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