kernel/rcu/rcuperf.c

   1 /*
   2  * Read-Copy Update module-based performance-test facility
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, you can access it online at
  16  * http://www.gnu.org/licenses/gpl-2.0.html.
  17  *
  18  * Copyright (C) IBM Corporation, 2015
  19  *
  20  * Authors: Paul E. McKenney <paulmck@us.ibm.com>
  21  */
  22 #include <linux/types.h>
  23 #include <linux/kernel.h>
  24 #include <linux/init.h>
  25 #include <linux/module.h>
  26 #include <linux/kthread.h>
  27 #include <linux/err.h>
  28 #include <linux/spinlock.h>
  29 #include <linux/smp.h>
  30 #include <linux/rcupdate.h>
  31 #include <linux/interrupt.h>
  32 #include <linux/sched.h>
  33 #include <uapi/linux/sched/types.h>
  34 #include <linux/atomic.h>
  35 #include <linux/bitops.h>
  36 #include <linux/completion.h>
  37 #include <linux/moduleparam.h>
  38 #include <linux/percpu.h>
  39 #include <linux/notifier.h>
  40 #include <linux/reboot.h>
  41 #include <linux/freezer.h>
  42 #include <linux/cpu.h>
  43 #include <linux/delay.h>
  44 #include <linux/stat.h>
  45 #include <linux/srcu.h>
  46 #include <linux/slab.h>
  47 #include <asm/byteorder.h>
  48 #include <linux/torture.h>
  49 #include <linux/vmalloc.h>
  50
  51 #include "rcu.h"
  52
  53 MODULE_LICENSE("GPL");
  54 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.vnet.ibm.com>");
  55
  56 #define PERF_FLAG "-perf:"
  57 #define PERFOUT_STRING(s) \
  58         pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
  59 #define VERBOSE_PERFOUT_STRING(s) \
  60         do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
  61 #define VERBOSE_PERFOUT_ERRSTRING(s) \
  62         do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
  63
  64 /*
  65  * The intended use cases for the nreaders and nwriters module parameters
  66  * are as follows:
  67  *
  68  * 1.   Specify only the nr_cpus kernel boot parameter.  This will
  69  *      set both nreaders and nwriters to the value specified by
  70  *      nr_cpus for a mixed reader/writer test.
  71  *
  72  * 2.   Specify the nr_cpus kernel boot parameter, but set
  73  *      rcuperf.nreaders to zero.  This will set nwriters to the
  74  *      value specified by nr_cpus for an update-only test.
  75  *
  76  * 3.   Specify the nr_cpus kernel boot parameter, but set
  77  *      rcuperf.nwriters to zero.  This will set nreaders to the
  78  *      value specified by nr_cpus for a read-only test.
  79  *
  80  * Various other use cases may of course be specified.
  81  */
  82
  83 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
  84 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
  85 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
  86 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
  87 torture_param(int, nreaders, -1, "Number of RCU reader threads");
  88 torture_param(int, nwriters, -1, "Number of RCU updater threads");
  89 torture_param(bool, shutdown, !IS_ENABLED(MODULE),
  90               "Shutdown at end of performance tests.");
  91 torture_param(bool, verbose, true, "Enable verbose debugging printk()s");
  92 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
  93
  94 static char *perf_type = "rcu";
  95 module_param(perf_type, charp, 0444);
  96 MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, rcu_bh, ...)");
  97
  98 static int nrealreaders;
  99 static int nrealwriters;
 100 static struct task_struct **writer_tasks;
 101 static struct task_struct **reader_tasks;
 102 static struct task_struct *shutdown_task;
 103
 104 static u64 **writer_durations;
 105 static int *writer_n_durations;
 106 static atomic_t n_rcu_perf_reader_started;
 107 static atomic_t n_rcu_perf_writer_started;
 108 static atomic_t n_rcu_perf_writer_finished;
 109 static wait_queue_head_t shutdown_wq;
 110 static u64 t_rcu_perf_writer_started;
 111 static u64 t_rcu_perf_writer_finished;
 112 static unsigned long b_rcu_perf_writer_started;
 113 static unsigned long b_rcu_perf_writer_finished;
 114 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
 115
 116 static int rcu_perf_writer_state;
 117 #define RTWS_INIT               0
 118 #define RTWS_ASYNC              1
 119 #define RTWS_BARRIER            2
 120 #define RTWS_EXP_SYNC           3
 121 #define RTWS_SYNC               4
 122 #define RTWS_IDLE               5
 123 #define RTWS_STOPPING           6
 124
 125 #define MAX_MEAS 10000
 126 #define MIN_MEAS 100
 127
 128 /*
 129  * Operations vector for selecting different types of tests.
 130  */
 131
 132 struct rcu_perf_ops {
 133         int ptype;
 134         void (*init)(void);
 135         void (*cleanup)(void);
 136         int (*readlock)(void);
 137         void (*readunlock)(int idx);
 138         unsigned long (*started)(void);
 139         unsigned long (*completed)(void);
 140         unsigned long (*exp_completed)(void);
 141         void (*async)(struct rcu_head *head, rcu_callback_t func);
 142         void (*gp_barrier)(void);
 143         void (*sync)(void);
 144         void (*exp_sync)(void);
 145         const char *name;
 146 };
 147
 148 static struct rcu_perf_ops *cur_ops;
 149
 150 /*
 151  * Definitions for rcu perf testing.
 152  */
 153
 154 static int rcu_perf_read_lock(void) __acquires(RCU)
 155 {
 156         rcu_read_lock();
 157         return 0;
 158 }
 159
 160 static void rcu_perf_read_unlock(int idx) __releases(RCU)
 161 {
 162         rcu_read_unlock();
 163 }
 164
 165 static unsigned long __maybe_unused rcu_no_completed(void)
 166 {
 167         return 0;
 168 }
 169
 170 static void rcu_sync_perf_init(void)
 171 {
 172 }
 173
 174 static struct rcu_perf_ops rcu_ops = {
 175         .ptype          = RCU_FLAVOR,
 176         .init           = rcu_sync_perf_init,
 177         .readlock       = rcu_perf_read_lock,
 178         .readunlock     = rcu_perf_read_unlock,
 179         .started        = rcu_batches_started,
 180         .completed      = rcu_batches_completed,
 181         .exp_completed  = rcu_exp_batches_completed,
 182         .async          = call_rcu,
 183         .gp_barrier     = rcu_barrier,
 184         .sync           = synchronize_rcu,
 185         .exp_sync       = synchronize_rcu_expedited,
 186         .name           = "rcu"
 187 };
 188
 189 /*
 190  * Definitions for rcu_bh perf testing.
 191  */
 192
 193 static int rcu_bh_perf_read_lock(void) __acquires(RCU_BH)
 194 {
 195         rcu_read_lock_bh();
 196         return 0;
 197 }
 198
 199 static void rcu_bh_perf_read_unlock(int idx) __releases(RCU_BH)
 200 {
 201         rcu_read_unlock_bh();
 202 }
 203
 204 static struct rcu_perf_ops rcu_bh_ops = {
 205         .ptype          = RCU_BH_FLAVOR,
 206         .init           = rcu_sync_perf_init,
 207         .readlock       = rcu_bh_perf_read_lock,
 208         .readunlock     = rcu_bh_perf_read_unlock,
 209         .started        = rcu_batches_started_bh,
 210         .completed      = rcu_batches_completed_bh,
 211         .exp_completed  = rcu_exp_batches_completed_sched,
 212         .async          = call_rcu_bh,
 213         .gp_barrier     = rcu_barrier_bh,
 214         .sync           = synchronize_rcu_bh,
 215         .exp_sync       = synchronize_rcu_bh_expedited,
 216         .name           = "rcu_bh"
 217 };
 218
 219 /*
 220  * Definitions for srcu perf testing.
 221  */
 222
 223 DEFINE_STATIC_SRCU(srcu_ctl_perf);
 224 static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
 225
 226 static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
 227 {
 228         return srcu_read_lock(srcu_ctlp);
 229 }
 230
 231 static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
 232 {
 233         srcu_read_unlock(srcu_ctlp, idx);
 234 }
 235
 236 static unsigned long srcu_perf_completed(void)
 237 {
 238         return srcu_batches_completed(srcu_ctlp);
 239 }
 240
 241 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
 242 {
 243         call_srcu(srcu_ctlp, head, func);
 244 }
 245
 246 static void srcu_rcu_barrier(void)
 247 {
 248         srcu_barrier(srcu_ctlp);
 249 }
 250
 251 static void srcu_perf_synchronize(void)
 252 {
 253         synchronize_srcu(srcu_ctlp);
 254 }
 255
 256 static void srcu_perf_synchronize_expedited(void)
 257 {
 258         synchronize_srcu_expedited(srcu_ctlp);
 259 }
 260
 261 static struct rcu_perf_ops srcu_ops = {
 262         .ptype          = SRCU_FLAVOR,
 263         .init           = rcu_sync_perf_init,
 264         .readlock       = srcu_perf_read_lock,
 265         .readunlock     = srcu_perf_read_unlock,
 266         .started        = NULL,
 267         .completed      = srcu_perf_completed,
 268         .exp_completed  = srcu_perf_completed,
 269         .async          = srcu_call_rcu,
 270         .gp_barrier     = srcu_rcu_barrier,
 271         .sync           = srcu_perf_synchronize,
 272         .exp_sync       = srcu_perf_synchronize_expedited,
 273         .name           = "srcu"
 274 };
 275
 276 static struct srcu_struct srcud;
 277
 278 static void srcu_sync_perf_init(void)
 279 {
 280         srcu_ctlp = &srcud;
 281         init_srcu_struct(srcu_ctlp);
 282 }
 283
 284 static void srcu_sync_perf_cleanup(void)
 285 {
 286         cleanup_srcu_struct(srcu_ctlp);
 287 }
 288
 289 static struct rcu_perf_ops srcud_ops = {
 290         .ptype          = SRCU_FLAVOR,
 291         .init           = srcu_sync_perf_init,
 292         .cleanup        = srcu_sync_perf_cleanup,
 293         .readlock       = srcu_perf_read_lock,
 294         .readunlock     = srcu_perf_read_unlock,
 295         .started        = NULL,
 296         .completed      = srcu_perf_completed,
 297         .exp_completed  = srcu_perf_completed,
 298         .async          = srcu_call_rcu,
 299         .gp_barrier     = srcu_rcu_barrier,
 300         .sync           = srcu_perf_synchronize,
 301         .exp_sync       = srcu_perf_synchronize_expedited,
 302         .name           = "srcud"
 303 };
 304
 305 /*
 306  * Definitions for sched perf testing.
 307  */
 308
 309 static int sched_perf_read_lock(void)
 310 {
 311         preempt_disable();
 312         return 0;
 313 }
 314
 315 static void sched_perf_read_unlock(int idx)
 316 {
 317         preempt_enable();
 318 }
 319
 320 static struct rcu_perf_ops sched_ops = {
 321         .ptype          = RCU_SCHED_FLAVOR,
 322         .init           = rcu_sync_perf_init,
 323         .readlock       = sched_perf_read_lock,
 324         .readunlock     = sched_perf_read_unlock,
 325         .started        = rcu_batches_started_sched,
 326         .completed      = rcu_batches_completed_sched,
 327         .exp_completed  = rcu_exp_batches_completed_sched,
 328         .async          = call_rcu_sched,
 329         .gp_barrier     = rcu_barrier_sched,
 330         .sync           = synchronize_sched,
 331         .exp_sync       = synchronize_sched_expedited,
 332         .name           = "sched"
 333 };
 334
 335 /*
 336  * Definitions for RCU-tasks perf testing.
 337  */
 338
 339 static int tasks_perf_read_lock(void)
 340 {
 341         return 0;
 342 }
 343
 344 static void tasks_perf_read_unlock(int idx)
 345 {
 346 }
 347
 348 static struct rcu_perf_ops tasks_ops = {
 349         .ptype          = RCU_TASKS_FLAVOR,
 350         .init           = rcu_sync_perf_init,
 351         .readlock       = tasks_perf_read_lock,
 352         .readunlock     = tasks_perf_read_unlock,
 353         .started        = rcu_no_completed,
 354         .completed      = rcu_no_completed,
 355         .async          = call_rcu_tasks,
 356         .gp_barrier     = rcu_barrier_tasks,
 357         .sync           = synchronize_rcu_tasks,
 358         .exp_sync       = synchronize_rcu_tasks,
 359         .name           = "tasks"
 360 };
 361
 362 static bool __maybe_unused torturing_tasks(void)
 363 {
 364         return cur_ops == &tasks_ops;
 365 }
 366
 367 /*
 368  * If performance tests complete, wait for shutdown to commence.
 369  */
 370 static void rcu_perf_wait_shutdown(void)
 371 {
 372         cond_resched_rcu_qs();
 373         if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
 374                 return;
 375         while (!torture_must_stop())
 376                 schedule_timeout_uninterruptible(1);
 377 }
 378
 379 /*
 380  * RCU perf reader kthread.  Repeatedly does empty RCU read-side
 381  * critical section, minimizing update-side interference.
 382  */
 383 static int
 384 rcu_perf_reader(void *arg)
 385 {
 386         unsigned long flags;
 387         int idx;
 388         long me = (long)arg;
 389
 390         VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
 391         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 392         set_user_nice(current, MAX_NICE);
 393         atomic_inc(&n_rcu_perf_reader_started);
 394
 395         do {
 396                 local_irq_save(flags);
 397                 idx = cur_ops->readlock();
 398                 cur_ops->readunlock(idx);
 399                 local_irq_restore(flags);
 400                 rcu_perf_wait_shutdown();
 401         } while (!torture_must_stop());
 402         torture_kthread_stopping("rcu_perf_reader");
 403         return 0;
 404 }
 405
 406 /*
 407  * Callback function for asynchronous grace periods from rcu_perf_writer().
 408  */
 409 static void rcu_perf_async_cb(struct rcu_head *rhp)
 410 {
 411         atomic_dec(this_cpu_ptr(&n_async_inflight));
 412         kfree(rhp);
 413 }
 414
 415 /*
 416  * RCU perf writer kthread.  Repeatedly does a grace period.
 417  */
 418 static int
 419 rcu_perf_writer(void *arg)
 420 {
 421         int i = 0;
 422         int i_max;
 423         long me = (long)arg;
 424         struct rcu_head *rhp = NULL;
 425         struct sched_param sp;
 426         bool started = false, done = false, alldone = false;
 427         u64 t;
 428         u64 *wdp;
 429         u64 *wdpp = writer_durations[me];
 430
 431         VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
 432         WARN_ON(!wdpp);
 433         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 434         sp.sched_priority = 1;
 435         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 436
 437         if (holdoff)
 438                 schedule_timeout_uninterruptible(holdoff * HZ);
 439
 440         t = ktime_get_mono_fast_ns();
 441         if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
 442                 t_rcu_perf_writer_started = t;
 443                 if (gp_exp) {
 444                         b_rcu_perf_writer_started =
 445                                 cur_ops->exp_completed() / 2;
 446                 } else {
 447                         b_rcu_perf_writer_started =
 448                                 cur_ops->completed();
 449                 }
 450         }
 451
 452         do {
 453                 if (writer_holdoff)
 454                         udelay(writer_holdoff);
 455                 wdp = &wdpp[i];
 456                 *wdp = ktime_get_mono_fast_ns();
 457                 if (gp_async) {
 458 retry:
 459                         if (!rhp)
 460                                 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
 461                         if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
 462                                 rcu_perf_writer_state = RTWS_ASYNC;
 463                                 atomic_inc(this_cpu_ptr(&n_async_inflight));
 464                                 cur_ops->async(rhp, rcu_perf_async_cb);
 465                                 rhp = NULL;
 466                         } else if (!kthread_should_stop()) {
 467                                 rcu_perf_writer_state = RTWS_BARRIER;
 468                                 cur_ops->gp_barrier();
 469                                 goto retry;
 470                         } else {
 471                                 kfree(rhp); /* Because we are stopping. */
 472                         }
 473                 } else if (gp_exp) {
 474                         rcu_perf_writer_state = RTWS_EXP_SYNC;
 475                         cur_ops->exp_sync();
 476                 } else {
 477                         rcu_perf_writer_state = RTWS_SYNC;
 478                         cur_ops->sync();
 479                 }
 480                 rcu_perf_writer_state = RTWS_IDLE;
 481                 t = ktime_get_mono_fast_ns();
 482                 *wdp = t - *wdp;
 483                 i_max = i;
 484                 if (!started &&
 485                     atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
 486                         started = true;
 487                 if (!done && i >= MIN_MEAS) {
 488                         done = true;
 489                         sp.sched_priority = 0;
 490                         sched_setscheduler_nocheck(current,
 491                                                    SCHED_NORMAL, &sp);
 492                         pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
 493                                  perf_type, PERF_FLAG, me, MIN_MEAS);
 494                         if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
 495                             nrealwriters) {
 496                                 schedule_timeout_interruptible(10);
 497                                 rcu_ftrace_dump(DUMP_ALL);
 498                                 PERFOUT_STRING("Test complete");
 499                                 t_rcu_perf_writer_finished = t;
 500                                 if (gp_exp) {
 501                                         b_rcu_perf_writer_finished =
 502                                                 cur_ops->exp_completed() / 2;
 503                                 } else {
 504                                         b_rcu_perf_writer_finished =
 505                                                 cur_ops->completed();
 506                                 }
 507                                 if (shutdown) {
 508                                         smp_mb(); /* Assign before wake. */
 509                                         wake_up(&shutdown_wq);
 510                                 }
 511                         }
 512                 }
 513                 if (done && !alldone &&
 514                     atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
 515                         alldone = true;
 516                 if (started && !alldone && i < MAX_MEAS - 1)
 517                         i++;
 518                 rcu_perf_wait_shutdown();
 519         } while (!torture_must_stop());
 520         if (gp_async) {
 521                 rcu_perf_writer_state = RTWS_BARRIER;
 522                 cur_ops->gp_barrier();
 523         }
 524         rcu_perf_writer_state = RTWS_STOPPING;
 525         writer_n_durations[me] = i_max;
 526         torture_kthread_stopping("rcu_perf_writer");
 527         return 0;
 528 }
 529
 530 static inline void
 531 rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
 532 {
 533         pr_alert("%s" PERF_FLAG
 534                  "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
 535                  perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
 536 }
 537
 538 static void
 539 rcu_perf_cleanup(void)
 540 {
 541         int i;
 542         int j;
 543         int ngps = 0;
 544         u64 *wdp;
 545         u64 *wdpp;
 546
 547         /*
 548          * Would like warning at start, but everything is expedited
 549          * during the mid-boot phase, so have to wait till the end.
 550          */
 551         if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
 552                 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
 553         if (rcu_gp_is_normal() && gp_exp)
 554                 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
 555         if (gp_exp && gp_async)
 556                 VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
 557
 558         if (torture_cleanup_begin())
 559                 return;
 560
 561         if (reader_tasks) {
 562                 for (i = 0; i < nrealreaders; i++)
 563                         torture_stop_kthread(rcu_perf_reader,
 564                                              reader_tasks[i]);
 565                 kfree(reader_tasks);
 566         }
 567
 568         if (writer_tasks) {
 569                 for (i = 0; i < nrealwriters; i++) {
 570                         torture_stop_kthread(rcu_perf_writer,
 571                                              writer_tasks[i]);
 572                         if (!writer_n_durations)
 573                                 continue;
 574                         j = writer_n_durations[i];
 575                         pr_alert("%s%s writer %d gps: %d\n",
 576                                  perf_type, PERF_FLAG, i, j);
 577                         ngps += j;
 578                 }
 579                 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
 580                          perf_type, PERF_FLAG,
 581                          t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
 582                          t_rcu_perf_writer_finished -
 583                          t_rcu_perf_writer_started,
 584                          ngps,
 585                          b_rcu_perf_writer_finished -
 586                          b_rcu_perf_writer_started);
 587                 for (i = 0; i < nrealwriters; i++) {
 588                         if (!writer_durations)
 589                                 break;
 590                         if (!writer_n_durations)
 591                                 continue;
 592                         wdpp = writer_durations[i];
 593                         if (!wdpp)
 594                                 continue;
 595                         for (j = 0; j <= writer_n_durations[i]; j++) {
 596                                 wdp = &wdpp[j];
 597                                 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
 598                                         perf_type, PERF_FLAG,
 599                                         i, j, *wdp);
 600                                 if (j % 100 == 0)
 601                                         schedule_timeout_uninterruptible(1);
 602                         }
 603                         kfree(writer_durations[i]);
 604                 }
 605                 kfree(writer_tasks);
 606                 kfree(writer_durations);
 607                 kfree(writer_n_durations);
 608         }
 609
 610         /* Do flavor-specific cleanup operations.  */
 611         if (cur_ops->cleanup != NULL)
 612                 cur_ops->cleanup();
 613
 614         torture_cleanup_end();
 615 }
 616
 617 /*
 618  * Return the number if non-negative.  If -1, the number of CPUs.
 619  * If less than -1, that much less than the number of CPUs, but
 620  * at least one.
 621  */
 622 static int compute_real(int n)
 623 {
 624         int nr;
 625
 626         if (n >= 0) {
 627                 nr = n;
 628         } else {
 629                 nr = num_online_cpus() + 1 + n;
 630                 if (nr <= 0)
 631                         nr = 1;
 632         }
 633         return nr;
 634 }
 635
 636 /*
 637  * RCU perf shutdown kthread.  Just waits to be awakened, then shuts
 638  * down system.
 639  */
 640 static int
 641 rcu_perf_shutdown(void *arg)
 642 {
 643         do {
 644                 wait_event(shutdown_wq,
 645                            atomic_read(&n_rcu_perf_writer_finished) >=
 646                            nrealwriters);
 647         } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters);
 648         smp_mb(); /* Wake before output. */
 649         rcu_perf_cleanup();
 650         kernel_power_off();
 651         return -EINVAL;
 652 }
 653
 654 static int __init
 655 rcu_perf_init(void)
 656 {
 657         long i;
 658         int firsterr = 0;
 659         static struct rcu_perf_ops *perf_ops[] = {
 660                 &rcu_ops, &rcu_bh_ops, &srcu_ops, &srcud_ops, &sched_ops,
 661                 &tasks_ops,
 662         };
 663
 664         if (!torture_init_begin(perf_type, verbose))
 665                 return -EBUSY;
 666
 667         /* Process args and tell the world that the perf'er is on the job. */
 668         for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
 669                 cur_ops = perf_ops[i];
 670                 if (strcmp(perf_type, cur_ops->name) == 0)
 671                         break;
 672         }
 673         if (i == ARRAY_SIZE(perf_ops)) {
 674                 pr_alert("rcu-perf: invalid perf type: \"%s\"\n",
 675                          perf_type);
 676                 pr_alert("rcu-perf types:");
 677                 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
 678                         pr_alert(" %s", perf_ops[i]->name);
 679                 pr_alert("\n");
 680                 firsterr = -EINVAL;
 681                 goto unwind;
 682         }
 683         if (cur_ops->init)
 684                 cur_ops->init();
 685
 686         nrealwriters = compute_real(nwriters);
 687         nrealreaders = compute_real(nreaders);
 688         atomic_set(&n_rcu_perf_reader_started, 0);
 689         atomic_set(&n_rcu_perf_writer_started, 0);
 690         atomic_set(&n_rcu_perf_writer_finished, 0);
 691         rcu_perf_print_module_parms(cur_ops, "Start of test");
 692
 693         /* Start up the kthreads. */
 694
 695         if (shutdown) {
 696                 init_waitqueue_head(&shutdown_wq);
 697                 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
 698                                                   shutdown_task);
 699                 if (firsterr)
 700                         goto unwind;
 701                 schedule_timeout_uninterruptible(1);
 702         }
 703         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
 704                                GFP_KERNEL);
 705         if (reader_tasks == NULL) {
 706                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 707                 firsterr = -ENOMEM;
 708                 goto unwind;
 709         }
 710         for (i = 0; i < nrealreaders; i++) {
 711                 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
 712                                                   reader_tasks[i]);
 713                 if (firsterr)
 714                         goto unwind;
 715         }
 716         while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
 717                 schedule_timeout_uninterruptible(1);
 718         writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
 719                                GFP_KERNEL);
 720         writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
 721                                    GFP_KERNEL);
 722         writer_n_durations =
 723                 kcalloc(nrealwriters, sizeof(*writer_n_durations),
 724                         GFP_KERNEL);
 725         if (!writer_tasks || !writer_durations || !writer_n_durations) {
 726                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 727                 firsterr = -ENOMEM;
 728                 goto unwind;
 729         }
 730         for (i = 0; i < nrealwriters; i++) {
 731                 writer_durations[i] =
 732                         kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
 733                                 GFP_KERNEL);
 734                 if (!writer_durations[i]) {
 735                         firsterr = -ENOMEM;
 736                         goto unwind;
 737                 }
 738                 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
 739                                                   writer_tasks[i]);
 740                 if (firsterr)
 741                         goto unwind;
 742         }
 743         torture_init_end();
 744         return 0;
 745
 746 unwind:
 747         torture_init_end();
 748         rcu_perf_cleanup();
 749         return firsterr;
 750 }
 751
 752 module_init(rcu_perf_init);
 753 module_exit(rcu_perf_cleanup);