xref: /linux/kernel/locking/locktorture.c (revision 662fa3d6099374c4615bf64d06895e3573b935b2)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Module-based torture test facility for locking
4  *
5  * Copyright (C) IBM Corporation, 2014
6  *
7  * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
8  *          Davidlohr Bueso <dave@stgolabs.net>
9  *	Based on kernel/rcu/torture.c.
10  */
11 
12 #define pr_fmt(fmt) fmt
13 
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/kthread.h>
17 #include <linux/sched/rt.h>
18 #include <linux/spinlock.h>
19 #include <linux/mutex.h>
20 #include <linux/rwsem.h>
21 #include <linux/smp.h>
22 #include <linux/interrupt.h>
23 #include <linux/sched.h>
24 #include <uapi/linux/sched/types.h>
25 #include <linux/rtmutex.h>
26 #include <linux/atomic.h>
27 #include <linux/moduleparam.h>
28 #include <linux/delay.h>
29 #include <linux/slab.h>
30 #include <linux/torture.h>
31 #include <linux/reboot.h>
32 
33 MODULE_LICENSE("GPL");
34 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
35 
36 torture_param(int, nwriters_stress, -1,
37 	     "Number of write-locking stress-test threads");
38 torture_param(int, nreaders_stress, -1,
39 	     "Number of read-locking stress-test threads");
40 torture_param(int, onoff_holdoff, 0, "Time after boot before CPU hotplugs (s)");
41 torture_param(int, onoff_interval, 0,
42 	     "Time between CPU hotplugs (s), 0=disable");
43 torture_param(int, shuffle_interval, 3,
44 	     "Number of jiffies between shuffles, 0=disable");
45 torture_param(int, shutdown_secs, 0, "Shutdown time (j), <= zero to disable.");
46 torture_param(int, stat_interval, 60,
47 	     "Number of seconds between stats printk()s");
48 torture_param(int, stutter, 5, "Number of jiffies to run/halt test, 0=disable");
49 torture_param(int, verbose, 1,
50 	     "Enable verbose debugging printk()s");
51 
52 static char *torture_type = "spin_lock";
53 module_param(torture_type, charp, 0444);
54 MODULE_PARM_DESC(torture_type,
55 		 "Type of lock to torture (spin_lock, spin_lock_irq, mutex_lock, ...)");
56 
57 static struct task_struct *stats_task;
58 static struct task_struct **writer_tasks;
59 static struct task_struct **reader_tasks;
60 
61 static bool lock_is_write_held;
62 static bool lock_is_read_held;
63 static unsigned long last_lock_release;
64 
65 struct lock_stress_stats {
66 	long n_lock_fail;
67 	long n_lock_acquired;
68 };
69 
70 /* Forward reference. */
71 static void lock_torture_cleanup(void);
72 
73 /*
74  * Operations vector for selecting different types of tests.
75  */
76 struct lock_torture_ops {
77 	void (*init)(void);
78 	void (*exit)(void);
79 	int (*writelock)(int tid);
80 	void (*write_delay)(struct torture_random_state *trsp);
81 	void (*task_boost)(struct torture_random_state *trsp);
82 	void (*writeunlock)(int tid);
83 	int (*readlock)(int tid);
84 	void (*read_delay)(struct torture_random_state *trsp);
85 	void (*readunlock)(int tid);
86 
87 	unsigned long flags; /* for irq spinlocks */
88 	const char *name;
89 };
90 
91 struct lock_torture_cxt {
92 	int nrealwriters_stress;
93 	int nrealreaders_stress;
94 	bool debug_lock;
95 	bool init_called;
96 	atomic_t n_lock_torture_errors;
97 	struct lock_torture_ops *cur_ops;
98 	struct lock_stress_stats *lwsa; /* writer statistics */
99 	struct lock_stress_stats *lrsa; /* reader statistics */
100 };
101 static struct lock_torture_cxt cxt = { 0, 0, false, false,
102 				       ATOMIC_INIT(0),
103 				       NULL, NULL};
104 /*
105  * Definitions for lock torture testing.
106  */
107 
108 static int torture_lock_busted_write_lock(int tid __maybe_unused)
109 {
110 	return 0;  /* BUGGY, do not use in real life!!! */
111 }
112 
113 static void torture_lock_busted_write_delay(struct torture_random_state *trsp)
114 {
115 	const unsigned long longdelay_ms = 100;
116 
117 	/* We want a long delay occasionally to force massive contention.  */
118 	if (!(torture_random(trsp) %
119 	      (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
120 		mdelay(longdelay_ms);
121 	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
122 		torture_preempt_schedule();  /* Allow test to be preempted. */
123 }
124 
125 static void torture_lock_busted_write_unlock(int tid __maybe_unused)
126 {
127 	  /* BUGGY, do not use in real life!!! */
128 }
129 
130 static void torture_boost_dummy(struct torture_random_state *trsp)
131 {
132 	/* Only rtmutexes care about priority */
133 }
134 
135 static struct lock_torture_ops lock_busted_ops = {
136 	.writelock	= torture_lock_busted_write_lock,
137 	.write_delay	= torture_lock_busted_write_delay,
138 	.task_boost     = torture_boost_dummy,
139 	.writeunlock	= torture_lock_busted_write_unlock,
140 	.readlock       = NULL,
141 	.read_delay     = NULL,
142 	.readunlock     = NULL,
143 	.name		= "lock_busted"
144 };
145 
146 static DEFINE_SPINLOCK(torture_spinlock);
147 
148 static int torture_spin_lock_write_lock(int tid __maybe_unused)
149 __acquires(torture_spinlock)
150 {
151 	spin_lock(&torture_spinlock);
152 	return 0;
153 }
154 
155 static void torture_spin_lock_write_delay(struct torture_random_state *trsp)
156 {
157 	const unsigned long shortdelay_us = 2;
158 	const unsigned long longdelay_ms = 100;
159 
160 	/* We want a short delay mostly to emulate likely code, and
161 	 * we want a long delay occasionally to force massive contention.
162 	 */
163 	if (!(torture_random(trsp) %
164 	      (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
165 		mdelay(longdelay_ms);
166 	if (!(torture_random(trsp) %
167 	      (cxt.nrealwriters_stress * 2 * shortdelay_us)))
168 		udelay(shortdelay_us);
169 	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
170 		torture_preempt_schedule();  /* Allow test to be preempted. */
171 }
172 
173 static void torture_spin_lock_write_unlock(int tid __maybe_unused)
174 __releases(torture_spinlock)
175 {
176 	spin_unlock(&torture_spinlock);
177 }
178 
179 static struct lock_torture_ops spin_lock_ops = {
180 	.writelock	= torture_spin_lock_write_lock,
181 	.write_delay	= torture_spin_lock_write_delay,
182 	.task_boost     = torture_boost_dummy,
183 	.writeunlock	= torture_spin_lock_write_unlock,
184 	.readlock       = NULL,
185 	.read_delay     = NULL,
186 	.readunlock     = NULL,
187 	.name		= "spin_lock"
188 };
189 
190 static int torture_spin_lock_write_lock_irq(int tid __maybe_unused)
191 __acquires(torture_spinlock)
192 {
193 	unsigned long flags;
194 
195 	spin_lock_irqsave(&torture_spinlock, flags);
196 	cxt.cur_ops->flags = flags;
197 	return 0;
198 }
199 
200 static void torture_lock_spin_write_unlock_irq(int tid __maybe_unused)
201 __releases(torture_spinlock)
202 {
203 	spin_unlock_irqrestore(&torture_spinlock, cxt.cur_ops->flags);
204 }
205 
206 static struct lock_torture_ops spin_lock_irq_ops = {
207 	.writelock	= torture_spin_lock_write_lock_irq,
208 	.write_delay	= torture_spin_lock_write_delay,
209 	.task_boost     = torture_boost_dummy,
210 	.writeunlock	= torture_lock_spin_write_unlock_irq,
211 	.readlock       = NULL,
212 	.read_delay     = NULL,
213 	.readunlock     = NULL,
214 	.name		= "spin_lock_irq"
215 };
216 
217 static DEFINE_RWLOCK(torture_rwlock);
218 
219 static int torture_rwlock_write_lock(int tid __maybe_unused)
220 __acquires(torture_rwlock)
221 {
222 	write_lock(&torture_rwlock);
223 	return 0;
224 }
225 
226 static void torture_rwlock_write_delay(struct torture_random_state *trsp)
227 {
228 	const unsigned long shortdelay_us = 2;
229 	const unsigned long longdelay_ms = 100;
230 
231 	/* We want a short delay mostly to emulate likely code, and
232 	 * we want a long delay occasionally to force massive contention.
233 	 */
234 	if (!(torture_random(trsp) %
235 	      (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
236 		mdelay(longdelay_ms);
237 	else
238 		udelay(shortdelay_us);
239 }
240 
241 static void torture_rwlock_write_unlock(int tid __maybe_unused)
242 __releases(torture_rwlock)
243 {
244 	write_unlock(&torture_rwlock);
245 }
246 
247 static int torture_rwlock_read_lock(int tid __maybe_unused)
248 __acquires(torture_rwlock)
249 {
250 	read_lock(&torture_rwlock);
251 	return 0;
252 }
253 
254 static void torture_rwlock_read_delay(struct torture_random_state *trsp)
255 {
256 	const unsigned long shortdelay_us = 10;
257 	const unsigned long longdelay_ms = 100;
258 
259 	/* We want a short delay mostly to emulate likely code, and
260 	 * we want a long delay occasionally to force massive contention.
261 	 */
262 	if (!(torture_random(trsp) %
263 	      (cxt.nrealreaders_stress * 2000 * longdelay_ms)))
264 		mdelay(longdelay_ms);
265 	else
266 		udelay(shortdelay_us);
267 }
268 
269 static void torture_rwlock_read_unlock(int tid __maybe_unused)
270 __releases(torture_rwlock)
271 {
272 	read_unlock(&torture_rwlock);
273 }
274 
275 static struct lock_torture_ops rw_lock_ops = {
276 	.writelock	= torture_rwlock_write_lock,
277 	.write_delay	= torture_rwlock_write_delay,
278 	.task_boost     = torture_boost_dummy,
279 	.writeunlock	= torture_rwlock_write_unlock,
280 	.readlock       = torture_rwlock_read_lock,
281 	.read_delay     = torture_rwlock_read_delay,
282 	.readunlock     = torture_rwlock_read_unlock,
283 	.name		= "rw_lock"
284 };
285 
286 static int torture_rwlock_write_lock_irq(int tid __maybe_unused)
287 __acquires(torture_rwlock)
288 {
289 	unsigned long flags;
290 
291 	write_lock_irqsave(&torture_rwlock, flags);
292 	cxt.cur_ops->flags = flags;
293 	return 0;
294 }
295 
296 static void torture_rwlock_write_unlock_irq(int tid __maybe_unused)
297 __releases(torture_rwlock)
298 {
299 	write_unlock_irqrestore(&torture_rwlock, cxt.cur_ops->flags);
300 }
301 
302 static int torture_rwlock_read_lock_irq(int tid __maybe_unused)
303 __acquires(torture_rwlock)
304 {
305 	unsigned long flags;
306 
307 	read_lock_irqsave(&torture_rwlock, flags);
308 	cxt.cur_ops->flags = flags;
309 	return 0;
310 }
311 
312 static void torture_rwlock_read_unlock_irq(int tid __maybe_unused)
313 __releases(torture_rwlock)
314 {
315 	read_unlock_irqrestore(&torture_rwlock, cxt.cur_ops->flags);
316 }
317 
318 static struct lock_torture_ops rw_lock_irq_ops = {
319 	.writelock	= torture_rwlock_write_lock_irq,
320 	.write_delay	= torture_rwlock_write_delay,
321 	.task_boost     = torture_boost_dummy,
322 	.writeunlock	= torture_rwlock_write_unlock_irq,
323 	.readlock       = torture_rwlock_read_lock_irq,
324 	.read_delay     = torture_rwlock_read_delay,
325 	.readunlock     = torture_rwlock_read_unlock_irq,
326 	.name		= "rw_lock_irq"
327 };
328 
329 static DEFINE_MUTEX(torture_mutex);
330 
331 static int torture_mutex_lock(int tid __maybe_unused)
332 __acquires(torture_mutex)
333 {
334 	mutex_lock(&torture_mutex);
335 	return 0;
336 }
337 
338 static void torture_mutex_delay(struct torture_random_state *trsp)
339 {
340 	const unsigned long longdelay_ms = 100;
341 
342 	/* We want a long delay occasionally to force massive contention.  */
343 	if (!(torture_random(trsp) %
344 	      (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
345 		mdelay(longdelay_ms * 5);
346 	else
347 		mdelay(longdelay_ms / 5);
348 	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
349 		torture_preempt_schedule();  /* Allow test to be preempted. */
350 }
351 
352 static void torture_mutex_unlock(int tid __maybe_unused)
353 __releases(torture_mutex)
354 {
355 	mutex_unlock(&torture_mutex);
356 }
357 
358 static struct lock_torture_ops mutex_lock_ops = {
359 	.writelock	= torture_mutex_lock,
360 	.write_delay	= torture_mutex_delay,
361 	.task_boost     = torture_boost_dummy,
362 	.writeunlock	= torture_mutex_unlock,
363 	.readlock       = NULL,
364 	.read_delay     = NULL,
365 	.readunlock     = NULL,
366 	.name		= "mutex_lock"
367 };
368 
369 #include <linux/ww_mutex.h>
370 /*
371  * The torture ww_mutexes should belong to the same lock class as
372  * torture_ww_class to avoid lockdep problem. The ww_mutex_init()
373  * function is called for initialization to ensure that.
374  */
375 static DEFINE_WD_CLASS(torture_ww_class);
376 static struct ww_mutex torture_ww_mutex_0, torture_ww_mutex_1, torture_ww_mutex_2;
377 static struct ww_acquire_ctx *ww_acquire_ctxs;
378 
379 static void torture_ww_mutex_init(void)
380 {
381 	ww_mutex_init(&torture_ww_mutex_0, &torture_ww_class);
382 	ww_mutex_init(&torture_ww_mutex_1, &torture_ww_class);
383 	ww_mutex_init(&torture_ww_mutex_2, &torture_ww_class);
384 
385 	ww_acquire_ctxs = kmalloc_array(cxt.nrealwriters_stress,
386 					sizeof(*ww_acquire_ctxs),
387 					GFP_KERNEL);
388 	if (!ww_acquire_ctxs)
389 		VERBOSE_TOROUT_STRING("ww_acquire_ctx: Out of memory");
390 }
391 
392 static void torture_ww_mutex_exit(void)
393 {
394 	kfree(ww_acquire_ctxs);
395 }
396 
397 static int torture_ww_mutex_lock(int tid)
398 __acquires(torture_ww_mutex_0)
399 __acquires(torture_ww_mutex_1)
400 __acquires(torture_ww_mutex_2)
401 {
402 	LIST_HEAD(list);
403 	struct reorder_lock {
404 		struct list_head link;
405 		struct ww_mutex *lock;
406 	} locks[3], *ll, *ln;
407 	struct ww_acquire_ctx *ctx = &ww_acquire_ctxs[tid];
408 
409 	locks[0].lock = &torture_ww_mutex_0;
410 	list_add(&locks[0].link, &list);
411 
412 	locks[1].lock = &torture_ww_mutex_1;
413 	list_add(&locks[1].link, &list);
414 
415 	locks[2].lock = &torture_ww_mutex_2;
416 	list_add(&locks[2].link, &list);
417 
418 	ww_acquire_init(ctx, &torture_ww_class);
419 
420 	list_for_each_entry(ll, &list, link) {
421 		int err;
422 
423 		err = ww_mutex_lock(ll->lock, ctx);
424 		if (!err)
425 			continue;
426 
427 		ln = ll;
428 		list_for_each_entry_continue_reverse(ln, &list, link)
429 			ww_mutex_unlock(ln->lock);
430 
431 		if (err != -EDEADLK)
432 			return err;
433 
434 		ww_mutex_lock_slow(ll->lock, ctx);
435 		list_move(&ll->link, &list);
436 	}
437 
438 	return 0;
439 }
440 
441 static void torture_ww_mutex_unlock(int tid)
442 __releases(torture_ww_mutex_0)
443 __releases(torture_ww_mutex_1)
444 __releases(torture_ww_mutex_2)
445 {
446 	struct ww_acquire_ctx *ctx = &ww_acquire_ctxs[tid];
447 
448 	ww_mutex_unlock(&torture_ww_mutex_0);
449 	ww_mutex_unlock(&torture_ww_mutex_1);
450 	ww_mutex_unlock(&torture_ww_mutex_2);
451 	ww_acquire_fini(ctx);
452 }
453 
454 static struct lock_torture_ops ww_mutex_lock_ops = {
455 	.init		= torture_ww_mutex_init,
456 	.exit		= torture_ww_mutex_exit,
457 	.writelock	= torture_ww_mutex_lock,
458 	.write_delay	= torture_mutex_delay,
459 	.task_boost     = torture_boost_dummy,
460 	.writeunlock	= torture_ww_mutex_unlock,
461 	.readlock       = NULL,
462 	.read_delay     = NULL,
463 	.readunlock     = NULL,
464 	.name		= "ww_mutex_lock"
465 };
466 
467 #ifdef CONFIG_RT_MUTEXES
468 static DEFINE_RT_MUTEX(torture_rtmutex);
469 
470 static int torture_rtmutex_lock(int tid __maybe_unused)
471 __acquires(torture_rtmutex)
472 {
473 	rt_mutex_lock(&torture_rtmutex);
474 	return 0;
475 }
476 
477 static void torture_rtmutex_boost(struct torture_random_state *trsp)
478 {
479 	const unsigned int factor = 50000; /* yes, quite arbitrary */
480 
481 	if (!rt_task(current)) {
482 		/*
483 		 * Boost priority once every ~50k operations. When the
484 		 * task tries to take the lock, the rtmutex it will account
485 		 * for the new priority, and do any corresponding pi-dance.
486 		 */
487 		if (trsp && !(torture_random(trsp) %
488 			      (cxt.nrealwriters_stress * factor))) {
489 			sched_set_fifo(current);
490 		} else /* common case, do nothing */
491 			return;
492 	} else {
493 		/*
494 		 * The task will remain boosted for another ~500k operations,
495 		 * then restored back to its original prio, and so forth.
496 		 *
497 		 * When @trsp is nil, we want to force-reset the task for
498 		 * stopping the kthread.
499 		 */
500 		if (!trsp || !(torture_random(trsp) %
501 			       (cxt.nrealwriters_stress * factor * 2))) {
502 			sched_set_normal(current, 0);
503 		} else /* common case, do nothing */
504 			return;
505 	}
506 }
507 
508 static void torture_rtmutex_delay(struct torture_random_state *trsp)
509 {
510 	const unsigned long shortdelay_us = 2;
511 	const unsigned long longdelay_ms = 100;
512 
513 	/*
514 	 * We want a short delay mostly to emulate likely code, and
515 	 * we want a long delay occasionally to force massive contention.
516 	 */
517 	if (!(torture_random(trsp) %
518 	      (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
519 		mdelay(longdelay_ms);
520 	if (!(torture_random(trsp) %
521 	      (cxt.nrealwriters_stress * 2 * shortdelay_us)))
522 		udelay(shortdelay_us);
523 	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
524 		torture_preempt_schedule();  /* Allow test to be preempted. */
525 }
526 
527 static void torture_rtmutex_unlock(int tid __maybe_unused)
528 __releases(torture_rtmutex)
529 {
530 	rt_mutex_unlock(&torture_rtmutex);
531 }
532 
533 static struct lock_torture_ops rtmutex_lock_ops = {
534 	.writelock	= torture_rtmutex_lock,
535 	.write_delay	= torture_rtmutex_delay,
536 	.task_boost     = torture_rtmutex_boost,
537 	.writeunlock	= torture_rtmutex_unlock,
538 	.readlock       = NULL,
539 	.read_delay     = NULL,
540 	.readunlock     = NULL,
541 	.name		= "rtmutex_lock"
542 };
543 #endif
544 
545 static DECLARE_RWSEM(torture_rwsem);
546 static int torture_rwsem_down_write(int tid __maybe_unused)
547 __acquires(torture_rwsem)
548 {
549 	down_write(&torture_rwsem);
550 	return 0;
551 }
552 
553 static void torture_rwsem_write_delay(struct torture_random_state *trsp)
554 {
555 	const unsigned long longdelay_ms = 100;
556 
557 	/* We want a long delay occasionally to force massive contention.  */
558 	if (!(torture_random(trsp) %
559 	      (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
560 		mdelay(longdelay_ms * 10);
561 	else
562 		mdelay(longdelay_ms / 10);
563 	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
564 		torture_preempt_schedule();  /* Allow test to be preempted. */
565 }
566 
567 static void torture_rwsem_up_write(int tid __maybe_unused)
568 __releases(torture_rwsem)
569 {
570 	up_write(&torture_rwsem);
571 }
572 
573 static int torture_rwsem_down_read(int tid __maybe_unused)
574 __acquires(torture_rwsem)
575 {
576 	down_read(&torture_rwsem);
577 	return 0;
578 }
579 
580 static void torture_rwsem_read_delay(struct torture_random_state *trsp)
581 {
582 	const unsigned long longdelay_ms = 100;
583 
584 	/* We want a long delay occasionally to force massive contention.  */
585 	if (!(torture_random(trsp) %
586 	      (cxt.nrealreaders_stress * 2000 * longdelay_ms)))
587 		mdelay(longdelay_ms * 2);
588 	else
589 		mdelay(longdelay_ms / 2);
590 	if (!(torture_random(trsp) % (cxt.nrealreaders_stress * 20000)))
591 		torture_preempt_schedule();  /* Allow test to be preempted. */
592 }
593 
594 static void torture_rwsem_up_read(int tid __maybe_unused)
595 __releases(torture_rwsem)
596 {
597 	up_read(&torture_rwsem);
598 }
599 
600 static struct lock_torture_ops rwsem_lock_ops = {
601 	.writelock	= torture_rwsem_down_write,
602 	.write_delay	= torture_rwsem_write_delay,
603 	.task_boost     = torture_boost_dummy,
604 	.writeunlock	= torture_rwsem_up_write,
605 	.readlock       = torture_rwsem_down_read,
606 	.read_delay     = torture_rwsem_read_delay,
607 	.readunlock     = torture_rwsem_up_read,
608 	.name		= "rwsem_lock"
609 };
610 
611 #include <linux/percpu-rwsem.h>
612 static struct percpu_rw_semaphore pcpu_rwsem;
613 
614 static void torture_percpu_rwsem_init(void)
615 {
616 	BUG_ON(percpu_init_rwsem(&pcpu_rwsem));
617 }
618 
619 static void torture_percpu_rwsem_exit(void)
620 {
621 	percpu_free_rwsem(&pcpu_rwsem);
622 }
623 
624 static int torture_percpu_rwsem_down_write(int tid __maybe_unused)
625 __acquires(pcpu_rwsem)
626 {
627 	percpu_down_write(&pcpu_rwsem);
628 	return 0;
629 }
630 
631 static void torture_percpu_rwsem_up_write(int tid __maybe_unused)
632 __releases(pcpu_rwsem)
633 {
634 	percpu_up_write(&pcpu_rwsem);
635 }
636 
637 static int torture_percpu_rwsem_down_read(int tid __maybe_unused)
638 __acquires(pcpu_rwsem)
639 {
640 	percpu_down_read(&pcpu_rwsem);
641 	return 0;
642 }
643 
644 static void torture_percpu_rwsem_up_read(int tid __maybe_unused)
645 __releases(pcpu_rwsem)
646 {
647 	percpu_up_read(&pcpu_rwsem);
648 }
649 
650 static struct lock_torture_ops percpu_rwsem_lock_ops = {
651 	.init		= torture_percpu_rwsem_init,
652 	.exit		= torture_percpu_rwsem_exit,
653 	.writelock	= torture_percpu_rwsem_down_write,
654 	.write_delay	= torture_rwsem_write_delay,
655 	.task_boost     = torture_boost_dummy,
656 	.writeunlock	= torture_percpu_rwsem_up_write,
657 	.readlock       = torture_percpu_rwsem_down_read,
658 	.read_delay     = torture_rwsem_read_delay,
659 	.readunlock     = torture_percpu_rwsem_up_read,
660 	.name		= "percpu_rwsem_lock"
661 };
662 
663 /*
664  * Lock torture writer kthread.  Repeatedly acquires and releases
665  * the lock, checking for duplicate acquisitions.
666  */
667 static int lock_torture_writer(void *arg)
668 {
669 	struct lock_stress_stats *lwsp = arg;
670 	int tid = lwsp - cxt.lwsa;
671 	DEFINE_TORTURE_RANDOM(rand);
672 
673 	VERBOSE_TOROUT_STRING("lock_torture_writer task started");
674 	set_user_nice(current, MAX_NICE);
675 
676 	do {
677 		if ((torture_random(&rand) & 0xfffff) == 0)
678 			schedule_timeout_uninterruptible(1);
679 
680 		cxt.cur_ops->task_boost(&rand);
681 		cxt.cur_ops->writelock(tid);
682 		if (WARN_ON_ONCE(lock_is_write_held))
683 			lwsp->n_lock_fail++;
684 		lock_is_write_held = true;
685 		if (WARN_ON_ONCE(lock_is_read_held))
686 			lwsp->n_lock_fail++; /* rare, but... */
687 
688 		lwsp->n_lock_acquired++;
689 		cxt.cur_ops->write_delay(&rand);
690 		lock_is_write_held = false;
691 		WRITE_ONCE(last_lock_release, jiffies);
692 		cxt.cur_ops->writeunlock(tid);
693 
694 		stutter_wait("lock_torture_writer");
695 	} while (!torture_must_stop());
696 
697 	cxt.cur_ops->task_boost(NULL); /* reset prio */
698 	torture_kthread_stopping("lock_torture_writer");
699 	return 0;
700 }
701 
702 /*
703  * Lock torture reader kthread.  Repeatedly acquires and releases
704  * the reader lock.
705  */
706 static int lock_torture_reader(void *arg)
707 {
708 	struct lock_stress_stats *lrsp = arg;
709 	int tid = lrsp - cxt.lrsa;
710 	DEFINE_TORTURE_RANDOM(rand);
711 
712 	VERBOSE_TOROUT_STRING("lock_torture_reader task started");
713 	set_user_nice(current, MAX_NICE);
714 
715 	do {
716 		if ((torture_random(&rand) & 0xfffff) == 0)
717 			schedule_timeout_uninterruptible(1);
718 
719 		cxt.cur_ops->readlock(tid);
720 		lock_is_read_held = true;
721 		if (WARN_ON_ONCE(lock_is_write_held))
722 			lrsp->n_lock_fail++; /* rare, but... */
723 
724 		lrsp->n_lock_acquired++;
725 		cxt.cur_ops->read_delay(&rand);
726 		lock_is_read_held = false;
727 		cxt.cur_ops->readunlock(tid);
728 
729 		stutter_wait("lock_torture_reader");
730 	} while (!torture_must_stop());
731 	torture_kthread_stopping("lock_torture_reader");
732 	return 0;
733 }
734 
735 /*
736  * Create an lock-torture-statistics message in the specified buffer.
737  */
738 static void __torture_print_stats(char *page,
739 				  struct lock_stress_stats *statp, bool write)
740 {
741 	bool fail = false;
742 	int i, n_stress;
743 	long max = 0, min = statp ? statp[0].n_lock_acquired : 0;
744 	long long sum = 0;
745 
746 	n_stress = write ? cxt.nrealwriters_stress : cxt.nrealreaders_stress;
747 	for (i = 0; i < n_stress; i++) {
748 		if (statp[i].n_lock_fail)
749 			fail = true;
750 		sum += statp[i].n_lock_acquired;
751 		if (max < statp[i].n_lock_acquired)
752 			max = statp[i].n_lock_acquired;
753 		if (min > statp[i].n_lock_acquired)
754 			min = statp[i].n_lock_acquired;
755 	}
756 	page += sprintf(page,
757 			"%s:  Total: %lld  Max/Min: %ld/%ld %s  Fail: %d %s\n",
758 			write ? "Writes" : "Reads ",
759 			sum, max, min,
760 			!onoff_interval && max / 2 > min ? "???" : "",
761 			fail, fail ? "!!!" : "");
762 	if (fail)
763 		atomic_inc(&cxt.n_lock_torture_errors);
764 }
765 
766 /*
767  * Print torture statistics.  Caller must ensure that there is only one
768  * call to this function at a given time!!!  This is normally accomplished
769  * by relying on the module system to only have one copy of the module
770  * loaded, and then by giving the lock_torture_stats kthread full control
771  * (or the init/cleanup functions when lock_torture_stats thread is not
772  * running).
773  */
774 static void lock_torture_stats_print(void)
775 {
776 	int size = cxt.nrealwriters_stress * 200 + 8192;
777 	char *buf;
778 
779 	if (cxt.cur_ops->readlock)
780 		size += cxt.nrealreaders_stress * 200 + 8192;
781 
782 	buf = kmalloc(size, GFP_KERNEL);
783 	if (!buf) {
784 		pr_err("lock_torture_stats_print: Out of memory, need: %d",
785 		       size);
786 		return;
787 	}
788 
789 	__torture_print_stats(buf, cxt.lwsa, true);
790 	pr_alert("%s", buf);
791 	kfree(buf);
792 
793 	if (cxt.cur_ops->readlock) {
794 		buf = kmalloc(size, GFP_KERNEL);
795 		if (!buf) {
796 			pr_err("lock_torture_stats_print: Out of memory, need: %d",
797 			       size);
798 			return;
799 		}
800 
801 		__torture_print_stats(buf, cxt.lrsa, false);
802 		pr_alert("%s", buf);
803 		kfree(buf);
804 	}
805 }
806 
807 /*
808  * Periodically prints torture statistics, if periodic statistics printing
809  * was specified via the stat_interval module parameter.
810  *
811  * No need to worry about fullstop here, since this one doesn't reference
812  * volatile state or register callbacks.
813  */
814 static int lock_torture_stats(void *arg)
815 {
816 	VERBOSE_TOROUT_STRING("lock_torture_stats task started");
817 	do {
818 		schedule_timeout_interruptible(stat_interval * HZ);
819 		lock_torture_stats_print();
820 		torture_shutdown_absorb("lock_torture_stats");
821 	} while (!torture_must_stop());
822 	torture_kthread_stopping("lock_torture_stats");
823 	return 0;
824 }
825 
826 static inline void
827 lock_torture_print_module_parms(struct lock_torture_ops *cur_ops,
828 				const char *tag)
829 {
830 	pr_alert("%s" TORTURE_FLAG
831 		 "--- %s%s: nwriters_stress=%d nreaders_stress=%d stat_interval=%d verbose=%d shuffle_interval=%d stutter=%d shutdown_secs=%d onoff_interval=%d onoff_holdoff=%d\n",
832 		 torture_type, tag, cxt.debug_lock ? " [debug]": "",
833 		 cxt.nrealwriters_stress, cxt.nrealreaders_stress, stat_interval,
834 		 verbose, shuffle_interval, stutter, shutdown_secs,
835 		 onoff_interval, onoff_holdoff);
836 }
837 
838 static void lock_torture_cleanup(void)
839 {
840 	int i;
841 
842 	if (torture_cleanup_begin())
843 		return;
844 
845 	/*
846 	 * Indicates early cleanup, meaning that the test has not run,
847 	 * such as when passing bogus args when loading the module.
848 	 * However cxt->cur_ops.init() may have been invoked, so beside
849 	 * perform the underlying torture-specific cleanups, cur_ops.exit()
850 	 * will be invoked if needed.
851 	 */
852 	if (!cxt.lwsa && !cxt.lrsa)
853 		goto end;
854 
855 	if (writer_tasks) {
856 		for (i = 0; i < cxt.nrealwriters_stress; i++)
857 			torture_stop_kthread(lock_torture_writer,
858 					     writer_tasks[i]);
859 		kfree(writer_tasks);
860 		writer_tasks = NULL;
861 	}
862 
863 	if (reader_tasks) {
864 		for (i = 0; i < cxt.nrealreaders_stress; i++)
865 			torture_stop_kthread(lock_torture_reader,
866 					     reader_tasks[i]);
867 		kfree(reader_tasks);
868 		reader_tasks = NULL;
869 	}
870 
871 	torture_stop_kthread(lock_torture_stats, stats_task);
872 	lock_torture_stats_print();  /* -After- the stats thread is stopped! */
873 
874 	if (atomic_read(&cxt.n_lock_torture_errors))
875 		lock_torture_print_module_parms(cxt.cur_ops,
876 						"End of test: FAILURE");
877 	else if (torture_onoff_failures())
878 		lock_torture_print_module_parms(cxt.cur_ops,
879 						"End of test: LOCK_HOTPLUG");
880 	else
881 		lock_torture_print_module_parms(cxt.cur_ops,
882 						"End of test: SUCCESS");
883 
884 	kfree(cxt.lwsa);
885 	cxt.lwsa = NULL;
886 	kfree(cxt.lrsa);
887 	cxt.lrsa = NULL;
888 
889 end:
890 	if (cxt.init_called) {
891 		if (cxt.cur_ops->exit)
892 			cxt.cur_ops->exit();
893 		cxt.init_called = false;
894 	}
895 	torture_cleanup_end();
896 }
897 
898 static int __init lock_torture_init(void)
899 {
900 	int i, j;
901 	int firsterr = 0;
902 	static struct lock_torture_ops *torture_ops[] = {
903 		&lock_busted_ops,
904 		&spin_lock_ops, &spin_lock_irq_ops,
905 		&rw_lock_ops, &rw_lock_irq_ops,
906 		&mutex_lock_ops,
907 		&ww_mutex_lock_ops,
908 #ifdef CONFIG_RT_MUTEXES
909 		&rtmutex_lock_ops,
910 #endif
911 		&rwsem_lock_ops,
912 		&percpu_rwsem_lock_ops,
913 	};
914 
915 	if (!torture_init_begin(torture_type, verbose))
916 		return -EBUSY;
917 
918 	/* Process args and tell the world that the torturer is on the job. */
919 	for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
920 		cxt.cur_ops = torture_ops[i];
921 		if (strcmp(torture_type, cxt.cur_ops->name) == 0)
922 			break;
923 	}
924 	if (i == ARRAY_SIZE(torture_ops)) {
925 		pr_alert("lock-torture: invalid torture type: \"%s\"\n",
926 			 torture_type);
927 		pr_alert("lock-torture types:");
928 		for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
929 			pr_alert(" %s", torture_ops[i]->name);
930 		pr_alert("\n");
931 		firsterr = -EINVAL;
932 		goto unwind;
933 	}
934 
935 	if (nwriters_stress == 0 &&
936 	    (!cxt.cur_ops->readlock || nreaders_stress == 0)) {
937 		pr_alert("lock-torture: must run at least one locking thread\n");
938 		firsterr = -EINVAL;
939 		goto unwind;
940 	}
941 
942 	if (nwriters_stress >= 0)
943 		cxt.nrealwriters_stress = nwriters_stress;
944 	else
945 		cxt.nrealwriters_stress = 2 * num_online_cpus();
946 
947 	if (cxt.cur_ops->init) {
948 		cxt.cur_ops->init();
949 		cxt.init_called = true;
950 	}
951 
952 #ifdef CONFIG_DEBUG_MUTEXES
953 	if (str_has_prefix(torture_type, "mutex"))
954 		cxt.debug_lock = true;
955 #endif
956 #ifdef CONFIG_DEBUG_RT_MUTEXES
957 	if (str_has_prefix(torture_type, "rtmutex"))
958 		cxt.debug_lock = true;
959 #endif
960 #ifdef CONFIG_DEBUG_SPINLOCK
961 	if ((str_has_prefix(torture_type, "spin")) ||
962 	    (str_has_prefix(torture_type, "rw_lock")))
963 		cxt.debug_lock = true;
964 #endif
965 
966 	/* Initialize the statistics so that each run gets its own numbers. */
967 	if (nwriters_stress) {
968 		lock_is_write_held = false;
969 		cxt.lwsa = kmalloc_array(cxt.nrealwriters_stress,
970 					 sizeof(*cxt.lwsa),
971 					 GFP_KERNEL);
972 		if (cxt.lwsa == NULL) {
973 			VERBOSE_TOROUT_STRING("cxt.lwsa: Out of memory");
974 			firsterr = -ENOMEM;
975 			goto unwind;
976 		}
977 
978 		for (i = 0; i < cxt.nrealwriters_stress; i++) {
979 			cxt.lwsa[i].n_lock_fail = 0;
980 			cxt.lwsa[i].n_lock_acquired = 0;
981 		}
982 	}
983 
984 	if (cxt.cur_ops->readlock) {
985 		if (nreaders_stress >= 0)
986 			cxt.nrealreaders_stress = nreaders_stress;
987 		else {
988 			/*
989 			 * By default distribute evenly the number of
990 			 * readers and writers. We still run the same number
991 			 * of threads as the writer-only locks default.
992 			 */
993 			if (nwriters_stress < 0) /* user doesn't care */
994 				cxt.nrealwriters_stress = num_online_cpus();
995 			cxt.nrealreaders_stress = cxt.nrealwriters_stress;
996 		}
997 
998 		if (nreaders_stress) {
999 			lock_is_read_held = false;
1000 			cxt.lrsa = kmalloc_array(cxt.nrealreaders_stress,
1001 						 sizeof(*cxt.lrsa),
1002 						 GFP_KERNEL);
1003 			if (cxt.lrsa == NULL) {
1004 				VERBOSE_TOROUT_STRING("cxt.lrsa: Out of memory");
1005 				firsterr = -ENOMEM;
1006 				kfree(cxt.lwsa);
1007 				cxt.lwsa = NULL;
1008 				goto unwind;
1009 			}
1010 
1011 			for (i = 0; i < cxt.nrealreaders_stress; i++) {
1012 				cxt.lrsa[i].n_lock_fail = 0;
1013 				cxt.lrsa[i].n_lock_acquired = 0;
1014 			}
1015 		}
1016 	}
1017 
1018 	lock_torture_print_module_parms(cxt.cur_ops, "Start of test");
1019 
1020 	/* Prepare torture context. */
1021 	if (onoff_interval > 0) {
1022 		firsterr = torture_onoff_init(onoff_holdoff * HZ,
1023 					      onoff_interval * HZ, NULL);
1024 		if (firsterr)
1025 			goto unwind;
1026 	}
1027 	if (shuffle_interval > 0) {
1028 		firsterr = torture_shuffle_init(shuffle_interval);
1029 		if (firsterr)
1030 			goto unwind;
1031 	}
1032 	if (shutdown_secs > 0) {
1033 		firsterr = torture_shutdown_init(shutdown_secs,
1034 						 lock_torture_cleanup);
1035 		if (firsterr)
1036 			goto unwind;
1037 	}
1038 	if (stutter > 0) {
1039 		firsterr = torture_stutter_init(stutter, stutter);
1040 		if (firsterr)
1041 			goto unwind;
1042 	}
1043 
1044 	if (nwriters_stress) {
1045 		writer_tasks = kcalloc(cxt.nrealwriters_stress,
1046 				       sizeof(writer_tasks[0]),
1047 				       GFP_KERNEL);
1048 		if (writer_tasks == NULL) {
1049 			VERBOSE_TOROUT_ERRSTRING("writer_tasks: Out of memory");
1050 			firsterr = -ENOMEM;
1051 			goto unwind;
1052 		}
1053 	}
1054 
1055 	if (cxt.cur_ops->readlock) {
1056 		reader_tasks = kcalloc(cxt.nrealreaders_stress,
1057 				       sizeof(reader_tasks[0]),
1058 				       GFP_KERNEL);
1059 		if (reader_tasks == NULL) {
1060 			VERBOSE_TOROUT_ERRSTRING("reader_tasks: Out of memory");
1061 			kfree(writer_tasks);
1062 			writer_tasks = NULL;
1063 			firsterr = -ENOMEM;
1064 			goto unwind;
1065 		}
1066 	}
1067 
1068 	/*
1069 	 * Create the kthreads and start torturing (oh, those poor little locks).
1070 	 *
1071 	 * TODO: Note that we interleave writers with readers, giving writers a
1072 	 * slight advantage, by creating its kthread first. This can be modified
1073 	 * for very specific needs, or even let the user choose the policy, if
1074 	 * ever wanted.
1075 	 */
1076 	for (i = 0, j = 0; i < cxt.nrealwriters_stress ||
1077 		    j < cxt.nrealreaders_stress; i++, j++) {
1078 		if (i >= cxt.nrealwriters_stress)
1079 			goto create_reader;
1080 
1081 		/* Create writer. */
1082 		firsterr = torture_create_kthread(lock_torture_writer, &cxt.lwsa[i],
1083 						  writer_tasks[i]);
1084 		if (firsterr)
1085 			goto unwind;
1086 
1087 	create_reader:
1088 		if (cxt.cur_ops->readlock == NULL || (j >= cxt.nrealreaders_stress))
1089 			continue;
1090 		/* Create reader. */
1091 		firsterr = torture_create_kthread(lock_torture_reader, &cxt.lrsa[j],
1092 						  reader_tasks[j]);
1093 		if (firsterr)
1094 			goto unwind;
1095 	}
1096 	if (stat_interval > 0) {
1097 		firsterr = torture_create_kthread(lock_torture_stats, NULL,
1098 						  stats_task);
1099 		if (firsterr)
1100 			goto unwind;
1101 	}
1102 	torture_init_end();
1103 	return 0;
1104 
1105 unwind:
1106 	torture_init_end();
1107 	lock_torture_cleanup();
1108 	if (shutdown_secs) {
1109 		WARN_ON(!IS_MODULE(CONFIG_LOCK_TORTURE_TEST));
1110 		kernel_power_off();
1111 	}
1112 	return firsterr;
1113 }
1114 
1115 module_init(lock_torture_init);
1116 module_exit(lock_torture_cleanup);
1117