kernel/rcu/rcutorture.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Read-Copy Update module-based torture test facility
   4  *
   5  * Copyright (C) IBM Corporation, 2005, 2006
   6  *
   7  * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
   8  *        Josh Triplett <josh@joshtriplett.org>
   9  *
  10  * See also:  Documentation/RCU/torture.rst
  11  */
  12
  13 #define pr_fmt(fmt) fmt
  14
  15 #include <linux/types.h>
  16 #include <linux/kernel.h>
  17 #include <linux/init.h>
  18 #include <linux/module.h>
  19 #include <linux/kthread.h>
  20 #include <linux/err.h>
  21 #include <linux/spinlock.h>
  22 #include <linux/smp.h>
  23 #include <linux/rcupdate_wait.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/sched/signal.h>
  26 #include <uapi/linux/sched/types.h>
  27 #include <linux/atomic.h>
  28 #include <linux/bitops.h>
  29 #include <linux/completion.h>
  30 #include <linux/moduleparam.h>
  31 #include <linux/percpu.h>
  32 #include <linux/notifier.h>
  33 #include <linux/reboot.h>
  34 #include <linux/freezer.h>
  35 #include <linux/cpu.h>
  36 #include <linux/delay.h>
  37 #include <linux/stat.h>
  38 #include <linux/srcu.h>
  39 #include <linux/slab.h>
  40 #include <linux/trace_clock.h>
  41 #include <asm/byteorder.h>
  42 #include <linux/torture.h>
  43 #include <linux/vmalloc.h>
  44 #include <linux/sched/debug.h>
  45 #include <linux/sched/sysctl.h>
  46 #include <linux/oom.h>
  47 #include <linux/tick.h>
  48 #include <linux/rcupdate_trace.h>
  49 #include <linux/nmi.h>
  50
  51 #include "rcu.h"
  52
  53 MODULE_LICENSE("GPL");
  54 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com> and Josh Triplett <josh@joshtriplett.org>");
  55
  56 /* Bits for ->extendables field, extendables param, and related definitions. */
  57 #define RCUTORTURE_RDR_SHIFT_1   8      /* Put SRCU index in upper bits. */
  58 #define RCUTORTURE_RDR_MASK_1    (1 << RCUTORTURE_RDR_SHIFT_1)
  59 #define RCUTORTURE_RDR_SHIFT_2   9      /* Put SRCU index in upper bits. */
  60 #define RCUTORTURE_RDR_MASK_2    (1 << RCUTORTURE_RDR_SHIFT_2)
  61 #define RCUTORTURE_RDR_BH        0x01   /* Extend readers by disabling bh. */
  62 #define RCUTORTURE_RDR_IRQ       0x02   /*  ... disabling interrupts. */
  63 #define RCUTORTURE_RDR_PREEMPT   0x04   /*  ... disabling preemption. */
  64 #define RCUTORTURE_RDR_RBH       0x08   /*  ... rcu_read_lock_bh(). */
  65 #define RCUTORTURE_RDR_SCHED     0x10   /*  ... rcu_read_lock_sched(). */
  66 #define RCUTORTURE_RDR_RCU_1     0x20   /*  ... entering another RCU reader. */
  67 #define RCUTORTURE_RDR_RCU_2     0x40   /*  ... entering another RCU reader. */
  68 #define RCUTORTURE_RDR_NBITS     7      /* Number of bits defined above. */
  69 #define RCUTORTURE_MAX_EXTEND    \
  70         (RCUTORTURE_RDR_BH | RCUTORTURE_RDR_IRQ | RCUTORTURE_RDR_PREEMPT | \
  71          RCUTORTURE_RDR_RBH | RCUTORTURE_RDR_SCHED)
  72 #define RCUTORTURE_RDR_MAX_LOOPS 0x7    /* Maximum reader extensions. */
  73                                         /* Must be power of two minus one. */
  74 #define RCUTORTURE_RDR_MAX_SEGS (RCUTORTURE_RDR_MAX_LOOPS + 3)
  75
  76 torture_param(int, extendables, RCUTORTURE_MAX_EXTEND,
  77               "Extend readers by disabling bh (1), irqs (2), or preempt (4)");
  78 torture_param(int, fqs_duration, 0,
  79               "Duration of fqs bursts (us), 0 to disable");
  80 torture_param(int, fqs_holdoff, 0, "Holdoff time within fqs bursts (us)");
  81 torture_param(int, fqs_stutter, 3, "Wait time between fqs bursts (s)");
  82 torture_param(int, fwd_progress, 1, "Test grace-period forward progress");
  83 torture_param(int, fwd_progress_div, 4, "Fraction of CPU stall to wait");
  84 torture_param(int, fwd_progress_holdoff, 60,
  85               "Time between forward-progress tests (s)");
  86 torture_param(bool, fwd_progress_need_resched, 1,
  87               "Hide cond_resched() behind need_resched()");
  88 torture_param(bool, gp_cond, false, "Use conditional/async GP wait primitives");
  89 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
  90 torture_param(bool, gp_normal, false,
  91              "Use normal (non-expedited) GP wait primitives");
  92 torture_param(bool, gp_poll, false, "Use polling GP wait primitives");
  93 torture_param(bool, gp_sync, false, "Use synchronous GP wait primitives");
  94 torture_param(int, irqreader, 1, "Allow RCU readers from irq handlers");
  95 torture_param(int, leakpointer, 0, "Leak pointer dereferences from readers");
  96 torture_param(int, n_barrier_cbs, 0,
  97              "# of callbacks/kthreads for barrier testing");
  98 torture_param(int, nfakewriters, 4, "Number of RCU fake writer threads");
  99 torture_param(int, nreaders, -1, "Number of RCU reader threads");
 100 torture_param(int, object_debug, 0,
 101              "Enable debug-object double call_rcu() testing");
 102 torture_param(int, onoff_holdoff, 0, "Time after boot before CPU hotplugs (s)");
 103 torture_param(int, onoff_interval, 0,
 104              "Time between CPU hotplugs (jiffies), 0=disable");
 105 torture_param(int, nocbs_nthreads, 0, "Number of NOCB toggle threads, 0 to disable");
 106 torture_param(int, nocbs_toggle, 1000, "Time between toggling nocb state (ms)");
 107 torture_param(int, read_exit_delay, 13,
 108               "Delay between read-then-exit episodes (s)");
 109 torture_param(int, read_exit_burst, 16,
 110               "# of read-then-exit bursts per episode, zero to disable");
 111 torture_param(int, shuffle_interval, 3, "Number of seconds between shuffles");
 112 torture_param(int, shutdown_secs, 0, "Shutdown time (s), <= zero to disable.");
 113 torture_param(int, stall_cpu, 0, "Stall duration (s), zero to disable.");
 114 torture_param(int, stall_cpu_holdoff, 10,
 115              "Time to wait before starting stall (s).");
 116 torture_param(bool, stall_no_softlockup, false,
 117              "Avoid softlockup warning during cpu stall.");
 118 torture_param(int, stall_cpu_irqsoff, 0, "Disable interrupts while stalling.");
 119 torture_param(int, stall_cpu_block, 0, "Sleep while stalling.");
 120 torture_param(int, stall_gp_kthread, 0,
 121               "Grace-period kthread stall duration (s).");
 122 torture_param(int, stat_interval, 60,
 123              "Number of seconds between stats printk()s");
 124 torture_param(int, stutter, 5, "Number of seconds to run/halt test");
 125 torture_param(int, test_boost, 1, "Test RCU prio boost: 0=no, 1=maybe, 2=yes.");
 126 torture_param(int, test_boost_duration, 4,
 127              "Duration of each boost test, seconds.");
 128 torture_param(int, test_boost_interval, 7,
 129              "Interval between boost tests, seconds.");
 130 torture_param(bool, test_no_idle_hz, true,
 131              "Test support for tickless idle CPUs");
 132 torture_param(int, verbose, 1,
 133              "Enable verbose debugging printk()s");
 134
 135 static char *torture_type = "rcu";
 136 module_param(torture_type, charp, 0444);
 137 MODULE_PARM_DESC(torture_type, "Type of RCU to torture (rcu, srcu, ...)");
 138
 139 static int nrealnocbers;
 140 static int nrealreaders;
 141 static struct task_struct *writer_task;
 142 static struct task_struct **fakewriter_tasks;
 143 static struct task_struct **reader_tasks;
 144 static struct task_struct **nocb_tasks;
 145 static struct task_struct *stats_task;
 146 static struct task_struct *fqs_task;
 147 static struct task_struct *boost_tasks[NR_CPUS];
 148 static struct task_struct *stall_task;
 149 static struct task_struct **fwd_prog_tasks;
 150 static struct task_struct **barrier_cbs_tasks;
 151 static struct task_struct *barrier_task;
 152 static struct task_struct *read_exit_task;
 153
 154 #define RCU_TORTURE_PIPE_LEN 10
 155
 156 // Mailbox-like structure to check RCU global memory ordering.
 157 struct rcu_torture_reader_check {
 158         unsigned long rtc_myloops;
 159         int rtc_chkrdr;
 160         unsigned long rtc_chkloops;
 161         int rtc_ready;
 162         struct rcu_torture_reader_check *rtc_assigner;
 163 } ____cacheline_internodealigned_in_smp;
 164
 165 // Update-side data structure used to check RCU readers.
 166 struct rcu_torture {
 167         struct rcu_head rtort_rcu;
 168         int rtort_pipe_count;
 169         struct list_head rtort_free;
 170         int rtort_mbtest;
 171         struct rcu_torture_reader_check *rtort_chkp;
 172 };
 173
 174 static LIST_HEAD(rcu_torture_freelist);
 175 static struct rcu_torture __rcu *rcu_torture_current;
 176 static unsigned long rcu_torture_current_version;
 177 static struct rcu_torture rcu_tortures[10 * RCU_TORTURE_PIPE_LEN];
 178 static DEFINE_SPINLOCK(rcu_torture_lock);
 179 static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1], rcu_torture_count);
 180 static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1], rcu_torture_batch);
 181 static atomic_t rcu_torture_wcount[RCU_TORTURE_PIPE_LEN + 1];
 182 static struct rcu_torture_reader_check *rcu_torture_reader_mbchk;
 183 static atomic_t n_rcu_torture_alloc;
 184 static atomic_t n_rcu_torture_alloc_fail;
 185 static atomic_t n_rcu_torture_free;
 186 static atomic_t n_rcu_torture_mberror;
 187 static atomic_t n_rcu_torture_mbchk_fail;
 188 static atomic_t n_rcu_torture_mbchk_tries;
 189 static atomic_t n_rcu_torture_error;
 190 static long n_rcu_torture_barrier_error;
 191 static long n_rcu_torture_boost_ktrerror;
 192 static long n_rcu_torture_boost_rterror;
 193 static long n_rcu_torture_boost_failure;
 194 static long n_rcu_torture_boosts;
 195 static atomic_long_t n_rcu_torture_timers;
 196 static long n_barrier_attempts;
 197 static long n_barrier_successes; /* did rcu_barrier test succeed? */
 198 static unsigned long n_read_exits;
 199 static struct list_head rcu_torture_removed;
 200 static unsigned long shutdown_jiffies;
 201 static unsigned long start_gp_seq;
 202 static atomic_long_t n_nocb_offload;
 203 static atomic_long_t n_nocb_deoffload;
 204
 205 static int rcu_torture_writer_state;
 206 #define RTWS_FIXED_DELAY        0
 207 #define RTWS_DELAY              1
 208 #define RTWS_REPLACE            2
 209 #define RTWS_DEF_FREE           3
 210 #define RTWS_EXP_SYNC           4
 211 #define RTWS_COND_GET           5
 212 #define RTWS_COND_SYNC          6
 213 #define RTWS_POLL_GET           7
 214 #define RTWS_POLL_WAIT          8
 215 #define RTWS_SYNC               9
 216 #define RTWS_STUTTER            10
 217 #define RTWS_STOPPING           11
 218 static const char * const rcu_torture_writer_state_names[] = {
 219         "RTWS_FIXED_DELAY",
 220         "RTWS_DELAY",
 221         "RTWS_REPLACE",
 222         "RTWS_DEF_FREE",
 223         "RTWS_EXP_SYNC",
 224         "RTWS_COND_GET",
 225         "RTWS_COND_SYNC",
 226         "RTWS_POLL_GET",
 227         "RTWS_POLL_WAIT",
 228         "RTWS_SYNC",
 229         "RTWS_STUTTER",
 230         "RTWS_STOPPING",
 231 };
 232
 233 /* Record reader segment types and duration for first failing read. */
 234 struct rt_read_seg {
 235         int rt_readstate;
 236         unsigned long rt_delay_jiffies;
 237         unsigned long rt_delay_ms;
 238         unsigned long rt_delay_us;
 239         bool rt_preempted;
 240 };
 241 static int err_segs_recorded;
 242 static struct rt_read_seg err_segs[RCUTORTURE_RDR_MAX_SEGS];
 243 static int rt_read_nsegs;
 244
 245 static const char *rcu_torture_writer_state_getname(void)
 246 {
 247         unsigned int i = READ_ONCE(rcu_torture_writer_state);
 248
 249         if (i >= ARRAY_SIZE(rcu_torture_writer_state_names))
 250                 return "???";
 251         return rcu_torture_writer_state_names[i];
 252 }
 253
 254 #ifdef CONFIG_RCU_TRACE
 255 static u64 notrace rcu_trace_clock_local(void)
 256 {
 257         u64 ts = trace_clock_local();
 258
 259         (void)do_div(ts, NSEC_PER_USEC);
 260         return ts;
 261 }
 262 #else /* #ifdef CONFIG_RCU_TRACE */
 263 static u64 notrace rcu_trace_clock_local(void)
 264 {
 265         return 0ULL;
 266 }
 267 #endif /* #else #ifdef CONFIG_RCU_TRACE */
 268
 269 /*
 270  * Stop aggressive CPU-hog tests a bit before the end of the test in order
 271  * to avoid interfering with test shutdown.
 272  */
 273 static bool shutdown_time_arrived(void)
 274 {
 275         return shutdown_secs && time_after(jiffies, shutdown_jiffies - 30 * HZ);
 276 }
 277
 278 static unsigned long boost_starttime;   /* jiffies of next boost test start. */
 279 static DEFINE_MUTEX(boost_mutex);       /* protect setting boost_starttime */
 280                                         /*  and boost task create/destroy. */
 281 static atomic_t barrier_cbs_count;      /* Barrier callbacks registered. */
 282 static bool barrier_phase;              /* Test phase. */
 283 static atomic_t barrier_cbs_invoked;    /* Barrier callbacks invoked. */
 284 static wait_queue_head_t *barrier_cbs_wq; /* Coordinate barrier testing. */
 285 static DECLARE_WAIT_QUEUE_HEAD(barrier_wq);
 286
 287 static atomic_t rcu_fwd_cb_nodelay;     /* Short rcu_torture_delay() delays. */
 288
 289 /*
 290  * Allocate an element from the rcu_tortures pool.
 291  */
 292 static struct rcu_torture *
 293 rcu_torture_alloc(void)
 294 {
 295         struct list_head *p;
 296
 297         spin_lock_bh(&rcu_torture_lock);
 298         if (list_empty(&rcu_torture_freelist)) {
 299                 atomic_inc(&n_rcu_torture_alloc_fail);
 300                 spin_unlock_bh(&rcu_torture_lock);
 301                 return NULL;
 302         }
 303         atomic_inc(&n_rcu_torture_alloc);
 304         p = rcu_torture_freelist.next;
 305         list_del_init(p);
 306         spin_unlock_bh(&rcu_torture_lock);
 307         return container_of(p, struct rcu_torture, rtort_free);
 308 }
 309
 310 /*
 311  * Free an element to the rcu_tortures pool.
 312  */
 313 static void
 314 rcu_torture_free(struct rcu_torture *p)
 315 {
 316         atomic_inc(&n_rcu_torture_free);
 317         spin_lock_bh(&rcu_torture_lock);
 318         list_add_tail(&p->rtort_free, &rcu_torture_freelist);
 319         spin_unlock_bh(&rcu_torture_lock);
 320 }
 321
 322 /*
 323  * Operations vector for selecting different types of tests.
 324  */
 325
 326 struct rcu_torture_ops {
 327         int ttype;
 328         void (*init)(void);
 329         void (*cleanup)(void);
 330         int (*readlock)(void);
 331         void (*read_delay)(struct torture_random_state *rrsp,
 332                            struct rt_read_seg *rtrsp);
 333         void (*readunlock)(int idx);
 334         int (*readlock_held)(void);
 335         unsigned long (*get_gp_seq)(void);
 336         unsigned long (*gp_diff)(unsigned long new, unsigned long old);
 337         void (*deferred_free)(struct rcu_torture *p);
 338         void (*sync)(void);
 339         void (*exp_sync)(void);
 340         unsigned long (*get_gp_state)(void);
 341         unsigned long (*start_gp_poll)(void);
 342         bool (*poll_gp_state)(unsigned long oldstate);
 343         void (*cond_sync)(unsigned long oldstate);
 344         call_rcu_func_t call;
 345         void (*cb_barrier)(void);
 346         void (*fqs)(void);
 347         void (*stats)(void);
 348         void (*gp_kthread_dbg)(void);
 349         bool (*check_boost_failed)(unsigned long gp_state, int *cpup);
 350         int (*stall_dur)(void);
 351         long cbflood_max;
 352         int irq_capable;
 353         int can_boost;
 354         int extendables;
 355         int slow_gps;
 356         int no_pi_lock;
 357         const char *name;
 358 };
 359
 360 static struct rcu_torture_ops *cur_ops;
 361
 362 /*
 363  * Definitions for rcu torture testing.
 364  */
 365
 366 static int torture_readlock_not_held(void)
 367 {
 368         return rcu_read_lock_bh_held() || rcu_read_lock_sched_held();
 369 }
 370
 371 static int rcu_torture_read_lock(void) __acquires(RCU)
 372 {
 373         rcu_read_lock();
 374         return 0;
 375 }
 376
 377 static void
 378 rcu_read_delay(struct torture_random_state *rrsp, struct rt_read_seg *rtrsp)
 379 {
 380         unsigned long started;
 381         unsigned long completed;
 382         const unsigned long shortdelay_us = 200;
 383         unsigned long longdelay_ms = 300;
 384         unsigned long long ts;
 385
 386         /* We want a short delay sometimes to make a reader delay the grace
 387          * period, and we want a long delay occasionally to trigger
 388          * force_quiescent_state. */
 389
 390         if (!atomic_read(&rcu_fwd_cb_nodelay) &&
 391             !(torture_random(rrsp) % (nrealreaders * 2000 * longdelay_ms))) {
 392                 started = cur_ops->get_gp_seq();
 393                 ts = rcu_trace_clock_local();
 394                 if (preempt_count() & (SOFTIRQ_MASK | HARDIRQ_MASK))
 395                         longdelay_ms = 5; /* Avoid triggering BH limits. */
 396                 mdelay(longdelay_ms);
 397                 rtrsp->rt_delay_ms = longdelay_ms;
 398                 completed = cur_ops->get_gp_seq();
 399                 do_trace_rcu_torture_read(cur_ops->name, NULL, ts,
 400                                           started, completed);
 401         }
 402         if (!(torture_random(rrsp) % (nrealreaders * 2 * shortdelay_us))) {
 403                 udelay(shortdelay_us);
 404                 rtrsp->rt_delay_us = shortdelay_us;
 405         }
 406         if (!preempt_count() &&
 407             !(torture_random(rrsp) % (nrealreaders * 500))) {
 408                 torture_preempt_schedule();  /* QS only if preemptible. */
 409                 rtrsp->rt_preempted = true;
 410         }
 411 }
 412
 413 static void rcu_torture_read_unlock(int idx) __releases(RCU)
 414 {
 415         rcu_read_unlock();
 416 }
 417
 418 /*
 419  * Update callback in the pipe.  This should be invoked after a grace period.
 420  */
 421 static bool
 422 rcu_torture_pipe_update_one(struct rcu_torture *rp)
 423 {
 424         int i;
 425         struct rcu_torture_reader_check *rtrcp = READ_ONCE(rp->rtort_chkp);
 426
 427         if (rtrcp) {
 428                 WRITE_ONCE(rp->rtort_chkp, NULL);
 429                 smp_store_release(&rtrcp->rtc_ready, 1); // Pair with smp_load_acquire().
 430         }
 431         i = READ_ONCE(rp->rtort_pipe_count);
 432         if (i > RCU_TORTURE_PIPE_LEN)
 433                 i = RCU_TORTURE_PIPE_LEN;
 434         atomic_inc(&rcu_torture_wcount[i]);
 435         WRITE_ONCE(rp->rtort_pipe_count, i + 1);
 436         if (rp->rtort_pipe_count >= RCU_TORTURE_PIPE_LEN) {
 437                 rp->rtort_mbtest = 0;
 438                 return true;
 439         }
 440         return false;
 441 }
 442
 443 /*
 444  * Update all callbacks in the pipe.  Suitable for synchronous grace-period
 445  * primitives.
 446  */
 447 static void
 448 rcu_torture_pipe_update(struct rcu_torture *old_rp)
 449 {
 450         struct rcu_torture *rp;
 451         struct rcu_torture *rp1;
 452
 453         if (old_rp)
 454                 list_add(&old_rp->rtort_free, &rcu_torture_removed);
 455         list_for_each_entry_safe(rp, rp1, &rcu_torture_removed, rtort_free) {
 456                 if (rcu_torture_pipe_update_one(rp)) {
 457                         list_del(&rp->rtort_free);
 458                         rcu_torture_free(rp);
 459                 }
 460         }
 461 }
 462
 463 static void
 464 rcu_torture_cb(struct rcu_head *p)
 465 {
 466         struct rcu_torture *rp = container_of(p, struct rcu_torture, rtort_rcu);
 467
 468         if (torture_must_stop_irq()) {
 469                 /* Test is ending, just drop callbacks on the floor. */
 470                 /* The next initialization will pick up the pieces. */
 471                 return;
 472         }
 473         if (rcu_torture_pipe_update_one(rp))
 474                 rcu_torture_free(rp);
 475         else
 476                 cur_ops->deferred_free(rp);
 477 }
 478
 479 static unsigned long rcu_no_completed(void)
 480 {
 481         return 0;
 482 }
 483
 484 static void rcu_torture_deferred_free(struct rcu_torture *p)
 485 {
 486         call_rcu(&p->rtort_rcu, rcu_torture_cb);
 487 }
 488
 489 static void rcu_sync_torture_init(void)
 490 {
 491         INIT_LIST_HEAD(&rcu_torture_removed);
 492 }
 493
 494 static struct rcu_torture_ops rcu_ops = {
 495         .ttype                  = RCU_FLAVOR,
 496         .init                   = rcu_sync_torture_init,
 497         .readlock               = rcu_torture_read_lock,
 498         .read_delay             = rcu_read_delay,
 499         .readunlock             = rcu_torture_read_unlock,
 500         .readlock_held          = torture_readlock_not_held,
 501         .get_gp_seq             = rcu_get_gp_seq,
 502         .gp_diff                = rcu_seq_diff,
 503         .deferred_free          = rcu_torture_deferred_free,
 504         .sync                   = synchronize_rcu,
 505         .exp_sync               = synchronize_rcu_expedited,
 506         .get_gp_state           = get_state_synchronize_rcu,
 507         .start_gp_poll          = start_poll_synchronize_rcu,
 508         .poll_gp_state          = poll_state_synchronize_rcu,
 509         .cond_sync              = cond_synchronize_rcu,
 510         .call                   = call_rcu,
 511         .cb_barrier             = rcu_barrier,
 512         .fqs                    = rcu_force_quiescent_state,
 513         .stats                  = NULL,
 514         .gp_kthread_dbg         = show_rcu_gp_kthreads,
 515         .check_boost_failed     = rcu_check_boost_fail,
 516         .stall_dur              = rcu_jiffies_till_stall_check,
 517         .irq_capable            = 1,
 518         .can_boost              = IS_ENABLED(CONFIG_RCU_BOOST),
 519         .extendables            = RCUTORTURE_MAX_EXTEND,
 520         .name                   = "rcu"
 521 };
 522
 523 /*
 524  * Don't even think about trying any of these in real life!!!
 525  * The names includes "busted", and they really means it!
 526  * The only purpose of these functions is to provide a buggy RCU
 527  * implementation to make sure that rcutorture correctly emits
 528  * buggy-RCU error messages.
 529  */
 530 static void rcu_busted_torture_deferred_free(struct rcu_torture *p)
 531 {
 532         /* This is a deliberate bug for testing purposes only! */
 533         rcu_torture_cb(&p->rtort_rcu);
 534 }
 535
 536 static void synchronize_rcu_busted(void)
 537 {
 538         /* This is a deliberate bug for testing purposes only! */
 539 }
 540
 541 static void
 542 call_rcu_busted(struct rcu_head *head, rcu_callback_t func)
 543 {
 544         /* This is a deliberate bug for testing purposes only! */
 545         func(head);
 546 }
 547
 548 static struct rcu_torture_ops rcu_busted_ops = {
 549         .ttype          = INVALID_RCU_FLAVOR,
 550         .init           = rcu_sync_torture_init,
 551         .readlock       = rcu_torture_read_lock,
 552         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 553         .readunlock     = rcu_torture_read_unlock,
 554         .readlock_held  = torture_readlock_not_held,
 555         .get_gp_seq     = rcu_no_completed,
 556         .deferred_free  = rcu_busted_torture_deferred_free,
 557         .sync           = synchronize_rcu_busted,
 558         .exp_sync       = synchronize_rcu_busted,
 559         .call           = call_rcu_busted,
 560         .cb_barrier     = NULL,
 561         .fqs            = NULL,
 562         .stats          = NULL,
 563         .irq_capable    = 1,
 564         .name           = "busted"
 565 };
 566
 567 /*
 568  * Definitions for srcu torture testing.
 569  */
 570
 571 DEFINE_STATIC_SRCU(srcu_ctl);
 572 static struct srcu_struct srcu_ctld;
 573 static struct srcu_struct *srcu_ctlp = &srcu_ctl;
 574
 575 static int srcu_torture_read_lock(void) __acquires(srcu_ctlp)
 576 {
 577         return srcu_read_lock(srcu_ctlp);
 578 }
 579
 580 static void
 581 srcu_read_delay(struct torture_random_state *rrsp, struct rt_read_seg *rtrsp)
 582 {
 583         long delay;
 584         const long uspertick = 1000000 / HZ;
 585         const long longdelay = 10;
 586
 587         /* We want there to be long-running readers, but not all the time. */
 588
 589         delay = torture_random(rrsp) %
 590                 (nrealreaders * 2 * longdelay * uspertick);
 591         if (!delay && in_task()) {
 592                 schedule_timeout_interruptible(longdelay);
 593                 rtrsp->rt_delay_jiffies = longdelay;
 594         } else {
 595                 rcu_read_delay(rrsp, rtrsp);
 596         }
 597 }
 598
 599 static void srcu_torture_read_unlock(int idx) __releases(srcu_ctlp)
 600 {
 601         srcu_read_unlock(srcu_ctlp, idx);
 602 }
 603
 604 static int torture_srcu_read_lock_held(void)
 605 {
 606         return srcu_read_lock_held(srcu_ctlp);
 607 }
 608
 609 static unsigned long srcu_torture_completed(void)
 610 {
 611         return srcu_batches_completed(srcu_ctlp);
 612 }
 613
 614 static void srcu_torture_deferred_free(struct rcu_torture *rp)
 615 {
 616         call_srcu(srcu_ctlp, &rp->rtort_rcu, rcu_torture_cb);
 617 }
 618
 619 static void srcu_torture_synchronize(void)
 620 {
 621         synchronize_srcu(srcu_ctlp);
 622 }
 623
 624 static unsigned long srcu_torture_get_gp_state(void)
 625 {
 626         return get_state_synchronize_srcu(srcu_ctlp);
 627 }
 628
 629 static unsigned long srcu_torture_start_gp_poll(void)
 630 {
 631         return start_poll_synchronize_srcu(srcu_ctlp);
 632 }
 633
 634 static bool srcu_torture_poll_gp_state(unsigned long oldstate)
 635 {
 636         return poll_state_synchronize_srcu(srcu_ctlp, oldstate);
 637 }
 638
 639 static void srcu_torture_call(struct rcu_head *head,
 640                               rcu_callback_t func)
 641 {
 642         call_srcu(srcu_ctlp, head, func);
 643 }
 644
 645 static void srcu_torture_barrier(void)
 646 {
 647         srcu_barrier(srcu_ctlp);
 648 }
 649
 650 static void srcu_torture_stats(void)
 651 {
 652         srcu_torture_stats_print(srcu_ctlp, torture_type, TORTURE_FLAG);
 653 }
 654
 655 static void srcu_torture_synchronize_expedited(void)
 656 {
 657         synchronize_srcu_expedited(srcu_ctlp);
 658 }
 659
 660 static struct rcu_torture_ops srcu_ops = {
 661         .ttype          = SRCU_FLAVOR,
 662         .init           = rcu_sync_torture_init,
 663         .readlock       = srcu_torture_read_lock,
 664         .read_delay     = srcu_read_delay,
 665         .readunlock     = srcu_torture_read_unlock,
 666         .readlock_held  = torture_srcu_read_lock_held,
 667         .get_gp_seq     = srcu_torture_completed,
 668         .deferred_free  = srcu_torture_deferred_free,
 669         .sync           = srcu_torture_synchronize,
 670         .exp_sync       = srcu_torture_synchronize_expedited,
 671         .get_gp_state   = srcu_torture_get_gp_state,
 672         .start_gp_poll  = srcu_torture_start_gp_poll,
 673         .poll_gp_state  = srcu_torture_poll_gp_state,
 674         .call           = srcu_torture_call,
 675         .cb_barrier     = srcu_torture_barrier,
 676         .stats          = srcu_torture_stats,
 677         .cbflood_max    = 50000,
 678         .irq_capable    = 1,
 679         .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
 680         .name           = "srcu"
 681 };
 682
 683 static void srcu_torture_init(void)
 684 {
 685         rcu_sync_torture_init();
 686         WARN_ON(init_srcu_struct(&srcu_ctld));
 687         srcu_ctlp = &srcu_ctld;
 688 }
 689
 690 static void srcu_torture_cleanup(void)
 691 {
 692         cleanup_srcu_struct(&srcu_ctld);
 693         srcu_ctlp = &srcu_ctl; /* In case of a later rcutorture run. */
 694 }
 695
 696 /* As above, but dynamically allocated. */
 697 static struct rcu_torture_ops srcud_ops = {
 698         .ttype          = SRCU_FLAVOR,
 699         .init           = srcu_torture_init,
 700         .cleanup        = srcu_torture_cleanup,
 701         .readlock       = srcu_torture_read_lock,
 702         .read_delay     = srcu_read_delay,
 703         .readunlock     = srcu_torture_read_unlock,
 704         .readlock_held  = torture_srcu_read_lock_held,
 705         .get_gp_seq     = srcu_torture_completed,
 706         .deferred_free  = srcu_torture_deferred_free,
 707         .sync           = srcu_torture_synchronize,
 708         .exp_sync       = srcu_torture_synchronize_expedited,
 709         .call           = srcu_torture_call,
 710         .cb_barrier     = srcu_torture_barrier,
 711         .stats          = srcu_torture_stats,
 712         .cbflood_max    = 50000,
 713         .irq_capable    = 1,
 714         .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
 715         .name           = "srcud"
 716 };
 717
 718 /* As above, but broken due to inappropriate reader extension. */
 719 static struct rcu_torture_ops busted_srcud_ops = {
 720         .ttype          = SRCU_FLAVOR,
 721         .init           = srcu_torture_init,
 722         .cleanup        = srcu_torture_cleanup,
 723         .readlock       = srcu_torture_read_lock,
 724         .read_delay     = rcu_read_delay,
 725         .readunlock     = srcu_torture_read_unlock,
 726         .readlock_held  = torture_srcu_read_lock_held,
 727         .get_gp_seq     = srcu_torture_completed,
 728         .deferred_free  = srcu_torture_deferred_free,
 729         .sync           = srcu_torture_synchronize,
 730         .exp_sync       = srcu_torture_synchronize_expedited,
 731         .call           = srcu_torture_call,
 732         .cb_barrier     = srcu_torture_barrier,
 733         .stats          = srcu_torture_stats,
 734         .irq_capable    = 1,
 735         .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
 736         .extendables    = RCUTORTURE_MAX_EXTEND,
 737         .name           = "busted_srcud"
 738 };
 739
 740 /*
 741  * Definitions for RCU-tasks torture testing.
 742  */
 743
 744 static int tasks_torture_read_lock(void)
 745 {
 746         return 0;
 747 }
 748
 749 static void tasks_torture_read_unlock(int idx)
 750 {
 751 }
 752
 753 static void rcu_tasks_torture_deferred_free(struct rcu_torture *p)
 754 {
 755         call_rcu_tasks(&p->rtort_rcu, rcu_torture_cb);
 756 }
 757
 758 static void synchronize_rcu_mult_test(void)
 759 {
 760         synchronize_rcu_mult(call_rcu_tasks, call_rcu);
 761 }
 762
 763 static struct rcu_torture_ops tasks_ops = {
 764         .ttype          = RCU_TASKS_FLAVOR,
 765         .init           = rcu_sync_torture_init,
 766         .readlock       = tasks_torture_read_lock,
 767         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 768         .readunlock     = tasks_torture_read_unlock,
 769         .get_gp_seq     = rcu_no_completed,
 770         .deferred_free  = rcu_tasks_torture_deferred_free,
 771         .sync           = synchronize_rcu_tasks,
 772         .exp_sync       = synchronize_rcu_mult_test,
 773         .call           = call_rcu_tasks,
 774         .cb_barrier     = rcu_barrier_tasks,
 775         .gp_kthread_dbg = show_rcu_tasks_classic_gp_kthread,
 776         .fqs            = NULL,
 777         .stats          = NULL,
 778         .irq_capable    = 1,
 779         .slow_gps       = 1,
 780         .name           = "tasks"
 781 };
 782
 783 /*
 784  * Definitions for trivial CONFIG_PREEMPT=n-only torture testing.
 785  * This implementation does not necessarily work well with CPU hotplug.
 786  */
 787
 788 static void synchronize_rcu_trivial(void)
 789 {
 790         int cpu;
 791
 792         for_each_online_cpu(cpu) {
 793                 rcutorture_sched_setaffinity(current->pid, cpumask_of(cpu));
 794                 WARN_ON_ONCE(raw_smp_processor_id() != cpu);
 795         }
 796 }
 797
 798 static int rcu_torture_read_lock_trivial(void) __acquires(RCU)
 799 {
 800         preempt_disable();
 801         return 0;
 802 }
 803
 804 static void rcu_torture_read_unlock_trivial(int idx) __releases(RCU)
 805 {
 806         preempt_enable();
 807 }
 808
 809 static struct rcu_torture_ops trivial_ops = {
 810         .ttype          = RCU_TRIVIAL_FLAVOR,
 811         .init           = rcu_sync_torture_init,
 812         .readlock       = rcu_torture_read_lock_trivial,
 813         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 814         .readunlock     = rcu_torture_read_unlock_trivial,
 815         .readlock_held  = torture_readlock_not_held,
 816         .get_gp_seq     = rcu_no_completed,
 817         .sync           = synchronize_rcu_trivial,
 818         .exp_sync       = synchronize_rcu_trivial,
 819         .fqs            = NULL,
 820         .stats          = NULL,
 821         .irq_capable    = 1,
 822         .name           = "trivial"
 823 };
 824
 825 /*
 826  * Definitions for rude RCU-tasks torture testing.
 827  */
 828
 829 static void rcu_tasks_rude_torture_deferred_free(struct rcu_torture *p)
 830 {
 831         call_rcu_tasks_rude(&p->rtort_rcu, rcu_torture_cb);
 832 }
 833
 834 static struct rcu_torture_ops tasks_rude_ops = {
 835         .ttype          = RCU_TASKS_RUDE_FLAVOR,
 836         .init           = rcu_sync_torture_init,
 837         .readlock       = rcu_torture_read_lock_trivial,
 838         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 839         .readunlock     = rcu_torture_read_unlock_trivial,
 840         .get_gp_seq     = rcu_no_completed,
 841         .deferred_free  = rcu_tasks_rude_torture_deferred_free,
 842         .sync           = synchronize_rcu_tasks_rude,
 843         .exp_sync       = synchronize_rcu_tasks_rude,
 844         .call           = call_rcu_tasks_rude,
 845         .cb_barrier     = rcu_barrier_tasks_rude,
 846         .gp_kthread_dbg = show_rcu_tasks_rude_gp_kthread,
 847         .cbflood_max    = 50000,
 848         .fqs            = NULL,
 849         .stats          = NULL,
 850         .irq_capable    = 1,
 851         .name           = "tasks-rude"
 852 };
 853
 854 /*
 855  * Definitions for tracing RCU-tasks torture testing.
 856  */
 857
 858 static int tasks_tracing_torture_read_lock(void)
 859 {
 860         rcu_read_lock_trace();
 861         return 0;
 862 }
 863
 864 static void tasks_tracing_torture_read_unlock(int idx)
 865 {
 866         rcu_read_unlock_trace();
 867 }
 868
 869 static void rcu_tasks_tracing_torture_deferred_free(struct rcu_torture *p)
 870 {
 871         call_rcu_tasks_trace(&p->rtort_rcu, rcu_torture_cb);
 872 }
 873
 874 static struct rcu_torture_ops tasks_tracing_ops = {
 875         .ttype          = RCU_TASKS_TRACING_FLAVOR,
 876         .init           = rcu_sync_torture_init,
 877         .readlock       = tasks_tracing_torture_read_lock,
 878         .read_delay     = srcu_read_delay,  /* just reuse srcu's version. */
 879         .readunlock     = tasks_tracing_torture_read_unlock,
 880         .readlock_held  = rcu_read_lock_trace_held,
 881         .get_gp_seq     = rcu_no_completed,
 882         .deferred_free  = rcu_tasks_tracing_torture_deferred_free,
 883         .sync           = synchronize_rcu_tasks_trace,
 884         .exp_sync       = synchronize_rcu_tasks_trace,
 885         .call           = call_rcu_tasks_trace,
 886         .cb_barrier     = rcu_barrier_tasks_trace,
 887         .gp_kthread_dbg = show_rcu_tasks_trace_gp_kthread,
 888         .cbflood_max    = 50000,
 889         .fqs            = NULL,
 890         .stats          = NULL,
 891         .irq_capable    = 1,
 892         .slow_gps       = 1,
 893         .name           = "tasks-tracing"
 894 };
 895
 896 static unsigned long rcutorture_seq_diff(unsigned long new, unsigned long old)
 897 {
 898         if (!cur_ops->gp_diff)
 899                 return new - old;
 900         return cur_ops->gp_diff(new, old);
 901 }
 902
 903 /*
 904  * RCU torture priority-boost testing.  Runs one real-time thread per
 905  * CPU for moderate bursts, repeatedly starting grace periods and waiting
 906  * for them to complete.  If a given grace period takes too long, we assume
 907  * that priority inversion has occurred.
 908  */
 909
 910 static int old_rt_runtime = -1;
 911
 912 static void rcu_torture_disable_rt_throttle(void)
 913 {
 914         /*
 915          * Disable RT throttling so that rcutorture's boost threads don't get
 916          * throttled. Only possible if rcutorture is built-in otherwise the
 917          * user should manually do this by setting the sched_rt_period_us and
 918          * sched_rt_runtime sysctls.
 919          */
 920         if (!IS_BUILTIN(CONFIG_RCU_TORTURE_TEST) || old_rt_runtime != -1)
 921                 return;
 922
 923         old_rt_runtime = sysctl_sched_rt_runtime;
 924         sysctl_sched_rt_runtime = -1;
 925 }
 926
 927 static void rcu_torture_enable_rt_throttle(void)
 928 {
 929         if (!IS_BUILTIN(CONFIG_RCU_TORTURE_TEST) || old_rt_runtime == -1)
 930                 return;
 931
 932         sysctl_sched_rt_runtime = old_rt_runtime;
 933         old_rt_runtime = -1;
 934 }
 935
 936 static bool rcu_torture_boost_failed(unsigned long gp_state, unsigned long *start)
 937 {
 938         int cpu;
 939         static int dbg_done;
 940         unsigned long end = jiffies;
 941         bool gp_done;
 942         unsigned long j;
 943         static unsigned long last_persist;
 944         unsigned long lp;
 945         unsigned long mininterval = test_boost_duration * HZ - HZ / 2;
 946
 947         if (end - *start > mininterval) {
 948                 // Recheck after checking time to avoid false positives.
 949                 smp_mb(); // Time check before grace-period check.
 950                 if (cur_ops->poll_gp_state(gp_state))
 951                         return false; // passed, though perhaps just barely
 952                 if (cur_ops->check_boost_failed && !cur_ops->check_boost_failed(gp_state, &cpu)) {
 953                         // At most one persisted message per boost test.
 954                         j = jiffies;
 955                         lp = READ_ONCE(last_persist);
 956                         if (time_after(j, lp + mininterval) && cmpxchg(&last_persist, lp, j) == lp)
 957                                 pr_info("Boost inversion persisted: No QS from CPU %d\n", cpu);
 958                         return false; // passed on a technicality
 959                 }
 960                 VERBOSE_TOROUT_STRING("rcu_torture_boost boosting failed");
 961                 n_rcu_torture_boost_failure++;
 962                 if (!xchg(&dbg_done, 1) && cur_ops->gp_kthread_dbg) {
 963                         pr_info("Boost inversion thread ->rt_priority %u gp_state %lu jiffies %lu\n",
 964                                 current->rt_priority, gp_state, end - *start);
 965                         cur_ops->gp_kthread_dbg();
 966                         // Recheck after print to flag grace period ending during splat.
 967                         gp_done = cur_ops->poll_gp_state(gp_state);
 968                         pr_info("Boost inversion: GP %lu %s.\n", gp_state,
 969                                 gp_done ? "ended already" : "still pending");
 970
 971                 }
 972
 973                 return true; // failed
 974         } else if (cur_ops->check_boost_failed && !cur_ops->check_boost_failed(gp_state, NULL)) {
 975                 *start = jiffies;
 976         }
 977
 978         return false; // passed
 979 }
 980
 981 static int rcu_torture_boost(void *arg)
 982 {
 983         unsigned long endtime;
 984         unsigned long gp_state;
 985         unsigned long gp_state_time;
 986         unsigned long oldstarttime;
 987
 988         VERBOSE_TOROUT_STRING("rcu_torture_boost started");
 989
 990         /* Set real-time priority. */
 991         sched_set_fifo_low(current);
 992
 993         /* Each pass through the following loop does one boost-test cycle. */
 994         do {
 995                 bool failed = false; // Test failed already in this test interval
 996                 bool gp_initiated = false;
 997
 998                 if (kthread_should_stop())
 999                         goto checkwait;
1000
1001                 /* Wait for the next test interval. */
1002                 oldstarttime = READ_ONCE(boost_starttime);
1003                 while (time_before(jiffies, oldstarttime)) {
1004                         schedule_timeout_interruptible(oldstarttime - jiffies);
1005                         if (stutter_wait("rcu_torture_boost"))
1006                                 sched_set_fifo_low(current);
1007                         if (torture_must_stop())
1008                                 goto checkwait;
1009                 }
1010
1011                 // Do one boost-test interval.
1012                 endtime = oldstarttime + test_boost_duration * HZ;
1013                 while (time_before(jiffies, endtime)) {
1014                         // Has current GP gone too long?
1015                         if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
1016                                 failed = rcu_torture_boost_failed(gp_state, &gp_state_time);
1017                         // If we don't have a grace period in flight, start one.
1018                         if (!gp_initiated || cur_ops->poll_gp_state(gp_state)) {
1019                                 gp_state = cur_ops->start_gp_poll();
1020                                 gp_initiated = true;
1021                                 gp_state_time = jiffies;
1022                         }
1023                         if (stutter_wait("rcu_torture_boost")) {
1024                                 sched_set_fifo_low(current);
1025                                 // If the grace period already ended,
1026                                 // we don't know when that happened, so
1027                                 // start over.
1028                                 if (cur_ops->poll_gp_state(gp_state))
1029                                         gp_initiated = false;
1030                         }
1031                         if (torture_must_stop())
1032                                 goto checkwait;
1033                 }
1034
1035                 // In case the grace period extended beyond the end of the loop.
1036                 if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
1037                         rcu_torture_boost_failed(gp_state, &gp_state_time);
1038
1039                 /*
1040                  * Set the start time of the next test interval.
1041                  * Yes, this is vulnerable to long delays, but such
1042                  * delays simply cause a false negative for the next
1043                  * interval.  Besides, we are running at RT priority,
1044                  * so delays should be relatively rare.
1045                  */
1046                 while (oldstarttime == READ_ONCE(boost_starttime) && !kthread_should_stop()) {
1047                         if (mutex_trylock(&boost_mutex)) {
1048                                 if (oldstarttime == boost_starttime) {
1049                                         WRITE_ONCE(boost_starttime,
1050                                                    jiffies + test_boost_interval * HZ);
1051                                         n_rcu_torture_boosts++;
1052                                 }
1053                                 mutex_unlock(&boost_mutex);
1054                                 break;
1055                         }
1056                         schedule_timeout_uninterruptible(1);
1057                 }
1058
1059                 /* Go do the stutter. */
1060 checkwait:      if (stutter_wait("rcu_torture_boost"))
1061                         sched_set_fifo_low(current);
1062         } while (!torture_must_stop());
1063
1064         /* Clean up and exit. */
1065         while (!kthread_should_stop()) {
1066                 torture_shutdown_absorb("rcu_torture_boost");
1067                 schedule_timeout_uninterruptible(1);
1068         }
1069         torture_kthread_stopping("rcu_torture_boost");
1070         return 0;
1071 }
1072
1073 /*
1074  * RCU torture force-quiescent-state kthread.  Repeatedly induces
1075  * bursts of calls to force_quiescent_state(), increasing the probability
1076  * of occurrence of some important types of race conditions.
1077  */
1078 static int
1079 rcu_torture_fqs(void *arg)
1080 {
1081         unsigned long fqs_resume_time;
1082         int fqs_burst_remaining;
1083         int oldnice = task_nice(current);
1084
1085         VERBOSE_TOROUT_STRING("rcu_torture_fqs task started");
1086         do {
1087                 fqs_resume_time = jiffies + fqs_stutter * HZ;
1088                 while (time_before(jiffies, fqs_resume_time) &&
1089                        !kthread_should_stop()) {
1090                         schedule_timeout_interruptible(1);
1091                 }
1092                 fqs_burst_remaining = fqs_duration;
1093                 while (fqs_burst_remaining > 0 &&
1094                        !kthread_should_stop()) {
1095                         cur_ops->fqs();
1096                         udelay(fqs_holdoff);
1097                         fqs_burst_remaining -= fqs_holdoff;
1098                 }
1099                 if (stutter_wait("rcu_torture_fqs"))
1100                         sched_set_normal(current, oldnice);
1101         } while (!torture_must_stop());
1102         torture_kthread_stopping("rcu_torture_fqs");
1103         return 0;
1104 }
1105
1106 // Used by writers to randomly choose from the available grace-period
1107 // primitives.  The only purpose of the initialization is to size the array.
1108 static int synctype[] = { RTWS_DEF_FREE, RTWS_EXP_SYNC, RTWS_COND_GET, RTWS_POLL_GET, RTWS_SYNC };
1109 static int nsynctypes;
1110
1111 /*
1112  * Determine which grace-period primitives are available.
1113  */
1114 static void rcu_torture_write_types(void)
1115 {
1116         bool gp_cond1 = gp_cond, gp_exp1 = gp_exp, gp_normal1 = gp_normal;
1117         bool gp_poll1 = gp_poll, gp_sync1 = gp_sync;
1118
1119         /* Initialize synctype[] array.  If none set, take default. */
1120         if (!gp_cond1 && !gp_exp1 && !gp_normal1 && !gp_poll1 && !gp_sync1)
1121                 gp_cond1 = gp_exp1 = gp_normal1 = gp_poll1 = gp_sync1 = true;
1122         if (gp_cond1 && cur_ops->get_gp_state && cur_ops->cond_sync) {
1123                 synctype[nsynctypes++] = RTWS_COND_GET;
1124                 pr_info("%s: Testing conditional GPs.\n", __func__);
1125         } else if (gp_cond && (!cur_ops->get_gp_state || !cur_ops->cond_sync)) {
1126                 pr_alert("%s: gp_cond without primitives.\n", __func__);
1127         }
1128         if (gp_exp1 && cur_ops->exp_sync) {
1129                 synctype[nsynctypes++] = RTWS_EXP_SYNC;
1130                 pr_info("%s: Testing expedited GPs.\n", __func__);
1131         } else if (gp_exp && !cur_ops->exp_sync) {
1132                 pr_alert("%s: gp_exp without primitives.\n", __func__);
1133         }
1134         if (gp_normal1 && cur_ops->deferred_free) {
1135                 synctype[nsynctypes++] = RTWS_DEF_FREE;
1136                 pr_info("%s: Testing asynchronous GPs.\n", __func__);
1137         } else if (gp_normal && !cur_ops->deferred_free) {
1138                 pr_alert("%s: gp_normal without primitives.\n", __func__);
1139         }
1140         if (gp_poll1 && cur_ops->start_gp_poll && cur_ops->poll_gp_state) {
1141                 synctype[nsynctypes++] = RTWS_POLL_GET;
1142                 pr_info("%s: Testing polling GPs.\n", __func__);
1143         } else if (gp_poll && (!cur_ops->start_gp_poll || !cur_ops->poll_gp_state)) {
1144                 pr_alert("%s: gp_poll without primitives.\n", __func__);
1145         }
1146         if (gp_sync1 && cur_ops->sync) {
1147                 synctype[nsynctypes++] = RTWS_SYNC;
1148                 pr_info("%s: Testing normal GPs.\n", __func__);
1149         } else if (gp_sync && !cur_ops->sync) {
1150                 pr_alert("%s: gp_sync without primitives.\n", __func__);
1151         }
1152 }
1153
1154 /*
1155  * RCU torture writer kthread.  Repeatedly substitutes a new structure
1156  * for that pointed to by rcu_torture_current, freeing the old structure
1157  * after a series of grace periods (the "pipeline").
1158  */
1159 static int
1160 rcu_torture_writer(void *arg)
1161 {
1162         bool boot_ended;
1163         bool can_expedite = !rcu_gp_is_expedited() && !rcu_gp_is_normal();
1164         unsigned long cookie;
1165         int expediting = 0;
1166         unsigned long gp_snap;
1167         int i;
1168         int idx;
1169         int oldnice = task_nice(current);
1170         struct rcu_torture *rp;
1171         struct rcu_torture *old_rp;
1172         static DEFINE_TORTURE_RANDOM(rand);
1173         bool stutter_waited;
1174
1175         VERBOSE_TOROUT_STRING("rcu_torture_writer task started");
1176         if (!can_expedite)
1177                 pr_alert("%s" TORTURE_FLAG
1178                          " GP expediting controlled from boot/sysfs for %s.\n",
1179                          torture_type, cur_ops->name);
1180         if (WARN_ONCE(nsynctypes == 0,
1181                       "rcu_torture_writer: No update-side primitives.\n")) {
1182                 /*
1183                  * No updates primitives, so don't try updating.
1184                  * The resulting test won't be testing much, hence the
1185                  * above WARN_ONCE().
1186                  */
1187                 rcu_torture_writer_state = RTWS_STOPPING;
1188                 torture_kthread_stopping("rcu_torture_writer");
1189         }
1190
1191         do {
1192                 rcu_torture_writer_state = RTWS_FIXED_DELAY;
1193                 torture_hrtimeout_us(500, 1000, &rand);
1194                 rp = rcu_torture_alloc();
1195                 if (rp == NULL)
1196                         continue;
1197                 rp->rtort_pipe_count = 0;
1198                 rcu_torture_writer_state = RTWS_DELAY;
1199                 udelay(torture_random(&rand) & 0x3ff);
1200                 rcu_torture_writer_state = RTWS_REPLACE;
1201                 old_rp = rcu_dereference_check(rcu_torture_current,
1202                                                current == writer_task);
1203                 rp->rtort_mbtest = 1;
1204                 rcu_assign_pointer(rcu_torture_current, rp);
1205                 smp_wmb(); /* Mods to old_rp must follow rcu_assign_pointer() */
1206                 if (old_rp) {
1207                         i = old_rp->rtort_pipe_count;
1208                         if (i > RCU_TORTURE_PIPE_LEN)
1209                                 i = RCU_TORTURE_PIPE_LEN;
1210                         atomic_inc(&rcu_torture_wcount[i]);
1211                         WRITE_ONCE(old_rp->rtort_pipe_count,
1212                                    old_rp->rtort_pipe_count + 1);
1213                         if (cur_ops->get_gp_state && cur_ops->poll_gp_state) {
1214                                 idx = cur_ops->readlock();
1215                                 cookie = cur_ops->get_gp_state();
1216                                 WARN_ONCE(rcu_torture_writer_state != RTWS_DEF_FREE &&
1217                                           cur_ops->poll_gp_state(cookie),
1218                                           "%s: Cookie check 1 failed %s(%d) %lu->%lu\n",
1219                                           __func__,
1220                                           rcu_torture_writer_state_getname(),
1221                                           rcu_torture_writer_state,
1222                                           cookie, cur_ops->get_gp_state());
1223                                 cur_ops->readunlock(idx);
1224                         }
1225                         switch (synctype[torture_random(&rand) % nsynctypes]) {
1226                         case RTWS_DEF_FREE:
1227                                 rcu_torture_writer_state = RTWS_DEF_FREE;
1228                                 cur_ops->deferred_free(old_rp);
1229                                 break;
1230                         case RTWS_EXP_SYNC:
1231                                 rcu_torture_writer_state = RTWS_EXP_SYNC;
1232                                 cur_ops->exp_sync();
1233                                 rcu_torture_pipe_update(old_rp);
1234                                 break;
1235                         case RTWS_COND_GET:
1236                                 rcu_torture_writer_state = RTWS_COND_GET;
1237                                 gp_snap = cur_ops->get_gp_state();
1238                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1239                                 rcu_torture_writer_state = RTWS_COND_SYNC;
1240                                 cur_ops->cond_sync(gp_snap);
1241                                 rcu_torture_pipe_update(old_rp);
1242                                 break;
1243                         case RTWS_POLL_GET:
1244                                 rcu_torture_writer_state = RTWS_POLL_GET;
1245                                 gp_snap = cur_ops->start_gp_poll();
1246                                 rcu_torture_writer_state = RTWS_POLL_WAIT;
1247                                 while (!cur_ops->poll_gp_state(gp_snap))
1248                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1249                                                                   &rand);
1250                                 rcu_torture_pipe_update(old_rp);
1251                                 break;
1252                         case RTWS_SYNC:
1253                                 rcu_torture_writer_state = RTWS_SYNC;
1254                                 cur_ops->sync();
1255                                 rcu_torture_pipe_update(old_rp);
1256                                 break;
1257                         default:
1258                                 WARN_ON_ONCE(1);
1259                                 break;
1260                         }
1261                 }
1262                 WRITE_ONCE(rcu_torture_current_version,
1263                            rcu_torture_current_version + 1);
1264                 /* Cycle through nesting levels of rcu_expedite_gp() calls. */
1265                 if (can_expedite &&
1266                     !(torture_random(&rand) & 0xff & (!!expediting - 1))) {
1267                         WARN_ON_ONCE(expediting == 0 && rcu_gp_is_expedited());
1268                         if (expediting >= 0)
1269                                 rcu_expedite_gp();
1270                         else
1271                                 rcu_unexpedite_gp();
1272                         if (++expediting > 3)
1273                                 expediting = -expediting;
1274                 } else if (!can_expedite) { /* Disabled during boot, recheck. */
1275                         can_expedite = !rcu_gp_is_expedited() &&
1276                                        !rcu_gp_is_normal();
1277                 }
1278                 rcu_torture_writer_state = RTWS_STUTTER;
1279                 boot_ended = rcu_inkernel_boot_has_ended();
1280                 stutter_waited = stutter_wait("rcu_torture_writer");
1281                 if (stutter_waited &&
1282                     !atomic_read(&rcu_fwd_cb_nodelay) &&
1283                     !cur_ops->slow_gps &&
1284                     !torture_must_stop() &&
1285                     boot_ended)
1286                         for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++)
1287                                 if (list_empty(&rcu_tortures[i].rtort_free) &&
1288                                     rcu_access_pointer(rcu_torture_current) !=
1289                                     &rcu_tortures[i]) {
1290                                         rcu_ftrace_dump(DUMP_ALL);
1291                                         WARN(1, "%s: rtort_pipe_count: %d\n", __func__, rcu_tortures[i].rtort_pipe_count);
1292                                 }
1293                 if (stutter_waited)
1294                         sched_set_normal(current, oldnice);
1295         } while (!torture_must_stop());
1296         rcu_torture_current = NULL;  // Let stats task know that we are done.
1297         /* Reset expediting back to unexpedited. */
1298         if (expediting > 0)
1299                 expediting = -expediting;
1300         while (can_expedite && expediting++ < 0)
1301                 rcu_unexpedite_gp();
1302         WARN_ON_ONCE(can_expedite && rcu_gp_is_expedited());
1303         if (!can_expedite)
1304                 pr_alert("%s" TORTURE_FLAG
1305                          " Dynamic grace-period expediting was disabled.\n",
1306                          torture_type);
1307         rcu_torture_writer_state = RTWS_STOPPING;
1308         torture_kthread_stopping("rcu_torture_writer");
1309         return 0;
1310 }
1311
1312 /*
1313  * RCU torture fake writer kthread.  Repeatedly calls sync, with a random
1314  * delay between calls.
1315  */
1316 static int
1317 rcu_torture_fakewriter(void *arg)
1318 {
1319         unsigned long gp_snap;
1320         DEFINE_TORTURE_RANDOM(rand);
1321
1322         VERBOSE_TOROUT_STRING("rcu_torture_fakewriter task started");
1323         set_user_nice(current, MAX_NICE);
1324
1325         do {
1326                 torture_hrtimeout_jiffies(torture_random(&rand) % 10, &rand);
1327                 if (cur_ops->cb_barrier != NULL &&
1328                     torture_random(&rand) % (nfakewriters * 8) == 0) {
1329                         cur_ops->cb_barrier();
1330                 } else {
1331                         switch (synctype[torture_random(&rand) % nsynctypes]) {
1332                         case RTWS_DEF_FREE:
1333                                 break;
1334                         case RTWS_EXP_SYNC:
1335                                 cur_ops->exp_sync();
1336                                 break;
1337                         case RTWS_COND_GET:
1338                                 gp_snap = cur_ops->get_gp_state();
1339                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1340                                 cur_ops->cond_sync(gp_snap);
1341                                 break;
1342                         case RTWS_POLL_GET:
1343                                 gp_snap = cur_ops->start_gp_poll();
1344                                 while (!cur_ops->poll_gp_state(gp_snap)) {
1345                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1346                                                                   &rand);
1347                                 }
1348                                 break;
1349                         case RTWS_SYNC:
1350                                 cur_ops->sync();
1351                                 break;
1352                         default:
1353                                 WARN_ON_ONCE(1);
1354                                 break;
1355                         }
1356                 }
1357                 stutter_wait("rcu_torture_fakewriter");
1358         } while (!torture_must_stop());
1359
1360         torture_kthread_stopping("rcu_torture_fakewriter");
1361         return 0;
1362 }
1363
1364 static void rcu_torture_timer_cb(struct rcu_head *rhp)
1365 {
1366         kfree(rhp);
1367 }
1368
1369 // Set up and carry out testing of RCU's global memory ordering
1370 static void rcu_torture_reader_do_mbchk(long myid, struct rcu_torture *rtp,
1371                                         struct torture_random_state *trsp)
1372 {
1373         unsigned long loops;
1374         int noc = torture_num_online_cpus();
1375         int rdrchked;
1376         int rdrchker;
1377         struct rcu_torture_reader_check *rtrcp; // Me.
1378         struct rcu_torture_reader_check *rtrcp_assigner; // Assigned us to do checking.
1379         struct rcu_torture_reader_check *rtrcp_chked; // Reader being checked.
1380         struct rcu_torture_reader_check *rtrcp_chker; // Reader doing checking when not me.
1381
1382         if (myid < 0)
1383                 return; // Don't try this from timer handlers.
1384
1385         // Increment my counter.
1386         rtrcp = &rcu_torture_reader_mbchk[myid];
1387         WRITE_ONCE(rtrcp->rtc_myloops, rtrcp->rtc_myloops + 1);
1388
1389         // Attempt to assign someone else some checking work.
1390         rdrchked = torture_random(trsp) % nrealreaders;
1391         rtrcp_chked = &rcu_torture_reader_mbchk[rdrchked];
1392         rdrchker = torture_random(trsp) % nrealreaders;
1393         rtrcp_chker = &rcu_torture_reader_mbchk[rdrchker];
1394         if (rdrchked != myid && rdrchked != rdrchker && noc >= rdrchked && noc >= rdrchker &&
1395             smp_load_acquire(&rtrcp->rtc_chkrdr) < 0 && // Pairs with smp_store_release below.
1396             !READ_ONCE(rtp->rtort_chkp) &&
1397             !smp_load_acquire(&rtrcp_chker->rtc_assigner)) { // Pairs with smp_store_release below.
1398                 rtrcp->rtc_chkloops = READ_ONCE(rtrcp_chked->rtc_myloops);
1399                 WARN_ON_ONCE(rtrcp->rtc_chkrdr >= 0);
1400                 rtrcp->rtc_chkrdr = rdrchked;
1401                 WARN_ON_ONCE(rtrcp->rtc_ready); // This gets set after the grace period ends.
1402                 if (cmpxchg_relaxed(&rtrcp_chker->rtc_assigner, NULL, rtrcp) ||
1403                     cmpxchg_relaxed(&rtp->rtort_chkp, NULL, rtrcp))
1404                         (void)cmpxchg_relaxed(&rtrcp_chker->rtc_assigner, rtrcp, NULL); // Back out.
1405         }
1406
1407         // If assigned some completed work, do it!
1408         rtrcp_assigner = READ_ONCE(rtrcp->rtc_assigner);
1409         if (!rtrcp_assigner || !smp_load_acquire(&rtrcp_assigner->rtc_ready))
1410                 return; // No work or work not yet ready.
1411         rdrchked = rtrcp_assigner->rtc_chkrdr;
1412         if (WARN_ON_ONCE(rdrchked < 0))
1413                 return;
1414         rtrcp_chked = &rcu_torture_reader_mbchk[rdrchked];
1415         loops = READ_ONCE(rtrcp_chked->rtc_myloops);
1416         atomic_inc(&n_rcu_torture_mbchk_tries);
1417         if (ULONG_CMP_LT(loops, rtrcp_assigner->rtc_chkloops))
1418                 atomic_inc(&n_rcu_torture_mbchk_fail);
1419         rtrcp_assigner->rtc_chkloops = loops + ULONG_MAX / 2;
1420         rtrcp_assigner->rtc_ready = 0;
1421         smp_store_release(&rtrcp->rtc_assigner, NULL); // Someone else can assign us work.
1422         smp_store_release(&rtrcp_assigner->rtc_chkrdr, -1); // Assigner can again assign.
1423 }
1424
1425 /*
1426  * Do one extension of an RCU read-side critical section using the
1427  * current reader state in readstate (set to zero for initial entry
1428  * to extended critical section), set the new state as specified by
1429  * newstate (set to zero for final exit from extended critical section),
1430  * and random-number-generator state in trsp.  If this is neither the
1431  * beginning or end of the critical section and if there was actually a
1432  * change, do a ->read_delay().
1433  */
1434 static void rcutorture_one_extend(int *readstate, int newstate,
1435                                   struct torture_random_state *trsp,
1436                                   struct rt_read_seg *rtrsp)
1437 {
1438         unsigned long flags;
1439         int idxnew1 = -1;
1440         int idxnew2 = -1;
1441         int idxold1 = *readstate;
1442         int idxold2 = idxold1;
1443         int statesnew = ~*readstate & newstate;
1444         int statesold = *readstate & ~newstate;
1445
1446         WARN_ON_ONCE(idxold2 < 0);
1447         WARN_ON_ONCE((idxold2 >> RCUTORTURE_RDR_SHIFT_2) > 1);
1448         rtrsp->rt_readstate = newstate;
1449
1450         /* First, put new protection in place to avoid critical-section gap. */
1451         if (statesnew & RCUTORTURE_RDR_BH)
1452                 local_bh_disable();
1453         if (statesnew & RCUTORTURE_RDR_RBH)
1454                 rcu_read_lock_bh();
1455         if (statesnew & RCUTORTURE_RDR_IRQ)
1456                 local_irq_disable();
1457         if (statesnew & RCUTORTURE_RDR_PREEMPT)
1458                 preempt_disable();
1459         if (statesnew & RCUTORTURE_RDR_SCHED)
1460                 rcu_read_lock_sched();
1461         if (statesnew & RCUTORTURE_RDR_RCU_1)
1462                 idxnew1 = (cur_ops->readlock() & 0x1) << RCUTORTURE_RDR_SHIFT_1;
1463         if (statesnew & RCUTORTURE_RDR_RCU_2)
1464                 idxnew2 = (cur_ops->readlock() & 0x1) << RCUTORTURE_RDR_SHIFT_2;
1465
1466         /*
1467          * Next, remove old protection, in decreasing order of strength
1468          * to avoid unlock paths that aren't safe in the stronger
1469          * context. Namely: BH can not be enabled with disabled interrupts.
1470          * Additionally PREEMPT_RT requires that BH is enabled in preemptible
1471          * context.
1472          */
1473         if (statesold & RCUTORTURE_RDR_IRQ)
1474                 local_irq_enable();
1475         if (statesold & RCUTORTURE_RDR_PREEMPT)
1476                 preempt_enable();
1477         if (statesold & RCUTORTURE_RDR_SCHED)
1478                 rcu_read_unlock_sched();
1479         if (statesold & RCUTORTURE_RDR_BH)
1480                 local_bh_enable();
1481         if (statesold & RCUTORTURE_RDR_RBH)
1482                 rcu_read_unlock_bh();
1483         if (statesold & RCUTORTURE_RDR_RCU_2) {
1484                 cur_ops->readunlock((idxold2 >> RCUTORTURE_RDR_SHIFT_2) & 0x1);
1485                 WARN_ON_ONCE(idxnew2 != -1);
1486                 idxold2 = 0;
1487         }
1488         if (statesold & RCUTORTURE_RDR_RCU_1) {
1489                 bool lockit;
1490
1491                 lockit = !cur_ops->no_pi_lock && !statesnew && !(torture_random(trsp) & 0xffff);
1492                 if (lockit)
1493                         raw_spin_lock_irqsave(&current->pi_lock, flags);
1494                 cur_ops->readunlock((idxold1 >> RCUTORTURE_RDR_SHIFT_1) & 0x1);
1495                 WARN_ON_ONCE(idxnew1 != -1);
1496                 idxold1 = 0;
1497                 if (lockit)
1498                         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
1499         }
1500
1501         /* Delay if neither beginning nor end and there was a change. */
1502         if ((statesnew || statesold) && *readstate && newstate)
1503                 cur_ops->read_delay(trsp, rtrsp);
1504
1505         /* Update the reader state. */
1506         if (idxnew1 == -1)
1507                 idxnew1 = idxold1 & RCUTORTURE_RDR_MASK_1;
1508         WARN_ON_ONCE(idxnew1 < 0);
1509         if (WARN_ON_ONCE((idxnew1 >> RCUTORTURE_RDR_SHIFT_1) > 1))
1510                 pr_info("Unexpected idxnew1 value of %#x\n", idxnew1);
1511         if (idxnew2 == -1)
1512                 idxnew2 = idxold2 & RCUTORTURE_RDR_MASK_2;
1513         WARN_ON_ONCE(idxnew2 < 0);
1514         WARN_ON_ONCE((idxnew2 >> RCUTORTURE_RDR_SHIFT_2) > 1);
1515         *readstate = idxnew1 | idxnew2 | newstate;
1516         WARN_ON_ONCE(*readstate < 0);
1517         if (WARN_ON_ONCE((*readstate >> RCUTORTURE_RDR_SHIFT_2) > 1))
1518                 pr_info("Unexpected idxnew2 value of %#x\n", idxnew2);
1519 }
1520
1521 /* Return the biggest extendables mask given current RCU and boot parameters. */
1522 static int rcutorture_extend_mask_max(void)
1523 {
1524         int mask;
1525
1526         WARN_ON_ONCE(extendables & ~RCUTORTURE_MAX_EXTEND);
1527         mask = extendables & RCUTORTURE_MAX_EXTEND & cur_ops->extendables;
1528         mask = mask | RCUTORTURE_RDR_RCU_1 | RCUTORTURE_RDR_RCU_2;
1529         return mask;
1530 }
1531
1532 /* Return a random protection state mask, but with at least one bit set. */
1533 static int
1534 rcutorture_extend_mask(int oldmask, struct torture_random_state *trsp)
1535 {
1536         int mask = rcutorture_extend_mask_max();
1537         unsigned long randmask1 = torture_random(trsp) >> 8;
1538         unsigned long randmask2 = randmask1 >> 3;
1539         unsigned long preempts = RCUTORTURE_RDR_PREEMPT | RCUTORTURE_RDR_SCHED;
1540         unsigned long preempts_irq = preempts | RCUTORTURE_RDR_IRQ;
1541         unsigned long bhs = RCUTORTURE_RDR_BH | RCUTORTURE_RDR_RBH;
1542
1543         WARN_ON_ONCE(mask >> RCUTORTURE_RDR_SHIFT_1);
1544         /* Mostly only one bit (need preemption!), sometimes lots of bits. */
1545         if (!(randmask1 & 0x7))
1546                 mask = mask & randmask2;
1547         else
1548                 mask = mask & (1 << (randmask2 % RCUTORTURE_RDR_NBITS));
1549
1550         // Can't have nested RCU reader without outer RCU reader.
1551         if (!(mask & RCUTORTURE_RDR_RCU_1) && (mask & RCUTORTURE_RDR_RCU_2)) {
1552                 if (oldmask & RCUTORTURE_RDR_RCU_1)
1553                         mask &= ~RCUTORTURE_RDR_RCU_2;
1554                 else
1555                         mask |= RCUTORTURE_RDR_RCU_1;
1556         }
1557
1558         /*
1559          * Can't enable bh w/irq disabled.
1560          */
1561         if (mask & RCUTORTURE_RDR_IRQ)
1562                 mask |= oldmask & bhs;
1563
1564         /*
1565          * Ideally these sequences would be detected in debug builds
1566          * (regardless of RT), but until then don't stop testing
1567          * them on non-RT.
1568          */
1569         if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
1570                 /* Can't modify BH in atomic context */
1571                 if (oldmask & preempts_irq)
1572                         mask &= ~bhs;
1573                 if ((oldmask | mask) & preempts_irq)
1574                         mask |= oldmask & bhs;
1575         }
1576
1577         return mask ?: RCUTORTURE_RDR_RCU_1;
1578 }
1579
1580 /*
1581  * Do a randomly selected number of extensions of an existing RCU read-side
1582  * critical section.
1583  */
1584 static struct rt_read_seg *
1585 rcutorture_loop_extend(int *readstate, struct torture_random_state *trsp,
1586                        struct rt_read_seg *rtrsp)
1587 {
1588         int i;
1589         int j;
1590         int mask = rcutorture_extend_mask_max();
1591
1592         WARN_ON_ONCE(!*readstate); /* -Existing- RCU read-side critsect! */
1593         if (!((mask - 1) & mask))
1594                 return rtrsp;  /* Current RCU reader not extendable. */
1595         /* Bias towards larger numbers of loops. */
1596         i = (torture_random(trsp) >> 3);
1597         i = ((i | (i >> 3)) & RCUTORTURE_RDR_MAX_LOOPS) + 1;
1598         for (j = 0; j < i; j++) {
1599                 mask = rcutorture_extend_mask(*readstate, trsp);
1600                 rcutorture_one_extend(readstate, mask, trsp, &rtrsp[j]);
1601         }
1602         return &rtrsp[j];
1603 }
1604
1605 /*
1606  * Do one read-side critical section, returning false if there was
1607  * no data to read.  Can be invoked both from process context and
1608  * from a timer handler.
1609  */
1610 static bool rcu_torture_one_read(struct torture_random_state *trsp, long myid)
1611 {
1612         unsigned long cookie;
1613         int i;
1614         unsigned long started;
1615         unsigned long completed;
1616         int newstate;
1617         struct rcu_torture *p;
1618         int pipe_count;
1619         int readstate = 0;
1620         struct rt_read_seg rtseg[RCUTORTURE_RDR_MAX_SEGS] = { { 0 } };
1621         struct rt_read_seg *rtrsp = &rtseg[0];
1622         struct rt_read_seg *rtrsp1;
1623         unsigned long long ts;
1624
1625         WARN_ON_ONCE(!rcu_is_watching());
1626         newstate = rcutorture_extend_mask(readstate, trsp);
1627         rcutorture_one_extend(&readstate, newstate, trsp, rtrsp++);
1628         if (cur_ops->get_gp_state && cur_ops->poll_gp_state)
1629                 cookie = cur_ops->get_gp_state();
1630         started = cur_ops->get_gp_seq();
1631         ts = rcu_trace_clock_local();
1632         p = rcu_dereference_check(rcu_torture_current,
1633                                   !cur_ops->readlock_held || cur_ops->readlock_held());
1634         if (p == NULL) {
1635                 /* Wait for rcu_torture_writer to get underway */
1636                 rcutorture_one_extend(&readstate, 0, trsp, rtrsp);
1637                 return false;
1638         }
1639         if (p->rtort_mbtest == 0)
1640                 atomic_inc(&n_rcu_torture_mberror);
1641         rcu_torture_reader_do_mbchk(myid, p, trsp);
1642         rtrsp = rcutorture_loop_extend(&readstate, trsp, rtrsp);
1643         preempt_disable();
1644         pipe_count = READ_ONCE(p->rtort_pipe_count);
1645         if (pipe_count > RCU_TORTURE_PIPE_LEN) {
1646                 /* Should not happen, but... */
1647                 pipe_count = RCU_TORTURE_PIPE_LEN;
1648         }
1649         completed = cur_ops->get_gp_seq();
1650         if (pipe_count > 1) {
1651                 do_trace_rcu_torture_read(cur_ops->name, &p->rtort_rcu,
1652                                           ts, started, completed);
1653                 rcu_ftrace_dump(DUMP_ALL);
1654         }
1655         __this_cpu_inc(rcu_torture_count[pipe_count]);
1656         completed = rcutorture_seq_diff(completed, started);
1657         if (completed > RCU_TORTURE_PIPE_LEN) {
1658                 /* Should not happen, but... */
1659                 completed = RCU_TORTURE_PIPE_LEN;
1660         }
1661         __this_cpu_inc(rcu_torture_batch[completed]);
1662         preempt_enable();
1663         if (cur_ops->get_gp_state && cur_ops->poll_gp_state)
1664                 WARN_ONCE(cur_ops->poll_gp_state(cookie),
1665                           "%s: Cookie check 2 failed %s(%d) %lu->%lu\n",
1666                           __func__,
1667                           rcu_torture_writer_state_getname(),
1668                           rcu_torture_writer_state,
1669                           cookie, cur_ops->get_gp_state());
1670         rcutorture_one_extend(&readstate, 0, trsp, rtrsp);
1671         WARN_ON_ONCE(readstate);
1672         // This next splat is expected behavior if leakpointer, especially
1673         // for CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels.
1674         WARN_ON_ONCE(leakpointer && READ_ONCE(p->rtort_pipe_count) > 1);
1675
1676         /* If error or close call, record the sequence of reader protections. */
1677         if ((pipe_count > 1 || completed > 1) && !xchg(&err_segs_recorded, 1)) {
1678                 i = 0;
1679                 for (rtrsp1 = &rtseg[0]; rtrsp1 < rtrsp; rtrsp1++)
1680                         err_segs[i++] = *rtrsp1;
1681                 rt_read_nsegs = i;
1682         }
1683
1684         return true;
1685 }
1686
1687 static DEFINE_TORTURE_RANDOM_PERCPU(rcu_torture_timer_rand);
1688
1689 /*
1690  * RCU torture reader from timer handler.  Dereferences rcu_torture_current,
1691  * incrementing the corresponding element of the pipeline array.  The
1692  * counter in the element should never be greater than 1, otherwise, the
1693  * RCU implementation is broken.
1694  */
1695 static void rcu_torture_timer(struct timer_list *unused)
1696 {
1697         atomic_long_inc(&n_rcu_torture_timers);
1698         (void)rcu_torture_one_read(this_cpu_ptr(&rcu_torture_timer_rand), -1);
1699
1700         /* Test call_rcu() invocation from interrupt handler. */
1701         if (cur_ops->call) {
1702                 struct rcu_head *rhp = kmalloc(sizeof(*rhp), GFP_NOWAIT);
1703
1704                 if (rhp)
1705                         cur_ops->call(rhp, rcu_torture_timer_cb);
1706         }
1707 }
1708
1709 /*
1710  * RCU torture reader kthread.  Repeatedly dereferences rcu_torture_current,
1711  * incrementing the corresponding element of the pipeline array.  The
1712  * counter in the element should never be greater than 1, otherwise, the
1713  * RCU implementation is broken.
1714  */
1715 static int
1716 rcu_torture_reader(void *arg)
1717 {
1718         unsigned long lastsleep = jiffies;
1719         long myid = (long)arg;
1720         int mynumonline = myid;
1721         DEFINE_TORTURE_RANDOM(rand);
1722         struct timer_list t;
1723
1724         VERBOSE_TOROUT_STRING("rcu_torture_reader task started");
1725         set_user_nice(current, MAX_NICE);
1726         if (irqreader && cur_ops->irq_capable)
1727                 timer_setup_on_stack(&t, rcu_torture_timer, 0);
1728         tick_dep_set_task(current, TICK_DEP_BIT_RCU);
1729         do {
1730                 if (irqreader && cur_ops->irq_capable) {
1731                         if (!timer_pending(&t))
1732                                 mod_timer(&t, jiffies + 1);
1733                 }
1734                 if (!rcu_torture_one_read(&rand, myid) && !torture_must_stop())
1735                         schedule_timeout_interruptible(HZ);
1736                 if (time_after(jiffies, lastsleep) && !torture_must_stop()) {
1737                         torture_hrtimeout_us(500, 1000, &rand);
1738                         lastsleep = jiffies + 10;
1739                 }
1740                 while (torture_num_online_cpus() < mynumonline && !torture_must_stop())
1741                         schedule_timeout_interruptible(HZ / 5);
1742                 stutter_wait("rcu_torture_reader");
1743         } while (!torture_must_stop());
1744         if (irqreader && cur_ops->irq_capable) {
1745                 del_timer_sync(&t);
1746                 destroy_timer_on_stack(&t);
1747         }
1748         tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
1749         torture_kthread_stopping("rcu_torture_reader");
1750         return 0;
1751 }
1752
1753 /*
1754  * Randomly Toggle CPUs' callback-offload state.  This uses hrtimers to
1755  * increase race probabilities and fuzzes the interval between toggling.
1756  */
1757 static int rcu_nocb_toggle(void *arg)
1758 {
1759         int cpu;
1760         int maxcpu = -1;
1761         int oldnice = task_nice(current);
1762         long r;
1763         DEFINE_TORTURE_RANDOM(rand);
1764         ktime_t toggle_delay;
1765         unsigned long toggle_fuzz;
1766         ktime_t toggle_interval = ms_to_ktime(nocbs_toggle);
1767
1768         VERBOSE_TOROUT_STRING("rcu_nocb_toggle task started");
1769         while (!rcu_inkernel_boot_has_ended())
1770                 schedule_timeout_interruptible(HZ / 10);
1771         for_each_online_cpu(cpu)
1772                 maxcpu = cpu;
1773         WARN_ON(maxcpu < 0);
1774         if (toggle_interval > ULONG_MAX)
1775                 toggle_fuzz = ULONG_MAX >> 3;
1776         else
1777                 toggle_fuzz = toggle_interval >> 3;
1778         if (toggle_fuzz <= 0)
1779                 toggle_fuzz = NSEC_PER_USEC;
1780         do {
1781                 r = torture_random(&rand);
1782                 cpu = (r >> 4) % (maxcpu + 1);
1783                 if (r & 0x1) {
1784                         rcu_nocb_cpu_offload(cpu);
1785                         atomic_long_inc(&n_nocb_offload);
1786                 } else {
1787                         rcu_nocb_cpu_deoffload(cpu);
1788                         atomic_long_inc(&n_nocb_deoffload);
1789                 }
1790                 toggle_delay = torture_random(&rand) % toggle_fuzz + toggle_interval;
1791                 set_current_state(TASK_INTERRUPTIBLE);
1792                 schedule_hrtimeout(&toggle_delay, HRTIMER_MODE_REL);
1793                 if (stutter_wait("rcu_nocb_toggle"))
1794                         sched_set_normal(current, oldnice);
1795         } while (!torture_must_stop());
1796         torture_kthread_stopping("rcu_nocb_toggle");
1797         return 0;
1798 }
1799
1800 /*
1801  * Print torture statistics.  Caller must ensure that there is only
1802  * one call to this function at a given time!!!  This is normally
1803  * accomplished by relying on the module system to only have one copy
1804  * of the module loaded, and then by giving the rcu_torture_stats
1805  * kthread full control (or the init/cleanup functions when rcu_torture_stats
1806  * thread is not running).
1807  */
1808 static void
1809 rcu_torture_stats_print(void)
1810 {
1811         int cpu;
1812         int i;
1813         long pipesummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
1814         long batchsummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
1815         struct rcu_torture *rtcp;
1816         static unsigned long rtcv_snap = ULONG_MAX;
1817         static bool splatted;
1818         struct task_struct *wtp;
1819
1820         for_each_possible_cpu(cpu) {
1821                 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
1822                         pipesummary[i] += READ_ONCE(per_cpu(rcu_torture_count, cpu)[i]);
1823                         batchsummary[i] += READ_ONCE(per_cpu(rcu_torture_batch, cpu)[i]);
1824                 }
1825         }
1826         for (i = RCU_TORTURE_PIPE_LEN - 1; i >= 0; i--) {
1827                 if (pipesummary[i] != 0)
1828                         break;
1829         }
1830
1831         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1832         rtcp = rcu_access_pointer(rcu_torture_current);
1833         pr_cont("rtc: %p %s: %lu tfle: %d rta: %d rtaf: %d rtf: %d ",
1834                 rtcp,
1835                 rtcp && !rcu_stall_is_suppressed_at_boot() ? "ver" : "VER",
1836                 rcu_torture_current_version,
1837                 list_empty(&rcu_torture_freelist),
1838                 atomic_read(&n_rcu_torture_alloc),
1839                 atomic_read(&n_rcu_torture_alloc_fail),
1840                 atomic_read(&n_rcu_torture_free));
1841         pr_cont("rtmbe: %d rtmbkf: %d/%d rtbe: %ld rtbke: %ld rtbre: %ld ",
1842                 atomic_read(&n_rcu_torture_mberror),
1843                 atomic_read(&n_rcu_torture_mbchk_fail), atomic_read(&n_rcu_torture_mbchk_tries),
1844                 n_rcu_torture_barrier_error,
1845                 n_rcu_torture_boost_ktrerror,
1846                 n_rcu_torture_boost_rterror);
1847         pr_cont("rtbf: %ld rtb: %ld nt: %ld ",
1848                 n_rcu_torture_boost_failure,
1849                 n_rcu_torture_boosts,
1850                 atomic_long_read(&n_rcu_torture_timers));
1851         torture_onoff_stats();
1852         pr_cont("barrier: %ld/%ld:%ld ",
1853                 data_race(n_barrier_successes),
1854                 data_race(n_barrier_attempts),
1855                 data_race(n_rcu_torture_barrier_error));
1856         pr_cont("read-exits: %ld ", data_race(n_read_exits)); // Statistic.
1857         pr_cont("nocb-toggles: %ld:%ld\n",
1858                 atomic_long_read(&n_nocb_offload), atomic_long_read(&n_nocb_deoffload));
1859
1860         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1861         if (atomic_read(&n_rcu_torture_mberror) ||
1862             atomic_read(&n_rcu_torture_mbchk_fail) ||
1863             n_rcu_torture_barrier_error || n_rcu_torture_boost_ktrerror ||
1864             n_rcu_torture_boost_rterror || n_rcu_torture_boost_failure ||
1865             i > 1) {
1866                 pr_cont("%s", "!!! ");
1867                 atomic_inc(&n_rcu_torture_error);
1868                 WARN_ON_ONCE(atomic_read(&n_rcu_torture_mberror));
1869                 WARN_ON_ONCE(atomic_read(&n_rcu_torture_mbchk_fail));
1870                 WARN_ON_ONCE(n_rcu_torture_barrier_error);  // rcu_barrier()
1871                 WARN_ON_ONCE(n_rcu_torture_boost_ktrerror); // no boost kthread
1872                 WARN_ON_ONCE(n_rcu_torture_boost_rterror); // can't set RT prio
1873                 WARN_ON_ONCE(n_rcu_torture_boost_failure); // boost failed (TIMER_SOFTIRQ RT prio?)
1874                 WARN_ON_ONCE(i > 1); // Too-short grace period
1875         }
1876         pr_cont("Reader Pipe: ");
1877         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
1878                 pr_cont(" %ld", pipesummary[i]);
1879         pr_cont("\n");
1880
1881         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1882         pr_cont("Reader Batch: ");
1883         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
1884                 pr_cont(" %ld", batchsummary[i]);
1885         pr_cont("\n");
1886
1887         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1888         pr_cont("Free-Block Circulation: ");
1889         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
1890                 pr_cont(" %d", atomic_read(&rcu_torture_wcount[i]));
1891         }
1892         pr_cont("\n");
1893
1894         if (cur_ops->stats)
1895                 cur_ops->stats();
1896         if (rtcv_snap == rcu_torture_current_version &&
1897             rcu_access_pointer(rcu_torture_current) &&
1898             !rcu_stall_is_suppressed()) {
1899                 int __maybe_unused flags = 0;
1900                 unsigned long __maybe_unused gp_seq = 0;
1901
1902                 rcutorture_get_gp_data(cur_ops->ttype,
1903                                        &flags, &gp_seq);
1904                 srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp,
1905                                         &flags, &gp_seq);
1906                 wtp = READ_ONCE(writer_task);
1907                 pr_alert("??? Writer stall state %s(%d) g%lu f%#x ->state %#x cpu %d\n",
1908                          rcu_torture_writer_state_getname(),
1909                          rcu_torture_writer_state, gp_seq, flags,
1910                          wtp == NULL ? ~0U : wtp->__state,
1911                          wtp == NULL ? -1 : (int)task_cpu(wtp));
1912                 if (!splatted && wtp) {
1913                         sched_show_task(wtp);
1914                         splatted = true;
1915                 }
1916                 if (cur_ops->gp_kthread_dbg)
1917                         cur_ops->gp_kthread_dbg();
1918                 rcu_ftrace_dump(DUMP_ALL);
1919         }
1920         rtcv_snap = rcu_torture_current_version;
1921 }
1922
1923 /*
1924  * Periodically prints torture statistics, if periodic statistics printing
1925  * was specified via the stat_interval module parameter.
1926  */
1927 static int
1928 rcu_torture_stats(void *arg)
1929 {
1930         VERBOSE_TOROUT_STRING("rcu_torture_stats task started");
1931         do {
1932                 schedule_timeout_interruptible(stat_interval * HZ);
1933                 rcu_torture_stats_print();
1934                 torture_shutdown_absorb("rcu_torture_stats");
1935         } while (!torture_must_stop());
1936         torture_kthread_stopping("rcu_torture_stats");
1937         return 0;
1938 }
1939
1940 /* Test mem_dump_obj() and friends.  */
1941 static void rcu_torture_mem_dump_obj(void)
1942 {
1943         struct rcu_head *rhp;
1944         struct kmem_cache *kcp;
1945         static int z;
1946
1947         kcp = kmem_cache_create("rcuscale", 136, 8, SLAB_STORE_USER, NULL);
1948         rhp = kmem_cache_alloc(kcp, GFP_KERNEL);
1949         pr_alert("mem_dump_obj() slab test: rcu_torture_stats = %px, &rhp = %px, rhp = %px, &z = %px\n", stats_task, &rhp, rhp, &z);
1950         pr_alert("mem_dump_obj(ZERO_SIZE_PTR):");
1951         mem_dump_obj(ZERO_SIZE_PTR);
1952         pr_alert("mem_dump_obj(NULL):");
1953         mem_dump_obj(NULL);
1954         pr_alert("mem_dump_obj(%px):", &rhp);
1955         mem_dump_obj(&rhp);
1956         pr_alert("mem_dump_obj(%px):", rhp);
1957         mem_dump_obj(rhp);
1958         pr_alert("mem_dump_obj(%px):", &rhp->func);
1959         mem_dump_obj(&rhp->func);
1960         pr_alert("mem_dump_obj(%px):", &z);
1961         mem_dump_obj(&z);
1962         kmem_cache_free(kcp, rhp);
1963         kmem_cache_destroy(kcp);
1964         rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
1965         pr_alert("mem_dump_obj() kmalloc test: rcu_torture_stats = %px, &rhp = %px, rhp = %px\n", stats_task, &rhp, rhp);
1966         pr_alert("mem_dump_obj(kmalloc %px):", rhp);
1967         mem_dump_obj(rhp);
1968         pr_alert("mem_dump_obj(kmalloc %px):", &rhp->func);
1969         mem_dump_obj(&rhp->func);
1970         kfree(rhp);
1971         rhp = vmalloc(4096);
1972         pr_alert("mem_dump_obj() vmalloc test: rcu_torture_stats = %px, &rhp = %px, rhp = %px\n", stats_task, &rhp, rhp);
1973         pr_alert("mem_dump_obj(vmalloc %px):", rhp);
1974         mem_dump_obj(rhp);
1975         pr_alert("mem_dump_obj(vmalloc %px):", &rhp->func);
1976         mem_dump_obj(&rhp->func);
1977         vfree(rhp);
1978 }
1979
1980 static void
1981 rcu_torture_print_module_parms(struct rcu_torture_ops *cur_ops, const char *tag)
1982 {
1983         pr_alert("%s" TORTURE_FLAG
1984                  "--- %s: nreaders=%d nfakewriters=%d "
1985                  "stat_interval=%d verbose=%d test_no_idle_hz=%d "
1986                  "shuffle_interval=%d stutter=%d irqreader=%d "
1987                  "fqs_duration=%d fqs_holdoff=%d fqs_stutter=%d "
1988                  "test_boost=%d/%d test_boost_interval=%d "
1989                  "test_boost_duration=%d shutdown_secs=%d "
1990                  "stall_cpu=%d stall_cpu_holdoff=%d stall_cpu_irqsoff=%d "
1991                  "stall_cpu_block=%d "
1992                  "n_barrier_cbs=%d "
1993                  "onoff_interval=%d onoff_holdoff=%d "
1994                  "read_exit_delay=%d read_exit_burst=%d "
1995                  "nocbs_nthreads=%d nocbs_toggle=%d\n",
1996                  torture_type, tag, nrealreaders, nfakewriters,
1997                  stat_interval, verbose, test_no_idle_hz, shuffle_interval,
1998                  stutter, irqreader, fqs_duration, fqs_holdoff, fqs_stutter,
1999                  test_boost, cur_ops->can_boost,
2000                  test_boost_interval, test_boost_duration, shutdown_secs,
2001                  stall_cpu, stall_cpu_holdoff, stall_cpu_irqsoff,
2002                  stall_cpu_block,
2003                  n_barrier_cbs,
2004                  onoff_interval, onoff_holdoff,
2005                  read_exit_delay, read_exit_burst,
2006                  nocbs_nthreads, nocbs_toggle);
2007 }
2008
2009 static int rcutorture_booster_cleanup(unsigned int cpu)
2010 {
2011         struct task_struct *t;
2012
2013         if (boost_tasks[cpu] == NULL)
2014                 return 0;
2015         mutex_lock(&boost_mutex);
2016         t = boost_tasks[cpu];
2017         boost_tasks[cpu] = NULL;
2018         rcu_torture_enable_rt_throttle();
2019         mutex_unlock(&boost_mutex);
2020
2021         /* This must be outside of the mutex, otherwise deadlock! */
2022         torture_stop_kthread(rcu_torture_boost, t);
2023         return 0;
2024 }
2025
2026 static int rcutorture_booster_init(unsigned int cpu)
2027 {
2028         int retval;
2029
2030         if (boost_tasks[cpu] != NULL)
2031                 return 0;  /* Already created, nothing more to do. */
2032
2033         /* Don't allow time recalculation while creating a new task. */
2034         mutex_lock(&boost_mutex);
2035         rcu_torture_disable_rt_throttle();
2036         VERBOSE_TOROUT_STRING("Creating rcu_torture_boost task");
2037         boost_tasks[cpu] = kthread_run_on_cpu(rcu_torture_boost, NULL,
2038                                               cpu, "rcu_torture_boost_%u");
2039         if (IS_ERR(boost_tasks[cpu])) {
2040                 retval = PTR_ERR(boost_tasks[cpu]);
2041                 VERBOSE_TOROUT_STRING("rcu_torture_boost task create failed");
2042                 n_rcu_torture_boost_ktrerror++;
2043                 boost_tasks[cpu] = NULL;
2044                 mutex_unlock(&boost_mutex);
2045                 return retval;
2046         }
2047         mutex_unlock(&boost_mutex);
2048         return 0;
2049 }
2050
2051 /*
2052  * CPU-stall kthread.  It waits as specified by stall_cpu_holdoff, then
2053  * induces a CPU stall for the time specified by stall_cpu.
2054  */
2055 static int rcu_torture_stall(void *args)
2056 {
2057         int idx;
2058         unsigned long stop_at;
2059
2060         VERBOSE_TOROUT_STRING("rcu_torture_stall task started");
2061         if (stall_cpu_holdoff > 0) {
2062                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin holdoff");
2063                 schedule_timeout_interruptible(stall_cpu_holdoff * HZ);
2064                 VERBOSE_TOROUT_STRING("rcu_torture_stall end holdoff");
2065         }
2066         if (!kthread_should_stop() && stall_gp_kthread > 0) {
2067                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin GP stall");
2068                 rcu_gp_set_torture_wait(stall_gp_kthread * HZ);
2069                 for (idx = 0; idx < stall_gp_kthread + 2; idx++) {
2070                         if (kthread_should_stop())
2071                                 break;
2072                         schedule_timeout_uninterruptible(HZ);
2073                 }
2074         }
2075         if (!kthread_should_stop() && stall_cpu > 0) {
2076                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin CPU stall");
2077                 stop_at = ktime_get_seconds() + stall_cpu;
2078                 /* RCU CPU stall is expected behavior in following code. */
2079                 idx = cur_ops->readlock();
2080                 if (stall_cpu_irqsoff)
2081                         local_irq_disable();
2082                 else if (!stall_cpu_block)
2083                         preempt_disable();
2084                 pr_alert("%s start on CPU %d.\n",
2085                           __func__, raw_smp_processor_id());
2086                 while (ULONG_CMP_LT((unsigned long)ktime_get_seconds(),
2087                                     stop_at))
2088                         if (stall_cpu_block) {
2089 #ifdef CONFIG_PREEMPTION
2090                                 preempt_schedule();
2091 #else
2092                                 schedule_timeout_uninterruptible(HZ);
2093 #endif
2094                         } else if (stall_no_softlockup) {
2095                                 touch_softlockup_watchdog();
2096                         }
2097                 if (stall_cpu_irqsoff)
2098                         local_irq_enable();
2099                 else if (!stall_cpu_block)
2100                         preempt_enable();
2101                 cur_ops->readunlock(idx);
2102         }
2103         pr_alert("%s end.\n", __func__);
2104         torture_shutdown_absorb("rcu_torture_stall");
2105         while (!kthread_should_stop())
2106                 schedule_timeout_interruptible(10 * HZ);
2107         return 0;
2108 }
2109
2110 /* Spawn CPU-stall kthread, if stall_cpu specified. */
2111 static int __init rcu_torture_stall_init(void)
2112 {
2113         if (stall_cpu <= 0 && stall_gp_kthread <= 0)
2114                 return 0;
2115         return torture_create_kthread(rcu_torture_stall, NULL, stall_task);
2116 }
2117
2118 /* State structure for forward-progress self-propagating RCU callback. */
2119 struct fwd_cb_state {
2120         struct rcu_head rh;
2121         int stop;
2122 };
2123
2124 /*
2125  * Forward-progress self-propagating RCU callback function.  Because
2126  * callbacks run from softirq, this function is an implicit RCU read-side
2127  * critical section.
2128  */
2129 static void rcu_torture_fwd_prog_cb(struct rcu_head *rhp)
2130 {
2131         struct fwd_cb_state *fcsp = container_of(rhp, struct fwd_cb_state, rh);
2132
2133         if (READ_ONCE(fcsp->stop)) {
2134                 WRITE_ONCE(fcsp->stop, 2);
2135                 return;
2136         }
2137         cur_ops->call(&fcsp->rh, rcu_torture_fwd_prog_cb);
2138 }
2139
2140 /* State for continuous-flood RCU callbacks. */
2141 struct rcu_fwd_cb {
2142         struct rcu_head rh;
2143         struct rcu_fwd_cb *rfc_next;
2144         struct rcu_fwd *rfc_rfp;
2145         int rfc_gps;
2146 };
2147
2148 #define MAX_FWD_CB_JIFFIES      (8 * HZ) /* Maximum CB test duration. */
2149 #define MIN_FWD_CB_LAUNDERS     3       /* This many CB invocations to count. */
2150 #define MIN_FWD_CBS_LAUNDERED   100     /* Number of counted CBs. */
2151 #define FWD_CBS_HIST_DIV        10      /* Histogram buckets/second. */
2152 #define N_LAUNDERS_HIST (2 * MAX_FWD_CB_JIFFIES / (HZ / FWD_CBS_HIST_DIV))
2153
2154 struct rcu_launder_hist {
2155         long n_launders;
2156         unsigned long launder_gp_seq;
2157 };
2158
2159 struct rcu_fwd {
2160         spinlock_t rcu_fwd_lock;
2161         struct rcu_fwd_cb *rcu_fwd_cb_head;
2162         struct rcu_fwd_cb **rcu_fwd_cb_tail;
2163         long n_launders_cb;
2164         unsigned long rcu_fwd_startat;
2165         struct rcu_launder_hist n_launders_hist[N_LAUNDERS_HIST];
2166         unsigned long rcu_launder_gp_seq_start;
2167         int rcu_fwd_id;
2168 };
2169
2170 static DEFINE_MUTEX(rcu_fwd_mutex);
2171 static struct rcu_fwd *rcu_fwds;
2172 static unsigned long rcu_fwd_seq;
2173 static atomic_long_t rcu_fwd_max_cbs;
2174 static bool rcu_fwd_emergency_stop;
2175
2176 static void rcu_torture_fwd_cb_hist(struct rcu_fwd *rfp)
2177 {
2178         unsigned long gps;
2179         unsigned long gps_old;
2180         int i;
2181         int j;
2182
2183         for (i = ARRAY_SIZE(rfp->n_launders_hist) - 1; i > 0; i--)
2184                 if (rfp->n_launders_hist[i].n_launders > 0)
2185                         break;
2186         pr_alert("%s: Callback-invocation histogram %d (duration %lu jiffies):",
2187                  __func__, rfp->rcu_fwd_id, jiffies - rfp->rcu_fwd_startat);
2188         gps_old = rfp->rcu_launder_gp_seq_start;
2189         for (j = 0; j <= i; j++) {
2190                 gps = rfp->n_launders_hist[j].launder_gp_seq;
2191                 pr_cont(" %ds/%d: %ld:%ld",
2192                         j + 1, FWD_CBS_HIST_DIV,
2193                         rfp->n_launders_hist[j].n_launders,
2194                         rcutorture_seq_diff(gps, gps_old));
2195                 gps_old = gps;
2196         }
2197         pr_cont("\n");
2198 }
2199
2200 /* Callback function for continuous-flood RCU callbacks. */
2201 static void rcu_torture_fwd_cb_cr(struct rcu_head *rhp)
2202 {
2203         unsigned long flags;
2204         int i;
2205         struct rcu_fwd_cb *rfcp = container_of(rhp, struct rcu_fwd_cb, rh);
2206         struct rcu_fwd_cb **rfcpp;
2207         struct rcu_fwd *rfp = rfcp->rfc_rfp;
2208
2209         rfcp->rfc_next = NULL;
2210         rfcp->rfc_gps++;
2211         spin_lock_irqsave(&rfp->rcu_fwd_lock, flags);
2212         rfcpp = rfp->rcu_fwd_cb_tail;
2213         rfp->rcu_fwd_cb_tail = &rfcp->rfc_next;
2214         WRITE_ONCE(*rfcpp, rfcp);
2215         WRITE_ONCE(rfp->n_launders_cb, rfp->n_launders_cb + 1);
2216         i = ((jiffies - rfp->rcu_fwd_startat) / (HZ / FWD_CBS_HIST_DIV));
2217         if (i >= ARRAY_SIZE(rfp->n_launders_hist))
2218                 i = ARRAY_SIZE(rfp->n_launders_hist) - 1;
2219         rfp->n_launders_hist[i].n_launders++;
2220         rfp->n_launders_hist[i].launder_gp_seq = cur_ops->get_gp_seq();
2221         spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2222 }
2223
2224 // Give the scheduler a chance, even on nohz_full CPUs.
2225 static void rcu_torture_fwd_prog_cond_resched(unsigned long iter)
2226 {
2227         if (IS_ENABLED(CONFIG_PREEMPTION) && IS_ENABLED(CONFIG_NO_HZ_FULL)) {
2228                 // Real call_rcu() floods hit userspace, so emulate that.
2229                 if (need_resched() || (iter & 0xfff))
2230                         schedule();
2231                 return;
2232         }
2233         // No userspace emulation: CB invocation throttles call_rcu()
2234         cond_resched();
2235 }
2236
2237 /*
2238  * Free all callbacks on the rcu_fwd_cb_head list, either because the
2239  * test is over or because we hit an OOM event.
2240  */
2241 static unsigned long rcu_torture_fwd_prog_cbfree(struct rcu_fwd *rfp)
2242 {
2243         unsigned long flags;
2244         unsigned long freed = 0;
2245         struct rcu_fwd_cb *rfcp;
2246
2247         for (;;) {
2248                 spin_lock_irqsave(&rfp->rcu_fwd_lock, flags);
2249                 rfcp = rfp->rcu_fwd_cb_head;
2250                 if (!rfcp) {
2251                         spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2252                         break;
2253                 }
2254                 rfp->rcu_fwd_cb_head = rfcp->rfc_next;
2255                 if (!rfp->rcu_fwd_cb_head)
2256                         rfp->rcu_fwd_cb_tail = &rfp->rcu_fwd_cb_head;
2257                 spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2258                 kfree(rfcp);
2259                 freed++;
2260                 rcu_torture_fwd_prog_cond_resched(freed);
2261                 if (tick_nohz_full_enabled()) {
2262                         local_irq_save(flags);
2263                         rcu_momentary_dyntick_idle();
2264                         local_irq_restore(flags);
2265                 }
2266         }
2267         return freed;
2268 }
2269
2270 /* Carry out need_resched()/cond_resched() forward-progress testing. */
2271 static void rcu_torture_fwd_prog_nr(struct rcu_fwd *rfp,
2272                                     int *tested, int *tested_tries)
2273 {
2274         unsigned long cver;
2275         unsigned long dur;
2276         struct fwd_cb_state fcs;
2277         unsigned long gps;
2278         int idx;
2279         int sd;
2280         int sd4;
2281         bool selfpropcb = false;
2282         unsigned long stopat;
2283         static DEFINE_TORTURE_RANDOM(trs);
2284
2285         pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2286         if (!cur_ops->sync)
2287                 return; // Cannot do need_resched() forward progress testing without ->sync.
2288         if (cur_ops->call && cur_ops->cb_barrier) {
2289                 init_rcu_head_on_stack(&fcs.rh);
2290                 selfpropcb = true;
2291         }
2292
2293         /* Tight loop containing cond_resched(). */
2294         atomic_inc(&rcu_fwd_cb_nodelay);
2295         cur_ops->sync(); /* Later readers see above write. */
2296         if  (selfpropcb) {
2297                 WRITE_ONCE(fcs.stop, 0);
2298                 cur_ops->call(&fcs.rh, rcu_torture_fwd_prog_cb);
2299         }
2300         cver = READ_ONCE(rcu_torture_current_version);
2301         gps = cur_ops->get_gp_seq();
2302         sd = cur_ops->stall_dur() + 1;
2303         sd4 = (sd + fwd_progress_div - 1) / fwd_progress_div;
2304         dur = sd4 + torture_random(&trs) % (sd - sd4);
2305         WRITE_ONCE(rfp->rcu_fwd_startat, jiffies);
2306         stopat = rfp->rcu_fwd_startat + dur;
2307         while (time_before(jiffies, stopat) &&
2308                !shutdown_time_arrived() &&
2309                !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2310                 idx = cur_ops->readlock();
2311                 udelay(10);
2312                 cur_ops->readunlock(idx);
2313                 if (!fwd_progress_need_resched || need_resched())
2314                         cond_resched();
2315         }
2316         (*tested_tries)++;
2317         if (!time_before(jiffies, stopat) &&
2318             !shutdown_time_arrived() &&
2319             !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2320                 (*tested)++;
2321                 cver = READ_ONCE(rcu_torture_current_version) - cver;
2322                 gps = rcutorture_seq_diff(cur_ops->get_gp_seq(), gps);
2323                 WARN_ON(!cver && gps < 2);
2324                 pr_alert("%s: %d Duration %ld cver %ld gps %ld\n", __func__,
2325                          rfp->rcu_fwd_id, dur, cver, gps);
2326         }
2327         if (selfpropcb) {
2328                 WRITE_ONCE(fcs.stop, 1);
2329                 cur_ops->sync(); /* Wait for running CB to complete. */
2330                 pr_alert("%s: Waiting for CBs: %pS() %d\n", __func__, cur_ops->cb_barrier, rfp->rcu_fwd_id);
2331                 cur_ops->cb_barrier(); /* Wait for queued callbacks. */
2332         }
2333
2334         if (selfpropcb) {
2335                 WARN_ON(READ_ONCE(fcs.stop) != 2);
2336                 destroy_rcu_head_on_stack(&fcs.rh);
2337         }
2338         schedule_timeout_uninterruptible(HZ / 10); /* Let kthreads recover. */
2339         atomic_dec(&rcu_fwd_cb_nodelay);
2340 }
2341
2342 /* Carry out call_rcu() forward-progress testing. */
2343 static void rcu_torture_fwd_prog_cr(struct rcu_fwd *rfp)
2344 {
2345         unsigned long cver;
2346         unsigned long flags;
2347         unsigned long gps;
2348         int i;
2349         long n_launders;
2350         long n_launders_cb_snap;
2351         long n_launders_sa;
2352         long n_max_cbs;
2353         long n_max_gps;
2354         struct rcu_fwd_cb *rfcp;
2355         struct rcu_fwd_cb *rfcpn;
2356         unsigned long stopat;
2357         unsigned long stoppedat;
2358
2359         pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2360         if (READ_ONCE(rcu_fwd_emergency_stop))
2361                 return; /* Get out of the way quickly, no GP wait! */
2362         if (!cur_ops->call)
2363                 return; /* Can't do call_rcu() fwd prog without ->call. */
2364
2365         /* Loop continuously posting RCU callbacks. */
2366         atomic_inc(&rcu_fwd_cb_nodelay);
2367         cur_ops->sync(); /* Later readers see above write. */
2368         WRITE_ONCE(rfp->rcu_fwd_startat, jiffies);
2369         stopat = rfp->rcu_fwd_startat + MAX_FWD_CB_JIFFIES;
2370         n_launders = 0;
2371         rfp->n_launders_cb = 0; // Hoist initialization for multi-kthread
2372         n_launders_sa = 0;
2373         n_max_cbs = 0;
2374         n_max_gps = 0;
2375         for (i = 0; i < ARRAY_SIZE(rfp->n_launders_hist); i++)
2376                 rfp->n_launders_hist[i].n_launders = 0;
2377         cver = READ_ONCE(rcu_torture_current_version);
2378         gps = cur_ops->get_gp_seq();
2379         rfp->rcu_launder_gp_seq_start = gps;
2380         tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2381         while (time_before(jiffies, stopat) &&
2382                !shutdown_time_arrived() &&
2383                !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2384                 rfcp = READ_ONCE(rfp->rcu_fwd_cb_head);
2385                 rfcpn = NULL;
2386                 if (rfcp)
2387                         rfcpn = READ_ONCE(rfcp->rfc_next);
2388                 if (rfcpn) {
2389                         if (rfcp->rfc_gps >= MIN_FWD_CB_LAUNDERS &&
2390                             ++n_max_gps >= MIN_FWD_CBS_LAUNDERED)
2391                                 break;
2392                         rfp->rcu_fwd_cb_head = rfcpn;
2393                         n_launders++;
2394                         n_launders_sa++;
2395                 } else if (!cur_ops->cbflood_max || cur_ops->cbflood_max > n_max_cbs) {
2396                         rfcp = kmalloc(sizeof(*rfcp), GFP_KERNEL);
2397                         if (WARN_ON_ONCE(!rfcp)) {
2398                                 schedule_timeout_interruptible(1);
2399                                 continue;
2400                         }
2401                         n_max_cbs++;
2402                         n_launders_sa = 0;
2403                         rfcp->rfc_gps = 0;
2404                         rfcp->rfc_rfp = rfp;
2405                 } else {
2406                         rfcp = NULL;
2407                 }
2408                 if (rfcp)
2409                         cur_ops->call(&rfcp->rh, rcu_torture_fwd_cb_cr);
2410                 rcu_torture_fwd_prog_cond_resched(n_launders + n_max_cbs);
2411                 if (tick_nohz_full_enabled()) {
2412                         local_irq_save(flags);
2413                         rcu_momentary_dyntick_idle();
2414                         local_irq_restore(flags);
2415                 }
2416         }
2417         stoppedat = jiffies;
2418         n_launders_cb_snap = READ_ONCE(rfp->n_launders_cb);
2419         cver = READ_ONCE(rcu_torture_current_version) - cver;
2420         gps = rcutorture_seq_diff(cur_ops->get_gp_seq(), gps);
2421         pr_alert("%s: Waiting for CBs: %pS() %d\n", __func__, cur_ops->cb_barrier, rfp->rcu_fwd_id);
2422         cur_ops->cb_barrier(); /* Wait for callbacks to be invoked. */
2423         (void)rcu_torture_fwd_prog_cbfree(rfp);
2424
2425         if (!torture_must_stop() && !READ_ONCE(rcu_fwd_emergency_stop) &&
2426             !shutdown_time_arrived()) {
2427                 WARN_ON(n_max_gps < MIN_FWD_CBS_LAUNDERED);
2428                 pr_alert("%s Duration %lu barrier: %lu pending %ld n_launders: %ld n_launders_sa: %ld n_max_gps: %ld n_max_cbs: %ld cver %ld gps %ld\n",
2429                          __func__,
2430                          stoppedat - rfp->rcu_fwd_startat, jiffies - stoppedat,
2431                          n_launders + n_max_cbs - n_launders_cb_snap,
2432                          n_launders, n_launders_sa,
2433                          n_max_gps, n_max_cbs, cver, gps);
2434                 atomic_long_add(n_max_cbs, &rcu_fwd_max_cbs);
2435                 mutex_lock(&rcu_fwd_mutex); // Serialize histograms.
2436                 rcu_torture_fwd_cb_hist(rfp);
2437                 mutex_unlock(&rcu_fwd_mutex);
2438         }
2439         schedule_timeout_uninterruptible(HZ); /* Let CBs drain. */
2440         tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2441         atomic_dec(&rcu_fwd_cb_nodelay);
2442 }
2443
2444
2445 /*
2446  * OOM notifier, but this only prints diagnostic information for the
2447  * current forward-progress test.
2448  */
2449 static int rcutorture_oom_notify(struct notifier_block *self,
2450                                  unsigned long notused, void *nfreed)
2451 {
2452         int i;
2453         long ncbs;
2454         struct rcu_fwd *rfp;
2455
2456         mutex_lock(&rcu_fwd_mutex);
2457         rfp = rcu_fwds;
2458         if (!rfp) {
2459                 mutex_unlock(&rcu_fwd_mutex);
2460                 return NOTIFY_OK;
2461         }
2462         WARN(1, "%s invoked upon OOM during forward-progress testing.\n",
2463              __func__);
2464         for (i = 0; i < fwd_progress; i++) {
2465                 rcu_torture_fwd_cb_hist(&rfp[i]);
2466                 rcu_fwd_progress_check(1 + (jiffies - READ_ONCE(rfp[i].rcu_fwd_startat)) / 2);
2467         }
2468         WRITE_ONCE(rcu_fwd_emergency_stop, true);
2469         smp_mb(); /* Emergency stop before free and wait to avoid hangs. */
2470         ncbs = 0;
2471         for (i = 0; i < fwd_progress; i++)
2472                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2473         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2474         rcu_barrier();
2475         ncbs = 0;
2476         for (i = 0; i < fwd_progress; i++)
2477                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2478         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2479         rcu_barrier();
2480         ncbs = 0;
2481         for (i = 0; i < fwd_progress; i++)
2482                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2483         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2484         smp_mb(); /* Frees before return to avoid redoing OOM. */
2485         (*(unsigned long *)nfreed)++; /* Forward progress CBs freed! */
2486         pr_info("%s returning after OOM processing.\n", __func__);
2487         mutex_unlock(&rcu_fwd_mutex);
2488         return NOTIFY_OK;
2489 }
2490
2491 static struct notifier_block rcutorture_oom_nb = {
2492         .notifier_call = rcutorture_oom_notify
2493 };
2494
2495 /* Carry out grace-period forward-progress testing. */
2496 static int rcu_torture_fwd_prog(void *args)
2497 {
2498         bool firsttime = true;
2499         long max_cbs;
2500         int oldnice = task_nice(current);
2501         unsigned long oldseq = READ_ONCE(rcu_fwd_seq);
2502         struct rcu_fwd *rfp = args;
2503         int tested = 0;
2504         int tested_tries = 0;
2505
2506         VERBOSE_TOROUT_STRING("rcu_torture_fwd_progress task started");
2507         rcu_bind_current_to_nocb();
2508         if (!IS_ENABLED(CONFIG_SMP) || !IS_ENABLED(CONFIG_RCU_BOOST))
2509                 set_user_nice(current, MAX_NICE);
2510         do {
2511                 if (!rfp->rcu_fwd_id) {
2512                         schedule_timeout_interruptible(fwd_progress_holdoff * HZ);
2513                         WRITE_ONCE(rcu_fwd_emergency_stop, false);
2514                         if (!firsttime) {
2515                                 max_cbs = atomic_long_xchg(&rcu_fwd_max_cbs, 0);
2516                                 pr_alert("%s n_max_cbs: %ld\n", __func__, max_cbs);
2517                         }
2518                         firsttime = false;
2519                         WRITE_ONCE(rcu_fwd_seq, rcu_fwd_seq + 1);
2520                 } else {
2521                         while (READ_ONCE(rcu_fwd_seq) == oldseq && !torture_must_stop())
2522                                 schedule_timeout_interruptible(1);
2523                         oldseq = READ_ONCE(rcu_fwd_seq);
2524                 }
2525                 pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2526                 if (rcu_inkernel_boot_has_ended() && torture_num_online_cpus() > rfp->rcu_fwd_id)
2527                         rcu_torture_fwd_prog_cr(rfp);
2528                 if ((cur_ops->stall_dur && cur_ops->stall_dur() > 0) &&
2529                     (!IS_ENABLED(CONFIG_TINY_RCU) ||
2530                      (rcu_inkernel_boot_has_ended() &&
2531                       torture_num_online_cpus() > rfp->rcu_fwd_id)))
2532                         rcu_torture_fwd_prog_nr(rfp, &tested, &tested_tries);
2533
2534                 /* Avoid slow periods, better to test when busy. */
2535                 if (stutter_wait("rcu_torture_fwd_prog"))
2536                         sched_set_normal(current, oldnice);
2537         } while (!torture_must_stop());
2538         /* Short runs might not contain a valid forward-progress attempt. */
2539         if (!rfp->rcu_fwd_id) {
2540                 WARN_ON(!tested && tested_tries >= 5);
2541                 pr_alert("%s: tested %d tested_tries %d\n", __func__, tested, tested_tries);
2542         }
2543         torture_kthread_stopping("rcu_torture_fwd_prog");
2544         return 0;
2545 }
2546
2547 /* If forward-progress checking is requested and feasible, spawn the thread. */
2548 static int __init rcu_torture_fwd_prog_init(void)
2549 {
2550         int i;
2551         int ret = 0;
2552         struct rcu_fwd *rfp;
2553
2554         if (!fwd_progress)
2555                 return 0; /* Not requested, so don't do it. */
2556         if (fwd_progress >= nr_cpu_ids) {
2557                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Limiting fwd_progress to # CPUs.\n");
2558                 fwd_progress = nr_cpu_ids;
2559         } else if (fwd_progress < 0) {
2560                 fwd_progress = nr_cpu_ids;
2561         }
2562         if ((!cur_ops->sync && !cur_ops->call) ||
2563             (!cur_ops->cbflood_max && (!cur_ops->stall_dur || cur_ops->stall_dur() <= 0)) ||
2564             cur_ops == &rcu_busted_ops) {
2565                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Disabled, unsupported by RCU flavor under test");
2566                 fwd_progress = 0;
2567                 return 0;
2568         }
2569         if (stall_cpu > 0) {
2570                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Disabled, conflicts with CPU-stall testing");
2571                 fwd_progress = 0;
2572                 if (IS_MODULE(CONFIG_RCU_TORTURE_TEST))
2573                         return -EINVAL; /* In module, can fail back to user. */
2574                 WARN_ON(1); /* Make sure rcutorture notices conflict. */
2575                 return 0;
2576         }
2577         if (fwd_progress_holdoff <= 0)
2578                 fwd_progress_holdoff = 1;
2579         if (fwd_progress_div <= 0)
2580                 fwd_progress_div = 4;
2581         rfp = kcalloc(fwd_progress, sizeof(*rfp), GFP_KERNEL);
2582         fwd_prog_tasks = kcalloc(fwd_progress, sizeof(*fwd_prog_tasks), GFP_KERNEL);
2583         if (!rfp || !fwd_prog_tasks) {
2584                 kfree(rfp);
2585                 kfree(fwd_prog_tasks);
2586                 fwd_prog_tasks = NULL;
2587                 fwd_progress = 0;
2588                 return -ENOMEM;
2589         }
2590         for (i = 0; i < fwd_progress; i++) {
2591                 spin_lock_init(&rfp[i].rcu_fwd_lock);
2592                 rfp[i].rcu_fwd_cb_tail = &rfp[i].rcu_fwd_cb_head;
2593                 rfp[i].rcu_fwd_id = i;
2594         }
2595         mutex_lock(&rcu_fwd_mutex);
2596         rcu_fwds = rfp;
2597         mutex_unlock(&rcu_fwd_mutex);
2598         register_oom_notifier(&rcutorture_oom_nb);
2599         for (i = 0; i < fwd_progress; i++) {
2600                 ret = torture_create_kthread(rcu_torture_fwd_prog, &rcu_fwds[i], fwd_prog_tasks[i]);
2601                 if (ret) {
2602                         fwd_progress = i;
2603                         return ret;
2604                 }
2605         }
2606         return 0;
2607 }
2608
2609 static void rcu_torture_fwd_prog_cleanup(void)
2610 {
2611         int i;
2612         struct rcu_fwd *rfp;
2613
2614         if (!rcu_fwds || !fwd_prog_tasks)
2615                 return;
2616         for (i = 0; i < fwd_progress; i++)
2617                 torture_stop_kthread(rcu_torture_fwd_prog, fwd_prog_tasks[i]);
2618         unregister_oom_notifier(&rcutorture_oom_nb);
2619         mutex_lock(&rcu_fwd_mutex);
2620         rfp = rcu_fwds;
2621         rcu_fwds = NULL;
2622         mutex_unlock(&rcu_fwd_mutex);
2623         kfree(rfp);
2624         kfree(fwd_prog_tasks);
2625         fwd_prog_tasks = NULL;
2626 }
2627
2628 /* Callback function for RCU barrier testing. */
2629 static void rcu_torture_barrier_cbf(struct rcu_head *rcu)
2630 {
2631         atomic_inc(&barrier_cbs_invoked);
2632 }
2633
2634 /* IPI handler to get callback posted on desired CPU, if online. */
2635 static void rcu_torture_barrier1cb(void *rcu_void)
2636 {
2637         struct rcu_head *rhp = rcu_void;
2638
2639         cur_ops->call(rhp, rcu_torture_barrier_cbf);
2640 }
2641
2642 /* kthread function to register callbacks used to test RCU barriers. */
2643 static int rcu_torture_barrier_cbs(void *arg)
2644 {
2645         long myid = (long)arg;
2646         bool lastphase = false;
2647         bool newphase;
2648         struct rcu_head rcu;
2649
2650         init_rcu_head_on_stack(&rcu);
2651         VERBOSE_TOROUT_STRING("rcu_torture_barrier_cbs task started");
2652         set_user_nice(current, MAX_NICE);
2653         do {
2654                 wait_event(barrier_cbs_wq[myid],
2655                            (newphase =
2656                             smp_load_acquire(&barrier_phase)) != lastphase ||
2657                            torture_must_stop());
2658                 lastphase = newphase;
2659                 if (torture_must_stop())
2660                         break;
2661                 /*
2662                  * The above smp_load_acquire() ensures barrier_phase load
2663                  * is ordered before the following ->call().
2664                  */
2665                 if (smp_call_function_single(myid, rcu_torture_barrier1cb,
2666                                              &rcu, 1)) {
2667                         // IPI failed, so use direct call from current CPU.
2668                         cur_ops->call(&rcu, rcu_torture_barrier_cbf);
2669                 }
2670                 if (atomic_dec_and_test(&barrier_cbs_count))
2671                         wake_up(&barrier_wq);
2672         } while (!torture_must_stop());
2673         if (cur_ops->cb_barrier != NULL)
2674                 cur_ops->cb_barrier();
2675         destroy_rcu_head_on_stack(&rcu);
2676         torture_kthread_stopping("rcu_torture_barrier_cbs");
2677         return 0;
2678 }
2679
2680 /* kthread function to drive and coordinate RCU barrier testing. */
2681 static int rcu_torture_barrier(void *arg)
2682 {
2683         int i;
2684
2685         VERBOSE_TOROUT_STRING("rcu_torture_barrier task starting");
2686         do {
2687                 atomic_set(&barrier_cbs_invoked, 0);
2688                 atomic_set(&barrier_cbs_count, n_barrier_cbs);
2689                 /* Ensure barrier_phase ordered after prior assignments. */
2690                 smp_store_release(&barrier_phase, !barrier_phase);
2691                 for (i = 0; i < n_barrier_cbs; i++)
2692                         wake_up(&barrier_cbs_wq[i]);
2693                 wait_event(barrier_wq,
2694                            atomic_read(&barrier_cbs_count) == 0 ||
2695                            torture_must_stop());
2696                 if (torture_must_stop())
2697                         break;
2698                 n_barrier_attempts++;
2699                 cur_ops->cb_barrier(); /* Implies smp_mb() for wait_event(). */
2700                 if (atomic_read(&barrier_cbs_invoked) != n_barrier_cbs) {
2701                         n_rcu_torture_barrier_error++;
2702                         pr_err("barrier_cbs_invoked = %d, n_barrier_cbs = %d\n",
2703                                atomic_read(&barrier_cbs_invoked),
2704                                n_barrier_cbs);
2705                         WARN_ON(1);
2706                         // Wait manually for the remaining callbacks
2707                         i = 0;
2708                         do {
2709                                 if (WARN_ON(i++ > HZ))
2710                                         i = INT_MIN;
2711                                 schedule_timeout_interruptible(1);
2712                                 cur_ops->cb_barrier();
2713                         } while (atomic_read(&barrier_cbs_invoked) !=
2714                                  n_barrier_cbs &&
2715                                  !torture_must_stop());
2716                         smp_mb(); // Can't trust ordering if broken.
2717                         if (!torture_must_stop())
2718                                 pr_err("Recovered: barrier_cbs_invoked = %d\n",
2719                                        atomic_read(&barrier_cbs_invoked));
2720                 } else {
2721                         n_barrier_successes++;
2722                 }
2723                 schedule_timeout_interruptible(HZ / 10);
2724         } while (!torture_must_stop());
2725         torture_kthread_stopping("rcu_torture_barrier");
2726         return 0;
2727 }
2728
2729 /* Initialize RCU barrier testing. */
2730 static int rcu_torture_barrier_init(void)
2731 {
2732         int i;
2733         int ret;
2734
2735         if (n_barrier_cbs <= 0)
2736                 return 0;
2737         if (cur_ops->call == NULL || cur_ops->cb_barrier == NULL) {
2738                 pr_alert("%s" TORTURE_FLAG
2739                          " Call or barrier ops missing for %s,\n",
2740                          torture_type, cur_ops->name);
2741                 pr_alert("%s" TORTURE_FLAG
2742                          " RCU barrier testing omitted from run.\n",
2743                          torture_type);
2744                 return 0;
2745         }
2746         atomic_set(&barrier_cbs_count, 0);
2747         atomic_set(&barrier_cbs_invoked, 0);
2748         barrier_cbs_tasks =
2749                 kcalloc(n_barrier_cbs, sizeof(barrier_cbs_tasks[0]),
2750                         GFP_KERNEL);
2751         barrier_cbs_wq =
2752                 kcalloc(n_barrier_cbs, sizeof(barrier_cbs_wq[0]), GFP_KERNEL);
2753         if (barrier_cbs_tasks == NULL || !barrier_cbs_wq)
2754                 return -ENOMEM;
2755         for (i = 0; i < n_barrier_cbs; i++) {
2756                 init_waitqueue_head(&barrier_cbs_wq[i]);
2757                 ret = torture_create_kthread(rcu_torture_barrier_cbs,
2758                                              (void *)(long)i,
2759                                              barrier_cbs_tasks[i]);
2760                 if (ret)
2761                         return ret;
2762         }
2763         return torture_create_kthread(rcu_torture_barrier, NULL, barrier_task);
2764 }
2765
2766 /* Clean up after RCU barrier testing. */
2767 static void rcu_torture_barrier_cleanup(void)
2768 {
2769         int i;
2770
2771         torture_stop_kthread(rcu_torture_barrier, barrier_task);
2772         if (barrier_cbs_tasks != NULL) {
2773                 for (i = 0; i < n_barrier_cbs; i++)
2774                         torture_stop_kthread(rcu_torture_barrier_cbs,
2775                                              barrier_cbs_tasks[i]);
2776                 kfree(barrier_cbs_tasks);
2777                 barrier_cbs_tasks = NULL;
2778         }
2779         if (barrier_cbs_wq != NULL) {
2780                 kfree(barrier_cbs_wq);
2781                 barrier_cbs_wq = NULL;
2782         }
2783 }
2784
2785 static bool rcu_torture_can_boost(void)
2786 {
2787         static int boost_warn_once;
2788         int prio;
2789
2790         if (!(test_boost == 1 && cur_ops->can_boost) && test_boost != 2)
2791                 return false;
2792         if (!cur_ops->start_gp_poll || !cur_ops->poll_gp_state)
2793                 return false;
2794
2795         prio = rcu_get_gp_kthreads_prio();
2796         if (!prio)
2797                 return false;
2798
2799         if (prio < 2) {
2800                 if (boost_warn_once == 1)
2801                         return false;
2802
2803                 pr_alert("%s: WARN: RCU kthread priority too low to test boosting.  Skipping RCU boost test. Try passing rcutree.kthread_prio > 1 on the kernel command line.\n", KBUILD_MODNAME);
2804                 boost_warn_once = 1;
2805                 return false;
2806         }
2807
2808         return true;
2809 }
2810
2811 static bool read_exit_child_stop;
2812 static bool read_exit_child_stopped;
2813 static wait_queue_head_t read_exit_wq;
2814
2815 // Child kthread which just does an rcutorture reader and exits.
2816 static int rcu_torture_read_exit_child(void *trsp_in)
2817 {
2818         struct torture_random_state *trsp = trsp_in;
2819
2820         set_user_nice(current, MAX_NICE);
2821         // Minimize time between reading and exiting.
2822         while (!kthread_should_stop())
2823                 schedule_timeout_uninterruptible(1);
2824         (void)rcu_torture_one_read(trsp, -1);
2825         return 0;
2826 }
2827
2828 // Parent kthread which creates and destroys read-exit child kthreads.
2829 static int rcu_torture_read_exit(void *unused)
2830 {
2831         int count = 0;
2832         bool errexit = false;
2833         int i;
2834         struct task_struct *tsp;
2835         DEFINE_TORTURE_RANDOM(trs);
2836
2837         // Allocate and initialize.
2838         set_user_nice(current, MAX_NICE);
2839         VERBOSE_TOROUT_STRING("rcu_torture_read_exit: Start of test");
2840
2841         // Each pass through this loop does one read-exit episode.
2842         do {
2843                 if (++count > read_exit_burst) {
2844                         VERBOSE_TOROUT_STRING("rcu_torture_read_exit: End of episode");
2845                         rcu_barrier(); // Wait for task_struct free, avoid OOM.
2846                         for (i = 0; i < read_exit_delay; i++) {
2847                                 schedule_timeout_uninterruptible(HZ);
2848                                 if (READ_ONCE(read_exit_child_stop))
2849                                         break;
2850                         }
2851                         if (!READ_ONCE(read_exit_child_stop))
2852                                 VERBOSE_TOROUT_STRING("rcu_torture_read_exit: Start of episode");
2853                         count = 0;
2854                 }
2855                 if (READ_ONCE(read_exit_child_stop))
2856                         break;
2857                 // Spawn child.
2858                 tsp = kthread_run(rcu_torture_read_exit_child,
2859                                      &trs, "%s",
2860                                      "rcu_torture_read_exit_child");
2861                 if (IS_ERR(tsp)) {
2862                         TOROUT_ERRSTRING("out of memory");
2863                         errexit = true;
2864                         tsp = NULL;
2865                         break;
2866                 }
2867                 cond_resched();
2868                 kthread_stop(tsp);
2869                 n_read_exits ++;
2870                 stutter_wait("rcu_torture_read_exit");
2871         } while (!errexit && !READ_ONCE(read_exit_child_stop));
2872
2873         // Clean up and exit.
2874         smp_store_release(&read_exit_child_stopped, true); // After reaping.
2875         smp_mb(); // Store before wakeup.
2876         wake_up(&read_exit_wq);
2877         while (!torture_must_stop())
2878                 schedule_timeout_uninterruptible(1);
2879         torture_kthread_stopping("rcu_torture_read_exit");
2880         return 0;
2881 }
2882
2883 static int rcu_torture_read_exit_init(void)
2884 {
2885         if (read_exit_burst <= 0)
2886                 return 0;
2887         init_waitqueue_head(&read_exit_wq);
2888         read_exit_child_stop = false;
2889         read_exit_child_stopped = false;
2890         return torture_create_kthread(rcu_torture_read_exit, NULL,
2891                                       read_exit_task);
2892 }
2893
2894 static void rcu_torture_read_exit_cleanup(void)
2895 {
2896         if (!read_exit_task)
2897                 return;
2898         WRITE_ONCE(read_exit_child_stop, true);
2899         smp_mb(); // Above write before wait.
2900         wait_event(read_exit_wq, smp_load_acquire(&read_exit_child_stopped));
2901         torture_stop_kthread(rcutorture_read_exit, read_exit_task);
2902 }
2903
2904 static enum cpuhp_state rcutor_hp;
2905
2906 static void
2907 rcu_torture_cleanup(void)
2908 {
2909         int firsttime;
2910         int flags = 0;
2911         unsigned long gp_seq = 0;
2912         int i;
2913
2914         if (torture_cleanup_begin()) {
2915                 if (cur_ops->cb_barrier != NULL) {
2916                         pr_info("%s: Invoking %pS().\n", __func__, cur_ops->cb_barrier);
2917                         cur_ops->cb_barrier();
2918                 }
2919                 return;
2920         }
2921         if (!cur_ops) {
2922                 torture_cleanup_end();
2923                 return;
2924         }
2925
2926         if (cur_ops->gp_kthread_dbg)
2927                 cur_ops->gp_kthread_dbg();
2928         rcu_torture_read_exit_cleanup();
2929         rcu_torture_barrier_cleanup();
2930         rcu_torture_fwd_prog_cleanup();
2931         torture_stop_kthread(rcu_torture_stall, stall_task);
2932         torture_stop_kthread(rcu_torture_writer, writer_task);
2933
2934         if (nocb_tasks) {
2935                 for (i = 0; i < nrealnocbers; i++)
2936                         torture_stop_kthread(rcu_nocb_toggle, nocb_tasks[i]);
2937                 kfree(nocb_tasks);
2938                 nocb_tasks = NULL;
2939         }
2940
2941         if (reader_tasks) {
2942                 for (i = 0; i < nrealreaders; i++)
2943                         torture_stop_kthread(rcu_torture_reader,
2944                                              reader_tasks[i]);
2945                 kfree(reader_tasks);
2946                 reader_tasks = NULL;
2947         }
2948         kfree(rcu_torture_reader_mbchk);
2949         rcu_torture_reader_mbchk = NULL;
2950
2951         if (fakewriter_tasks) {
2952                 for (i = 0; i < nfakewriters; i++)
2953                         torture_stop_kthread(rcu_torture_fakewriter,
2954                                              fakewriter_tasks[i]);
2955                 kfree(fakewriter_tasks);
2956                 fakewriter_tasks = NULL;
2957         }
2958
2959         rcutorture_get_gp_data(cur_ops->ttype, &flags, &gp_seq);
2960         srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp, &flags, &gp_seq);
2961         pr_alert("%s:  End-test grace-period state: g%ld f%#x total-gps=%ld\n",
2962                  cur_ops->name, (long)gp_seq, flags,
2963                  rcutorture_seq_diff(gp_seq, start_gp_seq));
2964         torture_stop_kthread(rcu_torture_stats, stats_task);
2965         torture_stop_kthread(rcu_torture_fqs, fqs_task);
2966         if (rcu_torture_can_boost() && rcutor_hp >= 0)
2967                 cpuhp_remove_state(rcutor_hp);
2968
2969         /*
2970          * Wait for all RCU callbacks to fire, then do torture-type-specific
2971          * cleanup operations.
2972          */
2973         if (cur_ops->cb_barrier != NULL) {
2974                 pr_info("%s: Invoking %pS().\n", __func__, cur_ops->cb_barrier);
2975                 cur_ops->cb_barrier();
2976         }
2977         if (cur_ops->cleanup != NULL)
2978                 cur_ops->cleanup();
2979
2980         rcu_torture_mem_dump_obj();
2981
2982         rcu_torture_stats_print();  /* -After- the stats thread is stopped! */
2983
2984         if (err_segs_recorded) {
2985                 pr_alert("Failure/close-call rcutorture reader segments:\n");
2986                 if (rt_read_nsegs == 0)
2987                         pr_alert("\t: No segments recorded!!!\n");
2988                 firsttime = 1;
2989                 for (i = 0; i < rt_read_nsegs; i++) {
2990                         pr_alert("\t%d: %#x ", i, err_segs[i].rt_readstate);
2991                         if (err_segs[i].rt_delay_jiffies != 0) {
2992                                 pr_cont("%s%ldjiffies", firsttime ? "" : "+",
2993                                         err_segs[i].rt_delay_jiffies);
2994                                 firsttime = 0;
2995                         }
2996                         if (err_segs[i].rt_delay_ms != 0) {
2997                                 pr_cont("%s%ldms", firsttime ? "" : "+",
2998                                         err_segs[i].rt_delay_ms);
2999                                 firsttime = 0;
3000                         }
3001                         if (err_segs[i].rt_delay_us != 0) {
3002                                 pr_cont("%s%ldus", firsttime ? "" : "+",
3003                                         err_segs[i].rt_delay_us);
3004                                 firsttime = 0;
3005                         }
3006                         pr_cont("%s\n",
3007                                 err_segs[i].rt_preempted ? "preempted" : "");
3008
3009                 }
3010         }
3011         if (atomic_read(&n_rcu_torture_error) || n_rcu_torture_barrier_error)
3012                 rcu_torture_print_module_parms(cur_ops, "End of test: FAILURE");
3013         else if (torture_onoff_failures())
3014                 rcu_torture_print_module_parms(cur_ops,
3015                                                "End of test: RCU_HOTPLUG");
3016         else
3017                 rcu_torture_print_module_parms(cur_ops, "End of test: SUCCESS");
3018         torture_cleanup_end();
3019 }
3020
3021 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3022 static void rcu_torture_leak_cb(struct rcu_head *rhp)
3023 {
3024 }
3025
3026 static void rcu_torture_err_cb(struct rcu_head *rhp)
3027 {
3028         /*
3029          * This -might- happen due to race conditions, but is unlikely.
3030          * The scenario that leads to this happening is that the
3031          * first of the pair of duplicate callbacks is queued,
3032          * someone else starts a grace period that includes that
3033          * callback, then the second of the pair must wait for the
3034          * next grace period.  Unlikely, but can happen.  If it
3035          * does happen, the debug-objects subsystem won't have splatted.
3036          */
3037         pr_alert("%s: duplicated callback was invoked.\n", KBUILD_MODNAME);
3038 }
3039 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3040
3041 /*
3042  * Verify that double-free causes debug-objects to complain, but only
3043  * if CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.  Otherwise, say that the test
3044  * cannot be carried out.
3045  */
3046 static void rcu_test_debug_objects(void)
3047 {
3048 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3049         struct rcu_head rh1;
3050         struct rcu_head rh2;
3051         struct rcu_head *rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
3052
3053         init_rcu_head_on_stack(&rh1);
3054         init_rcu_head_on_stack(&rh2);
3055         pr_alert("%s: WARN: Duplicate call_rcu() test starting.\n", KBUILD_MODNAME);
3056
3057         /* Try to queue the rh2 pair of callbacks for the same grace period. */
3058         preempt_disable(); /* Prevent preemption from interrupting test. */
3059         rcu_read_lock(); /* Make it impossible to finish a grace period. */
3060         call_rcu(&rh1, rcu_torture_leak_cb); /* Start grace period. */
3061         local_irq_disable(); /* Make it harder to start a new grace period. */
3062         call_rcu(&rh2, rcu_torture_leak_cb);
3063         call_rcu(&rh2, rcu_torture_err_cb); /* Duplicate callback. */
3064         if (rhp) {
3065                 call_rcu(rhp, rcu_torture_leak_cb);
3066                 call_rcu(rhp, rcu_torture_err_cb); /* Another duplicate callback. */
3067         }
3068         local_irq_enable();
3069         rcu_read_unlock();
3070         preempt_enable();
3071
3072         /* Wait for them all to get done so we can safely return. */
3073         rcu_barrier();
3074         pr_alert("%s: WARN: Duplicate call_rcu() test complete.\n", KBUILD_MODNAME);
3075         destroy_rcu_head_on_stack(&rh1);
3076         destroy_rcu_head_on_stack(&rh2);
3077 #else /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3078         pr_alert("%s: !CONFIG_DEBUG_OBJECTS_RCU_HEAD, not testing duplicate call_rcu()\n", KBUILD_MODNAME);
3079 #endif /* #else #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3080 }
3081
3082 static void rcutorture_sync(void)
3083 {
3084         static unsigned long n;
3085
3086         if (cur_ops->sync && !(++n & 0xfff))
3087                 cur_ops->sync();
3088 }
3089
3090 static int __init
3091 rcu_torture_init(void)
3092 {
3093         long i;
3094         int cpu;
3095         int firsterr = 0;
3096         int flags = 0;
3097         unsigned long gp_seq = 0;
3098         static struct rcu_torture_ops *torture_ops[] = {
3099                 &rcu_ops, &rcu_busted_ops, &srcu_ops, &srcud_ops,
3100                 &busted_srcud_ops, &tasks_ops, &tasks_rude_ops,
3101                 &tasks_tracing_ops, &trivial_ops,
3102         };
3103
3104         if (!torture_init_begin(torture_type, verbose))
3105                 return -EBUSY;
3106
3107         /* Process args and tell the world that the torturer is on the job. */
3108         for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
3109                 cur_ops = torture_ops[i];
3110                 if (strcmp(torture_type, cur_ops->name) == 0)
3111                         break;
3112         }
3113         if (i == ARRAY_SIZE(torture_ops)) {
3114                 pr_alert("rcu-torture: invalid torture type: \"%s\"\n",
3115                          torture_type);
3116                 pr_alert("rcu-torture types:");
3117                 for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
3118                         pr_cont(" %s", torture_ops[i]->name);
3119                 pr_cont("\n");
3120                 firsterr = -EINVAL;
3121                 cur_ops = NULL;
3122                 goto unwind;
3123         }
3124         if (cur_ops->fqs == NULL && fqs_duration != 0) {
3125                 pr_alert("rcu-torture: ->fqs NULL and non-zero fqs_duration, fqs disabled.\n");
3126                 fqs_duration = 0;
3127         }
3128         if (cur_ops->init)
3129                 cur_ops->init();
3130
3131         if (nreaders >= 0) {
3132                 nrealreaders = nreaders;
3133         } else {
3134                 nrealreaders = num_online_cpus() - 2 - nreaders;
3135                 if (nrealreaders <= 0)
3136                         nrealreaders = 1;
3137         }
3138         rcu_torture_print_module_parms(cur_ops, "Start of test");
3139         rcutorture_get_gp_data(cur_ops->ttype, &flags, &gp_seq);
3140         srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp, &flags, &gp_seq);
3141         start_gp_seq = gp_seq;
3142         pr_alert("%s:  Start-test grace-period state: g%ld f%#x\n",
3143                  cur_ops->name, (long)gp_seq, flags);
3144
3145         /* Set up the freelist. */
3146
3147         INIT_LIST_HEAD(&rcu_torture_freelist);
3148         for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++) {
3149                 rcu_tortures[i].rtort_mbtest = 0;
3150                 list_add_tail(&rcu_tortures[i].rtort_free,
3151                               &rcu_torture_freelist);
3152         }
3153
3154         /* Initialize the statistics so that each run gets its own numbers. */
3155
3156         rcu_torture_current = NULL;
3157         rcu_torture_current_version = 0;
3158         atomic_set(&n_rcu_torture_alloc, 0);
3159         atomic_set(&n_rcu_torture_alloc_fail, 0);
3160         atomic_set(&n_rcu_torture_free, 0);
3161         atomic_set(&n_rcu_torture_mberror, 0);
3162         atomic_set(&n_rcu_torture_mbchk_fail, 0);
3163         atomic_set(&n_rcu_torture_mbchk_tries, 0);
3164         atomic_set(&n_rcu_torture_error, 0);
3165         n_rcu_torture_barrier_error = 0;
3166         n_rcu_torture_boost_ktrerror = 0;
3167         n_rcu_torture_boost_rterror = 0;
3168         n_rcu_torture_boost_failure = 0;
3169         n_rcu_torture_boosts = 0;
3170         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
3171                 atomic_set(&rcu_torture_wcount[i], 0);
3172         for_each_possible_cpu(cpu) {
3173                 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
3174                         per_cpu(rcu_torture_count, cpu)[i] = 0;
3175                         per_cpu(rcu_torture_batch, cpu)[i] = 0;
3176                 }
3177         }
3178         err_segs_recorded = 0;
3179         rt_read_nsegs = 0;
3180
3181         /* Start up the kthreads. */
3182
3183         rcu_torture_write_types();
3184         firsterr = torture_create_kthread(rcu_torture_writer, NULL,
3185                                           writer_task);
3186         if (torture_init_error(firsterr))
3187                 goto unwind;
3188         if (nfakewriters > 0) {
3189                 fakewriter_tasks = kcalloc(nfakewriters,
3190                                            sizeof(fakewriter_tasks[0]),
3191                                            GFP_KERNEL);
3192                 if (fakewriter_tasks == NULL) {
3193                         TOROUT_ERRSTRING("out of memory");
3194                         firsterr = -ENOMEM;
3195                         goto unwind;
3196                 }
3197         }
3198         for (i = 0; i < nfakewriters; i++) {
3199                 firsterr = torture_create_kthread(rcu_torture_fakewriter,
3200                                                   NULL, fakewriter_tasks[i]);
3201                 if (torture_init_error(firsterr))
3202                         goto unwind;
3203         }
3204         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
3205                                GFP_KERNEL);
3206         rcu_torture_reader_mbchk = kcalloc(nrealreaders, sizeof(*rcu_torture_reader_mbchk),
3207                                            GFP_KERNEL);
3208         if (!reader_tasks || !rcu_torture_reader_mbchk) {
3209                 TOROUT_ERRSTRING("out of memory");
3210                 firsterr = -ENOMEM;
3211                 goto unwind;
3212         }
3213         for (i = 0; i < nrealreaders; i++) {
3214                 rcu_torture_reader_mbchk[i].rtc_chkrdr = -1;
3215                 firsterr = torture_create_kthread(rcu_torture_reader, (void *)i,
3216                                                   reader_tasks[i]);
3217                 if (torture_init_error(firsterr))
3218                         goto unwind;
3219         }
3220         nrealnocbers = nocbs_nthreads;
3221         if (WARN_ON(nrealnocbers < 0))
3222                 nrealnocbers = 1;
3223         if (WARN_ON(nocbs_toggle < 0))
3224                 nocbs_toggle = HZ;
3225         if (nrealnocbers > 0) {
3226                 nocb_tasks = kcalloc(nrealnocbers, sizeof(nocb_tasks[0]), GFP_KERNEL);
3227                 if (nocb_tasks == NULL) {
3228                         TOROUT_ERRSTRING("out of memory");
3229                         firsterr = -ENOMEM;
3230                         goto unwind;
3231                 }
3232         } else {
3233                 nocb_tasks = NULL;
3234         }
3235         for (i = 0; i < nrealnocbers; i++) {
3236                 firsterr = torture_create_kthread(rcu_nocb_toggle, NULL, nocb_tasks[i]);
3237                 if (torture_init_error(firsterr))
3238                         goto unwind;
3239         }
3240         if (stat_interval > 0) {
3241                 firsterr = torture_create_kthread(rcu_torture_stats, NULL,
3242                                                   stats_task);
3243                 if (torture_init_error(firsterr))
3244                         goto unwind;
3245         }
3246         if (test_no_idle_hz && shuffle_interval > 0) {
3247                 firsterr = torture_shuffle_init(shuffle_interval * HZ);
3248                 if (torture_init_error(firsterr))
3249                         goto unwind;
3250         }
3251         if (stutter < 0)
3252                 stutter = 0;
3253         if (stutter) {
3254                 int t;
3255
3256                 t = cur_ops->stall_dur ? cur_ops->stall_dur() : stutter * HZ;
3257                 firsterr = torture_stutter_init(stutter * HZ, t);
3258                 if (torture_init_error(firsterr))
3259                         goto unwind;
3260         }
3261         if (fqs_duration < 0)
3262                 fqs_duration = 0;
3263         if (fqs_duration) {
3264                 /* Create the fqs thread */
3265                 firsterr = torture_create_kthread(rcu_torture_fqs, NULL,
3266                                                   fqs_task);
3267                 if (torture_init_error(firsterr))
3268                         goto unwind;
3269         }
3270         if (test_boost_interval < 1)
3271                 test_boost_interval = 1;
3272         if (test_boost_duration < 2)
3273                 test_boost_duration = 2;
3274         if (rcu_torture_can_boost()) {
3275
3276                 boost_starttime = jiffies + test_boost_interval * HZ;
3277
3278                 firsterr = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "RCU_TORTURE",
3279                                              rcutorture_booster_init,
3280                                              rcutorture_booster_cleanup);
3281                 rcutor_hp = firsterr;
3282                 if (torture_init_error(firsterr))
3283                         goto unwind;
3284
3285                 // Testing RCU priority boosting requires rcutorture do
3286                 // some serious abuse.  Counter this by running ksoftirqd
3287                 // at higher priority.
3288                 if (IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)) {
3289                         for_each_online_cpu(cpu) {
3290                                 struct sched_param sp;
3291                                 struct task_struct *t;
3292
3293                                 t = per_cpu(ksoftirqd, cpu);
3294                                 WARN_ON_ONCE(!t);
3295                                 sp.sched_priority = 2;
3296                                 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3297                         }
3298                 }
3299         }
3300         shutdown_jiffies = jiffies + shutdown_secs * HZ;
3301         firsterr = torture_shutdown_init(shutdown_secs, rcu_torture_cleanup);
3302         if (torture_init_error(firsterr))
3303                 goto unwind;
3304         firsterr = torture_onoff_init(onoff_holdoff * HZ, onoff_interval,
3305                                       rcutorture_sync);
3306         if (torture_init_error(firsterr))
3307                 goto unwind;
3308         firsterr = rcu_torture_stall_init();
3309         if (torture_init_error(firsterr))
3310                 goto unwind;
3311         firsterr = rcu_torture_fwd_prog_init();
3312         if (torture_init_error(firsterr))
3313                 goto unwind;
3314         firsterr = rcu_torture_barrier_init();
3315         if (torture_init_error(firsterr))
3316                 goto unwind;
3317         firsterr = rcu_torture_read_exit_init();
3318         if (torture_init_error(firsterr))
3319                 goto unwind;
3320         if (object_debug)
3321                 rcu_test_debug_objects();
3322         torture_init_end();
3323         return 0;
3324
3325 unwind:
3326         torture_init_end();
3327         rcu_torture_cleanup();
3328         if (shutdown_secs) {
3329                 WARN_ON(!IS_MODULE(CONFIG_RCU_TORTURE_TEST));
3330                 kernel_power_off();
3331         }
3332         return firsterr;
3333 }
3334
3335 module_init(rcu_torture_init);
3336 module_exit(rcu_torture_cleanup);