Merge tag 'm68k-for-v6.4-tag1' of git://git.kernel.org/pub/scm/linux/kernel/git/geert...
[linux-block.git] / kernel / locking / locktorture.c
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, rt_boost, 2,
50                 "Do periodic rt-boost. 0=Disable, 1=Only for rt_mutex, 2=For all lock types.");
51 torture_param(int, rt_boost_factor, 50, "A factor determining how often rt-boost happens.");
52 torture_param(int, verbose, 1,
53              "Enable verbose debugging printk()s");
54 torture_param(int, nested_locks, 0, "Number of nested locks (max = 8)");
55 /* Going much higher trips "BUG: MAX_LOCKDEP_CHAIN_HLOCKS too low!" errors */
56 #define MAX_NESTED_LOCKS 8
57
58 static char *torture_type = IS_ENABLED(CONFIG_PREEMPT_RT) ? "raw_spin_lock" : "spin_lock";
59 module_param(torture_type, charp, 0444);
60 MODULE_PARM_DESC(torture_type,
61                  "Type of lock to torture (spin_lock, spin_lock_irq, mutex_lock, ...)");
62
63 static struct task_struct *stats_task;
64 static struct task_struct **writer_tasks;
65 static struct task_struct **reader_tasks;
66
67 static bool lock_is_write_held;
68 static atomic_t lock_is_read_held;
69 static unsigned long last_lock_release;
70
71 struct lock_stress_stats {
72         long n_lock_fail;
73         long n_lock_acquired;
74 };
75
76 /* Forward reference. */
77 static void lock_torture_cleanup(void);
78
79 /*
80  * Operations vector for selecting different types of tests.
81  */
82 struct lock_torture_ops {
83         void (*init)(void);
84         void (*exit)(void);
85         int (*nested_lock)(int tid, u32 lockset);
86         int (*writelock)(int tid);
87         void (*write_delay)(struct torture_random_state *trsp);
88         void (*task_boost)(struct torture_random_state *trsp);
89         void (*writeunlock)(int tid);
90         void (*nested_unlock)(int tid, u32 lockset);
91         int (*readlock)(int tid);
92         void (*read_delay)(struct torture_random_state *trsp);
93         void (*readunlock)(int tid);
94
95         unsigned long flags; /* for irq spinlocks */
96         const char *name;
97 };
98
99 struct lock_torture_cxt {
100         int nrealwriters_stress;
101         int nrealreaders_stress;
102         bool debug_lock;
103         bool init_called;
104         atomic_t n_lock_torture_errors;
105         struct lock_torture_ops *cur_ops;
106         struct lock_stress_stats *lwsa; /* writer statistics */
107         struct lock_stress_stats *lrsa; /* reader statistics */
108 };
109 static struct lock_torture_cxt cxt = { 0, 0, false, false,
110                                        ATOMIC_INIT(0),
111                                        NULL, NULL};
112 /*
113  * Definitions for lock torture testing.
114  */
115
116 static int torture_lock_busted_write_lock(int tid __maybe_unused)
117 {
118         return 0;  /* BUGGY, do not use in real life!!! */
119 }
120
121 static void torture_lock_busted_write_delay(struct torture_random_state *trsp)
122 {
123         const unsigned long longdelay_ms = 100;
124
125         /* We want a long delay occasionally to force massive contention.  */
126         if (!(torture_random(trsp) %
127               (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
128                 mdelay(longdelay_ms);
129         if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
130                 torture_preempt_schedule();  /* Allow test to be preempted. */
131 }
132
133 static void torture_lock_busted_write_unlock(int tid __maybe_unused)
134 {
135           /* BUGGY, do not use in real life!!! */
136 }
137
138 static void __torture_rt_boost(struct torture_random_state *trsp)
139 {
140         const unsigned int factor = rt_boost_factor;
141
142         if (!rt_task(current)) {
143                 /*
144                  * Boost priority once every rt_boost_factor operations. When
145                  * the task tries to take the lock, the rtmutex it will account
146                  * for the new priority, and do any corresponding pi-dance.
147                  */
148                 if (trsp && !(torture_random(trsp) %
149                               (cxt.nrealwriters_stress * factor))) {
150                         sched_set_fifo(current);
151                 } else /* common case, do nothing */
152                         return;
153         } else {
154                 /*
155                  * The task will remain boosted for another 10 * rt_boost_factor
156                  * operations, then restored back to its original prio, and so
157                  * forth.
158                  *
159                  * When @trsp is nil, we want to force-reset the task for
160                  * stopping the kthread.
161                  */
162                 if (!trsp || !(torture_random(trsp) %
163                                (cxt.nrealwriters_stress * factor * 2))) {
164                         sched_set_normal(current, 0);
165                 } else /* common case, do nothing */
166                         return;
167         }
168 }
169
170 static void torture_rt_boost(struct torture_random_state *trsp)
171 {
172         if (rt_boost != 2)
173                 return;
174
175         __torture_rt_boost(trsp);
176 }
177
178 static struct lock_torture_ops lock_busted_ops = {
179         .writelock      = torture_lock_busted_write_lock,
180         .write_delay    = torture_lock_busted_write_delay,
181         .task_boost     = torture_rt_boost,
182         .writeunlock    = torture_lock_busted_write_unlock,
183         .readlock       = NULL,
184         .read_delay     = NULL,
185         .readunlock     = NULL,
186         .name           = "lock_busted"
187 };
188
189 static DEFINE_SPINLOCK(torture_spinlock);
190
191 static int torture_spin_lock_write_lock(int tid __maybe_unused)
192 __acquires(torture_spinlock)
193 {
194         spin_lock(&torture_spinlock);
195         return 0;
196 }
197
198 static void torture_spin_lock_write_delay(struct torture_random_state *trsp)
199 {
200         const unsigned long shortdelay_us = 2;
201         const unsigned long longdelay_ms = 100;
202
203         /* We want a short delay mostly to emulate likely code, and
204          * we want a long delay occasionally to force massive contention.
205          */
206         if (!(torture_random(trsp) %
207               (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
208                 mdelay(longdelay_ms);
209         if (!(torture_random(trsp) %
210               (cxt.nrealwriters_stress * 2 * shortdelay_us)))
211                 udelay(shortdelay_us);
212         if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
213                 torture_preempt_schedule();  /* Allow test to be preempted. */
214 }
215
216 static void torture_spin_lock_write_unlock(int tid __maybe_unused)
217 __releases(torture_spinlock)
218 {
219         spin_unlock(&torture_spinlock);
220 }
221
222 static struct lock_torture_ops spin_lock_ops = {
223         .writelock      = torture_spin_lock_write_lock,
224         .write_delay    = torture_spin_lock_write_delay,
225         .task_boost     = torture_rt_boost,
226         .writeunlock    = torture_spin_lock_write_unlock,
227         .readlock       = NULL,
228         .read_delay     = NULL,
229         .readunlock     = NULL,
230         .name           = "spin_lock"
231 };
232
233 static int torture_spin_lock_write_lock_irq(int tid __maybe_unused)
234 __acquires(torture_spinlock)
235 {
236         unsigned long flags;
237
238         spin_lock_irqsave(&torture_spinlock, flags);
239         cxt.cur_ops->flags = flags;
240         return 0;
241 }
242
243 static void torture_lock_spin_write_unlock_irq(int tid __maybe_unused)
244 __releases(torture_spinlock)
245 {
246         spin_unlock_irqrestore(&torture_spinlock, cxt.cur_ops->flags);
247 }
248
249 static struct lock_torture_ops spin_lock_irq_ops = {
250         .writelock      = torture_spin_lock_write_lock_irq,
251         .write_delay    = torture_spin_lock_write_delay,
252         .task_boost     = torture_rt_boost,
253         .writeunlock    = torture_lock_spin_write_unlock_irq,
254         .readlock       = NULL,
255         .read_delay     = NULL,
256         .readunlock     = NULL,
257         .name           = "spin_lock_irq"
258 };
259
260 static DEFINE_RAW_SPINLOCK(torture_raw_spinlock);
261
262 static int torture_raw_spin_lock_write_lock(int tid __maybe_unused)
263 __acquires(torture_raw_spinlock)
264 {
265         raw_spin_lock(&torture_raw_spinlock);
266         return 0;
267 }
268
269 static void torture_raw_spin_lock_write_unlock(int tid __maybe_unused)
270 __releases(torture_raw_spinlock)
271 {
272         raw_spin_unlock(&torture_raw_spinlock);
273 }
274
275 static struct lock_torture_ops raw_spin_lock_ops = {
276         .writelock      = torture_raw_spin_lock_write_lock,
277         .write_delay    = torture_spin_lock_write_delay,
278         .task_boost     = torture_rt_boost,
279         .writeunlock    = torture_raw_spin_lock_write_unlock,
280         .readlock       = NULL,
281         .read_delay     = NULL,
282         .readunlock     = NULL,
283         .name           = "raw_spin_lock"
284 };
285
286 static int torture_raw_spin_lock_write_lock_irq(int tid __maybe_unused)
287 __acquires(torture_raw_spinlock)
288 {
289         unsigned long flags;
290
291         raw_spin_lock_irqsave(&torture_raw_spinlock, flags);
292         cxt.cur_ops->flags = flags;
293         return 0;
294 }
295
296 static void torture_raw_spin_lock_write_unlock_irq(int tid __maybe_unused)
297 __releases(torture_raw_spinlock)
298 {
299         raw_spin_unlock_irqrestore(&torture_raw_spinlock, cxt.cur_ops->flags);
300 }
301
302 static struct lock_torture_ops raw_spin_lock_irq_ops = {
303         .writelock      = torture_raw_spin_lock_write_lock_irq,
304         .write_delay    = torture_spin_lock_write_delay,
305         .task_boost     = torture_rt_boost,
306         .writeunlock    = torture_raw_spin_lock_write_unlock_irq,
307         .readlock       = NULL,
308         .read_delay     = NULL,
309         .readunlock     = NULL,
310         .name           = "raw_spin_lock_irq"
311 };
312
313 static DEFINE_RWLOCK(torture_rwlock);
314
315 static int torture_rwlock_write_lock(int tid __maybe_unused)
316 __acquires(torture_rwlock)
317 {
318         write_lock(&torture_rwlock);
319         return 0;
320 }
321
322 static void torture_rwlock_write_delay(struct torture_random_state *trsp)
323 {
324         const unsigned long shortdelay_us = 2;
325         const unsigned long longdelay_ms = 100;
326
327         /* We want a short delay mostly to emulate likely code, and
328          * we want a long delay occasionally to force massive contention.
329          */
330         if (!(torture_random(trsp) %
331               (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
332                 mdelay(longdelay_ms);
333         else
334                 udelay(shortdelay_us);
335 }
336
337 static void torture_rwlock_write_unlock(int tid __maybe_unused)
338 __releases(torture_rwlock)
339 {
340         write_unlock(&torture_rwlock);
341 }
342
343 static int torture_rwlock_read_lock(int tid __maybe_unused)
344 __acquires(torture_rwlock)
345 {
346         read_lock(&torture_rwlock);
347         return 0;
348 }
349
350 static void torture_rwlock_read_delay(struct torture_random_state *trsp)
351 {
352         const unsigned long shortdelay_us = 10;
353         const unsigned long longdelay_ms = 100;
354
355         /* We want a short delay mostly to emulate likely code, and
356          * we want a long delay occasionally to force massive contention.
357          */
358         if (!(torture_random(trsp) %
359               (cxt.nrealreaders_stress * 2000 * longdelay_ms)))
360                 mdelay(longdelay_ms);
361         else
362                 udelay(shortdelay_us);
363 }
364
365 static void torture_rwlock_read_unlock(int tid __maybe_unused)
366 __releases(torture_rwlock)
367 {
368         read_unlock(&torture_rwlock);
369 }
370
371 static struct lock_torture_ops rw_lock_ops = {
372         .writelock      = torture_rwlock_write_lock,
373         .write_delay    = torture_rwlock_write_delay,
374         .task_boost     = torture_rt_boost,
375         .writeunlock    = torture_rwlock_write_unlock,
376         .readlock       = torture_rwlock_read_lock,
377         .read_delay     = torture_rwlock_read_delay,
378         .readunlock     = torture_rwlock_read_unlock,
379         .name           = "rw_lock"
380 };
381
382 static int torture_rwlock_write_lock_irq(int tid __maybe_unused)
383 __acquires(torture_rwlock)
384 {
385         unsigned long flags;
386
387         write_lock_irqsave(&torture_rwlock, flags);
388         cxt.cur_ops->flags = flags;
389         return 0;
390 }
391
392 static void torture_rwlock_write_unlock_irq(int tid __maybe_unused)
393 __releases(torture_rwlock)
394 {
395         write_unlock_irqrestore(&torture_rwlock, cxt.cur_ops->flags);
396 }
397
398 static int torture_rwlock_read_lock_irq(int tid __maybe_unused)
399 __acquires(torture_rwlock)
400 {
401         unsigned long flags;
402
403         read_lock_irqsave(&torture_rwlock, flags);
404         cxt.cur_ops->flags = flags;
405         return 0;
406 }
407
408 static void torture_rwlock_read_unlock_irq(int tid __maybe_unused)
409 __releases(torture_rwlock)
410 {
411         read_unlock_irqrestore(&torture_rwlock, cxt.cur_ops->flags);
412 }
413
414 static struct lock_torture_ops rw_lock_irq_ops = {
415         .writelock      = torture_rwlock_write_lock_irq,
416         .write_delay    = torture_rwlock_write_delay,
417         .task_boost     = torture_rt_boost,
418         .writeunlock    = torture_rwlock_write_unlock_irq,
419         .readlock       = torture_rwlock_read_lock_irq,
420         .read_delay     = torture_rwlock_read_delay,
421         .readunlock     = torture_rwlock_read_unlock_irq,
422         .name           = "rw_lock_irq"
423 };
424
425 static DEFINE_MUTEX(torture_mutex);
426 static struct mutex torture_nested_mutexes[MAX_NESTED_LOCKS];
427 static struct lock_class_key nested_mutex_keys[MAX_NESTED_LOCKS];
428
429 static void torture_mutex_init(void)
430 {
431         int i;
432
433         for (i = 0; i < MAX_NESTED_LOCKS; i++)
434                 __mutex_init(&torture_nested_mutexes[i], __func__,
435                              &nested_mutex_keys[i]);
436 }
437
438 static int torture_mutex_nested_lock(int tid __maybe_unused,
439                                      u32 lockset)
440 {
441         int i;
442
443         for (i = 0; i < nested_locks; i++)
444                 if (lockset & (1 << i))
445                         mutex_lock(&torture_nested_mutexes[i]);
446         return 0;
447 }
448
449 static int torture_mutex_lock(int tid __maybe_unused)
450 __acquires(torture_mutex)
451 {
452         mutex_lock(&torture_mutex);
453         return 0;
454 }
455
456 static void torture_mutex_delay(struct torture_random_state *trsp)
457 {
458         const unsigned long longdelay_ms = 100;
459
460         /* We want a long delay occasionally to force massive contention.  */
461         if (!(torture_random(trsp) %
462               (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
463                 mdelay(longdelay_ms * 5);
464         else
465                 mdelay(longdelay_ms / 5);
466         if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
467                 torture_preempt_schedule();  /* Allow test to be preempted. */
468 }
469
470 static void torture_mutex_unlock(int tid __maybe_unused)
471 __releases(torture_mutex)
472 {
473         mutex_unlock(&torture_mutex);
474 }
475
476 static void torture_mutex_nested_unlock(int tid __maybe_unused,
477                                         u32 lockset)
478 {
479         int i;
480
481         for (i = nested_locks - 1; i >= 0; i--)
482                 if (lockset & (1 << i))
483                         mutex_unlock(&torture_nested_mutexes[i]);
484 }
485
486 static struct lock_torture_ops mutex_lock_ops = {
487         .init           = torture_mutex_init,
488         .nested_lock    = torture_mutex_nested_lock,
489         .writelock      = torture_mutex_lock,
490         .write_delay    = torture_mutex_delay,
491         .task_boost     = torture_rt_boost,
492         .writeunlock    = torture_mutex_unlock,
493         .nested_unlock  = torture_mutex_nested_unlock,
494         .readlock       = NULL,
495         .read_delay     = NULL,
496         .readunlock     = NULL,
497         .name           = "mutex_lock"
498 };
499
500 #include <linux/ww_mutex.h>
501 /*
502  * The torture ww_mutexes should belong to the same lock class as
503  * torture_ww_class to avoid lockdep problem. The ww_mutex_init()
504  * function is called for initialization to ensure that.
505  */
506 static DEFINE_WD_CLASS(torture_ww_class);
507 static struct ww_mutex torture_ww_mutex_0, torture_ww_mutex_1, torture_ww_mutex_2;
508 static struct ww_acquire_ctx *ww_acquire_ctxs;
509
510 static void torture_ww_mutex_init(void)
511 {
512         ww_mutex_init(&torture_ww_mutex_0, &torture_ww_class);
513         ww_mutex_init(&torture_ww_mutex_1, &torture_ww_class);
514         ww_mutex_init(&torture_ww_mutex_2, &torture_ww_class);
515
516         ww_acquire_ctxs = kmalloc_array(cxt.nrealwriters_stress,
517                                         sizeof(*ww_acquire_ctxs),
518                                         GFP_KERNEL);
519         if (!ww_acquire_ctxs)
520                 VERBOSE_TOROUT_STRING("ww_acquire_ctx: Out of memory");
521 }
522
523 static void torture_ww_mutex_exit(void)
524 {
525         kfree(ww_acquire_ctxs);
526 }
527
528 static int torture_ww_mutex_lock(int tid)
529 __acquires(torture_ww_mutex_0)
530 __acquires(torture_ww_mutex_1)
531 __acquires(torture_ww_mutex_2)
532 {
533         LIST_HEAD(list);
534         struct reorder_lock {
535                 struct list_head link;
536                 struct ww_mutex *lock;
537         } locks[3], *ll, *ln;
538         struct ww_acquire_ctx *ctx = &ww_acquire_ctxs[tid];
539
540         locks[0].lock = &torture_ww_mutex_0;
541         list_add(&locks[0].link, &list);
542
543         locks[1].lock = &torture_ww_mutex_1;
544         list_add(&locks[1].link, &list);
545
546         locks[2].lock = &torture_ww_mutex_2;
547         list_add(&locks[2].link, &list);
548
549         ww_acquire_init(ctx, &torture_ww_class);
550
551         list_for_each_entry(ll, &list, link) {
552                 int err;
553
554                 err = ww_mutex_lock(ll->lock, ctx);
555                 if (!err)
556                         continue;
557
558                 ln = ll;
559                 list_for_each_entry_continue_reverse(ln, &list, link)
560                         ww_mutex_unlock(ln->lock);
561
562                 if (err != -EDEADLK)
563                         return err;
564
565                 ww_mutex_lock_slow(ll->lock, ctx);
566                 list_move(&ll->link, &list);
567         }
568
569         return 0;
570 }
571
572 static void torture_ww_mutex_unlock(int tid)
573 __releases(torture_ww_mutex_0)
574 __releases(torture_ww_mutex_1)
575 __releases(torture_ww_mutex_2)
576 {
577         struct ww_acquire_ctx *ctx = &ww_acquire_ctxs[tid];
578
579         ww_mutex_unlock(&torture_ww_mutex_0);
580         ww_mutex_unlock(&torture_ww_mutex_1);
581         ww_mutex_unlock(&torture_ww_mutex_2);
582         ww_acquire_fini(ctx);
583 }
584
585 static struct lock_torture_ops ww_mutex_lock_ops = {
586         .init           = torture_ww_mutex_init,
587         .exit           = torture_ww_mutex_exit,
588         .writelock      = torture_ww_mutex_lock,
589         .write_delay    = torture_mutex_delay,
590         .task_boost     = torture_rt_boost,
591         .writeunlock    = torture_ww_mutex_unlock,
592         .readlock       = NULL,
593         .read_delay     = NULL,
594         .readunlock     = NULL,
595         .name           = "ww_mutex_lock"
596 };
597
598 #ifdef CONFIG_RT_MUTEXES
599 static DEFINE_RT_MUTEX(torture_rtmutex);
600 static struct rt_mutex torture_nested_rtmutexes[MAX_NESTED_LOCKS];
601 static struct lock_class_key nested_rtmutex_keys[MAX_NESTED_LOCKS];
602
603 static void torture_rtmutex_init(void)
604 {
605         int i;
606
607         for (i = 0; i < MAX_NESTED_LOCKS; i++)
608                 __rt_mutex_init(&torture_nested_rtmutexes[i], __func__,
609                                 &nested_rtmutex_keys[i]);
610 }
611
612 static int torture_rtmutex_nested_lock(int tid __maybe_unused,
613                                        u32 lockset)
614 {
615         int i;
616
617         for (i = 0; i < nested_locks; i++)
618                 if (lockset & (1 << i))
619                         rt_mutex_lock(&torture_nested_rtmutexes[i]);
620         return 0;
621 }
622
623 static int torture_rtmutex_lock(int tid __maybe_unused)
624 __acquires(torture_rtmutex)
625 {
626         rt_mutex_lock(&torture_rtmutex);
627         return 0;
628 }
629
630 static void torture_rtmutex_delay(struct torture_random_state *trsp)
631 {
632         const unsigned long shortdelay_us = 2;
633         const unsigned long longdelay_ms = 100;
634
635         /*
636          * We want a short delay mostly to emulate likely code, and
637          * we want a long delay occasionally to force massive contention.
638          */
639         if (!(torture_random(trsp) %
640               (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
641                 mdelay(longdelay_ms);
642         if (!(torture_random(trsp) %
643               (cxt.nrealwriters_stress * 2 * shortdelay_us)))
644                 udelay(shortdelay_us);
645         if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
646                 torture_preempt_schedule();  /* Allow test to be preempted. */
647 }
648
649 static void torture_rtmutex_unlock(int tid __maybe_unused)
650 __releases(torture_rtmutex)
651 {
652         rt_mutex_unlock(&torture_rtmutex);
653 }
654
655 static void torture_rt_boost_rtmutex(struct torture_random_state *trsp)
656 {
657         if (!rt_boost)
658                 return;
659
660         __torture_rt_boost(trsp);
661 }
662
663 static void torture_rtmutex_nested_unlock(int tid __maybe_unused,
664                                           u32 lockset)
665 {
666         int i;
667
668         for (i = nested_locks - 1; i >= 0; i--)
669                 if (lockset & (1 << i))
670                         rt_mutex_unlock(&torture_nested_rtmutexes[i]);
671 }
672
673 static struct lock_torture_ops rtmutex_lock_ops = {
674         .init           = torture_rtmutex_init,
675         .nested_lock    = torture_rtmutex_nested_lock,
676         .writelock      = torture_rtmutex_lock,
677         .write_delay    = torture_rtmutex_delay,
678         .task_boost     = torture_rt_boost_rtmutex,
679         .writeunlock    = torture_rtmutex_unlock,
680         .nested_unlock  = torture_rtmutex_nested_unlock,
681         .readlock       = NULL,
682         .read_delay     = NULL,
683         .readunlock     = NULL,
684         .name           = "rtmutex_lock"
685 };
686 #endif
687
688 static DECLARE_RWSEM(torture_rwsem);
689 static int torture_rwsem_down_write(int tid __maybe_unused)
690 __acquires(torture_rwsem)
691 {
692         down_write(&torture_rwsem);
693         return 0;
694 }
695
696 static void torture_rwsem_write_delay(struct torture_random_state *trsp)
697 {
698         const unsigned long longdelay_ms = 100;
699
700         /* We want a long delay occasionally to force massive contention.  */
701         if (!(torture_random(trsp) %
702               (cxt.nrealwriters_stress * 2000 * longdelay_ms)))
703                 mdelay(longdelay_ms * 10);
704         else
705                 mdelay(longdelay_ms / 10);
706         if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
707                 torture_preempt_schedule();  /* Allow test to be preempted. */
708 }
709
710 static void torture_rwsem_up_write(int tid __maybe_unused)
711 __releases(torture_rwsem)
712 {
713         up_write(&torture_rwsem);
714 }
715
716 static int torture_rwsem_down_read(int tid __maybe_unused)
717 __acquires(torture_rwsem)
718 {
719         down_read(&torture_rwsem);
720         return 0;
721 }
722
723 static void torture_rwsem_read_delay(struct torture_random_state *trsp)
724 {
725         const unsigned long longdelay_ms = 100;
726
727         /* We want a long delay occasionally to force massive contention.  */
728         if (!(torture_random(trsp) %
729               (cxt.nrealreaders_stress * 2000 * longdelay_ms)))
730                 mdelay(longdelay_ms * 2);
731         else
732                 mdelay(longdelay_ms / 2);
733         if (!(torture_random(trsp) % (cxt.nrealreaders_stress * 20000)))
734                 torture_preempt_schedule();  /* Allow test to be preempted. */
735 }
736
737 static void torture_rwsem_up_read(int tid __maybe_unused)
738 __releases(torture_rwsem)
739 {
740         up_read(&torture_rwsem);
741 }
742
743 static struct lock_torture_ops rwsem_lock_ops = {
744         .writelock      = torture_rwsem_down_write,
745         .write_delay    = torture_rwsem_write_delay,
746         .task_boost     = torture_rt_boost,
747         .writeunlock    = torture_rwsem_up_write,
748         .readlock       = torture_rwsem_down_read,
749         .read_delay     = torture_rwsem_read_delay,
750         .readunlock     = torture_rwsem_up_read,
751         .name           = "rwsem_lock"
752 };
753
754 #include <linux/percpu-rwsem.h>
755 static struct percpu_rw_semaphore pcpu_rwsem;
756
757 static void torture_percpu_rwsem_init(void)
758 {
759         BUG_ON(percpu_init_rwsem(&pcpu_rwsem));
760 }
761
762 static void torture_percpu_rwsem_exit(void)
763 {
764         percpu_free_rwsem(&pcpu_rwsem);
765 }
766
767 static int torture_percpu_rwsem_down_write(int tid __maybe_unused)
768 __acquires(pcpu_rwsem)
769 {
770         percpu_down_write(&pcpu_rwsem);
771         return 0;
772 }
773
774 static void torture_percpu_rwsem_up_write(int tid __maybe_unused)
775 __releases(pcpu_rwsem)
776 {
777         percpu_up_write(&pcpu_rwsem);
778 }
779
780 static int torture_percpu_rwsem_down_read(int tid __maybe_unused)
781 __acquires(pcpu_rwsem)
782 {
783         percpu_down_read(&pcpu_rwsem);
784         return 0;
785 }
786
787 static void torture_percpu_rwsem_up_read(int tid __maybe_unused)
788 __releases(pcpu_rwsem)
789 {
790         percpu_up_read(&pcpu_rwsem);
791 }
792
793 static struct lock_torture_ops percpu_rwsem_lock_ops = {
794         .init           = torture_percpu_rwsem_init,
795         .exit           = torture_percpu_rwsem_exit,
796         .writelock      = torture_percpu_rwsem_down_write,
797         .write_delay    = torture_rwsem_write_delay,
798         .task_boost     = torture_rt_boost,
799         .writeunlock    = torture_percpu_rwsem_up_write,
800         .readlock       = torture_percpu_rwsem_down_read,
801         .read_delay     = torture_rwsem_read_delay,
802         .readunlock     = torture_percpu_rwsem_up_read,
803         .name           = "percpu_rwsem_lock"
804 };
805
806 /*
807  * Lock torture writer kthread.  Repeatedly acquires and releases
808  * the lock, checking for duplicate acquisitions.
809  */
810 static int lock_torture_writer(void *arg)
811 {
812         struct lock_stress_stats *lwsp = arg;
813         int tid = lwsp - cxt.lwsa;
814         DEFINE_TORTURE_RANDOM(rand);
815         u32 lockset_mask;
816         bool skip_main_lock;
817
818         VERBOSE_TOROUT_STRING("lock_torture_writer task started");
819         set_user_nice(current, MAX_NICE);
820
821         do {
822                 if ((torture_random(&rand) & 0xfffff) == 0)
823                         schedule_timeout_uninterruptible(1);
824
825                 lockset_mask = torture_random(&rand);
826                 /*
827                  * When using nested_locks, we want to occasionally
828                  * skip the main lock so we can avoid always serializing
829                  * the lock chains on that central lock. By skipping the
830                  * main lock occasionally, we can create different
831                  * contention patterns (allowing for multiple disjoint
832                  * blocked trees)
833                  */
834                 skip_main_lock = (nested_locks &&
835                                  !(torture_random(&rand) % 100));
836
837                 cxt.cur_ops->task_boost(&rand);
838                 if (cxt.cur_ops->nested_lock)
839                         cxt.cur_ops->nested_lock(tid, lockset_mask);
840
841                 if (!skip_main_lock) {
842                         cxt.cur_ops->writelock(tid);
843                         if (WARN_ON_ONCE(lock_is_write_held))
844                                 lwsp->n_lock_fail++;
845                         lock_is_write_held = true;
846                         if (WARN_ON_ONCE(atomic_read(&lock_is_read_held)))
847                                 lwsp->n_lock_fail++; /* rare, but... */
848
849                         lwsp->n_lock_acquired++;
850                 }
851                 cxt.cur_ops->write_delay(&rand);
852                 if (!skip_main_lock) {
853                         lock_is_write_held = false;
854                         WRITE_ONCE(last_lock_release, jiffies);
855                         cxt.cur_ops->writeunlock(tid);
856                 }
857                 if (cxt.cur_ops->nested_unlock)
858                         cxt.cur_ops->nested_unlock(tid, lockset_mask);
859
860                 stutter_wait("lock_torture_writer");
861         } while (!torture_must_stop());
862
863         cxt.cur_ops->task_boost(NULL); /* reset prio */
864         torture_kthread_stopping("lock_torture_writer");
865         return 0;
866 }
867
868 /*
869  * Lock torture reader kthread.  Repeatedly acquires and releases
870  * the reader lock.
871  */
872 static int lock_torture_reader(void *arg)
873 {
874         struct lock_stress_stats *lrsp = arg;
875         int tid = lrsp - cxt.lrsa;
876         DEFINE_TORTURE_RANDOM(rand);
877
878         VERBOSE_TOROUT_STRING("lock_torture_reader task started");
879         set_user_nice(current, MAX_NICE);
880
881         do {
882                 if ((torture_random(&rand) & 0xfffff) == 0)
883                         schedule_timeout_uninterruptible(1);
884
885                 cxt.cur_ops->readlock(tid);
886                 atomic_inc(&lock_is_read_held);
887                 if (WARN_ON_ONCE(lock_is_write_held))
888                         lrsp->n_lock_fail++; /* rare, but... */
889
890                 lrsp->n_lock_acquired++;
891                 cxt.cur_ops->read_delay(&rand);
892                 atomic_dec(&lock_is_read_held);
893                 cxt.cur_ops->readunlock(tid);
894
895                 stutter_wait("lock_torture_reader");
896         } while (!torture_must_stop());
897         torture_kthread_stopping("lock_torture_reader");
898         return 0;
899 }
900
901 /*
902  * Create an lock-torture-statistics message in the specified buffer.
903  */
904 static void __torture_print_stats(char *page,
905                                   struct lock_stress_stats *statp, bool write)
906 {
907         long cur;
908         bool fail = false;
909         int i, n_stress;
910         long max = 0, min = statp ? data_race(statp[0].n_lock_acquired) : 0;
911         long long sum = 0;
912
913         n_stress = write ? cxt.nrealwriters_stress : cxt.nrealreaders_stress;
914         for (i = 0; i < n_stress; i++) {
915                 if (data_race(statp[i].n_lock_fail))
916                         fail = true;
917                 cur = data_race(statp[i].n_lock_acquired);
918                 sum += cur;
919                 if (max < cur)
920                         max = cur;
921                 if (min > cur)
922                         min = cur;
923         }
924         page += sprintf(page,
925                         "%s:  Total: %lld  Max/Min: %ld/%ld %s  Fail: %d %s\n",
926                         write ? "Writes" : "Reads ",
927                         sum, max, min,
928                         !onoff_interval && max / 2 > min ? "???" : "",
929                         fail, fail ? "!!!" : "");
930         if (fail)
931                 atomic_inc(&cxt.n_lock_torture_errors);
932 }
933
934 /*
935  * Print torture statistics.  Caller must ensure that there is only one
936  * call to this function at a given time!!!  This is normally accomplished
937  * by relying on the module system to only have one copy of the module
938  * loaded, and then by giving the lock_torture_stats kthread full control
939  * (or the init/cleanup functions when lock_torture_stats thread is not
940  * running).
941  */
942 static void lock_torture_stats_print(void)
943 {
944         int size = cxt.nrealwriters_stress * 200 + 8192;
945         char *buf;
946
947         if (cxt.cur_ops->readlock)
948                 size += cxt.nrealreaders_stress * 200 + 8192;
949
950         buf = kmalloc(size, GFP_KERNEL);
951         if (!buf) {
952                 pr_err("lock_torture_stats_print: Out of memory, need: %d",
953                        size);
954                 return;
955         }
956
957         __torture_print_stats(buf, cxt.lwsa, true);
958         pr_alert("%s", buf);
959         kfree(buf);
960
961         if (cxt.cur_ops->readlock) {
962                 buf = kmalloc(size, GFP_KERNEL);
963                 if (!buf) {
964                         pr_err("lock_torture_stats_print: Out of memory, need: %d",
965                                size);
966                         return;
967                 }
968
969                 __torture_print_stats(buf, cxt.lrsa, false);
970                 pr_alert("%s", buf);
971                 kfree(buf);
972         }
973 }
974
975 /*
976  * Periodically prints torture statistics, if periodic statistics printing
977  * was specified via the stat_interval module parameter.
978  *
979  * No need to worry about fullstop here, since this one doesn't reference
980  * volatile state or register callbacks.
981  */
982 static int lock_torture_stats(void *arg)
983 {
984         VERBOSE_TOROUT_STRING("lock_torture_stats task started");
985         do {
986                 schedule_timeout_interruptible(stat_interval * HZ);
987                 lock_torture_stats_print();
988                 torture_shutdown_absorb("lock_torture_stats");
989         } while (!torture_must_stop());
990         torture_kthread_stopping("lock_torture_stats");
991         return 0;
992 }
993
994 static inline void
995 lock_torture_print_module_parms(struct lock_torture_ops *cur_ops,
996                                 const char *tag)
997 {
998         pr_alert("%s" TORTURE_FLAG
999                  "--- %s%s: nwriters_stress=%d nreaders_stress=%d nested_locks=%d stat_interval=%d verbose=%d shuffle_interval=%d stutter=%d shutdown_secs=%d onoff_interval=%d onoff_holdoff=%d\n",
1000                  torture_type, tag, cxt.debug_lock ? " [debug]": "",
1001                  cxt.nrealwriters_stress, cxt.nrealreaders_stress,
1002                  nested_locks, stat_interval, verbose, shuffle_interval,
1003                  stutter, shutdown_secs, onoff_interval, onoff_holdoff);
1004 }
1005
1006 static void lock_torture_cleanup(void)
1007 {
1008         int i;
1009
1010         if (torture_cleanup_begin())
1011                 return;
1012
1013         /*
1014          * Indicates early cleanup, meaning that the test has not run,
1015          * such as when passing bogus args when loading the module.
1016          * However cxt->cur_ops.init() may have been invoked, so beside
1017          * perform the underlying torture-specific cleanups, cur_ops.exit()
1018          * will be invoked if needed.
1019          */
1020         if (!cxt.lwsa && !cxt.lrsa)
1021                 goto end;
1022
1023         if (writer_tasks) {
1024                 for (i = 0; i < cxt.nrealwriters_stress; i++)
1025                         torture_stop_kthread(lock_torture_writer,
1026                                              writer_tasks[i]);
1027                 kfree(writer_tasks);
1028                 writer_tasks = NULL;
1029         }
1030
1031         if (reader_tasks) {
1032                 for (i = 0; i < cxt.nrealreaders_stress; i++)
1033                         torture_stop_kthread(lock_torture_reader,
1034                                              reader_tasks[i]);
1035                 kfree(reader_tasks);
1036                 reader_tasks = NULL;
1037         }
1038
1039         torture_stop_kthread(lock_torture_stats, stats_task);
1040         lock_torture_stats_print();  /* -After- the stats thread is stopped! */
1041
1042         if (atomic_read(&cxt.n_lock_torture_errors))
1043                 lock_torture_print_module_parms(cxt.cur_ops,
1044                                                 "End of test: FAILURE");
1045         else if (torture_onoff_failures())
1046                 lock_torture_print_module_parms(cxt.cur_ops,
1047                                                 "End of test: LOCK_HOTPLUG");
1048         else
1049                 lock_torture_print_module_parms(cxt.cur_ops,
1050                                                 "End of test: SUCCESS");
1051
1052         kfree(cxt.lwsa);
1053         cxt.lwsa = NULL;
1054         kfree(cxt.lrsa);
1055         cxt.lrsa = NULL;
1056
1057 end:
1058         if (cxt.init_called) {
1059                 if (cxt.cur_ops->exit)
1060                         cxt.cur_ops->exit();
1061                 cxt.init_called = false;
1062         }
1063         torture_cleanup_end();
1064 }
1065
1066 static int __init lock_torture_init(void)
1067 {
1068         int i, j;
1069         int firsterr = 0;
1070         static struct lock_torture_ops *torture_ops[] = {
1071                 &lock_busted_ops,
1072                 &spin_lock_ops, &spin_lock_irq_ops,
1073                 &raw_spin_lock_ops, &raw_spin_lock_irq_ops,
1074                 &rw_lock_ops, &rw_lock_irq_ops,
1075                 &mutex_lock_ops,
1076                 &ww_mutex_lock_ops,
1077 #ifdef CONFIG_RT_MUTEXES
1078                 &rtmutex_lock_ops,
1079 #endif
1080                 &rwsem_lock_ops,
1081                 &percpu_rwsem_lock_ops,
1082         };
1083
1084         if (!torture_init_begin(torture_type, verbose))
1085                 return -EBUSY;
1086
1087         /* Process args and tell the world that the torturer is on the job. */
1088         for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
1089                 cxt.cur_ops = torture_ops[i];
1090                 if (strcmp(torture_type, cxt.cur_ops->name) == 0)
1091                         break;
1092         }
1093         if (i == ARRAY_SIZE(torture_ops)) {
1094                 pr_alert("lock-torture: invalid torture type: \"%s\"\n",
1095                          torture_type);
1096                 pr_alert("lock-torture types:");
1097                 for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
1098                         pr_alert(" %s", torture_ops[i]->name);
1099                 pr_alert("\n");
1100                 firsterr = -EINVAL;
1101                 goto unwind;
1102         }
1103
1104         if (nwriters_stress == 0 &&
1105             (!cxt.cur_ops->readlock || nreaders_stress == 0)) {
1106                 pr_alert("lock-torture: must run at least one locking thread\n");
1107                 firsterr = -EINVAL;
1108                 goto unwind;
1109         }
1110
1111         if (nwriters_stress >= 0)
1112                 cxt.nrealwriters_stress = nwriters_stress;
1113         else
1114                 cxt.nrealwriters_stress = 2 * num_online_cpus();
1115
1116         if (cxt.cur_ops->init) {
1117                 cxt.cur_ops->init();
1118                 cxt.init_called = true;
1119         }
1120
1121 #ifdef CONFIG_DEBUG_MUTEXES
1122         if (str_has_prefix(torture_type, "mutex"))
1123                 cxt.debug_lock = true;
1124 #endif
1125 #ifdef CONFIG_DEBUG_RT_MUTEXES
1126         if (str_has_prefix(torture_type, "rtmutex"))
1127                 cxt.debug_lock = true;
1128 #endif
1129 #ifdef CONFIG_DEBUG_SPINLOCK
1130         if ((str_has_prefix(torture_type, "spin")) ||
1131             (str_has_prefix(torture_type, "rw_lock")))
1132                 cxt.debug_lock = true;
1133 #endif
1134
1135         /* Initialize the statistics so that each run gets its own numbers. */
1136         if (nwriters_stress) {
1137                 lock_is_write_held = false;
1138                 cxt.lwsa = kmalloc_array(cxt.nrealwriters_stress,
1139                                          sizeof(*cxt.lwsa),
1140                                          GFP_KERNEL);
1141                 if (cxt.lwsa == NULL) {
1142                         VERBOSE_TOROUT_STRING("cxt.lwsa: Out of memory");
1143                         firsterr = -ENOMEM;
1144                         goto unwind;
1145                 }
1146
1147                 for (i = 0; i < cxt.nrealwriters_stress; i++) {
1148                         cxt.lwsa[i].n_lock_fail = 0;
1149                         cxt.lwsa[i].n_lock_acquired = 0;
1150                 }
1151         }
1152
1153         if (cxt.cur_ops->readlock) {
1154                 if (nreaders_stress >= 0)
1155                         cxt.nrealreaders_stress = nreaders_stress;
1156                 else {
1157                         /*
1158                          * By default distribute evenly the number of
1159                          * readers and writers. We still run the same number
1160                          * of threads as the writer-only locks default.
1161                          */
1162                         if (nwriters_stress < 0) /* user doesn't care */
1163                                 cxt.nrealwriters_stress = num_online_cpus();
1164                         cxt.nrealreaders_stress = cxt.nrealwriters_stress;
1165                 }
1166
1167                 if (nreaders_stress) {
1168                         cxt.lrsa = kmalloc_array(cxt.nrealreaders_stress,
1169                                                  sizeof(*cxt.lrsa),
1170                                                  GFP_KERNEL);
1171                         if (cxt.lrsa == NULL) {
1172                                 VERBOSE_TOROUT_STRING("cxt.lrsa: Out of memory");
1173                                 firsterr = -ENOMEM;
1174                                 kfree(cxt.lwsa);
1175                                 cxt.lwsa = NULL;
1176                                 goto unwind;
1177                         }
1178
1179                         for (i = 0; i < cxt.nrealreaders_stress; i++) {
1180                                 cxt.lrsa[i].n_lock_fail = 0;
1181                                 cxt.lrsa[i].n_lock_acquired = 0;
1182                         }
1183                 }
1184         }
1185
1186         lock_torture_print_module_parms(cxt.cur_ops, "Start of test");
1187
1188         /* Prepare torture context. */
1189         if (onoff_interval > 0) {
1190                 firsterr = torture_onoff_init(onoff_holdoff * HZ,
1191                                               onoff_interval * HZ, NULL);
1192                 if (torture_init_error(firsterr))
1193                         goto unwind;
1194         }
1195         if (shuffle_interval > 0) {
1196                 firsterr = torture_shuffle_init(shuffle_interval);
1197                 if (torture_init_error(firsterr))
1198                         goto unwind;
1199         }
1200         if (shutdown_secs > 0) {
1201                 firsterr = torture_shutdown_init(shutdown_secs,
1202                                                  lock_torture_cleanup);
1203                 if (torture_init_error(firsterr))
1204                         goto unwind;
1205         }
1206         if (stutter > 0) {
1207                 firsterr = torture_stutter_init(stutter, stutter);
1208                 if (torture_init_error(firsterr))
1209                         goto unwind;
1210         }
1211
1212         if (nwriters_stress) {
1213                 writer_tasks = kcalloc(cxt.nrealwriters_stress,
1214                                        sizeof(writer_tasks[0]),
1215                                        GFP_KERNEL);
1216                 if (writer_tasks == NULL) {
1217                         TOROUT_ERRSTRING("writer_tasks: Out of memory");
1218                         firsterr = -ENOMEM;
1219                         goto unwind;
1220                 }
1221         }
1222
1223         /* cap nested_locks to MAX_NESTED_LOCKS */
1224         if (nested_locks > MAX_NESTED_LOCKS)
1225                 nested_locks = MAX_NESTED_LOCKS;
1226
1227         if (cxt.cur_ops->readlock) {
1228                 reader_tasks = kcalloc(cxt.nrealreaders_stress,
1229                                        sizeof(reader_tasks[0]),
1230                                        GFP_KERNEL);
1231                 if (reader_tasks == NULL) {
1232                         TOROUT_ERRSTRING("reader_tasks: Out of memory");
1233                         kfree(writer_tasks);
1234                         writer_tasks = NULL;
1235                         firsterr = -ENOMEM;
1236                         goto unwind;
1237                 }
1238         }
1239
1240         /*
1241          * Create the kthreads and start torturing (oh, those poor little locks).
1242          *
1243          * TODO: Note that we interleave writers with readers, giving writers a
1244          * slight advantage, by creating its kthread first. This can be modified
1245          * for very specific needs, or even let the user choose the policy, if
1246          * ever wanted.
1247          */
1248         for (i = 0, j = 0; i < cxt.nrealwriters_stress ||
1249                     j < cxt.nrealreaders_stress; i++, j++) {
1250                 if (i >= cxt.nrealwriters_stress)
1251                         goto create_reader;
1252
1253                 /* Create writer. */
1254                 firsterr = torture_create_kthread(lock_torture_writer, &cxt.lwsa[i],
1255                                                   writer_tasks[i]);
1256                 if (torture_init_error(firsterr))
1257                         goto unwind;
1258
1259         create_reader:
1260                 if (cxt.cur_ops->readlock == NULL || (j >= cxt.nrealreaders_stress))
1261                         continue;
1262                 /* Create reader. */
1263                 firsterr = torture_create_kthread(lock_torture_reader, &cxt.lrsa[j],
1264                                                   reader_tasks[j]);
1265                 if (torture_init_error(firsterr))
1266                         goto unwind;
1267         }
1268         if (stat_interval > 0) {
1269                 firsterr = torture_create_kthread(lock_torture_stats, NULL,
1270                                                   stats_task);
1271                 if (torture_init_error(firsterr))
1272                         goto unwind;
1273         }
1274         torture_init_end();
1275         return 0;
1276
1277 unwind:
1278         torture_init_end();
1279         lock_torture_cleanup();
1280         if (shutdown_secs) {
1281                 WARN_ON(!IS_MODULE(CONFIG_LOCK_TORTURE_TEST));
1282                 kernel_power_off();
1283         }
1284         return firsterr;
1285 }
1286
1287 module_init(lock_torture_init);
1288 module_exit(lock_torture_cleanup);