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