
363	struct rcu_head *rhp = NULL;
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	sched_set_fifo_low(current);
373
374	if (holdoff)
375		schedule_timeout_uninterruptible(holdoff * HZ);
376
377	/*
378	 * Wait until rcu_end_inkernel_boot() is called for normal GP tests
379	 * so that RCU is not always expedited for normal GP tests.
380	 * The system_state test is approximate, but works well in practice.
381	 */
382	while (!gp_exp && system_state != SYSTEM_RUNNING)
383		schedule_timeout_uninterruptible(1);
384
385	t = ktime_get_mono_fast_ns();
386	if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
387		t_rcu_perf_writer_started = t;
388		if (gp_exp) {
389			b_rcu_gp_test_started =
390				cur_ops->exp_completed() / 2;
391		} else {
392			b_rcu_gp_test_started = cur_ops->get_gp_seq();
393		}
394	}
395
396	do {
397		if (writer_holdoff)
398			udelay(writer_holdoff);
399		wdp = &wdpp[i];
400		*wdp = ktime_get_mono_fast_ns();
401		if (gp_async) {
402retry:
403			if (!rhp)
404				rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
405			if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
406				atomic_inc(this_cpu_ptr(&n_async_inflight));
407				cur_ops->async(rhp, rcu_perf_async_cb);
408				rhp = NULL;
409			} else if (!kthread_should_stop()) {
410				cur_ops->gp_barrier();
411				goto retry;
412			} else {
413				kfree(rhp); /* Because we are stopping. */
414			}
415		} else if (gp_exp) {
416			cur_ops->exp_sync();
417		} else {
418			cur_ops->sync();
419		}
420		t = ktime_get_mono_fast_ns();
421		*wdp = t - *wdp;
422		i_max = i;
423		if (!started &&
424		    atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
425			started = true;
426		if (!done && i >= MIN_MEAS) {
427			done = true;
428			sched_set_normal(current, 0);
429			pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
430				 perf_type, PERF_FLAG, me, MIN_MEAS);
431			if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
432			    nrealwriters) {
433				schedule_timeout_interruptible(10);
434				rcu_ftrace_dump(DUMP_ALL);
435				PERFOUT_STRING("Test complete");
436				t_rcu_perf_writer_finished = t;
437				if (gp_exp) {
438					b_rcu_gp_test_finished =
439						cur_ops->exp_completed() / 2;
440				} else {
441					b_rcu_gp_test_finished =
442						cur_ops->get_gp_seq();
443				}
444				if (shutdown) {
445					smp_mb(); /* Assign before wake. */
446					wake_up(&shutdown_wq);
447				}
448			}
449		}
450		if (done && !alldone &&
451		    atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
452			alldone = true;
453		if (started && !alldone && i < MAX_MEAS - 1)
454			i++;
455		rcu_perf_wait_shutdown();
456	} while (!torture_must_stop());
457	if (gp_async) {
458		cur_ops->gp_barrier();
459	}
460	writer_n_durations[me] = i_max;
461	torture_kthread_stopping("rcu_perf_writer");
462	return 0;
463}
464
465static void
466rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
467{
468	pr_alert("%s" PERF_FLAG
469		 "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
470		 perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
471}
472
473static void
474rcu_perf_cleanup(void)
475{
476	int i;
477	int j;
478	int ngps = 0;
479	u64 *wdp;
480	u64 *wdpp;
481
482	/*
483	 * Would like warning at start, but everything is expedited
484	 * during the mid-boot phase, so have to wait till the end.
485	 */
486	if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
487		VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
488	if (rcu_gp_is_normal() && gp_exp)
489		VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
490	if (gp_exp && gp_async)
491		VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
492
493	if (torture_cleanup_begin())
494		return;
495	if (!cur_ops) {
496		torture_cleanup_end();
497		return;
498	}
499
500	if (reader_tasks) {
501		for (i = 0; i < nrealreaders; i++)
502			torture_stop_kthread(rcu_perf_reader,
503					     reader_tasks[i]);
504		kfree(reader_tasks);
505	}
506
507	if (writer_tasks) {
508		for (i = 0; i < nrealwriters; i++) {
509			torture_stop_kthread(rcu_perf_writer,
510					     writer_tasks[i]);
511			if (!writer_n_durations)
512				continue;
513			j = writer_n_durations[i];
514			pr_alert("%s%s writer %d gps: %d\n",
515				 perf_type, PERF_FLAG, i, j);
516			ngps += j;
517		}
518		pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
519			 perf_type, PERF_FLAG,
520			 t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
521			 t_rcu_perf_writer_finished -
522			 t_rcu_perf_writer_started,
523			 ngps,
524			 rcuperf_seq_diff(b_rcu_gp_test_finished,
525					  b_rcu_gp_test_started));
526		for (i = 0; i < nrealwriters; i++) {
527			if (!writer_durations)
528				break;
529			if (!writer_n_durations)
530				continue;
531			wdpp = writer_durations[i];
532			if (!wdpp)
533				continue;
534			for (j = 0; j <= writer_n_durations[i]; j++) {
535				wdp = &wdpp[j];
536				pr_alert("%s%s %4d writer-duration: %5d %llu\n",
537					perf_type, PERF_FLAG,
538					i, j, *wdp);
539				if (j % 100 == 0)
540					schedule_timeout_uninterruptible(1);
541			}
542			kfree(writer_durations[i]);
543		}
544		kfree(writer_tasks);
545		kfree(writer_durations);
546		kfree(writer_n_durations);
547	}
548
549	/* Do torture-type-specific cleanup operations.  */
550	if (cur_ops->cleanup != NULL)
551		cur_ops->cleanup();
552
553	torture_cleanup_end();
554}
555
556/*
557 * Return the number if non-negative.  If -1, the number of CPUs.
558 * If less than -1, that much less than the number of CPUs, but
559 * at least one.
560 */
561static int compute_real(int n)
562{
563	int nr;
564
565	if (n >= 0) {
566		nr = n;
567	} else {
568		nr = num_online_cpus() + 1 + n;
569		if (nr <= 0)
570			nr = 1;
571	}
572	return nr;
573}
574
575/*
576 * RCU perf shutdown kthread.  Just waits to be awakened, then shuts
577 * down system.
578 */
579static int
580rcu_perf_shutdown(void *arg)
581{
582	wait_event(shutdown_wq,
583		   atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters);
584	smp_mb(); /* Wake before output. */
585	rcu_perf_cleanup();
586	kernel_power_off();
587	return -EINVAL;
588}
589
590/*
591 * kfree_rcu() performance tests: Start a kfree_rcu() loop on all CPUs for number
592 * of iterations and measure total time and number of GP for all iterations to complete.
593 */
594
595torture_param(int, kfree_nthreads, -1, "Number of threads running loops of kfree_rcu().");
596torture_param(int, kfree_alloc_num, 8000, "Number of allocations and frees done in an iteration.");
597torture_param(int, kfree_loops, 10, "Number of loops doing kfree_alloc_num allocations and frees.");
598
599static struct task_struct **kfree_reader_tasks;
600static int kfree_nrealthreads;
601static atomic_t n_kfree_perf_thread_started;
602static atomic_t n_kfree_perf_thread_ended;
603
604struct kfree_obj {
605	char kfree_obj[8];
606	struct rcu_head rh;
607};
608
609static int
610kfree_perf_thread(void *arg)
611{
612	int i, loop = 0;
613	long me = (long)arg;
614	struct kfree_obj *alloc_ptr;
615	u64 start_time, end_time;
616	long long mem_begin, mem_during = 0;
617
618	VERBOSE_PERFOUT_STRING("kfree_perf_thread task started");
619	set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
620	set_user_nice(current, MAX_NICE);
621
622	start_time = ktime_get_mono_fast_ns();
623
624	if (atomic_inc_return(&n_kfree_perf_thread_started) >= kfree_nrealthreads) {
625		if (gp_exp)
626			b_rcu_gp_test_started = cur_ops->exp_completed() / 2;
627		else
628			b_rcu_gp_test_started = cur_ops->get_gp_seq();
629	}
630
631	do {
632		if (!mem_during) {
633			mem_during = mem_begin = si_mem_available();
634		} else if (loop % (kfree_loops / 4) == 0) {
635			mem_during = (mem_during + si_mem_available()) / 2;
636		}
637
638		for (i = 0; i < kfree_alloc_num; i++) {
639			alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL);
640			if (!alloc_ptr)
641				return -ENOMEM;
642
643			kfree_rcu(alloc_ptr, rh);
644		}
645
646		cond_resched();
647	} while (!torture_must_stop() && ++loop < kfree_loops);
648
649	if (atomic_inc_return(&n_kfree_perf_thread_ended) >= kfree_nrealthreads) {
650		end_time = ktime_get_mono_fast_ns();
651
652		if (gp_exp)
653			b_rcu_gp_test_finished = cur_ops->exp_completed() / 2;
654		else
655			b_rcu_gp_test_finished = cur_ops->get_gp_seq();
656
657		pr_alert("Total time taken by all kfree'ers: %llu ns, loops: %d, batches: %ld, memory footprint: %lldMB\n",
658		       (unsigned long long)(end_time - start_time), kfree_loops,
659		       rcuperf_seq_diff(b_rcu_gp_test_finished, b_rcu_gp_test_started),
660		       (mem_begin - mem_during) >> (20 - PAGE_SHIFT));
661
662		if (shutdown) {
663			smp_mb(); /* Assign before wake. */
664			wake_up(&shutdown_wq);
665		}
666	}
667
668	torture_kthread_stopping("kfree_perf_thread");
669	return 0;
670}
671
672static void
673kfree_perf_cleanup(void)
674{
675	int i;
676
677	if (torture_cleanup_begin())
678		return;
679
680	if (kfree_reader_tasks) {
681		for (i = 0; i < kfree_nrealthreads; i++)
682			torture_stop_kthread(kfree_perf_thread,
683					     kfree_reader_tasks[i]);
684		kfree(kfree_reader_tasks);
685	}
686
687	torture_cleanup_end();
688}
689
690/*
691 * shutdown kthread.  Just waits to be awakened, then shuts down system.
692 */
693static int
694kfree_perf_shutdown(void *arg)
695{
696	wait_event(shutdown_wq,
697		   atomic_read(&n_kfree_perf_thread_ended) >= kfree_nrealthreads);
698
699	smp_mb(); /* Wake before output. */
700
701	kfree_perf_cleanup();
702	kernel_power_off();
703	return -EINVAL;
704}
705
706static int __init
707kfree_perf_init(void)
708{
709	long i;
710	int firsterr = 0;
711
712	kfree_nrealthreads = compute_real(kfree_nthreads);
713	/* Start up the kthreads. */
714	if (shutdown) {
715		init_waitqueue_head(&shutdown_wq);
716		firsterr = torture_create_kthread(kfree_perf_shutdown, NULL,
717						  shutdown_task);
718		if (firsterr)
719			goto unwind;
720		schedule_timeout_uninterruptible(1);
721	}
722
723	pr_alert("kfree object size=%zu\n", kfree_mult * sizeof(struct kfree_obj));
724
725	kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]),
726			       GFP_KERNEL);
727	if (kfree_reader_tasks == NULL) {
728		firsterr = -ENOMEM;
729		goto unwind;
730	}
731
732	for (i = 0; i < kfree_nrealthreads; i++) {
733		firsterr = torture_create_kthread(kfree_perf_thread, (void *)i,
734						  kfree_reader_tasks[i]);
735		if (firsterr)
736			goto unwind;
737	}
738
739	while (atomic_read(&n_kfree_perf_thread_started) < kfree_nrealthreads)
740		schedule_timeout_uninterruptible(1);
741
742	torture_init_end();
743	return 0;
744
745unwind:
746	torture_init_end();
747	kfree_perf_cleanup();
748	return firsterr;
749}
750
751static int __init
752rcu_perf_init(void)
753{
754	long i;
755	int firsterr = 0;
756	static struct rcu_perf_ops *perf_ops[] = {
757		&rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops,
758	};
759
760	if (!torture_init_begin(perf_type, verbose))
761		return -EBUSY;
762
763	/* Process args and tell the world that the perf'er is on the job. */
764	for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
765		cur_ops = perf_ops[i];
766		if (strcmp(perf_type, cur_ops->name) == 0)
767			break;
768	}
769	if (i == ARRAY_SIZE(perf_ops)) {
770		pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type);
771		pr_alert("rcu-perf types:");
772		for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
773			pr_cont(" %s", perf_ops[i]->name);
774		pr_cont("\n");
775		WARN_ON(!IS_MODULE(CONFIG_RCU_PERF_TEST));
776		firsterr = -EINVAL;
777		cur_ops = NULL;
778		goto unwind;
779	}
780	if (cur_ops->init)
781		cur_ops->init();
782
783	if (kfree_rcu_test)
784		return kfree_perf_init();
785
786	nrealwriters = compute_real(nwriters);
787	nrealreaders = compute_real(nreaders);
788	atomic_set(&n_rcu_perf_reader_started, 0);
789	atomic_set(&n_rcu_perf_writer_started, 0);
790	atomic_set(&n_rcu_perf_writer_finished, 0);
791	rcu_perf_print_module_parms(cur_ops, "Start of test");
792
793	/* Start up the kthreads. */
794
795	if (shutdown) {
796		init_waitqueue_head(&shutdown_wq);
797		firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
798						  shutdown_task);
799		if (firsterr)
800			goto unwind;
801		schedule_timeout_uninterruptible(1);
802	}
803	reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
804			       GFP_KERNEL);
805	if (reader_tasks == NULL) {
806		VERBOSE_PERFOUT_ERRSTRING("out of memory");
807		firsterr = -ENOMEM;
808		goto unwind;
809	}
810	for (i = 0; i < nrealreaders; i++) {
811		firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
812						  reader_tasks[i]);
813		if (firsterr)
814			goto unwind;
815	}
816	while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
817		schedule_timeout_uninterruptible(1);
818	writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
819			       GFP_KERNEL);
820	writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
821				   GFP_KERNEL);
822	writer_n_durations =
823		kcalloc(nrealwriters, sizeof(*writer_n_durations),
824			GFP_KERNEL);
825	if (!writer_tasks || !writer_durations || !writer_n_durations) {
826		VERBOSE_PERFOUT_ERRSTRING("out of memory");
827		firsterr = -ENOMEM;
828		goto unwind;
829	}
830	for (i = 0; i < nrealwriters; i++) {
831		writer_durations[i] =
832			kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
833				GFP_KERNEL);
834		if (!writer_durations[i]) {
835			firsterr = -ENOMEM;
836			goto unwind;
837		}
838		firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
839						  writer_tasks[i]);
840		if (firsterr)
841			goto unwind;
842	}
843	torture_init_end();
844	return 0;
845
846unwind:
847	torture_init_end();
848	rcu_perf_cleanup();
849	return firsterr;
850}
851
852module_init(rcu_perf_init);
853module_exit(rcu_perf_cleanup);