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 trivial CONFIG_PREEMPT=n-only torture testing.
 742  * This implementation does not necessarily work well with CPU hotplug.
 743  */
 744
 745 static void synchronize_rcu_trivial(void)
 746 {
 747         int cpu;
 748
 749         for_each_online_cpu(cpu) {
 750                 rcutorture_sched_setaffinity(current->pid, cpumask_of(cpu));
 751                 WARN_ON_ONCE(raw_smp_processor_id() != cpu);
 752         }
 753 }
 754
 755 static int rcu_torture_read_lock_trivial(void) __acquires(RCU)
 756 {
 757         preempt_disable();
 758         return 0;
 759 }
 760
 761 static void rcu_torture_read_unlock_trivial(int idx) __releases(RCU)
 762 {
 763         preempt_enable();
 764 }
 765
 766 static struct rcu_torture_ops trivial_ops = {
 767         .ttype          = RCU_TRIVIAL_FLAVOR,
 768         .init           = rcu_sync_torture_init,
 769         .readlock       = rcu_torture_read_lock_trivial,
 770         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 771         .readunlock     = rcu_torture_read_unlock_trivial,
 772         .readlock_held  = torture_readlock_not_held,
 773         .get_gp_seq     = rcu_no_completed,
 774         .sync           = synchronize_rcu_trivial,
 775         .exp_sync       = synchronize_rcu_trivial,
 776         .fqs            = NULL,
 777         .stats          = NULL,
 778         .irq_capable    = 1,
 779         .name           = "trivial"
 780 };
 781
 782 #ifdef CONFIG_TASKS_RCU
 783
 784 /*
 785  * Definitions for RCU-tasks torture testing.
 786  */
 787
 788 static int tasks_torture_read_lock(void)
 789 {
 790         return 0;
 791 }
 792
 793 static void tasks_torture_read_unlock(int idx)
 794 {
 795 }
 796
 797 static void rcu_tasks_torture_deferred_free(struct rcu_torture *p)
 798 {
 799         call_rcu_tasks(&p->rtort_rcu, rcu_torture_cb);
 800 }
 801
 802 static void synchronize_rcu_mult_test(void)
 803 {
 804         synchronize_rcu_mult(call_rcu_tasks, call_rcu);
 805 }
 806
 807 static struct rcu_torture_ops tasks_ops = {
 808         .ttype          = RCU_TASKS_FLAVOR,
 809         .init           = rcu_sync_torture_init,
 810         .readlock       = tasks_torture_read_lock,
 811         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 812         .readunlock     = tasks_torture_read_unlock,
 813         .get_gp_seq     = rcu_no_completed,
 814         .deferred_free  = rcu_tasks_torture_deferred_free,
 815         .sync           = synchronize_rcu_tasks,
 816         .exp_sync       = synchronize_rcu_mult_test,
 817         .call           = call_rcu_tasks,
 818         .cb_barrier     = rcu_barrier_tasks,
 819         .gp_kthread_dbg = show_rcu_tasks_classic_gp_kthread,
 820         .fqs            = NULL,
 821         .stats          = NULL,
 822         .irq_capable    = 1,
 823         .slow_gps       = 1,
 824         .name           = "tasks"
 825 };
 826
 827 #define TASKS_OPS &tasks_ops,
 828
 829 #else // #ifdef CONFIG_TASKS_RCU
 830
 831 #define TASKS_OPS
 832
 833 #endif // #else #ifdef CONFIG_TASKS_RCU
 834
 835
 836 #ifdef CONFIG_TASKS_RUDE_RCU
 837
 838 /*
 839  * Definitions for rude RCU-tasks torture testing.
 840  */
 841
 842 static void rcu_tasks_rude_torture_deferred_free(struct rcu_torture *p)
 843 {
 844         call_rcu_tasks_rude(&p->rtort_rcu, rcu_torture_cb);
 845 }
 846
 847 static struct rcu_torture_ops tasks_rude_ops = {
 848         .ttype          = RCU_TASKS_RUDE_FLAVOR,
 849         .init           = rcu_sync_torture_init,
 850         .readlock       = rcu_torture_read_lock_trivial,
 851         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 852         .readunlock     = rcu_torture_read_unlock_trivial,
 853         .get_gp_seq     = rcu_no_completed,
 854         .deferred_free  = rcu_tasks_rude_torture_deferred_free,
 855         .sync           = synchronize_rcu_tasks_rude,
 856         .exp_sync       = synchronize_rcu_tasks_rude,
 857         .call           = call_rcu_tasks_rude,
 858         .cb_barrier     = rcu_barrier_tasks_rude,
 859         .gp_kthread_dbg = show_rcu_tasks_rude_gp_kthread,
 860         .cbflood_max    = 50000,
 861         .fqs            = NULL,
 862         .stats          = NULL,
 863         .irq_capable    = 1,
 864         .name           = "tasks-rude"
 865 };
 866
 867 #define TASKS_RUDE_OPS &tasks_rude_ops,
 868
 869 #else // #ifdef CONFIG_TASKS_RUDE_RCU
 870
 871 #define TASKS_RUDE_OPS
 872
 873 #endif // #else #ifdef CONFIG_TASKS_RUDE_RCU
 874
 875
 876 #ifdef CONFIG_TASKS_TRACE_RCU
 877
 878 /*
 879  * Definitions for tracing RCU-tasks torture testing.
 880  */
 881
 882 static int tasks_tracing_torture_read_lock(void)
 883 {
 884         rcu_read_lock_trace();
 885         return 0;
 886 }
 887
 888 static void tasks_tracing_torture_read_unlock(int idx)
 889 {
 890         rcu_read_unlock_trace();
 891 }
 892
 893 static void rcu_tasks_tracing_torture_deferred_free(struct rcu_torture *p)
 894 {
 895         call_rcu_tasks_trace(&p->rtort_rcu, rcu_torture_cb);
 896 }
 897
 898 static struct rcu_torture_ops tasks_tracing_ops = {
 899         .ttype          = RCU_TASKS_TRACING_FLAVOR,
 900         .init           = rcu_sync_torture_init,
 901         .readlock       = tasks_tracing_torture_read_lock,
 902         .read_delay     = srcu_read_delay,  /* just reuse srcu's version. */
 903         .readunlock     = tasks_tracing_torture_read_unlock,
 904         .readlock_held  = rcu_read_lock_trace_held,
 905         .get_gp_seq     = rcu_no_completed,
 906         .deferred_free  = rcu_tasks_tracing_torture_deferred_free,
 907         .sync           = synchronize_rcu_tasks_trace,
 908         .exp_sync       = synchronize_rcu_tasks_trace,
 909         .call           = call_rcu_tasks_trace,
 910         .cb_barrier     = rcu_barrier_tasks_trace,
 911         .gp_kthread_dbg = show_rcu_tasks_trace_gp_kthread,
 912         .cbflood_max    = 50000,
 913         .fqs            = NULL,
 914         .stats          = NULL,
 915         .irq_capable    = 1,
 916         .slow_gps       = 1,
 917         .name           = "tasks-tracing"
 918 };
 919
 920 #define TASKS_TRACING_OPS &tasks_tracing_ops,
 921
 922 #else // #ifdef CONFIG_TASKS_TRACE_RCU
 923
 924 #define TASKS_TRACING_OPS
 925
 926 #endif // #else #ifdef CONFIG_TASKS_TRACE_RCU
 927
 928
 929 static unsigned long rcutorture_seq_diff(unsigned long new, unsigned long old)
 930 {
 931         if (!cur_ops->gp_diff)
 932                 return new - old;
 933         return cur_ops->gp_diff(new, old);
 934 }
 935
 936 /*
 937  * RCU torture priority-boost testing.  Runs one real-time thread per
 938  * CPU for moderate bursts, repeatedly starting grace periods and waiting
 939  * for them to complete.  If a given grace period takes too long, we assume
 940  * that priority inversion has occurred.
 941  */
 942
 943 static int old_rt_runtime = -1;
 944
 945 static void rcu_torture_disable_rt_throttle(void)
 946 {
 947         /*
 948          * Disable RT throttling so that rcutorture's boost threads don't get
 949          * throttled. Only possible if rcutorture is built-in otherwise the
 950          * user should manually do this by setting the sched_rt_period_us and
 951          * sched_rt_runtime sysctls.
 952          */
 953         if (!IS_BUILTIN(CONFIG_RCU_TORTURE_TEST) || old_rt_runtime != -1)
 954                 return;
 955
 956         old_rt_runtime = sysctl_sched_rt_runtime;
 957         sysctl_sched_rt_runtime = -1;
 958 }
 959
 960 static void rcu_torture_enable_rt_throttle(void)
 961 {
 962         if (!IS_BUILTIN(CONFIG_RCU_TORTURE_TEST) || old_rt_runtime == -1)
 963                 return;
 964
 965         sysctl_sched_rt_runtime = old_rt_runtime;
 966         old_rt_runtime = -1;
 967 }
 968
 969 static bool rcu_torture_boost_failed(unsigned long gp_state, unsigned long *start)
 970 {
 971         int cpu;
 972         static int dbg_done;
 973         unsigned long end = jiffies;
 974         bool gp_done;
 975         unsigned long j;
 976         static unsigned long last_persist;
 977         unsigned long lp;
 978         unsigned long mininterval = test_boost_duration * HZ - HZ / 2;
 979
 980         if (end - *start > mininterval) {
 981                 // Recheck after checking time to avoid false positives.
 982                 smp_mb(); // Time check before grace-period check.
 983                 if (cur_ops->poll_gp_state(gp_state))
 984                         return false; // passed, though perhaps just barely
 985                 if (cur_ops->check_boost_failed && !cur_ops->check_boost_failed(gp_state, &cpu)) {
 986                         // At most one persisted message per boost test.
 987                         j = jiffies;
 988                         lp = READ_ONCE(last_persist);
 989                         if (time_after(j, lp + mininterval) && cmpxchg(&last_persist, lp, j) == lp)
 990                                 pr_info("Boost inversion persisted: No QS from CPU %d\n", cpu);
 991                         return false; // passed on a technicality
 992                 }
 993                 VERBOSE_TOROUT_STRING("rcu_torture_boost boosting failed");
 994                 n_rcu_torture_boost_failure++;
 995                 if (!xchg(&dbg_done, 1) && cur_ops->gp_kthread_dbg) {
 996                         pr_info("Boost inversion thread ->rt_priority %u gp_state %lu jiffies %lu\n",
 997                                 current->rt_priority, gp_state, end - *start);
 998                         cur_ops->gp_kthread_dbg();
 999                         // Recheck after print to flag grace period ending during splat.
1000                         gp_done = cur_ops->poll_gp_state(gp_state);
1001                         pr_info("Boost inversion: GP %lu %s.\n", gp_state,
1002                                 gp_done ? "ended already" : "still pending");
1003
1004                 }
1005
1006                 return true; // failed
1007         } else if (cur_ops->check_boost_failed && !cur_ops->check_boost_failed(gp_state, NULL)) {
1008                 *start = jiffies;
1009         }
1010
1011         return false; // passed
1012 }
1013
1014 static int rcu_torture_boost(void *arg)
1015 {
1016         unsigned long endtime;
1017         unsigned long gp_state;
1018         unsigned long gp_state_time;
1019         unsigned long oldstarttime;
1020
1021         VERBOSE_TOROUT_STRING("rcu_torture_boost started");
1022
1023         /* Set real-time priority. */
1024         sched_set_fifo_low(current);
1025
1026         /* Each pass through the following loop does one boost-test cycle. */
1027         do {
1028                 bool failed = false; // Test failed already in this test interval
1029                 bool gp_initiated = false;
1030
1031                 if (kthread_should_stop())
1032                         goto checkwait;
1033
1034                 /* Wait for the next test interval. */
1035                 oldstarttime = READ_ONCE(boost_starttime);
1036                 while (time_before(jiffies, oldstarttime)) {
1037                         schedule_timeout_interruptible(oldstarttime - jiffies);
1038                         if (stutter_wait("rcu_torture_boost"))
1039                                 sched_set_fifo_low(current);
1040                         if (torture_must_stop())
1041                                 goto checkwait;
1042                 }
1043
1044                 // Do one boost-test interval.
1045                 endtime = oldstarttime + test_boost_duration * HZ;
1046                 while (time_before(jiffies, endtime)) {
1047                         // Has current GP gone too long?
1048                         if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
1049                                 failed = rcu_torture_boost_failed(gp_state, &gp_state_time);
1050                         // If we don't have a grace period in flight, start one.
1051                         if (!gp_initiated || cur_ops->poll_gp_state(gp_state)) {
1052                                 gp_state = cur_ops->start_gp_poll();
1053                                 gp_initiated = true;
1054                                 gp_state_time = jiffies;
1055                         }
1056                         if (stutter_wait("rcu_torture_boost")) {
1057                                 sched_set_fifo_low(current);
1058                                 // If the grace period already ended,
1059                                 // we don't know when that happened, so
1060                                 // start over.
1061                                 if (cur_ops->poll_gp_state(gp_state))
1062                                         gp_initiated = false;
1063                         }
1064                         if (torture_must_stop())
1065                                 goto checkwait;
1066                 }
1067
1068                 // In case the grace period extended beyond the end of the loop.
1069                 if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
1070                         rcu_torture_boost_failed(gp_state, &gp_state_time);
1071
1072                 /*
1073                  * Set the start time of the next test interval.
1074                  * Yes, this is vulnerable to long delays, but such
1075                  * delays simply cause a false negative for the next
1076                  * interval.  Besides, we are running at RT priority,
1077                  * so delays should be relatively rare.
1078                  */
1079                 while (oldstarttime == READ_ONCE(boost_starttime) && !kthread_should_stop()) {
1080                         if (mutex_trylock(&boost_mutex)) {
1081                                 if (oldstarttime == boost_starttime) {
1082                                         WRITE_ONCE(boost_starttime,
1083                                                    jiffies + test_boost_interval * HZ);
1084                                         n_rcu_torture_boosts++;
1085                                 }
1086                                 mutex_unlock(&boost_mutex);
1087                                 break;
1088                         }
1089                         schedule_timeout_uninterruptible(1);
1090                 }
1091
1092                 /* Go do the stutter. */
1093 checkwait:      if (stutter_wait("rcu_torture_boost"))
1094                         sched_set_fifo_low(current);
1095         } while (!torture_must_stop());
1096
1097         /* Clean up and exit. */
1098         while (!kthread_should_stop()) {
1099                 torture_shutdown_absorb("rcu_torture_boost");
1100                 schedule_timeout_uninterruptible(1);
1101         }
1102         torture_kthread_stopping("rcu_torture_boost");
1103         return 0;
1104 }
1105
1106 /*
1107  * RCU torture force-quiescent-state kthread.  Repeatedly induces
1108  * bursts of calls to force_quiescent_state(), increasing the probability
1109  * of occurrence of some important types of race conditions.
1110  */
1111 static int
1112 rcu_torture_fqs(void *arg)
1113 {
1114         unsigned long fqs_resume_time;
1115         int fqs_burst_remaining;
1116         int oldnice = task_nice(current);
1117
1118         VERBOSE_TOROUT_STRING("rcu_torture_fqs task started");
1119         do {
1120                 fqs_resume_time = jiffies + fqs_stutter * HZ;
1121                 while (time_before(jiffies, fqs_resume_time) &&
1122                        !kthread_should_stop()) {
1123                         schedule_timeout_interruptible(1);
1124                 }
1125                 fqs_burst_remaining = fqs_duration;
1126                 while (fqs_burst_remaining > 0 &&
1127                        !kthread_should_stop()) {
1128                         cur_ops->fqs();
1129                         udelay(fqs_holdoff);
1130                         fqs_burst_remaining -= fqs_holdoff;
1131                 }
1132                 if (stutter_wait("rcu_torture_fqs"))
1133                         sched_set_normal(current, oldnice);
1134         } while (!torture_must_stop());
1135         torture_kthread_stopping("rcu_torture_fqs");
1136         return 0;
1137 }
1138
1139 // Used by writers to randomly choose from the available grace-period
1140 // primitives.  The only purpose of the initialization is to size the array.
1141 static int synctype[] = { RTWS_DEF_FREE, RTWS_EXP_SYNC, RTWS_COND_GET, RTWS_POLL_GET, RTWS_SYNC };
1142 static int nsynctypes;
1143
1144 /*
1145  * Determine which grace-period primitives are available.
1146  */
1147 static void rcu_torture_write_types(void)
1148 {
1149         bool gp_cond1 = gp_cond, gp_exp1 = gp_exp, gp_normal1 = gp_normal;
1150         bool gp_poll1 = gp_poll, gp_sync1 = gp_sync;
1151
1152         /* Initialize synctype[] array.  If none set, take default. */
1153         if (!gp_cond1 && !gp_exp1 && !gp_normal1 && !gp_poll1 && !gp_sync1)
1154                 gp_cond1 = gp_exp1 = gp_normal1 = gp_poll1 = gp_sync1 = true;
1155         if (gp_cond1 && cur_ops->get_gp_state && cur_ops->cond_sync) {
1156                 synctype[nsynctypes++] = RTWS_COND_GET;
1157                 pr_info("%s: Testing conditional GPs.\n", __func__);
1158         } else if (gp_cond && (!cur_ops->get_gp_state || !cur_ops->cond_sync)) {
1159                 pr_alert("%s: gp_cond without primitives.\n", __func__);
1160         }
1161         if (gp_exp1 && cur_ops->exp_sync) {
1162                 synctype[nsynctypes++] = RTWS_EXP_SYNC;
1163                 pr_info("%s: Testing expedited GPs.\n", __func__);
1164         } else if (gp_exp && !cur_ops->exp_sync) {
1165                 pr_alert("%s: gp_exp without primitives.\n", __func__);
1166         }
1167         if (gp_normal1 && cur_ops->deferred_free) {
1168                 synctype[nsynctypes++] = RTWS_DEF_FREE;
1169                 pr_info("%s: Testing asynchronous GPs.\n", __func__);
1170         } else if (gp_normal && !cur_ops->deferred_free) {
1171                 pr_alert("%s: gp_normal without primitives.\n", __func__);
1172         }
1173         if (gp_poll1 && cur_ops->start_gp_poll && cur_ops->poll_gp_state) {
1174                 synctype[nsynctypes++] = RTWS_POLL_GET;
1175                 pr_info("%s: Testing polling GPs.\n", __func__);
1176         } else if (gp_poll && (!cur_ops->start_gp_poll || !cur_ops->poll_gp_state)) {
1177                 pr_alert("%s: gp_poll without primitives.\n", __func__);
1178         }
1179         if (gp_sync1 && cur_ops->sync) {
1180                 synctype[nsynctypes++] = RTWS_SYNC;
1181                 pr_info("%s: Testing normal GPs.\n", __func__);
1182         } else if (gp_sync && !cur_ops->sync) {
1183                 pr_alert("%s: gp_sync without primitives.\n", __func__);
1184         }
1185 }
1186
1187 /*
1188  * RCU torture writer kthread.  Repeatedly substitutes a new structure
1189  * for that pointed to by rcu_torture_current, freeing the old structure
1190  * after a series of grace periods (the "pipeline").
1191  */
1192 static int
1193 rcu_torture_writer(void *arg)
1194 {
1195         bool boot_ended;
1196         bool can_expedite = !rcu_gp_is_expedited() && !rcu_gp_is_normal();
1197         unsigned long cookie;
1198         int expediting = 0;
1199         unsigned long gp_snap;
1200         int i;
1201         int idx;
1202         int oldnice = task_nice(current);
1203         struct rcu_torture *rp;
1204         struct rcu_torture *old_rp;
1205         static DEFINE_TORTURE_RANDOM(rand);
1206         bool stutter_waited;
1207
1208         VERBOSE_TOROUT_STRING("rcu_torture_writer task started");
1209         if (!can_expedite)
1210                 pr_alert("%s" TORTURE_FLAG
1211                          " GP expediting controlled from boot/sysfs for %s.\n",
1212                          torture_type, cur_ops->name);
1213         if (WARN_ONCE(nsynctypes == 0,
1214                       "%s: No update-side primitives.\n", __func__)) {
1215                 /*
1216                  * No updates primitives, so don't try updating.
1217                  * The resulting test won't be testing much, hence the
1218                  * above WARN_ONCE().
1219                  */
1220                 rcu_torture_writer_state = RTWS_STOPPING;
1221                 torture_kthread_stopping("rcu_torture_writer");
1222                 return 0;
1223         }
1224
1225         do {
1226                 rcu_torture_writer_state = RTWS_FIXED_DELAY;
1227                 torture_hrtimeout_us(500, 1000, &rand);
1228                 rp = rcu_torture_alloc();
1229                 if (rp == NULL)
1230                         continue;
1231                 rp->rtort_pipe_count = 0;
1232                 rcu_torture_writer_state = RTWS_DELAY;
1233                 udelay(torture_random(&rand) & 0x3ff);
1234                 rcu_torture_writer_state = RTWS_REPLACE;
1235                 old_rp = rcu_dereference_check(rcu_torture_current,
1236                                                current == writer_task);
1237                 rp->rtort_mbtest = 1;
1238                 rcu_assign_pointer(rcu_torture_current, rp);
1239                 smp_wmb(); /* Mods to old_rp must follow rcu_assign_pointer() */
1240                 if (old_rp) {
1241                         i = old_rp->rtort_pipe_count;
1242                         if (i > RCU_TORTURE_PIPE_LEN)
1243                                 i = RCU_TORTURE_PIPE_LEN;
1244                         atomic_inc(&rcu_torture_wcount[i]);
1245                         WRITE_ONCE(old_rp->rtort_pipe_count,
1246                                    old_rp->rtort_pipe_count + 1);
1247                         if (cur_ops->get_gp_state && cur_ops->poll_gp_state) {
1248                                 idx = cur_ops->readlock();
1249                                 cookie = cur_ops->get_gp_state();
1250                                 WARN_ONCE(rcu_torture_writer_state != RTWS_DEF_FREE &&
1251                                           cur_ops->poll_gp_state(cookie),
1252                                           "%s: Cookie check 1 failed %s(%d) %lu->%lu\n",
1253                                           __func__,
1254                                           rcu_torture_writer_state_getname(),
1255                                           rcu_torture_writer_state,
1256                                           cookie, cur_ops->get_gp_state());
1257                                 cur_ops->readunlock(idx);
1258                         }
1259                         switch (synctype[torture_random(&rand) % nsynctypes]) {
1260                         case RTWS_DEF_FREE:
1261                                 rcu_torture_writer_state = RTWS_DEF_FREE;
1262                                 cur_ops->deferred_free(old_rp);
1263                                 break;
1264                         case RTWS_EXP_SYNC:
1265                                 rcu_torture_writer_state = RTWS_EXP_SYNC;
1266                                 cur_ops->exp_sync();
1267                                 rcu_torture_pipe_update(old_rp);
1268                                 break;
1269                         case RTWS_COND_GET:
1270                                 rcu_torture_writer_state = RTWS_COND_GET;
1271                                 gp_snap = cur_ops->get_gp_state();
1272                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1273                                 rcu_torture_writer_state = RTWS_COND_SYNC;
1274                                 cur_ops->cond_sync(gp_snap);
1275                                 rcu_torture_pipe_update(old_rp);
1276                                 break;
1277                         case RTWS_POLL_GET:
1278                                 rcu_torture_writer_state = RTWS_POLL_GET;
1279                                 gp_snap = cur_ops->start_gp_poll();
1280                                 rcu_torture_writer_state = RTWS_POLL_WAIT;
1281                                 while (!cur_ops->poll_gp_state(gp_snap))
1282                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1283                                                                   &rand);
1284                                 rcu_torture_pipe_update(old_rp);
1285                                 break;
1286                         case RTWS_SYNC:
1287                                 rcu_torture_writer_state = RTWS_SYNC;
1288                                 cur_ops->sync();
1289                                 rcu_torture_pipe_update(old_rp);
1290                                 break;
1291                         default:
1292                                 WARN_ON_ONCE(1);
1293                                 break;
1294                         }
1295                 }
1296                 WRITE_ONCE(rcu_torture_current_version,
1297                            rcu_torture_current_version + 1);
1298                 /* Cycle through nesting levels of rcu_expedite_gp() calls. */
1299                 if (can_expedite &&
1300                     !(torture_random(&rand) & 0xff & (!!expediting - 1))) {
1301                         WARN_ON_ONCE(expediting == 0 && rcu_gp_is_expedited());
1302                         if (expediting >= 0)
1303                                 rcu_expedite_gp();
1304                         else
1305                                 rcu_unexpedite_gp();
1306                         if (++expediting > 3)
1307                                 expediting = -expediting;
1308                 } else if (!can_expedite) { /* Disabled during boot, recheck. */
1309                         can_expedite = !rcu_gp_is_expedited() &&
1310                                        !rcu_gp_is_normal();
1311                 }
1312                 rcu_torture_writer_state = RTWS_STUTTER;
1313                 boot_ended = rcu_inkernel_boot_has_ended();
1314                 stutter_waited = stutter_wait("rcu_torture_writer");
1315                 if (stutter_waited &&
1316                     !atomic_read(&rcu_fwd_cb_nodelay) &&
1317                     !cur_ops->slow_gps &&
1318                     !torture_must_stop() &&
1319                     boot_ended)
1320                         for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++)
1321                                 if (list_empty(&rcu_tortures[i].rtort_free) &&
1322                                     rcu_access_pointer(rcu_torture_current) !=
1323                                     &rcu_tortures[i]) {
1324                                         rcu_ftrace_dump(DUMP_ALL);
1325                                         WARN(1, "%s: rtort_pipe_count: %d\n", __func__, rcu_tortures[i].rtort_pipe_count);
1326                                 }
1327                 if (stutter_waited)
1328                         sched_set_normal(current, oldnice);
1329         } while (!torture_must_stop());
1330         rcu_torture_current = NULL;  // Let stats task know that we are done.
1331         /* Reset expediting back to unexpedited. */
1332         if (expediting > 0)
1333                 expediting = -expediting;
1334         while (can_expedite && expediting++ < 0)
1335                 rcu_unexpedite_gp();
1336         WARN_ON_ONCE(can_expedite && rcu_gp_is_expedited());
1337         if (!can_expedite)
1338                 pr_alert("%s" TORTURE_FLAG
1339                          " Dynamic grace-period expediting was disabled.\n",
1340                          torture_type);
1341         rcu_torture_writer_state = RTWS_STOPPING;
1342         torture_kthread_stopping("rcu_torture_writer");
1343         return 0;
1344 }
1345
1346 /*
1347  * RCU torture fake writer kthread.  Repeatedly calls sync, with a random
1348  * delay between calls.
1349  */
1350 static int
1351 rcu_torture_fakewriter(void *arg)
1352 {
1353         unsigned long gp_snap;
1354         DEFINE_TORTURE_RANDOM(rand);
1355
1356         VERBOSE_TOROUT_STRING("rcu_torture_fakewriter task started");
1357         set_user_nice(current, MAX_NICE);
1358
1359         if (WARN_ONCE(nsynctypes == 0,
1360                       "%s: No update-side primitives.\n", __func__)) {
1361                 /*
1362                  * No updates primitives, so don't try updating.
1363                  * The resulting test won't be testing much, hence the
1364                  * above WARN_ONCE().
1365                  */
1366                 torture_kthread_stopping("rcu_torture_fakewriter");
1367                 return 0;
1368         }
1369
1370         do {
1371                 torture_hrtimeout_jiffies(torture_random(&rand) % 10, &rand);
1372                 if (cur_ops->cb_barrier != NULL &&
1373                     torture_random(&rand) % (nfakewriters * 8) == 0) {
1374                         cur_ops->cb_barrier();
1375                 } else {
1376                         switch (synctype[torture_random(&rand) % nsynctypes]) {
1377                         case RTWS_DEF_FREE:
1378                                 break;
1379                         case RTWS_EXP_SYNC:
1380                                 cur_ops->exp_sync();
1381                                 break;
1382                         case RTWS_COND_GET:
1383                                 gp_snap = cur_ops->get_gp_state();
1384                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1385                                 cur_ops->cond_sync(gp_snap);
1386                                 break;
1387                         case RTWS_POLL_GET:
1388                                 gp_snap = cur_ops->start_gp_poll();
1389                                 while (!cur_ops->poll_gp_state(gp_snap)) {
1390                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1391                                                                   &rand);
1392                                 }
1393                                 break;
1394                         case RTWS_SYNC:
1395                                 cur_ops->sync();
1396                                 break;
1397                         default:
1398                                 WARN_ON_ONCE(1);
1399                                 break;
1400                         }
1401                 }
1402                 stutter_wait("rcu_torture_fakewriter");
1403         } while (!torture_must_stop());
1404
1405         torture_kthread_stopping("rcu_torture_fakewriter");
1406         return 0;
1407 }
1408
1409 static void rcu_torture_timer_cb(struct rcu_head *rhp)
1410 {
1411         kfree(rhp);
1412 }
1413
1414 // Set up and carry out testing of RCU's global memory ordering
1415 static void rcu_torture_reader_do_mbchk(long myid, struct rcu_torture *rtp,
1416                                         struct torture_random_state *trsp)
1417 {
1418         unsigned long loops;
1419         int noc = torture_num_online_cpus();
1420         int rdrchked;
1421         int rdrchker;
1422         struct rcu_torture_reader_check *rtrcp; // Me.
1423         struct rcu_torture_reader_check *rtrcp_assigner; // Assigned us to do checking.
1424         struct rcu_torture_reader_check *rtrcp_chked; // Reader being checked.
1425         struct rcu_torture_reader_check *rtrcp_chker; // Reader doing checking when not me.
1426
1427         if (myid < 0)
1428                 return; // Don't try this from timer handlers.
1429
1430         // Increment my counter.
1431         rtrcp = &rcu_torture_reader_mbchk[myid];
1432         WRITE_ONCE(rtrcp->rtc_myloops, rtrcp->rtc_myloops + 1);
1433
1434         // Attempt to assign someone else some checking work.
1435         rdrchked = torture_random(trsp) % nrealreaders;
1436         rtrcp_chked = &rcu_torture_reader_mbchk[rdrchked];
1437         rdrchker = torture_random(trsp) % nrealreaders;
1438         rtrcp_chker = &rcu_torture_reader_mbchk[rdrchker];
1439         if (rdrchked != myid && rdrchked != rdrchker && noc >= rdrchked && noc >= rdrchker &&
1440             smp_load_acquire(&rtrcp->rtc_chkrdr) < 0 && // Pairs with smp_store_release below.
1441             !READ_ONCE(rtp->rtort_chkp) &&
1442             !smp_load_acquire(&rtrcp_chker->rtc_assigner)) { // Pairs with smp_store_release below.
1443                 rtrcp->rtc_chkloops = READ_ONCE(rtrcp_chked->rtc_myloops);
1444                 WARN_ON_ONCE(rtrcp->rtc_chkrdr >= 0);
1445                 rtrcp->rtc_chkrdr = rdrchked;
1446                 WARN_ON_ONCE(rtrcp->rtc_ready); // This gets set after the grace period ends.
1447                 if (cmpxchg_relaxed(&rtrcp_chker->rtc_assigner, NULL, rtrcp) ||
1448                     cmpxchg_relaxed(&rtp->rtort_chkp, NULL, rtrcp))
1449                         (void)cmpxchg_relaxed(&rtrcp_chker->rtc_assigner, rtrcp, NULL); // Back out.
1450         }
1451
1452         // If assigned some completed work, do it!
1453         rtrcp_assigner = READ_ONCE(rtrcp->rtc_assigner);
1454         if (!rtrcp_assigner || !smp_load_acquire(&rtrcp_assigner->rtc_ready))
1455                 return; // No work or work not yet ready.
1456         rdrchked = rtrcp_assigner->rtc_chkrdr;
1457         if (WARN_ON_ONCE(rdrchked < 0))
1458                 return;
1459         rtrcp_chked = &rcu_torture_reader_mbchk[rdrchked];
1460         loops = READ_ONCE(rtrcp_chked->rtc_myloops);
1461         atomic_inc(&n_rcu_torture_mbchk_tries);
1462         if (ULONG_CMP_LT(loops, rtrcp_assigner->rtc_chkloops))
1463                 atomic_inc(&n_rcu_torture_mbchk_fail);
1464         rtrcp_assigner->rtc_chkloops = loops + ULONG_MAX / 2;
1465         rtrcp_assigner->rtc_ready = 0;
1466         smp_store_release(&rtrcp->rtc_assigner, NULL); // Someone else can assign us work.
1467         smp_store_release(&rtrcp_assigner->rtc_chkrdr, -1); // Assigner can again assign.
1468 }
1469
1470 /*
1471  * Do one extension of an RCU read-side critical section using the
1472  * current reader state in readstate (set to zero for initial entry
1473  * to extended critical section), set the new state as specified by
1474  * newstate (set to zero for final exit from extended critical section),
1475  * and random-number-generator state in trsp.  If this is neither the
1476  * beginning or end of the critical section and if there was actually a
1477  * change, do a ->read_delay().
1478  */
1479 static void rcutorture_one_extend(int *readstate, int newstate,
1480                                   struct torture_random_state *trsp,
1481                                   struct rt_read_seg *rtrsp)
1482 {
1483         unsigned long flags;
1484         int idxnew1 = -1;
1485         int idxnew2 = -1;
1486         int idxold1 = *readstate;
1487         int idxold2 = idxold1;
1488         int statesnew = ~*readstate & newstate;
1489         int statesold = *readstate & ~newstate;
1490
1491         WARN_ON_ONCE(idxold2 < 0);
1492         WARN_ON_ONCE((idxold2 >> RCUTORTURE_RDR_SHIFT_2) > 1);
1493         rtrsp->rt_readstate = newstate;
1494
1495         /* First, put new protection in place to avoid critical-section gap. */
1496         if (statesnew & RCUTORTURE_RDR_BH)
1497                 local_bh_disable();
1498         if (statesnew & RCUTORTURE_RDR_RBH)
1499                 rcu_read_lock_bh();
1500         if (statesnew & RCUTORTURE_RDR_IRQ)
1501                 local_irq_disable();
1502         if (statesnew & RCUTORTURE_RDR_PREEMPT)
1503                 preempt_disable();
1504         if (statesnew & RCUTORTURE_RDR_SCHED)
1505                 rcu_read_lock_sched();
1506         if (statesnew & RCUTORTURE_RDR_RCU_1)
1507                 idxnew1 = (cur_ops->readlock() & 0x1) << RCUTORTURE_RDR_SHIFT_1;
1508         if (statesnew & RCUTORTURE_RDR_RCU_2)
1509                 idxnew2 = (cur_ops->readlock() & 0x1) << RCUTORTURE_RDR_SHIFT_2;
1510
1511         /*
1512          * Next, remove old protection, in decreasing order of strength
1513          * to avoid unlock paths that aren't safe in the stronger
1514          * context. Namely: BH can not be enabled with disabled interrupts.
1515          * Additionally PREEMPT_RT requires that BH is enabled in preemptible
1516          * context.
1517          */
1518         if (statesold & RCUTORTURE_RDR_IRQ)
1519                 local_irq_enable();
1520         if (statesold & RCUTORTURE_RDR_PREEMPT)
1521                 preempt_enable();
1522         if (statesold & RCUTORTURE_RDR_SCHED)
1523                 rcu_read_unlock_sched();
1524         if (statesold & RCUTORTURE_RDR_BH)
1525                 local_bh_enable();
1526         if (statesold & RCUTORTURE_RDR_RBH)
1527                 rcu_read_unlock_bh();
1528         if (statesold & RCUTORTURE_RDR_RCU_2) {
1529                 cur_ops->readunlock((idxold2 >> RCUTORTURE_RDR_SHIFT_2) & 0x1);
1530                 WARN_ON_ONCE(idxnew2 != -1);
1531                 idxold2 = 0;
1532         }
1533         if (statesold & RCUTORTURE_RDR_RCU_1) {
1534                 bool lockit;
1535
1536                 lockit = !cur_ops->no_pi_lock && !statesnew && !(torture_random(trsp) & 0xffff);
1537                 if (lockit)
1538                         raw_spin_lock_irqsave(&current->pi_lock, flags);
1539                 cur_ops->readunlock((idxold1 >> RCUTORTURE_RDR_SHIFT_1) & 0x1);
1540                 WARN_ON_ONCE(idxnew1 != -1);
1541                 idxold1 = 0;
1542                 if (lockit)
1543                         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
1544         }
1545
1546         /* Delay if neither beginning nor end and there was a change. */
1547         if ((statesnew || statesold) && *readstate && newstate)
1548                 cur_ops->read_delay(trsp, rtrsp);
1549
1550         /* Update the reader state. */
1551         if (idxnew1 == -1)
1552                 idxnew1 = idxold1 & RCUTORTURE_RDR_MASK_1;
1553         WARN_ON_ONCE(idxnew1 < 0);
1554         if (WARN_ON_ONCE((idxnew1 >> RCUTORTURE_RDR_SHIFT_1) > 1))
1555                 pr_info("Unexpected idxnew1 value of %#x\n", idxnew1);
1556         if (idxnew2 == -1)
1557                 idxnew2 = idxold2 & RCUTORTURE_RDR_MASK_2;
1558         WARN_ON_ONCE(idxnew2 < 0);
1559         WARN_ON_ONCE((idxnew2 >> RCUTORTURE_RDR_SHIFT_2) > 1);
1560         *readstate = idxnew1 | idxnew2 | newstate;
1561         WARN_ON_ONCE(*readstate < 0);
1562         if (WARN_ON_ONCE((*readstate >> RCUTORTURE_RDR_SHIFT_2) > 1))
1563                 pr_info("Unexpected idxnew2 value of %#x\n", idxnew2);
1564 }
1565
1566 /* Return the biggest extendables mask given current RCU and boot parameters. */
1567 static int rcutorture_extend_mask_max(void)
1568 {
1569         int mask;
1570
1571         WARN_ON_ONCE(extendables & ~RCUTORTURE_MAX_EXTEND);
1572         mask = extendables & RCUTORTURE_MAX_EXTEND & cur_ops->extendables;
1573         mask = mask | RCUTORTURE_RDR_RCU_1 | RCUTORTURE_RDR_RCU_2;
1574         return mask;
1575 }
1576
1577 /* Return a random protection state mask, but with at least one bit set. */
1578 static int
1579 rcutorture_extend_mask(int oldmask, struct torture_random_state *trsp)
1580 {
1581         int mask = rcutorture_extend_mask_max();
1582         unsigned long randmask1 = torture_random(trsp) >> 8;
1583         unsigned long randmask2 = randmask1 >> 3;
1584         unsigned long preempts = RCUTORTURE_RDR_PREEMPT | RCUTORTURE_RDR_SCHED;
1585         unsigned long preempts_irq = preempts | RCUTORTURE_RDR_IRQ;
1586         unsigned long bhs = RCUTORTURE_RDR_BH | RCUTORTURE_RDR_RBH;
1587
1588         WARN_ON_ONCE(mask >> RCUTORTURE_RDR_SHIFT_1);
1589         /* Mostly only one bit (need preemption!), sometimes lots of bits. */
1590         if (!(randmask1 & 0x7))
1591                 mask = mask & randmask2;
1592         else
1593                 mask = mask & (1 << (randmask2 % RCUTORTURE_RDR_NBITS));
1594
1595         // Can't have nested RCU reader without outer RCU reader.
1596         if (!(mask & RCUTORTURE_RDR_RCU_1) && (mask & RCUTORTURE_RDR_RCU_2)) {
1597                 if (oldmask & RCUTORTURE_RDR_RCU_1)
1598                         mask &= ~RCUTORTURE_RDR_RCU_2;
1599                 else
1600                         mask |= RCUTORTURE_RDR_RCU_1;
1601         }
1602
1603         /*
1604          * Can't enable bh w/irq disabled.
1605          */
1606         if (mask & RCUTORTURE_RDR_IRQ)
1607                 mask |= oldmask & bhs;
1608
1609         /*
1610          * Ideally these sequences would be detected in debug builds
1611          * (regardless of RT), but until then don't stop testing
1612          * them on non-RT.
1613          */
1614         if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
1615                 /* Can't modify BH in atomic context */
1616                 if (oldmask & preempts_irq)
1617                         mask &= ~bhs;
1618                 if ((oldmask | mask) & preempts_irq)
1619                         mask |= oldmask & bhs;
1620         }
1621
1622         return mask ?: RCUTORTURE_RDR_RCU_1;
1623 }
1624
1625 /*
1626  * Do a randomly selected number of extensions of an existing RCU read-side
1627  * critical section.
1628  */
1629 static struct rt_read_seg *
1630 rcutorture_loop_extend(int *readstate, struct torture_random_state *trsp,
1631                        struct rt_read_seg *rtrsp)
1632 {
1633         int i;
1634         int j;
1635         int mask = rcutorture_extend_mask_max();
1636
1637         WARN_ON_ONCE(!*readstate); /* -Existing- RCU read-side critsect! */
1638         if (!((mask - 1) & mask))
1639                 return rtrsp;  /* Current RCU reader not extendable. */
1640         /* Bias towards larger numbers of loops. */
1641         i = (torture_random(trsp) >> 3);
1642         i = ((i | (i >> 3)) & RCUTORTURE_RDR_MAX_LOOPS) + 1;
1643         for (j = 0; j < i; j++) {
1644                 mask = rcutorture_extend_mask(*readstate, trsp);
1645                 rcutorture_one_extend(readstate, mask, trsp, &rtrsp[j]);
1646         }
1647         return &rtrsp[j];
1648 }
1649
1650 /*
1651  * Do one read-side critical section, returning false if there was
1652  * no data to read.  Can be invoked both from process context and
1653  * from a timer handler.
1654  */
1655 static bool rcu_torture_one_read(struct torture_random_state *trsp, long myid)
1656 {
1657         unsigned long cookie;
1658         int i;
1659         unsigned long started;
1660         unsigned long completed;
1661         int newstate;
1662         struct rcu_torture *p;
1663         int pipe_count;
1664         int readstate = 0;
1665         struct rt_read_seg rtseg[RCUTORTURE_RDR_MAX_SEGS] = { { 0 } };
1666         struct rt_read_seg *rtrsp = &rtseg[0];
1667         struct rt_read_seg *rtrsp1;
1668         unsigned long long ts;
1669
1670         WARN_ON_ONCE(!rcu_is_watching());
1671         newstate = rcutorture_extend_mask(readstate, trsp);
1672         rcutorture_one_extend(&readstate, newstate, trsp, rtrsp++);
1673         if (cur_ops->get_gp_state && cur_ops->poll_gp_state)
1674                 cookie = cur_ops->get_gp_state();
1675         started = cur_ops->get_gp_seq();
1676         ts = rcu_trace_clock_local();
1677         p = rcu_dereference_check(rcu_torture_current,
1678                                   !cur_ops->readlock_held || cur_ops->readlock_held());
1679         if (p == NULL) {
1680                 /* Wait for rcu_torture_writer to get underway */
1681                 rcutorture_one_extend(&readstate, 0, trsp, rtrsp);
1682                 return false;
1683         }
1684         if (p->rtort_mbtest == 0)
1685                 atomic_inc(&n_rcu_torture_mberror);
1686         rcu_torture_reader_do_mbchk(myid, p, trsp);
1687         rtrsp = rcutorture_loop_extend(&readstate, trsp, rtrsp);
1688         preempt_disable();
1689         pipe_count = READ_ONCE(p->rtort_pipe_count);
1690         if (pipe_count > RCU_TORTURE_PIPE_LEN) {
1691                 /* Should not happen, but... */
1692                 pipe_count = RCU_TORTURE_PIPE_LEN;
1693         }
1694         completed = cur_ops->get_gp_seq();
1695         if (pipe_count > 1) {
1696                 do_trace_rcu_torture_read(cur_ops->name, &p->rtort_rcu,
1697                                           ts, started, completed);
1698                 rcu_ftrace_dump(DUMP_ALL);
1699         }
1700         __this_cpu_inc(rcu_torture_count[pipe_count]);
1701         completed = rcutorture_seq_diff(completed, started);
1702         if (completed > RCU_TORTURE_PIPE_LEN) {
1703                 /* Should not happen, but... */
1704                 completed = RCU_TORTURE_PIPE_LEN;
1705         }
1706         __this_cpu_inc(rcu_torture_batch[completed]);
1707         preempt_enable();
1708         if (cur_ops->get_gp_state && cur_ops->poll_gp_state)
1709                 WARN_ONCE(cur_ops->poll_gp_state(cookie),
1710                           "%s: Cookie check 2 failed %s(%d) %lu->%lu\n",
1711                           __func__,
1712                           rcu_torture_writer_state_getname(),
1713                           rcu_torture_writer_state,
1714                           cookie, cur_ops->get_gp_state());
1715         rcutorture_one_extend(&readstate, 0, trsp, rtrsp);
1716         WARN_ON_ONCE(readstate);
1717         // This next splat is expected behavior if leakpointer, especially
1718         // for CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels.
1719         WARN_ON_ONCE(leakpointer && READ_ONCE(p->rtort_pipe_count) > 1);
1720
1721         /* If error or close call, record the sequence of reader protections. */
1722         if ((pipe_count > 1 || completed > 1) && !xchg(&err_segs_recorded, 1)) {
1723                 i = 0;
1724                 for (rtrsp1 = &rtseg[0]; rtrsp1 < rtrsp; rtrsp1++)
1725                         err_segs[i++] = *rtrsp1;
1726                 rt_read_nsegs = i;
1727         }
1728
1729         return true;
1730 }
1731
1732 static DEFINE_TORTURE_RANDOM_PERCPU(rcu_torture_timer_rand);
1733
1734 /*
1735  * RCU torture reader from timer handler.  Dereferences rcu_torture_current,
1736  * incrementing the corresponding element of the pipeline array.  The
1737  * counter in the element should never be greater than 1, otherwise, the
1738  * RCU implementation is broken.
1739  */
1740 static void rcu_torture_timer(struct timer_list *unused)
1741 {
1742         atomic_long_inc(&n_rcu_torture_timers);
1743         (void)rcu_torture_one_read(this_cpu_ptr(&rcu_torture_timer_rand), -1);
1744
1745         /* Test call_rcu() invocation from interrupt handler. */
1746         if (cur_ops->call) {
1747                 struct rcu_head *rhp = kmalloc(sizeof(*rhp), GFP_NOWAIT);
1748
1749                 if (rhp)
1750                         cur_ops->call(rhp, rcu_torture_timer_cb);
1751         }
1752 }
1753
1754 /*
1755  * RCU torture reader kthread.  Repeatedly dereferences rcu_torture_current,
1756  * incrementing the corresponding element of the pipeline array.  The
1757  * counter in the element should never be greater than 1, otherwise, the
1758  * RCU implementation is broken.
1759  */
1760 static int
1761 rcu_torture_reader(void *arg)
1762 {
1763         unsigned long lastsleep = jiffies;
1764         long myid = (long)arg;
1765         int mynumonline = myid;
1766         DEFINE_TORTURE_RANDOM(rand);
1767         struct timer_list t;
1768
1769         VERBOSE_TOROUT_STRING("rcu_torture_reader task started");
1770         set_user_nice(current, MAX_NICE);
1771         if (irqreader && cur_ops->irq_capable)
1772                 timer_setup_on_stack(&t, rcu_torture_timer, 0);
1773         tick_dep_set_task(current, TICK_DEP_BIT_RCU);
1774         do {
1775                 if (irqreader && cur_ops->irq_capable) {
1776                         if (!timer_pending(&t))
1777                                 mod_timer(&t, jiffies + 1);
1778                 }
1779                 if (!rcu_torture_one_read(&rand, myid) && !torture_must_stop())
1780                         schedule_timeout_interruptible(HZ);
1781                 if (time_after(jiffies, lastsleep) && !torture_must_stop()) {
1782                         torture_hrtimeout_us(500, 1000, &rand);
1783                         lastsleep = jiffies + 10;
1784                 }
1785                 while (torture_num_online_cpus() < mynumonline && !torture_must_stop())
1786                         schedule_timeout_interruptible(HZ / 5);
1787                 stutter_wait("rcu_torture_reader");
1788         } while (!torture_must_stop());
1789         if (irqreader && cur_ops->irq_capable) {
1790                 del_timer_sync(&t);
1791                 destroy_timer_on_stack(&t);
1792         }
1793         tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
1794         torture_kthread_stopping("rcu_torture_reader");
1795         return 0;
1796 }
1797
1798 /*
1799  * Randomly Toggle CPUs' callback-offload state.  This uses hrtimers to
1800  * increase race probabilities and fuzzes the interval between toggling.
1801  */
1802 static int rcu_nocb_toggle(void *arg)
1803 {
1804         int cpu;
1805         int maxcpu = -1;
1806         int oldnice = task_nice(current);
1807         long r;
1808         DEFINE_TORTURE_RANDOM(rand);
1809         ktime_t toggle_delay;
1810         unsigned long toggle_fuzz;
1811         ktime_t toggle_interval = ms_to_ktime(nocbs_toggle);
1812
1813         VERBOSE_TOROUT_STRING("rcu_nocb_toggle task started");
1814         while (!rcu_inkernel_boot_has_ended())
1815                 schedule_timeout_interruptible(HZ / 10);
1816         for_each_online_cpu(cpu)
1817                 maxcpu = cpu;
1818         WARN_ON(maxcpu < 0);
1819         if (toggle_interval > ULONG_MAX)
1820                 toggle_fuzz = ULONG_MAX >> 3;
1821         else
1822                 toggle_fuzz = toggle_interval >> 3;
1823         if (toggle_fuzz <= 0)
1824                 toggle_fuzz = NSEC_PER_USEC;
1825         do {
1826                 r = torture_random(&rand);
1827                 cpu = (r >> 4) % (maxcpu + 1);
1828                 if (r & 0x1) {
1829                         rcu_nocb_cpu_offload(cpu);
1830                         atomic_long_inc(&n_nocb_offload);
1831                 } else {
1832                         rcu_nocb_cpu_deoffload(cpu);
1833                         atomic_long_inc(&n_nocb_deoffload);
1834                 }
1835                 toggle_delay = torture_random(&rand) % toggle_fuzz + toggle_interval;
1836                 set_current_state(TASK_INTERRUPTIBLE);
1837                 schedule_hrtimeout(&toggle_delay, HRTIMER_MODE_REL);
1838                 if (stutter_wait("rcu_nocb_toggle"))
1839                         sched_set_normal(current, oldnice);
1840         } while (!torture_must_stop());
1841         torture_kthread_stopping("rcu_nocb_toggle");
1842         return 0;
1843 }
1844
1845 /*
1846  * Print torture statistics.  Caller must ensure that there is only
1847  * one call to this function at a given time!!!  This is normally
1848  * accomplished by relying on the module system to only have one copy
1849  * of the module loaded, and then by giving the rcu_torture_stats
1850  * kthread full control (or the init/cleanup functions when rcu_torture_stats
1851  * thread is not running).
1852  */
1853 static void
1854 rcu_torture_stats_print(void)
1855 {
1856         int cpu;
1857         int i;
1858         long pipesummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
1859         long batchsummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
1860         struct rcu_torture *rtcp;
1861         static unsigned long rtcv_snap = ULONG_MAX;
1862         static bool splatted;
1863         struct task_struct *wtp;
1864
1865         for_each_possible_cpu(cpu) {
1866                 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
1867                         pipesummary[i] += READ_ONCE(per_cpu(rcu_torture_count, cpu)[i]);
1868                         batchsummary[i] += READ_ONCE(per_cpu(rcu_torture_batch, cpu)[i]);
1869                 }
1870         }
1871         for (i = RCU_TORTURE_PIPE_LEN - 1; i >= 0; i--) {
1872                 if (pipesummary[i] != 0)
1873                         break;
1874         }
1875
1876         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1877         rtcp = rcu_access_pointer(rcu_torture_current);
1878         pr_cont("rtc: %p %s: %lu tfle: %d rta: %d rtaf: %d rtf: %d ",
1879                 rtcp,
1880                 rtcp && !rcu_stall_is_suppressed_at_boot() ? "ver" : "VER",
1881                 rcu_torture_current_version,
1882                 list_empty(&rcu_torture_freelist),
1883                 atomic_read(&n_rcu_torture_alloc),
1884                 atomic_read(&n_rcu_torture_alloc_fail),
1885                 atomic_read(&n_rcu_torture_free));
1886         pr_cont("rtmbe: %d rtmbkf: %d/%d rtbe: %ld rtbke: %ld rtbre: %ld ",
1887                 atomic_read(&n_rcu_torture_mberror),
1888                 atomic_read(&n_rcu_torture_mbchk_fail), atomic_read(&n_rcu_torture_mbchk_tries),
1889                 n_rcu_torture_barrier_error,
1890                 n_rcu_torture_boost_ktrerror,
1891                 n_rcu_torture_boost_rterror);
1892         pr_cont("rtbf: %ld rtb: %ld nt: %ld ",
1893                 n_rcu_torture_boost_failure,
1894                 n_rcu_torture_boosts,
1895                 atomic_long_read(&n_rcu_torture_timers));
1896         torture_onoff_stats();
1897         pr_cont("barrier: %ld/%ld:%ld ",
1898                 data_race(n_barrier_successes),
1899                 data_race(n_barrier_attempts),
1900                 data_race(n_rcu_torture_barrier_error));
1901         pr_cont("read-exits: %ld ", data_race(n_read_exits)); // Statistic.
1902         pr_cont("nocb-toggles: %ld:%ld\n",
1903                 atomic_long_read(&n_nocb_offload), atomic_long_read(&n_nocb_deoffload));
1904
1905         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1906         if (atomic_read(&n_rcu_torture_mberror) ||
1907             atomic_read(&n_rcu_torture_mbchk_fail) ||
1908             n_rcu_torture_barrier_error || n_rcu_torture_boost_ktrerror ||
1909             n_rcu_torture_boost_rterror || n_rcu_torture_boost_failure ||
1910             i > 1) {
1911                 pr_cont("%s", "!!! ");
1912                 atomic_inc(&n_rcu_torture_error);
1913                 WARN_ON_ONCE(atomic_read(&n_rcu_torture_mberror));
1914                 WARN_ON_ONCE(atomic_read(&n_rcu_torture_mbchk_fail));
1915                 WARN_ON_ONCE(n_rcu_torture_barrier_error);  // rcu_barrier()
1916                 WARN_ON_ONCE(n_rcu_torture_boost_ktrerror); // no boost kthread
1917                 WARN_ON_ONCE(n_rcu_torture_boost_rterror); // can't set RT prio
1918                 WARN_ON_ONCE(n_rcu_torture_boost_failure); // boost failed (TIMER_SOFTIRQ RT prio?)
1919                 WARN_ON_ONCE(i > 1); // Too-short grace period
1920         }
1921         pr_cont("Reader Pipe: ");
1922         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
1923                 pr_cont(" %ld", pipesummary[i]);
1924         pr_cont("\n");
1925
1926         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1927         pr_cont("Reader Batch: ");
1928         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
1929                 pr_cont(" %ld", batchsummary[i]);
1930         pr_cont("\n");
1931
1932         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
1933         pr_cont("Free-Block Circulation: ");
1934         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
1935                 pr_cont(" %d", atomic_read(&rcu_torture_wcount[i]));
1936         }
1937         pr_cont("\n");
1938
1939         if (cur_ops->stats)
1940                 cur_ops->stats();
1941         if (rtcv_snap == rcu_torture_current_version &&
1942             rcu_access_pointer(rcu_torture_current) &&
1943             !rcu_stall_is_suppressed()) {
1944                 int __maybe_unused flags = 0;
1945                 unsigned long __maybe_unused gp_seq = 0;
1946
1947                 rcutorture_get_gp_data(cur_ops->ttype,
1948                                        &flags, &gp_seq);
1949                 srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp,
1950                                         &flags, &gp_seq);
1951                 wtp = READ_ONCE(writer_task);
1952                 pr_alert("??? Writer stall state %s(%d) g%lu f%#x ->state %#x cpu %d\n",
1953                          rcu_torture_writer_state_getname(),
1954                          rcu_torture_writer_state, gp_seq, flags,
1955                          wtp == NULL ? ~0U : wtp->__state,
1956                          wtp == NULL ? -1 : (int)task_cpu(wtp));
1957                 if (!splatted && wtp) {
1958                         sched_show_task(wtp);
1959                         splatted = true;
1960                 }
1961                 if (cur_ops->gp_kthread_dbg)
1962                         cur_ops->gp_kthread_dbg();
1963                 rcu_ftrace_dump(DUMP_ALL);
1964         }
1965         rtcv_snap = rcu_torture_current_version;
1966 }
1967
1968 /*
1969  * Periodically prints torture statistics, if periodic statistics printing
1970  * was specified via the stat_interval module parameter.
1971  */
1972 static int
1973 rcu_torture_stats(void *arg)
1974 {
1975         VERBOSE_TOROUT_STRING("rcu_torture_stats task started");
1976         do {
1977                 schedule_timeout_interruptible(stat_interval * HZ);
1978                 rcu_torture_stats_print();
1979                 torture_shutdown_absorb("rcu_torture_stats");
1980         } while (!torture_must_stop());
1981         torture_kthread_stopping("rcu_torture_stats");
1982         return 0;
1983 }
1984
1985 /* Test mem_dump_obj() and friends.  */
1986 static void rcu_torture_mem_dump_obj(void)
1987 {
1988         struct rcu_head *rhp;
1989         struct kmem_cache *kcp;
1990         static int z;
1991
1992         kcp = kmem_cache_create("rcuscale", 136, 8, SLAB_STORE_USER, NULL);
1993         rhp = kmem_cache_alloc(kcp, GFP_KERNEL);
1994         pr_alert("mem_dump_obj() slab test: rcu_torture_stats = %px, &rhp = %px, rhp = %px, &z = %px\n", stats_task, &rhp, rhp, &z);
1995         pr_alert("mem_dump_obj(ZERO_SIZE_PTR):");
1996         mem_dump_obj(ZERO_SIZE_PTR);
1997         pr_alert("mem_dump_obj(NULL):");
1998         mem_dump_obj(NULL);
1999         pr_alert("mem_dump_obj(%px):", &rhp);
2000         mem_dump_obj(&rhp);
2001         pr_alert("mem_dump_obj(%px):", rhp);
2002         mem_dump_obj(rhp);
2003         pr_alert("mem_dump_obj(%px):", &rhp->func);
2004         mem_dump_obj(&rhp->func);
2005         pr_alert("mem_dump_obj(%px):", &z);
2006         mem_dump_obj(&z);
2007         kmem_cache_free(kcp, rhp);
2008         kmem_cache_destroy(kcp);
2009         rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
2010         pr_alert("mem_dump_obj() kmalloc test: rcu_torture_stats = %px, &rhp = %px, rhp = %px\n", stats_task, &rhp, rhp);
2011         pr_alert("mem_dump_obj(kmalloc %px):", rhp);
2012         mem_dump_obj(rhp);
2013         pr_alert("mem_dump_obj(kmalloc %px):", &rhp->func);
2014         mem_dump_obj(&rhp->func);
2015         kfree(rhp);
2016         rhp = vmalloc(4096);
2017         pr_alert("mem_dump_obj() vmalloc test: rcu_torture_stats = %px, &rhp = %px, rhp = %px\n", stats_task, &rhp, rhp);
2018         pr_alert("mem_dump_obj(vmalloc %px):", rhp);
2019         mem_dump_obj(rhp);
2020         pr_alert("mem_dump_obj(vmalloc %px):", &rhp->func);
2021         mem_dump_obj(&rhp->func);
2022         vfree(rhp);
2023 }
2024
2025 static void
2026 rcu_torture_print_module_parms(struct rcu_torture_ops *cur_ops, const char *tag)
2027 {
2028         pr_alert("%s" TORTURE_FLAG
2029                  "--- %s: nreaders=%d nfakewriters=%d "
2030                  "stat_interval=%d verbose=%d test_no_idle_hz=%d "
2031                  "shuffle_interval=%d stutter=%d irqreader=%d "
2032                  "fqs_duration=%d fqs_holdoff=%d fqs_stutter=%d "
2033                  "test_boost=%d/%d test_boost_interval=%d "
2034                  "test_boost_duration=%d shutdown_secs=%d "
2035                  "stall_cpu=%d stall_cpu_holdoff=%d stall_cpu_irqsoff=%d "
2036                  "stall_cpu_block=%d "
2037                  "n_barrier_cbs=%d "
2038                  "onoff_interval=%d onoff_holdoff=%d "
2039                  "read_exit_delay=%d read_exit_burst=%d "
2040                  "nocbs_nthreads=%d nocbs_toggle=%d\n",
2041                  torture_type, tag, nrealreaders, nfakewriters,
2042                  stat_interval, verbose, test_no_idle_hz, shuffle_interval,
2043                  stutter, irqreader, fqs_duration, fqs_holdoff, fqs_stutter,
2044                  test_boost, cur_ops->can_boost,
2045                  test_boost_interval, test_boost_duration, shutdown_secs,
2046                  stall_cpu, stall_cpu_holdoff, stall_cpu_irqsoff,
2047                  stall_cpu_block,
2048                  n_barrier_cbs,
2049                  onoff_interval, onoff_holdoff,
2050                  read_exit_delay, read_exit_burst,
2051                  nocbs_nthreads, nocbs_toggle);
2052 }
2053
2054 static int rcutorture_booster_cleanup(unsigned int cpu)
2055 {
2056         struct task_struct *t;
2057
2058         if (boost_tasks[cpu] == NULL)
2059                 return 0;
2060         mutex_lock(&boost_mutex);
2061         t = boost_tasks[cpu];
2062         boost_tasks[cpu] = NULL;
2063         rcu_torture_enable_rt_throttle();
2064         mutex_unlock(&boost_mutex);
2065
2066         /* This must be outside of the mutex, otherwise deadlock! */
2067         torture_stop_kthread(rcu_torture_boost, t);
2068         return 0;
2069 }
2070
2071 static int rcutorture_booster_init(unsigned int cpu)
2072 {
2073         int retval;
2074
2075         if (boost_tasks[cpu] != NULL)
2076                 return 0;  /* Already created, nothing more to do. */
2077
2078         /* Don't allow time recalculation while creating a new task. */
2079         mutex_lock(&boost_mutex);
2080         rcu_torture_disable_rt_throttle();
2081         VERBOSE_TOROUT_STRING("Creating rcu_torture_boost task");
2082         boost_tasks[cpu] = kthread_run_on_cpu(rcu_torture_boost, NULL,
2083                                               cpu, "rcu_torture_boost_%u");
2084         if (IS_ERR(boost_tasks[cpu])) {
2085                 retval = PTR_ERR(boost_tasks[cpu]);
2086                 VERBOSE_TOROUT_STRING("rcu_torture_boost task create failed");
2087                 n_rcu_torture_boost_ktrerror++;
2088                 boost_tasks[cpu] = NULL;
2089                 mutex_unlock(&boost_mutex);
2090                 return retval;
2091         }
2092         mutex_unlock(&boost_mutex);
2093         return 0;
2094 }
2095
2096 /*
2097  * CPU-stall kthread.  It waits as specified by stall_cpu_holdoff, then
2098  * induces a CPU stall for the time specified by stall_cpu.
2099  */
2100 static int rcu_torture_stall(void *args)
2101 {
2102         int idx;
2103         unsigned long stop_at;
2104
2105         VERBOSE_TOROUT_STRING("rcu_torture_stall task started");
2106         if (stall_cpu_holdoff > 0) {
2107                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin holdoff");
2108                 schedule_timeout_interruptible(stall_cpu_holdoff * HZ);
2109                 VERBOSE_TOROUT_STRING("rcu_torture_stall end holdoff");
2110         }
2111         if (!kthread_should_stop() && stall_gp_kthread > 0) {
2112                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin GP stall");
2113                 rcu_gp_set_torture_wait(stall_gp_kthread * HZ);
2114                 for (idx = 0; idx < stall_gp_kthread + 2; idx++) {
2115                         if (kthread_should_stop())
2116                                 break;
2117                         schedule_timeout_uninterruptible(HZ);
2118                 }
2119         }
2120         if (!kthread_should_stop() && stall_cpu > 0) {
2121                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin CPU stall");
2122                 stop_at = ktime_get_seconds() + stall_cpu;
2123                 /* RCU CPU stall is expected behavior in following code. */
2124                 idx = cur_ops->readlock();
2125                 if (stall_cpu_irqsoff)
2126                         local_irq_disable();
2127                 else if (!stall_cpu_block)
2128                         preempt_disable();
2129                 pr_alert("%s start on CPU %d.\n",
2130                           __func__, raw_smp_processor_id());
2131                 while (ULONG_CMP_LT((unsigned long)ktime_get_seconds(),
2132                                     stop_at))
2133                         if (stall_cpu_block) {
2134 #ifdef CONFIG_PREEMPTION
2135                                 preempt_schedule();
2136 #else
2137                                 schedule_timeout_uninterruptible(HZ);
2138 #endif
2139                         } else if (stall_no_softlockup) {
2140                                 touch_softlockup_watchdog();
2141                         }
2142                 if (stall_cpu_irqsoff)
2143                         local_irq_enable();
2144                 else if (!stall_cpu_block)
2145                         preempt_enable();
2146                 cur_ops->readunlock(idx);
2147         }
2148         pr_alert("%s end.\n", __func__);
2149         torture_shutdown_absorb("rcu_torture_stall");
2150         while (!kthread_should_stop())
2151                 schedule_timeout_interruptible(10 * HZ);
2152         return 0;
2153 }
2154
2155 /* Spawn CPU-stall kthread, if stall_cpu specified. */
2156 static int __init rcu_torture_stall_init(void)
2157 {
2158         if (stall_cpu <= 0 && stall_gp_kthread <= 0)
2159                 return 0;
2160         return torture_create_kthread(rcu_torture_stall, NULL, stall_task);
2161 }
2162
2163 /* State structure for forward-progress self-propagating RCU callback. */
2164 struct fwd_cb_state {
2165         struct rcu_head rh;
2166         int stop;
2167 };
2168
2169 /*
2170  * Forward-progress self-propagating RCU callback function.  Because
2171  * callbacks run from softirq, this function is an implicit RCU read-side
2172  * critical section.
2173  */
2174 static void rcu_torture_fwd_prog_cb(struct rcu_head *rhp)
2175 {
2176         struct fwd_cb_state *fcsp = container_of(rhp, struct fwd_cb_state, rh);
2177
2178         if (READ_ONCE(fcsp->stop)) {
2179                 WRITE_ONCE(fcsp->stop, 2);
2180                 return;
2181         }
2182         cur_ops->call(&fcsp->rh, rcu_torture_fwd_prog_cb);
2183 }
2184
2185 /* State for continuous-flood RCU callbacks. */
2186 struct rcu_fwd_cb {
2187         struct rcu_head rh;
2188         struct rcu_fwd_cb *rfc_next;
2189         struct rcu_fwd *rfc_rfp;
2190         int rfc_gps;
2191 };
2192
2193 #define MAX_FWD_CB_JIFFIES      (8 * HZ) /* Maximum CB test duration. */
2194 #define MIN_FWD_CB_LAUNDERS     3       /* This many CB invocations to count. */
2195 #define MIN_FWD_CBS_LAUNDERED   100     /* Number of counted CBs. */
2196 #define FWD_CBS_HIST_DIV        10      /* Histogram buckets/second. */
2197 #define N_LAUNDERS_HIST (2 * MAX_FWD_CB_JIFFIES / (HZ / FWD_CBS_HIST_DIV))
2198
2199 struct rcu_launder_hist {
2200         long n_launders;
2201         unsigned long launder_gp_seq;
2202 };
2203
2204 struct rcu_fwd {
2205         spinlock_t rcu_fwd_lock;
2206         struct rcu_fwd_cb *rcu_fwd_cb_head;
2207         struct rcu_fwd_cb **rcu_fwd_cb_tail;
2208         long n_launders_cb;
2209         unsigned long rcu_fwd_startat;
2210         struct rcu_launder_hist n_launders_hist[N_LAUNDERS_HIST];
2211         unsigned long rcu_launder_gp_seq_start;
2212         int rcu_fwd_id;
2213 };
2214
2215 static DEFINE_MUTEX(rcu_fwd_mutex);
2216 static struct rcu_fwd *rcu_fwds;
2217 static unsigned long rcu_fwd_seq;
2218 static atomic_long_t rcu_fwd_max_cbs;
2219 static bool rcu_fwd_emergency_stop;
2220
2221 static void rcu_torture_fwd_cb_hist(struct rcu_fwd *rfp)
2222 {
2223         unsigned long gps;
2224         unsigned long gps_old;
2225         int i;
2226         int j;
2227
2228         for (i = ARRAY_SIZE(rfp->n_launders_hist) - 1; i > 0; i--)
2229                 if (rfp->n_launders_hist[i].n_launders > 0)
2230                         break;
2231         pr_alert("%s: Callback-invocation histogram %d (duration %lu jiffies):",
2232                  __func__, rfp->rcu_fwd_id, jiffies - rfp->rcu_fwd_startat);
2233         gps_old = rfp->rcu_launder_gp_seq_start;
2234         for (j = 0; j <= i; j++) {
2235                 gps = rfp->n_launders_hist[j].launder_gp_seq;
2236                 pr_cont(" %ds/%d: %ld:%ld",
2237                         j + 1, FWD_CBS_HIST_DIV,
2238                         rfp->n_launders_hist[j].n_launders,
2239                         rcutorture_seq_diff(gps, gps_old));
2240                 gps_old = gps;
2241         }
2242         pr_cont("\n");
2243 }
2244
2245 /* Callback function for continuous-flood RCU callbacks. */
2246 static void rcu_torture_fwd_cb_cr(struct rcu_head *rhp)
2247 {
2248         unsigned long flags;
2249         int i;
2250         struct rcu_fwd_cb *rfcp = container_of(rhp, struct rcu_fwd_cb, rh);
2251         struct rcu_fwd_cb **rfcpp;
2252         struct rcu_fwd *rfp = rfcp->rfc_rfp;
2253
2254         rfcp->rfc_next = NULL;
2255         rfcp->rfc_gps++;
2256         spin_lock_irqsave(&rfp->rcu_fwd_lock, flags);
2257         rfcpp = rfp->rcu_fwd_cb_tail;
2258         rfp->rcu_fwd_cb_tail = &rfcp->rfc_next;
2259         WRITE_ONCE(*rfcpp, rfcp);
2260         WRITE_ONCE(rfp->n_launders_cb, rfp->n_launders_cb + 1);
2261         i = ((jiffies - rfp->rcu_fwd_startat) / (HZ / FWD_CBS_HIST_DIV));
2262         if (i >= ARRAY_SIZE(rfp->n_launders_hist))
2263                 i = ARRAY_SIZE(rfp->n_launders_hist) - 1;
2264         rfp->n_launders_hist[i].n_launders++;
2265         rfp->n_launders_hist[i].launder_gp_seq = cur_ops->get_gp_seq();
2266         spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2267 }
2268
2269 // Give the scheduler a chance, even on nohz_full CPUs.
2270 static void rcu_torture_fwd_prog_cond_resched(unsigned long iter)
2271 {
2272         if (IS_ENABLED(CONFIG_PREEMPTION) && IS_ENABLED(CONFIG_NO_HZ_FULL)) {
2273                 // Real call_rcu() floods hit userspace, so emulate that.
2274                 if (need_resched() || (iter & 0xfff))
2275                         schedule();
2276                 return;
2277         }
2278         // No userspace emulation: CB invocation throttles call_rcu()
2279         cond_resched();
2280 }
2281
2282 /*
2283  * Free all callbacks on the rcu_fwd_cb_head list, either because the
2284  * test is over or because we hit an OOM event.
2285  */
2286 static unsigned long rcu_torture_fwd_prog_cbfree(struct rcu_fwd *rfp)
2287 {
2288         unsigned long flags;
2289         unsigned long freed = 0;
2290         struct rcu_fwd_cb *rfcp;
2291
2292         for (;;) {
2293                 spin_lock_irqsave(&rfp->rcu_fwd_lock, flags);
2294                 rfcp = rfp->rcu_fwd_cb_head;
2295                 if (!rfcp) {
2296                         spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2297                         break;
2298                 }
2299                 rfp->rcu_fwd_cb_head = rfcp->rfc_next;
2300                 if (!rfp->rcu_fwd_cb_head)
2301                         rfp->rcu_fwd_cb_tail = &rfp->rcu_fwd_cb_head;
2302                 spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2303                 kfree(rfcp);
2304                 freed++;
2305                 rcu_torture_fwd_prog_cond_resched(freed);
2306                 if (tick_nohz_full_enabled()) {
2307                         local_irq_save(flags);
2308                         rcu_momentary_dyntick_idle();
2309                         local_irq_restore(flags);
2310                 }
2311         }
2312         return freed;
2313 }
2314
2315 /* Carry out need_resched()/cond_resched() forward-progress testing. */
2316 static void rcu_torture_fwd_prog_nr(struct rcu_fwd *rfp,
2317                                     int *tested, int *tested_tries)
2318 {
2319         unsigned long cver;
2320         unsigned long dur;
2321         struct fwd_cb_state fcs;
2322         unsigned long gps;
2323         int idx;
2324         int sd;
2325         int sd4;
2326         bool selfpropcb = false;
2327         unsigned long stopat;
2328         static DEFINE_TORTURE_RANDOM(trs);
2329
2330         pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2331         if (!cur_ops->sync)
2332                 return; // Cannot do need_resched() forward progress testing without ->sync.
2333         if (cur_ops->call && cur_ops->cb_barrier) {
2334                 init_rcu_head_on_stack(&fcs.rh);
2335                 selfpropcb = true;
2336         }
2337
2338         /* Tight loop containing cond_resched(). */
2339         atomic_inc(&rcu_fwd_cb_nodelay);
2340         cur_ops->sync(); /* Later readers see above write. */
2341         if  (selfpropcb) {
2342                 WRITE_ONCE(fcs.stop, 0);
2343                 cur_ops->call(&fcs.rh, rcu_torture_fwd_prog_cb);
2344         }
2345         cver = READ_ONCE(rcu_torture_current_version);
2346         gps = cur_ops->get_gp_seq();
2347         sd = cur_ops->stall_dur() + 1;
2348         sd4 = (sd + fwd_progress_div - 1) / fwd_progress_div;
2349         dur = sd4 + torture_random(&trs) % (sd - sd4);
2350         WRITE_ONCE(rfp->rcu_fwd_startat, jiffies);
2351         stopat = rfp->rcu_fwd_startat + dur;
2352         while (time_before(jiffies, stopat) &&
2353                !shutdown_time_arrived() &&
2354                !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2355                 idx = cur_ops->readlock();
2356                 udelay(10);
2357                 cur_ops->readunlock(idx);
2358                 if (!fwd_progress_need_resched || need_resched())
2359                         cond_resched();
2360         }
2361         (*tested_tries)++;
2362         if (!time_before(jiffies, stopat) &&
2363             !shutdown_time_arrived() &&
2364             !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2365                 (*tested)++;
2366                 cver = READ_ONCE(rcu_torture_current_version) - cver;
2367                 gps = rcutorture_seq_diff(cur_ops->get_gp_seq(), gps);
2368                 WARN_ON(!cver && gps < 2);
2369                 pr_alert("%s: %d Duration %ld cver %ld gps %ld\n", __func__,
2370                          rfp->rcu_fwd_id, dur, cver, gps);
2371         }
2372         if (selfpropcb) {
2373                 WRITE_ONCE(fcs.stop, 1);
2374                 cur_ops->sync(); /* Wait for running CB to complete. */
2375                 pr_alert("%s: Waiting for CBs: %pS() %d\n", __func__, cur_ops->cb_barrier, rfp->rcu_fwd_id);
2376                 cur_ops->cb_barrier(); /* Wait for queued callbacks. */
2377         }
2378
2379         if (selfpropcb) {
2380                 WARN_ON(READ_ONCE(fcs.stop) != 2);
2381                 destroy_rcu_head_on_stack(&fcs.rh);
2382         }
2383         schedule_timeout_uninterruptible(HZ / 10); /* Let kthreads recover. */
2384         atomic_dec(&rcu_fwd_cb_nodelay);
2385 }
2386
2387 /* Carry out call_rcu() forward-progress testing. */
2388 static void rcu_torture_fwd_prog_cr(struct rcu_fwd *rfp)
2389 {
2390         unsigned long cver;
2391         unsigned long flags;
2392         unsigned long gps;
2393         int i;
2394         long n_launders;
2395         long n_launders_cb_snap;
2396         long n_launders_sa;
2397         long n_max_cbs;
2398         long n_max_gps;
2399         struct rcu_fwd_cb *rfcp;
2400         struct rcu_fwd_cb *rfcpn;
2401         unsigned long stopat;
2402         unsigned long stoppedat;
2403
2404         pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2405         if (READ_ONCE(rcu_fwd_emergency_stop))
2406                 return; /* Get out of the way quickly, no GP wait! */
2407         if (!cur_ops->call)
2408                 return; /* Can't do call_rcu() fwd prog without ->call. */
2409
2410         /* Loop continuously posting RCU callbacks. */
2411         atomic_inc(&rcu_fwd_cb_nodelay);
2412         cur_ops->sync(); /* Later readers see above write. */
2413         WRITE_ONCE(rfp->rcu_fwd_startat, jiffies);
2414         stopat = rfp->rcu_fwd_startat + MAX_FWD_CB_JIFFIES;
2415         n_launders = 0;
2416         rfp->n_launders_cb = 0; // Hoist initialization for multi-kthread
2417         n_launders_sa = 0;
2418         n_max_cbs = 0;
2419         n_max_gps = 0;
2420         for (i = 0; i < ARRAY_SIZE(rfp->n_launders_hist); i++)
2421                 rfp->n_launders_hist[i].n_launders = 0;
2422         cver = READ_ONCE(rcu_torture_current_version);
2423         gps = cur_ops->get_gp_seq();
2424         rfp->rcu_launder_gp_seq_start = gps;
2425         tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2426         while (time_before(jiffies, stopat) &&
2427                !shutdown_time_arrived() &&
2428                !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2429                 rfcp = READ_ONCE(rfp->rcu_fwd_cb_head);
2430                 rfcpn = NULL;
2431                 if (rfcp)
2432                         rfcpn = READ_ONCE(rfcp->rfc_next);
2433                 if (rfcpn) {
2434                         if (rfcp->rfc_gps >= MIN_FWD_CB_LAUNDERS &&
2435                             ++n_max_gps >= MIN_FWD_CBS_LAUNDERED)
2436                                 break;
2437                         rfp->rcu_fwd_cb_head = rfcpn;
2438                         n_launders++;
2439                         n_launders_sa++;
2440                 } else if (!cur_ops->cbflood_max || cur_ops->cbflood_max > n_max_cbs) {
2441                         rfcp = kmalloc(sizeof(*rfcp), GFP_KERNEL);
2442                         if (WARN_ON_ONCE(!rfcp)) {
2443                                 schedule_timeout_interruptible(1);
2444                                 continue;
2445                         }
2446                         n_max_cbs++;
2447                         n_launders_sa = 0;
2448                         rfcp->rfc_gps = 0;
2449                         rfcp->rfc_rfp = rfp;
2450                 } else {
2451                         rfcp = NULL;
2452                 }
2453                 if (rfcp)
2454                         cur_ops->call(&rfcp->rh, rcu_torture_fwd_cb_cr);
2455                 rcu_torture_fwd_prog_cond_resched(n_launders + n_max_cbs);
2456                 if (tick_nohz_full_enabled()) {
2457                         local_irq_save(flags);
2458                         rcu_momentary_dyntick_idle();
2459                         local_irq_restore(flags);
2460                 }
2461         }
2462         stoppedat = jiffies;
2463         n_launders_cb_snap = READ_ONCE(rfp->n_launders_cb);
2464         cver = READ_ONCE(rcu_torture_current_version) - cver;
2465         gps = rcutorture_seq_diff(cur_ops->get_gp_seq(), gps);
2466         pr_alert("%s: Waiting for CBs: %pS() %d\n", __func__, cur_ops->cb_barrier, rfp->rcu_fwd_id);
2467         cur_ops->cb_barrier(); /* Wait for callbacks to be invoked. */
2468         (void)rcu_torture_fwd_prog_cbfree(rfp);
2469
2470         if (!torture_must_stop() && !READ_ONCE(rcu_fwd_emergency_stop) &&
2471             !shutdown_time_arrived()) {
2472                 WARN_ON(n_max_gps < MIN_FWD_CBS_LAUNDERED);
2473                 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",
2474                          __func__,
2475                          stoppedat - rfp->rcu_fwd_startat, jiffies - stoppedat,
2476                          n_launders + n_max_cbs - n_launders_cb_snap,
2477                          n_launders, n_launders_sa,
2478                          n_max_gps, n_max_cbs, cver, gps);
2479                 atomic_long_add(n_max_cbs, &rcu_fwd_max_cbs);
2480                 mutex_lock(&rcu_fwd_mutex); // Serialize histograms.
2481                 rcu_torture_fwd_cb_hist(rfp);
2482                 mutex_unlock(&rcu_fwd_mutex);
2483         }
2484         schedule_timeout_uninterruptible(HZ); /* Let CBs drain. */
2485         tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2486         atomic_dec(&rcu_fwd_cb_nodelay);
2487 }
2488
2489
2490 /*
2491  * OOM notifier, but this only prints diagnostic information for the
2492  * current forward-progress test.
2493  */
2494 static int rcutorture_oom_notify(struct notifier_block *self,
2495                                  unsigned long notused, void *nfreed)
2496 {
2497         int i;
2498         long ncbs;
2499         struct rcu_fwd *rfp;
2500
2501         mutex_lock(&rcu_fwd_mutex);
2502         rfp = rcu_fwds;
2503         if (!rfp) {
2504                 mutex_unlock(&rcu_fwd_mutex);
2505                 return NOTIFY_OK;
2506         }
2507         WARN(1, "%s invoked upon OOM during forward-progress testing.\n",
2508              __func__);
2509         for (i = 0; i < fwd_progress; i++) {
2510                 rcu_torture_fwd_cb_hist(&rfp[i]);
2511                 rcu_fwd_progress_check(1 + (jiffies - READ_ONCE(rfp[i].rcu_fwd_startat)) / 2);
2512         }
2513         WRITE_ONCE(rcu_fwd_emergency_stop, true);
2514         smp_mb(); /* Emergency stop before free and wait to avoid hangs. */
2515         ncbs = 0;
2516         for (i = 0; i < fwd_progress; i++)
2517                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2518         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2519         rcu_barrier();
2520         ncbs = 0;
2521         for (i = 0; i < fwd_progress; i++)
2522                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2523         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2524         rcu_barrier();
2525         ncbs = 0;
2526         for (i = 0; i < fwd_progress; i++)
2527                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2528         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2529         smp_mb(); /* Frees before return to avoid redoing OOM. */
2530         (*(unsigned long *)nfreed)++; /* Forward progress CBs freed! */
2531         pr_info("%s returning after OOM processing.\n", __func__);
2532         mutex_unlock(&rcu_fwd_mutex);
2533         return NOTIFY_OK;
2534 }
2535
2536 static struct notifier_block rcutorture_oom_nb = {
2537         .notifier_call = rcutorture_oom_notify
2538 };
2539
2540 /* Carry out grace-period forward-progress testing. */
2541 static int rcu_torture_fwd_prog(void *args)
2542 {
2543         bool firsttime = true;
2544         long max_cbs;
2545         int oldnice = task_nice(current);
2546         unsigned long oldseq = READ_ONCE(rcu_fwd_seq);
2547         struct rcu_fwd *rfp = args;
2548         int tested = 0;
2549         int tested_tries = 0;
2550
2551         VERBOSE_TOROUT_STRING("rcu_torture_fwd_progress task started");
2552         rcu_bind_current_to_nocb();
2553         if (!IS_ENABLED(CONFIG_SMP) || !IS_ENABLED(CONFIG_RCU_BOOST))
2554                 set_user_nice(current, MAX_NICE);
2555         do {
2556                 if (!rfp->rcu_fwd_id) {
2557                         schedule_timeout_interruptible(fwd_progress_holdoff * HZ);
2558                         WRITE_ONCE(rcu_fwd_emergency_stop, false);
2559                         if (!firsttime) {
2560                                 max_cbs = atomic_long_xchg(&rcu_fwd_max_cbs, 0);
2561                                 pr_alert("%s n_max_cbs: %ld\n", __func__, max_cbs);
2562                         }
2563                         firsttime = false;
2564                         WRITE_ONCE(rcu_fwd_seq, rcu_fwd_seq + 1);
2565                 } else {
2566                         while (READ_ONCE(rcu_fwd_seq) == oldseq && !torture_must_stop())
2567                                 schedule_timeout_interruptible(1);
2568                         oldseq = READ_ONCE(rcu_fwd_seq);
2569                 }
2570                 pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2571                 if (rcu_inkernel_boot_has_ended() && torture_num_online_cpus() > rfp->rcu_fwd_id)
2572                         rcu_torture_fwd_prog_cr(rfp);
2573                 if ((cur_ops->stall_dur && cur_ops->stall_dur() > 0) &&
2574                     (!IS_ENABLED(CONFIG_TINY_RCU) ||
2575                      (rcu_inkernel_boot_has_ended() &&
2576                       torture_num_online_cpus() > rfp->rcu_fwd_id)))
2577                         rcu_torture_fwd_prog_nr(rfp, &tested, &tested_tries);
2578
2579                 /* Avoid slow periods, better to test when busy. */
2580                 if (stutter_wait("rcu_torture_fwd_prog"))
2581                         sched_set_normal(current, oldnice);
2582         } while (!torture_must_stop());
2583         /* Short runs might not contain a valid forward-progress attempt. */
2584         if (!rfp->rcu_fwd_id) {
2585                 WARN_ON(!tested && tested_tries >= 5);
2586                 pr_alert("%s: tested %d tested_tries %d\n", __func__, tested, tested_tries);
2587         }
2588         torture_kthread_stopping("rcu_torture_fwd_prog");
2589         return 0;
2590 }
2591
2592 /* If forward-progress checking is requested and feasible, spawn the thread. */
2593 static int __init rcu_torture_fwd_prog_init(void)
2594 {
2595         int i;
2596         int ret = 0;
2597         struct rcu_fwd *rfp;
2598
2599         if (!fwd_progress)
2600                 return 0; /* Not requested, so don't do it. */
2601         if (fwd_progress >= nr_cpu_ids) {
2602                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Limiting fwd_progress to # CPUs.\n");
2603                 fwd_progress = nr_cpu_ids;
2604         } else if (fwd_progress < 0) {
2605                 fwd_progress = nr_cpu_ids;
2606         }
2607         if ((!cur_ops->sync && !cur_ops->call) ||
2608             (!cur_ops->cbflood_max && (!cur_ops->stall_dur || cur_ops->stall_dur() <= 0)) ||
2609             cur_ops == &rcu_busted_ops) {
2610                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Disabled, unsupported by RCU flavor under test");
2611                 fwd_progress = 0;
2612                 return 0;
2613         }
2614         if (stall_cpu > 0) {
2615                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Disabled, conflicts with CPU-stall testing");
2616                 fwd_progress = 0;
2617                 if (IS_MODULE(CONFIG_RCU_TORTURE_TEST))
2618                         return -EINVAL; /* In module, can fail back to user. */
2619                 WARN_ON(1); /* Make sure rcutorture notices conflict. */
2620                 return 0;
2621         }
2622         if (fwd_progress_holdoff <= 0)
2623                 fwd_progress_holdoff = 1;
2624         if (fwd_progress_div <= 0)
2625                 fwd_progress_div = 4;
2626         rfp = kcalloc(fwd_progress, sizeof(*rfp), GFP_KERNEL);
2627         fwd_prog_tasks = kcalloc(fwd_progress, sizeof(*fwd_prog_tasks), GFP_KERNEL);
2628         if (!rfp || !fwd_prog_tasks) {
2629                 kfree(rfp);
2630                 kfree(fwd_prog_tasks);
2631                 fwd_prog_tasks = NULL;
2632                 fwd_progress = 0;
2633                 return -ENOMEM;
2634         }
2635         for (i = 0; i < fwd_progress; i++) {
2636                 spin_lock_init(&rfp[i].rcu_fwd_lock);
2637                 rfp[i].rcu_fwd_cb_tail = &rfp[i].rcu_fwd_cb_head;
2638                 rfp[i].rcu_fwd_id = i;
2639         }
2640         mutex_lock(&rcu_fwd_mutex);
2641         rcu_fwds = rfp;
2642         mutex_unlock(&rcu_fwd_mutex);
2643         register_oom_notifier(&rcutorture_oom_nb);
2644         for (i = 0; i < fwd_progress; i++) {
2645                 ret = torture_create_kthread(rcu_torture_fwd_prog, &rcu_fwds[i], fwd_prog_tasks[i]);
2646                 if (ret) {
2647                         fwd_progress = i;
2648                         return ret;
2649                 }
2650         }
2651         return 0;
2652 }
2653
2654 static void rcu_torture_fwd_prog_cleanup(void)
2655 {
2656         int i;
2657         struct rcu_fwd *rfp;
2658
2659         if (!rcu_fwds || !fwd_prog_tasks)
2660                 return;
2661         for (i = 0; i < fwd_progress; i++)
2662                 torture_stop_kthread(rcu_torture_fwd_prog, fwd_prog_tasks[i]);
2663         unregister_oom_notifier(&rcutorture_oom_nb);
2664         mutex_lock(&rcu_fwd_mutex);
2665         rfp = rcu_fwds;
2666         rcu_fwds = NULL;
2667         mutex_unlock(&rcu_fwd_mutex);
2668         kfree(rfp);
2669         kfree(fwd_prog_tasks);
2670         fwd_prog_tasks = NULL;
2671 }
2672
2673 /* Callback function for RCU barrier testing. */
2674 static void rcu_torture_barrier_cbf(struct rcu_head *rcu)
2675 {
2676         atomic_inc(&barrier_cbs_invoked);
2677 }
2678
2679 /* IPI handler to get callback posted on desired CPU, if online. */
2680 static void rcu_torture_barrier1cb(void *rcu_void)
2681 {
2682         struct rcu_head *rhp = rcu_void;
2683
2684         cur_ops->call(rhp, rcu_torture_barrier_cbf);
2685 }
2686
2687 /* kthread function to register callbacks used to test RCU barriers. */
2688 static int rcu_torture_barrier_cbs(void *arg)
2689 {
2690         long myid = (long)arg;
2691         bool lastphase = false;
2692         bool newphase;
2693         struct rcu_head rcu;
2694
2695         init_rcu_head_on_stack(&rcu);
2696         VERBOSE_TOROUT_STRING("rcu_torture_barrier_cbs task started");
2697         set_user_nice(current, MAX_NICE);
2698         do {
2699                 wait_event(barrier_cbs_wq[myid],
2700                            (newphase =
2701                             smp_load_acquire(&barrier_phase)) != lastphase ||
2702                            torture_must_stop());
2703                 lastphase = newphase;
2704                 if (torture_must_stop())
2705                         break;
2706                 /*
2707                  * The above smp_load_acquire() ensures barrier_phase load
2708                  * is ordered before the following ->call().
2709                  */
2710                 if (smp_call_function_single(myid, rcu_torture_barrier1cb,
2711                                              &rcu, 1)) {
2712                         // IPI failed, so use direct call from current CPU.
2713                         cur_ops->call(&rcu, rcu_torture_barrier_cbf);
2714                 }
2715                 if (atomic_dec_and_test(&barrier_cbs_count))
2716                         wake_up(&barrier_wq);
2717         } while (!torture_must_stop());
2718         if (cur_ops->cb_barrier != NULL)
2719                 cur_ops->cb_barrier();
2720         destroy_rcu_head_on_stack(&rcu);
2721         torture_kthread_stopping("rcu_torture_barrier_cbs");
2722         return 0;
2723 }
2724
2725 /* kthread function to drive and coordinate RCU barrier testing. */
2726 static int rcu_torture_barrier(void *arg)
2727 {
2728         int i;
2729
2730         VERBOSE_TOROUT_STRING("rcu_torture_barrier task starting");
2731         do {
2732                 atomic_set(&barrier_cbs_invoked, 0);
2733                 atomic_set(&barrier_cbs_count, n_barrier_cbs);
2734                 /* Ensure barrier_phase ordered after prior assignments. */
2735                 smp_store_release(&barrier_phase, !barrier_phase);
2736                 for (i = 0; i < n_barrier_cbs; i++)
2737                         wake_up(&barrier_cbs_wq[i]);
2738                 wait_event(barrier_wq,
2739                            atomic_read(&barrier_cbs_count) == 0 ||
2740                            torture_must_stop());
2741                 if (torture_must_stop())
2742                         break;
2743                 n_barrier_attempts++;
2744                 cur_ops->cb_barrier(); /* Implies smp_mb() for wait_event(). */
2745                 if (atomic_read(&barrier_cbs_invoked) != n_barrier_cbs) {
2746                         n_rcu_torture_barrier_error++;
2747                         pr_err("barrier_cbs_invoked = %d, n_barrier_cbs = %d\n",
2748                                atomic_read(&barrier_cbs_invoked),
2749                                n_barrier_cbs);
2750                         WARN_ON(1);
2751                         // Wait manually for the remaining callbacks
2752                         i = 0;
2753                         do {
2754                                 if (WARN_ON(i++ > HZ))
2755                                         i = INT_MIN;
2756                                 schedule_timeout_interruptible(1);
2757                                 cur_ops->cb_barrier();
2758                         } while (atomic_read(&barrier_cbs_invoked) !=
2759                                  n_barrier_cbs &&
2760                                  !torture_must_stop());
2761                         smp_mb(); // Can't trust ordering if broken.
2762                         if (!torture_must_stop())
2763                                 pr_err("Recovered: barrier_cbs_invoked = %d\n",
2764                                        atomic_read(&barrier_cbs_invoked));
2765                 } else {
2766                         n_barrier_successes++;
2767                 }
2768                 schedule_timeout_interruptible(HZ / 10);
2769         } while (!torture_must_stop());
2770         torture_kthread_stopping("rcu_torture_barrier");
2771         return 0;
2772 }
2773
2774 /* Initialize RCU barrier testing. */
2775 static int rcu_torture_barrier_init(void)
2776 {
2777         int i;
2778         int ret;
2779
2780         if (n_barrier_cbs <= 0)
2781                 return 0;
2782         if (cur_ops->call == NULL || cur_ops->cb_barrier == NULL) {
2783                 pr_alert("%s" TORTURE_FLAG
2784                          " Call or barrier ops missing for %s,\n",
2785                          torture_type, cur_ops->name);
2786                 pr_alert("%s" TORTURE_FLAG
2787                          " RCU barrier testing omitted from run.\n",
2788                          torture_type);
2789                 return 0;
2790         }
2791         atomic_set(&barrier_cbs_count, 0);
2792         atomic_set(&barrier_cbs_invoked, 0);
2793         barrier_cbs_tasks =
2794                 kcalloc(n_barrier_cbs, sizeof(barrier_cbs_tasks[0]),
2795                         GFP_KERNEL);
2796         barrier_cbs_wq =
2797                 kcalloc(n_barrier_cbs, sizeof(barrier_cbs_wq[0]), GFP_KERNEL);
2798         if (barrier_cbs_tasks == NULL || !barrier_cbs_wq)
2799                 return -ENOMEM;
2800         for (i = 0; i < n_barrier_cbs; i++) {
2801                 init_waitqueue_head(&barrier_cbs_wq[i]);
2802                 ret = torture_create_kthread(rcu_torture_barrier_cbs,
2803                                              (void *)(long)i,
2804                                              barrier_cbs_tasks[i]);
2805                 if (ret)
2806                         return ret;
2807         }
2808         return torture_create_kthread(rcu_torture_barrier, NULL, barrier_task);
2809 }
2810
2811 /* Clean up after RCU barrier testing. */
2812 static void rcu_torture_barrier_cleanup(void)
2813 {
2814         int i;
2815
2816         torture_stop_kthread(rcu_torture_barrier, barrier_task);
2817         if (barrier_cbs_tasks != NULL) {
2818                 for (i = 0; i < n_barrier_cbs; i++)
2819                         torture_stop_kthread(rcu_torture_barrier_cbs,
2820                                              barrier_cbs_tasks[i]);
2821                 kfree(barrier_cbs_tasks);
2822                 barrier_cbs_tasks = NULL;
2823         }
2824         if (barrier_cbs_wq != NULL) {
2825                 kfree(barrier_cbs_wq);
2826                 barrier_cbs_wq = NULL;
2827         }
2828 }
2829
2830 static bool rcu_torture_can_boost(void)
2831 {
2832         static int boost_warn_once;
2833         int prio;
2834
2835         if (!(test_boost == 1 && cur_ops->can_boost) && test_boost != 2)
2836                 return false;
2837         if (!cur_ops->start_gp_poll || !cur_ops->poll_gp_state)
2838                 return false;
2839
2840         prio = rcu_get_gp_kthreads_prio();
2841         if (!prio)
2842                 return false;
2843
2844         if (prio < 2) {
2845                 if (boost_warn_once == 1)
2846                         return false;
2847
2848                 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);
2849                 boost_warn_once = 1;
2850                 return false;
2851         }
2852
2853         return true;
2854 }
2855
2856 static bool read_exit_child_stop;
2857 static bool read_exit_child_stopped;
2858 static wait_queue_head_t read_exit_wq;
2859
2860 // Child kthread which just does an rcutorture reader and exits.
2861 static int rcu_torture_read_exit_child(void *trsp_in)
2862 {
2863         struct torture_random_state *trsp = trsp_in;
2864
2865         set_user_nice(current, MAX_NICE);
2866         // Minimize time between reading and exiting.
2867         while (!kthread_should_stop())
2868                 schedule_timeout_uninterruptible(1);
2869         (void)rcu_torture_one_read(trsp, -1);
2870         return 0;
2871 }
2872
2873 // Parent kthread which creates and destroys read-exit child kthreads.
2874 static int rcu_torture_read_exit(void *unused)
2875 {
2876         int count = 0;
2877         bool errexit = false;
2878         int i;
2879         struct task_struct *tsp;
2880         DEFINE_TORTURE_RANDOM(trs);
2881
2882         // Allocate and initialize.
2883         set_user_nice(current, MAX_NICE);
2884         VERBOSE_TOROUT_STRING("rcu_torture_read_exit: Start of test");
2885
2886         // Each pass through this loop does one read-exit episode.
2887         do {
2888                 if (++count > read_exit_burst) {
2889                         VERBOSE_TOROUT_STRING("rcu_torture_read_exit: End of episode");
2890                         rcu_barrier(); // Wait for task_struct free, avoid OOM.
2891                         for (i = 0; i < read_exit_delay; i++) {
2892                                 schedule_timeout_uninterruptible(HZ);
2893                                 if (READ_ONCE(read_exit_child_stop))
2894                                         break;
2895                         }
2896                         if (!READ_ONCE(read_exit_child_stop))
2897                                 VERBOSE_TOROUT_STRING("rcu_torture_read_exit: Start of episode");
2898                         count = 0;
2899                 }
2900                 if (READ_ONCE(read_exit_child_stop))
2901                         break;
2902                 // Spawn child.
2903                 tsp = kthread_run(rcu_torture_read_exit_child,
2904                                      &trs, "%s",
2905                                      "rcu_torture_read_exit_child");
2906                 if (IS_ERR(tsp)) {
2907                         TOROUT_ERRSTRING("out of memory");
2908                         errexit = true;
2909                         tsp = NULL;
2910                         break;
2911                 }
2912                 cond_resched();
2913                 kthread_stop(tsp);
2914                 n_read_exits ++;
2915                 stutter_wait("rcu_torture_read_exit");
2916         } while (!errexit && !READ_ONCE(read_exit_child_stop));
2917
2918         // Clean up and exit.
2919         smp_store_release(&read_exit_child_stopped, true); // After reaping.
2920         smp_mb(); // Store before wakeup.
2921         wake_up(&read_exit_wq);
2922         while (!torture_must_stop())
2923                 schedule_timeout_uninterruptible(1);
2924         torture_kthread_stopping("rcu_torture_read_exit");
2925         return 0;
2926 }
2927
2928 static int rcu_torture_read_exit_init(void)
2929 {
2930         if (read_exit_burst <= 0)
2931                 return 0;
2932         init_waitqueue_head(&read_exit_wq);
2933         read_exit_child_stop = false;
2934         read_exit_child_stopped = false;
2935         return torture_create_kthread(rcu_torture_read_exit, NULL,
2936                                       read_exit_task);
2937 }
2938
2939 static void rcu_torture_read_exit_cleanup(void)
2940 {
2941         if (!read_exit_task)
2942                 return;
2943         WRITE_ONCE(read_exit_child_stop, true);
2944         smp_mb(); // Above write before wait.
2945         wait_event(read_exit_wq, smp_load_acquire(&read_exit_child_stopped));
2946         torture_stop_kthread(rcutorture_read_exit, read_exit_task);
2947 }
2948
2949 static enum cpuhp_state rcutor_hp;
2950
2951 static void
2952 rcu_torture_cleanup(void)
2953 {
2954         int firsttime;
2955         int flags = 0;
2956         unsigned long gp_seq = 0;
2957         int i;
2958
2959         if (torture_cleanup_begin()) {
2960                 if (cur_ops->cb_barrier != NULL) {
2961                         pr_info("%s: Invoking %pS().\n", __func__, cur_ops->cb_barrier);
2962                         cur_ops->cb_barrier();
2963                 }
2964                 rcu_gp_slow_unregister(NULL);
2965                 return;
2966         }
2967         if (!cur_ops) {
2968                 torture_cleanup_end();
2969                 rcu_gp_slow_unregister(NULL);
2970                 return;
2971         }
2972
2973         if (cur_ops->gp_kthread_dbg)
2974                 cur_ops->gp_kthread_dbg();
2975         rcu_torture_read_exit_cleanup();
2976         rcu_torture_barrier_cleanup();
2977         rcu_torture_fwd_prog_cleanup();
2978         torture_stop_kthread(rcu_torture_stall, stall_task);
2979         torture_stop_kthread(rcu_torture_writer, writer_task);
2980
2981         if (nocb_tasks) {
2982                 for (i = 0; i < nrealnocbers; i++)
2983                         torture_stop_kthread(rcu_nocb_toggle, nocb_tasks[i]);
2984                 kfree(nocb_tasks);
2985                 nocb_tasks = NULL;
2986         }
2987
2988         if (reader_tasks) {
2989                 for (i = 0; i < nrealreaders; i++)
2990                         torture_stop_kthread(rcu_torture_reader,
2991                                              reader_tasks[i]);
2992                 kfree(reader_tasks);
2993                 reader_tasks = NULL;
2994         }
2995         kfree(rcu_torture_reader_mbchk);
2996         rcu_torture_reader_mbchk = NULL;
2997
2998         if (fakewriter_tasks) {
2999                 for (i = 0; i < nfakewriters; i++)
3000                         torture_stop_kthread(rcu_torture_fakewriter,
3001                                              fakewriter_tasks[i]);
3002                 kfree(fakewriter_tasks);
3003                 fakewriter_tasks = NULL;
3004         }
3005
3006         rcutorture_get_gp_data(cur_ops->ttype, &flags, &gp_seq);
3007         srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp, &flags, &gp_seq);
3008         pr_alert("%s:  End-test grace-period state: g%ld f%#x total-gps=%ld\n",
3009                  cur_ops->name, (long)gp_seq, flags,
3010                  rcutorture_seq_diff(gp_seq, start_gp_seq));
3011         torture_stop_kthread(rcu_torture_stats, stats_task);
3012         torture_stop_kthread(rcu_torture_fqs, fqs_task);
3013         if (rcu_torture_can_boost() && rcutor_hp >= 0)
3014                 cpuhp_remove_state(rcutor_hp);
3015
3016         /*
3017          * Wait for all RCU callbacks to fire, then do torture-type-specific
3018          * cleanup operations.
3019          */
3020         if (cur_ops->cb_barrier != NULL) {
3021                 pr_info("%s: Invoking %pS().\n", __func__, cur_ops->cb_barrier);
3022                 cur_ops->cb_barrier();
3023         }
3024         if (cur_ops->cleanup != NULL)
3025                 cur_ops->cleanup();
3026
3027         rcu_torture_mem_dump_obj();
3028
3029         rcu_torture_stats_print();  /* -After- the stats thread is stopped! */
3030
3031         if (err_segs_recorded) {
3032                 pr_alert("Failure/close-call rcutorture reader segments:\n");
3033                 if (rt_read_nsegs == 0)
3034                         pr_alert("\t: No segments recorded!!!\n");
3035                 firsttime = 1;
3036                 for (i = 0; i < rt_read_nsegs; i++) {
3037                         pr_alert("\t%d: %#x ", i, err_segs[i].rt_readstate);
3038                         if (err_segs[i].rt_delay_jiffies != 0) {
3039                                 pr_cont("%s%ldjiffies", firsttime ? "" : "+",
3040                                         err_segs[i].rt_delay_jiffies);
3041                                 firsttime = 0;
3042                         }
3043                         if (err_segs[i].rt_delay_ms != 0) {
3044                                 pr_cont("%s%ldms", firsttime ? "" : "+",
3045                                         err_segs[i].rt_delay_ms);
3046                                 firsttime = 0;
3047                         }
3048                         if (err_segs[i].rt_delay_us != 0) {
3049                                 pr_cont("%s%ldus", firsttime ? "" : "+",
3050                                         err_segs[i].rt_delay_us);
3051                                 firsttime = 0;
3052                         }
3053                         pr_cont("%s\n",
3054                                 err_segs[i].rt_preempted ? "preempted" : "");
3055
3056                 }
3057         }
3058         if (atomic_read(&n_rcu_torture_error) || n_rcu_torture_barrier_error)
3059                 rcu_torture_print_module_parms(cur_ops, "End of test: FAILURE");
3060         else if (torture_onoff_failures())
3061                 rcu_torture_print_module_parms(cur_ops,
3062                                                "End of test: RCU_HOTPLUG");
3063         else
3064                 rcu_torture_print_module_parms(cur_ops, "End of test: SUCCESS");
3065         torture_cleanup_end();
3066         rcu_gp_slow_unregister(&rcu_fwd_cb_nodelay);
3067 }
3068
3069 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3070 static void rcu_torture_leak_cb(struct rcu_head *rhp)
3071 {
3072 }
3073
3074 static void rcu_torture_err_cb(struct rcu_head *rhp)
3075 {
3076         /*
3077          * This -might- happen due to race conditions, but is unlikely.
3078          * The scenario that leads to this happening is that the
3079          * first of the pair of duplicate callbacks is queued,
3080          * someone else starts a grace period that includes that
3081          * callback, then the second of the pair must wait for the
3082          * next grace period.  Unlikely, but can happen.  If it
3083          * does happen, the debug-objects subsystem won't have splatted.
3084          */
3085         pr_alert("%s: duplicated callback was invoked.\n", KBUILD_MODNAME);
3086 }
3087 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3088
3089 /*
3090  * Verify that double-free causes debug-objects to complain, but only
3091  * if CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.  Otherwise, say that the test
3092  * cannot be carried out.
3093  */
3094 static void rcu_test_debug_objects(void)
3095 {
3096 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3097         struct rcu_head rh1;
3098         struct rcu_head rh2;
3099         struct rcu_head *rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
3100
3101         init_rcu_head_on_stack(&rh1);
3102         init_rcu_head_on_stack(&rh2);
3103         pr_alert("%s: WARN: Duplicate call_rcu() test starting.\n", KBUILD_MODNAME);
3104
3105         /* Try to queue the rh2 pair of callbacks for the same grace period. */
3106         preempt_disable(); /* Prevent preemption from interrupting test. */
3107         rcu_read_lock(); /* Make it impossible to finish a grace period. */
3108         call_rcu(&rh1, rcu_torture_leak_cb); /* Start grace period. */
3109         local_irq_disable(); /* Make it harder to start a new grace period. */
3110         call_rcu(&rh2, rcu_torture_leak_cb);
3111         call_rcu(&rh2, rcu_torture_err_cb); /* Duplicate callback. */
3112         if (rhp) {
3113                 call_rcu(rhp, rcu_torture_leak_cb);
3114                 call_rcu(rhp, rcu_torture_err_cb); /* Another duplicate callback. */
3115         }
3116         local_irq_enable();
3117         rcu_read_unlock();
3118         preempt_enable();
3119
3120         /* Wait for them all to get done so we can safely return. */
3121         rcu_barrier();
3122         pr_alert("%s: WARN: Duplicate call_rcu() test complete.\n", KBUILD_MODNAME);
3123         destroy_rcu_head_on_stack(&rh1);
3124         destroy_rcu_head_on_stack(&rh2);
3125 #else /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3126         pr_alert("%s: !CONFIG_DEBUG_OBJECTS_RCU_HEAD, not testing duplicate call_rcu()\n", KBUILD_MODNAME);
3127 #endif /* #else #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3128 }
3129
3130 static void rcutorture_sync(void)
3131 {
3132         static unsigned long n;
3133
3134         if (cur_ops->sync && !(++n & 0xfff))
3135                 cur_ops->sync();
3136 }
3137
3138 static int __init
3139 rcu_torture_init(void)
3140 {
3141         long i;
3142         int cpu;
3143         int firsterr = 0;
3144         int flags = 0;
3145         unsigned long gp_seq = 0;
3146         static struct rcu_torture_ops *torture_ops[] = {
3147                 &rcu_ops, &rcu_busted_ops, &srcu_ops, &srcud_ops, &busted_srcud_ops,
3148                 TASKS_OPS TASKS_RUDE_OPS TASKS_TRACING_OPS
3149                 &trivial_ops,
3150         };
3151
3152         if (!torture_init_begin(torture_type, verbose))
3153                 return -EBUSY;
3154
3155         /* Process args and tell the world that the torturer is on the job. */
3156         for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
3157                 cur_ops = torture_ops[i];
3158                 if (strcmp(torture_type, cur_ops->name) == 0)
3159                         break;
3160         }
3161         if (i == ARRAY_SIZE(torture_ops)) {
3162                 pr_alert("rcu-torture: invalid torture type: \"%s\"\n",
3163                          torture_type);
3164                 pr_alert("rcu-torture types:");
3165                 for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
3166                         pr_cont(" %s", torture_ops[i]->name);
3167                 pr_cont("\n");
3168                 firsterr = -EINVAL;
3169                 cur_ops = NULL;
3170                 goto unwind;
3171         }
3172         if (cur_ops->fqs == NULL && fqs_duration != 0) {
3173                 pr_alert("rcu-torture: ->fqs NULL and non-zero fqs_duration, fqs disabled.\n");
3174                 fqs_duration = 0;
3175         }
3176         if (cur_ops->init)
3177                 cur_ops->init();
3178
3179         if (nreaders >= 0) {
3180                 nrealreaders = nreaders;
3181         } else {
3182                 nrealreaders = num_online_cpus() - 2 - nreaders;
3183                 if (nrealreaders <= 0)
3184                         nrealreaders = 1;
3185         }
3186         rcu_torture_print_module_parms(cur_ops, "Start of test");
3187         rcutorture_get_gp_data(cur_ops->ttype, &flags, &gp_seq);
3188         srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp, &flags, &gp_seq);
3189         start_gp_seq = gp_seq;
3190         pr_alert("%s:  Start-test grace-period state: g%ld f%#x\n",
3191                  cur_ops->name, (long)gp_seq, flags);
3192
3193         /* Set up the freelist. */
3194
3195         INIT_LIST_HEAD(&rcu_torture_freelist);
3196         for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++) {
3197                 rcu_tortures[i].rtort_mbtest = 0;
3198                 list_add_tail(&rcu_tortures[i].rtort_free,
3199                               &rcu_torture_freelist);
3200         }
3201
3202         /* Initialize the statistics so that each run gets its own numbers. */
3203
3204         rcu_torture_current = NULL;
3205         rcu_torture_current_version = 0;
3206         atomic_set(&n_rcu_torture_alloc, 0);
3207         atomic_set(&n_rcu_torture_alloc_fail, 0);
3208         atomic_set(&n_rcu_torture_free, 0);
3209         atomic_set(&n_rcu_torture_mberror, 0);
3210         atomic_set(&n_rcu_torture_mbchk_fail, 0);
3211         atomic_set(&n_rcu_torture_mbchk_tries, 0);
3212         atomic_set(&n_rcu_torture_error, 0);
3213         n_rcu_torture_barrier_error = 0;
3214         n_rcu_torture_boost_ktrerror = 0;
3215         n_rcu_torture_boost_rterror = 0;
3216         n_rcu_torture_boost_failure = 0;
3217         n_rcu_torture_boosts = 0;
3218         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
3219                 atomic_set(&rcu_torture_wcount[i], 0);
3220         for_each_possible_cpu(cpu) {
3221                 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
3222                         per_cpu(rcu_torture_count, cpu)[i] = 0;
3223                         per_cpu(rcu_torture_batch, cpu)[i] = 0;
3224                 }
3225         }
3226         err_segs_recorded = 0;
3227         rt_read_nsegs = 0;
3228
3229         /* Start up the kthreads. */
3230
3231         rcu_torture_write_types();
3232         firsterr = torture_create_kthread(rcu_torture_writer, NULL,
3233                                           writer_task);
3234         if (torture_init_error(firsterr))
3235                 goto unwind;
3236         if (nfakewriters > 0) {
3237                 fakewriter_tasks = kcalloc(nfakewriters,
3238                                            sizeof(fakewriter_tasks[0]),
3239                                            GFP_KERNEL);
3240                 if (fakewriter_tasks == NULL) {
3241                         TOROUT_ERRSTRING("out of memory");
3242                         firsterr = -ENOMEM;
3243                         goto unwind;
3244                 }
3245         }
3246         for (i = 0; i < nfakewriters; i++) {
3247                 firsterr = torture_create_kthread(rcu_torture_fakewriter,
3248                                                   NULL, fakewriter_tasks[i]);
3249                 if (torture_init_error(firsterr))
3250                         goto unwind;
3251         }
3252         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
3253                                GFP_KERNEL);
3254         rcu_torture_reader_mbchk = kcalloc(nrealreaders, sizeof(*rcu_torture_reader_mbchk),
3255                                            GFP_KERNEL);
3256         if (!reader_tasks || !rcu_torture_reader_mbchk) {
3257                 TOROUT_ERRSTRING("out of memory");
3258                 firsterr = -ENOMEM;
3259                 goto unwind;
3260         }
3261         for (i = 0; i < nrealreaders; i++) {
3262                 rcu_torture_reader_mbchk[i].rtc_chkrdr = -1;
3263                 firsterr = torture_create_kthread(rcu_torture_reader, (void *)i,
3264                                                   reader_tasks[i]);
3265                 if (torture_init_error(firsterr))
3266                         goto unwind;
3267         }
3268         nrealnocbers = nocbs_nthreads;
3269         if (WARN_ON(nrealnocbers < 0))
3270                 nrealnocbers = 1;
3271         if (WARN_ON(nocbs_toggle < 0))
3272                 nocbs_toggle = HZ;
3273         if (nrealnocbers > 0) {
3274                 nocb_tasks = kcalloc(nrealnocbers, sizeof(nocb_tasks[0]), GFP_KERNEL);
3275                 if (nocb_tasks == NULL) {
3276                         TOROUT_ERRSTRING("out of memory");
3277                         firsterr = -ENOMEM;
3278                         goto unwind;
3279                 }
3280         } else {
3281                 nocb_tasks = NULL;
3282         }
3283         for (i = 0; i < nrealnocbers; i++) {
3284                 firsterr = torture_create_kthread(rcu_nocb_toggle, NULL, nocb_tasks[i]);
3285                 if (torture_init_error(firsterr))
3286                         goto unwind;
3287         }
3288         if (stat_interval > 0) {
3289                 firsterr = torture_create_kthread(rcu_torture_stats, NULL,
3290                                                   stats_task);
3291                 if (torture_init_error(firsterr))
3292                         goto unwind;
3293         }
3294         if (test_no_idle_hz && shuffle_interval > 0) {
3295                 firsterr = torture_shuffle_init(shuffle_interval * HZ);
3296                 if (torture_init_error(firsterr))
3297                         goto unwind;
3298         }
3299         if (stutter < 0)
3300                 stutter = 0;
3301         if (stutter) {
3302                 int t;
3303
3304                 t = cur_ops->stall_dur ? cur_ops->stall_dur() : stutter * HZ;
3305                 firsterr = torture_stutter_init(stutter * HZ, t);
3306                 if (torture_init_error(firsterr))
3307                         goto unwind;
3308         }
3309         if (fqs_duration < 0)
3310                 fqs_duration = 0;
3311         if (fqs_duration) {
3312                 /* Create the fqs thread */
3313                 firsterr = torture_create_kthread(rcu_torture_fqs, NULL,
3314                                                   fqs_task);
3315                 if (torture_init_error(firsterr))
3316                         goto unwind;
3317         }
3318         if (test_boost_interval < 1)
3319                 test_boost_interval = 1;
3320         if (test_boost_duration < 2)
3321                 test_boost_duration = 2;
3322         if (rcu_torture_can_boost()) {
3323
3324                 boost_starttime = jiffies + test_boost_interval * HZ;
3325
3326                 firsterr = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "RCU_TORTURE",
3327                                              rcutorture_booster_init,
3328                                              rcutorture_booster_cleanup);
3329                 rcutor_hp = firsterr;
3330                 if (torture_init_error(firsterr))
3331                         goto unwind;
3332
3333                 // Testing RCU priority boosting requires rcutorture do
3334                 // some serious abuse.  Counter this by running ksoftirqd
3335                 // at higher priority.
3336                 if (IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)) {
3337                         for_each_online_cpu(cpu) {
3338                                 struct sched_param sp;
3339                                 struct task_struct *t;
3340
3341                                 t = per_cpu(ksoftirqd, cpu);
3342                                 WARN_ON_ONCE(!t);
3343                                 sp.sched_priority = 2;
3344                                 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3345                         }
3346                 }
3347         }
3348         shutdown_jiffies = jiffies + shutdown_secs * HZ;
3349         firsterr = torture_shutdown_init(shutdown_secs, rcu_torture_cleanup);
3350         if (torture_init_error(firsterr))
3351                 goto unwind;
3352         firsterr = torture_onoff_init(onoff_holdoff * HZ, onoff_interval,
3353                                       rcutorture_sync);
3354         if (torture_init_error(firsterr))
3355                 goto unwind;
3356         firsterr = rcu_torture_stall_init();
3357         if (torture_init_error(firsterr))
3358                 goto unwind;
3359         firsterr = rcu_torture_fwd_prog_init();
3360         if (torture_init_error(firsterr))
3361                 goto unwind;
3362         firsterr = rcu_torture_barrier_init();
3363         if (torture_init_error(firsterr))
3364                 goto unwind;
3365         firsterr = rcu_torture_read_exit_init();
3366         if (torture_init_error(firsterr))
3367                 goto unwind;
3368         if (object_debug)
3369                 rcu_test_debug_objects();
3370         torture_init_end();
3371         rcu_gp_slow_register(&rcu_fwd_cb_nodelay);
3372         return 0;
3373
3374 unwind:
3375         torture_init_end();
3376         rcu_torture_cleanup();
3377         if (shutdown_secs) {
3378                 WARN_ON(!IS_MODULE(CONFIG_RCU_TORTURE_TEST));
3379                 kernel_power_off();
3380         }
3381         return firsterr;
3382 }
3383
3384 module_init(rcu_torture_init);
3385 module_exit(rcu_torture_cleanup);