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/rcu_notifier.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/sched/signal.h>
  27 #include <uapi/linux/sched/types.h>
  28 #include <linux/atomic.h>
  29 #include <linux/bitops.h>
  30 #include <linux/completion.h>
  31 #include <linux/moduleparam.h>
  32 #include <linux/percpu.h>
  33 #include <linux/notifier.h>
  34 #include <linux/reboot.h>
  35 #include <linux/freezer.h>
  36 #include <linux/cpu.h>
  37 #include <linux/delay.h>
  38 #include <linux/stat.h>
  39 #include <linux/srcu.h>
  40 #include <linux/slab.h>
  41 #include <linux/trace_clock.h>
  42 #include <asm/byteorder.h>
  43 #include <linux/torture.h>
  44 #include <linux/vmalloc.h>
  45 #include <linux/sched/debug.h>
  46 #include <linux/sched/sysctl.h>
  47 #include <linux/oom.h>
  48 #include <linux/tick.h>
  49 #include <linux/rcupdate_trace.h>
  50 #include <linux/nmi.h>
  51
  52 #include "rcu.h"
  53
  54 MODULE_LICENSE("GPL");
  55 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com> and Josh Triplett <josh@joshtriplett.org>");
  56
  57 /* Bits for ->extendables field, extendables param, and related definitions. */
  58 #define RCUTORTURE_RDR_SHIFT_1   8      /* Put SRCU index in upper bits. */
  59 #define RCUTORTURE_RDR_MASK_1    (1 << RCUTORTURE_RDR_SHIFT_1)
  60 #define RCUTORTURE_RDR_SHIFT_2   9      /* Put SRCU index in upper bits. */
  61 #define RCUTORTURE_RDR_MASK_2    (1 << RCUTORTURE_RDR_SHIFT_2)
  62 #define RCUTORTURE_RDR_BH        0x01   /* Extend readers by disabling bh. */
  63 #define RCUTORTURE_RDR_IRQ       0x02   /*  ... disabling interrupts. */
  64 #define RCUTORTURE_RDR_PREEMPT   0x04   /*  ... disabling preemption. */
  65 #define RCUTORTURE_RDR_RBH       0x08   /*  ... rcu_read_lock_bh(). */
  66 #define RCUTORTURE_RDR_SCHED     0x10   /*  ... rcu_read_lock_sched(). */
  67 #define RCUTORTURE_RDR_RCU_1     0x20   /*  ... entering another RCU reader. */
  68 #define RCUTORTURE_RDR_RCU_2     0x40   /*  ... entering another RCU reader. */
  69 #define RCUTORTURE_RDR_NBITS     7      /* Number of bits defined above. */
  70 #define RCUTORTURE_MAX_EXTEND    \
  71         (RCUTORTURE_RDR_BH | RCUTORTURE_RDR_IRQ | RCUTORTURE_RDR_PREEMPT | \
  72          RCUTORTURE_RDR_RBH | RCUTORTURE_RDR_SCHED)
  73 #define RCUTORTURE_RDR_MAX_LOOPS 0x7    /* Maximum reader extensions. */
  74                                         /* Must be power of two minus one. */
  75 #define RCUTORTURE_RDR_MAX_SEGS (RCUTORTURE_RDR_MAX_LOOPS + 3)
  76
  77 torture_param(int, extendables, RCUTORTURE_MAX_EXTEND,
  78               "Extend readers by disabling bh (1), irqs (2), or preempt (4)");
  79 torture_param(int, fqs_duration, 0, "Duration of fqs bursts (us), 0 to disable");
  80 torture_param(int, fqs_holdoff, 0, "Holdoff time within fqs bursts (us)");
  81 torture_param(int, fqs_stutter, 3, "Wait time between fqs bursts (s)");
  82 torture_param(int, fwd_progress, 1, "Number of grace-period forward progress tasks (0 to disable)");
  83 torture_param(int, fwd_progress_div, 4, "Fraction of CPU stall to wait");
  84 torture_param(int, fwd_progress_holdoff, 60, "Time between forward-progress tests (s)");
  85 torture_param(bool, fwd_progress_need_resched, 1, "Hide cond_resched() behind need_resched()");
  86 torture_param(bool, gp_cond, false, "Use conditional/async GP wait primitives");
  87 torture_param(bool, gp_cond_exp, false, "Use conditional/async expedited GP wait primitives");
  88 torture_param(bool, gp_cond_full, false, "Use conditional/async full-state GP wait primitives");
  89 torture_param(bool, gp_cond_exp_full, false,
  90                     "Use conditional/async full-stateexpedited GP wait primitives");
  91 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
  92 torture_param(bool, gp_normal, false, "Use normal (non-expedited) GP wait primitives");
  93 torture_param(bool, gp_poll, false, "Use polling GP wait primitives");
  94 torture_param(bool, gp_poll_exp, false, "Use polling expedited GP wait primitives");
  95 torture_param(bool, gp_poll_full, false, "Use polling full-state GP wait primitives");
  96 torture_param(bool, gp_poll_exp_full, false, "Use polling full-state expedited GP wait primitives");
  97 torture_param(bool, gp_sync, false, "Use synchronous GP wait primitives");
  98 torture_param(int, irqreader, 1, "Allow RCU readers from irq handlers");
  99 torture_param(int, leakpointer, 0, "Leak pointer dereferences from readers");
 100 torture_param(int, n_barrier_cbs, 0, "# of callbacks/kthreads for barrier testing");
 101 torture_param(int, nfakewriters, 4, "Number of RCU fake writer threads");
 102 torture_param(int, nreaders, -1, "Number of RCU reader threads");
 103 torture_param(int, object_debug, 0, "Enable debug-object double call_rcu() testing");
 104 torture_param(int, onoff_holdoff, 0, "Time after boot before CPU hotplugs (s)");
 105 torture_param(int, onoff_interval, 0, "Time between CPU hotplugs (jiffies), 0=disable");
 106 torture_param(int, nocbs_nthreads, 0, "Number of NOCB toggle threads, 0 to disable");
 107 torture_param(int, nocbs_toggle, 1000, "Time between toggling nocb state (ms)");
 108 torture_param(int, read_exit_delay, 13, "Delay between read-then-exit episodes (s)");
 109 torture_param(int, read_exit_burst, 16, "# of read-then-exit bursts per episode, zero to disable");
 110 torture_param(int, shuffle_interval, 3, "Number of seconds between shuffles");
 111 torture_param(int, shutdown_secs, 0, "Shutdown time (s), <= zero to disable.");
 112 torture_param(int, stall_cpu, 0, "Stall duration (s), zero to disable.");
 113 torture_param(int, stall_cpu_holdoff, 10, "Time to wait before starting stall (s).");
 114 torture_param(bool, stall_no_softlockup, false, "Avoid softlockup warning during cpu stall.");
 115 torture_param(int, stall_cpu_irqsoff, 0, "Disable interrupts while stalling.");
 116 torture_param(int, stall_cpu_block, 0, "Sleep while stalling.");
 117 torture_param(int, stall_gp_kthread, 0, "Grace-period kthread stall duration (s).");
 118 torture_param(int, stat_interval, 60, "Number of seconds between stats printk()s");
 119 torture_param(int, stutter, 5, "Number of seconds to run/halt test");
 120 torture_param(int, test_boost, 1, "Test RCU prio boost: 0=no, 1=maybe, 2=yes.");
 121 torture_param(int, test_boost_duration, 4, "Duration of each boost test, seconds.");
 122 torture_param(int, test_boost_interval, 7, "Interval between boost tests, seconds.");
 123 torture_param(int, test_nmis, 0, "End-test NMI tests, 0 to disable.");
 124 torture_param(bool, test_no_idle_hz, true, "Test support for tickless idle CPUs");
 125 torture_param(int, test_srcu_lockdep, 0, "Test specified SRCU deadlock scenario.");
 126 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
 127
 128 static char *torture_type = "rcu";
 129 module_param(torture_type, charp, 0444);
 130 MODULE_PARM_DESC(torture_type, "Type of RCU to torture (rcu, srcu, ...)");
 131
 132 static int nrealnocbers;
 133 static int nrealreaders;
 134 static struct task_struct *writer_task;
 135 static struct task_struct **fakewriter_tasks;
 136 static struct task_struct **reader_tasks;
 137 static struct task_struct **nocb_tasks;
 138 static struct task_struct *stats_task;
 139 static struct task_struct *fqs_task;
 140 static struct task_struct *boost_tasks[NR_CPUS];
 141 static struct task_struct *stall_task;
 142 static struct task_struct **fwd_prog_tasks;
 143 static struct task_struct **barrier_cbs_tasks;
 144 static struct task_struct *barrier_task;
 145 static struct task_struct *read_exit_task;
 146
 147 #define RCU_TORTURE_PIPE_LEN 10
 148
 149 // Mailbox-like structure to check RCU global memory ordering.
 150 struct rcu_torture_reader_check {
 151         unsigned long rtc_myloops;
 152         int rtc_chkrdr;
 153         unsigned long rtc_chkloops;
 154         int rtc_ready;
 155         struct rcu_torture_reader_check *rtc_assigner;
 156 } ____cacheline_internodealigned_in_smp;
 157
 158 // Update-side data structure used to check RCU readers.
 159 struct rcu_torture {
 160         struct rcu_head rtort_rcu;
 161         int rtort_pipe_count;
 162         struct list_head rtort_free;
 163         int rtort_mbtest;
 164         struct rcu_torture_reader_check *rtort_chkp;
 165 };
 166
 167 static LIST_HEAD(rcu_torture_freelist);
 168 static struct rcu_torture __rcu *rcu_torture_current;
 169 static unsigned long rcu_torture_current_version;
 170 static struct rcu_torture rcu_tortures[10 * RCU_TORTURE_PIPE_LEN];
 171 static DEFINE_SPINLOCK(rcu_torture_lock);
 172 static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1], rcu_torture_count);
 173 static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1], rcu_torture_batch);
 174 static atomic_t rcu_torture_wcount[RCU_TORTURE_PIPE_LEN + 1];
 175 static struct rcu_torture_reader_check *rcu_torture_reader_mbchk;
 176 static atomic_t n_rcu_torture_alloc;
 177 static atomic_t n_rcu_torture_alloc_fail;
 178 static atomic_t n_rcu_torture_free;
 179 static atomic_t n_rcu_torture_mberror;
 180 static atomic_t n_rcu_torture_mbchk_fail;
 181 static atomic_t n_rcu_torture_mbchk_tries;
 182 static atomic_t n_rcu_torture_error;
 183 static long n_rcu_torture_barrier_error;
 184 static long n_rcu_torture_boost_ktrerror;
 185 static long n_rcu_torture_boost_failure;
 186 static long n_rcu_torture_boosts;
 187 static atomic_long_t n_rcu_torture_timers;
 188 static long n_barrier_attempts;
 189 static long n_barrier_successes; /* did rcu_barrier test succeed? */
 190 static unsigned long n_read_exits;
 191 static struct list_head rcu_torture_removed;
 192 static unsigned long shutdown_jiffies;
 193 static unsigned long start_gp_seq;
 194 static atomic_long_t n_nocb_offload;
 195 static atomic_long_t n_nocb_deoffload;
 196
 197 static int rcu_torture_writer_state;
 198 #define RTWS_FIXED_DELAY        0
 199 #define RTWS_DELAY              1
 200 #define RTWS_REPLACE            2
 201 #define RTWS_DEF_FREE           3
 202 #define RTWS_EXP_SYNC           4
 203 #define RTWS_COND_GET           5
 204 #define RTWS_COND_GET_FULL      6
 205 #define RTWS_COND_GET_EXP       7
 206 #define RTWS_COND_GET_EXP_FULL  8
 207 #define RTWS_COND_SYNC          9
 208 #define RTWS_COND_SYNC_FULL     10
 209 #define RTWS_COND_SYNC_EXP      11
 210 #define RTWS_COND_SYNC_EXP_FULL 12
 211 #define RTWS_POLL_GET           13
 212 #define RTWS_POLL_GET_FULL      14
 213 #define RTWS_POLL_GET_EXP       15
 214 #define RTWS_POLL_GET_EXP_FULL  16
 215 #define RTWS_POLL_WAIT          17
 216 #define RTWS_POLL_WAIT_FULL     18
 217 #define RTWS_POLL_WAIT_EXP      19
 218 #define RTWS_POLL_WAIT_EXP_FULL 20
 219 #define RTWS_SYNC               21
 220 #define RTWS_STUTTER            22
 221 #define RTWS_STOPPING           23
 222 static const char * const rcu_torture_writer_state_names[] = {
 223         "RTWS_FIXED_DELAY",
 224         "RTWS_DELAY",
 225         "RTWS_REPLACE",
 226         "RTWS_DEF_FREE",
 227         "RTWS_EXP_SYNC",
 228         "RTWS_COND_GET",
 229         "RTWS_COND_GET_FULL",
 230         "RTWS_COND_GET_EXP",
 231         "RTWS_COND_GET_EXP_FULL",
 232         "RTWS_COND_SYNC",
 233         "RTWS_COND_SYNC_FULL",
 234         "RTWS_COND_SYNC_EXP",
 235         "RTWS_COND_SYNC_EXP_FULL",
 236         "RTWS_POLL_GET",
 237         "RTWS_POLL_GET_FULL",
 238         "RTWS_POLL_GET_EXP",
 239         "RTWS_POLL_GET_EXP_FULL",
 240         "RTWS_POLL_WAIT",
 241         "RTWS_POLL_WAIT_FULL",
 242         "RTWS_POLL_WAIT_EXP",
 243         "RTWS_POLL_WAIT_EXP_FULL",
 244         "RTWS_SYNC",
 245         "RTWS_STUTTER",
 246         "RTWS_STOPPING",
 247 };
 248
 249 /* Record reader segment types and duration for first failing read. */
 250 struct rt_read_seg {
 251         int rt_readstate;
 252         unsigned long rt_delay_jiffies;
 253         unsigned long rt_delay_ms;
 254         unsigned long rt_delay_us;
 255         bool rt_preempted;
 256 };
 257 static int err_segs_recorded;
 258 static struct rt_read_seg err_segs[RCUTORTURE_RDR_MAX_SEGS];
 259 static int rt_read_nsegs;
 260
 261 static const char *rcu_torture_writer_state_getname(void)
 262 {
 263         unsigned int i = READ_ONCE(rcu_torture_writer_state);
 264
 265         if (i >= ARRAY_SIZE(rcu_torture_writer_state_names))
 266                 return "???";
 267         return rcu_torture_writer_state_names[i];
 268 }
 269
 270 #ifdef CONFIG_RCU_TRACE
 271 static u64 notrace rcu_trace_clock_local(void)
 272 {
 273         u64 ts = trace_clock_local();
 274
 275         (void)do_div(ts, NSEC_PER_USEC);
 276         return ts;
 277 }
 278 #else /* #ifdef CONFIG_RCU_TRACE */
 279 static u64 notrace rcu_trace_clock_local(void)
 280 {
 281         return 0ULL;
 282 }
 283 #endif /* #else #ifdef CONFIG_RCU_TRACE */
 284
 285 /*
 286  * Stop aggressive CPU-hog tests a bit before the end of the test in order
 287  * to avoid interfering with test shutdown.
 288  */
 289 static bool shutdown_time_arrived(void)
 290 {
 291         return shutdown_secs && time_after(jiffies, shutdown_jiffies - 30 * HZ);
 292 }
 293
 294 static unsigned long boost_starttime;   /* jiffies of next boost test start. */
 295 static DEFINE_MUTEX(boost_mutex);       /* protect setting boost_starttime */
 296                                         /*  and boost task create/destroy. */
 297 static atomic_t barrier_cbs_count;      /* Barrier callbacks registered. */
 298 static bool barrier_phase;              /* Test phase. */
 299 static atomic_t barrier_cbs_invoked;    /* Barrier callbacks invoked. */
 300 static wait_queue_head_t *barrier_cbs_wq; /* Coordinate barrier testing. */
 301 static DECLARE_WAIT_QUEUE_HEAD(barrier_wq);
 302
 303 static atomic_t rcu_fwd_cb_nodelay;     /* Short rcu_torture_delay() delays. */
 304
 305 /*
 306  * Allocate an element from the rcu_tortures pool.
 307  */
 308 static struct rcu_torture *
 309 rcu_torture_alloc(void)
 310 {
 311         struct list_head *p;
 312
 313         spin_lock_bh(&rcu_torture_lock);
 314         if (list_empty(&rcu_torture_freelist)) {
 315                 atomic_inc(&n_rcu_torture_alloc_fail);
 316                 spin_unlock_bh(&rcu_torture_lock);
 317                 return NULL;
 318         }
 319         atomic_inc(&n_rcu_torture_alloc);
 320         p = rcu_torture_freelist.next;
 321         list_del_init(p);
 322         spin_unlock_bh(&rcu_torture_lock);
 323         return container_of(p, struct rcu_torture, rtort_free);
 324 }
 325
 326 /*
 327  * Free an element to the rcu_tortures pool.
 328  */
 329 static void
 330 rcu_torture_free(struct rcu_torture *p)
 331 {
 332         atomic_inc(&n_rcu_torture_free);
 333         spin_lock_bh(&rcu_torture_lock);
 334         list_add_tail(&p->rtort_free, &rcu_torture_freelist);
 335         spin_unlock_bh(&rcu_torture_lock);
 336 }
 337
 338 /*
 339  * Operations vector for selecting different types of tests.
 340  */
 341
 342 struct rcu_torture_ops {
 343         int ttype;
 344         void (*init)(void);
 345         void (*cleanup)(void);
 346         int (*readlock)(void);
 347         void (*read_delay)(struct torture_random_state *rrsp,
 348                            struct rt_read_seg *rtrsp);
 349         void (*readunlock)(int idx);
 350         int (*readlock_held)(void);
 351         unsigned long (*get_gp_seq)(void);
 352         unsigned long (*gp_diff)(unsigned long new, unsigned long old);
 353         void (*deferred_free)(struct rcu_torture *p);
 354         void (*sync)(void);
 355         void (*exp_sync)(void);
 356         unsigned long (*get_gp_state_exp)(void);
 357         unsigned long (*start_gp_poll_exp)(void);
 358         void (*start_gp_poll_exp_full)(struct rcu_gp_oldstate *rgosp);
 359         bool (*poll_gp_state_exp)(unsigned long oldstate);
 360         void (*cond_sync_exp)(unsigned long oldstate);
 361         void (*cond_sync_exp_full)(struct rcu_gp_oldstate *rgosp);
 362         unsigned long (*get_comp_state)(void);
 363         void (*get_comp_state_full)(struct rcu_gp_oldstate *rgosp);
 364         bool (*same_gp_state)(unsigned long oldstate1, unsigned long oldstate2);
 365         bool (*same_gp_state_full)(struct rcu_gp_oldstate *rgosp1, struct rcu_gp_oldstate *rgosp2);
 366         unsigned long (*get_gp_state)(void);
 367         void (*get_gp_state_full)(struct rcu_gp_oldstate *rgosp);
 368         unsigned long (*get_gp_completed)(void);
 369         void (*get_gp_completed_full)(struct rcu_gp_oldstate *rgosp);
 370         unsigned long (*start_gp_poll)(void);
 371         void (*start_gp_poll_full)(struct rcu_gp_oldstate *rgosp);
 372         bool (*poll_gp_state)(unsigned long oldstate);
 373         bool (*poll_gp_state_full)(struct rcu_gp_oldstate *rgosp);
 374         bool (*poll_need_2gp)(bool poll, bool poll_full);
 375         void (*cond_sync)(unsigned long oldstate);
 376         void (*cond_sync_full)(struct rcu_gp_oldstate *rgosp);
 377         call_rcu_func_t call;
 378         void (*cb_barrier)(void);
 379         void (*fqs)(void);
 380         void (*stats)(void);
 381         void (*gp_kthread_dbg)(void);
 382         bool (*check_boost_failed)(unsigned long gp_state, int *cpup);
 383         int (*stall_dur)(void);
 384         long cbflood_max;
 385         int irq_capable;
 386         int can_boost;
 387         int extendables;
 388         int slow_gps;
 389         int no_pi_lock;
 390         const char *name;
 391 };
 392
 393 static struct rcu_torture_ops *cur_ops;
 394
 395 /*
 396  * Definitions for rcu torture testing.
 397  */
 398
 399 static int torture_readlock_not_held(void)
 400 {
 401         return rcu_read_lock_bh_held() || rcu_read_lock_sched_held();
 402 }
 403
 404 static int rcu_torture_read_lock(void)
 405 {
 406         rcu_read_lock();
 407         return 0;
 408 }
 409
 410 static void
 411 rcu_read_delay(struct torture_random_state *rrsp, struct rt_read_seg *rtrsp)
 412 {
 413         unsigned long started;
 414         unsigned long completed;
 415         const unsigned long shortdelay_us = 200;
 416         unsigned long longdelay_ms = 300;
 417         unsigned long long ts;
 418
 419         /* We want a short delay sometimes to make a reader delay the grace
 420          * period, and we want a long delay occasionally to trigger
 421          * force_quiescent_state. */
 422
 423         if (!atomic_read(&rcu_fwd_cb_nodelay) &&
 424             !(torture_random(rrsp) % (nrealreaders * 2000 * longdelay_ms))) {
 425                 started = cur_ops->get_gp_seq();
 426                 ts = rcu_trace_clock_local();
 427                 if (preempt_count() & (SOFTIRQ_MASK | HARDIRQ_MASK))
 428                         longdelay_ms = 5; /* Avoid triggering BH limits. */
 429                 mdelay(longdelay_ms);
 430                 rtrsp->rt_delay_ms = longdelay_ms;
 431                 completed = cur_ops->get_gp_seq();
 432                 do_trace_rcu_torture_read(cur_ops->name, NULL, ts,
 433                                           started, completed);
 434         }
 435         if (!(torture_random(rrsp) % (nrealreaders * 2 * shortdelay_us))) {
 436                 udelay(shortdelay_us);
 437                 rtrsp->rt_delay_us = shortdelay_us;
 438         }
 439         if (!preempt_count() &&
 440             !(torture_random(rrsp) % (nrealreaders * 500))) {
 441                 torture_preempt_schedule();  /* QS only if preemptible. */
 442                 rtrsp->rt_preempted = true;
 443         }
 444 }
 445
 446 static void rcu_torture_read_unlock(int idx)
 447 {
 448         rcu_read_unlock();
 449 }
 450
 451 /*
 452  * Update callback in the pipe.  This should be invoked after a grace period.
 453  */
 454 static bool
 455 rcu_torture_pipe_update_one(struct rcu_torture *rp)
 456 {
 457         int i;
 458         struct rcu_torture_reader_check *rtrcp = READ_ONCE(rp->rtort_chkp);
 459
 460         if (rtrcp) {
 461                 WRITE_ONCE(rp->rtort_chkp, NULL);
 462                 smp_store_release(&rtrcp->rtc_ready, 1); // Pair with smp_load_acquire().
 463         }
 464         i = READ_ONCE(rp->rtort_pipe_count);
 465         if (i > RCU_TORTURE_PIPE_LEN)
 466                 i = RCU_TORTURE_PIPE_LEN;
 467         atomic_inc(&rcu_torture_wcount[i]);
 468         WRITE_ONCE(rp->rtort_pipe_count, i + 1);
 469         if (rp->rtort_pipe_count >= RCU_TORTURE_PIPE_LEN) {
 470                 rp->rtort_mbtest = 0;
 471                 return true;
 472         }
 473         return false;
 474 }
 475
 476 /*
 477  * Update all callbacks in the pipe.  Suitable for synchronous grace-period
 478  * primitives.
 479  */
 480 static void
 481 rcu_torture_pipe_update(struct rcu_torture *old_rp)
 482 {
 483         struct rcu_torture *rp;
 484         struct rcu_torture *rp1;
 485
 486         if (old_rp)
 487                 list_add(&old_rp->rtort_free, &rcu_torture_removed);
 488         list_for_each_entry_safe(rp, rp1, &rcu_torture_removed, rtort_free) {
 489                 if (rcu_torture_pipe_update_one(rp)) {
 490                         list_del(&rp->rtort_free);
 491                         rcu_torture_free(rp);
 492                 }
 493         }
 494 }
 495
 496 static void
 497 rcu_torture_cb(struct rcu_head *p)
 498 {
 499         struct rcu_torture *rp = container_of(p, struct rcu_torture, rtort_rcu);
 500
 501         if (torture_must_stop_irq()) {
 502                 /* Test is ending, just drop callbacks on the floor. */
 503                 /* The next initialization will pick up the pieces. */
 504                 return;
 505         }
 506         if (rcu_torture_pipe_update_one(rp))
 507                 rcu_torture_free(rp);
 508         else
 509                 cur_ops->deferred_free(rp);
 510 }
 511
 512 static unsigned long rcu_no_completed(void)
 513 {
 514         return 0;
 515 }
 516
 517 static void rcu_torture_deferred_free(struct rcu_torture *p)
 518 {
 519         call_rcu_hurry(&p->rtort_rcu, rcu_torture_cb);
 520 }
 521
 522 static void rcu_sync_torture_init(void)
 523 {
 524         INIT_LIST_HEAD(&rcu_torture_removed);
 525 }
 526
 527 static bool rcu_poll_need_2gp(bool poll, bool poll_full)
 528 {
 529         return poll;
 530 }
 531
 532 static struct rcu_torture_ops rcu_ops = {
 533         .ttype                  = RCU_FLAVOR,
 534         .init                   = rcu_sync_torture_init,
 535         .readlock               = rcu_torture_read_lock,
 536         .read_delay             = rcu_read_delay,
 537         .readunlock             = rcu_torture_read_unlock,
 538         .readlock_held          = torture_readlock_not_held,
 539         .get_gp_seq             = rcu_get_gp_seq,
 540         .gp_diff                = rcu_seq_diff,
 541         .deferred_free          = rcu_torture_deferred_free,
 542         .sync                   = synchronize_rcu,
 543         .exp_sync               = synchronize_rcu_expedited,
 544         .same_gp_state          = same_state_synchronize_rcu,
 545         .same_gp_state_full     = same_state_synchronize_rcu_full,
 546         .get_comp_state         = get_completed_synchronize_rcu,
 547         .get_comp_state_full    = get_completed_synchronize_rcu_full,
 548         .get_gp_state           = get_state_synchronize_rcu,
 549         .get_gp_state_full      = get_state_synchronize_rcu_full,
 550         .get_gp_completed       = get_completed_synchronize_rcu,
 551         .get_gp_completed_full  = get_completed_synchronize_rcu_full,
 552         .start_gp_poll          = start_poll_synchronize_rcu,
 553         .start_gp_poll_full     = start_poll_synchronize_rcu_full,
 554         .poll_gp_state          = poll_state_synchronize_rcu,
 555         .poll_gp_state_full     = poll_state_synchronize_rcu_full,
 556         .poll_need_2gp          = rcu_poll_need_2gp,
 557         .cond_sync              = cond_synchronize_rcu,
 558         .cond_sync_full         = cond_synchronize_rcu_full,
 559         .get_gp_state_exp       = get_state_synchronize_rcu,
 560         .start_gp_poll_exp      = start_poll_synchronize_rcu_expedited,
 561         .start_gp_poll_exp_full = start_poll_synchronize_rcu_expedited_full,
 562         .poll_gp_state_exp      = poll_state_synchronize_rcu,
 563         .cond_sync_exp          = cond_synchronize_rcu_expedited,
 564         .call                   = call_rcu_hurry,
 565         .cb_barrier             = rcu_barrier,
 566         .fqs                    = rcu_force_quiescent_state,
 567         .stats                  = NULL,
 568         .gp_kthread_dbg         = show_rcu_gp_kthreads,
 569         .check_boost_failed     = rcu_check_boost_fail,
 570         .stall_dur              = rcu_jiffies_till_stall_check,
 571         .irq_capable            = 1,
 572         .can_boost              = IS_ENABLED(CONFIG_RCU_BOOST),
 573         .extendables            = RCUTORTURE_MAX_EXTEND,
 574         .name                   = "rcu"
 575 };
 576
 577 /*
 578  * Don't even think about trying any of these in real life!!!
 579  * The names includes "busted", and they really means it!
 580  * The only purpose of these functions is to provide a buggy RCU
 581  * implementation to make sure that rcutorture correctly emits
 582  * buggy-RCU error messages.
 583  */
 584 static void rcu_busted_torture_deferred_free(struct rcu_torture *p)
 585 {
 586         /* This is a deliberate bug for testing purposes only! */
 587         rcu_torture_cb(&p->rtort_rcu);
 588 }
 589
 590 static void synchronize_rcu_busted(void)
 591 {
 592         /* This is a deliberate bug for testing purposes only! */
 593 }
 594
 595 static void
 596 call_rcu_busted(struct rcu_head *head, rcu_callback_t func)
 597 {
 598         /* This is a deliberate bug for testing purposes only! */
 599         func(head);
 600 }
 601
 602 static struct rcu_torture_ops rcu_busted_ops = {
 603         .ttype          = INVALID_RCU_FLAVOR,
 604         .init           = rcu_sync_torture_init,
 605         .readlock       = rcu_torture_read_lock,
 606         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 607         .readunlock     = rcu_torture_read_unlock,
 608         .readlock_held  = torture_readlock_not_held,
 609         .get_gp_seq     = rcu_no_completed,
 610         .deferred_free  = rcu_busted_torture_deferred_free,
 611         .sync           = synchronize_rcu_busted,
 612         .exp_sync       = synchronize_rcu_busted,
 613         .call           = call_rcu_busted,
 614         .cb_barrier     = NULL,
 615         .fqs            = NULL,
 616         .stats          = NULL,
 617         .irq_capable    = 1,
 618         .name           = "busted"
 619 };
 620
 621 /*
 622  * Definitions for srcu torture testing.
 623  */
 624
 625 DEFINE_STATIC_SRCU(srcu_ctl);
 626 static struct srcu_struct srcu_ctld;
 627 static struct srcu_struct *srcu_ctlp = &srcu_ctl;
 628 static struct rcu_torture_ops srcud_ops;
 629
 630 static int srcu_torture_read_lock(void)
 631 {
 632         if (cur_ops == &srcud_ops)
 633                 return srcu_read_lock_nmisafe(srcu_ctlp);
 634         else
 635                 return srcu_read_lock(srcu_ctlp);
 636 }
 637
 638 static void
 639 srcu_read_delay(struct torture_random_state *rrsp, struct rt_read_seg *rtrsp)
 640 {
 641         long delay;
 642         const long uspertick = 1000000 / HZ;
 643         const long longdelay = 10;
 644
 645         /* We want there to be long-running readers, but not all the time. */
 646
 647         delay = torture_random(rrsp) %
 648                 (nrealreaders * 2 * longdelay * uspertick);
 649         if (!delay && in_task()) {
 650                 schedule_timeout_interruptible(longdelay);
 651                 rtrsp->rt_delay_jiffies = longdelay;
 652         } else {
 653                 rcu_read_delay(rrsp, rtrsp);
 654         }
 655 }
 656
 657 static void srcu_torture_read_unlock(int idx)
 658 {
 659         if (cur_ops == &srcud_ops)
 660                 srcu_read_unlock_nmisafe(srcu_ctlp, idx);
 661         else
 662                 srcu_read_unlock(srcu_ctlp, idx);
 663 }
 664
 665 static int torture_srcu_read_lock_held(void)
 666 {
 667         return srcu_read_lock_held(srcu_ctlp);
 668 }
 669
 670 static unsigned long srcu_torture_completed(void)
 671 {
 672         return srcu_batches_completed(srcu_ctlp);
 673 }
 674
 675 static void srcu_torture_deferred_free(struct rcu_torture *rp)
 676 {
 677         call_srcu(srcu_ctlp, &rp->rtort_rcu, rcu_torture_cb);
 678 }
 679
 680 static void srcu_torture_synchronize(void)
 681 {
 682         synchronize_srcu(srcu_ctlp);
 683 }
 684
 685 static unsigned long srcu_torture_get_gp_state(void)
 686 {
 687         return get_state_synchronize_srcu(srcu_ctlp);
 688 }
 689
 690 static unsigned long srcu_torture_start_gp_poll(void)
 691 {
 692         return start_poll_synchronize_srcu(srcu_ctlp);
 693 }
 694
 695 static bool srcu_torture_poll_gp_state(unsigned long oldstate)
 696 {
 697         return poll_state_synchronize_srcu(srcu_ctlp, oldstate);
 698 }
 699
 700 static void srcu_torture_call(struct rcu_head *head,
 701                               rcu_callback_t func)
 702 {
 703         call_srcu(srcu_ctlp, head, func);
 704 }
 705
 706 static void srcu_torture_barrier(void)
 707 {
 708         srcu_barrier(srcu_ctlp);
 709 }
 710
 711 static void srcu_torture_stats(void)
 712 {
 713         srcu_torture_stats_print(srcu_ctlp, torture_type, TORTURE_FLAG);
 714 }
 715
 716 static void srcu_torture_synchronize_expedited(void)
 717 {
 718         synchronize_srcu_expedited(srcu_ctlp);
 719 }
 720
 721 static struct rcu_torture_ops srcu_ops = {
 722         .ttype          = SRCU_FLAVOR,
 723         .init           = rcu_sync_torture_init,
 724         .readlock       = srcu_torture_read_lock,
 725         .read_delay     = srcu_read_delay,
 726         .readunlock     = srcu_torture_read_unlock,
 727         .readlock_held  = torture_srcu_read_lock_held,
 728         .get_gp_seq     = srcu_torture_completed,
 729         .deferred_free  = srcu_torture_deferred_free,
 730         .sync           = srcu_torture_synchronize,
 731         .exp_sync       = srcu_torture_synchronize_expedited,
 732         .get_gp_state   = srcu_torture_get_gp_state,
 733         .start_gp_poll  = srcu_torture_start_gp_poll,
 734         .poll_gp_state  = srcu_torture_poll_gp_state,
 735         .call           = srcu_torture_call,
 736         .cb_barrier     = srcu_torture_barrier,
 737         .stats          = srcu_torture_stats,
 738         .cbflood_max    = 50000,
 739         .irq_capable    = 1,
 740         .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
 741         .name           = "srcu"
 742 };
 743
 744 static void srcu_torture_init(void)
 745 {
 746         rcu_sync_torture_init();
 747         WARN_ON(init_srcu_struct(&srcu_ctld));
 748         srcu_ctlp = &srcu_ctld;
 749 }
 750
 751 static void srcu_torture_cleanup(void)
 752 {
 753         cleanup_srcu_struct(&srcu_ctld);
 754         srcu_ctlp = &srcu_ctl; /* In case of a later rcutorture run. */
 755 }
 756
 757 /* As above, but dynamically allocated. */
 758 static struct rcu_torture_ops srcud_ops = {
 759         .ttype          = SRCU_FLAVOR,
 760         .init           = srcu_torture_init,
 761         .cleanup        = srcu_torture_cleanup,
 762         .readlock       = srcu_torture_read_lock,
 763         .read_delay     = srcu_read_delay,
 764         .readunlock     = srcu_torture_read_unlock,
 765         .readlock_held  = torture_srcu_read_lock_held,
 766         .get_gp_seq     = srcu_torture_completed,
 767         .deferred_free  = srcu_torture_deferred_free,
 768         .sync           = srcu_torture_synchronize,
 769         .exp_sync       = srcu_torture_synchronize_expedited,
 770         .get_gp_state   = srcu_torture_get_gp_state,
 771         .start_gp_poll  = srcu_torture_start_gp_poll,
 772         .poll_gp_state  = srcu_torture_poll_gp_state,
 773         .call           = srcu_torture_call,
 774         .cb_barrier     = srcu_torture_barrier,
 775         .stats          = srcu_torture_stats,
 776         .cbflood_max    = 50000,
 777         .irq_capable    = 1,
 778         .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
 779         .name           = "srcud"
 780 };
 781
 782 /* As above, but broken due to inappropriate reader extension. */
 783 static struct rcu_torture_ops busted_srcud_ops = {
 784         .ttype          = SRCU_FLAVOR,
 785         .init           = srcu_torture_init,
 786         .cleanup        = srcu_torture_cleanup,
 787         .readlock       = srcu_torture_read_lock,
 788         .read_delay     = rcu_read_delay,
 789         .readunlock     = srcu_torture_read_unlock,
 790         .readlock_held  = torture_srcu_read_lock_held,
 791         .get_gp_seq     = srcu_torture_completed,
 792         .deferred_free  = srcu_torture_deferred_free,
 793         .sync           = srcu_torture_synchronize,
 794         .exp_sync       = srcu_torture_synchronize_expedited,
 795         .call           = srcu_torture_call,
 796         .cb_barrier     = srcu_torture_barrier,
 797         .stats          = srcu_torture_stats,
 798         .irq_capable    = 1,
 799         .no_pi_lock     = IS_ENABLED(CONFIG_TINY_SRCU),
 800         .extendables    = RCUTORTURE_MAX_EXTEND,
 801         .name           = "busted_srcud"
 802 };
 803
 804 /*
 805  * Definitions for trivial CONFIG_PREEMPT=n-only torture testing.
 806  * This implementation does not necessarily work well with CPU hotplug.
 807  */
 808
 809 static void synchronize_rcu_trivial(void)
 810 {
 811         int cpu;
 812
 813         for_each_online_cpu(cpu) {
 814                 torture_sched_setaffinity(current->pid, cpumask_of(cpu));
 815                 WARN_ON_ONCE(raw_smp_processor_id() != cpu);
 816         }
 817 }
 818
 819 static int rcu_torture_read_lock_trivial(void)
 820 {
 821         preempt_disable();
 822         return 0;
 823 }
 824
 825 static void rcu_torture_read_unlock_trivial(int idx)
 826 {
 827         preempt_enable();
 828 }
 829
 830 static struct rcu_torture_ops trivial_ops = {
 831         .ttype          = RCU_TRIVIAL_FLAVOR,
 832         .init           = rcu_sync_torture_init,
 833         .readlock       = rcu_torture_read_lock_trivial,
 834         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 835         .readunlock     = rcu_torture_read_unlock_trivial,
 836         .readlock_held  = torture_readlock_not_held,
 837         .get_gp_seq     = rcu_no_completed,
 838         .sync           = synchronize_rcu_trivial,
 839         .exp_sync       = synchronize_rcu_trivial,
 840         .fqs            = NULL,
 841         .stats          = NULL,
 842         .irq_capable    = 1,
 843         .name           = "trivial"
 844 };
 845
 846 #ifdef CONFIG_TASKS_RCU
 847
 848 /*
 849  * Definitions for RCU-tasks torture testing.
 850  */
 851
 852 static int tasks_torture_read_lock(void)
 853 {
 854         return 0;
 855 }
 856
 857 static void tasks_torture_read_unlock(int idx)
 858 {
 859 }
 860
 861 static void rcu_tasks_torture_deferred_free(struct rcu_torture *p)
 862 {
 863         call_rcu_tasks(&p->rtort_rcu, rcu_torture_cb);
 864 }
 865
 866 static void synchronize_rcu_mult_test(void)
 867 {
 868         synchronize_rcu_mult(call_rcu_tasks, call_rcu_hurry);
 869 }
 870
 871 static struct rcu_torture_ops tasks_ops = {
 872         .ttype          = RCU_TASKS_FLAVOR,
 873         .init           = rcu_sync_torture_init,
 874         .readlock       = tasks_torture_read_lock,
 875         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 876         .readunlock     = tasks_torture_read_unlock,
 877         .get_gp_seq     = rcu_no_completed,
 878         .deferred_free  = rcu_tasks_torture_deferred_free,
 879         .sync           = synchronize_rcu_tasks,
 880         .exp_sync       = synchronize_rcu_mult_test,
 881         .call           = call_rcu_tasks,
 882         .cb_barrier     = rcu_barrier_tasks,
 883         .gp_kthread_dbg = show_rcu_tasks_classic_gp_kthread,
 884         .fqs            = NULL,
 885         .stats          = NULL,
 886         .irq_capable    = 1,
 887         .slow_gps       = 1,
 888         .name           = "tasks"
 889 };
 890
 891 #define TASKS_OPS &tasks_ops,
 892
 893 #else // #ifdef CONFIG_TASKS_RCU
 894
 895 #define TASKS_OPS
 896
 897 #endif // #else #ifdef CONFIG_TASKS_RCU
 898
 899
 900 #ifdef CONFIG_TASKS_RUDE_RCU
 901
 902 /*
 903  * Definitions for rude RCU-tasks torture testing.
 904  */
 905
 906 static void rcu_tasks_rude_torture_deferred_free(struct rcu_torture *p)
 907 {
 908         call_rcu_tasks_rude(&p->rtort_rcu, rcu_torture_cb);
 909 }
 910
 911 static struct rcu_torture_ops tasks_rude_ops = {
 912         .ttype          = RCU_TASKS_RUDE_FLAVOR,
 913         .init           = rcu_sync_torture_init,
 914         .readlock       = rcu_torture_read_lock_trivial,
 915         .read_delay     = rcu_read_delay,  /* just reuse rcu's version. */
 916         .readunlock     = rcu_torture_read_unlock_trivial,
 917         .get_gp_seq     = rcu_no_completed,
 918         .deferred_free  = rcu_tasks_rude_torture_deferred_free,
 919         .sync           = synchronize_rcu_tasks_rude,
 920         .exp_sync       = synchronize_rcu_tasks_rude,
 921         .call           = call_rcu_tasks_rude,
 922         .cb_barrier     = rcu_barrier_tasks_rude,
 923         .gp_kthread_dbg = show_rcu_tasks_rude_gp_kthread,
 924         .cbflood_max    = 50000,
 925         .fqs            = NULL,
 926         .stats          = NULL,
 927         .irq_capable    = 1,
 928         .name           = "tasks-rude"
 929 };
 930
 931 #define TASKS_RUDE_OPS &tasks_rude_ops,
 932
 933 #else // #ifdef CONFIG_TASKS_RUDE_RCU
 934
 935 #define TASKS_RUDE_OPS
 936
 937 #endif // #else #ifdef CONFIG_TASKS_RUDE_RCU
 938
 939
 940 #ifdef CONFIG_TASKS_TRACE_RCU
 941
 942 /*
 943  * Definitions for tracing RCU-tasks torture testing.
 944  */
 945
 946 static int tasks_tracing_torture_read_lock(void)
 947 {
 948         rcu_read_lock_trace();
 949         return 0;
 950 }
 951
 952 static void tasks_tracing_torture_read_unlock(int idx)
 953 {
 954         rcu_read_unlock_trace();
 955 }
 956
 957 static void rcu_tasks_tracing_torture_deferred_free(struct rcu_torture *p)
 958 {
 959         call_rcu_tasks_trace(&p->rtort_rcu, rcu_torture_cb);
 960 }
 961
 962 static struct rcu_torture_ops tasks_tracing_ops = {
 963         .ttype          = RCU_TASKS_TRACING_FLAVOR,
 964         .init           = rcu_sync_torture_init,
 965         .readlock       = tasks_tracing_torture_read_lock,
 966         .read_delay     = srcu_read_delay,  /* just reuse srcu's version. */
 967         .readunlock     = tasks_tracing_torture_read_unlock,
 968         .readlock_held  = rcu_read_lock_trace_held,
 969         .get_gp_seq     = rcu_no_completed,
 970         .deferred_free  = rcu_tasks_tracing_torture_deferred_free,
 971         .sync           = synchronize_rcu_tasks_trace,
 972         .exp_sync       = synchronize_rcu_tasks_trace,
 973         .call           = call_rcu_tasks_trace,
 974         .cb_barrier     = rcu_barrier_tasks_trace,
 975         .gp_kthread_dbg = show_rcu_tasks_trace_gp_kthread,
 976         .cbflood_max    = 50000,
 977         .fqs            = NULL,
 978         .stats          = NULL,
 979         .irq_capable    = 1,
 980         .slow_gps       = 1,
 981         .name           = "tasks-tracing"
 982 };
 983
 984 #define TASKS_TRACING_OPS &tasks_tracing_ops,
 985
 986 #else // #ifdef CONFIG_TASKS_TRACE_RCU
 987
 988 #define TASKS_TRACING_OPS
 989
 990 #endif // #else #ifdef CONFIG_TASKS_TRACE_RCU
 991
 992
 993 static unsigned long rcutorture_seq_diff(unsigned long new, unsigned long old)
 994 {
 995         if (!cur_ops->gp_diff)
 996                 return new - old;
 997         return cur_ops->gp_diff(new, old);
 998 }
 999
1000 /*
1001  * RCU torture priority-boost testing.  Runs one real-time thread per
1002  * CPU for moderate bursts, repeatedly starting grace periods and waiting
1003  * for them to complete.  If a given grace period takes too long, we assume
1004  * that priority inversion has occurred.
1005  */
1006
1007 static int old_rt_runtime = -1;
1008
1009 static void rcu_torture_disable_rt_throttle(void)
1010 {
1011         /*
1012          * Disable RT throttling so that rcutorture's boost threads don't get
1013          * throttled. Only possible if rcutorture is built-in otherwise the
1014          * user should manually do this by setting the sched_rt_period_us and
1015          * sched_rt_runtime sysctls.
1016          */
1017         if (!IS_BUILTIN(CONFIG_RCU_TORTURE_TEST) || old_rt_runtime != -1)
1018                 return;
1019
1020         old_rt_runtime = sysctl_sched_rt_runtime;
1021         sysctl_sched_rt_runtime = -1;
1022 }
1023
1024 static void rcu_torture_enable_rt_throttle(void)
1025 {
1026         if (!IS_BUILTIN(CONFIG_RCU_TORTURE_TEST) || old_rt_runtime == -1)
1027                 return;
1028
1029         sysctl_sched_rt_runtime = old_rt_runtime;
1030         old_rt_runtime = -1;
1031 }
1032
1033 static bool rcu_torture_boost_failed(unsigned long gp_state, unsigned long *start)
1034 {
1035         int cpu;
1036         static int dbg_done;
1037         unsigned long end = jiffies;
1038         bool gp_done;
1039         unsigned long j;
1040         static unsigned long last_persist;
1041         unsigned long lp;
1042         unsigned long mininterval = test_boost_duration * HZ - HZ / 2;
1043
1044         if (end - *start > mininterval) {
1045                 // Recheck after checking time to avoid false positives.
1046                 smp_mb(); // Time check before grace-period check.
1047                 if (cur_ops->poll_gp_state(gp_state))
1048                         return false; // passed, though perhaps just barely
1049                 if (cur_ops->check_boost_failed && !cur_ops->check_boost_failed(gp_state, &cpu)) {
1050                         // At most one persisted message per boost test.
1051                         j = jiffies;
1052                         lp = READ_ONCE(last_persist);
1053                         if (time_after(j, lp + mininterval) && cmpxchg(&last_persist, lp, j) == lp)
1054                                 pr_info("Boost inversion persisted: No QS from CPU %d\n", cpu);
1055                         return false; // passed on a technicality
1056                 }
1057                 VERBOSE_TOROUT_STRING("rcu_torture_boost boosting failed");
1058                 n_rcu_torture_boost_failure++;
1059                 if (!xchg(&dbg_done, 1) && cur_ops->gp_kthread_dbg) {
1060                         pr_info("Boost inversion thread ->rt_priority %u gp_state %lu jiffies %lu\n",
1061                                 current->rt_priority, gp_state, end - *start);
1062                         cur_ops->gp_kthread_dbg();
1063                         // Recheck after print to flag grace period ending during splat.
1064                         gp_done = cur_ops->poll_gp_state(gp_state);
1065                         pr_info("Boost inversion: GP %lu %s.\n", gp_state,
1066                                 gp_done ? "ended already" : "still pending");
1067
1068                 }
1069
1070                 return true; // failed
1071         } else if (cur_ops->check_boost_failed && !cur_ops->check_boost_failed(gp_state, NULL)) {
1072                 *start = jiffies;
1073         }
1074
1075         return false; // passed
1076 }
1077
1078 static int rcu_torture_boost(void *arg)
1079 {
1080         unsigned long endtime;
1081         unsigned long gp_state;
1082         unsigned long gp_state_time;
1083         unsigned long oldstarttime;
1084
1085         VERBOSE_TOROUT_STRING("rcu_torture_boost started");
1086
1087         /* Set real-time priority. */
1088         sched_set_fifo_low(current);
1089
1090         /* Each pass through the following loop does one boost-test cycle. */
1091         do {
1092                 bool failed = false; // Test failed already in this test interval
1093                 bool gp_initiated = false;
1094
1095                 if (kthread_should_stop())
1096                         goto checkwait;
1097
1098                 /* Wait for the next test interval. */
1099                 oldstarttime = READ_ONCE(boost_starttime);
1100                 while (time_before(jiffies, oldstarttime)) {
1101                         schedule_timeout_interruptible(oldstarttime - jiffies);
1102                         if (stutter_wait("rcu_torture_boost"))
1103                                 sched_set_fifo_low(current);
1104                         if (torture_must_stop())
1105                                 goto checkwait;
1106                 }
1107
1108                 // Do one boost-test interval.
1109                 endtime = oldstarttime + test_boost_duration * HZ;
1110                 while (time_before(jiffies, endtime)) {
1111                         // Has current GP gone too long?
1112                         if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
1113                                 failed = rcu_torture_boost_failed(gp_state, &gp_state_time);
1114                         // If we don't have a grace period in flight, start one.
1115                         if (!gp_initiated || cur_ops->poll_gp_state(gp_state)) {
1116                                 gp_state = cur_ops->start_gp_poll();
1117                                 gp_initiated = true;
1118                                 gp_state_time = jiffies;
1119                         }
1120                         if (stutter_wait("rcu_torture_boost")) {
1121                                 sched_set_fifo_low(current);
1122                                 // If the grace period already ended,
1123                                 // we don't know when that happened, so
1124                                 // start over.
1125                                 if (cur_ops->poll_gp_state(gp_state))
1126                                         gp_initiated = false;
1127                         }
1128                         if (torture_must_stop())
1129                                 goto checkwait;
1130                 }
1131
1132                 // In case the grace period extended beyond the end of the loop.
1133                 if (gp_initiated && !failed && !cur_ops->poll_gp_state(gp_state))
1134                         rcu_torture_boost_failed(gp_state, &gp_state_time);
1135
1136                 /*
1137                  * Set the start time of the next test interval.
1138                  * Yes, this is vulnerable to long delays, but such
1139                  * delays simply cause a false negative for the next
1140                  * interval.  Besides, we are running at RT priority,
1141                  * so delays should be relatively rare.
1142                  */
1143                 while (oldstarttime == READ_ONCE(boost_starttime) && !kthread_should_stop()) {
1144                         if (mutex_trylock(&boost_mutex)) {
1145                                 if (oldstarttime == boost_starttime) {
1146                                         WRITE_ONCE(boost_starttime,
1147                                                    jiffies + test_boost_interval * HZ);
1148                                         n_rcu_torture_boosts++;
1149                                 }
1150                                 mutex_unlock(&boost_mutex);
1151                                 break;
1152                         }
1153                         schedule_timeout_uninterruptible(HZ / 20);
1154                 }
1155
1156                 /* Go do the stutter. */
1157 checkwait:      if (stutter_wait("rcu_torture_boost"))
1158                         sched_set_fifo_low(current);
1159         } while (!torture_must_stop());
1160
1161         /* Clean up and exit. */
1162         while (!kthread_should_stop()) {
1163                 torture_shutdown_absorb("rcu_torture_boost");
1164                 schedule_timeout_uninterruptible(HZ / 20);
1165         }
1166         torture_kthread_stopping("rcu_torture_boost");
1167         return 0;
1168 }
1169
1170 /*
1171  * RCU torture force-quiescent-state kthread.  Repeatedly induces
1172  * bursts of calls to force_quiescent_state(), increasing the probability
1173  * of occurrence of some important types of race conditions.
1174  */
1175 static int
1176 rcu_torture_fqs(void *arg)
1177 {
1178         unsigned long fqs_resume_time;
1179         int fqs_burst_remaining;
1180         int oldnice = task_nice(current);
1181
1182         VERBOSE_TOROUT_STRING("rcu_torture_fqs task started");
1183         do {
1184                 fqs_resume_time = jiffies + fqs_stutter * HZ;
1185                 while (time_before(jiffies, fqs_resume_time) &&
1186                        !kthread_should_stop()) {
1187                         schedule_timeout_interruptible(HZ / 20);
1188                 }
1189                 fqs_burst_remaining = fqs_duration;
1190                 while (fqs_burst_remaining > 0 &&
1191                        !kthread_should_stop()) {
1192                         cur_ops->fqs();
1193                         udelay(fqs_holdoff);
1194                         fqs_burst_remaining -= fqs_holdoff;
1195                 }
1196                 if (stutter_wait("rcu_torture_fqs"))
1197                         sched_set_normal(current, oldnice);
1198         } while (!torture_must_stop());
1199         torture_kthread_stopping("rcu_torture_fqs");
1200         return 0;
1201 }
1202
1203 // Used by writers to randomly choose from the available grace-period primitives.
1204 static int synctype[ARRAY_SIZE(rcu_torture_writer_state_names)] = { };
1205 static int nsynctypes;
1206
1207 /*
1208  * Determine which grace-period primitives are available.
1209  */
1210 static void rcu_torture_write_types(void)
1211 {
1212         bool gp_cond1 = gp_cond, gp_cond_exp1 = gp_cond_exp, gp_cond_full1 = gp_cond_full;
1213         bool gp_cond_exp_full1 = gp_cond_exp_full, gp_exp1 = gp_exp, gp_poll_exp1 = gp_poll_exp;
1214         bool gp_poll_exp_full1 = gp_poll_exp_full, gp_normal1 = gp_normal, gp_poll1 = gp_poll;
1215         bool gp_poll_full1 = gp_poll_full, gp_sync1 = gp_sync;
1216
1217         /* Initialize synctype[] array.  If none set, take default. */
1218         if (!gp_cond1 &&
1219             !gp_cond_exp1 &&
1220             !gp_cond_full1 &&
1221             !gp_cond_exp_full1 &&
1222             !gp_exp1 &&
1223             !gp_poll_exp1 &&
1224             !gp_poll_exp_full1 &&
1225             !gp_normal1 &&
1226             !gp_poll1 &&
1227             !gp_poll_full1 &&
1228             !gp_sync1) {
1229                 gp_cond1 = true;
1230                 gp_cond_exp1 = true;
1231                 gp_cond_full1 = true;
1232                 gp_cond_exp_full1 = true;
1233                 gp_exp1 = true;
1234                 gp_poll_exp1 = true;
1235                 gp_poll_exp_full1 = true;
1236                 gp_normal1 = true;
1237                 gp_poll1 = true;
1238                 gp_poll_full1 = true;
1239                 gp_sync1 = true;
1240         }
1241         if (gp_cond1 && cur_ops->get_gp_state && cur_ops->cond_sync) {
1242                 synctype[nsynctypes++] = RTWS_COND_GET;
1243                 pr_info("%s: Testing conditional GPs.\n", __func__);
1244         } else if (gp_cond && (!cur_ops->get_gp_state || !cur_ops->cond_sync)) {
1245                 pr_alert("%s: gp_cond without primitives.\n", __func__);
1246         }
1247         if (gp_cond_exp1 && cur_ops->get_gp_state_exp && cur_ops->cond_sync_exp) {
1248                 synctype[nsynctypes++] = RTWS_COND_GET_EXP;
1249                 pr_info("%s: Testing conditional expedited GPs.\n", __func__);
1250         } else if (gp_cond_exp && (!cur_ops->get_gp_state_exp || !cur_ops->cond_sync_exp)) {
1251                 pr_alert("%s: gp_cond_exp without primitives.\n", __func__);
1252         }
1253         if (gp_cond_full1 && cur_ops->get_gp_state && cur_ops->cond_sync_full) {
1254                 synctype[nsynctypes++] = RTWS_COND_GET_FULL;
1255                 pr_info("%s: Testing conditional full-state GPs.\n", __func__);
1256         } else if (gp_cond_full && (!cur_ops->get_gp_state || !cur_ops->cond_sync_full)) {
1257                 pr_alert("%s: gp_cond_full without primitives.\n", __func__);
1258         }
1259         if (gp_cond_exp_full1 && cur_ops->get_gp_state_exp && cur_ops->cond_sync_exp_full) {
1260                 synctype[nsynctypes++] = RTWS_COND_GET_EXP_FULL;
1261                 pr_info("%s: Testing conditional full-state expedited GPs.\n", __func__);
1262         } else if (gp_cond_exp_full &&
1263                    (!cur_ops->get_gp_state_exp || !cur_ops->cond_sync_exp_full)) {
1264                 pr_alert("%s: gp_cond_exp_full without primitives.\n", __func__);
1265         }
1266         if (gp_exp1 && cur_ops->exp_sync) {
1267                 synctype[nsynctypes++] = RTWS_EXP_SYNC;
1268                 pr_info("%s: Testing expedited GPs.\n", __func__);
1269         } else if (gp_exp && !cur_ops->exp_sync) {
1270                 pr_alert("%s: gp_exp without primitives.\n", __func__);
1271         }
1272         if (gp_normal1 && cur_ops->deferred_free) {
1273                 synctype[nsynctypes++] = RTWS_DEF_FREE;
1274                 pr_info("%s: Testing asynchronous GPs.\n", __func__);
1275         } else if (gp_normal && !cur_ops->deferred_free) {
1276                 pr_alert("%s: gp_normal without primitives.\n", __func__);
1277         }
1278         if (gp_poll1 && cur_ops->get_comp_state && cur_ops->same_gp_state &&
1279             cur_ops->start_gp_poll && cur_ops->poll_gp_state) {
1280                 synctype[nsynctypes++] = RTWS_POLL_GET;
1281                 pr_info("%s: Testing polling GPs.\n", __func__);
1282         } else if (gp_poll && (!cur_ops->start_gp_poll || !cur_ops->poll_gp_state)) {
1283                 pr_alert("%s: gp_poll without primitives.\n", __func__);
1284         }
1285         if (gp_poll_full1 && cur_ops->get_comp_state_full && cur_ops->same_gp_state_full
1286             && cur_ops->start_gp_poll_full && cur_ops->poll_gp_state_full) {
1287                 synctype[nsynctypes++] = RTWS_POLL_GET_FULL;
1288                 pr_info("%s: Testing polling full-state GPs.\n", __func__);
1289         } else if (gp_poll_full && (!cur_ops->start_gp_poll_full || !cur_ops->poll_gp_state_full)) {
1290                 pr_alert("%s: gp_poll_full without primitives.\n", __func__);
1291         }
1292         if (gp_poll_exp1 && cur_ops->start_gp_poll_exp && cur_ops->poll_gp_state_exp) {
1293                 synctype[nsynctypes++] = RTWS_POLL_GET_EXP;
1294                 pr_info("%s: Testing polling expedited GPs.\n", __func__);
1295         } else if (gp_poll_exp && (!cur_ops->start_gp_poll_exp || !cur_ops->poll_gp_state_exp)) {
1296                 pr_alert("%s: gp_poll_exp without primitives.\n", __func__);
1297         }
1298         if (gp_poll_exp_full1 && cur_ops->start_gp_poll_exp_full && cur_ops->poll_gp_state_full) {
1299                 synctype[nsynctypes++] = RTWS_POLL_GET_EXP_FULL;
1300                 pr_info("%s: Testing polling full-state expedited GPs.\n", __func__);
1301         } else if (gp_poll_exp_full &&
1302                    (!cur_ops->start_gp_poll_exp_full || !cur_ops->poll_gp_state_full)) {
1303                 pr_alert("%s: gp_poll_exp_full without primitives.\n", __func__);
1304         }
1305         if (gp_sync1 && cur_ops->sync) {
1306                 synctype[nsynctypes++] = RTWS_SYNC;
1307                 pr_info("%s: Testing normal GPs.\n", __func__);
1308         } else if (gp_sync && !cur_ops->sync) {
1309                 pr_alert("%s: gp_sync without primitives.\n", __func__);
1310         }
1311 }
1312
1313 /*
1314  * Do the specified rcu_torture_writer() synchronous grace period,
1315  * while also testing out the polled APIs.  Note well that the single-CPU
1316  * grace-period optimizations must be accounted for.
1317  */
1318 static void do_rtws_sync(struct torture_random_state *trsp, void (*sync)(void))
1319 {
1320         unsigned long cookie;
1321         struct rcu_gp_oldstate cookie_full;
1322         bool dopoll;
1323         bool dopoll_full;
1324         unsigned long r = torture_random(trsp);
1325
1326         dopoll = cur_ops->get_gp_state && cur_ops->poll_gp_state && !(r & 0x300);
1327         dopoll_full = cur_ops->get_gp_state_full && cur_ops->poll_gp_state_full && !(r & 0xc00);
1328         if (dopoll || dopoll_full)
1329                 cpus_read_lock();
1330         if (dopoll)
1331                 cookie = cur_ops->get_gp_state();
1332         if (dopoll_full)
1333                 cur_ops->get_gp_state_full(&cookie_full);
1334         if (cur_ops->poll_need_2gp && cur_ops->poll_need_2gp(dopoll, dopoll_full))
1335                 sync();
1336         sync();
1337         WARN_ONCE(dopoll && !cur_ops->poll_gp_state(cookie),
1338                   "%s: Cookie check 3 failed %pS() online %*pbl.",
1339                   __func__, sync, cpumask_pr_args(cpu_online_mask));
1340         WARN_ONCE(dopoll_full && !cur_ops->poll_gp_state_full(&cookie_full),
1341                   "%s: Cookie check 4 failed %pS() online %*pbl",
1342                   __func__, sync, cpumask_pr_args(cpu_online_mask));
1343         if (dopoll || dopoll_full)
1344                 cpus_read_unlock();
1345 }
1346
1347 /*
1348  * RCU torture writer kthread.  Repeatedly substitutes a new structure
1349  * for that pointed to by rcu_torture_current, freeing the old structure
1350  * after a series of grace periods (the "pipeline").
1351  */
1352 static int
1353 rcu_torture_writer(void *arg)
1354 {
1355         bool boot_ended;
1356         bool can_expedite = !rcu_gp_is_expedited() && !rcu_gp_is_normal();
1357         unsigned long cookie;
1358         struct rcu_gp_oldstate cookie_full;
1359         int expediting = 0;
1360         unsigned long gp_snap;
1361         unsigned long gp_snap1;
1362         struct rcu_gp_oldstate gp_snap_full;
1363         struct rcu_gp_oldstate gp_snap1_full;
1364         int i;
1365         int idx;
1366         int oldnice = task_nice(current);
1367         struct rcu_gp_oldstate rgo[NUM_ACTIVE_RCU_POLL_FULL_OLDSTATE];
1368         struct rcu_torture *rp;
1369         struct rcu_torture *old_rp;
1370         static DEFINE_TORTURE_RANDOM(rand);
1371         unsigned long stallsdone = jiffies;
1372         bool stutter_waited;
1373         unsigned long ulo[NUM_ACTIVE_RCU_POLL_OLDSTATE];
1374
1375         // If a new stall test is added, this must be adjusted.
1376         if (stall_cpu_holdoff + stall_gp_kthread + stall_cpu)
1377                 stallsdone += (stall_cpu_holdoff + stall_gp_kthread + stall_cpu + 60) * HZ;
1378         VERBOSE_TOROUT_STRING("rcu_torture_writer task started");
1379         if (!can_expedite)
1380                 pr_alert("%s" TORTURE_FLAG
1381                          " GP expediting controlled from boot/sysfs for %s.\n",
1382                          torture_type, cur_ops->name);
1383         if (WARN_ONCE(nsynctypes == 0,
1384                       "%s: No update-side primitives.\n", __func__)) {
1385                 /*
1386                  * No updates primitives, so don't try updating.
1387                  * The resulting test won't be testing much, hence the
1388                  * above WARN_ONCE().
1389                  */
1390                 rcu_torture_writer_state = RTWS_STOPPING;
1391                 torture_kthread_stopping("rcu_torture_writer");
1392                 return 0;
1393         }
1394
1395         do {
1396                 rcu_torture_writer_state = RTWS_FIXED_DELAY;
1397                 torture_hrtimeout_us(500, 1000, &rand);
1398                 rp = rcu_torture_alloc();
1399                 if (rp == NULL)
1400                         continue;
1401                 rp->rtort_pipe_count = 0;
1402                 rcu_torture_writer_state = RTWS_DELAY;
1403                 udelay(torture_random(&rand) & 0x3ff);
1404                 rcu_torture_writer_state = RTWS_REPLACE;
1405                 old_rp = rcu_dereference_check(rcu_torture_current,
1406                                                current == writer_task);
1407                 rp->rtort_mbtest = 1;
1408                 rcu_assign_pointer(rcu_torture_current, rp);
1409                 smp_wmb(); /* Mods to old_rp must follow rcu_assign_pointer() */
1410                 if (old_rp) {
1411                         i = old_rp->rtort_pipe_count;
1412                         if (i > RCU_TORTURE_PIPE_LEN)
1413                                 i = RCU_TORTURE_PIPE_LEN;
1414                         atomic_inc(&rcu_torture_wcount[i]);
1415                         WRITE_ONCE(old_rp->rtort_pipe_count,
1416                                    old_rp->rtort_pipe_count + 1);
1417
1418                         // Make sure readers block polled grace periods.
1419                         if (cur_ops->get_gp_state && cur_ops->poll_gp_state) {
1420                                 idx = cur_ops->readlock();
1421                                 cookie = cur_ops->get_gp_state();
1422                                 WARN_ONCE(cur_ops->poll_gp_state(cookie),
1423                                           "%s: Cookie check 1 failed %s(%d) %lu->%lu\n",
1424                                           __func__,
1425                                           rcu_torture_writer_state_getname(),
1426                                           rcu_torture_writer_state,
1427                                           cookie, cur_ops->get_gp_state());
1428                                 if (cur_ops->get_gp_completed) {
1429                                         cookie = cur_ops->get_gp_completed();
1430                                         WARN_ON_ONCE(!cur_ops->poll_gp_state(cookie));
1431                                 }
1432                                 cur_ops->readunlock(idx);
1433                         }
1434                         if (cur_ops->get_gp_state_full && cur_ops->poll_gp_state_full) {
1435                                 idx = cur_ops->readlock();
1436                                 cur_ops->get_gp_state_full(&cookie_full);
1437                                 WARN_ONCE(cur_ops->poll_gp_state_full(&cookie_full),
1438                                           "%s: Cookie check 5 failed %s(%d) online %*pbl\n",
1439                                           __func__,
1440                                           rcu_torture_writer_state_getname(),
1441                                           rcu_torture_writer_state,
1442                                           cpumask_pr_args(cpu_online_mask));
1443                                 if (cur_ops->get_gp_completed_full) {
1444                                         cur_ops->get_gp_completed_full(&cookie_full);
1445                                         WARN_ON_ONCE(!cur_ops->poll_gp_state_full(&cookie_full));
1446                                 }
1447                                 cur_ops->readunlock(idx);
1448                         }
1449                         switch (synctype[torture_random(&rand) % nsynctypes]) {
1450                         case RTWS_DEF_FREE:
1451                                 rcu_torture_writer_state = RTWS_DEF_FREE;
1452                                 cur_ops->deferred_free(old_rp);
1453                                 break;
1454                         case RTWS_EXP_SYNC:
1455                                 rcu_torture_writer_state = RTWS_EXP_SYNC;
1456                                 do_rtws_sync(&rand, cur_ops->exp_sync);
1457                                 rcu_torture_pipe_update(old_rp);
1458                                 break;
1459                         case RTWS_COND_GET:
1460                                 rcu_torture_writer_state = RTWS_COND_GET;
1461                                 gp_snap = cur_ops->get_gp_state();
1462                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1463                                 rcu_torture_writer_state = RTWS_COND_SYNC;
1464                                 cur_ops->cond_sync(gp_snap);
1465                                 rcu_torture_pipe_update(old_rp);
1466                                 break;
1467                         case RTWS_COND_GET_EXP:
1468                                 rcu_torture_writer_state = RTWS_COND_GET_EXP;
1469                                 gp_snap = cur_ops->get_gp_state_exp();
1470                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1471                                 rcu_torture_writer_state = RTWS_COND_SYNC_EXP;
1472                                 cur_ops->cond_sync_exp(gp_snap);
1473                                 rcu_torture_pipe_update(old_rp);
1474                                 break;
1475                         case RTWS_COND_GET_FULL:
1476                                 rcu_torture_writer_state = RTWS_COND_GET_FULL;
1477                                 cur_ops->get_gp_state_full(&gp_snap_full);
1478                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1479                                 rcu_torture_writer_state = RTWS_COND_SYNC_FULL;
1480                                 cur_ops->cond_sync_full(&gp_snap_full);
1481                                 rcu_torture_pipe_update(old_rp);
1482                                 break;
1483                         case RTWS_COND_GET_EXP_FULL:
1484                                 rcu_torture_writer_state = RTWS_COND_GET_EXP_FULL;
1485                                 cur_ops->get_gp_state_full(&gp_snap_full);
1486                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1487                                 rcu_torture_writer_state = RTWS_COND_SYNC_EXP_FULL;
1488                                 cur_ops->cond_sync_exp_full(&gp_snap_full);
1489                                 rcu_torture_pipe_update(old_rp);
1490                                 break;
1491                         case RTWS_POLL_GET:
1492                                 rcu_torture_writer_state = RTWS_POLL_GET;
1493                                 for (i = 0; i < ARRAY_SIZE(ulo); i++)
1494                                         ulo[i] = cur_ops->get_comp_state();
1495                                 gp_snap = cur_ops->start_gp_poll();
1496                                 rcu_torture_writer_state = RTWS_POLL_WAIT;
1497                                 while (!cur_ops->poll_gp_state(gp_snap)) {
1498                                         gp_snap1 = cur_ops->get_gp_state();
1499                                         for (i = 0; i < ARRAY_SIZE(ulo); i++)
1500                                                 if (cur_ops->poll_gp_state(ulo[i]) ||
1501                                                     cur_ops->same_gp_state(ulo[i], gp_snap1)) {
1502                                                         ulo[i] = gp_snap1;
1503                                                         break;
1504                                                 }
1505                                         WARN_ON_ONCE(i >= ARRAY_SIZE(ulo));
1506                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1507                                                                   &rand);
1508                                 }
1509                                 rcu_torture_pipe_update(old_rp);
1510                                 break;
1511                         case RTWS_POLL_GET_FULL:
1512                                 rcu_torture_writer_state = RTWS_POLL_GET_FULL;
1513                                 for (i = 0; i < ARRAY_SIZE(rgo); i++)
1514                                         cur_ops->get_comp_state_full(&rgo[i]);
1515                                 cur_ops->start_gp_poll_full(&gp_snap_full);
1516                                 rcu_torture_writer_state = RTWS_POLL_WAIT_FULL;
1517                                 while (!cur_ops->poll_gp_state_full(&gp_snap_full)) {
1518                                         cur_ops->get_gp_state_full(&gp_snap1_full);
1519                                         for (i = 0; i < ARRAY_SIZE(rgo); i++)
1520                                                 if (cur_ops->poll_gp_state_full(&rgo[i]) ||
1521                                                     cur_ops->same_gp_state_full(&rgo[i],
1522                                                                                 &gp_snap1_full)) {
1523                                                         rgo[i] = gp_snap1_full;
1524                                                         break;
1525                                                 }
1526                                         WARN_ON_ONCE(i >= ARRAY_SIZE(rgo));
1527                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1528                                                                   &rand);
1529                                 }
1530                                 rcu_torture_pipe_update(old_rp);
1531                                 break;
1532                         case RTWS_POLL_GET_EXP:
1533                                 rcu_torture_writer_state = RTWS_POLL_GET_EXP;
1534                                 gp_snap = cur_ops->start_gp_poll_exp();
1535                                 rcu_torture_writer_state = RTWS_POLL_WAIT_EXP;
1536                                 while (!cur_ops->poll_gp_state_exp(gp_snap))
1537                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1538                                                                   &rand);
1539                                 rcu_torture_pipe_update(old_rp);
1540                                 break;
1541                         case RTWS_POLL_GET_EXP_FULL:
1542                                 rcu_torture_writer_state = RTWS_POLL_GET_EXP_FULL;
1543                                 cur_ops->start_gp_poll_exp_full(&gp_snap_full);
1544                                 rcu_torture_writer_state = RTWS_POLL_WAIT_EXP_FULL;
1545                                 while (!cur_ops->poll_gp_state_full(&gp_snap_full))
1546                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1547                                                                   &rand);
1548                                 rcu_torture_pipe_update(old_rp);
1549                                 break;
1550                         case RTWS_SYNC:
1551                                 rcu_torture_writer_state = RTWS_SYNC;
1552                                 do_rtws_sync(&rand, cur_ops->sync);
1553                                 rcu_torture_pipe_update(old_rp);
1554                                 break;
1555                         default:
1556                                 WARN_ON_ONCE(1);
1557                                 break;
1558                         }
1559                 }
1560                 WRITE_ONCE(rcu_torture_current_version,
1561                            rcu_torture_current_version + 1);
1562                 /* Cycle through nesting levels of rcu_expedite_gp() calls. */
1563                 if (can_expedite &&
1564                     !(torture_random(&rand) & 0xff & (!!expediting - 1))) {
1565                         WARN_ON_ONCE(expediting == 0 && rcu_gp_is_expedited());
1566                         if (expediting >= 0)
1567                                 rcu_expedite_gp();
1568                         else
1569                                 rcu_unexpedite_gp();
1570                         if (++expediting > 3)
1571                                 expediting = -expediting;
1572                 } else if (!can_expedite) { /* Disabled during boot, recheck. */
1573                         can_expedite = !rcu_gp_is_expedited() &&
1574                                        !rcu_gp_is_normal();
1575                 }
1576                 rcu_torture_writer_state = RTWS_STUTTER;
1577                 boot_ended = rcu_inkernel_boot_has_ended();
1578                 stutter_waited = stutter_wait("rcu_torture_writer");
1579                 if (stutter_waited &&
1580                     !atomic_read(&rcu_fwd_cb_nodelay) &&
1581                     !cur_ops->slow_gps &&
1582                     !torture_must_stop() &&
1583                     boot_ended &&
1584                     time_after(jiffies, stallsdone))
1585                         for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++)
1586                                 if (list_empty(&rcu_tortures[i].rtort_free) &&
1587                                     rcu_access_pointer(rcu_torture_current) != &rcu_tortures[i]) {
1588                                         tracing_off();
1589                                         show_rcu_gp_kthreads();
1590                                         WARN(1, "%s: rtort_pipe_count: %d\n", __func__, rcu_tortures[i].rtort_pipe_count);
1591                                         rcu_ftrace_dump(DUMP_ALL);
1592                                 }
1593                 if (stutter_waited)
1594                         sched_set_normal(current, oldnice);
1595         } while (!torture_must_stop());
1596         rcu_torture_current = NULL;  // Let stats task know that we are done.
1597         /* Reset expediting back to unexpedited. */
1598         if (expediting > 0)
1599                 expediting = -expediting;
1600         while (can_expedite && expediting++ < 0)
1601                 rcu_unexpedite_gp();
1602         WARN_ON_ONCE(can_expedite && rcu_gp_is_expedited());
1603         if (!can_expedite)
1604                 pr_alert("%s" TORTURE_FLAG
1605                          " Dynamic grace-period expediting was disabled.\n",
1606                          torture_type);
1607         rcu_torture_writer_state = RTWS_STOPPING;
1608         torture_kthread_stopping("rcu_torture_writer");
1609         return 0;
1610 }
1611
1612 /*
1613  * RCU torture fake writer kthread.  Repeatedly calls sync, with a random
1614  * delay between calls.
1615  */
1616 static int
1617 rcu_torture_fakewriter(void *arg)
1618 {
1619         unsigned long gp_snap;
1620         struct rcu_gp_oldstate gp_snap_full;
1621         DEFINE_TORTURE_RANDOM(rand);
1622
1623         VERBOSE_TOROUT_STRING("rcu_torture_fakewriter task started");
1624         set_user_nice(current, MAX_NICE);
1625
1626         if (WARN_ONCE(nsynctypes == 0,
1627                       "%s: No update-side primitives.\n", __func__)) {
1628                 /*
1629                  * No updates primitives, so don't try updating.
1630                  * The resulting test won't be testing much, hence the
1631                  * above WARN_ONCE().
1632                  */
1633                 torture_kthread_stopping("rcu_torture_fakewriter");
1634                 return 0;
1635         }
1636
1637         do {
1638                 torture_hrtimeout_jiffies(torture_random(&rand) % 10, &rand);
1639                 if (cur_ops->cb_barrier != NULL &&
1640                     torture_random(&rand) % (nfakewriters * 8) == 0) {
1641                         cur_ops->cb_barrier();
1642                 } else {
1643                         switch (synctype[torture_random(&rand) % nsynctypes]) {
1644                         case RTWS_DEF_FREE:
1645                                 break;
1646                         case RTWS_EXP_SYNC:
1647                                 cur_ops->exp_sync();
1648                                 break;
1649                         case RTWS_COND_GET:
1650                                 gp_snap = cur_ops->get_gp_state();
1651                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1652                                 cur_ops->cond_sync(gp_snap);
1653                                 break;
1654                         case RTWS_COND_GET_EXP:
1655                                 gp_snap = cur_ops->get_gp_state_exp();
1656                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1657                                 cur_ops->cond_sync_exp(gp_snap);
1658                                 break;
1659                         case RTWS_COND_GET_FULL:
1660                                 cur_ops->get_gp_state_full(&gp_snap_full);
1661                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1662                                 cur_ops->cond_sync_full(&gp_snap_full);
1663                                 break;
1664                         case RTWS_COND_GET_EXP_FULL:
1665                                 cur_ops->get_gp_state_full(&gp_snap_full);
1666                                 torture_hrtimeout_jiffies(torture_random(&rand) % 16, &rand);
1667                                 cur_ops->cond_sync_exp_full(&gp_snap_full);
1668                                 break;
1669                         case RTWS_POLL_GET:
1670                                 gp_snap = cur_ops->start_gp_poll();
1671                                 while (!cur_ops->poll_gp_state(gp_snap)) {
1672                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1673                                                                   &rand);
1674                                 }
1675                                 break;
1676                         case RTWS_POLL_GET_FULL:
1677                                 cur_ops->start_gp_poll_full(&gp_snap_full);
1678                                 while (!cur_ops->poll_gp_state_full(&gp_snap_full)) {
1679                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1680                                                                   &rand);
1681                                 }
1682                                 break;
1683                         case RTWS_POLL_GET_EXP:
1684                                 gp_snap = cur_ops->start_gp_poll_exp();
1685                                 while (!cur_ops->poll_gp_state_exp(gp_snap)) {
1686                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1687                                                                   &rand);
1688                                 }
1689                                 break;
1690                         case RTWS_POLL_GET_EXP_FULL:
1691                                 cur_ops->start_gp_poll_exp_full(&gp_snap_full);
1692                                 while (!cur_ops->poll_gp_state_full(&gp_snap_full)) {
1693                                         torture_hrtimeout_jiffies(torture_random(&rand) % 16,
1694                                                                   &rand);
1695                                 }
1696                                 break;
1697                         case RTWS_SYNC:
1698                                 cur_ops->sync();
1699                                 break;
1700                         default:
1701                                 WARN_ON_ONCE(1);
1702                                 break;
1703                         }
1704                 }
1705                 stutter_wait("rcu_torture_fakewriter");
1706         } while (!torture_must_stop());
1707
1708         torture_kthread_stopping("rcu_torture_fakewriter");
1709         return 0;
1710 }
1711
1712 static void rcu_torture_timer_cb(struct rcu_head *rhp)
1713 {
1714         kfree(rhp);
1715 }
1716
1717 // Set up and carry out testing of RCU's global memory ordering
1718 static void rcu_torture_reader_do_mbchk(long myid, struct rcu_torture *rtp,
1719                                         struct torture_random_state *trsp)
1720 {
1721         unsigned long loops;
1722         int noc = torture_num_online_cpus();
1723         int rdrchked;
1724         int rdrchker;
1725         struct rcu_torture_reader_check *rtrcp; // Me.
1726         struct rcu_torture_reader_check *rtrcp_assigner; // Assigned us to do checking.
1727         struct rcu_torture_reader_check *rtrcp_chked; // Reader being checked.
1728         struct rcu_torture_reader_check *rtrcp_chker; // Reader doing checking when not me.
1729
1730         if (myid < 0)
1731                 return; // Don't try this from timer handlers.
1732
1733         // Increment my counter.
1734         rtrcp = &rcu_torture_reader_mbchk[myid];
1735         WRITE_ONCE(rtrcp->rtc_myloops, rtrcp->rtc_myloops + 1);
1736
1737         // Attempt to assign someone else some checking work.
1738         rdrchked = torture_random(trsp) % nrealreaders;
1739         rtrcp_chked = &rcu_torture_reader_mbchk[rdrchked];
1740         rdrchker = torture_random(trsp) % nrealreaders;
1741         rtrcp_chker = &rcu_torture_reader_mbchk[rdrchker];
1742         if (rdrchked != myid && rdrchked != rdrchker && noc >= rdrchked && noc >= rdrchker &&
1743             smp_load_acquire(&rtrcp->rtc_chkrdr) < 0 && // Pairs with smp_store_release below.
1744             !READ_ONCE(rtp->rtort_chkp) &&
1745             !smp_load_acquire(&rtrcp_chker->rtc_assigner)) { // Pairs with smp_store_release below.
1746                 rtrcp->rtc_chkloops = READ_ONCE(rtrcp_chked->rtc_myloops);
1747                 WARN_ON_ONCE(rtrcp->rtc_chkrdr >= 0);
1748                 rtrcp->rtc_chkrdr = rdrchked;
1749                 WARN_ON_ONCE(rtrcp->rtc_ready); // This gets set after the grace period ends.
1750                 if (cmpxchg_relaxed(&rtrcp_chker->rtc_assigner, NULL, rtrcp) ||
1751                     cmpxchg_relaxed(&rtp->rtort_chkp, NULL, rtrcp))
1752                         (void)cmpxchg_relaxed(&rtrcp_chker->rtc_assigner, rtrcp, NULL); // Back out.
1753         }
1754
1755         // If assigned some completed work, do it!
1756         rtrcp_assigner = READ_ONCE(rtrcp->rtc_assigner);
1757         if (!rtrcp_assigner || !smp_load_acquire(&rtrcp_assigner->rtc_ready))
1758                 return; // No work or work not yet ready.
1759         rdrchked = rtrcp_assigner->rtc_chkrdr;
1760         if (WARN_ON_ONCE(rdrchked < 0))
1761                 return;
1762         rtrcp_chked = &rcu_torture_reader_mbchk[rdrchked];
1763         loops = READ_ONCE(rtrcp_chked->rtc_myloops);
1764         atomic_inc(&n_rcu_torture_mbchk_tries);
1765         if (ULONG_CMP_LT(loops, rtrcp_assigner->rtc_chkloops))
1766                 atomic_inc(&n_rcu_torture_mbchk_fail);
1767         rtrcp_assigner->rtc_chkloops = loops + ULONG_MAX / 2;
1768         rtrcp_assigner->rtc_ready = 0;
1769         smp_store_release(&rtrcp->rtc_assigner, NULL); // Someone else can assign us work.
1770         smp_store_release(&rtrcp_assigner->rtc_chkrdr, -1); // Assigner can again assign.
1771 }
1772
1773 /*
1774  * Do one extension of an RCU read-side critical section using the
1775  * current reader state in readstate (set to zero for initial entry
1776  * to extended critical section), set the new state as specified by
1777  * newstate (set to zero for final exit from extended critical section),
1778  * and random-number-generator state in trsp.  If this is neither the
1779  * beginning or end of the critical section and if there was actually a
1780  * change, do a ->read_delay().
1781  */
1782 static void rcutorture_one_extend(int *readstate, int newstate,
1783                                   struct torture_random_state *trsp,
1784                                   struct rt_read_seg *rtrsp)
1785 {
1786         unsigned long flags;
1787         int idxnew1 = -1;
1788         int idxnew2 = -1;
1789         int idxold1 = *readstate;
1790         int idxold2 = idxold1;
1791         int statesnew = ~*readstate & newstate;
1792         int statesold = *readstate & ~newstate;
1793
1794         WARN_ON_ONCE(idxold2 < 0);
1795         WARN_ON_ONCE((idxold2 >> RCUTORTURE_RDR_SHIFT_2) > 1);
1796         rtrsp->rt_readstate = newstate;
1797
1798         /* First, put new protection in place to avoid critical-section gap. */
1799         if (statesnew & RCUTORTURE_RDR_BH)
1800                 local_bh_disable();
1801         if (statesnew & RCUTORTURE_RDR_RBH)
1802                 rcu_read_lock_bh();
1803         if (statesnew & RCUTORTURE_RDR_IRQ)
1804                 local_irq_disable();
1805         if (statesnew & RCUTORTURE_RDR_PREEMPT)
1806                 preempt_disable();
1807         if (statesnew & RCUTORTURE_RDR_SCHED)
1808                 rcu_read_lock_sched();
1809         if (statesnew & RCUTORTURE_RDR_RCU_1)
1810                 idxnew1 = (cur_ops->readlock() & 0x1) << RCUTORTURE_RDR_SHIFT_1;
1811         if (statesnew & RCUTORTURE_RDR_RCU_2)
1812                 idxnew2 = (cur_ops->readlock() & 0x1) << RCUTORTURE_RDR_SHIFT_2;
1813
1814         /*
1815          * Next, remove old protection, in decreasing order of strength
1816          * to avoid unlock paths that aren't safe in the stronger
1817          * context. Namely: BH can not be enabled with disabled interrupts.
1818          * Additionally PREEMPT_RT requires that BH is enabled in preemptible
1819          * context.
1820          */
1821         if (statesold & RCUTORTURE_RDR_IRQ)
1822                 local_irq_enable();
1823         if (statesold & RCUTORTURE_RDR_PREEMPT)
1824                 preempt_enable();
1825         if (statesold & RCUTORTURE_RDR_SCHED)
1826                 rcu_read_unlock_sched();
1827         if (statesold & RCUTORTURE_RDR_BH)
1828                 local_bh_enable();
1829         if (statesold & RCUTORTURE_RDR_RBH)
1830                 rcu_read_unlock_bh();
1831         if (statesold & RCUTORTURE_RDR_RCU_2) {
1832                 cur_ops->readunlock((idxold2 >> RCUTORTURE_RDR_SHIFT_2) & 0x1);
1833                 WARN_ON_ONCE(idxnew2 != -1);
1834                 idxold2 = 0;
1835         }
1836         if (statesold & RCUTORTURE_RDR_RCU_1) {
1837                 bool lockit;
1838
1839                 lockit = !cur_ops->no_pi_lock && !statesnew && !(torture_random(trsp) & 0xffff);
1840                 if (lockit)
1841                         raw_spin_lock_irqsave(&current->pi_lock, flags);
1842                 cur_ops->readunlock((idxold1 >> RCUTORTURE_RDR_SHIFT_1) & 0x1);
1843                 WARN_ON_ONCE(idxnew1 != -1);
1844                 idxold1 = 0;
1845                 if (lockit)
1846                         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
1847         }
1848
1849         /* Delay if neither beginning nor end and there was a change. */
1850         if ((statesnew || statesold) && *readstate && newstate)
1851                 cur_ops->read_delay(trsp, rtrsp);
1852
1853         /* Update the reader state. */
1854         if (idxnew1 == -1)
1855                 idxnew1 = idxold1 & RCUTORTURE_RDR_MASK_1;
1856         WARN_ON_ONCE(idxnew1 < 0);
1857         if (WARN_ON_ONCE((idxnew1 >> RCUTORTURE_RDR_SHIFT_1) > 1))
1858                 pr_info("Unexpected idxnew1 value of %#x\n", idxnew1);
1859         if (idxnew2 == -1)
1860                 idxnew2 = idxold2 & RCUTORTURE_RDR_MASK_2;
1861         WARN_ON_ONCE(idxnew2 < 0);
1862         WARN_ON_ONCE((idxnew2 >> RCUTORTURE_RDR_SHIFT_2) > 1);
1863         *readstate = idxnew1 | idxnew2 | newstate;
1864         WARN_ON_ONCE(*readstate < 0);
1865         if (WARN_ON_ONCE((*readstate >> RCUTORTURE_RDR_SHIFT_2) > 1))
1866                 pr_info("Unexpected idxnew2 value of %#x\n", idxnew2);
1867 }
1868
1869 /* Return the biggest extendables mask given current RCU and boot parameters. */
1870 static int rcutorture_extend_mask_max(void)
1871 {
1872         int mask;
1873
1874         WARN_ON_ONCE(extendables & ~RCUTORTURE_MAX_EXTEND);
1875         mask = extendables & RCUTORTURE_MAX_EXTEND & cur_ops->extendables;
1876         mask = mask | RCUTORTURE_RDR_RCU_1 | RCUTORTURE_RDR_RCU_2;
1877         return mask;
1878 }
1879
1880 /* Return a random protection state mask, but with at least one bit set. */
1881 static int
1882 rcutorture_extend_mask(int oldmask, struct torture_random_state *trsp)
1883 {
1884         int mask = rcutorture_extend_mask_max();
1885         unsigned long randmask1 = torture_random(trsp);
1886         unsigned long randmask2 = randmask1 >> 3;
1887         unsigned long preempts = RCUTORTURE_RDR_PREEMPT | RCUTORTURE_RDR_SCHED;
1888         unsigned long preempts_irq = preempts | RCUTORTURE_RDR_IRQ;
1889         unsigned long bhs = RCUTORTURE_RDR_BH | RCUTORTURE_RDR_RBH;
1890
1891         WARN_ON_ONCE(mask >> RCUTORTURE_RDR_SHIFT_1);
1892         /* Mostly only one bit (need preemption!), sometimes lots of bits. */
1893         if (!(randmask1 & 0x7))
1894                 mask = mask & randmask2;
1895         else
1896                 mask = mask & (1 << (randmask2 % RCUTORTURE_RDR_NBITS));
1897
1898         // Can't have nested RCU reader without outer RCU reader.
1899         if (!(mask & RCUTORTURE_RDR_RCU_1) && (mask & RCUTORTURE_RDR_RCU_2)) {
1900                 if (oldmask & RCUTORTURE_RDR_RCU_1)
1901                         mask &= ~RCUTORTURE_RDR_RCU_2;
1902                 else
1903                         mask |= RCUTORTURE_RDR_RCU_1;
1904         }
1905
1906         /*
1907          * Can't enable bh w/irq disabled.
1908          */
1909         if (mask & RCUTORTURE_RDR_IRQ)
1910                 mask |= oldmask & bhs;
1911
1912         /*
1913          * Ideally these sequences would be detected in debug builds
1914          * (regardless of RT), but until then don't stop testing
1915          * them on non-RT.
1916          */
1917         if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
1918                 /* Can't modify BH in atomic context */
1919                 if (oldmask & preempts_irq)
1920                         mask &= ~bhs;
1921                 if ((oldmask | mask) & preempts_irq)
1922                         mask |= oldmask & bhs;
1923         }
1924
1925         return mask ?: RCUTORTURE_RDR_RCU_1;
1926 }
1927
1928 /*
1929  * Do a randomly selected number of extensions of an existing RCU read-side
1930  * critical section.
1931  */
1932 static struct rt_read_seg *
1933 rcutorture_loop_extend(int *readstate, struct torture_random_state *trsp,
1934                        struct rt_read_seg *rtrsp)
1935 {
1936         int i;
1937         int j;
1938         int mask = rcutorture_extend_mask_max();
1939
1940         WARN_ON_ONCE(!*readstate); /* -Existing- RCU read-side critsect! */
1941         if (!((mask - 1) & mask))
1942                 return rtrsp;  /* Current RCU reader not extendable. */
1943         /* Bias towards larger numbers of loops. */
1944         i = torture_random(trsp);
1945         i = ((i | (i >> 3)) & RCUTORTURE_RDR_MAX_LOOPS) + 1;
1946         for (j = 0; j < i; j++) {
1947                 mask = rcutorture_extend_mask(*readstate, trsp);
1948                 rcutorture_one_extend(readstate, mask, trsp, &rtrsp[j]);
1949         }
1950         return &rtrsp[j];
1951 }
1952
1953 /*
1954  * Do one read-side critical section, returning false if there was
1955  * no data to read.  Can be invoked both from process context and
1956  * from a timer handler.
1957  */
1958 static bool rcu_torture_one_read(struct torture_random_state *trsp, long myid)
1959 {
1960         bool checkpolling = !(torture_random(trsp) & 0xfff);
1961         unsigned long cookie;
1962         struct rcu_gp_oldstate cookie_full;
1963         int i;
1964         unsigned long started;
1965         unsigned long completed;
1966         int newstate;
1967         struct rcu_torture *p;
1968         int pipe_count;
1969         int readstate = 0;
1970         struct rt_read_seg rtseg[RCUTORTURE_RDR_MAX_SEGS] = { { 0 } };
1971         struct rt_read_seg *rtrsp = &rtseg[0];
1972         struct rt_read_seg *rtrsp1;
1973         unsigned long long ts;
1974
1975         WARN_ON_ONCE(!rcu_is_watching());
1976         newstate = rcutorture_extend_mask(readstate, trsp);
1977         rcutorture_one_extend(&readstate, newstate, trsp, rtrsp++);
1978         if (checkpolling) {
1979                 if (cur_ops->get_gp_state && cur_ops->poll_gp_state)
1980                         cookie = cur_ops->get_gp_state();
1981                 if (cur_ops->get_gp_state_full && cur_ops->poll_gp_state_full)
1982                         cur_ops->get_gp_state_full(&cookie_full);
1983         }
1984         started = cur_ops->get_gp_seq();
1985         ts = rcu_trace_clock_local();
1986         p = rcu_dereference_check(rcu_torture_current,
1987                                   !cur_ops->readlock_held || cur_ops->readlock_held());
1988         if (p == NULL) {
1989                 /* Wait for rcu_torture_writer to get underway */
1990                 rcutorture_one_extend(&readstate, 0, trsp, rtrsp);
1991                 return false;
1992         }
1993         if (p->rtort_mbtest == 0)
1994                 atomic_inc(&n_rcu_torture_mberror);
1995         rcu_torture_reader_do_mbchk(myid, p, trsp);
1996         rtrsp = rcutorture_loop_extend(&readstate, trsp, rtrsp);
1997         preempt_disable();
1998         pipe_count = READ_ONCE(p->rtort_pipe_count);
1999         if (pipe_count > RCU_TORTURE_PIPE_LEN) {
2000                 /* Should not happen, but... */
2001                 pipe_count = RCU_TORTURE_PIPE_LEN;
2002         }
2003         completed = cur_ops->get_gp_seq();
2004         if (pipe_count > 1) {
2005                 do_trace_rcu_torture_read(cur_ops->name, &p->rtort_rcu,
2006                                           ts, started, completed);
2007                 rcu_ftrace_dump(DUMP_ALL);
2008         }
2009         __this_cpu_inc(rcu_torture_count[pipe_count]);
2010         completed = rcutorture_seq_diff(completed, started);
2011         if (completed > RCU_TORTURE_PIPE_LEN) {
2012                 /* Should not happen, but... */
2013                 completed = RCU_TORTURE_PIPE_LEN;
2014         }
2015         __this_cpu_inc(rcu_torture_batch[completed]);
2016         preempt_enable();
2017         if (checkpolling) {
2018                 if (cur_ops->get_gp_state && cur_ops->poll_gp_state)
2019                         WARN_ONCE(cur_ops->poll_gp_state(cookie),
2020                                   "%s: Cookie check 2 failed %s(%d) %lu->%lu\n",
2021                                   __func__,
2022                                   rcu_torture_writer_state_getname(),
2023                                   rcu_torture_writer_state,
2024                                   cookie, cur_ops->get_gp_state());
2025                 if (cur_ops->get_gp_state_full && cur_ops->poll_gp_state_full)
2026                         WARN_ONCE(cur_ops->poll_gp_state_full(&cookie_full),
2027                                   "%s: Cookie check 6 failed %s(%d) online %*pbl\n",
2028                                   __func__,
2029                                   rcu_torture_writer_state_getname(),
2030                                   rcu_torture_writer_state,
2031                                   cpumask_pr_args(cpu_online_mask));
2032         }
2033         rcutorture_one_extend(&readstate, 0, trsp, rtrsp);
2034         WARN_ON_ONCE(readstate);
2035         // This next splat is expected behavior if leakpointer, especially
2036         // for CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels.
2037         WARN_ON_ONCE(leakpointer && READ_ONCE(p->rtort_pipe_count) > 1);
2038
2039         /* If error or close call, record the sequence of reader protections. */
2040         if ((pipe_count > 1 || completed > 1) && !xchg(&err_segs_recorded, 1)) {
2041                 i = 0;
2042                 for (rtrsp1 = &rtseg[0]; rtrsp1 < rtrsp; rtrsp1++)
2043                         err_segs[i++] = *rtrsp1;
2044                 rt_read_nsegs = i;
2045         }
2046
2047         return true;
2048 }
2049
2050 static DEFINE_TORTURE_RANDOM_PERCPU(rcu_torture_timer_rand);
2051
2052 /*
2053  * RCU torture reader from timer handler.  Dereferences rcu_torture_current,
2054  * incrementing the corresponding element of the pipeline array.  The
2055  * counter in the element should never be greater than 1, otherwise, the
2056  * RCU implementation is broken.
2057  */
2058 static void rcu_torture_timer(struct timer_list *unused)
2059 {
2060         atomic_long_inc(&n_rcu_torture_timers);
2061         (void)rcu_torture_one_read(this_cpu_ptr(&rcu_torture_timer_rand), -1);
2062
2063         /* Test call_rcu() invocation from interrupt handler. */
2064         if (cur_ops->call) {
2065                 struct rcu_head *rhp = kmalloc(sizeof(*rhp), GFP_NOWAIT);
2066
2067                 if (rhp)
2068                         cur_ops->call(rhp, rcu_torture_timer_cb);
2069         }
2070 }
2071
2072 /*
2073  * RCU torture reader kthread.  Repeatedly dereferences rcu_torture_current,
2074  * incrementing the corresponding element of the pipeline array.  The
2075  * counter in the element should never be greater than 1, otherwise, the
2076  * RCU implementation is broken.
2077  */
2078 static int
2079 rcu_torture_reader(void *arg)
2080 {
2081         unsigned long lastsleep = jiffies;
2082         long myid = (long)arg;
2083         int mynumonline = myid;
2084         DEFINE_TORTURE_RANDOM(rand);
2085         struct timer_list t;
2086
2087         VERBOSE_TOROUT_STRING("rcu_torture_reader task started");
2088         set_user_nice(current, MAX_NICE);
2089         if (irqreader && cur_ops->irq_capable)
2090                 timer_setup_on_stack(&t, rcu_torture_timer, 0);
2091         tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2092         do {
2093                 if (irqreader && cur_ops->irq_capable) {
2094                         if (!timer_pending(&t))
2095                                 mod_timer(&t, jiffies + 1);
2096                 }
2097                 if (!rcu_torture_one_read(&rand, myid) && !torture_must_stop())
2098                         schedule_timeout_interruptible(HZ);
2099                 if (time_after(jiffies, lastsleep) && !torture_must_stop()) {
2100                         torture_hrtimeout_us(500, 1000, &rand);
2101                         lastsleep = jiffies + 10;
2102                 }
2103                 while (torture_num_online_cpus() < mynumonline && !torture_must_stop())
2104                         schedule_timeout_interruptible(HZ / 5);
2105                 stutter_wait("rcu_torture_reader");
2106         } while (!torture_must_stop());
2107         if (irqreader && cur_ops->irq_capable) {
2108                 del_timer_sync(&t);
2109                 destroy_timer_on_stack(&t);
2110         }
2111         tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2112         torture_kthread_stopping("rcu_torture_reader");
2113         return 0;
2114 }
2115
2116 /*
2117  * Randomly Toggle CPUs' callback-offload state.  This uses hrtimers to
2118  * increase race probabilities and fuzzes the interval between toggling.
2119  */
2120 static int rcu_nocb_toggle(void *arg)
2121 {
2122         int cpu;
2123         int maxcpu = -1;
2124         int oldnice = task_nice(current);
2125         long r;
2126         DEFINE_TORTURE_RANDOM(rand);
2127         ktime_t toggle_delay;
2128         unsigned long toggle_fuzz;
2129         ktime_t toggle_interval = ms_to_ktime(nocbs_toggle);
2130
2131         VERBOSE_TOROUT_STRING("rcu_nocb_toggle task started");
2132         while (!rcu_inkernel_boot_has_ended())
2133                 schedule_timeout_interruptible(HZ / 10);
2134         for_each_possible_cpu(cpu)
2135                 maxcpu = cpu;
2136         WARN_ON(maxcpu < 0);
2137         if (toggle_interval > ULONG_MAX)
2138                 toggle_fuzz = ULONG_MAX >> 3;
2139         else
2140                 toggle_fuzz = toggle_interval >> 3;
2141         if (toggle_fuzz <= 0)
2142                 toggle_fuzz = NSEC_PER_USEC;
2143         do {
2144                 r = torture_random(&rand);
2145                 cpu = (r >> 1) % (maxcpu + 1);
2146                 if (r & 0x1) {
2147                         rcu_nocb_cpu_offload(cpu);
2148                         atomic_long_inc(&n_nocb_offload);
2149                 } else {
2150                         rcu_nocb_cpu_deoffload(cpu);
2151                         atomic_long_inc(&n_nocb_deoffload);
2152                 }
2153                 toggle_delay = torture_random(&rand) % toggle_fuzz + toggle_interval;
2154                 set_current_state(TASK_INTERRUPTIBLE);
2155                 schedule_hrtimeout(&toggle_delay, HRTIMER_MODE_REL);
2156                 if (stutter_wait("rcu_nocb_toggle"))
2157                         sched_set_normal(current, oldnice);
2158         } while (!torture_must_stop());
2159         torture_kthread_stopping("rcu_nocb_toggle");
2160         return 0;
2161 }
2162
2163 /*
2164  * Print torture statistics.  Caller must ensure that there is only
2165  * one call to this function at a given time!!!  This is normally
2166  * accomplished by relying on the module system to only have one copy
2167  * of the module loaded, and then by giving the rcu_torture_stats
2168  * kthread full control (or the init/cleanup functions when rcu_torture_stats
2169  * thread is not running).
2170  */
2171 static void
2172 rcu_torture_stats_print(void)
2173 {
2174         int cpu;
2175         int i;
2176         long pipesummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
2177         long batchsummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
2178         struct rcu_torture *rtcp;
2179         static unsigned long rtcv_snap = ULONG_MAX;
2180         static bool splatted;
2181         struct task_struct *wtp;
2182
2183         for_each_possible_cpu(cpu) {
2184                 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
2185                         pipesummary[i] += READ_ONCE(per_cpu(rcu_torture_count, cpu)[i]);
2186                         batchsummary[i] += READ_ONCE(per_cpu(rcu_torture_batch, cpu)[i]);
2187                 }
2188         }
2189         for (i = RCU_TORTURE_PIPE_LEN; i >= 0; i--) {
2190                 if (pipesummary[i] != 0)
2191                         break;
2192         }
2193
2194         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
2195         rtcp = rcu_access_pointer(rcu_torture_current);
2196         pr_cont("rtc: %p %s: %lu tfle: %d rta: %d rtaf: %d rtf: %d ",
2197                 rtcp,
2198                 rtcp && !rcu_stall_is_suppressed_at_boot() ? "ver" : "VER",
2199                 rcu_torture_current_version,
2200                 list_empty(&rcu_torture_freelist),
2201                 atomic_read(&n_rcu_torture_alloc),
2202                 atomic_read(&n_rcu_torture_alloc_fail),
2203                 atomic_read(&n_rcu_torture_free));
2204         pr_cont("rtmbe: %d rtmbkf: %d/%d rtbe: %ld rtbke: %ld ",
2205                 atomic_read(&n_rcu_torture_mberror),
2206                 atomic_read(&n_rcu_torture_mbchk_fail), atomic_read(&n_rcu_torture_mbchk_tries),
2207                 n_rcu_torture_barrier_error,
2208                 n_rcu_torture_boost_ktrerror);
2209         pr_cont("rtbf: %ld rtb: %ld nt: %ld ",
2210                 n_rcu_torture_boost_failure,
2211                 n_rcu_torture_boosts,
2212                 atomic_long_read(&n_rcu_torture_timers));
2213         torture_onoff_stats();
2214         pr_cont("barrier: %ld/%ld:%ld ",
2215                 data_race(n_barrier_successes),
2216                 data_race(n_barrier_attempts),
2217                 data_race(n_rcu_torture_barrier_error));
2218         pr_cont("read-exits: %ld ", data_race(n_read_exits)); // Statistic.
2219         pr_cont("nocb-toggles: %ld:%ld\n",
2220                 atomic_long_read(&n_nocb_offload), atomic_long_read(&n_nocb_deoffload));
2221
2222         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
2223         if (atomic_read(&n_rcu_torture_mberror) ||
2224             atomic_read(&n_rcu_torture_mbchk_fail) ||
2225             n_rcu_torture_barrier_error || n_rcu_torture_boost_ktrerror ||
2226             n_rcu_torture_boost_failure || i > 1) {
2227                 pr_cont("%s", "!!! ");
2228                 atomic_inc(&n_rcu_torture_error);
2229                 WARN_ON_ONCE(atomic_read(&n_rcu_torture_mberror));
2230                 WARN_ON_ONCE(atomic_read(&n_rcu_torture_mbchk_fail));
2231                 WARN_ON_ONCE(n_rcu_torture_barrier_error);  // rcu_barrier()
2232                 WARN_ON_ONCE(n_rcu_torture_boost_ktrerror); // no boost kthread
2233                 WARN_ON_ONCE(n_rcu_torture_boost_failure); // boost failed (TIMER_SOFTIRQ RT prio?)
2234                 WARN_ON_ONCE(i > 1); // Too-short grace period
2235         }
2236         pr_cont("Reader Pipe: ");
2237         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
2238                 pr_cont(" %ld", pipesummary[i]);
2239         pr_cont("\n");
2240
2241         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
2242         pr_cont("Reader Batch: ");
2243         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
2244                 pr_cont(" %ld", batchsummary[i]);
2245         pr_cont("\n");
2246
2247         pr_alert("%s%s ", torture_type, TORTURE_FLAG);
2248         pr_cont("Free-Block Circulation: ");
2249         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
2250                 pr_cont(" %d", atomic_read(&rcu_torture_wcount[i]));
2251         }
2252         pr_cont("\n");
2253
2254         if (cur_ops->stats)
2255                 cur_ops->stats();
2256         if (rtcv_snap == rcu_torture_current_version &&
2257             rcu_access_pointer(rcu_torture_current) &&
2258             !rcu_stall_is_suppressed()) {
2259                 int __maybe_unused flags = 0;
2260                 unsigned long __maybe_unused gp_seq = 0;
2261
2262                 rcutorture_get_gp_data(cur_ops->ttype,
2263                                        &flags, &gp_seq);
2264                 srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp,
2265                                         &flags, &gp_seq);
2266                 wtp = READ_ONCE(writer_task);
2267                 pr_alert("??? Writer stall state %s(%d) g%lu f%#x ->state %#x cpu %d\n",
2268                          rcu_torture_writer_state_getname(),
2269                          rcu_torture_writer_state, gp_seq, flags,
2270                          wtp == NULL ? ~0U : wtp->__state,
2271                          wtp == NULL ? -1 : (int)task_cpu(wtp));
2272                 if (!splatted && wtp) {
2273                         sched_show_task(wtp);
2274                         splatted = true;
2275                 }
2276                 if (cur_ops->gp_kthread_dbg)
2277                         cur_ops->gp_kthread_dbg();
2278                 rcu_ftrace_dump(DUMP_ALL);
2279         }
2280         rtcv_snap = rcu_torture_current_version;
2281 }
2282
2283 /*
2284  * Periodically prints torture statistics, if periodic statistics printing
2285  * was specified via the stat_interval module parameter.
2286  */
2287 static int
2288 rcu_torture_stats(void *arg)
2289 {
2290         VERBOSE_TOROUT_STRING("rcu_torture_stats task started");
2291         do {
2292                 schedule_timeout_interruptible(stat_interval * HZ);
2293                 rcu_torture_stats_print();
2294                 torture_shutdown_absorb("rcu_torture_stats");
2295         } while (!torture_must_stop());
2296         torture_kthread_stopping("rcu_torture_stats");
2297         return 0;
2298 }
2299
2300 /* Test mem_dump_obj() and friends.  */
2301 static void rcu_torture_mem_dump_obj(void)
2302 {
2303         struct rcu_head *rhp;
2304         struct kmem_cache *kcp;
2305         static int z;
2306
2307         kcp = kmem_cache_create("rcuscale", 136, 8, SLAB_STORE_USER, NULL);
2308         if (WARN_ON_ONCE(!kcp))
2309                 return;
2310         rhp = kmem_cache_alloc(kcp, GFP_KERNEL);
2311         if (WARN_ON_ONCE(!rhp)) {
2312                 kmem_cache_destroy(kcp);
2313                 return;
2314         }
2315         pr_alert("mem_dump_obj() slab test: rcu_torture_stats = %px, &rhp = %px, rhp = %px, &z = %px\n", stats_task, &rhp, rhp, &z);
2316         pr_alert("mem_dump_obj(ZERO_SIZE_PTR):");
2317         mem_dump_obj(ZERO_SIZE_PTR);
2318         pr_alert("mem_dump_obj(NULL):");
2319         mem_dump_obj(NULL);
2320         pr_alert("mem_dump_obj(%px):", &rhp);
2321         mem_dump_obj(&rhp);
2322         pr_alert("mem_dump_obj(%px):", rhp);
2323         mem_dump_obj(rhp);
2324         pr_alert("mem_dump_obj(%px):", &rhp->func);
2325         mem_dump_obj(&rhp->func);
2326         pr_alert("mem_dump_obj(%px):", &z);
2327         mem_dump_obj(&z);
2328         kmem_cache_free(kcp, rhp);
2329         kmem_cache_destroy(kcp);
2330         rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
2331         if (WARN_ON_ONCE(!rhp))
2332                 return;
2333         pr_alert("mem_dump_obj() kmalloc test: rcu_torture_stats = %px, &rhp = %px, rhp = %px\n", stats_task, &rhp, rhp);
2334         pr_alert("mem_dump_obj(kmalloc %px):", rhp);
2335         mem_dump_obj(rhp);
2336         pr_alert("mem_dump_obj(kmalloc %px):", &rhp->func);
2337         mem_dump_obj(&rhp->func);
2338         kfree(rhp);
2339         rhp = vmalloc(4096);
2340         if (WARN_ON_ONCE(!rhp))
2341                 return;
2342         pr_alert("mem_dump_obj() vmalloc test: rcu_torture_stats = %px, &rhp = %px, rhp = %px\n", stats_task, &rhp, rhp);
2343         pr_alert("mem_dump_obj(vmalloc %px):", rhp);
2344         mem_dump_obj(rhp);
2345         pr_alert("mem_dump_obj(vmalloc %px):", &rhp->func);
2346         mem_dump_obj(&rhp->func);
2347         vfree(rhp);
2348 }
2349
2350 static void
2351 rcu_torture_print_module_parms(struct rcu_torture_ops *cur_ops, const char *tag)
2352 {
2353         pr_alert("%s" TORTURE_FLAG
2354                  "--- %s: nreaders=%d nfakewriters=%d "
2355                  "stat_interval=%d verbose=%d test_no_idle_hz=%d "
2356                  "shuffle_interval=%d stutter=%d irqreader=%d "
2357                  "fqs_duration=%d fqs_holdoff=%d fqs_stutter=%d "
2358                  "test_boost=%d/%d test_boost_interval=%d "
2359                  "test_boost_duration=%d shutdown_secs=%d "
2360                  "stall_cpu=%d stall_cpu_holdoff=%d stall_cpu_irqsoff=%d "
2361                  "stall_cpu_block=%d "
2362                  "n_barrier_cbs=%d "
2363                  "onoff_interval=%d onoff_holdoff=%d "
2364                  "read_exit_delay=%d read_exit_burst=%d "
2365                  "nocbs_nthreads=%d nocbs_toggle=%d "
2366                  "test_nmis=%d\n",
2367                  torture_type, tag, nrealreaders, nfakewriters,
2368                  stat_interval, verbose, test_no_idle_hz, shuffle_interval,
2369                  stutter, irqreader, fqs_duration, fqs_holdoff, fqs_stutter,
2370                  test_boost, cur_ops->can_boost,
2371                  test_boost_interval, test_boost_duration, shutdown_secs,
2372                  stall_cpu, stall_cpu_holdoff, stall_cpu_irqsoff,
2373                  stall_cpu_block,
2374                  n_barrier_cbs,
2375                  onoff_interval, onoff_holdoff,
2376                  read_exit_delay, read_exit_burst,
2377                  nocbs_nthreads, nocbs_toggle,
2378                  test_nmis);
2379 }
2380
2381 static int rcutorture_booster_cleanup(unsigned int cpu)
2382 {
2383         struct task_struct *t;
2384
2385         if (boost_tasks[cpu] == NULL)
2386                 return 0;
2387         mutex_lock(&boost_mutex);
2388         t = boost_tasks[cpu];
2389         boost_tasks[cpu] = NULL;
2390         rcu_torture_enable_rt_throttle();
2391         mutex_unlock(&boost_mutex);
2392
2393         /* This must be outside of the mutex, otherwise deadlock! */
2394         torture_stop_kthread(rcu_torture_boost, t);
2395         return 0;
2396 }
2397
2398 static int rcutorture_booster_init(unsigned int cpu)
2399 {
2400         int retval;
2401
2402         if (boost_tasks[cpu] != NULL)
2403                 return 0;  /* Already created, nothing more to do. */
2404
2405         // Testing RCU priority boosting requires rcutorture do
2406         // some serious abuse.  Counter this by running ksoftirqd
2407         // at higher priority.
2408         if (IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)) {
2409                 struct sched_param sp;
2410                 struct task_struct *t;
2411
2412                 t = per_cpu(ksoftirqd, cpu);
2413                 WARN_ON_ONCE(!t);
2414                 sp.sched_priority = 2;
2415                 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
2416         }
2417
2418         /* Don't allow time recalculation while creating a new task. */
2419         mutex_lock(&boost_mutex);
2420         rcu_torture_disable_rt_throttle();
2421         VERBOSE_TOROUT_STRING("Creating rcu_torture_boost task");
2422         boost_tasks[cpu] = kthread_run_on_cpu(rcu_torture_boost, NULL,
2423                                               cpu, "rcu_torture_boost_%u");
2424         if (IS_ERR(boost_tasks[cpu])) {
2425                 retval = PTR_ERR(boost_tasks[cpu]);
2426                 VERBOSE_TOROUT_STRING("rcu_torture_boost task create failed");
2427                 n_rcu_torture_boost_ktrerror++;
2428                 boost_tasks[cpu] = NULL;
2429                 mutex_unlock(&boost_mutex);
2430                 return retval;
2431         }
2432         mutex_unlock(&boost_mutex);
2433         return 0;
2434 }
2435
2436 static int rcu_torture_stall_nf(struct notifier_block *nb, unsigned long v, void *ptr)
2437 {
2438         pr_info("%s: v=%lu, duration=%lu.\n", __func__, v, (unsigned long)ptr);
2439         return NOTIFY_OK;
2440 }
2441
2442 static struct notifier_block rcu_torture_stall_block = {
2443         .notifier_call = rcu_torture_stall_nf,
2444 };
2445
2446 /*
2447  * CPU-stall kthread.  It waits as specified by stall_cpu_holdoff, then
2448  * induces a CPU stall for the time specified by stall_cpu.  If a new
2449  * stall test is added, stallsdone in rcu_torture_writer() must be adjusted.
2450  */
2451 static int rcu_torture_stall(void *args)
2452 {
2453         int idx;
2454         int ret;
2455         unsigned long stop_at;
2456
2457         VERBOSE_TOROUT_STRING("rcu_torture_stall task started");
2458         if (rcu_cpu_stall_notifiers) {
2459                 ret = rcu_stall_chain_notifier_register(&rcu_torture_stall_block);
2460                 if (ret)
2461                         pr_info("%s: rcu_stall_chain_notifier_register() returned %d, %sexpected.\n",
2462                                 __func__, ret, !IS_ENABLED(CONFIG_RCU_STALL_COMMON) ? "un" : "");
2463         }
2464         if (stall_cpu_holdoff > 0) {
2465                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin holdoff");
2466                 schedule_timeout_interruptible(stall_cpu_holdoff * HZ);
2467                 VERBOSE_TOROUT_STRING("rcu_torture_stall end holdoff");
2468         }
2469         if (!kthread_should_stop() && stall_gp_kthread > 0) {
2470                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin GP stall");
2471                 rcu_gp_set_torture_wait(stall_gp_kthread * HZ);
2472                 for (idx = 0; idx < stall_gp_kthread + 2; idx++) {
2473                         if (kthread_should_stop())
2474                                 break;
2475                         schedule_timeout_uninterruptible(HZ);
2476                 }
2477         }
2478         if (!kthread_should_stop() && stall_cpu > 0) {
2479                 VERBOSE_TOROUT_STRING("rcu_torture_stall begin CPU stall");
2480                 stop_at = ktime_get_seconds() + stall_cpu;
2481                 /* RCU CPU stall is expected behavior in following code. */
2482                 idx = cur_ops->readlock();
2483                 if (stall_cpu_irqsoff)
2484                         local_irq_disable();
2485                 else if (!stall_cpu_block)
2486                         preempt_disable();
2487                 pr_alert("%s start on CPU %d.\n",
2488                           __func__, raw_smp_processor_id());
2489                 while (ULONG_CMP_LT((unsigned long)ktime_get_seconds(),
2490                                     stop_at))
2491                         if (stall_cpu_block) {
2492 #ifdef CONFIG_PREEMPTION
2493                                 preempt_schedule();
2494 #else
2495                                 schedule_timeout_uninterruptible(HZ);
2496 #endif
2497                         } else if (stall_no_softlockup) {
2498                                 touch_softlockup_watchdog();
2499                         }
2500                 if (stall_cpu_irqsoff)
2501                         local_irq_enable();
2502                 else if (!stall_cpu_block)
2503                         preempt_enable();
2504                 cur_ops->readunlock(idx);
2505         }
2506         pr_alert("%s end.\n", __func__);
2507         if (rcu_cpu_stall_notifiers && !ret) {
2508                 ret = rcu_stall_chain_notifier_unregister(&rcu_torture_stall_block);
2509                 if (ret)
2510                         pr_info("%s: rcu_stall_chain_notifier_unregister() returned %d.\n", __func__, ret);
2511         }
2512         torture_shutdown_absorb("rcu_torture_stall");
2513         while (!kthread_should_stop())
2514                 schedule_timeout_interruptible(10 * HZ);
2515         return 0;
2516 }
2517
2518 /* Spawn CPU-stall kthread, if stall_cpu specified. */
2519 static int __init rcu_torture_stall_init(void)
2520 {
2521         if (stall_cpu <= 0 && stall_gp_kthread <= 0)
2522                 return 0;
2523         return torture_create_kthread(rcu_torture_stall, NULL, stall_task);
2524 }
2525
2526 /* State structure for forward-progress self-propagating RCU callback. */
2527 struct fwd_cb_state {
2528         struct rcu_head rh;
2529         int stop;
2530 };
2531
2532 /*
2533  * Forward-progress self-propagating RCU callback function.  Because
2534  * callbacks run from softirq, this function is an implicit RCU read-side
2535  * critical section.
2536  */
2537 static void rcu_torture_fwd_prog_cb(struct rcu_head *rhp)
2538 {
2539         struct fwd_cb_state *fcsp = container_of(rhp, struct fwd_cb_state, rh);
2540
2541         if (READ_ONCE(fcsp->stop)) {
2542                 WRITE_ONCE(fcsp->stop, 2);
2543                 return;
2544         }
2545         cur_ops->call(&fcsp->rh, rcu_torture_fwd_prog_cb);
2546 }
2547
2548 /* State for continuous-flood RCU callbacks. */
2549 struct rcu_fwd_cb {
2550         struct rcu_head rh;
2551         struct rcu_fwd_cb *rfc_next;
2552         struct rcu_fwd *rfc_rfp;
2553         int rfc_gps;
2554 };
2555
2556 #define MAX_FWD_CB_JIFFIES      (8 * HZ) /* Maximum CB test duration. */
2557 #define MIN_FWD_CB_LAUNDERS     3       /* This many CB invocations to count. */
2558 #define MIN_FWD_CBS_LAUNDERED   100     /* Number of counted CBs. */
2559 #define FWD_CBS_HIST_DIV        10      /* Histogram buckets/second. */
2560 #define N_LAUNDERS_HIST (2 * MAX_FWD_CB_JIFFIES / (HZ / FWD_CBS_HIST_DIV))
2561
2562 struct rcu_launder_hist {
2563         long n_launders;
2564         unsigned long launder_gp_seq;
2565 };
2566
2567 struct rcu_fwd {
2568         spinlock_t rcu_fwd_lock;
2569         struct rcu_fwd_cb *rcu_fwd_cb_head;
2570         struct rcu_fwd_cb **rcu_fwd_cb_tail;
2571         long n_launders_cb;
2572         unsigned long rcu_fwd_startat;
2573         struct rcu_launder_hist n_launders_hist[N_LAUNDERS_HIST];
2574         unsigned long rcu_launder_gp_seq_start;
2575         int rcu_fwd_id;
2576 };
2577
2578 static DEFINE_MUTEX(rcu_fwd_mutex);
2579 static struct rcu_fwd *rcu_fwds;
2580 static unsigned long rcu_fwd_seq;
2581 static atomic_long_t rcu_fwd_max_cbs;
2582 static bool rcu_fwd_emergency_stop;
2583
2584 static void rcu_torture_fwd_cb_hist(struct rcu_fwd *rfp)
2585 {
2586         unsigned long gps;
2587         unsigned long gps_old;
2588         int i;
2589         int j;
2590
2591         for (i = ARRAY_SIZE(rfp->n_launders_hist) - 1; i > 0; i--)
2592                 if (rfp->n_launders_hist[i].n_launders > 0)
2593                         break;
2594         pr_alert("%s: Callback-invocation histogram %d (duration %lu jiffies):",
2595                  __func__, rfp->rcu_fwd_id, jiffies - rfp->rcu_fwd_startat);
2596         gps_old = rfp->rcu_launder_gp_seq_start;
2597         for (j = 0; j <= i; j++) {
2598                 gps = rfp->n_launders_hist[j].launder_gp_seq;
2599                 pr_cont(" %ds/%d: %ld:%ld",
2600                         j + 1, FWD_CBS_HIST_DIV,
2601                         rfp->n_launders_hist[j].n_launders,
2602                         rcutorture_seq_diff(gps, gps_old));
2603                 gps_old = gps;
2604         }
2605         pr_cont("\n");
2606 }
2607
2608 /* Callback function for continuous-flood RCU callbacks. */
2609 static void rcu_torture_fwd_cb_cr(struct rcu_head *rhp)
2610 {
2611         unsigned long flags;
2612         int i;
2613         struct rcu_fwd_cb *rfcp = container_of(rhp, struct rcu_fwd_cb, rh);
2614         struct rcu_fwd_cb **rfcpp;
2615         struct rcu_fwd *rfp = rfcp->rfc_rfp;
2616
2617         rfcp->rfc_next = NULL;
2618         rfcp->rfc_gps++;
2619         spin_lock_irqsave(&rfp->rcu_fwd_lock, flags);
2620         rfcpp = rfp->rcu_fwd_cb_tail;
2621         rfp->rcu_fwd_cb_tail = &rfcp->rfc_next;
2622         WRITE_ONCE(*rfcpp, rfcp);
2623         WRITE_ONCE(rfp->n_launders_cb, rfp->n_launders_cb + 1);
2624         i = ((jiffies - rfp->rcu_fwd_startat) / (HZ / FWD_CBS_HIST_DIV));
2625         if (i >= ARRAY_SIZE(rfp->n_launders_hist))
2626                 i = ARRAY_SIZE(rfp->n_launders_hist) - 1;
2627         rfp->n_launders_hist[i].n_launders++;
2628         rfp->n_launders_hist[i].launder_gp_seq = cur_ops->get_gp_seq();
2629         spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2630 }
2631
2632 // Give the scheduler a chance, even on nohz_full CPUs.
2633 static void rcu_torture_fwd_prog_cond_resched(unsigned long iter)
2634 {
2635         if (IS_ENABLED(CONFIG_PREEMPTION) && IS_ENABLED(CONFIG_NO_HZ_FULL)) {
2636                 // Real call_rcu() floods hit userspace, so emulate that.
2637                 if (need_resched() || (iter & 0xfff))
2638                         schedule();
2639                 return;
2640         }
2641         // No userspace emulation: CB invocation throttles call_rcu()
2642         cond_resched();
2643 }
2644
2645 /*
2646  * Free all callbacks on the rcu_fwd_cb_head list, either because the
2647  * test is over or because we hit an OOM event.
2648  */
2649 static unsigned long rcu_torture_fwd_prog_cbfree(struct rcu_fwd *rfp)
2650 {
2651         unsigned long flags;
2652         unsigned long freed = 0;
2653         struct rcu_fwd_cb *rfcp;
2654
2655         for (;;) {
2656                 spin_lock_irqsave(&rfp->rcu_fwd_lock, flags);
2657                 rfcp = rfp->rcu_fwd_cb_head;
2658                 if (!rfcp) {
2659                         spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2660                         break;
2661                 }
2662                 rfp->rcu_fwd_cb_head = rfcp->rfc_next;
2663                 if (!rfp->rcu_fwd_cb_head)
2664                         rfp->rcu_fwd_cb_tail = &rfp->rcu_fwd_cb_head;
2665                 spin_unlock_irqrestore(&rfp->rcu_fwd_lock, flags);
2666                 kfree(rfcp);
2667                 freed++;
2668                 rcu_torture_fwd_prog_cond_resched(freed);
2669                 if (tick_nohz_full_enabled()) {
2670                         local_irq_save(flags);
2671                         rcu_momentary_dyntick_idle();
2672                         local_irq_restore(flags);
2673                 }
2674         }
2675         return freed;
2676 }
2677
2678 /* Carry out need_resched()/cond_resched() forward-progress testing. */
2679 static void rcu_torture_fwd_prog_nr(struct rcu_fwd *rfp,
2680                                     int *tested, int *tested_tries)
2681 {
2682         unsigned long cver;
2683         unsigned long dur;
2684         struct fwd_cb_state fcs;
2685         unsigned long gps;
2686         int idx;
2687         int sd;
2688         int sd4;
2689         bool selfpropcb = false;
2690         unsigned long stopat;
2691         static DEFINE_TORTURE_RANDOM(trs);
2692
2693         pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2694         if (!cur_ops->sync)
2695                 return; // Cannot do need_resched() forward progress testing without ->sync.
2696         if (cur_ops->call && cur_ops->cb_barrier) {
2697                 init_rcu_head_on_stack(&fcs.rh);
2698                 selfpropcb = true;
2699         }
2700
2701         /* Tight loop containing cond_resched(). */
2702         atomic_inc(&rcu_fwd_cb_nodelay);
2703         cur_ops->sync(); /* Later readers see above write. */
2704         if  (selfpropcb) {
2705                 WRITE_ONCE(fcs.stop, 0);
2706                 cur_ops->call(&fcs.rh, rcu_torture_fwd_prog_cb);
2707         }
2708         cver = READ_ONCE(rcu_torture_current_version);
2709         gps = cur_ops->get_gp_seq();
2710         sd = cur_ops->stall_dur() + 1;
2711         sd4 = (sd + fwd_progress_div - 1) / fwd_progress_div;
2712         dur = sd4 + torture_random(&trs) % (sd - sd4);
2713         WRITE_ONCE(rfp->rcu_fwd_startat, jiffies);
2714         stopat = rfp->rcu_fwd_startat + dur;
2715         while (time_before(jiffies, stopat) &&
2716                !shutdown_time_arrived() &&
2717                !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2718                 idx = cur_ops->readlock();
2719                 udelay(10);
2720                 cur_ops->readunlock(idx);
2721                 if (!fwd_progress_need_resched || need_resched())
2722                         cond_resched();
2723         }
2724         (*tested_tries)++;
2725         if (!time_before(jiffies, stopat) &&
2726             !shutdown_time_arrived() &&
2727             !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2728                 (*tested)++;
2729                 cver = READ_ONCE(rcu_torture_current_version) - cver;
2730                 gps = rcutorture_seq_diff(cur_ops->get_gp_seq(), gps);
2731                 WARN_ON(!cver && gps < 2);
2732                 pr_alert("%s: %d Duration %ld cver %ld gps %ld\n", __func__,
2733                          rfp->rcu_fwd_id, dur, cver, gps);
2734         }
2735         if (selfpropcb) {
2736                 WRITE_ONCE(fcs.stop, 1);
2737                 cur_ops->sync(); /* Wait for running CB to complete. */
2738                 pr_alert("%s: Waiting for CBs: %pS() %d\n", __func__, cur_ops->cb_barrier, rfp->rcu_fwd_id);
2739                 cur_ops->cb_barrier(); /* Wait for queued callbacks. */
2740         }
2741
2742         if (selfpropcb) {
2743                 WARN_ON(READ_ONCE(fcs.stop) != 2);
2744                 destroy_rcu_head_on_stack(&fcs.rh);
2745         }
2746         schedule_timeout_uninterruptible(HZ / 10); /* Let kthreads recover. */
2747         atomic_dec(&rcu_fwd_cb_nodelay);
2748 }
2749
2750 /* Carry out call_rcu() forward-progress testing. */
2751 static void rcu_torture_fwd_prog_cr(struct rcu_fwd *rfp)
2752 {
2753         unsigned long cver;
2754         unsigned long flags;
2755         unsigned long gps;
2756         int i;
2757         long n_launders;
2758         long n_launders_cb_snap;
2759         long n_launders_sa;
2760         long n_max_cbs;
2761         long n_max_gps;
2762         struct rcu_fwd_cb *rfcp;
2763         struct rcu_fwd_cb *rfcpn;
2764         unsigned long stopat;
2765         unsigned long stoppedat;
2766
2767         pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2768         if (READ_ONCE(rcu_fwd_emergency_stop))
2769                 return; /* Get out of the way quickly, no GP wait! */
2770         if (!cur_ops->call)
2771                 return; /* Can't do call_rcu() fwd prog without ->call. */
2772
2773         /* Loop continuously posting RCU callbacks. */
2774         atomic_inc(&rcu_fwd_cb_nodelay);
2775         cur_ops->sync(); /* Later readers see above write. */
2776         WRITE_ONCE(rfp->rcu_fwd_startat, jiffies);
2777         stopat = rfp->rcu_fwd_startat + MAX_FWD_CB_JIFFIES;
2778         n_launders = 0;
2779         rfp->n_launders_cb = 0; // Hoist initialization for multi-kthread
2780         n_launders_sa = 0;
2781         n_max_cbs = 0;
2782         n_max_gps = 0;
2783         for (i = 0; i < ARRAY_SIZE(rfp->n_launders_hist); i++)
2784                 rfp->n_launders_hist[i].n_launders = 0;
2785         cver = READ_ONCE(rcu_torture_current_version);
2786         gps = cur_ops->get_gp_seq();
2787         rfp->rcu_launder_gp_seq_start = gps;
2788         tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2789         while (time_before(jiffies, stopat) &&
2790                !shutdown_time_arrived() &&
2791                !READ_ONCE(rcu_fwd_emergency_stop) && !torture_must_stop()) {
2792                 rfcp = READ_ONCE(rfp->rcu_fwd_cb_head);
2793                 rfcpn = NULL;
2794                 if (rfcp)
2795                         rfcpn = READ_ONCE(rfcp->rfc_next);
2796                 if (rfcpn) {
2797                         if (rfcp->rfc_gps >= MIN_FWD_CB_LAUNDERS &&
2798                             ++n_max_gps >= MIN_FWD_CBS_LAUNDERED)
2799                                 break;
2800                         rfp->rcu_fwd_cb_head = rfcpn;
2801                         n_launders++;
2802                         n_launders_sa++;
2803                 } else if (!cur_ops->cbflood_max || cur_ops->cbflood_max > n_max_cbs) {
2804                         rfcp = kmalloc(sizeof(*rfcp), GFP_KERNEL);
2805                         if (WARN_ON_ONCE(!rfcp)) {
2806                                 schedule_timeout_interruptible(1);
2807                                 continue;
2808                         }
2809                         n_max_cbs++;
2810                         n_launders_sa = 0;
2811                         rfcp->rfc_gps = 0;
2812                         rfcp->rfc_rfp = rfp;
2813                 } else {
2814                         rfcp = NULL;
2815                 }
2816                 if (rfcp)
2817                         cur_ops->call(&rfcp->rh, rcu_torture_fwd_cb_cr);
2818                 rcu_torture_fwd_prog_cond_resched(n_launders + n_max_cbs);
2819                 if (tick_nohz_full_enabled()) {
2820                         local_irq_save(flags);
2821                         rcu_momentary_dyntick_idle();
2822                         local_irq_restore(flags);
2823                 }
2824         }
2825         stoppedat = jiffies;
2826         n_launders_cb_snap = READ_ONCE(rfp->n_launders_cb);
2827         cver = READ_ONCE(rcu_torture_current_version) - cver;
2828         gps = rcutorture_seq_diff(cur_ops->get_gp_seq(), gps);
2829         pr_alert("%s: Waiting for CBs: %pS() %d\n", __func__, cur_ops->cb_barrier, rfp->rcu_fwd_id);
2830         cur_ops->cb_barrier(); /* Wait for callbacks to be invoked. */
2831         (void)rcu_torture_fwd_prog_cbfree(rfp);
2832
2833         if (!torture_must_stop() && !READ_ONCE(rcu_fwd_emergency_stop) &&
2834             !shutdown_time_arrived()) {
2835                 WARN_ON(n_max_gps < MIN_FWD_CBS_LAUNDERED);
2836                 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",
2837                          __func__,
2838                          stoppedat - rfp->rcu_fwd_startat, jiffies - stoppedat,
2839                          n_launders + n_max_cbs - n_launders_cb_snap,
2840                          n_launders, n_launders_sa,
2841                          n_max_gps, n_max_cbs, cver, gps);
2842                 atomic_long_add(n_max_cbs, &rcu_fwd_max_cbs);
2843                 mutex_lock(&rcu_fwd_mutex); // Serialize histograms.
2844                 rcu_torture_fwd_cb_hist(rfp);
2845                 mutex_unlock(&rcu_fwd_mutex);
2846         }
2847         schedule_timeout_uninterruptible(HZ); /* Let CBs drain. */
2848         tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2849         atomic_dec(&rcu_fwd_cb_nodelay);
2850 }
2851
2852
2853 /*
2854  * OOM notifier, but this only prints diagnostic information for the
2855  * current forward-progress test.
2856  */
2857 static int rcutorture_oom_notify(struct notifier_block *self,
2858                                  unsigned long notused, void *nfreed)
2859 {
2860         int i;
2861         long ncbs;
2862         struct rcu_fwd *rfp;
2863
2864         mutex_lock(&rcu_fwd_mutex);
2865         rfp = rcu_fwds;
2866         if (!rfp) {
2867                 mutex_unlock(&rcu_fwd_mutex);
2868                 return NOTIFY_OK;
2869         }
2870         WARN(1, "%s invoked upon OOM during forward-progress testing.\n",
2871              __func__);
2872         for (i = 0; i < fwd_progress; i++) {
2873                 rcu_torture_fwd_cb_hist(&rfp[i]);
2874                 rcu_fwd_progress_check(1 + (jiffies - READ_ONCE(rfp[i].rcu_fwd_startat)) / 2);
2875         }
2876         WRITE_ONCE(rcu_fwd_emergency_stop, true);
2877         smp_mb(); /* Emergency stop before free and wait to avoid hangs. */
2878         ncbs = 0;
2879         for (i = 0; i < fwd_progress; i++)
2880                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2881         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2882         cur_ops->cb_barrier();
2883         ncbs = 0;
2884         for (i = 0; i < fwd_progress; i++)
2885                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2886         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2887         cur_ops->cb_barrier();
2888         ncbs = 0;
2889         for (i = 0; i < fwd_progress; i++)
2890                 ncbs += rcu_torture_fwd_prog_cbfree(&rfp[i]);
2891         pr_info("%s: Freed %lu RCU callbacks.\n", __func__, ncbs);
2892         smp_mb(); /* Frees before return to avoid redoing OOM. */
2893         (*(unsigned long *)nfreed)++; /* Forward progress CBs freed! */
2894         pr_info("%s returning after OOM processing.\n", __func__);
2895         mutex_unlock(&rcu_fwd_mutex);
2896         return NOTIFY_OK;
2897 }
2898
2899 static struct notifier_block rcutorture_oom_nb = {
2900         .notifier_call = rcutorture_oom_notify
2901 };
2902
2903 /* Carry out grace-period forward-progress testing. */
2904 static int rcu_torture_fwd_prog(void *args)
2905 {
2906         bool firsttime = true;
2907         long max_cbs;
2908         int oldnice = task_nice(current);
2909         unsigned long oldseq = READ_ONCE(rcu_fwd_seq);
2910         struct rcu_fwd *rfp = args;
2911         int tested = 0;
2912         int tested_tries = 0;
2913
2914         VERBOSE_TOROUT_STRING("rcu_torture_fwd_progress task started");
2915         rcu_bind_current_to_nocb();
2916         if (!IS_ENABLED(CONFIG_SMP) || !IS_ENABLED(CONFIG_RCU_BOOST))
2917                 set_user_nice(current, MAX_NICE);
2918         do {
2919                 if (!rfp->rcu_fwd_id) {
2920                         schedule_timeout_interruptible(fwd_progress_holdoff * HZ);
2921                         WRITE_ONCE(rcu_fwd_emergency_stop, false);
2922                         if (!firsttime) {
2923                                 max_cbs = atomic_long_xchg(&rcu_fwd_max_cbs, 0);
2924                                 pr_alert("%s n_max_cbs: %ld\n", __func__, max_cbs);
2925                         }
2926                         firsttime = false;
2927                         WRITE_ONCE(rcu_fwd_seq, rcu_fwd_seq + 1);
2928                 } else {
2929                         while (READ_ONCE(rcu_fwd_seq) == oldseq && !torture_must_stop())
2930                                 schedule_timeout_interruptible(HZ / 20);
2931                         oldseq = READ_ONCE(rcu_fwd_seq);
2932                 }
2933                 pr_alert("%s: Starting forward-progress test %d\n", __func__, rfp->rcu_fwd_id);
2934                 if (rcu_inkernel_boot_has_ended() && torture_num_online_cpus() > rfp->rcu_fwd_id)
2935                         rcu_torture_fwd_prog_cr(rfp);
2936                 if ((cur_ops->stall_dur && cur_ops->stall_dur() > 0) &&
2937                     (!IS_ENABLED(CONFIG_TINY_RCU) ||
2938                      (rcu_inkernel_boot_has_ended() &&
2939                       torture_num_online_cpus() > rfp->rcu_fwd_id)))
2940                         rcu_torture_fwd_prog_nr(rfp, &tested, &tested_tries);
2941
2942                 /* Avoid slow periods, better to test when busy. */
2943                 if (stutter_wait("rcu_torture_fwd_prog"))
2944                         sched_set_normal(current, oldnice);
2945         } while (!torture_must_stop());
2946         /* Short runs might not contain a valid forward-progress attempt. */
2947         if (!rfp->rcu_fwd_id) {
2948                 WARN_ON(!tested && tested_tries >= 5);
2949                 pr_alert("%s: tested %d tested_tries %d\n", __func__, tested, tested_tries);
2950         }
2951         torture_kthread_stopping("rcu_torture_fwd_prog");
2952         return 0;
2953 }
2954
2955 /* If forward-progress checking is requested and feasible, spawn the thread. */
2956 static int __init rcu_torture_fwd_prog_init(void)
2957 {
2958         int i;
2959         int ret = 0;
2960         struct rcu_fwd *rfp;
2961
2962         if (!fwd_progress)
2963                 return 0; /* Not requested, so don't do it. */
2964         if (fwd_progress >= nr_cpu_ids) {
2965                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Limiting fwd_progress to # CPUs.\n");
2966                 fwd_progress = nr_cpu_ids;
2967         } else if (fwd_progress < 0) {
2968                 fwd_progress = nr_cpu_ids;
2969         }
2970         if ((!cur_ops->sync && !cur_ops->call) ||
2971             (!cur_ops->cbflood_max && (!cur_ops->stall_dur || cur_ops->stall_dur() <= 0)) ||
2972             cur_ops == &rcu_busted_ops) {
2973                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Disabled, unsupported by RCU flavor under test");
2974                 fwd_progress = 0;
2975                 return 0;
2976         }
2977         if (stall_cpu > 0) {
2978                 VERBOSE_TOROUT_STRING("rcu_torture_fwd_prog_init: Disabled, conflicts with CPU-stall testing");
2979                 fwd_progress = 0;
2980                 if (IS_MODULE(CONFIG_RCU_TORTURE_TEST))
2981                         return -EINVAL; /* In module, can fail back to user. */
2982                 WARN_ON(1); /* Make sure rcutorture notices conflict. */
2983                 return 0;
2984         }
2985         if (fwd_progress_holdoff <= 0)
2986                 fwd_progress_holdoff = 1;
2987         if (fwd_progress_div <= 0)
2988                 fwd_progress_div = 4;
2989         rfp = kcalloc(fwd_progress, sizeof(*rfp), GFP_KERNEL);
2990         fwd_prog_tasks = kcalloc(fwd_progress, sizeof(*fwd_prog_tasks), GFP_KERNEL);
2991         if (!rfp || !fwd_prog_tasks) {
2992                 kfree(rfp);
2993                 kfree(fwd_prog_tasks);
2994                 fwd_prog_tasks = NULL;
2995                 fwd_progress = 0;
2996                 return -ENOMEM;
2997         }
2998         for (i = 0; i < fwd_progress; i++) {
2999                 spin_lock_init(&rfp[i].rcu_fwd_lock);
3000                 rfp[i].rcu_fwd_cb_tail = &rfp[i].rcu_fwd_cb_head;
3001                 rfp[i].rcu_fwd_id = i;
3002         }
3003         mutex_lock(&rcu_fwd_mutex);
3004         rcu_fwds = rfp;
3005         mutex_unlock(&rcu_fwd_mutex);
3006         register_oom_notifier(&rcutorture_oom_nb);
3007         for (i = 0; i < fwd_progress; i++) {
3008                 ret = torture_create_kthread(rcu_torture_fwd_prog, &rcu_fwds[i], fwd_prog_tasks[i]);
3009                 if (ret) {
3010                         fwd_progress = i;
3011                         return ret;
3012                 }
3013         }
3014         return 0;
3015 }
3016
3017 static void rcu_torture_fwd_prog_cleanup(void)
3018 {
3019         int i;
3020         struct rcu_fwd *rfp;
3021
3022         if (!rcu_fwds || !fwd_prog_tasks)
3023                 return;
3024         for (i = 0; i < fwd_progress; i++)
3025                 torture_stop_kthread(rcu_torture_fwd_prog, fwd_prog_tasks[i]);
3026         unregister_oom_notifier(&rcutorture_oom_nb);
3027         mutex_lock(&rcu_fwd_mutex);
3028         rfp = rcu_fwds;
3029         rcu_fwds = NULL;
3030         mutex_unlock(&rcu_fwd_mutex);
3031         kfree(rfp);
3032         kfree(fwd_prog_tasks);
3033         fwd_prog_tasks = NULL;
3034 }
3035
3036 /* Callback function for RCU barrier testing. */
3037 static void rcu_torture_barrier_cbf(struct rcu_head *rcu)
3038 {
3039         atomic_inc(&barrier_cbs_invoked);
3040 }
3041
3042 /* IPI handler to get callback posted on desired CPU, if online. */
3043 static void rcu_torture_barrier1cb(void *rcu_void)
3044 {
3045         struct rcu_head *rhp = rcu_void;
3046
3047         cur_ops->call(rhp, rcu_torture_barrier_cbf);
3048 }
3049
3050 /* kthread function to register callbacks used to test RCU barriers. */
3051 static int rcu_torture_barrier_cbs(void *arg)
3052 {
3053         long myid = (long)arg;
3054         bool lastphase = false;
3055         bool newphase;
3056         struct rcu_head rcu;
3057
3058         init_rcu_head_on_stack(&rcu);
3059         VERBOSE_TOROUT_STRING("rcu_torture_barrier_cbs task started");
3060         set_user_nice(current, MAX_NICE);
3061         do {
3062                 wait_event(barrier_cbs_wq[myid],
3063                            (newphase =
3064                             smp_load_acquire(&barrier_phase)) != lastphase ||
3065                            torture_must_stop());
3066                 lastphase = newphase;
3067                 if (torture_must_stop())
3068                         break;
3069                 /*
3070                  * The above smp_load_acquire() ensures barrier_phase load
3071                  * is ordered before the following ->call().
3072                  */
3073                 if (smp_call_function_single(myid, rcu_torture_barrier1cb,
3074                                              &rcu, 1)) {
3075                         // IPI failed, so use direct call from current CPU.
3076                         cur_ops->call(&rcu, rcu_torture_barrier_cbf);
3077                 }
3078                 if (atomic_dec_and_test(&barrier_cbs_count))
3079                         wake_up(&barrier_wq);
3080         } while (!torture_must_stop());
3081         if (cur_ops->cb_barrier != NULL)
3082                 cur_ops->cb_barrier();
3083         destroy_rcu_head_on_stack(&rcu);
3084         torture_kthread_stopping("rcu_torture_barrier_cbs");
3085         return 0;
3086 }
3087
3088 /* kthread function to drive and coordinate RCU barrier testing. */
3089 static int rcu_torture_barrier(void *arg)
3090 {
3091         int i;
3092
3093         VERBOSE_TOROUT_STRING("rcu_torture_barrier task starting");
3094         do {
3095                 atomic_set(&barrier_cbs_invoked, 0);
3096                 atomic_set(&barrier_cbs_count, n_barrier_cbs);
3097                 /* Ensure barrier_phase ordered after prior assignments. */
3098                 smp_store_release(&barrier_phase, !barrier_phase);
3099                 for (i = 0; i < n_barrier_cbs; i++)
3100                         wake_up(&barrier_cbs_wq[i]);
3101                 wait_event(barrier_wq,
3102                            atomic_read(&barrier_cbs_count) == 0 ||
3103                            torture_must_stop());
3104                 if (torture_must_stop())
3105                         break;
3106                 n_barrier_attempts++;
3107                 cur_ops->cb_barrier(); /* Implies smp_mb() for wait_event(). */
3108                 if (atomic_read(&barrier_cbs_invoked) != n_barrier_cbs) {
3109                         n_rcu_torture_barrier_error++;
3110                         pr_err("barrier_cbs_invoked = %d, n_barrier_cbs = %d\n",
3111                                atomic_read(&barrier_cbs_invoked),
3112                                n_barrier_cbs);
3113                         WARN_ON(1);
3114                         // Wait manually for the remaining callbacks
3115                         i = 0;
3116                         do {
3117                                 if (WARN_ON(i++ > HZ))
3118                                         i = INT_MIN;
3119                                 schedule_timeout_interruptible(1);
3120                                 cur_ops->cb_barrier();
3121                         } while (atomic_read(&barrier_cbs_invoked) !=
3122                                  n_barrier_cbs &&
3123                                  !torture_must_stop());
3124                         smp_mb(); // Can't trust ordering if broken.
3125                         if (!torture_must_stop())
3126                                 pr_err("Recovered: barrier_cbs_invoked = %d\n",
3127                                        atomic_read(&barrier_cbs_invoked));
3128                 } else {
3129                         n_barrier_successes++;
3130                 }
3131                 schedule_timeout_interruptible(HZ / 10);
3132         } while (!torture_must_stop());
3133         torture_kthread_stopping("rcu_torture_barrier");
3134         return 0;
3135 }
3136
3137 /* Initialize RCU barrier testing. */
3138 static int rcu_torture_barrier_init(void)
3139 {
3140         int i;
3141         int ret;
3142
3143         if (n_barrier_cbs <= 0)
3144                 return 0;
3145         if (cur_ops->call == NULL || cur_ops->cb_barrier == NULL) {
3146                 pr_alert("%s" TORTURE_FLAG
3147                          " Call or barrier ops missing for %s,\n",
3148                          torture_type, cur_ops->name);
3149                 pr_alert("%s" TORTURE_FLAG
3150                          " RCU barrier testing omitted from run.\n",
3151                          torture_type);
3152                 return 0;
3153         }
3154         atomic_set(&barrier_cbs_count, 0);
3155         atomic_set(&barrier_cbs_invoked, 0);
3156         barrier_cbs_tasks =
3157                 kcalloc(n_barrier_cbs, sizeof(barrier_cbs_tasks[0]),
3158                         GFP_KERNEL);
3159         barrier_cbs_wq =
3160                 kcalloc(n_barrier_cbs, sizeof(barrier_cbs_wq[0]), GFP_KERNEL);
3161         if (barrier_cbs_tasks == NULL || !barrier_cbs_wq)
3162                 return -ENOMEM;
3163         for (i = 0; i < n_barrier_cbs; i++) {
3164                 init_waitqueue_head(&barrier_cbs_wq[i]);
3165                 ret = torture_create_kthread(rcu_torture_barrier_cbs,
3166                                              (void *)(long)i,
3167                                              barrier_cbs_tasks[i]);
3168                 if (ret)
3169                         return ret;
3170         }
3171         return torture_create_kthread(rcu_torture_barrier, NULL, barrier_task);
3172 }
3173
3174 /* Clean up after RCU barrier testing. */
3175 static void rcu_torture_barrier_cleanup(void)
3176 {
3177         int i;
3178
3179         torture_stop_kthread(rcu_torture_barrier, barrier_task);
3180         if (barrier_cbs_tasks != NULL) {
3181                 for (i = 0; i < n_barrier_cbs; i++)
3182                         torture_stop_kthread(rcu_torture_barrier_cbs,
3183                                              barrier_cbs_tasks[i]);
3184                 kfree(barrier_cbs_tasks);
3185                 barrier_cbs_tasks = NULL;
3186         }
3187         if (barrier_cbs_wq != NULL) {
3188                 kfree(barrier_cbs_wq);
3189                 barrier_cbs_wq = NULL;
3190         }
3191 }
3192
3193 static bool rcu_torture_can_boost(void)
3194 {
3195         static int boost_warn_once;
3196         int prio;
3197
3198         if (!(test_boost == 1 && cur_ops->can_boost) && test_boost != 2)
3199                 return false;
3200         if (!cur_ops->start_gp_poll || !cur_ops->poll_gp_state)
3201                 return false;
3202
3203         prio = rcu_get_gp_kthreads_prio();
3204         if (!prio)
3205                 return false;
3206
3207         if (prio < 2) {
3208                 if (boost_warn_once == 1)
3209                         return false;
3210
3211                 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);
3212                 boost_warn_once = 1;
3213                 return false;
3214         }
3215
3216         return true;
3217 }
3218
3219 static bool read_exit_child_stop;
3220 static bool read_exit_child_stopped;
3221 static wait_queue_head_t read_exit_wq;
3222
3223 // Child kthread which just does an rcutorture reader and exits.
3224 static int rcu_torture_read_exit_child(void *trsp_in)
3225 {
3226         struct torture_random_state *trsp = trsp_in;
3227
3228         set_user_nice(current, MAX_NICE);
3229         // Minimize time between reading and exiting.
3230         while (!kthread_should_stop())
3231                 schedule_timeout_uninterruptible(HZ / 20);
3232         (void)rcu_torture_one_read(trsp, -1);
3233         return 0;
3234 }
3235
3236 // Parent kthread which creates and destroys read-exit child kthreads.
3237 static int rcu_torture_read_exit(void *unused)
3238 {
3239         bool errexit = false;
3240         int i;
3241         struct task_struct *tsp;
3242         DEFINE_TORTURE_RANDOM(trs);
3243
3244         // Allocate and initialize.
3245         set_user_nice(current, MAX_NICE);
3246         VERBOSE_TOROUT_STRING("rcu_torture_read_exit: Start of test");
3247
3248         // Each pass through this loop does one read-exit episode.
3249         do {
3250                 VERBOSE_TOROUT_STRING("rcu_torture_read_exit: Start of episode");
3251                 for (i = 0; i < read_exit_burst; i++) {
3252                         if (READ_ONCE(read_exit_child_stop))
3253                                 break;
3254                         stutter_wait("rcu_torture_read_exit");
3255                         // Spawn child.
3256                         tsp = kthread_run(rcu_torture_read_exit_child,
3257                                           &trs, "%s", "rcu_torture_read_exit_child");
3258                         if (IS_ERR(tsp)) {
3259                                 TOROUT_ERRSTRING("out of memory");
3260                                 errexit = true;
3261                                 break;
3262                         }
3263                         cond_resched();
3264                         kthread_stop(tsp);
3265                         n_read_exits++;
3266                 }
3267                 VERBOSE_TOROUT_STRING("rcu_torture_read_exit: End of episode");
3268                 rcu_barrier(); // Wait for task_struct free, avoid OOM.
3269                 i = 0;
3270                 for (; !errexit && !READ_ONCE(read_exit_child_stop) && i < read_exit_delay; i++)
3271                         schedule_timeout_uninterruptible(HZ);
3272         } while (!errexit && !READ_ONCE(read_exit_child_stop));
3273
3274         // Clean up and exit.
3275         smp_store_release(&read_exit_child_stopped, true); // After reaping.
3276         smp_mb(); // Store before wakeup.
3277         wake_up(&read_exit_wq);
3278         while (!torture_must_stop())
3279                 schedule_timeout_uninterruptible(HZ / 20);
3280         torture_kthread_stopping("rcu_torture_read_exit");
3281         return 0;
3282 }
3283
3284 static int rcu_torture_read_exit_init(void)
3285 {
3286         if (read_exit_burst <= 0)
3287                 return 0;
3288         init_waitqueue_head(&read_exit_wq);
3289         read_exit_child_stop = false;
3290         read_exit_child_stopped = false;
3291         return torture_create_kthread(rcu_torture_read_exit, NULL,
3292                                       read_exit_task);
3293 }
3294
3295 static void rcu_torture_read_exit_cleanup(void)
3296 {
3297         if (!read_exit_task)
3298                 return;
3299         WRITE_ONCE(read_exit_child_stop, true);
3300         smp_mb(); // Above write before wait.
3301         wait_event(read_exit_wq, smp_load_acquire(&read_exit_child_stopped));
3302         torture_stop_kthread(rcutorture_read_exit, read_exit_task);
3303 }
3304
3305 static void rcutorture_test_nmis(int n)
3306 {
3307 #if IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)
3308         int cpu;
3309         int dumpcpu;
3310         int i;
3311
3312         for (i = 0; i < n; i++) {
3313                 preempt_disable();
3314                 cpu = smp_processor_id();
3315                 dumpcpu = cpu + 1;
3316                 if (dumpcpu >= nr_cpu_ids)
3317                         dumpcpu = 0;
3318                 pr_alert("%s: CPU %d invoking dump_cpu_task(%d)\n", __func__, cpu, dumpcpu);
3319                 dump_cpu_task(dumpcpu);
3320                 preempt_enable();
3321                 schedule_timeout_uninterruptible(15 * HZ);
3322         }
3323 #else // #if IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)
3324         WARN_ONCE(n, "Non-zero rcutorture.test_nmis=%d permitted only when rcutorture is built in.\n", test_nmis);
3325 #endif // #else // #if IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)
3326 }
3327
3328 static enum cpuhp_state rcutor_hp;
3329
3330 static void
3331 rcu_torture_cleanup(void)
3332 {
3333         int firsttime;
3334         int flags = 0;
3335         unsigned long gp_seq = 0;
3336         int i;
3337
3338         if (torture_cleanup_begin()) {
3339                 if (cur_ops->cb_barrier != NULL) {
3340                         pr_info("%s: Invoking %pS().\n", __func__, cur_ops->cb_barrier);
3341                         cur_ops->cb_barrier();
3342                 }
3343                 rcu_gp_slow_unregister(NULL);
3344                 return;
3345         }
3346         if (!cur_ops) {
3347                 torture_cleanup_end();
3348                 rcu_gp_slow_unregister(NULL);
3349                 return;
3350         }
3351
3352         rcutorture_test_nmis(test_nmis);
3353
3354         if (cur_ops->gp_kthread_dbg)
3355                 cur_ops->gp_kthread_dbg();
3356         rcu_torture_read_exit_cleanup();
3357         rcu_torture_barrier_cleanup();
3358         rcu_torture_fwd_prog_cleanup();
3359         torture_stop_kthread(rcu_torture_stall, stall_task);
3360         torture_stop_kthread(rcu_torture_writer, writer_task);
3361
3362         if (nocb_tasks) {
3363                 for (i = 0; i < nrealnocbers; i++)
3364                         torture_stop_kthread(rcu_nocb_toggle, nocb_tasks[i]);
3365                 kfree(nocb_tasks);
3366                 nocb_tasks = NULL;
3367         }
3368
3369         if (reader_tasks) {
3370                 for (i = 0; i < nrealreaders; i++)
3371                         torture_stop_kthread(rcu_torture_reader,
3372                                              reader_tasks[i]);
3373                 kfree(reader_tasks);
3374                 reader_tasks = NULL;
3375         }
3376         kfree(rcu_torture_reader_mbchk);
3377         rcu_torture_reader_mbchk = NULL;
3378
3379         if (fakewriter_tasks) {
3380                 for (i = 0; i < nfakewriters; i++)
3381                         torture_stop_kthread(rcu_torture_fakewriter,
3382                                              fakewriter_tasks[i]);
3383                 kfree(fakewriter_tasks);
3384                 fakewriter_tasks = NULL;
3385         }
3386
3387         rcutorture_get_gp_data(cur_ops->ttype, &flags, &gp_seq);
3388         srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp, &flags, &gp_seq);
3389         pr_alert("%s:  End-test grace-period state: g%ld f%#x total-gps=%ld\n",
3390                  cur_ops->name, (long)gp_seq, flags,
3391                  rcutorture_seq_diff(gp_seq, start_gp_seq));
3392         torture_stop_kthread(rcu_torture_stats, stats_task);
3393         torture_stop_kthread(rcu_torture_fqs, fqs_task);
3394         if (rcu_torture_can_boost() && rcutor_hp >= 0)
3395                 cpuhp_remove_state(rcutor_hp);
3396
3397         /*
3398          * Wait for all RCU callbacks to fire, then do torture-type-specific
3399          * cleanup operations.
3400          */
3401         if (cur_ops->cb_barrier != NULL) {
3402                 pr_info("%s: Invoking %pS().\n", __func__, cur_ops->cb_barrier);
3403                 cur_ops->cb_barrier();
3404         }
3405         if (cur_ops->cleanup != NULL)
3406                 cur_ops->cleanup();
3407
3408         rcu_torture_mem_dump_obj();
3409
3410         rcu_torture_stats_print();  /* -After- the stats thread is stopped! */
3411
3412         if (err_segs_recorded) {
3413                 pr_alert("Failure/close-call rcutorture reader segments:\n");
3414                 if (rt_read_nsegs == 0)
3415                         pr_alert("\t: No segments recorded!!!\n");
3416                 firsttime = 1;
3417                 for (i = 0; i < rt_read_nsegs; i++) {
3418                         pr_alert("\t%d: %#x ", i, err_segs[i].rt_readstate);
3419                         if (err_segs[i].rt_delay_jiffies != 0) {
3420                                 pr_cont("%s%ldjiffies", firsttime ? "" : "+",
3421                                         err_segs[i].rt_delay_jiffies);
3422                                 firsttime = 0;
3423                         }
3424                         if (err_segs[i].rt_delay_ms != 0) {
3425                                 pr_cont("%s%ldms", firsttime ? "" : "+",
3426                                         err_segs[i].rt_delay_ms);
3427                                 firsttime = 0;
3428                         }
3429                         if (err_segs[i].rt_delay_us != 0) {
3430                                 pr_cont("%s%ldus", firsttime ? "" : "+",
3431                                         err_segs[i].rt_delay_us);
3432                                 firsttime = 0;
3433                         }
3434                         pr_cont("%s\n",
3435                                 err_segs[i].rt_preempted ? "preempted" : "");
3436
3437                 }
3438         }
3439         if (atomic_read(&n_rcu_torture_error) || n_rcu_torture_barrier_error)
3440                 rcu_torture_print_module_parms(cur_ops, "End of test: FAILURE");
3441         else if (torture_onoff_failures())
3442                 rcu_torture_print_module_parms(cur_ops,
3443                                                "End of test: RCU_HOTPLUG");
3444         else
3445                 rcu_torture_print_module_parms(cur_ops, "End of test: SUCCESS");
3446         torture_cleanup_end();
3447         rcu_gp_slow_unregister(&rcu_fwd_cb_nodelay);
3448 }
3449
3450 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3451 static void rcu_torture_leak_cb(struct rcu_head *rhp)
3452 {
3453 }
3454
3455 static void rcu_torture_err_cb(struct rcu_head *rhp)
3456 {
3457         /*
3458          * This -might- happen due to race conditions, but is unlikely.
3459          * The scenario that leads to this happening is that the
3460          * first of the pair of duplicate callbacks is queued,
3461          * someone else starts a grace period that includes that
3462          * callback, then the second of the pair must wait for the
3463          * next grace period.  Unlikely, but can happen.  If it
3464          * does happen, the debug-objects subsystem won't have splatted.
3465          */
3466         pr_alert("%s: duplicated callback was invoked.\n", KBUILD_MODNAME);
3467 }
3468 #endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3469
3470 /*
3471  * Verify that double-free causes debug-objects to complain, but only
3472  * if CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.  Otherwise, say that the test
3473  * cannot be carried out.
3474  */
3475 static void rcu_test_debug_objects(void)
3476 {
3477 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3478         struct rcu_head rh1;
3479         struct rcu_head rh2;
3480         struct rcu_head *rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
3481
3482         init_rcu_head_on_stack(&rh1);
3483         init_rcu_head_on_stack(&rh2);
3484         pr_alert("%s: WARN: Duplicate call_rcu() test starting.\n", KBUILD_MODNAME);
3485
3486         /* Try to queue the rh2 pair of callbacks for the same grace period. */
3487         preempt_disable(); /* Prevent preemption from interrupting test. */
3488         rcu_read_lock(); /* Make it impossible to finish a grace period. */
3489         call_rcu_hurry(&rh1, rcu_torture_leak_cb); /* Start grace period. */
3490         local_irq_disable(); /* Make it harder to start a new grace period. */
3491         call_rcu_hurry(&rh2, rcu_torture_leak_cb);
3492         call_rcu_hurry(&rh2, rcu_torture_err_cb); /* Duplicate callback. */
3493         if (rhp) {
3494                 call_rcu_hurry(rhp, rcu_torture_leak_cb);
3495                 call_rcu_hurry(rhp, rcu_torture_err_cb); /* Another duplicate callback. */
3496         }
3497         local_irq_enable();
3498         rcu_read_unlock();
3499         preempt_enable();
3500
3501         /* Wait for them all to get done so we can safely return. */
3502         rcu_barrier();
3503         pr_alert("%s: WARN: Duplicate call_rcu() test complete.\n", KBUILD_MODNAME);
3504         destroy_rcu_head_on_stack(&rh1);
3505         destroy_rcu_head_on_stack(&rh2);
3506         kfree(rhp);
3507 #else /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3508         pr_alert("%s: !CONFIG_DEBUG_OBJECTS_RCU_HEAD, not testing duplicate call_rcu()\n", KBUILD_MODNAME);
3509 #endif /* #else #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
3510 }
3511
3512 static void rcutorture_sync(void)
3513 {
3514         static unsigned long n;
3515
3516         if (cur_ops->sync && !(++n & 0xfff))
3517                 cur_ops->sync();
3518 }
3519
3520 static DEFINE_MUTEX(mut0);
3521 static DEFINE_MUTEX(mut1);
3522 static DEFINE_MUTEX(mut2);
3523 static DEFINE_MUTEX(mut3);
3524 static DEFINE_MUTEX(mut4);
3525 static DEFINE_MUTEX(mut5);
3526 static DEFINE_MUTEX(mut6);
3527 static DEFINE_MUTEX(mut7);
3528 static DEFINE_MUTEX(mut8);
3529 static DEFINE_MUTEX(mut9);
3530
3531 static DECLARE_RWSEM(rwsem0);
3532 static DECLARE_RWSEM(rwsem1);
3533 static DECLARE_RWSEM(rwsem2);
3534 static DECLARE_RWSEM(rwsem3);
3535 static DECLARE_RWSEM(rwsem4);
3536 static DECLARE_RWSEM(rwsem5);
3537 static DECLARE_RWSEM(rwsem6);
3538 static DECLARE_RWSEM(rwsem7);
3539 static DECLARE_RWSEM(rwsem8);
3540 static DECLARE_RWSEM(rwsem9);
3541
3542 DEFINE_STATIC_SRCU(srcu0);
3543 DEFINE_STATIC_SRCU(srcu1);
3544 DEFINE_STATIC_SRCU(srcu2);
3545 DEFINE_STATIC_SRCU(srcu3);
3546 DEFINE_STATIC_SRCU(srcu4);
3547 DEFINE_STATIC_SRCU(srcu5);
3548 DEFINE_STATIC_SRCU(srcu6);
3549 DEFINE_STATIC_SRCU(srcu7);
3550 DEFINE_STATIC_SRCU(srcu8);
3551 DEFINE_STATIC_SRCU(srcu9);
3552
3553 static int srcu_lockdep_next(const char *f, const char *fl, const char *fs, const char *fu, int i,
3554                              int cyclelen, int deadlock)
3555 {
3556         int j = i + 1;
3557
3558         if (j >= cyclelen)
3559                 j = deadlock ? 0 : -1;
3560         if (j >= 0)
3561                 pr_info("%s: %s(%d), %s(%d), %s(%d)\n", f, fl, i, fs, j, fu, i);
3562         else
3563                 pr_info("%s: %s(%d), %s(%d)\n", f, fl, i, fu, i);
3564         return j;
3565 }
3566
3567 // Test lockdep on SRCU-based deadlock scenarios.
3568 static void rcu_torture_init_srcu_lockdep(void)
3569 {
3570         int cyclelen;
3571         int deadlock;
3572         bool err = false;
3573         int i;
3574         int j;
3575         int idx;
3576         struct mutex *muts[] = { &mut0, &mut1, &mut2, &mut3, &mut4,
3577                                  &mut5, &mut6, &mut7, &mut8, &mut9 };
3578         struct rw_semaphore *rwsems[] = { &rwsem0, &rwsem1, &rwsem2, &rwsem3, &rwsem4,
3579                                           &rwsem5, &rwsem6, &rwsem7, &rwsem8, &rwsem9 };
3580         struct srcu_struct *srcus[] = { &srcu0, &srcu1, &srcu2, &srcu3, &srcu4,
3581                                         &srcu5, &srcu6, &srcu7, &srcu8, &srcu9 };
3582         int testtype;
3583
3584         if (!test_srcu_lockdep)
3585                 return;
3586
3587         deadlock = test_srcu_lockdep / 1000;
3588         testtype = (test_srcu_lockdep / 10) % 100;
3589         cyclelen = test_srcu_lockdep % 10;
3590         WARN_ON_ONCE(ARRAY_SIZE(muts) != ARRAY_SIZE(srcus));
3591         if (WARN_ONCE(deadlock != !!deadlock,
3592                       "%s: test_srcu_lockdep=%d and deadlock digit %d must be zero or one.\n",
3593                       __func__, test_srcu_lockdep, deadlock))
3594                 err = true;
3595         if (WARN_ONCE(cyclelen <= 0,
3596                       "%s: test_srcu_lockdep=%d and cycle-length digit %d must be greater than zero.\n",
3597                       __func__, test_srcu_lockdep, cyclelen))
3598                 err = true;
3599         if (err)
3600                 goto err_out;
3601
3602         if (testtype == 0) {
3603                 pr_info("%s: test_srcu_lockdep = %05d: SRCU %d-way %sdeadlock.\n",
3604                         __func__, test_srcu_lockdep, cyclelen, deadlock ? "" : "non-");
3605                 if (deadlock && cyclelen == 1)
3606                         pr_info("%s: Expect hang.\n", __func__);
3607                 for (i = 0; i < cyclelen; i++) {
3608                         j = srcu_lockdep_next(__func__, "srcu_read_lock", "synchronize_srcu",
3609                                               "srcu_read_unlock", i, cyclelen, deadlock);
3610                         idx = srcu_read_lock(srcus[i]);
3611                         if (j >= 0)
3612                                 synchronize_srcu(srcus[j]);
3613                         srcu_read_unlock(srcus[i], idx);
3614                 }
3615                 return;
3616         }
3617
3618         if (testtype == 1) {
3619                 pr_info("%s: test_srcu_lockdep = %05d: SRCU/mutex %d-way %sdeadlock.\n",
3620                         __func__, test_srcu_lockdep, cyclelen, deadlock ? "" : "non-");
3621                 for (i = 0; i < cyclelen; i++) {
3622                         pr_info("%s: srcu_read_lock(%d), mutex_lock(%d), mutex_unlock(%d), srcu_read_unlock(%d)\n",
3623                                 __func__, i, i, i, i);
3624                         idx = srcu_read_lock(srcus[i]);
3625                         mutex_lock(muts[i]);
3626                         mutex_unlock(muts[i]);
3627                         srcu_read_unlock(srcus[i], idx);
3628
3629                         j = srcu_lockdep_next(__func__, "mutex_lock", "synchronize_srcu",
3630                                               "mutex_unlock", i, cyclelen, deadlock);
3631                         mutex_lock(muts[i]);
3632                         if (j >= 0)
3633                                 synchronize_srcu(srcus[j]);
3634                         mutex_unlock(muts[i]);
3635                 }
3636                 return;
3637         }
3638
3639         if (testtype == 2) {
3640                 pr_info("%s: test_srcu_lockdep = %05d: SRCU/rwsem %d-way %sdeadlock.\n",
3641                         __func__, test_srcu_lockdep, cyclelen, deadlock ? "" : "non-");
3642                 for (i = 0; i < cyclelen; i++) {
3643                         pr_info("%s: srcu_read_lock(%d), down_read(%d), up_read(%d), srcu_read_unlock(%d)\n",
3644                                 __func__, i, i, i, i);
3645                         idx = srcu_read_lock(srcus[i]);
3646                         down_read(rwsems[i]);
3647                         up_read(rwsems[i]);
3648                         srcu_read_unlock(srcus[i], idx);
3649
3650                         j = srcu_lockdep_next(__func__, "down_write", "synchronize_srcu",
3651                                               "up_write", i, cyclelen, deadlock);
3652                         down_write(rwsems[i]);
3653                         if (j >= 0)
3654                                 synchronize_srcu(srcus[j]);
3655                         up_write(rwsems[i]);
3656                 }
3657                 return;
3658         }
3659
3660 #ifdef CONFIG_TASKS_TRACE_RCU
3661         if (testtype == 3) {
3662                 pr_info("%s: test_srcu_lockdep = %05d: SRCU and Tasks Trace RCU %d-way %sdeadlock.\n",
3663                         __func__, test_srcu_lockdep, cyclelen, deadlock ? "" : "non-");
3664                 if (deadlock && cyclelen == 1)
3665                         pr_info("%s: Expect hang.\n", __func__);
3666                 for (i = 0; i < cyclelen; i++) {
3667                         char *fl = i == 0 ? "rcu_read_lock_trace" : "srcu_read_lock";
3668                         char *fs = i == cyclelen - 1 ? "synchronize_rcu_tasks_trace"
3669                                                      : "synchronize_srcu";
3670                         char *fu = i == 0 ? "rcu_read_unlock_trace" : "srcu_read_unlock";
3671
3672                         j = srcu_lockdep_next(__func__, fl, fs, fu, i, cyclelen, deadlock);
3673                         if (i == 0)
3674                                 rcu_read_lock_trace();
3675                         else
3676                                 idx = srcu_read_lock(srcus[i]);
3677                         if (j >= 0) {
3678                                 if (i == cyclelen - 1)
3679                                         synchronize_rcu_tasks_trace();
3680                                 else
3681                                         synchronize_srcu(srcus[j]);
3682                         }
3683                         if (i == 0)
3684                                 rcu_read_unlock_trace();
3685                         else
3686                                 srcu_read_unlock(srcus[i], idx);
3687                 }
3688                 return;
3689         }
3690 #endif // #ifdef CONFIG_TASKS_TRACE_RCU
3691
3692 err_out:
3693         pr_info("%s: test_srcu_lockdep = %05d does nothing.\n", __func__, test_srcu_lockdep);
3694         pr_info("%s: test_srcu_lockdep = DNNL.\n", __func__);
3695         pr_info("%s: D: Deadlock if nonzero.\n", __func__);
3696         pr_info("%s: NN: Test number, 0=SRCU, 1=SRCU/mutex, 2=SRCU/rwsem, 3=SRCU/Tasks Trace RCU.\n", __func__);
3697         pr_info("%s: L: Cycle length.\n", __func__);
3698         if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU))
3699                 pr_info("%s: NN=3 disallowed because kernel is built with CONFIG_TASKS_TRACE_RCU=n\n", __func__);
3700 }
3701
3702 static int __init
3703 rcu_torture_init(void)
3704 {
3705         long i;
3706         int cpu;
3707         int firsterr = 0;
3708         int flags = 0;
3709         unsigned long gp_seq = 0;
3710         static struct rcu_torture_ops *torture_ops[] = {
3711                 &rcu_ops, &rcu_busted_ops, &srcu_ops, &srcud_ops, &busted_srcud_ops,
3712                 TASKS_OPS TASKS_RUDE_OPS TASKS_TRACING_OPS
3713                 &trivial_ops,
3714         };
3715
3716         if (!torture_init_begin(torture_type, verbose))
3717                 return -EBUSY;
3718
3719         /* Process args and tell the world that the torturer is on the job. */
3720         for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
3721                 cur_ops = torture_ops[i];
3722                 if (strcmp(torture_type, cur_ops->name) == 0)
3723                         break;
3724         }
3725         if (i == ARRAY_SIZE(torture_ops)) {
3726                 pr_alert("rcu-torture: invalid torture type: \"%s\"\n",
3727                          torture_type);
3728                 pr_alert("rcu-torture types:");
3729                 for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
3730                         pr_cont(" %s", torture_ops[i]->name);
3731                 pr_cont("\n");
3732                 firsterr = -EINVAL;
3733                 cur_ops = NULL;
3734                 goto unwind;
3735         }
3736         if (cur_ops->fqs == NULL && fqs_duration != 0) {
3737                 pr_alert("rcu-torture: ->fqs NULL and non-zero fqs_duration, fqs disabled.\n");
3738                 fqs_duration = 0;
3739         }
3740         if (nocbs_nthreads != 0 && (cur_ops != &rcu_ops ||
3741                                     !IS_ENABLED(CONFIG_RCU_NOCB_CPU))) {
3742                 pr_alert("rcu-torture types: %s and CONFIG_RCU_NOCB_CPU=%d, nocb toggle disabled.\n",
3743                          cur_ops->name, IS_ENABLED(CONFIG_RCU_NOCB_CPU));
3744                 nocbs_nthreads = 0;
3745         }
3746         if (cur_ops->init)
3747                 cur_ops->init();
3748
3749         rcu_torture_init_srcu_lockdep();
3750
3751         if (nreaders >= 0) {
3752                 nrealreaders = nreaders;
3753         } else {
3754                 nrealreaders = num_online_cpus() - 2 - nreaders;
3755                 if (nrealreaders <= 0)
3756                         nrealreaders = 1;
3757         }
3758         rcu_torture_print_module_parms(cur_ops, "Start of test");
3759         rcutorture_get_gp_data(cur_ops->ttype, &flags, &gp_seq);
3760         srcutorture_get_gp_data(cur_ops->ttype, srcu_ctlp, &flags, &gp_seq);
3761         start_gp_seq = gp_seq;
3762         pr_alert("%s:  Start-test grace-period state: g%ld f%#x\n",
3763                  cur_ops->name, (long)gp_seq, flags);
3764
3765         /* Set up the freelist. */
3766
3767         INIT_LIST_HEAD(&rcu_torture_freelist);
3768         for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++) {
3769                 rcu_tortures[i].rtort_mbtest = 0;
3770                 list_add_tail(&rcu_tortures[i].rtort_free,
3771                               &rcu_torture_freelist);
3772         }
3773
3774         /* Initialize the statistics so that each run gets its own numbers. */
3775
3776         rcu_torture_current = NULL;
3777         rcu_torture_current_version = 0;
3778         atomic_set(&n_rcu_torture_alloc, 0);
3779         atomic_set(&n_rcu_torture_alloc_fail, 0);
3780         atomic_set(&n_rcu_torture_free, 0);
3781         atomic_set(&n_rcu_torture_mberror, 0);
3782         atomic_set(&n_rcu_torture_mbchk_fail, 0);
3783         atomic_set(&n_rcu_torture_mbchk_tries, 0);
3784         atomic_set(&n_rcu_torture_error, 0);
3785         n_rcu_torture_barrier_error = 0;
3786         n_rcu_torture_boost_ktrerror = 0;
3787         n_rcu_torture_boost_failure = 0;
3788         n_rcu_torture_boosts = 0;
3789         for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
3790                 atomic_set(&rcu_torture_wcount[i], 0);
3791         for_each_possible_cpu(cpu) {
3792                 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
3793                         per_cpu(rcu_torture_count, cpu)[i] = 0;
3794                         per_cpu(rcu_torture_batch, cpu)[i] = 0;
3795                 }
3796         }
3797         err_segs_recorded = 0;
3798         rt_read_nsegs = 0;
3799
3800         /* Start up the kthreads. */
3801
3802         rcu_torture_write_types();
3803         firsterr = torture_create_kthread(rcu_torture_writer, NULL,
3804                                           writer_task);
3805         if (torture_init_error(firsterr))
3806                 goto unwind;
3807         if (nfakewriters > 0) {
3808                 fakewriter_tasks = kcalloc(nfakewriters,
3809                                            sizeof(fakewriter_tasks[0]),
3810                                            GFP_KERNEL);
3811                 if (fakewriter_tasks == NULL) {
3812                         TOROUT_ERRSTRING("out of memory");
3813                         firsterr = -ENOMEM;
3814                         goto unwind;
3815                 }
3816         }
3817         for (i = 0; i < nfakewriters; i++) {
3818                 firsterr = torture_create_kthread(rcu_torture_fakewriter,
3819                                                   NULL, fakewriter_tasks[i]);
3820                 if (torture_init_error(firsterr))
3821                         goto unwind;
3822         }
3823         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
3824                                GFP_KERNEL);
3825         rcu_torture_reader_mbchk = kcalloc(nrealreaders, sizeof(*rcu_torture_reader_mbchk),
3826                                            GFP_KERNEL);
3827         if (!reader_tasks || !rcu_torture_reader_mbchk) {
3828                 TOROUT_ERRSTRING("out of memory");
3829                 firsterr = -ENOMEM;
3830                 goto unwind;
3831         }
3832         for (i = 0; i < nrealreaders; i++) {
3833                 rcu_torture_reader_mbchk[i].rtc_chkrdr = -1;
3834                 firsterr = torture_create_kthread(rcu_torture_reader, (void *)i,
3835                                                   reader_tasks[i]);
3836                 if (torture_init_error(firsterr))
3837                         goto unwind;
3838         }
3839         nrealnocbers = nocbs_nthreads;
3840         if (WARN_ON(nrealnocbers < 0))
3841                 nrealnocbers = 1;
3842         if (WARN_ON(nocbs_toggle < 0))
3843                 nocbs_toggle = HZ;
3844         if (nrealnocbers > 0) {
3845                 nocb_tasks = kcalloc(nrealnocbers, sizeof(nocb_tasks[0]), GFP_KERNEL);
3846                 if (nocb_tasks == NULL) {
3847                         TOROUT_ERRSTRING("out of memory");
3848                         firsterr = -ENOMEM;
3849                         goto unwind;
3850                 }
3851         } else {
3852                 nocb_tasks = NULL;
3853         }
3854         for (i = 0; i < nrealnocbers; i++) {
3855                 firsterr = torture_create_kthread(rcu_nocb_toggle, NULL, nocb_tasks[i]);
3856                 if (torture_init_error(firsterr))
3857                         goto unwind;
3858         }
3859         if (stat_interval > 0) {
3860                 firsterr = torture_create_kthread(rcu_torture_stats, NULL,
3861                                                   stats_task);
3862                 if (torture_init_error(firsterr))
3863                         goto unwind;
3864         }
3865         if (test_no_idle_hz && shuffle_interval > 0) {
3866                 firsterr = torture_shuffle_init(shuffle_interval * HZ);
3867                 if (torture_init_error(firsterr))
3868                         goto unwind;
3869         }
3870         if (stutter < 0)
3871                 stutter = 0;
3872         if (stutter) {
3873                 int t;
3874
3875                 t = cur_ops->stall_dur ? cur_ops->stall_dur() : stutter * HZ;
3876                 firsterr = torture_stutter_init(stutter * HZ, t);
3877                 if (torture_init_error(firsterr))
3878                         goto unwind;
3879         }
3880         if (fqs_duration < 0)
3881                 fqs_duration = 0;
3882         if (fqs_holdoff < 0)
3883                 fqs_holdoff = 0;
3884         if (fqs_duration && fqs_holdoff) {
3885                 /* Create the fqs thread */
3886                 firsterr = torture_create_kthread(rcu_torture_fqs, NULL,
3887                                                   fqs_task);
3888                 if (torture_init_error(firsterr))
3889                         goto unwind;
3890         }
3891         if (test_boost_interval < 1)
3892                 test_boost_interval = 1;
3893         if (test_boost_duration < 2)
3894                 test_boost_duration = 2;
3895         if (rcu_torture_can_boost()) {
3896
3897                 boost_starttime = jiffies + test_boost_interval * HZ;
3898
3899                 firsterr = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "RCU_TORTURE",
3900                                              rcutorture_booster_init,
3901                                              rcutorture_booster_cleanup);
3902                 rcutor_hp = firsterr;
3903                 if (torture_init_error(firsterr))
3904                         goto unwind;
3905         }
3906         shutdown_jiffies = jiffies + shutdown_secs * HZ;
3907         firsterr = torture_shutdown_init(shutdown_secs, rcu_torture_cleanup);
3908         if (torture_init_error(firsterr))
3909                 goto unwind;
3910         firsterr = torture_onoff_init(onoff_holdoff * HZ, onoff_interval,
3911                                       rcutorture_sync);
3912         if (torture_init_error(firsterr))
3913                 goto unwind;
3914         firsterr = rcu_torture_stall_init();
3915         if (torture_init_error(firsterr))
3916                 goto unwind;
3917         firsterr = rcu_torture_fwd_prog_init();
3918         if (torture_init_error(firsterr))
3919                 goto unwind;
3920         firsterr = rcu_torture_barrier_init();
3921         if (torture_init_error(firsterr))
3922                 goto unwind;
3923         firsterr = rcu_torture_read_exit_init();
3924         if (torture_init_error(firsterr))
3925                 goto unwind;
3926         if (object_debug)
3927                 rcu_test_debug_objects();
3928         torture_init_end();
3929         rcu_gp_slow_register(&rcu_fwd_cb_nodelay);
3930         return 0;
3931
3932 unwind:
3933         torture_init_end();
3934         rcu_torture_cleanup();
3935         if (shutdown_secs) {
3936                 WARN_ON(!IS_MODULE(CONFIG_RCU_TORTURE_TEST));
3937                 kernel_power_off();
3938         }
3939         return firsterr;
3940 }
3941
3942 module_init(rcu_torture_init);
3943 module_exit(rcu_torture_cleanup);