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
  23 #define pr_fmt(fmt) fmt
  24
  25 #include <linux/types.h>
  26 #include <linux/kernel.h>
  27 #include <linux/init.h>
  28 #include <linux/module.h>
  29 #include <linux/kthread.h>
  30 #include <linux/err.h>
  31 #include <linux/spinlock.h>
  32 #include <linux/smp.h>
  33 #include <linux/rcupdate.h>
  34 #include <linux/interrupt.h>
  35 #include <linux/sched.h>
  36 #include <uapi/linux/sched/types.h>
  37 #include <linux/atomic.h>
  38 #include <linux/bitops.h>
  39 #include <linux/completion.h>
  40 #include <linux/moduleparam.h>
  41 #include <linux/percpu.h>
  42 #include <linux/notifier.h>
  43 #include <linux/reboot.h>
  44 #include <linux/freezer.h>
  45 #include <linux/cpu.h>
  46 #include <linux/delay.h>
  47 #include <linux/stat.h>
  48 #include <linux/srcu.h>
  49 #include <linux/slab.h>
  50 #include <asm/byteorder.h>
  51 #include <linux/torture.h>
  52 #include <linux/vmalloc.h>
  53
  54 #include "rcu.h"
  55
  56 MODULE_LICENSE("GPL");
  57 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.vnet.ibm.com>");
  58
  59 #define PERF_FLAG "-perf:"
  60 #define PERFOUT_STRING(s) \
  61         pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
  62 #define VERBOSE_PERFOUT_STRING(s) \
  63         do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
  64 #define VERBOSE_PERFOUT_ERRSTRING(s) \
  65         do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
  66
  67 /*
  68  * The intended use cases for the nreaders and nwriters module parameters
  69  * are as follows:
  70  *
  71  * 1.   Specify only the nr_cpus kernel boot parameter.  This will
  72  *      set both nreaders and nwriters to the value specified by
  73  *      nr_cpus for a mixed reader/writer test.
  74  *
  75  * 2.   Specify the nr_cpus kernel boot parameter, but set
  76  *      rcuperf.nreaders to zero.  This will set nwriters to the
  77  *      value specified by nr_cpus for an update-only test.
  78  *
  79  * 3.   Specify the nr_cpus kernel boot parameter, but set
  80  *      rcuperf.nwriters to zero.  This will set nreaders to the
  81  *      value specified by nr_cpus for a read-only test.
  82  *
  83  * Various other use cases may of course be specified.
  84  */
  85
  86 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
  87 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
  88 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
  89 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
  90 torture_param(int, nreaders, -1, "Number of RCU reader threads");
  91 torture_param(int, nwriters, -1, "Number of RCU updater threads");
  92 torture_param(bool, shutdown, !IS_ENABLED(MODULE),
  93               "Shutdown at end of performance tests.");
  94 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
  95 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
  96
  97 static char *perf_type = "rcu";
  98 module_param(perf_type, charp, 0444);
  99 MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, rcu_bh, ...)");
 100
 101 static int nrealreaders;
 102 static int nrealwriters;
 103 static struct task_struct **writer_tasks;
 104 static struct task_struct **reader_tasks;
 105 static struct task_struct *shutdown_task;
 106
 107 static u64 **writer_durations;
 108 static int *writer_n_durations;
 109 static atomic_t n_rcu_perf_reader_started;
 110 static atomic_t n_rcu_perf_writer_started;
 111 static atomic_t n_rcu_perf_writer_finished;
 112 static wait_queue_head_t shutdown_wq;
 113 static u64 t_rcu_perf_writer_started;
 114 static u64 t_rcu_perf_writer_finished;
 115 static unsigned long b_rcu_perf_writer_started;
 116 static unsigned long b_rcu_perf_writer_finished;
 117 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
 118
 119 static int rcu_perf_writer_state;
 120 #define RTWS_INIT               0
 121 #define RTWS_ASYNC              1
 122 #define RTWS_BARRIER            2
 123 #define RTWS_EXP_SYNC           3
 124 #define RTWS_SYNC               4
 125 #define RTWS_IDLE               5
 126 #define RTWS_STOPPING           6
 127
 128 #define MAX_MEAS 10000
 129 #define MIN_MEAS 100
 130
 131 /*
 132  * Operations vector for selecting different types of tests.
 133  */
 134
 135 struct rcu_perf_ops {
 136         int ptype;
 137         void (*init)(void);
 138         void (*cleanup)(void);
 139         int (*readlock)(void);
 140         void (*readunlock)(int idx);
 141         unsigned long (*get_gp_seq)(void);
 142         unsigned long (*gp_diff)(unsigned long new, unsigned long old);
 143         unsigned long (*exp_completed)(void);
 144         void (*async)(struct rcu_head *head, rcu_callback_t func);
 145         void (*gp_barrier)(void);
 146         void (*sync)(void);
 147         void (*exp_sync)(void);
 148         const char *name;
 149 };
 150
 151 static struct rcu_perf_ops *cur_ops;
 152
 153 /*
 154  * Definitions for rcu perf testing.
 155  */
 156
 157 static int rcu_perf_read_lock(void) __acquires(RCU)
 158 {
 159         rcu_read_lock();
 160         return 0;
 161 }
 162
 163 static void rcu_perf_read_unlock(int idx) __releases(RCU)
 164 {
 165         rcu_read_unlock();
 166 }
 167
 168 static unsigned long __maybe_unused rcu_no_completed(void)
 169 {
 170         return 0;
 171 }
 172
 173 static void rcu_sync_perf_init(void)
 174 {
 175 }
 176
 177 static struct rcu_perf_ops rcu_ops = {
 178         .ptype          = RCU_FLAVOR,
 179         .init           = rcu_sync_perf_init,
 180         .readlock       = rcu_perf_read_lock,
 181         .readunlock     = rcu_perf_read_unlock,
 182         .get_gp_seq     = rcu_get_gp_seq,
 183         .gp_diff        = rcu_seq_diff,
 184         .exp_completed  = rcu_exp_batches_completed,
 185         .async          = call_rcu,
 186         .gp_barrier     = rcu_barrier,
 187         .sync           = synchronize_rcu,
 188         .exp_sync       = synchronize_rcu_expedited,
 189         .name           = "rcu"
 190 };
 191
 192 /*
 193  * Definitions for rcu_bh perf testing.
 194  */
 195
 196 static int rcu_bh_perf_read_lock(void) __acquires(RCU_BH)
 197 {
 198         rcu_read_lock_bh();
 199         return 0;
 200 }
 201
 202 static void rcu_bh_perf_read_unlock(int idx) __releases(RCU_BH)
 203 {
 204         rcu_read_unlock_bh();
 205 }
 206
 207 static struct rcu_perf_ops rcu_bh_ops = {
 208         .ptype          = RCU_BH_FLAVOR,
 209         .init           = rcu_sync_perf_init,
 210         .readlock       = rcu_bh_perf_read_lock,
 211         .readunlock     = rcu_bh_perf_read_unlock,
 212         .get_gp_seq     = rcu_bh_get_gp_seq,
 213         .gp_diff        = rcu_seq_diff,
 214         .exp_completed  = rcu_exp_batches_completed_sched,
 215         .async          = call_rcu_bh,
 216         .gp_barrier     = rcu_barrier_bh,
 217         .sync           = synchronize_rcu_bh,
 218         .exp_sync       = synchronize_rcu_bh_expedited,
 219         .name           = "rcu_bh"
 220 };
 221
 222 /*
 223  * Definitions for srcu perf testing.
 224  */
 225
 226 DEFINE_STATIC_SRCU(srcu_ctl_perf);
 227 static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
 228
 229 static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
 230 {
 231         return srcu_read_lock(srcu_ctlp);
 232 }
 233
 234 static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
 235 {
 236         srcu_read_unlock(srcu_ctlp, idx);
 237 }
 238
 239 static unsigned long srcu_perf_completed(void)
 240 {
 241         return srcu_batches_completed(srcu_ctlp);
 242 }
 243
 244 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
 245 {
 246         call_srcu(srcu_ctlp, head, func);
 247 }
 248
 249 static void srcu_rcu_barrier(void)
 250 {
 251         srcu_barrier(srcu_ctlp);
 252 }
 253
 254 static void srcu_perf_synchronize(void)
 255 {
 256         synchronize_srcu(srcu_ctlp);
 257 }
 258
 259 static void srcu_perf_synchronize_expedited(void)
 260 {
 261         synchronize_srcu_expedited(srcu_ctlp);
 262 }
 263
 264 static struct rcu_perf_ops srcu_ops = {
 265         .ptype          = SRCU_FLAVOR,
 266         .init           = rcu_sync_perf_init,
 267         .readlock       = srcu_perf_read_lock,
 268         .readunlock     = srcu_perf_read_unlock,
 269         .get_gp_seq     = srcu_perf_completed,
 270         .gp_diff        = rcu_seq_diff,
 271         .exp_completed  = srcu_perf_completed,
 272         .async          = srcu_call_rcu,
 273         .gp_barrier     = srcu_rcu_barrier,
 274         .sync           = srcu_perf_synchronize,
 275         .exp_sync       = srcu_perf_synchronize_expedited,
 276         .name           = "srcu"
 277 };
 278
 279 static struct srcu_struct srcud;
 280
 281 static void srcu_sync_perf_init(void)
 282 {
 283         srcu_ctlp = &srcud;
 284         init_srcu_struct(srcu_ctlp);
 285 }
 286
 287 static void srcu_sync_perf_cleanup(void)
 288 {
 289         cleanup_srcu_struct(srcu_ctlp);
 290 }
 291
 292 static struct rcu_perf_ops srcud_ops = {
 293         .ptype          = SRCU_FLAVOR,
 294         .init           = srcu_sync_perf_init,
 295         .cleanup        = srcu_sync_perf_cleanup,
 296         .readlock       = srcu_perf_read_lock,
 297         .readunlock     = srcu_perf_read_unlock,
 298         .get_gp_seq     = srcu_perf_completed,
 299         .gp_diff        = rcu_seq_diff,
 300         .exp_completed  = srcu_perf_completed,
 301         .async          = srcu_call_rcu,
 302         .gp_barrier     = srcu_rcu_barrier,
 303         .sync           = srcu_perf_synchronize,
 304         .exp_sync       = srcu_perf_synchronize_expedited,
 305         .name           = "srcud"
 306 };
 307
 308 /*
 309  * Definitions for sched perf testing.
 310  */
 311
 312 static int sched_perf_read_lock(void)
 313 {
 314         preempt_disable();
 315         return 0;
 316 }
 317
 318 static void sched_perf_read_unlock(int idx)
 319 {
 320         preempt_enable();
 321 }
 322
 323 static struct rcu_perf_ops sched_ops = {
 324         .ptype          = RCU_SCHED_FLAVOR,
 325         .init           = rcu_sync_perf_init,
 326         .readlock       = sched_perf_read_lock,
 327         .readunlock     = sched_perf_read_unlock,
 328         .get_gp_seq     = rcu_sched_get_gp_seq,
 329         .gp_diff        = rcu_seq_diff,
 330         .exp_completed  = rcu_exp_batches_completed_sched,
 331         .async          = call_rcu_sched,
 332         .gp_barrier     = rcu_barrier_sched,
 333         .sync           = synchronize_sched,
 334         .exp_sync       = synchronize_sched_expedited,
 335         .name           = "sched"
 336 };
 337
 338 /*
 339  * Definitions for RCU-tasks perf testing.
 340  */
 341
 342 static int tasks_perf_read_lock(void)
 343 {
 344         return 0;
 345 }
 346
 347 static void tasks_perf_read_unlock(int idx)
 348 {
 349 }
 350
 351 static struct rcu_perf_ops tasks_ops = {
 352         .ptype          = RCU_TASKS_FLAVOR,
 353         .init           = rcu_sync_perf_init,
 354         .readlock       = tasks_perf_read_lock,
 355         .readunlock     = tasks_perf_read_unlock,
 356         .get_gp_seq     = rcu_no_completed,
 357         .gp_diff        = rcu_seq_diff,
 358         .async          = call_rcu_tasks,
 359         .gp_barrier     = rcu_barrier_tasks,
 360         .sync           = synchronize_rcu_tasks,
 361         .exp_sync       = synchronize_rcu_tasks,
 362         .name           = "tasks"
 363 };
 364
 365 static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old)
 366 {
 367         if (!cur_ops->gp_diff)
 368                 return new - old;
 369         return cur_ops->gp_diff(new, old);
 370 }
 371
 372 /*
 373  * If performance tests complete, wait for shutdown to commence.
 374  */
 375 static void rcu_perf_wait_shutdown(void)
 376 {
 377         cond_resched_tasks_rcu_qs();
 378         if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
 379                 return;
 380         while (!torture_must_stop())
 381                 schedule_timeout_uninterruptible(1);
 382 }
 383
 384 /*
 385  * RCU perf reader kthread.  Repeatedly does empty RCU read-side
 386  * critical section, minimizing update-side interference.
 387  */
 388 static int
 389 rcu_perf_reader(void *arg)
 390 {
 391         unsigned long flags;
 392         int idx;
 393         long me = (long)arg;
 394
 395         VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
 396         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 397         set_user_nice(current, MAX_NICE);
 398         atomic_inc(&n_rcu_perf_reader_started);
 399
 400         do {
 401                 local_irq_save(flags);
 402                 idx = cur_ops->readlock();
 403                 cur_ops->readunlock(idx);
 404                 local_irq_restore(flags);
 405                 rcu_perf_wait_shutdown();
 406         } while (!torture_must_stop());
 407         torture_kthread_stopping("rcu_perf_reader");
 408         return 0;
 409 }
 410
 411 /*
 412  * Callback function for asynchronous grace periods from rcu_perf_writer().
 413  */
 414 static void rcu_perf_async_cb(struct rcu_head *rhp)
 415 {
 416         atomic_dec(this_cpu_ptr(&n_async_inflight));
 417         kfree(rhp);
 418 }
 419
 420 /*
 421  * RCU perf writer kthread.  Repeatedly does a grace period.
 422  */
 423 static int
 424 rcu_perf_writer(void *arg)
 425 {
 426         int i = 0;
 427         int i_max;
 428         long me = (long)arg;
 429         struct rcu_head *rhp = NULL;
 430         struct sched_param sp;
 431         bool started = false, done = false, alldone = false;
 432         u64 t;
 433         u64 *wdp;
 434         u64 *wdpp = writer_durations[me];
 435
 436         VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
 437         WARN_ON(!wdpp);
 438         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 439         sp.sched_priority = 1;
 440         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 441
 442         if (holdoff)
 443                 schedule_timeout_uninterruptible(holdoff * HZ);
 444
 445         t = ktime_get_mono_fast_ns();
 446         if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
 447                 t_rcu_perf_writer_started = t;
 448                 if (gp_exp) {
 449                         b_rcu_perf_writer_started =
 450                                 cur_ops->exp_completed() / 2;
 451                 } else {
 452                         b_rcu_perf_writer_started = cur_ops->get_gp_seq();
 453                 }
 454         }
 455
 456         do {
 457                 if (writer_holdoff)
 458                         udelay(writer_holdoff);
 459                 wdp = &wdpp[i];
 460                 *wdp = ktime_get_mono_fast_ns();
 461                 if (gp_async) {
 462 retry:
 463                         if (!rhp)
 464                                 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
 465                         if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
 466                                 rcu_perf_writer_state = RTWS_ASYNC;
 467                                 atomic_inc(this_cpu_ptr(&n_async_inflight));
 468                                 cur_ops->async(rhp, rcu_perf_async_cb);
 469                                 rhp = NULL;
 470                         } else if (!kthread_should_stop()) {
 471                                 rcu_perf_writer_state = RTWS_BARRIER;
 472                                 cur_ops->gp_barrier();
 473                                 goto retry;
 474                         } else {
 475                                 kfree(rhp); /* Because we are stopping. */
 476                         }
 477                 } else if (gp_exp) {
 478                         rcu_perf_writer_state = RTWS_EXP_SYNC;
 479                         cur_ops->exp_sync();
 480                 } else {
 481                         rcu_perf_writer_state = RTWS_SYNC;
 482                         cur_ops->sync();
 483                 }
 484                 rcu_perf_writer_state = RTWS_IDLE;
 485                 t = ktime_get_mono_fast_ns();
 486                 *wdp = t - *wdp;
 487                 i_max = i;
 488                 if (!started &&
 489                     atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
 490                         started = true;
 491                 if (!done && i >= MIN_MEAS) {
 492                         done = true;
 493                         sp.sched_priority = 0;
 494                         sched_setscheduler_nocheck(current,
 495                                                    SCHED_NORMAL, &sp);
 496                         pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
 497                                  perf_type, PERF_FLAG, me, MIN_MEAS);
 498                         if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
 499                             nrealwriters) {
 500                                 schedule_timeout_interruptible(10);
 501                                 rcu_ftrace_dump(DUMP_ALL);
 502                                 PERFOUT_STRING("Test complete");
 503                                 t_rcu_perf_writer_finished = t;
 504                                 if (gp_exp) {
 505                                         b_rcu_perf_writer_finished =
 506                                                 cur_ops->exp_completed() / 2;
 507                                 } else {
 508                                         b_rcu_perf_writer_finished =
 509                                                 cur_ops->get_gp_seq();
 510                                 }
 511                                 if (shutdown) {
 512                                         smp_mb(); /* Assign before wake. */
 513                                         wake_up(&shutdown_wq);
 514                                 }
 515                         }
 516                 }
 517                 if (done && !alldone &&
 518                     atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
 519                         alldone = true;
 520                 if (started && !alldone && i < MAX_MEAS - 1)
 521                         i++;
 522                 rcu_perf_wait_shutdown();
 523         } while (!torture_must_stop());
 524         if (gp_async) {
 525                 rcu_perf_writer_state = RTWS_BARRIER;
 526                 cur_ops->gp_barrier();
 527         }
 528         rcu_perf_writer_state = RTWS_STOPPING;
 529         writer_n_durations[me] = i_max;
 530         torture_kthread_stopping("rcu_perf_writer");
 531         return 0;
 532 }
 533
 534 static void
 535 rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
 536 {
 537         pr_alert("%s" PERF_FLAG
 538                  "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
 539                  perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
 540 }
 541
 542 static void
 543 rcu_perf_cleanup(void)
 544 {
 545         int i;
 546         int j;
 547         int ngps = 0;
 548         u64 *wdp;
 549         u64 *wdpp;
 550
 551         /*
 552          * Would like warning at start, but everything is expedited
 553          * during the mid-boot phase, so have to wait till the end.
 554          */
 555         if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
 556                 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
 557         if (rcu_gp_is_normal() && gp_exp)
 558                 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
 559         if (gp_exp && gp_async)
 560                 VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
 561
 562         if (torture_cleanup_begin())
 563                 return;
 564
 565         if (reader_tasks) {
 566                 for (i = 0; i < nrealreaders; i++)
 567                         torture_stop_kthread(rcu_perf_reader,
 568                                              reader_tasks[i]);
 569                 kfree(reader_tasks);
 570         }
 571
 572         if (writer_tasks) {
 573                 for (i = 0; i < nrealwriters; i++) {
 574                         torture_stop_kthread(rcu_perf_writer,
 575                                              writer_tasks[i]);
 576                         if (!writer_n_durations)
 577                                 continue;
 578                         j = writer_n_durations[i];
 579                         pr_alert("%s%s writer %d gps: %d\n",
 580                                  perf_type, PERF_FLAG, i, j);
 581                         ngps += j;
 582                 }
 583                 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
 584                          perf_type, PERF_FLAG,
 585                          t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
 586                          t_rcu_perf_writer_finished -
 587                          t_rcu_perf_writer_started,
 588                          ngps,
 589                          rcuperf_seq_diff(b_rcu_perf_writer_finished,
 590                                           b_rcu_perf_writer_started));
 591                 for (i = 0; i < nrealwriters; i++) {
 592                         if (!writer_durations)
 593                                 break;
 594                         if (!writer_n_durations)
 595                                 continue;
 596                         wdpp = writer_durations[i];
 597                         if (!wdpp)
 598                                 continue;
 599                         for (j = 0; j <= writer_n_durations[i]; j++) {
 600                                 wdp = &wdpp[j];
 601                                 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
 602                                         perf_type, PERF_FLAG,
 603                                         i, j, *wdp);
 604                                 if (j % 100 == 0)
 605                                         schedule_timeout_uninterruptible(1);
 606                         }
 607                         kfree(writer_durations[i]);
 608                 }
 609                 kfree(writer_tasks);
 610                 kfree(writer_durations);
 611                 kfree(writer_n_durations);
 612         }
 613
 614         /* Do flavor-specific cleanup operations.  */
 615         if (cur_ops->cleanup != NULL)
 616                 cur_ops->cleanup();
 617
 618         torture_cleanup_end();
 619 }
 620
 621 /*
 622  * Return the number if non-negative.  If -1, the number of CPUs.
 623  * If less than -1, that much less than the number of CPUs, but
 624  * at least one.
 625  */
 626 static int compute_real(int n)
 627 {
 628         int nr;
 629
 630         if (n >= 0) {
 631                 nr = n;
 632         } else {
 633                 nr = num_online_cpus() + 1 + n;
 634                 if (nr <= 0)
 635                         nr = 1;
 636         }
 637         return nr;
 638 }
 639
 640 /*
 641  * RCU perf shutdown kthread.  Just waits to be awakened, then shuts
 642  * down system.
 643  */
 644 static int
 645 rcu_perf_shutdown(void *arg)
 646 {
 647         do {
 648                 wait_event(shutdown_wq,
 649                            atomic_read(&n_rcu_perf_writer_finished) >=
 650                            nrealwriters);
 651         } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters);
 652         smp_mb(); /* Wake before output. */
 653         rcu_perf_cleanup();
 654         kernel_power_off();
 655         return -EINVAL;
 656 }
 657
 658 static int __init
 659 rcu_perf_init(void)
 660 {
 661         long i;
 662         int firsterr = 0;
 663         static struct rcu_perf_ops *perf_ops[] = {
 664                 &rcu_ops, &rcu_bh_ops, &srcu_ops, &srcud_ops, &sched_ops,
 665                 &tasks_ops,
 666         };
 667
 668         if (!torture_init_begin(perf_type, verbose))
 669                 return -EBUSY;
 670
 671         /* Process args and tell the world that the perf'er is on the job. */
 672         for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
 673                 cur_ops = perf_ops[i];
 674                 if (strcmp(perf_type, cur_ops->name) == 0)
 675                         break;
 676         }
 677         if (i == ARRAY_SIZE(perf_ops)) {
 678                 pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type);
 679                 pr_alert("rcu-perf types:");
 680                 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
 681                         pr_cont(" %s", perf_ops[i]->name);
 682                 pr_cont("\n");
 683                 firsterr = -EINVAL;
 684                 goto unwind;
 685         }
 686         if (cur_ops->init)
 687                 cur_ops->init();
 688
 689         nrealwriters = compute_real(nwriters);
 690         nrealreaders = compute_real(nreaders);
 691         atomic_set(&n_rcu_perf_reader_started, 0);
 692         atomic_set(&n_rcu_perf_writer_started, 0);
 693         atomic_set(&n_rcu_perf_writer_finished, 0);
 694         rcu_perf_print_module_parms(cur_ops, "Start of test");
 695
 696         /* Start up the kthreads. */
 697
 698         if (shutdown) {
 699                 init_waitqueue_head(&shutdown_wq);
 700                 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
 701                                                   shutdown_task);
 702                 if (firsterr)
 703                         goto unwind;
 704                 schedule_timeout_uninterruptible(1);
 705         }
 706         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
 707                                GFP_KERNEL);
 708         if (reader_tasks == NULL) {
 709                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 710                 firsterr = -ENOMEM;
 711                 goto unwind;
 712         }
 713         for (i = 0; i < nrealreaders; i++) {
 714                 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
 715                                                   reader_tasks[i]);
 716                 if (firsterr)
 717                         goto unwind;
 718         }
 719         while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
 720                 schedule_timeout_uninterruptible(1);
 721         writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
 722                                GFP_KERNEL);
 723         writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
 724                                    GFP_KERNEL);
 725         writer_n_durations =
 726                 kcalloc(nrealwriters, sizeof(*writer_n_durations),
 727                         GFP_KERNEL);
 728         if (!writer_tasks || !writer_durations || !writer_n_durations) {
 729                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 730                 firsterr = -ENOMEM;
 731                 goto unwind;
 732         }
 733         for (i = 0; i < nrealwriters; i++) {
 734                 writer_durations[i] =
 735                         kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
 736                                 GFP_KERNEL);
 737                 if (!writer_durations[i]) {
 738                         firsterr = -ENOMEM;
 739                         goto unwind;
 740                 }
 741                 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
 742                                                   writer_tasks[i]);
 743                 if (firsterr)
 744                         goto unwind;
 745         }
 746         torture_init_end();
 747         return 0;
 748
 749 unwind:
 750         torture_init_end();
 751         rcu_perf_cleanup();
 752         return firsterr;
 753 }
 754
 755 module_init(rcu_perf_init);
 756 module_exit(rcu_perf_cleanup);