xref: /openbmc/linux/kernel/rcu/rcuscale.c (revision 36de991e)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Read-Copy Update module-based scalability-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 #include <linux/rcupdate_trace.h>
42 
43 #include "rcu.h"
44 
45 MODULE_LICENSE("GPL");
46 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
47 
48 #define SCALE_FLAG "-scale:"
49 #define SCALEOUT_STRING(s) \
50 	pr_alert("%s" SCALE_FLAG " %s\n", scale_type, s)
51 #define VERBOSE_SCALEOUT_STRING(s) \
52 	do { if (verbose) pr_alert("%s" SCALE_FLAG " %s\n", scale_type, s); } while (0)
53 #define VERBOSE_SCALEOUT_ERRSTRING(s) \
54 	do { if (verbose) pr_alert("%s" SCALE_FLAG "!!! %s\n", scale_type, s); } while (0)
55 
56 /*
57  * The intended use cases for the nreaders and nwriters module parameters
58  * are as follows:
59  *
60  * 1.	Specify only the nr_cpus kernel boot parameter.  This will
61  *	set both nreaders and nwriters to the value specified by
62  *	nr_cpus for a mixed reader/writer test.
63  *
64  * 2.	Specify the nr_cpus kernel boot parameter, but set
65  *	rcuscale.nreaders to zero.  This will set nwriters to the
66  *	value specified by nr_cpus for an update-only test.
67  *
68  * 3.	Specify the nr_cpus kernel boot parameter, but set
69  *	rcuscale.nwriters to zero.  This will set nreaders to the
70  *	value specified by nr_cpus for a read-only test.
71  *
72  * Various other use cases may of course be specified.
73  *
74  * Note that this test's readers are intended only as a test load for
75  * the writers.  The reader scalability statistics will be overly
76  * pessimistic due to the per-critical-section interrupt disabling,
77  * test-end checks, and the pair of calls through pointers.
78  */
79 
80 #ifdef MODULE
81 # define RCUSCALE_SHUTDOWN 0
82 #else
83 # define RCUSCALE_SHUTDOWN 1
84 #endif
85 
86 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
87 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
88 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
89 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
90 torture_param(int, nreaders, -1, "Number of RCU reader threads");
91 torture_param(int, nwriters, -1, "Number of RCU updater threads");
92 torture_param(bool, shutdown, RCUSCALE_SHUTDOWN,
93 	      "Shutdown at end of scalability tests.");
94 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
95 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
96 torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() scale test?");
97 torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate.");
98 
99 static char *scale_type = "rcu";
100 module_param(scale_type, charp, 0444);
101 MODULE_PARM_DESC(scale_type, "Type of RCU to scalability-test (rcu, srcu, ...)");
102 
103 static int nrealreaders;
104 static int nrealwriters;
105 static struct task_struct **writer_tasks;
106 static struct task_struct **reader_tasks;
107 static struct task_struct *shutdown_task;
108 
109 static u64 **writer_durations;
110 static int *writer_n_durations;
111 static atomic_t n_rcu_scale_reader_started;
112 static atomic_t n_rcu_scale_writer_started;
113 static atomic_t n_rcu_scale_writer_finished;
114 static wait_queue_head_t shutdown_wq;
115 static u64 t_rcu_scale_writer_started;
116 static u64 t_rcu_scale_writer_finished;
117 static unsigned long b_rcu_gp_test_started;
118 static unsigned long b_rcu_gp_test_finished;
119 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
120 
121 #define MAX_MEAS 10000
122 #define MIN_MEAS 100
123 
124 /*
125  * Operations vector for selecting different types of tests.
126  */
127 
128 struct rcu_scale_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_scale_ops *cur_ops;
145 
146 /*
147  * Definitions for rcu scalability testing.
148  */
149 
150 static int rcu_scale_read_lock(void) __acquires(RCU)
151 {
152 	rcu_read_lock();
153 	return 0;
154 }
155 
156 static void rcu_scale_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_scale_init(void)
167 {
168 }
169 
170 static struct rcu_scale_ops rcu_ops = {
171 	.ptype		= RCU_FLAVOR,
172 	.init		= rcu_sync_scale_init,
173 	.readlock	= rcu_scale_read_lock,
174 	.readunlock	= rcu_scale_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 scalability testing.
187  */
188 
189 DEFINE_STATIC_SRCU(srcu_ctl_scale);
190 static struct srcu_struct *srcu_ctlp = &srcu_ctl_scale;
191 
192 static int srcu_scale_read_lock(void) __acquires(srcu_ctlp)
193 {
194 	return srcu_read_lock(srcu_ctlp);
195 }
196 
197 static void srcu_scale_read_unlock(int idx) __releases(srcu_ctlp)
198 {
199 	srcu_read_unlock(srcu_ctlp, idx);
200 }
201 
202 static unsigned long srcu_scale_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_scale_synchronize(void)
218 {
219 	synchronize_srcu(srcu_ctlp);
220 }
221 
222 static void srcu_scale_synchronize_expedited(void)
223 {
224 	synchronize_srcu_expedited(srcu_ctlp);
225 }
226 
227 static struct rcu_scale_ops srcu_ops = {
228 	.ptype		= SRCU_FLAVOR,
229 	.init		= rcu_sync_scale_init,
230 	.readlock	= srcu_scale_read_lock,
231 	.readunlock	= srcu_scale_read_unlock,
232 	.get_gp_seq	= srcu_scale_completed,
233 	.gp_diff	= rcu_seq_diff,
234 	.exp_completed	= srcu_scale_completed,
235 	.async		= srcu_call_rcu,
236 	.gp_barrier	= srcu_rcu_barrier,
237 	.sync		= srcu_scale_synchronize,
238 	.exp_sync	= srcu_scale_synchronize_expedited,
239 	.name		= "srcu"
240 };
241 
242 static struct srcu_struct srcud;
243 
244 static void srcu_sync_scale_init(void)
245 {
246 	srcu_ctlp = &srcud;
247 	init_srcu_struct(srcu_ctlp);
248 }
249 
250 static void srcu_sync_scale_cleanup(void)
251 {
252 	cleanup_srcu_struct(srcu_ctlp);
253 }
254 
255 static struct rcu_scale_ops srcud_ops = {
256 	.ptype		= SRCU_FLAVOR,
257 	.init		= srcu_sync_scale_init,
258 	.cleanup	= srcu_sync_scale_cleanup,
259 	.readlock	= srcu_scale_read_lock,
260 	.readunlock	= srcu_scale_read_unlock,
261 	.get_gp_seq	= srcu_scale_completed,
262 	.gp_diff	= rcu_seq_diff,
263 	.exp_completed	= srcu_scale_completed,
264 	.async		= srcu_call_rcu,
265 	.gp_barrier	= srcu_rcu_barrier,
266 	.sync		= srcu_scale_synchronize,
267 	.exp_sync	= srcu_scale_synchronize_expedited,
268 	.name		= "srcud"
269 };
270 
271 /*
272  * Definitions for RCU-tasks scalability testing.
273  */
274 
275 static int tasks_scale_read_lock(void)
276 {
277 	return 0;
278 }
279 
280 static void tasks_scale_read_unlock(int idx)
281 {
282 }
283 
284 static struct rcu_scale_ops tasks_ops = {
285 	.ptype		= RCU_TASKS_FLAVOR,
286 	.init		= rcu_sync_scale_init,
287 	.readlock	= tasks_scale_read_lock,
288 	.readunlock	= tasks_scale_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 /*
299  * Definitions for RCU-tasks-trace scalability testing.
300  */
301 
302 static int tasks_trace_scale_read_lock(void)
303 {
304 	rcu_read_lock_trace();
305 	return 0;
306 }
307 
308 static void tasks_trace_scale_read_unlock(int idx)
309 {
310 	rcu_read_unlock_trace();
311 }
312 
313 static struct rcu_scale_ops tasks_tracing_ops = {
314 	.ptype		= RCU_TASKS_FLAVOR,
315 	.init		= rcu_sync_scale_init,
316 	.readlock	= tasks_trace_scale_read_lock,
317 	.readunlock	= tasks_trace_scale_read_unlock,
318 	.get_gp_seq	= rcu_no_completed,
319 	.gp_diff	= rcu_seq_diff,
320 	.async		= call_rcu_tasks_trace,
321 	.gp_barrier	= rcu_barrier_tasks_trace,
322 	.sync		= synchronize_rcu_tasks_trace,
323 	.exp_sync	= synchronize_rcu_tasks_trace,
324 	.name		= "tasks-tracing"
325 };
326 
327 static unsigned long rcuscale_seq_diff(unsigned long new, unsigned long old)
328 {
329 	if (!cur_ops->gp_diff)
330 		return new - old;
331 	return cur_ops->gp_diff(new, old);
332 }
333 
334 /*
335  * If scalability tests complete, wait for shutdown to commence.
336  */
337 static void rcu_scale_wait_shutdown(void)
338 {
339 	cond_resched_tasks_rcu_qs();
340 	if (atomic_read(&n_rcu_scale_writer_finished) < nrealwriters)
341 		return;
342 	while (!torture_must_stop())
343 		schedule_timeout_uninterruptible(1);
344 }
345 
346 /*
347  * RCU scalability reader kthread.  Repeatedly does empty RCU read-side
348  * critical section, minimizing update-side interference.  However, the
349  * point of this test is not to evaluate reader scalability, but instead
350  * to serve as a test load for update-side scalability testing.
351  */
352 static int
353 rcu_scale_reader(void *arg)
354 {
355 	unsigned long flags;
356 	int idx;
357 	long me = (long)arg;
358 
359 	VERBOSE_SCALEOUT_STRING("rcu_scale_reader task started");
360 	set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
361 	set_user_nice(current, MAX_NICE);
362 	atomic_inc(&n_rcu_scale_reader_started);
363 
364 	do {
365 		local_irq_save(flags);
366 		idx = cur_ops->readlock();
367 		cur_ops->readunlock(idx);
368 		local_irq_restore(flags);
369 		rcu_scale_wait_shutdown();
370 	} while (!torture_must_stop());
371 	torture_kthread_stopping("rcu_scale_reader");
372 	return 0;
373 }
374 
375 /*
376  * Callback function for asynchronous grace periods from rcu_scale_writer().
377  */
378 static void rcu_scale_async_cb(struct rcu_head *rhp)
379 {
380 	atomic_dec(this_cpu_ptr(&n_async_inflight));
381 	kfree(rhp);
382 }
383 
384 /*
385  * RCU scale writer kthread.  Repeatedly does a grace period.
386  */
387 static int
388 rcu_scale_writer(void *arg)
389 {
390 	int i = 0;
391 	int i_max;
392 	long me = (long)arg;
393 	struct rcu_head *rhp = NULL;
394 	bool started = false, done = false, alldone = false;
395 	u64 t;
396 	u64 *wdp;
397 	u64 *wdpp = writer_durations[me];
398 
399 	VERBOSE_SCALEOUT_STRING("rcu_scale_writer task started");
400 	WARN_ON(!wdpp);
401 	set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
402 	sched_set_fifo_low(current);
403 
404 	if (holdoff)
405 		schedule_timeout_uninterruptible(holdoff * HZ);
406 
407 	/*
408 	 * Wait until rcu_end_inkernel_boot() is called for normal GP tests
409 	 * so that RCU is not always expedited for normal GP tests.
410 	 * The system_state test is approximate, but works well in practice.
411 	 */
412 	while (!gp_exp && system_state != SYSTEM_RUNNING)
413 		schedule_timeout_uninterruptible(1);
414 
415 	t = ktime_get_mono_fast_ns();
416 	if (atomic_inc_return(&n_rcu_scale_writer_started) >= nrealwriters) {
417 		t_rcu_scale_writer_started = t;
418 		if (gp_exp) {
419 			b_rcu_gp_test_started =
420 				cur_ops->exp_completed() / 2;
421 		} else {
422 			b_rcu_gp_test_started = cur_ops->get_gp_seq();
423 		}
424 	}
425 
426 	do {
427 		if (writer_holdoff)
428 			udelay(writer_holdoff);
429 		wdp = &wdpp[i];
430 		*wdp = ktime_get_mono_fast_ns();
431 		if (gp_async) {
432 retry:
433 			if (!rhp)
434 				rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
435 			if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
436 				atomic_inc(this_cpu_ptr(&n_async_inflight));
437 				cur_ops->async(rhp, rcu_scale_async_cb);
438 				rhp = NULL;
439 			} else if (!kthread_should_stop()) {
440 				cur_ops->gp_barrier();
441 				goto retry;
442 			} else {
443 				kfree(rhp); /* Because we are stopping. */
444 			}
445 		} else if (gp_exp) {
446 			cur_ops->exp_sync();
447 		} else {
448 			cur_ops->sync();
449 		}
450 		t = ktime_get_mono_fast_ns();
451 		*wdp = t - *wdp;
452 		i_max = i;
453 		if (!started &&
454 		    atomic_read(&n_rcu_scale_writer_started) >= nrealwriters)
455 			started = true;
456 		if (!done && i >= MIN_MEAS) {
457 			done = true;
458 			sched_set_normal(current, 0);
459 			pr_alert("%s%s rcu_scale_writer %ld has %d measurements\n",
460 				 scale_type, SCALE_FLAG, me, MIN_MEAS);
461 			if (atomic_inc_return(&n_rcu_scale_writer_finished) >=
462 			    nrealwriters) {
463 				schedule_timeout_interruptible(10);
464 				rcu_ftrace_dump(DUMP_ALL);
465 				SCALEOUT_STRING("Test complete");
466 				t_rcu_scale_writer_finished = t;
467 				if (gp_exp) {
468 					b_rcu_gp_test_finished =
469 						cur_ops->exp_completed() / 2;
470 				} else {
471 					b_rcu_gp_test_finished =
472 						cur_ops->get_gp_seq();
473 				}
474 				if (shutdown) {
475 					smp_mb(); /* Assign before wake. */
476 					wake_up(&shutdown_wq);
477 				}
478 			}
479 		}
480 		if (done && !alldone &&
481 		    atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters)
482 			alldone = true;
483 		if (started && !alldone && i < MAX_MEAS - 1)
484 			i++;
485 		rcu_scale_wait_shutdown();
486 	} while (!torture_must_stop());
487 	if (gp_async) {
488 		cur_ops->gp_barrier();
489 	}
490 	writer_n_durations[me] = i_max + 1;
491 	torture_kthread_stopping("rcu_scale_writer");
492 	return 0;
493 }
494 
495 static void
496 rcu_scale_print_module_parms(struct rcu_scale_ops *cur_ops, const char *tag)
497 {
498 	pr_alert("%s" SCALE_FLAG
499 		 "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
500 		 scale_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
501 }
502 
503 static void
504 rcu_scale_cleanup(void)
505 {
506 	int i;
507 	int j;
508 	int ngps = 0;
509 	u64 *wdp;
510 	u64 *wdpp;
511 
512 	/*
513 	 * Would like warning at start, but everything is expedited
514 	 * during the mid-boot phase, so have to wait till the end.
515 	 */
516 	if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
517 		VERBOSE_SCALEOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
518 	if (rcu_gp_is_normal() && gp_exp)
519 		VERBOSE_SCALEOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
520 	if (gp_exp && gp_async)
521 		VERBOSE_SCALEOUT_ERRSTRING("No expedited async GPs, so went with async!");
522 
523 	if (torture_cleanup_begin())
524 		return;
525 	if (!cur_ops) {
526 		torture_cleanup_end();
527 		return;
528 	}
529 
530 	if (reader_tasks) {
531 		for (i = 0; i < nrealreaders; i++)
532 			torture_stop_kthread(rcu_scale_reader,
533 					     reader_tasks[i]);
534 		kfree(reader_tasks);
535 	}
536 
537 	if (writer_tasks) {
538 		for (i = 0; i < nrealwriters; i++) {
539 			torture_stop_kthread(rcu_scale_writer,
540 					     writer_tasks[i]);
541 			if (!writer_n_durations)
542 				continue;
543 			j = writer_n_durations[i];
544 			pr_alert("%s%s writer %d gps: %d\n",
545 				 scale_type, SCALE_FLAG, i, j);
546 			ngps += j;
547 		}
548 		pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
549 			 scale_type, SCALE_FLAG,
550 			 t_rcu_scale_writer_started, t_rcu_scale_writer_finished,
551 			 t_rcu_scale_writer_finished -
552 			 t_rcu_scale_writer_started,
553 			 ngps,
554 			 rcuscale_seq_diff(b_rcu_gp_test_finished,
555 					   b_rcu_gp_test_started));
556 		for (i = 0; i < nrealwriters; i++) {
557 			if (!writer_durations)
558 				break;
559 			if (!writer_n_durations)
560 				continue;
561 			wdpp = writer_durations[i];
562 			if (!wdpp)
563 				continue;
564 			for (j = 0; j < writer_n_durations[i]; j++) {
565 				wdp = &wdpp[j];
566 				pr_alert("%s%s %4d writer-duration: %5d %llu\n",
567 					scale_type, SCALE_FLAG,
568 					i, j, *wdp);
569 				if (j % 100 == 0)
570 					schedule_timeout_uninterruptible(1);
571 			}
572 			kfree(writer_durations[i]);
573 		}
574 		kfree(writer_tasks);
575 		kfree(writer_durations);
576 		kfree(writer_n_durations);
577 	}
578 
579 	/* Do torture-type-specific cleanup operations.  */
580 	if (cur_ops->cleanup != NULL)
581 		cur_ops->cleanup();
582 
583 	torture_cleanup_end();
584 }
585 
586 /*
587  * Return the number if non-negative.  If -1, the number of CPUs.
588  * If less than -1, that much less than the number of CPUs, but
589  * at least one.
590  */
591 static int compute_real(int n)
592 {
593 	int nr;
594 
595 	if (n >= 0) {
596 		nr = n;
597 	} else {
598 		nr = num_online_cpus() + 1 + n;
599 		if (nr <= 0)
600 			nr = 1;
601 	}
602 	return nr;
603 }
604 
605 /*
606  * RCU scalability shutdown kthread.  Just waits to be awakened, then shuts
607  * down system.
608  */
609 static int
610 rcu_scale_shutdown(void *arg)
611 {
612 	wait_event(shutdown_wq,
613 		   atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters);
614 	smp_mb(); /* Wake before output. */
615 	rcu_scale_cleanup();
616 	kernel_power_off();
617 	return -EINVAL;
618 }
619 
620 /*
621  * kfree_rcu() scalability tests: Start a kfree_rcu() loop on all CPUs for number
622  * of iterations and measure total time and number of GP for all iterations to complete.
623  */
624 
625 torture_param(int, kfree_nthreads, -1, "Number of threads running loops of kfree_rcu().");
626 torture_param(int, kfree_alloc_num, 8000, "Number of allocations and frees done in an iteration.");
627 torture_param(int, kfree_loops, 10, "Number of loops doing kfree_alloc_num allocations and frees.");
628 torture_param(bool, kfree_rcu_test_double, false, "Do we run a kfree_rcu() double-argument scale test?");
629 torture_param(bool, kfree_rcu_test_single, false, "Do we run a kfree_rcu() single-argument scale test?");
630 
631 static struct task_struct **kfree_reader_tasks;
632 static int kfree_nrealthreads;
633 static atomic_t n_kfree_scale_thread_started;
634 static atomic_t n_kfree_scale_thread_ended;
635 
636 struct kfree_obj {
637 	char kfree_obj[8];
638 	struct rcu_head rh;
639 };
640 
641 static int
642 kfree_scale_thread(void *arg)
643 {
644 	int i, loop = 0;
645 	long me = (long)arg;
646 	struct kfree_obj *alloc_ptr;
647 	u64 start_time, end_time;
648 	long long mem_begin, mem_during = 0;
649 	bool kfree_rcu_test_both;
650 	DEFINE_TORTURE_RANDOM(tr);
651 
652 	VERBOSE_SCALEOUT_STRING("kfree_scale_thread task started");
653 	set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
654 	set_user_nice(current, MAX_NICE);
655 	kfree_rcu_test_both = (kfree_rcu_test_single == kfree_rcu_test_double);
656 
657 	start_time = ktime_get_mono_fast_ns();
658 
659 	if (atomic_inc_return(&n_kfree_scale_thread_started) >= kfree_nrealthreads) {
660 		if (gp_exp)
661 			b_rcu_gp_test_started = cur_ops->exp_completed() / 2;
662 		else
663 			b_rcu_gp_test_started = cur_ops->get_gp_seq();
664 	}
665 
666 	do {
667 		if (!mem_during) {
668 			mem_during = mem_begin = si_mem_available();
669 		} else if (loop % (kfree_loops / 4) == 0) {
670 			mem_during = (mem_during + si_mem_available()) / 2;
671 		}
672 
673 		for (i = 0; i < kfree_alloc_num; i++) {
674 			alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL);
675 			if (!alloc_ptr)
676 				return -ENOMEM;
677 
678 			// By default kfree_rcu_test_single and kfree_rcu_test_double are
679 			// initialized to false. If both have the same value (false or true)
680 			// both are randomly tested, otherwise only the one with value true
681 			// is tested.
682 			if ((kfree_rcu_test_single && !kfree_rcu_test_double) ||
683 					(kfree_rcu_test_both && torture_random(&tr) & 0x800))
684 				kfree_rcu(alloc_ptr);
685 			else
686 				kfree_rcu(alloc_ptr, rh);
687 		}
688 
689 		cond_resched();
690 	} while (!torture_must_stop() && ++loop < kfree_loops);
691 
692 	if (atomic_inc_return(&n_kfree_scale_thread_ended) >= kfree_nrealthreads) {
693 		end_time = ktime_get_mono_fast_ns();
694 
695 		if (gp_exp)
696 			b_rcu_gp_test_finished = cur_ops->exp_completed() / 2;
697 		else
698 			b_rcu_gp_test_finished = cur_ops->get_gp_seq();
699 
700 		pr_alert("Total time taken by all kfree'ers: %llu ns, loops: %d, batches: %ld, memory footprint: %lldMB\n",
701 		       (unsigned long long)(end_time - start_time), kfree_loops,
702 		       rcuscale_seq_diff(b_rcu_gp_test_finished, b_rcu_gp_test_started),
703 		       (mem_begin - mem_during) >> (20 - PAGE_SHIFT));
704 
705 		if (shutdown) {
706 			smp_mb(); /* Assign before wake. */
707 			wake_up(&shutdown_wq);
708 		}
709 	}
710 
711 	torture_kthread_stopping("kfree_scale_thread");
712 	return 0;
713 }
714 
715 static void
716 kfree_scale_cleanup(void)
717 {
718 	int i;
719 
720 	if (torture_cleanup_begin())
721 		return;
722 
723 	if (kfree_reader_tasks) {
724 		for (i = 0; i < kfree_nrealthreads; i++)
725 			torture_stop_kthread(kfree_scale_thread,
726 					     kfree_reader_tasks[i]);
727 		kfree(kfree_reader_tasks);
728 	}
729 
730 	torture_cleanup_end();
731 }
732 
733 /*
734  * shutdown kthread.  Just waits to be awakened, then shuts down system.
735  */
736 static int
737 kfree_scale_shutdown(void *arg)
738 {
739 	wait_event(shutdown_wq,
740 		   atomic_read(&n_kfree_scale_thread_ended) >= kfree_nrealthreads);
741 
742 	smp_mb(); /* Wake before output. */
743 
744 	kfree_scale_cleanup();
745 	kernel_power_off();
746 	return -EINVAL;
747 }
748 
749 static int __init
750 kfree_scale_init(void)
751 {
752 	long i;
753 	int firsterr = 0;
754 
755 	kfree_nrealthreads = compute_real(kfree_nthreads);
756 	/* Start up the kthreads. */
757 	if (shutdown) {
758 		init_waitqueue_head(&shutdown_wq);
759 		firsterr = torture_create_kthread(kfree_scale_shutdown, NULL,
760 						  shutdown_task);
761 		if (torture_init_error(firsterr))
762 			goto unwind;
763 		schedule_timeout_uninterruptible(1);
764 	}
765 
766 	pr_alert("kfree object size=%zu\n", kfree_mult * sizeof(struct kfree_obj));
767 
768 	kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]),
769 			       GFP_KERNEL);
770 	if (kfree_reader_tasks == NULL) {
771 		firsterr = -ENOMEM;
772 		goto unwind;
773 	}
774 
775 	for (i = 0; i < kfree_nrealthreads; i++) {
776 		firsterr = torture_create_kthread(kfree_scale_thread, (void *)i,
777 						  kfree_reader_tasks[i]);
778 		if (torture_init_error(firsterr))
779 			goto unwind;
780 	}
781 
782 	while (atomic_read(&n_kfree_scale_thread_started) < kfree_nrealthreads)
783 		schedule_timeout_uninterruptible(1);
784 
785 	torture_init_end();
786 	return 0;
787 
788 unwind:
789 	torture_init_end();
790 	kfree_scale_cleanup();
791 	return firsterr;
792 }
793 
794 static int __init
795 rcu_scale_init(void)
796 {
797 	long i;
798 	int firsterr = 0;
799 	static struct rcu_scale_ops *scale_ops[] = {
800 		&rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops, &tasks_tracing_ops
801 	};
802 
803 	if (!torture_init_begin(scale_type, verbose))
804 		return -EBUSY;
805 
806 	/* Process args and announce that the scalability'er is on the job. */
807 	for (i = 0; i < ARRAY_SIZE(scale_ops); i++) {
808 		cur_ops = scale_ops[i];
809 		if (strcmp(scale_type, cur_ops->name) == 0)
810 			break;
811 	}
812 	if (i == ARRAY_SIZE(scale_ops)) {
813 		pr_alert("rcu-scale: invalid scale type: \"%s\"\n", scale_type);
814 		pr_alert("rcu-scale types:");
815 		for (i = 0; i < ARRAY_SIZE(scale_ops); i++)
816 			pr_cont(" %s", scale_ops[i]->name);
817 		pr_cont("\n");
818 		firsterr = -EINVAL;
819 		cur_ops = NULL;
820 		goto unwind;
821 	}
822 	if (cur_ops->init)
823 		cur_ops->init();
824 
825 	if (kfree_rcu_test)
826 		return kfree_scale_init();
827 
828 	nrealwriters = compute_real(nwriters);
829 	nrealreaders = compute_real(nreaders);
830 	atomic_set(&n_rcu_scale_reader_started, 0);
831 	atomic_set(&n_rcu_scale_writer_started, 0);
832 	atomic_set(&n_rcu_scale_writer_finished, 0);
833 	rcu_scale_print_module_parms(cur_ops, "Start of test");
834 
835 	/* Start up the kthreads. */
836 
837 	if (shutdown) {
838 		init_waitqueue_head(&shutdown_wq);
839 		firsterr = torture_create_kthread(rcu_scale_shutdown, NULL,
840 						  shutdown_task);
841 		if (torture_init_error(firsterr))
842 			goto unwind;
843 		schedule_timeout_uninterruptible(1);
844 	}
845 	reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
846 			       GFP_KERNEL);
847 	if (reader_tasks == NULL) {
848 		VERBOSE_SCALEOUT_ERRSTRING("out of memory");
849 		firsterr = -ENOMEM;
850 		goto unwind;
851 	}
852 	for (i = 0; i < nrealreaders; i++) {
853 		firsterr = torture_create_kthread(rcu_scale_reader, (void *)i,
854 						  reader_tasks[i]);
855 		if (torture_init_error(firsterr))
856 			goto unwind;
857 	}
858 	while (atomic_read(&n_rcu_scale_reader_started) < nrealreaders)
859 		schedule_timeout_uninterruptible(1);
860 	writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
861 			       GFP_KERNEL);
862 	writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
863 				   GFP_KERNEL);
864 	writer_n_durations =
865 		kcalloc(nrealwriters, sizeof(*writer_n_durations),
866 			GFP_KERNEL);
867 	if (!writer_tasks || !writer_durations || !writer_n_durations) {
868 		VERBOSE_SCALEOUT_ERRSTRING("out of memory");
869 		firsterr = -ENOMEM;
870 		goto unwind;
871 	}
872 	for (i = 0; i < nrealwriters; i++) {
873 		writer_durations[i] =
874 			kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
875 				GFP_KERNEL);
876 		if (!writer_durations[i]) {
877 			firsterr = -ENOMEM;
878 			goto unwind;
879 		}
880 		firsterr = torture_create_kthread(rcu_scale_writer, (void *)i,
881 						  writer_tasks[i]);
882 		if (torture_init_error(firsterr))
883 			goto unwind;
884 	}
885 	torture_init_end();
886 	return 0;
887 
888 unwind:
889 	torture_init_end();
890 	rcu_scale_cleanup();
891 	if (shutdown) {
892 		WARN_ON(!IS_MODULE(CONFIG_RCU_SCALE_TEST));
893 		kernel_power_off();
894 	}
895 	return firsterr;
896 }
897 
898 module_init(rcu_scale_init);
899 module_exit(rcu_scale_cleanup);
900