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
  33MODULE_LICENSE("GPL");
  34MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
  35
  36torture_param(int, acq_writer_lim, 0, "Write_acquisition time limit (jiffies).");
  37torture_param(int, call_rcu_chains, 0, "Self-propagate call_rcu() chains during test (0=disable).");
  38torture_param(int, long_hold, 100, "Do occasional long hold of lock (ms), 0=disable");
  39torture_param(int, nested_locks, 0, "Number of nested locks (max = 8)");
  40torture_param(int, nreaders_stress, -1, "Number of read-locking stress-test threads");
  41torture_param(int, nwriters_stress, -1, "Number of write-locking stress-test threads");
  42torture_param(int, onoff_holdoff, 0, "Time after boot before CPU hotplugs (s)");
  43torture_param(int, onoff_interval, 0, "Time between CPU hotplugs (s), 0=disable");
  44torture_param(int, rt_boost, 2,
  45		   "Do periodic rt-boost. 0=Disable, 1=Only for rt_mutex, 2=For all lock types.");
  46torture_param(int, rt_boost_factor, 50, "A factor determining how often rt-boost happens.");
  47torture_param(int, shuffle_interval, 3, "Number of jiffies between shuffles, 0=disable");
  48torture_param(int, shutdown_secs, 0, "Shutdown time (j), <= zero to disable.");
  49torture_param(int, stat_interval, 60, "Number of seconds between stats printk()s");
  50torture_param(int, stutter, 5, "Number of jiffies to run/halt test, 0=disable");
  51torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
  52torture_param(int, writer_fifo, 0, "Run writers at sched_set_fifo() priority");
  53/* Going much higher trips "BUG: MAX_LOCKDEP_CHAIN_HLOCKS too low!" errors */
  54#define MAX_NESTED_LOCKS 8
  55
  56static char *torture_type = IS_ENABLED(CONFIG_PREEMPT_RT) ? "raw_spin_lock" : "spin_lock";
  57module_param(torture_type, charp, 0444);
  58MODULE_PARM_DESC(torture_type,
  59		 "Type of lock to torture (spin_lock, spin_lock_irq, mutex_lock, ...)");
  60
  61static cpumask_var_t bind_readers; // Bind the readers to the specified set of CPUs.
  62static cpumask_var_t bind_writers; // Bind the writers to the specified set of CPUs.
  63
  64// Parse a cpumask kernel parameter.  If there are more users later on,
  65// this might need to got to a more central location.
  66static int param_set_cpumask(const char *val, const struct kernel_param *kp)
  67{
  68	cpumask_var_t *cm_bind = kp->arg;
  69	int ret;
  70	char *s;
  71
  72	if (!alloc_cpumask_var(cm_bind, GFP_KERNEL)) {
  73		s = "Out of memory";
  74		ret = -ENOMEM;
  75		goto out_err;
  76	}
  77	ret = cpulist_parse(val, *cm_bind);
  78	if (!ret)
  79		return ret;
  80	s = "Bad CPU range";
  81out_err:
  82	pr_warn("%s: %s, all CPUs set\n", kp->name, s);
  83	cpumask_setall(*cm_bind);
  84	return ret;
  85}
  86
  87// Output a cpumask kernel parameter.
  88static int param_get_cpumask(char *buffer, const struct kernel_param *kp)
  89{
  90	cpumask_var_t *cm_bind = kp->arg;
  91
  92	return sprintf(buffer, "%*pbl", cpumask_pr_args(*cm_bind));
  93}
  94
  95static bool cpumask_nonempty(cpumask_var_t mask)
  96{
  97	return cpumask_available(mask) && !cpumask_empty(mask);
  98}
  99
 100static const struct kernel_param_ops lt_bind_ops = {
 101	.set = param_set_cpumask,
 102	.get = param_get_cpumask,
 103};
 104
 105module_param_cb(bind_readers, &lt_bind_ops, &bind_readers, 0644);
 106module_param_cb(bind_writers, &lt_bind_ops, &bind_writers, 0644);
 107
 108long torture_sched_setaffinity(pid_t pid, const struct cpumask *in_mask);
 109
 110static struct task_struct *stats_task;
 111static struct task_struct **writer_tasks;
 112static struct task_struct **reader_tasks;
 113
 114static bool lock_is_write_held;
 115static atomic_t lock_is_read_held;
 116static unsigned long last_lock_release;
 117
 118struct lock_stress_stats {
 119	long n_lock_fail;
 120	long n_lock_acquired;
 121};
 122
 123struct call_rcu_chain {
 124	struct rcu_head crc_rh;
 125	bool crc_stop;
 126};
 127struct call_rcu_chain *call_rcu_chain_list;
 128
 129/* Forward reference. */
 130static void lock_torture_cleanup(void);
 131
 132/*
 133 * Operations vector for selecting different types of tests.
 134 */
 135struct lock_torture_ops {
 136	void (*init)(void);
 137	void (*exit)(void);
 138	int (*nested_lock)(int tid, u32 lockset);
 139	int (*writelock)(int tid);
 140	void (*write_delay)(struct torture_random_state *trsp);
 141	void (*task_boost)(struct torture_random_state *trsp);
 142	void (*writeunlock)(int tid);
 143	void (*nested_unlock)(int tid, u32 lockset);
 144	int (*readlock)(int tid);
 145	void (*read_delay)(struct torture_random_state *trsp);
 146	void (*readunlock)(int tid);
 147
 148	unsigned long flags; /* for irq spinlocks */
 149	const char *name;
 150};
 151
 152struct lock_torture_cxt {
 153	int nrealwriters_stress;
 154	int nrealreaders_stress;
 155	bool debug_lock;
 156	bool init_called;
 157	atomic_t n_lock_torture_errors;
 158	struct lock_torture_ops *cur_ops;
 159	struct lock_stress_stats *lwsa; /* writer statistics */
 160	struct lock_stress_stats *lrsa; /* reader statistics */
 161};
 162static struct lock_torture_cxt cxt = { 0, 0, false, false,
 163				       ATOMIC_INIT(0),
 164				       NULL, NULL};
 165/*
 166 * Definitions for lock torture testing.
 167 */
 168
 169static int torture_lock_busted_write_lock(int tid __maybe_unused)
 170{
 171	return 0;  /* BUGGY, do not use in real life!!! */
 172}
 173
 174static void torture_lock_busted_write_delay(struct torture_random_state *trsp)
 175{
 176	/* We want a long delay occasionally to force massive contention.  */
 177	if (long_hold && !(torture_random(trsp) % (cxt.nrealwriters_stress * 2000 * long_hold)))
 178		mdelay(long_hold);
 179	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
 180		torture_preempt_schedule();  /* Allow test to be preempted. */
 181}
 182
 183static void torture_lock_busted_write_unlock(int tid __maybe_unused)
 184{
 185	  /* BUGGY, do not use in real life!!! */
 186}
 187
 188static void __torture_rt_boost(struct torture_random_state *trsp)
 189{
 190	const unsigned int factor = rt_boost_factor;
 191
 192	if (!rt_task(current)) {
 193		/*
 194		 * Boost priority once every rt_boost_factor operations. When
 195		 * the task tries to take the lock, the rtmutex it will account
 196		 * for the new priority, and do any corresponding pi-dance.
 197		 */
 198		if (trsp && !(torture_random(trsp) %
 199			      (cxt.nrealwriters_stress * factor))) {
 200			sched_set_fifo(current);
 201		} else /* common case, do nothing */
 202			return;
 203	} else {
 204		/*
 205		 * The task will remain boosted for another 10 * rt_boost_factor
 206		 * operations, then restored back to its original prio, and so
 207		 * forth.
 208		 *
 209		 * When @trsp is nil, we want to force-reset the task for
 210		 * stopping the kthread.
 211		 */
 212		if (!trsp || !(torture_random(trsp) %
 213			       (cxt.nrealwriters_stress * factor * 2))) {
 214			sched_set_normal(current, 0);
 215		} else /* common case, do nothing */
 216			return;
 217	}
 218}
 219
 220static void torture_rt_boost(struct torture_random_state *trsp)
 221{
 222	if (rt_boost != 2)
 223		return;
 224
 225	__torture_rt_boost(trsp);
 226}
 227
 228static struct lock_torture_ops lock_busted_ops = {
 229	.writelock	= torture_lock_busted_write_lock,
 230	.write_delay	= torture_lock_busted_write_delay,
 231	.task_boost     = torture_rt_boost,
 232	.writeunlock	= torture_lock_busted_write_unlock,
 233	.readlock       = NULL,
 234	.read_delay     = NULL,
 235	.readunlock     = NULL,
 236	.name		= "lock_busted"
 237};
 238
 239static DEFINE_SPINLOCK(torture_spinlock);
 240
 241static int torture_spin_lock_write_lock(int tid __maybe_unused)
 242__acquires(torture_spinlock)
 243{
 244	spin_lock(&torture_spinlock);
 245	return 0;
 246}
 247
 248static void torture_spin_lock_write_delay(struct torture_random_state *trsp)
 249{
 250	const unsigned long shortdelay_us = 2;
 251	unsigned long j;
 252
 253	/* We want a short delay mostly to emulate likely code, and
 254	 * we want a long delay occasionally to force massive contention.
 255	 */
 256	if (long_hold && !(torture_random(trsp) % (cxt.nrealwriters_stress * 2000 * long_hold))) {
 257		j = jiffies;
 258		mdelay(long_hold);
 259		pr_alert("%s: delay = %lu jiffies.\n", __func__, jiffies - j);
 260	}
 261	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 200 * shortdelay_us)))
 262		udelay(shortdelay_us);
 263	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
 264		torture_preempt_schedule();  /* Allow test to be preempted. */
 265}
 266
 267static void torture_spin_lock_write_unlock(int tid __maybe_unused)
 268__releases(torture_spinlock)
 269{
 270	spin_unlock(&torture_spinlock);
 271}
 272
 273static struct lock_torture_ops spin_lock_ops = {
 274	.writelock	= torture_spin_lock_write_lock,
 275	.write_delay	= torture_spin_lock_write_delay,
 276	.task_boost     = torture_rt_boost,
 277	.writeunlock	= torture_spin_lock_write_unlock,
 278	.readlock       = NULL,
 279	.read_delay     = NULL,
 280	.readunlock     = NULL,
 281	.name		= "spin_lock"
 282};
 283
 284static int torture_spin_lock_write_lock_irq(int tid __maybe_unused)
 285__acquires(torture_spinlock)
 286{
 287	unsigned long flags;
 288
 289	spin_lock_irqsave(&torture_spinlock, flags);
 290	cxt.cur_ops->flags = flags;
 291	return 0;
 292}
 293
 294static void torture_lock_spin_write_unlock_irq(int tid __maybe_unused)
 295__releases(torture_spinlock)
 296{
 297	spin_unlock_irqrestore(&torture_spinlock, cxt.cur_ops->flags);
 298}
 299
 300static struct lock_torture_ops spin_lock_irq_ops = {
 301	.writelock	= torture_spin_lock_write_lock_irq,
 302	.write_delay	= torture_spin_lock_write_delay,
 303	.task_boost     = torture_rt_boost,
 304	.writeunlock	= torture_lock_spin_write_unlock_irq,
 305	.readlock       = NULL,
 306	.read_delay     = NULL,
 307	.readunlock     = NULL,
 308	.name		= "spin_lock_irq"
 309};
 310
 311static DEFINE_RAW_SPINLOCK(torture_raw_spinlock);
 312
 313static int torture_raw_spin_lock_write_lock(int tid __maybe_unused)
 314__acquires(torture_raw_spinlock)
 315{
 316	raw_spin_lock(&torture_raw_spinlock);
 317	return 0;
 318}
 319
 320static void torture_raw_spin_lock_write_unlock(int tid __maybe_unused)
 321__releases(torture_raw_spinlock)
 322{
 323	raw_spin_unlock(&torture_raw_spinlock);
 324}
 325
 326static struct lock_torture_ops raw_spin_lock_ops = {
 327	.writelock	= torture_raw_spin_lock_write_lock,
 328	.write_delay	= torture_spin_lock_write_delay,
 329	.task_boost	= torture_rt_boost,
 330	.writeunlock	= torture_raw_spin_lock_write_unlock,
 331	.readlock	= NULL,
 332	.read_delay	= NULL,
 333	.readunlock	= NULL,
 334	.name		= "raw_spin_lock"
 335};
 336
 337static int torture_raw_spin_lock_write_lock_irq(int tid __maybe_unused)
 338__acquires(torture_raw_spinlock)
 339{
 340	unsigned long flags;
 341
 342	raw_spin_lock_irqsave(&torture_raw_spinlock, flags);
 343	cxt.cur_ops->flags = flags;
 344	return 0;
 345}
 346
 347static void torture_raw_spin_lock_write_unlock_irq(int tid __maybe_unused)
 348__releases(torture_raw_spinlock)
 349{
 350	raw_spin_unlock_irqrestore(&torture_raw_spinlock, cxt.cur_ops->flags);
 351}
 352
 353static struct lock_torture_ops raw_spin_lock_irq_ops = {
 354	.writelock	= torture_raw_spin_lock_write_lock_irq,
 355	.write_delay	= torture_spin_lock_write_delay,
 356	.task_boost	= torture_rt_boost,
 357	.writeunlock	= torture_raw_spin_lock_write_unlock_irq,
 358	.readlock	= NULL,
 359	.read_delay	= NULL,
 360	.readunlock	= NULL,
 361	.name		= "raw_spin_lock_irq"
 362};
 363
 364static DEFINE_RWLOCK(torture_rwlock);
 365
 366static int torture_rwlock_write_lock(int tid __maybe_unused)
 367__acquires(torture_rwlock)
 368{
 369	write_lock(&torture_rwlock);
 370	return 0;
 371}
 372
 373static void torture_rwlock_write_delay(struct torture_random_state *trsp)
 374{
 375	const unsigned long shortdelay_us = 2;
 376
 377	/* We want a short delay mostly to emulate likely code, and
 378	 * we want a long delay occasionally to force massive contention.
 379	 */
 380	if (long_hold && !(torture_random(trsp) % (cxt.nrealwriters_stress * 2000 * long_hold)))
 381		mdelay(long_hold);
 382	else
 383		udelay(shortdelay_us);
 384}
 385
 386static void torture_rwlock_write_unlock(int tid __maybe_unused)
 387__releases(torture_rwlock)
 388{
 389	write_unlock(&torture_rwlock);
 390}
 391
 392static int torture_rwlock_read_lock(int tid __maybe_unused)
 393__acquires(torture_rwlock)
 394{
 395	read_lock(&torture_rwlock);
 396	return 0;
 397}
 398
 399static void torture_rwlock_read_delay(struct torture_random_state *trsp)
 400{
 401	const unsigned long shortdelay_us = 10;
 402
 403	/* We want a short delay mostly to emulate likely code, and
 404	 * we want a long delay occasionally to force massive contention.
 405	 */
 406	if (long_hold && !(torture_random(trsp) % (cxt.nrealreaders_stress * 2000 * long_hold)))
 407		mdelay(long_hold);
 408	else
 409		udelay(shortdelay_us);
 410}
 411
 412static void torture_rwlock_read_unlock(int tid __maybe_unused)
 413__releases(torture_rwlock)
 414{
 415	read_unlock(&torture_rwlock);
 416}
 417
 418static struct lock_torture_ops rw_lock_ops = {
 419	.writelock	= torture_rwlock_write_lock,
 420	.write_delay	= torture_rwlock_write_delay,
 421	.task_boost     = torture_rt_boost,
 422	.writeunlock	= torture_rwlock_write_unlock,
 423	.readlock       = torture_rwlock_read_lock,
 424	.read_delay     = torture_rwlock_read_delay,
 425	.readunlock     = torture_rwlock_read_unlock,
 426	.name		= "rw_lock"
 427};
 428
 429static int torture_rwlock_write_lock_irq(int tid __maybe_unused)
 430__acquires(torture_rwlock)
 431{
 432	unsigned long flags;
 433
 434	write_lock_irqsave(&torture_rwlock, flags);
 435	cxt.cur_ops->flags = flags;
 436	return 0;
 437}
 438
 439static void torture_rwlock_write_unlock_irq(int tid __maybe_unused)
 440__releases(torture_rwlock)
 441{
 442	write_unlock_irqrestore(&torture_rwlock, cxt.cur_ops->flags);
 443}
 444
 445static int torture_rwlock_read_lock_irq(int tid __maybe_unused)
 446__acquires(torture_rwlock)
 447{
 448	unsigned long flags;
 449
 450	read_lock_irqsave(&torture_rwlock, flags);
 451	cxt.cur_ops->flags = flags;
 452	return 0;
 453}
 454
 455static void torture_rwlock_read_unlock_irq(int tid __maybe_unused)
 456__releases(torture_rwlock)
 457{
 458	read_unlock_irqrestore(&torture_rwlock, cxt.cur_ops->flags);
 459}
 460
 461static struct lock_torture_ops rw_lock_irq_ops = {
 462	.writelock	= torture_rwlock_write_lock_irq,
 463	.write_delay	= torture_rwlock_write_delay,
 464	.task_boost     = torture_rt_boost,
 465	.writeunlock	= torture_rwlock_write_unlock_irq,
 466	.readlock       = torture_rwlock_read_lock_irq,
 467	.read_delay     = torture_rwlock_read_delay,
 468	.readunlock     = torture_rwlock_read_unlock_irq,
 469	.name		= "rw_lock_irq"
 470};
 471
 472static DEFINE_MUTEX(torture_mutex);
 473static struct mutex torture_nested_mutexes[MAX_NESTED_LOCKS];
 474static struct lock_class_key nested_mutex_keys[MAX_NESTED_LOCKS];
 475
 476static void torture_mutex_init(void)
 477{
 478	int i;
 479
 480	for (i = 0; i < MAX_NESTED_LOCKS; i++)
 481		__mutex_init(&torture_nested_mutexes[i], __func__,
 482			     &nested_mutex_keys[i]);
 483}
 484
 485static int torture_mutex_nested_lock(int tid __maybe_unused,
 486				     u32 lockset)
 487{
 488	int i;
 489
 490	for (i = 0; i < nested_locks; i++)
 491		if (lockset & (1 << i))
 492			mutex_lock(&torture_nested_mutexes[i]);
 493	return 0;
 494}
 495
 496static int torture_mutex_lock(int tid __maybe_unused)
 497__acquires(torture_mutex)
 498{
 499	mutex_lock(&torture_mutex);
 500	return 0;
 501}
 502
 503static void torture_mutex_delay(struct torture_random_state *trsp)
 504{
 505	/* We want a long delay occasionally to force massive contention.  */
 506	if (long_hold && !(torture_random(trsp) % (cxt.nrealwriters_stress * 2000 * long_hold)))
 507		mdelay(long_hold * 5);
 508	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
 509		torture_preempt_schedule();  /* Allow test to be preempted. */
 510}
 511
 512static void torture_mutex_unlock(int tid __maybe_unused)
 513__releases(torture_mutex)
 514{
 515	mutex_unlock(&torture_mutex);
 516}
 517
 518static void torture_mutex_nested_unlock(int tid __maybe_unused,
 519					u32 lockset)
 520{
 521	int i;
 522
 523	for (i = nested_locks - 1; i >= 0; i--)
 524		if (lockset & (1 << i))
 525			mutex_unlock(&torture_nested_mutexes[i]);
 526}
 527
 528static struct lock_torture_ops mutex_lock_ops = {
 529	.init		= torture_mutex_init,
 530	.nested_lock	= torture_mutex_nested_lock,
 531	.writelock	= torture_mutex_lock,
 532	.write_delay	= torture_mutex_delay,
 533	.task_boost     = torture_rt_boost,
 534	.writeunlock	= torture_mutex_unlock,
 535	.nested_unlock	= torture_mutex_nested_unlock,
 536	.readlock       = NULL,
 537	.read_delay     = NULL,
 538	.readunlock     = NULL,
 539	.name		= "mutex_lock"
 540};
 541
 542#include <linux/ww_mutex.h>
 543/*
 544 * The torture ww_mutexes should belong to the same lock class as
 545 * torture_ww_class to avoid lockdep problem. The ww_mutex_init()
 546 * function is called for initialization to ensure that.
 547 */
 548static DEFINE_WD_CLASS(torture_ww_class);
 549static struct ww_mutex torture_ww_mutex_0, torture_ww_mutex_1, torture_ww_mutex_2;
 550static struct ww_acquire_ctx *ww_acquire_ctxs;
 551
 552static void torture_ww_mutex_init(void)
 553{
 554	ww_mutex_init(&torture_ww_mutex_0, &torture_ww_class);
 555	ww_mutex_init(&torture_ww_mutex_1, &torture_ww_class);
 556	ww_mutex_init(&torture_ww_mutex_2, &torture_ww_class);
 557
 558	ww_acquire_ctxs = kmalloc_array(cxt.nrealwriters_stress,
 559					sizeof(*ww_acquire_ctxs),
 560					GFP_KERNEL);
 561	if (!ww_acquire_ctxs)
 562		VERBOSE_TOROUT_STRING("ww_acquire_ctx: Out of memory");
 563}
 564
 565static void torture_ww_mutex_exit(void)
 566{
 567	kfree(ww_acquire_ctxs);
 568}
 569
 570static int torture_ww_mutex_lock(int tid)
 571__acquires(torture_ww_mutex_0)
 572__acquires(torture_ww_mutex_1)
 573__acquires(torture_ww_mutex_2)
 574{
 575	LIST_HEAD(list);
 576	struct reorder_lock {
 577		struct list_head link;
 578		struct ww_mutex *lock;
 579	} locks[3], *ll, *ln;
 580	struct ww_acquire_ctx *ctx = &ww_acquire_ctxs[tid];
 581
 582	locks[0].lock = &torture_ww_mutex_0;
 583	list_add(&locks[0].link, &list);
 584
 585	locks[1].lock = &torture_ww_mutex_1;
 586	list_add(&locks[1].link, &list);
 587
 588	locks[2].lock = &torture_ww_mutex_2;
 589	list_add(&locks[2].link, &list);
 590
 591	ww_acquire_init(ctx, &torture_ww_class);
 592
 593	list_for_each_entry(ll, &list, link) {
 594		int err;
 595
 596		err = ww_mutex_lock(ll->lock, ctx);
 597		if (!err)
 598			continue;
 599
 600		ln = ll;
 601		list_for_each_entry_continue_reverse(ln, &list, link)
 602			ww_mutex_unlock(ln->lock);
 603
 604		if (err != -EDEADLK)
 605			return err;
 606
 607		ww_mutex_lock_slow(ll->lock, ctx);
 608		list_move(&ll->link, &list);
 609	}
 610
 611	return 0;
 612}
 613
 614static void torture_ww_mutex_unlock(int tid)
 615__releases(torture_ww_mutex_0)
 616__releases(torture_ww_mutex_1)
 617__releases(torture_ww_mutex_2)
 618{
 619	struct ww_acquire_ctx *ctx = &ww_acquire_ctxs[tid];
 620
 621	ww_mutex_unlock(&torture_ww_mutex_0);
 622	ww_mutex_unlock(&torture_ww_mutex_1);
 623	ww_mutex_unlock(&torture_ww_mutex_2);
 624	ww_acquire_fini(ctx);
 625}
 626
 627static struct lock_torture_ops ww_mutex_lock_ops = {
 628	.init		= torture_ww_mutex_init,
 629	.exit		= torture_ww_mutex_exit,
 630	.writelock	= torture_ww_mutex_lock,
 631	.write_delay	= torture_mutex_delay,
 632	.task_boost     = torture_rt_boost,
 633	.writeunlock	= torture_ww_mutex_unlock,
 634	.readlock       = NULL,
 635	.read_delay     = NULL,
 636	.readunlock     = NULL,
 637	.name		= "ww_mutex_lock"
 638};
 639
 640#ifdef CONFIG_RT_MUTEXES
 641static DEFINE_RT_MUTEX(torture_rtmutex);
 642static struct rt_mutex torture_nested_rtmutexes[MAX_NESTED_LOCKS];
 643static struct lock_class_key nested_rtmutex_keys[MAX_NESTED_LOCKS];
 644
 645static void torture_rtmutex_init(void)
 646{
 647	int i;
 648
 649	for (i = 0; i < MAX_NESTED_LOCKS; i++)
 650		__rt_mutex_init(&torture_nested_rtmutexes[i], __func__,
 651				&nested_rtmutex_keys[i]);
 652}
 653
 654static int torture_rtmutex_nested_lock(int tid __maybe_unused,
 655				       u32 lockset)
 656{
 657	int i;
 658
 659	for (i = 0; i < nested_locks; i++)
 660		if (lockset & (1 << i))
 661			rt_mutex_lock(&torture_nested_rtmutexes[i]);
 662	return 0;
 663}
 664
 665static int torture_rtmutex_lock(int tid __maybe_unused)
 666__acquires(torture_rtmutex)
 667{
 668	rt_mutex_lock(&torture_rtmutex);
 669	return 0;
 670}
 671
 672static void torture_rtmutex_delay(struct torture_random_state *trsp)
 673{
 674	const unsigned long shortdelay_us = 2;
 675
 676	/*
 677	 * We want a short delay mostly to emulate likely code, and
 678	 * we want a long delay occasionally to force massive contention.
 679	 */
 680	if (long_hold && !(torture_random(trsp) % (cxt.nrealwriters_stress * 2000 * long_hold)))
 681		mdelay(long_hold);
 682	if (!(torture_random(trsp) %
 683	      (cxt.nrealwriters_stress * 200 * shortdelay_us)))
 684		udelay(shortdelay_us);
 685	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
 686		torture_preempt_schedule();  /* Allow test to be preempted. */
 687}
 688
 689static void torture_rtmutex_unlock(int tid __maybe_unused)
 690__releases(torture_rtmutex)
 691{
 692	rt_mutex_unlock(&torture_rtmutex);
 693}
 694
 695static void torture_rt_boost_rtmutex(struct torture_random_state *trsp)
 696{
 697	if (!rt_boost)
 698		return;
 699
 700	__torture_rt_boost(trsp);
 701}
 702
 703static void torture_rtmutex_nested_unlock(int tid __maybe_unused,
 704					  u32 lockset)
 705{
 706	int i;
 707
 708	for (i = nested_locks - 1; i >= 0; i--)
 709		if (lockset & (1 << i))
 710			rt_mutex_unlock(&torture_nested_rtmutexes[i]);
 711}
 712
 713static struct lock_torture_ops rtmutex_lock_ops = {
 714	.init		= torture_rtmutex_init,
 715	.nested_lock	= torture_rtmutex_nested_lock,
 716	.writelock	= torture_rtmutex_lock,
 717	.write_delay	= torture_rtmutex_delay,
 718	.task_boost     = torture_rt_boost_rtmutex,
 719	.writeunlock	= torture_rtmutex_unlock,
 720	.nested_unlock	= torture_rtmutex_nested_unlock,
 721	.readlock       = NULL,
 722	.read_delay     = NULL,
 723	.readunlock     = NULL,
 724	.name		= "rtmutex_lock"
 725};
 726#endif
 727
 728static DECLARE_RWSEM(torture_rwsem);
 729static int torture_rwsem_down_write(int tid __maybe_unused)
 730__acquires(torture_rwsem)
 731{
 732	down_write(&torture_rwsem);
 733	return 0;
 734}
 735
 736static void torture_rwsem_write_delay(struct torture_random_state *trsp)
 737{
 738	/* We want a long delay occasionally to force massive contention.  */
 739	if (long_hold && !(torture_random(trsp) % (cxt.nrealwriters_stress * 2000 * long_hold)))
 740		mdelay(long_hold * 10);
 741	if (!(torture_random(trsp) % (cxt.nrealwriters_stress * 20000)))
 742		torture_preempt_schedule();  /* Allow test to be preempted. */
 743}
 744
 745static void torture_rwsem_up_write(int tid __maybe_unused)
 746__releases(torture_rwsem)
 747{
 748	up_write(&torture_rwsem);
 749}
 750
 751static int torture_rwsem_down_read(int tid __maybe_unused)
 752__acquires(torture_rwsem)
 753{
 754	down_read(&torture_rwsem);
 755	return 0;
 756}
 757
 758static void torture_rwsem_read_delay(struct torture_random_state *trsp)
 759{
 760	/* We want a long delay occasionally to force massive contention.  */
 761	if (long_hold && !(torture_random(trsp) % (cxt.nrealreaders_stress * 2000 * long_hold)))
 762		mdelay(long_hold * 2);
 763	else
 764		mdelay(long_hold / 2);
 765	if (!(torture_random(trsp) % (cxt.nrealreaders_stress * 20000)))
 766		torture_preempt_schedule();  /* Allow test to be preempted. */
 767}
 768
 769static void torture_rwsem_up_read(int tid __maybe_unused)
 770__releases(torture_rwsem)
 771{
 772	up_read(&torture_rwsem);
 773}
 774
 775static struct lock_torture_ops rwsem_lock_ops = {
 776	.writelock	= torture_rwsem_down_write,
 777	.write_delay	= torture_rwsem_write_delay,
 778	.task_boost     = torture_rt_boost,
 779	.writeunlock	= torture_rwsem_up_write,
 780	.readlock       = torture_rwsem_down_read,
 781	.read_delay     = torture_rwsem_read_delay,
 782	.readunlock     = torture_rwsem_up_read,
 783	.name		= "rwsem_lock"
 784};
 785
 786#include <linux/percpu-rwsem.h>
 787static struct percpu_rw_semaphore pcpu_rwsem;
 788
 789static void torture_percpu_rwsem_init(void)
 790{
 791	BUG_ON(percpu_init_rwsem(&pcpu_rwsem));
 792}
 793
 794static void torture_percpu_rwsem_exit(void)
 795{
 796	percpu_free_rwsem(&pcpu_rwsem);
 797}
 798
 799static int torture_percpu_rwsem_down_write(int tid __maybe_unused)
 800__acquires(pcpu_rwsem)
 801{
 802	percpu_down_write(&pcpu_rwsem);
 803	return 0;
 804}
 805
 806static void torture_percpu_rwsem_up_write(int tid __maybe_unused)
 807__releases(pcpu_rwsem)
 808{
 809	percpu_up_write(&pcpu_rwsem);
 810}
 811
 812static int torture_percpu_rwsem_down_read(int tid __maybe_unused)
 813__acquires(pcpu_rwsem)
 814{
 815	percpu_down_read(&pcpu_rwsem);
 816	return 0;
 817}
 818
 819static void torture_percpu_rwsem_up_read(int tid __maybe_unused)
 820__releases(pcpu_rwsem)
 821{
 822	percpu_up_read(&pcpu_rwsem);
 823}
 824
 825static struct lock_torture_ops percpu_rwsem_lock_ops = {
 826	.init		= torture_percpu_rwsem_init,
 827	.exit		= torture_percpu_rwsem_exit,
 828	.writelock	= torture_percpu_rwsem_down_write,
 829	.write_delay	= torture_rwsem_write_delay,
 830	.task_boost     = torture_rt_boost,
 831	.writeunlock	= torture_percpu_rwsem_up_write,
 832	.readlock       = torture_percpu_rwsem_down_read,
 833	.read_delay     = torture_rwsem_read_delay,
 834	.readunlock     = torture_percpu_rwsem_up_read,
 835	.name		= "percpu_rwsem_lock"
 836};
 837
 838/*
 839 * Lock torture writer kthread.  Repeatedly acquires and releases
 840 * the lock, checking for duplicate acquisitions.
 841 */
 842static int lock_torture_writer(void *arg)
 843{
 844	unsigned long j;
 845	unsigned long j1;
 846	u32 lockset_mask;
 847	struct lock_stress_stats *lwsp = arg;
 848	DEFINE_TORTURE_RANDOM(rand);
 849	bool skip_main_lock;
 850	int tid = lwsp - cxt.lwsa;
 851
 852	VERBOSE_TOROUT_STRING("lock_torture_writer task started");
 853	if (!rt_task(current))
 854		set_user_nice(current, MAX_NICE);
 855
 856	do {
 857		if ((torture_random(&rand) & 0xfffff) == 0)
 858			schedule_timeout_uninterruptible(1);
 859
 860		lockset_mask = torture_random(&rand);
 861		/*
 862		 * When using nested_locks, we want to occasionally
 863		 * skip the main lock so we can avoid always serializing
 864		 * the lock chains on that central lock. By skipping the
 865		 * main lock occasionally, we can create different
 866		 * contention patterns (allowing for multiple disjoint
 867		 * blocked trees)
 868		 */
 869		skip_main_lock = (nested_locks &&
 870				 !(torture_random(&rand) % 100));
 871
 872		cxt.cur_ops->task_boost(&rand);
 873		if (cxt.cur_ops->nested_lock)
 874			cxt.cur_ops->nested_lock(tid, lockset_mask);
 875
 876		if (!skip_main_lock) {
 877			if (acq_writer_lim > 0)
 878				j = jiffies;
 879			cxt.cur_ops->writelock(tid);
 880			if (WARN_ON_ONCE(lock_is_write_held))
 881				lwsp->n_lock_fail++;
 882			lock_is_write_held = true;
 883			if (WARN_ON_ONCE(atomic_read(&lock_is_read_held)))
 884				lwsp->n_lock_fail++; /* rare, but... */
 885			if (acq_writer_lim > 0) {
 886				j1 = jiffies;
 887				WARN_ONCE(time_after(j1, j + acq_writer_lim),
 888					  "%s: Lock acquisition took %lu jiffies.\n",
 889					  __func__, j1 - j);
 890			}
 891			lwsp->n_lock_acquired++;
 892
 893			cxt.cur_ops->write_delay(&rand);
 894
 895			lock_is_write_held = false;
 896			WRITE_ONCE(last_lock_release, jiffies);
 897			cxt.cur_ops->writeunlock(tid);
 898		}
 899		if (cxt.cur_ops->nested_unlock)
 900			cxt.cur_ops->nested_unlock(tid, lockset_mask);
 901
 902		stutter_wait("lock_torture_writer");
 903	} while (!torture_must_stop());
 904
 905	cxt.cur_ops->task_boost(NULL); /* reset prio */
 906	torture_kthread_stopping("lock_torture_writer");
 907	return 0;
 908}
 909
 910/*
 911 * Lock torture reader kthread.  Repeatedly acquires and releases
 912 * the reader lock.
 913 */
 914static int lock_torture_reader(void *arg)
 915{
 916	struct lock_stress_stats *lrsp = arg;
 917	int tid = lrsp - cxt.lrsa;
 918	DEFINE_TORTURE_RANDOM(rand);
 919
 920	VERBOSE_TOROUT_STRING("lock_torture_reader task started");
 921	set_user_nice(current, MAX_NICE);
 922
 923	do {
 924		if ((torture_random(&rand) & 0xfffff) == 0)
 925			schedule_timeout_uninterruptible(1);
 926
 927		cxt.cur_ops->readlock(tid);
 928		atomic_inc(&lock_is_read_held);
 929		if (WARN_ON_ONCE(lock_is_write_held))
 930			lrsp->n_lock_fail++; /* rare, but... */
 931
 932		lrsp->n_lock_acquired++;
 933		cxt.cur_ops->read_delay(&rand);
 934		atomic_dec(&lock_is_read_held);
 935		cxt.cur_ops->readunlock(tid);
 936
 937		stutter_wait("lock_torture_reader");
 938	} while (!torture_must_stop());
 939	torture_kthread_stopping("lock_torture_reader");
 940	return 0;
 941}
 942
 943/*
 944 * Create an lock-torture-statistics message in the specified buffer.
 945 */
 946static void __torture_print_stats(char *page,
 947				  struct lock_stress_stats *statp, bool write)
 948{
 949	long cur;
 950	bool fail = false;
 951	int i, n_stress;
 952	long max = 0, min = statp ? data_race(statp[0].n_lock_acquired) : 0;
 953	long long sum = 0;
 954
 955	n_stress = write ? cxt.nrealwriters_stress : cxt.nrealreaders_stress;
 956	for (i = 0; i < n_stress; i++) {
 957		if (data_race(statp[i].n_lock_fail))
 958			fail = true;
 959		cur = data_race(statp[i].n_lock_acquired);
 960		sum += cur;
 961		if (max < cur)
 962			max = cur;
 963		if (min > cur)
 964			min = cur;
 965	}
 966	page += sprintf(page,
 967			"%s:  Total: %lld  Max/Min: %ld/%ld %s  Fail: %d %s\n",
 968			write ? "Writes" : "Reads ",
 969			sum, max, min,
 970			!onoff_interval && max / 2 > min ? "???" : "",
 971			fail, fail ? "!!!" : "");
 972	if (fail)
 973		atomic_inc(&cxt.n_lock_torture_errors);
 974}
 975
 976/*
 977 * Print torture statistics.  Caller must ensure that there is only one
 978 * call to this function at a given time!!!  This is normally accomplished
 979 * by relying on the module system to only have one copy of the module
 980 * loaded, and then by giving the lock_torture_stats kthread full control
 981 * (or the init/cleanup functions when lock_torture_stats thread is not
 982 * running).
 983 */
 984static void lock_torture_stats_print(void)
 985{
 986	int size = cxt.nrealwriters_stress * 200 + 8192;
 987	char *buf;
 988
 989	if (cxt.cur_ops->readlock)
 990		size += cxt.nrealreaders_stress * 200 + 8192;
 991
 992	buf = kmalloc(size, GFP_KERNEL);
 993	if (!buf) {
 994		pr_err("lock_torture_stats_print: Out of memory, need: %d",
 995		       size);
 996		return;
 997	}
 998
 999	__torture_print_stats(buf, cxt.lwsa, true);
1000	pr_alert("%s", buf);
1001	kfree(buf);
1002
1003	if (cxt.cur_ops->readlock) {
1004		buf = kmalloc(size, GFP_KERNEL);
1005		if (!buf) {
1006			pr_err("lock_torture_stats_print: Out of memory, need: %d",
1007			       size);
1008			return;
1009		}
1010
1011		__torture_print_stats(buf, cxt.lrsa, false);
1012		pr_alert("%s", buf);
1013		kfree(buf);
1014	}
1015}
1016
1017/*
1018 * Periodically prints torture statistics, if periodic statistics printing
1019 * was specified via the stat_interval module parameter.
1020 *
1021 * No need to worry about fullstop here, since this one doesn't reference
1022 * volatile state or register callbacks.
1023 */
1024static int lock_torture_stats(void *arg)
1025{
1026	VERBOSE_TOROUT_STRING("lock_torture_stats task started");
1027	do {
1028		schedule_timeout_interruptible(stat_interval * HZ);
1029		lock_torture_stats_print();
1030		torture_shutdown_absorb("lock_torture_stats");
1031	} while (!torture_must_stop());
1032	torture_kthread_stopping("lock_torture_stats");
1033	return 0;
1034}
1035
1036
1037static inline void
1038lock_torture_print_module_parms(struct lock_torture_ops *cur_ops,
1039				const char *tag)
1040{
1041	static cpumask_t cpumask_all;
1042	cpumask_t *rcmp = cpumask_nonempty(bind_readers) ? bind_readers : &cpumask_all;
1043	cpumask_t *wcmp = cpumask_nonempty(bind_writers) ? bind_writers : &cpumask_all;
1044
1045	cpumask_setall(&cpumask_all);
1046	pr_alert("%s" TORTURE_FLAG
1047		 "--- %s%s: acq_writer_lim=%d bind_readers=%*pbl bind_writers=%*pbl call_rcu_chains=%d long_hold=%d nested_locks=%d nreaders_stress=%d nwriters_stress=%d onoff_holdoff=%d onoff_interval=%d rt_boost=%d rt_boost_factor=%d shuffle_interval=%d shutdown_secs=%d stat_interval=%d stutter=%d verbose=%d writer_fifo=%d\n",
1048		 torture_type, tag, cxt.debug_lock ? " [debug]": "",
1049		 acq_writer_lim, cpumask_pr_args(rcmp), cpumask_pr_args(wcmp),
1050		 call_rcu_chains, long_hold, nested_locks, cxt.nrealreaders_stress,
1051		 cxt.nrealwriters_stress, onoff_holdoff, onoff_interval, rt_boost,
1052		 rt_boost_factor, shuffle_interval, shutdown_secs, stat_interval, stutter,
1053		 verbose, writer_fifo);
1054}
1055
1056// If requested, maintain call_rcu() chains to keep a grace period always
1057// in flight.  These increase the probability of getting an RCU CPU stall
1058// warning and associated diagnostics when a locking primitive stalls.
1059
1060static void call_rcu_chain_cb(struct rcu_head *rhp)
1061{
1062	struct call_rcu_chain *crcp = container_of(rhp, struct call_rcu_chain, crc_rh);
1063
1064	if (!smp_load_acquire(&crcp->crc_stop)) {
1065		(void)start_poll_synchronize_rcu(); // Start one grace period...
1066		call_rcu(&crcp->crc_rh, call_rcu_chain_cb); // ... and later start another.
1067	}
1068}
1069
1070// Start the requested number of call_rcu() chains.
1071static int call_rcu_chain_init(void)
1072{
1073	int i;
1074
1075	if (call_rcu_chains <= 0)
1076		return 0;
1077	call_rcu_chain_list = kcalloc(call_rcu_chains, sizeof(*call_rcu_chain_list), GFP_KERNEL);
1078	if (!call_rcu_chain_list)
1079		return -ENOMEM;
1080	for (i = 0; i < call_rcu_chains; i++) {
1081		call_rcu_chain_list[i].crc_stop = false;
1082		call_rcu(&call_rcu_chain_list[i].crc_rh, call_rcu_chain_cb);
1083	}
1084	return 0;
1085}
1086
1087// Stop all of the call_rcu() chains.
1088static void call_rcu_chain_cleanup(void)
1089{
1090	int i;
1091
1092	if (!call_rcu_chain_list)
1093		return;
1094	for (i = 0; i < call_rcu_chains; i++)
1095		smp_store_release(&call_rcu_chain_list[i].crc_stop, true);
1096	rcu_barrier();
1097	kfree(call_rcu_chain_list);
1098	call_rcu_chain_list = NULL;
1099}
1100
1101static void lock_torture_cleanup(void)
1102{
1103	int i;
1104
1105	if (torture_cleanup_begin())
1106		return;
1107
1108	/*
1109	 * Indicates early cleanup, meaning that the test has not run,
1110	 * such as when passing bogus args when loading the module.
1111	 * However cxt->cur_ops.init() may have been invoked, so beside
1112	 * perform the underlying torture-specific cleanups, cur_ops.exit()
1113	 * will be invoked if needed.
1114	 */
1115	if (!cxt.lwsa && !cxt.lrsa)
1116		goto end;
1117
1118	if (writer_tasks) {
1119		for (i = 0; i < cxt.nrealwriters_stress; i++)
1120			torture_stop_kthread(lock_torture_writer, writer_tasks[i]);
1121		kfree(writer_tasks);
1122		writer_tasks = NULL;
1123	}
1124
1125	if (reader_tasks) {
1126		for (i = 0; i < cxt.nrealreaders_stress; i++)
1127			torture_stop_kthread(lock_torture_reader,
1128					     reader_tasks[i]);
1129		kfree(reader_tasks);
1130		reader_tasks = NULL;
1131	}
1132
1133	torture_stop_kthread(lock_torture_stats, stats_task);
1134	lock_torture_stats_print();  /* -After- the stats thread is stopped! */
1135
1136	if (atomic_read(&cxt.n_lock_torture_errors))
1137		lock_torture_print_module_parms(cxt.cur_ops,
1138						"End of test: FAILURE");
1139	else if (torture_onoff_failures())
1140		lock_torture_print_module_parms(cxt.cur_ops,
1141						"End of test: LOCK_HOTPLUG");
1142	else
1143		lock_torture_print_module_parms(cxt.cur_ops,
1144						"End of test: SUCCESS");
1145
1146	kfree(cxt.lwsa);
1147	cxt.lwsa = NULL;
1148	kfree(cxt.lrsa);
1149	cxt.lrsa = NULL;
1150
1151	call_rcu_chain_cleanup();
1152
1153end:
1154	if (cxt.init_called) {
1155		if (cxt.cur_ops->exit)
1156			cxt.cur_ops->exit();
1157		cxt.init_called = false;
1158	}
1159	torture_cleanup_end();
1160}
1161
1162static int __init lock_torture_init(void)
1163{
1164	int i, j;
1165	int firsterr = 0;
1166	static struct lock_torture_ops *torture_ops[] = {
1167		&lock_busted_ops,
1168		&spin_lock_ops, &spin_lock_irq_ops,
1169		&raw_spin_lock_ops, &raw_spin_lock_irq_ops,
1170		&rw_lock_ops, &rw_lock_irq_ops,
1171		&mutex_lock_ops,
1172		&ww_mutex_lock_ops,
1173#ifdef CONFIG_RT_MUTEXES
1174		&rtmutex_lock_ops,
1175#endif
1176		&rwsem_lock_ops,
1177		&percpu_rwsem_lock_ops,
1178	};
1179
1180	if (!torture_init_begin(torture_type, verbose))
1181		return -EBUSY;
1182
1183	/* Process args and tell the world that the torturer is on the job. */
1184	for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
1185		cxt.cur_ops = torture_ops[i];
1186		if (strcmp(torture_type, cxt.cur_ops->name) == 0)
1187			break;
1188	}
1189	if (i == ARRAY_SIZE(torture_ops)) {
1190		pr_alert("lock-torture: invalid torture type: \"%s\"\n",
1191			 torture_type);
1192		pr_alert("lock-torture types:");
1193		for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
1194			pr_alert(" %s", torture_ops[i]->name);
1195		pr_alert("\n");
1196		firsterr = -EINVAL;
1197		goto unwind;
1198	}
1199
1200	if (nwriters_stress == 0 &&
1201	    (!cxt.cur_ops->readlock || nreaders_stress == 0)) {
1202		pr_alert("lock-torture: must run at least one locking thread\n");
1203		firsterr = -EINVAL;
1204		goto unwind;
1205	}
1206
1207	if (nwriters_stress >= 0)
1208		cxt.nrealwriters_stress = nwriters_stress;
1209	else
1210		cxt.nrealwriters_stress = 2 * num_online_cpus();
1211
1212	if (cxt.cur_ops->init) {
1213		cxt.cur_ops->init();
1214		cxt.init_called = true;
1215	}
1216
1217#ifdef CONFIG_DEBUG_MUTEXES
1218	if (str_has_prefix(torture_type, "mutex"))
1219		cxt.debug_lock = true;
1220#endif
1221#ifdef CONFIG_DEBUG_RT_MUTEXES
1222	if (str_has_prefix(torture_type, "rtmutex"))
1223		cxt.debug_lock = true;
1224#endif
1225#ifdef CONFIG_DEBUG_SPINLOCK
1226	if ((str_has_prefix(torture_type, "spin")) ||
1227	    (str_has_prefix(torture_type, "rw_lock")))
1228		cxt.debug_lock = true;
1229#endif
1230
1231	/* Initialize the statistics so that each run gets its own numbers. */
1232	if (nwriters_stress) {
1233		lock_is_write_held = false;
1234		cxt.lwsa = kmalloc_array(cxt.nrealwriters_stress,
1235					 sizeof(*cxt.lwsa),
1236					 GFP_KERNEL);
1237		if (cxt.lwsa == NULL) {
1238			VERBOSE_TOROUT_STRING("cxt.lwsa: Out of memory");
1239			firsterr = -ENOMEM;
1240			goto unwind;
1241		}
1242
1243		for (i = 0; i < cxt.nrealwriters_stress; i++) {
1244			cxt.lwsa[i].n_lock_fail = 0;
1245			cxt.lwsa[i].n_lock_acquired = 0;
1246		}
1247	}
1248
1249	if (cxt.cur_ops->readlock) {
1250		if (nreaders_stress >= 0)
1251			cxt.nrealreaders_stress = nreaders_stress;
1252		else {
1253			/*
1254			 * By default distribute evenly the number of
1255			 * readers and writers. We still run the same number
1256			 * of threads as the writer-only locks default.
1257			 */
1258			if (nwriters_stress < 0) /* user doesn't care */
1259				cxt.nrealwriters_stress = num_online_cpus();
1260			cxt.nrealreaders_stress = cxt.nrealwriters_stress;
1261		}
1262
1263		if (nreaders_stress) {
1264			cxt.lrsa = kmalloc_array(cxt.nrealreaders_stress,
1265						 sizeof(*cxt.lrsa),
1266						 GFP_KERNEL);
1267			if (cxt.lrsa == NULL) {
1268				VERBOSE_TOROUT_STRING("cxt.lrsa: Out of memory");
1269				firsterr = -ENOMEM;
1270				kfree(cxt.lwsa);
1271				cxt.lwsa = NULL;
1272				goto unwind;
1273			}
1274
1275			for (i = 0; i < cxt.nrealreaders_stress; i++) {
1276				cxt.lrsa[i].n_lock_fail = 0;
1277				cxt.lrsa[i].n_lock_acquired = 0;
1278			}
1279		}
1280	}
1281
1282	firsterr = call_rcu_chain_init();
1283	if (torture_init_error(firsterr))
1284		goto unwind;
1285
1286	lock_torture_print_module_parms(cxt.cur_ops, "Start of test");
1287
1288	/* Prepare torture context. */
1289	if (onoff_interval > 0) {
1290		firsterr = torture_onoff_init(onoff_holdoff * HZ,
1291					      onoff_interval * HZ, NULL);
1292		if (torture_init_error(firsterr))
1293			goto unwind;
1294	}
1295	if (shuffle_interval > 0) {
1296		firsterr = torture_shuffle_init(shuffle_interval);
1297		if (torture_init_error(firsterr))
1298			goto unwind;
1299	}
1300	if (shutdown_secs > 0) {
1301		firsterr = torture_shutdown_init(shutdown_secs,
1302						 lock_torture_cleanup);
1303		if (torture_init_error(firsterr))
1304			goto unwind;
1305	}
1306	if (stutter > 0) {
1307		firsterr = torture_stutter_init(stutter, stutter);
1308		if (torture_init_error(firsterr))
1309			goto unwind;
1310	}
1311
1312	if (nwriters_stress) {
1313		writer_tasks = kcalloc(cxt.nrealwriters_stress,
1314				       sizeof(writer_tasks[0]),
1315				       GFP_KERNEL);
1316		if (writer_tasks == NULL) {
1317			TOROUT_ERRSTRING("writer_tasks: Out of memory");
1318			firsterr = -ENOMEM;
1319			goto unwind;
1320		}
1321	}
1322
1323	/* cap nested_locks to MAX_NESTED_LOCKS */
1324	if (nested_locks > MAX_NESTED_LOCKS)
1325		nested_locks = MAX_NESTED_LOCKS;
1326
1327	if (cxt.cur_ops->readlock) {
1328		reader_tasks = kcalloc(cxt.nrealreaders_stress,
1329				       sizeof(reader_tasks[0]),
1330				       GFP_KERNEL);
1331		if (reader_tasks == NULL) {
1332			TOROUT_ERRSTRING("reader_tasks: Out of memory");
1333			kfree(writer_tasks);
1334			writer_tasks = NULL;
1335			firsterr = -ENOMEM;
1336			goto unwind;
1337		}
1338	}
1339
1340	/*
1341	 * Create the kthreads and start torturing (oh, those poor little locks).
1342	 *
1343	 * TODO: Note that we interleave writers with readers, giving writers a
1344	 * slight advantage, by creating its kthread first. This can be modified
1345	 * for very specific needs, or even let the user choose the policy, if
1346	 * ever wanted.
1347	 */
1348	for (i = 0, j = 0; i < cxt.nrealwriters_stress ||
1349		    j < cxt.nrealreaders_stress; i++, j++) {
1350		if (i >= cxt.nrealwriters_stress)
1351			goto create_reader;
1352
1353		/* Create writer. */
1354		firsterr = torture_create_kthread_cb(lock_torture_writer, &cxt.lwsa[i],
1355						     writer_tasks[i],
1356						     writer_fifo ? sched_set_fifo : NULL);
1357		if (torture_init_error(firsterr))
1358			goto unwind;
1359		if (cpumask_nonempty(bind_writers))
1360			torture_sched_setaffinity(writer_tasks[i]->pid, bind_writers);
1361
1362	create_reader:
1363		if (cxt.cur_ops->readlock == NULL || (j >= cxt.nrealreaders_stress))
1364			continue;
1365		/* Create reader. */
1366		firsterr = torture_create_kthread(lock_torture_reader, &cxt.lrsa[j],
1367						  reader_tasks[j]);
1368		if (torture_init_error(firsterr))
1369			goto unwind;
1370		if (cpumask_nonempty(bind_readers))
1371			torture_sched_setaffinity(reader_tasks[j]->pid, bind_readers);
1372	}
1373	if (stat_interval > 0) {
1374		firsterr = torture_create_kthread(lock_torture_stats, NULL,
1375						  stats_task);
1376		if (torture_init_error(firsterr))
1377			goto unwind;
1378	}
1379	torture_init_end();
1380	return 0;
1381
1382unwind:
1383	torture_init_end();
1384	lock_torture_cleanup();
1385	if (shutdown_secs) {
1386		WARN_ON(!IS_MODULE(CONFIG_LOCK_TORTURE_TEST));
1387		kernel_power_off();
1388	}
1389	return firsterr;
1390}
1391
1392module_init(lock_torture_init);
1393module_exit(lock_torture_cleanup);