kernel/rcu/rcuperf.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Read-Copy Update module-based performance-test facility
   4  *
   5  * Copyright (C) IBM Corporation, 2015
   6  *
   7  * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
   8  */
   9
  10 #define pr_fmt(fmt) fmt
  11
  12 #include <linux/types.h>
  13 #include <linux/kernel.h>
  14 #include <linux/init.h>
  15 #include <linux/module.h>
  16 #include <linux/kthread.h>
  17 #include <linux/err.h>
  18 #include <linux/spinlock.h>
  19 #include <linux/smp.h>
  20 #include <linux/rcupdate.h>
  21 #include <linux/interrupt.h>
  22 #include <linux/sched.h>
  23 #include <uapi/linux/sched/types.h>
  24 #include <linux/atomic.h>
  25 #include <linux/bitops.h>
  26 #include <linux/completion.h>
  27 #include <linux/moduleparam.h>
  28 #include <linux/percpu.h>
  29 #include <linux/notifier.h>
  30 #include <linux/reboot.h>
  31 #include <linux/freezer.h>
  32 #include <linux/cpu.h>
  33 #include <linux/delay.h>
  34 #include <linux/stat.h>
  35 #include <linux/srcu.h>
  36 #include <linux/slab.h>
  37 #include <asm/byteorder.h>
  38 #include <linux/torture.h>
  39 #include <linux/vmalloc.h>
  40
  41 #include "rcu.h"
  42
  43 MODULE_LICENSE("GPL");
  44 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
  45
  46 #define PERF_FLAG "-perf:"
  47 #define PERFOUT_STRING(s) \
  48         pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
  49 #define VERBOSE_PERFOUT_STRING(s) \
  50         do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
  51 #define VERBOSE_PERFOUT_ERRSTRING(s) \
  52         do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
  53
  54 /*
  55  * The intended use cases for the nreaders and nwriters module parameters
  56  * are as follows:
  57  *
  58  * 1.   Specify only the nr_cpus kernel boot parameter.  This will
  59  *      set both nreaders and nwriters to the value specified by
  60  *      nr_cpus for a mixed reader/writer test.
  61  *
  62  * 2.   Specify the nr_cpus kernel boot parameter, but set
  63  *      rcuperf.nreaders to zero.  This will set nwriters to the
  64  *      value specified by nr_cpus for an update-only test.
  65  *
  66  * 3.   Specify the nr_cpus kernel boot parameter, but set
  67  *      rcuperf.nwriters to zero.  This will set nreaders to the
  68  *      value specified by nr_cpus for a read-only test.
  69  *
  70  * Various other use cases may of course be specified.
  71  */
  72
  73 #ifdef MODULE
  74 # define RCUPERF_SHUTDOWN 0
  75 #else
  76 # define RCUPERF_SHUTDOWN 1
  77 #endif
  78
  79 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
  80 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
  81 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
  82 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
  83 torture_param(int, nreaders, -1, "Number of RCU reader threads");
  84 torture_param(int, nwriters, -1, "Number of RCU updater threads");
  85 torture_param(bool, shutdown, RCUPERF_SHUTDOWN,
  86               "Shutdown at end of performance tests.");
  87 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
  88 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
  89
  90 static char *perf_type = "rcu";
  91 module_param(perf_type, charp, 0444);
  92 MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, srcu, ...)");
  93
  94 static int nrealreaders;
  95 static int nrealwriters;
  96 static struct task_struct **writer_tasks;
  97 static struct task_struct **reader_tasks;
  98 static struct task_struct *shutdown_task;
  99
 100 static u64 **writer_durations;
 101 static int *writer_n_durations;
 102 static atomic_t n_rcu_perf_reader_started;
 103 static atomic_t n_rcu_perf_writer_started;
 104 static atomic_t n_rcu_perf_writer_finished;
 105 static wait_queue_head_t shutdown_wq;
 106 static u64 t_rcu_perf_writer_started;
 107 static u64 t_rcu_perf_writer_finished;
 108 static unsigned long b_rcu_perf_writer_started;
 109 static unsigned long b_rcu_perf_writer_finished;
 110 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
 111
 112 #define MAX_MEAS 10000
 113 #define MIN_MEAS 100
 114
 115 /*
 116  * Operations vector for selecting different types of tests.
 117  */
 118
 119 struct rcu_perf_ops {
 120         int ptype;
 121         void (*init)(void);
 122         void (*cleanup)(void);
 123         int (*readlock)(void);
 124         void (*readunlock)(int idx);
 125         unsigned long (*get_gp_seq)(void);
 126         unsigned long (*gp_diff)(unsigned long new, unsigned long old);
 127         unsigned long (*exp_completed)(void);
 128         void (*async)(struct rcu_head *head, rcu_callback_t func);
 129         void (*gp_barrier)(void);
 130         void (*sync)(void);
 131         void (*exp_sync)(void);
 132         const char *name;
 133 };
 134
 135 static struct rcu_perf_ops *cur_ops;
 136
 137 /*
 138  * Definitions for rcu perf testing.
 139  */
 140
 141 static int rcu_perf_read_lock(void) __acquires(RCU)
 142 {
 143         rcu_read_lock();
 144         return 0;
 145 }
 146
 147 static void rcu_perf_read_unlock(int idx) __releases(RCU)
 148 {
 149         rcu_read_unlock();
 150 }
 151
 152 static unsigned long __maybe_unused rcu_no_completed(void)
 153 {
 154         return 0;
 155 }
 156
 157 static void rcu_sync_perf_init(void)
 158 {
 159 }
 160
 161 static struct rcu_perf_ops rcu_ops = {
 162         .ptype          = RCU_FLAVOR,
 163         .init           = rcu_sync_perf_init,
 164         .readlock       = rcu_perf_read_lock,
 165         .readunlock     = rcu_perf_read_unlock,
 166         .get_gp_seq     = rcu_get_gp_seq,
 167         .gp_diff        = rcu_seq_diff,
 168         .exp_completed  = rcu_exp_batches_completed,
 169         .async          = call_rcu,
 170         .gp_barrier     = rcu_barrier,
 171         .sync           = synchronize_rcu,
 172         .exp_sync       = synchronize_rcu_expedited,
 173         .name           = "rcu"
 174 };
 175
 176 /*
 177  * Definitions for srcu perf testing.
 178  */
 179
 180 DEFINE_STATIC_SRCU(srcu_ctl_perf);
 181 static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
 182
 183 static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
 184 {
 185         return srcu_read_lock(srcu_ctlp);
 186 }
 187
 188 static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
 189 {
 190         srcu_read_unlock(srcu_ctlp, idx);
 191 }
 192
 193 static unsigned long srcu_perf_completed(void)
 194 {
 195         return srcu_batches_completed(srcu_ctlp);
 196 }
 197
 198 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
 199 {
 200         call_srcu(srcu_ctlp, head, func);
 201 }
 202
 203 static void srcu_rcu_barrier(void)
 204 {
 205         srcu_barrier(srcu_ctlp);
 206 }
 207
 208 static void srcu_perf_synchronize(void)
 209 {
 210         synchronize_srcu(srcu_ctlp);
 211 }
 212
 213 static void srcu_perf_synchronize_expedited(void)
 214 {
 215         synchronize_srcu_expedited(srcu_ctlp);
 216 }
 217
 218 static struct rcu_perf_ops srcu_ops = {
 219         .ptype          = SRCU_FLAVOR,
 220         .init           = rcu_sync_perf_init,
 221         .readlock       = srcu_perf_read_lock,
 222         .readunlock     = srcu_perf_read_unlock,
 223         .get_gp_seq     = srcu_perf_completed,
 224         .gp_diff        = rcu_seq_diff,
 225         .exp_completed  = srcu_perf_completed,
 226         .async          = srcu_call_rcu,
 227         .gp_barrier     = srcu_rcu_barrier,
 228         .sync           = srcu_perf_synchronize,
 229         .exp_sync       = srcu_perf_synchronize_expedited,
 230         .name           = "srcu"
 231 };
 232
 233 static struct srcu_struct srcud;
 234
 235 static void srcu_sync_perf_init(void)
 236 {
 237         srcu_ctlp = &srcud;
 238         init_srcu_struct(srcu_ctlp);
 239 }
 240
 241 static void srcu_sync_perf_cleanup(void)
 242 {
 243         cleanup_srcu_struct(srcu_ctlp);
 244 }
 245
 246 static struct rcu_perf_ops srcud_ops = {
 247         .ptype          = SRCU_FLAVOR,
 248         .init           = srcu_sync_perf_init,
 249         .cleanup        = srcu_sync_perf_cleanup,
 250         .readlock       = srcu_perf_read_lock,
 251         .readunlock     = srcu_perf_read_unlock,
 252         .get_gp_seq     = srcu_perf_completed,
 253         .gp_diff        = rcu_seq_diff,
 254         .exp_completed  = srcu_perf_completed,
 255         .async          = srcu_call_rcu,
 256         .gp_barrier     = srcu_rcu_barrier,
 257         .sync           = srcu_perf_synchronize,
 258         .exp_sync       = srcu_perf_synchronize_expedited,
 259         .name           = "srcud"
 260 };
 261
 262 /*
 263  * Definitions for RCU-tasks perf testing.
 264  */
 265
 266 static int tasks_perf_read_lock(void)
 267 {
 268         return 0;
 269 }
 270
 271 static void tasks_perf_read_unlock(int idx)
 272 {
 273 }
 274
 275 static struct rcu_perf_ops tasks_ops = {
 276         .ptype          = RCU_TASKS_FLAVOR,
 277         .init           = rcu_sync_perf_init,
 278         .readlock       = tasks_perf_read_lock,
 279         .readunlock     = tasks_perf_read_unlock,
 280         .get_gp_seq     = rcu_no_completed,
 281         .gp_diff        = rcu_seq_diff,
 282         .async          = call_rcu_tasks,
 283         .gp_barrier     = rcu_barrier_tasks,
 284         .sync           = synchronize_rcu_tasks,
 285         .exp_sync       = synchronize_rcu_tasks,
 286         .name           = "tasks"
 287 };
 288
 289 static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old)
 290 {
 291         if (!cur_ops->gp_diff)
 292                 return new - old;
 293         return cur_ops->gp_diff(new, old);
 294 }
 295
 296 /*
 297  * If performance tests complete, wait for shutdown to commence.
 298  */
 299 static void rcu_perf_wait_shutdown(void)
 300 {
 301         cond_resched_tasks_rcu_qs();
 302         if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
 303                 return;
 304         while (!torture_must_stop())
 305                 schedule_timeout_uninterruptible(1);
 306 }
 307
 308 /*
 309  * RCU perf reader kthread.  Repeatedly does empty RCU read-side
 310  * critical section, minimizing update-side interference.
 311  */
 312 static int
 313 rcu_perf_reader(void *arg)
 314 {
 315         unsigned long flags;
 316         int idx;
 317         long me = (long)arg;
 318
 319         VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
 320         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 321         set_user_nice(current, MAX_NICE);
 322         atomic_inc(&n_rcu_perf_reader_started);
 323
 324         do {
 325                 local_irq_save(flags);
 326                 idx = cur_ops->readlock();
 327                 cur_ops->readunlock(idx);
 328                 local_irq_restore(flags);
 329                 rcu_perf_wait_shutdown();
 330         } while (!torture_must_stop());
 331         torture_kthread_stopping("rcu_perf_reader");
 332         return 0;
 333 }
 334
 335 /*
 336  * Callback function for asynchronous grace periods from rcu_perf_writer().
 337  */
 338 static void rcu_perf_async_cb(struct rcu_head *rhp)
 339 {
 340         atomic_dec(this_cpu_ptr(&n_async_inflight));
 341         kfree(rhp);
 342 }
 343
 344 /*
 345  * RCU perf writer kthread.  Repeatedly does a grace period.
 346  */
 347 static int
 348 rcu_perf_writer(void *arg)
 349 {
 350         int i = 0;
 351         int i_max;
 352         long me = (long)arg;
 353         struct rcu_head *rhp = NULL;
 354         struct sched_param sp;
 355         bool started = false, done = false, alldone = false;
 356         u64 t;
 357         u64 *wdp;
 358         u64 *wdpp = writer_durations[me];
 359
 360         VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
 361         WARN_ON(!wdpp);
 362         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 363         sp.sched_priority = 1;
 364         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 365
 366         if (holdoff)
 367                 schedule_timeout_uninterruptible(holdoff * HZ);
 368
 369         /*
 370          * Wait until rcu_end_inkernel_boot() is called for normal GP tests
 371          * so that RCU is not always expedited for normal GP tests.
 372          * The system_state test is approximate, but works well in practice.
 373          */
 374         while (!gp_exp && system_state != SYSTEM_RUNNING)
 375                 schedule_timeout_uninterruptible(1);
 376
 377         t = ktime_get_mono_fast_ns();
 378         if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
 379                 t_rcu_perf_writer_started = t;
 380                 if (gp_exp) {
 381                         b_rcu_perf_writer_started =
 382                                 cur_ops->exp_completed() / 2;
 383                 } else {
 384                         b_rcu_perf_writer_started = cur_ops->get_gp_seq();
 385                 }
 386         }
 387
 388         do {
 389                 if (writer_holdoff)
 390                         udelay(writer_holdoff);
 391                 wdp = &wdpp[i];
 392                 *wdp = ktime_get_mono_fast_ns();
 393                 if (gp_async) {
 394 retry:
 395                         if (!rhp)
 396                                 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
 397                         if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
 398                                 atomic_inc(this_cpu_ptr(&n_async_inflight));
 399                                 cur_ops->async(rhp, rcu_perf_async_cb);
 400                                 rhp = NULL;
 401                         } else if (!kthread_should_stop()) {
 402                                 cur_ops->gp_barrier();
 403                                 goto retry;
 404                         } else {
 405                                 kfree(rhp); /* Because we are stopping. */
 406                         }
 407                 } else if (gp_exp) {
 408                         cur_ops->exp_sync();
 409                 } else {
 410                         cur_ops->sync();
 411                 }
 412                 t = ktime_get_mono_fast_ns();
 413                 *wdp = t - *wdp;
 414                 i_max = i;
 415                 if (!started &&
 416                     atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
 417                         started = true;
 418                 if (!done && i >= MIN_MEAS) {
 419                         done = true;
 420                         sp.sched_priority = 0;
 421                         sched_setscheduler_nocheck(current,
 422                                                    SCHED_NORMAL, &sp);
 423                         pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
 424                                  perf_type, PERF_FLAG, me, MIN_MEAS);
 425                         if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
 426                             nrealwriters) {
 427                                 schedule_timeout_interruptible(10);
 428                                 rcu_ftrace_dump(DUMP_ALL);
 429                                 PERFOUT_STRING("Test complete");
 430                                 t_rcu_perf_writer_finished = t;
 431                                 if (gp_exp) {
 432                                         b_rcu_perf_writer_finished =
 433                                                 cur_ops->exp_completed() / 2;
 434                                 } else {
 435                                         b_rcu_perf_writer_finished =
 436                                                 cur_ops->get_gp_seq();
 437                                 }
 438                                 if (shutdown) {
 439                                         smp_mb(); /* Assign before wake. */
 440                                         wake_up(&shutdown_wq);
 441                                 }
 442                         }
 443                 }
 444                 if (done && !alldone &&
 445                     atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
 446                         alldone = true;
 447                 if (started && !alldone && i < MAX_MEAS - 1)
 448                         i++;
 449                 rcu_perf_wait_shutdown();
 450         } while (!torture_must_stop());
 451         if (gp_async) {
 452                 cur_ops->gp_barrier();
 453         }
 454         writer_n_durations[me] = i_max;
 455         torture_kthread_stopping("rcu_perf_writer");
 456         return 0;
 457 }
 458
 459 static void
 460 rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
 461 {
 462         pr_alert("%s" PERF_FLAG
 463                  "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
 464                  perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
 465 }
 466
 467 static void
 468 rcu_perf_cleanup(void)
 469 {
 470         int i;
 471         int j;
 472         int ngps = 0;
 473         u64 *wdp;
 474         u64 *wdpp;
 475
 476         /*
 477          * Would like warning at start, but everything is expedited
 478          * during the mid-boot phase, so have to wait till the end.
 479          */
 480         if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
 481                 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
 482         if (rcu_gp_is_normal() && gp_exp)
 483                 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
 484         if (gp_exp && gp_async)
 485                 VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
 486
 487         if (torture_cleanup_begin())
 488                 return;
 489         if (!cur_ops) {
 490                 torture_cleanup_end();
 491                 return;
 492         }
 493
 494         if (reader_tasks) {
 495                 for (i = 0; i < nrealreaders; i++)
 496                         torture_stop_kthread(rcu_perf_reader,
 497                                              reader_tasks[i]);
 498                 kfree(reader_tasks);
 499         }
 500
 501         if (writer_tasks) {
 502                 for (i = 0; i < nrealwriters; i++) {
 503                         torture_stop_kthread(rcu_perf_writer,
 504                                              writer_tasks[i]);
 505                         if (!writer_n_durations)
 506                                 continue;
 507                         j = writer_n_durations[i];
 508                         pr_alert("%s%s writer %d gps: %d\n",
 509                                  perf_type, PERF_FLAG, i, j);
 510                         ngps += j;
 511                 }
 512                 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
 513                          perf_type, PERF_FLAG,
 514                          t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
 515                          t_rcu_perf_writer_finished -
 516                          t_rcu_perf_writer_started,
 517                          ngps,
 518                          rcuperf_seq_diff(b_rcu_perf_writer_finished,
 519                                           b_rcu_perf_writer_started));
 520                 for (i = 0; i < nrealwriters; i++) {
 521                         if (!writer_durations)
 522                                 break;
 523                         if (!writer_n_durations)
 524                                 continue;
 525                         wdpp = writer_durations[i];
 526                         if (!wdpp)
 527                                 continue;
 528                         for (j = 0; j <= writer_n_durations[i]; j++) {
 529                                 wdp = &wdpp[j];
 530                                 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
 531                                         perf_type, PERF_FLAG,
 532                                         i, j, *wdp);
 533                                 if (j % 100 == 0)
 534                                         schedule_timeout_uninterruptible(1);
 535                         }
 536                         kfree(writer_durations[i]);
 537                 }
 538                 kfree(writer_tasks);
 539                 kfree(writer_durations);
 540                 kfree(writer_n_durations);
 541         }
 542
 543         /* Do torture-type-specific cleanup operations.  */
 544         if (cur_ops->cleanup != NULL)
 545                 cur_ops->cleanup();
 546
 547         torture_cleanup_end();
 548 }
 549
 550 /*
 551  * Return the number if non-negative.  If -1, the number of CPUs.
 552  * If less than -1, that much less than the number of CPUs, but
 553  * at least one.
 554  */
 555 static int compute_real(int n)
 556 {
 557         int nr;
 558
 559         if (n >= 0) {
 560                 nr = n;
 561         } else {
 562                 nr = num_online_cpus() + 1 + n;
 563                 if (nr <= 0)
 564                         nr = 1;
 565         }
 566         return nr;
 567 }
 568
 569 /*
 570  * RCU perf shutdown kthread.  Just waits to be awakened, then shuts
 571  * down system.
 572  */
 573 static int
 574 rcu_perf_shutdown(void *arg)
 575 {
 576         do {
 577                 wait_event(shutdown_wq,
 578                            atomic_read(&n_rcu_perf_writer_finished) >=
 579                            nrealwriters);
 580         } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters);
 581         smp_mb(); /* Wake before output. */
 582         rcu_perf_cleanup();
 583         kernel_power_off();
 584         return -EINVAL;
 585 }
 586
 587 static int __init
 588 rcu_perf_init(void)
 589 {
 590         long i;
 591         int firsterr = 0;
 592         static struct rcu_perf_ops *perf_ops[] = {
 593                 &rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops,
 594         };
 595
 596         if (!torture_init_begin(perf_type, verbose))
 597                 return -EBUSY;
 598
 599         /* Process args and tell the world that the perf'er is on the job. */
 600         for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
 601                 cur_ops = perf_ops[i];
 602                 if (strcmp(perf_type, cur_ops->name) == 0)
 603                         break;
 604         }
 605         if (i == ARRAY_SIZE(perf_ops)) {
 606                 pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type);
 607                 pr_alert("rcu-perf types:");
 608                 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
 609                         pr_cont(" %s", perf_ops[i]->name);
 610                 pr_cont("\n");
 611                 WARN_ON(!IS_MODULE(CONFIG_RCU_PERF_TEST));
 612                 firsterr = -EINVAL;
 613                 cur_ops = NULL;
 614                 goto unwind;
 615         }
 616         if (cur_ops->init)
 617                 cur_ops->init();
 618
 619         nrealwriters = compute_real(nwriters);
 620         nrealreaders = compute_real(nreaders);
 621         atomic_set(&n_rcu_perf_reader_started, 0);
 622         atomic_set(&n_rcu_perf_writer_started, 0);
 623         atomic_set(&n_rcu_perf_writer_finished, 0);
 624         rcu_perf_print_module_parms(cur_ops, "Start of test");
 625
 626         /* Start up the kthreads. */
 627
 628         if (shutdown) {
 629                 init_waitqueue_head(&shutdown_wq);
 630                 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
 631                                                   shutdown_task);
 632                 if (firsterr)
 633                         goto unwind;
 634                 schedule_timeout_uninterruptible(1);
 635         }
 636         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
 637                                GFP_KERNEL);
 638         if (reader_tasks == NULL) {
 639                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 640                 firsterr = -ENOMEM;
 641                 goto unwind;
 642         }
 643         for (i = 0; i < nrealreaders; i++) {
 644                 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
 645                                                   reader_tasks[i]);
 646                 if (firsterr)
 647                         goto unwind;
 648         }
 649         while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
 650                 schedule_timeout_uninterruptible(1);
 651         writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
 652                                GFP_KERNEL);
 653         writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
 654                                    GFP_KERNEL);
 655         writer_n_durations =
 656                 kcalloc(nrealwriters, sizeof(*writer_n_durations),
 657                         GFP_KERNEL);
 658         if (!writer_tasks || !writer_durations || !writer_n_durations) {
 659                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 660                 firsterr = -ENOMEM;
 661                 goto unwind;
 662         }
 663         for (i = 0; i < nrealwriters; i++) {
 664                 writer_durations[i] =
 665                         kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
 666                                 GFP_KERNEL);
 667                 if (!writer_durations[i]) {
 668                         firsterr = -ENOMEM;
 669                         goto unwind;
 670                 }
 671                 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
 672                                                   writer_tasks[i]);
 673                 if (firsterr)
 674                         goto unwind;
 675         }
 676         torture_init_end();
 677         return 0;
 678
 679 unwind:
 680         torture_init_end();
 681         rcu_perf_cleanup();
 682         return firsterr;
 683 }
 684
 685 module_init(rcu_perf_init);
 686 module_exit(rcu_perf_cleanup);