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, rcu_bh, ...)");
  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 static int rcu_perf_writer_state;
 113 #define RTWS_INIT               0
 114 #define RTWS_ASYNC              1
 115 #define RTWS_BARRIER            2
 116 #define RTWS_EXP_SYNC           3
 117 #define RTWS_SYNC               4
 118 #define RTWS_IDLE               5
 119 #define RTWS_STOPPING           6
 120
 121 #define MAX_MEAS 10000
 122 #define MIN_MEAS 100
 123
 124 /*
 125  * Operations vector for selecting different types of tests.
 126  */
 127
 128 struct rcu_perf_ops {
 129         int ptype;
 130         void (*init)(void);
 131         void (*cleanup)(void);
 132         int (*readlock)(void);
 133         void (*readunlock)(int idx);
 134         unsigned long (*get_gp_seq)(void);
 135         unsigned long (*gp_diff)(unsigned long new, unsigned long old);
 136         unsigned long (*exp_completed)(void);
 137         void (*async)(struct rcu_head *head, rcu_callback_t func);
 138         void (*gp_barrier)(void);
 139         void (*sync)(void);
 140         void (*exp_sync)(void);
 141         const char *name;
 142 };
 143
 144 static struct rcu_perf_ops *cur_ops;
 145
 146 /*
 147  * Definitions for rcu perf testing.
 148  */
 149
 150 static int rcu_perf_read_lock(void) __acquires(RCU)
 151 {
 152         rcu_read_lock();
 153         return 0;
 154 }
 155
 156 static void rcu_perf_read_unlock(int idx) __releases(RCU)
 157 {
 158         rcu_read_unlock();
 159 }
 160
 161 static unsigned long __maybe_unused rcu_no_completed(void)
 162 {
 163         return 0;
 164 }
 165
 166 static void rcu_sync_perf_init(void)
 167 {
 168 }
 169
 170 static struct rcu_perf_ops rcu_ops = {
 171         .ptype          = RCU_FLAVOR,
 172         .init           = rcu_sync_perf_init,
 173         .readlock       = rcu_perf_read_lock,
 174         .readunlock     = rcu_perf_read_unlock,
 175         .get_gp_seq     = rcu_get_gp_seq,
 176         .gp_diff        = rcu_seq_diff,
 177         .exp_completed  = rcu_exp_batches_completed,
 178         .async          = call_rcu,
 179         .gp_barrier     = rcu_barrier,
 180         .sync           = synchronize_rcu,
 181         .exp_sync       = synchronize_rcu_expedited,
 182         .name           = "rcu"
 183 };
 184
 185 /*
 186  * Definitions for srcu perf testing.
 187  */
 188
 189 DEFINE_STATIC_SRCU(srcu_ctl_perf);
 190 static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
 191
 192 static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
 193 {
 194         return srcu_read_lock(srcu_ctlp);
 195 }
 196
 197 static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
 198 {
 199         srcu_read_unlock(srcu_ctlp, idx);
 200 }
 201
 202 static unsigned long srcu_perf_completed(void)
 203 {
 204         return srcu_batches_completed(srcu_ctlp);
 205 }
 206
 207 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
 208 {
 209         call_srcu(srcu_ctlp, head, func);
 210 }
 211
 212 static void srcu_rcu_barrier(void)
 213 {
 214         srcu_barrier(srcu_ctlp);
 215 }
 216
 217 static void srcu_perf_synchronize(void)
 218 {
 219         synchronize_srcu(srcu_ctlp);
 220 }
 221
 222 static void srcu_perf_synchronize_expedited(void)
 223 {
 224         synchronize_srcu_expedited(srcu_ctlp);
 225 }
 226
 227 static struct rcu_perf_ops srcu_ops = {
 228         .ptype          = SRCU_FLAVOR,
 229         .init           = rcu_sync_perf_init,
 230         .readlock       = srcu_perf_read_lock,
 231         .readunlock     = srcu_perf_read_unlock,
 232         .get_gp_seq     = srcu_perf_completed,
 233         .gp_diff        = rcu_seq_diff,
 234         .exp_completed  = srcu_perf_completed,
 235         .async          = srcu_call_rcu,
 236         .gp_barrier     = srcu_rcu_barrier,
 237         .sync           = srcu_perf_synchronize,
 238         .exp_sync       = srcu_perf_synchronize_expedited,
 239         .name           = "srcu"
 240 };
 241
 242 static struct srcu_struct srcud;
 243
 244 static void srcu_sync_perf_init(void)
 245 {
 246         srcu_ctlp = &srcud;
 247         init_srcu_struct(srcu_ctlp);
 248 }
 249
 250 static void srcu_sync_perf_cleanup(void)
 251 {
 252         cleanup_srcu_struct(srcu_ctlp);
 253 }
 254
 255 static struct rcu_perf_ops srcud_ops = {
 256         .ptype          = SRCU_FLAVOR,
 257         .init           = srcu_sync_perf_init,
 258         .cleanup        = srcu_sync_perf_cleanup,
 259         .readlock       = srcu_perf_read_lock,
 260         .readunlock     = srcu_perf_read_unlock,
 261         .get_gp_seq     = srcu_perf_completed,
 262         .gp_diff        = rcu_seq_diff,
 263         .exp_completed  = srcu_perf_completed,
 264         .async          = srcu_call_rcu,
 265         .gp_barrier     = srcu_rcu_barrier,
 266         .sync           = srcu_perf_synchronize,
 267         .exp_sync       = srcu_perf_synchronize_expedited,
 268         .name           = "srcud"
 269 };
 270
 271 /*
 272  * Definitions for RCU-tasks perf testing.
 273  */
 274
 275 static int tasks_perf_read_lock(void)
 276 {
 277         return 0;
 278 }
 279
 280 static void tasks_perf_read_unlock(int idx)
 281 {
 282 }
 283
 284 static struct rcu_perf_ops tasks_ops = {
 285         .ptype          = RCU_TASKS_FLAVOR,
 286         .init           = rcu_sync_perf_init,
 287         .readlock       = tasks_perf_read_lock,
 288         .readunlock     = tasks_perf_read_unlock,
 289         .get_gp_seq     = rcu_no_completed,
 290         .gp_diff        = rcu_seq_diff,
 291         .async          = call_rcu_tasks,
 292         .gp_barrier     = rcu_barrier_tasks,
 293         .sync           = synchronize_rcu_tasks,
 294         .exp_sync       = synchronize_rcu_tasks,
 295         .name           = "tasks"
 296 };
 297
 298 static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old)
 299 {
 300         if (!cur_ops->gp_diff)
 301                 return new - old;
 302         return cur_ops->gp_diff(new, old);
 303 }
 304
 305 /*
 306  * If performance tests complete, wait for shutdown to commence.
 307  */
 308 static void rcu_perf_wait_shutdown(void)
 309 {
 310         cond_resched_tasks_rcu_qs();
 311         if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
 312                 return;
 313         while (!torture_must_stop())
 314                 schedule_timeout_uninterruptible(1);
 315 }
 316
 317 /*
 318  * RCU perf reader kthread.  Repeatedly does empty RCU read-side
 319  * critical section, minimizing update-side interference.
 320  */
 321 static int
 322 rcu_perf_reader(void *arg)
 323 {
 324         unsigned long flags;
 325         int idx;
 326         long me = (long)arg;
 327
 328         VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
 329         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 330         set_user_nice(current, MAX_NICE);
 331         atomic_inc(&n_rcu_perf_reader_started);
 332
 333         do {
 334                 local_irq_save(flags);
 335                 idx = cur_ops->readlock();
 336                 cur_ops->readunlock(idx);
 337                 local_irq_restore(flags);
 338                 rcu_perf_wait_shutdown();
 339         } while (!torture_must_stop());
 340         torture_kthread_stopping("rcu_perf_reader");
 341         return 0;
 342 }
 343
 344 /*
 345  * Callback function for asynchronous grace periods from rcu_perf_writer().
 346  */
 347 static void rcu_perf_async_cb(struct rcu_head *rhp)
 348 {
 349         atomic_dec(this_cpu_ptr(&n_async_inflight));
 350         kfree(rhp);
 351 }
 352
 353 /*
 354  * RCU perf writer kthread.  Repeatedly does a grace period.
 355  */
 356 static int
 357 rcu_perf_writer(void *arg)
 358 {
 359         int i = 0;
 360         int i_max;
 361         long me = (long)arg;
 362         struct rcu_head *rhp = NULL;
 363         struct sched_param sp;
 364         bool started = false, done = false, alldone = false;
 365         u64 t;
 366         u64 *wdp;
 367         u64 *wdpp = writer_durations[me];
 368
 369         VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
 370         WARN_ON(!wdpp);
 371         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 372         sp.sched_priority = 1;
 373         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 374
 375         if (holdoff)
 376                 schedule_timeout_uninterruptible(holdoff * HZ);
 377
 378         t = ktime_get_mono_fast_ns();
 379         if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
 380                 t_rcu_perf_writer_started = t;
 381                 if (gp_exp) {
 382                         b_rcu_perf_writer_started =
 383                                 cur_ops->exp_completed() / 2;
 384                 } else {
 385                         b_rcu_perf_writer_started = cur_ops->get_gp_seq();
 386                 }
 387         }
 388
 389         do {
 390                 if (writer_holdoff)
 391                         udelay(writer_holdoff);
 392                 wdp = &wdpp[i];
 393                 *wdp = ktime_get_mono_fast_ns();
 394                 if (gp_async) {
 395 retry:
 396                         if (!rhp)
 397                                 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
 398                         if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
 399                                 rcu_perf_writer_state = RTWS_ASYNC;
 400                                 atomic_inc(this_cpu_ptr(&n_async_inflight));
 401                                 cur_ops->async(rhp, rcu_perf_async_cb);
 402                                 rhp = NULL;
 403                         } else if (!kthread_should_stop()) {
 404                                 rcu_perf_writer_state = RTWS_BARRIER;
 405                                 cur_ops->gp_barrier();
 406                                 goto retry;
 407                         } else {
 408                                 kfree(rhp); /* Because we are stopping. */
 409                         }
 410                 } else if (gp_exp) {
 411                         rcu_perf_writer_state = RTWS_EXP_SYNC;
 412                         cur_ops->exp_sync();
 413                 } else {
 414                         rcu_perf_writer_state = RTWS_SYNC;
 415                         cur_ops->sync();
 416                 }
 417                 rcu_perf_writer_state = RTWS_IDLE;
 418                 t = ktime_get_mono_fast_ns();
 419                 *wdp = t - *wdp;
 420                 i_max = i;
 421                 if (!started &&
 422                     atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
 423                         started = true;
 424                 if (!done && i >= MIN_MEAS) {
 425                         done = true;
 426                         sp.sched_priority = 0;
 427                         sched_setscheduler_nocheck(current,
 428                                                    SCHED_NORMAL, &sp);
 429                         pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
 430                                  perf_type, PERF_FLAG, me, MIN_MEAS);
 431                         if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
 432                             nrealwriters) {
 433                                 schedule_timeout_interruptible(10);
 434                                 rcu_ftrace_dump(DUMP_ALL);
 435                                 PERFOUT_STRING("Test complete");
 436                                 t_rcu_perf_writer_finished = t;
 437                                 if (gp_exp) {
 438                                         b_rcu_perf_writer_finished =
 439                                                 cur_ops->exp_completed() / 2;
 440                                 } else {
 441                                         b_rcu_perf_writer_finished =
 442                                                 cur_ops->get_gp_seq();
 443                                 }
 444                                 if (shutdown) {
 445                                         smp_mb(); /* Assign before wake. */
 446                                         wake_up(&shutdown_wq);
 447                                 }
 448                         }
 449                 }
 450                 if (done && !alldone &&
 451                     atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
 452                         alldone = true;
 453                 if (started && !alldone && i < MAX_MEAS - 1)
 454                         i++;
 455                 rcu_perf_wait_shutdown();
 456         } while (!torture_must_stop());
 457         if (gp_async) {
 458                 rcu_perf_writer_state = RTWS_BARRIER;
 459                 cur_ops->gp_barrier();
 460         }
 461         rcu_perf_writer_state = RTWS_STOPPING;
 462         writer_n_durations[me] = i_max;
 463         torture_kthread_stopping("rcu_perf_writer");
 464         return 0;
 465 }
 466
 467 static void
 468 rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
 469 {
 470         pr_alert("%s" PERF_FLAG
 471                  "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
 472                  perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
 473 }
 474
 475 static void
 476 rcu_perf_cleanup(void)
 477 {
 478         int i;
 479         int j;
 480         int ngps = 0;
 481         u64 *wdp;
 482         u64 *wdpp;
 483
 484         /*
 485          * Would like warning at start, but everything is expedited
 486          * during the mid-boot phase, so have to wait till the end.
 487          */
 488         if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
 489                 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
 490         if (rcu_gp_is_normal() && gp_exp)
 491                 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
 492         if (gp_exp && gp_async)
 493                 VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
 494
 495         if (torture_cleanup_begin())
 496                 return;
 497
 498         if (reader_tasks) {
 499                 for (i = 0; i < nrealreaders; i++)
 500                         torture_stop_kthread(rcu_perf_reader,
 501                                              reader_tasks[i]);
 502                 kfree(reader_tasks);
 503         }
 504
 505         if (writer_tasks) {
 506                 for (i = 0; i < nrealwriters; i++) {
 507                         torture_stop_kthread(rcu_perf_writer,
 508                                              writer_tasks[i]);
 509                         if (!writer_n_durations)
 510                                 continue;
 511                         j = writer_n_durations[i];
 512                         pr_alert("%s%s writer %d gps: %d\n",
 513                                  perf_type, PERF_FLAG, i, j);
 514                         ngps += j;
 515                 }
 516                 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
 517                          perf_type, PERF_FLAG,
 518                          t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
 519                          t_rcu_perf_writer_finished -
 520                          t_rcu_perf_writer_started,
 521                          ngps,
 522                          rcuperf_seq_diff(b_rcu_perf_writer_finished,
 523                                           b_rcu_perf_writer_started));
 524                 for (i = 0; i < nrealwriters; i++) {
 525                         if (!writer_durations)
 526                                 break;
 527                         if (!writer_n_durations)
 528                                 continue;
 529                         wdpp = writer_durations[i];
 530                         if (!wdpp)
 531                                 continue;
 532                         for (j = 0; j <= writer_n_durations[i]; j++) {
 533                                 wdp = &wdpp[j];
 534                                 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
 535                                         perf_type, PERF_FLAG,
 536                                         i, j, *wdp);
 537                                 if (j % 100 == 0)
 538                                         schedule_timeout_uninterruptible(1);
 539                         }
 540                         kfree(writer_durations[i]);
 541                 }
 542                 kfree(writer_tasks);
 543                 kfree(writer_durations);
 544                 kfree(writer_n_durations);
 545         }
 546
 547         /* Do torture-type-specific cleanup operations.  */
 548         if (cur_ops->cleanup != NULL)
 549                 cur_ops->cleanup();
 550
 551         torture_cleanup_end();
 552 }
 553
 554 /*
 555  * Return the number if non-negative.  If -1, the number of CPUs.
 556  * If less than -1, that much less than the number of CPUs, but
 557  * at least one.
 558  */
 559 static int compute_real(int n)
 560 {
 561         int nr;
 562
 563         if (n >= 0) {
 564                 nr = n;
 565         } else {
 566                 nr = num_online_cpus() + 1 + n;
 567                 if (nr <= 0)
 568                         nr = 1;
 569         }
 570         return nr;
 571 }
 572
 573 /*
 574  * RCU perf shutdown kthread.  Just waits to be awakened, then shuts
 575  * down system.
 576  */
 577 static int
 578 rcu_perf_shutdown(void *arg)
 579 {
 580         do {
 581                 wait_event(shutdown_wq,
 582                            atomic_read(&n_rcu_perf_writer_finished) >=
 583                            nrealwriters);
 584         } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters);
 585         smp_mb(); /* Wake before output. */
 586         rcu_perf_cleanup();
 587         kernel_power_off();
 588         return -EINVAL;
 589 }
 590
 591 static int __init
 592 rcu_perf_init(void)
 593 {
 594         long i;
 595         int firsterr = 0;
 596         static struct rcu_perf_ops *perf_ops[] = {
 597                 &rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops,
 598         };
 599
 600         if (!torture_init_begin(perf_type, verbose))
 601                 return -EBUSY;
 602
 603         /* Process args and tell the world that the perf'er is on the job. */
 604         for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
 605                 cur_ops = perf_ops[i];
 606                 if (strcmp(perf_type, cur_ops->name) == 0)
 607                         break;
 608         }
 609         if (i == ARRAY_SIZE(perf_ops)) {
 610                 pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type);
 611                 pr_alert("rcu-perf types:");
 612                 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
 613                         pr_cont(" %s", perf_ops[i]->name);
 614                 pr_cont("\n");
 615                 WARN_ON(!IS_MODULE(CONFIG_RCU_PERF_TEST));
 616                 firsterr = -EINVAL;
 617                 goto unwind;
 618         }
 619         if (cur_ops->init)
 620                 cur_ops->init();
 621
 622         nrealwriters = compute_real(nwriters);
 623         nrealreaders = compute_real(nreaders);
 624         atomic_set(&n_rcu_perf_reader_started, 0);
 625         atomic_set(&n_rcu_perf_writer_started, 0);
 626         atomic_set(&n_rcu_perf_writer_finished, 0);
 627         rcu_perf_print_module_parms(cur_ops, "Start of test");
 628
 629         /* Start up the kthreads. */
 630
 631         if (shutdown) {
 632                 init_waitqueue_head(&shutdown_wq);
 633                 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
 634                                                   shutdown_task);
 635                 if (firsterr)
 636                         goto unwind;
 637                 schedule_timeout_uninterruptible(1);
 638         }
 639         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
 640                                GFP_KERNEL);
 641         if (reader_tasks == NULL) {
 642                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 643                 firsterr = -ENOMEM;
 644                 goto unwind;
 645         }
 646         for (i = 0; i < nrealreaders; i++) {
 647                 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
 648                                                   reader_tasks[i]);
 649                 if (firsterr)
 650                         goto unwind;
 651         }
 652         while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
 653                 schedule_timeout_uninterruptible(1);
 654         writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
 655                                GFP_KERNEL);
 656         writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
 657                                    GFP_KERNEL);
 658         writer_n_durations =
 659                 kcalloc(nrealwriters, sizeof(*writer_n_durations),
 660                         GFP_KERNEL);
 661         if (!writer_tasks || !writer_durations || !writer_n_durations) {
 662                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 663                 firsterr = -ENOMEM;
 664                 goto unwind;
 665         }
 666         for (i = 0; i < nrealwriters; i++) {
 667                 writer_durations[i] =
 668                         kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
 669                                 GFP_KERNEL);
 670                 if (!writer_durations[i]) {
 671                         firsterr = -ENOMEM;
 672                         goto unwind;
 673                 }
 674                 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
 675                                                   writer_tasks[i]);
 676                 if (firsterr)
 677                         goto unwind;
 678         }
 679         torture_init_end();
 680         return 0;
 681
 682 unwind:
 683         torture_init_end();
 684         rcu_perf_cleanup();
 685         return firsterr;
 686 }
 687
 688 module_init(rcu_perf_init);
 689 module_exit(rcu_perf_cleanup);