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/mm.h>
  16 #include <linux/module.h>
  17 #include <linux/kthread.h>
  18 #include <linux/err.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/smp.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/sched.h>
  24 #include <uapi/linux/sched/types.h>
  25 #include <linux/atomic.h>
  26 #include <linux/bitops.h>
  27 #include <linux/completion.h>
  28 #include <linux/moduleparam.h>
  29 #include <linux/percpu.h>
  30 #include <linux/notifier.h>
  31 #include <linux/reboot.h>
  32 #include <linux/freezer.h>
  33 #include <linux/cpu.h>
  34 #include <linux/delay.h>
  35 #include <linux/stat.h>
  36 #include <linux/srcu.h>
  37 #include <linux/slab.h>
  38 #include <asm/byteorder.h>
  39 #include <linux/torture.h>
  40 #include <linux/vmalloc.h>
  41
  42 #include "rcu.h"
  43
  44 MODULE_LICENSE("GPL");
  45 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
  46
  47 #define PERF_FLAG "-perf:"
  48 #define PERFOUT_STRING(s) \
  49         pr_alert("%s" PERF_FLAG " %s\n", perf_type, s)
  50 #define VERBOSE_PERFOUT_STRING(s) \
  51         do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0)
  52 #define VERBOSE_PERFOUT_ERRSTRING(s) \
  53         do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0)
  54
  55 /*
  56  * The intended use cases for the nreaders and nwriters module parameters
  57  * are as follows:
  58  *
  59  * 1.   Specify only the nr_cpus kernel boot parameter.  This will
  60  *      set both nreaders and nwriters to the value specified by
  61  *      nr_cpus for a mixed reader/writer test.
  62  *
  63  * 2.   Specify the nr_cpus kernel boot parameter, but set
  64  *      rcuperf.nreaders to zero.  This will set nwriters to the
  65  *      value specified by nr_cpus for an update-only test.
  66  *
  67  * 3.   Specify the nr_cpus kernel boot parameter, but set
  68  *      rcuperf.nwriters to zero.  This will set nreaders to the
  69  *      value specified by nr_cpus for a read-only test.
  70  *
  71  * Various other use cases may of course be specified.
  72  *
  73  * Note that this test's readers are intended only as a test load for
  74  * the writers.  The reader performance statistics will be overly
  75  * pessimistic due to the per-critical-section interrupt disabling,
  76  * test-end checks, and the pair of calls through pointers.
  77  */
  78
  79 #ifdef MODULE
  80 # define RCUPERF_SHUTDOWN 0
  81 #else
  82 # define RCUPERF_SHUTDOWN 1
  83 #endif
  84
  85 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
  86 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader");
  87 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
  88 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
  89 torture_param(int, nreaders, -1, "Number of RCU reader threads");
  90 torture_param(int, nwriters, -1, "Number of RCU updater threads");
  91 torture_param(bool, shutdown, RCUPERF_SHUTDOWN,
  92               "Shutdown at end of performance tests.");
  93 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
  94 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
  95 torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() perf test?");
  96 torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate.");
  97
  98 static char *perf_type = "rcu";
  99 module_param(perf_type, charp, 0444);
 100 MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, srcu, ...)");
 101
 102 static int nrealreaders;
 103 static int nrealwriters;
 104 static struct task_struct **writer_tasks;
 105 static struct task_struct **reader_tasks;
 106 static struct task_struct *shutdown_task;
 107
 108 static u64 **writer_durations;
 109 static int *writer_n_durations;
 110 static atomic_t n_rcu_perf_reader_started;
 111 static atomic_t n_rcu_perf_writer_started;
 112 static atomic_t n_rcu_perf_writer_finished;
 113 static wait_queue_head_t shutdown_wq;
 114 static u64 t_rcu_perf_writer_started;
 115 static u64 t_rcu_perf_writer_finished;
 116 static unsigned long b_rcu_gp_test_started;
 117 static unsigned long b_rcu_gp_test_finished;
 118 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
 119
 120 #define MAX_MEAS 10000
 121 #define MIN_MEAS 100
 122
 123 /*
 124  * Operations vector for selecting different types of tests.
 125  */
 126
 127 struct rcu_perf_ops {
 128         int ptype;
 129         void (*init)(void);
 130         void (*cleanup)(void);
 131         int (*readlock)(void);
 132         void (*readunlock)(int idx);
 133         unsigned long (*get_gp_seq)(void);
 134         unsigned long (*gp_diff)(unsigned long new, unsigned long old);
 135         unsigned long (*exp_completed)(void);
 136         void (*async)(struct rcu_head *head, rcu_callback_t func);
 137         void (*gp_barrier)(void);
 138         void (*sync)(void);
 139         void (*exp_sync)(void);
 140         const char *name;
 141 };
 142
 143 static struct rcu_perf_ops *cur_ops;
 144
 145 /*
 146  * Definitions for rcu perf testing.
 147  */
 148
 149 static int rcu_perf_read_lock(void) __acquires(RCU)
 150 {
 151         rcu_read_lock();
 152         return 0;
 153 }
 154
 155 static void rcu_perf_read_unlock(int idx) __releases(RCU)
 156 {
 157         rcu_read_unlock();
 158 }
 159
 160 static unsigned long __maybe_unused rcu_no_completed(void)
 161 {
 162         return 0;
 163 }
 164
 165 static void rcu_sync_perf_init(void)
 166 {
 167 }
 168
 169 static struct rcu_perf_ops rcu_ops = {
 170         .ptype          = RCU_FLAVOR,
 171         .init           = rcu_sync_perf_init,
 172         .readlock       = rcu_perf_read_lock,
 173         .readunlock     = rcu_perf_read_unlock,
 174         .get_gp_seq     = rcu_get_gp_seq,
 175         .gp_diff        = rcu_seq_diff,
 176         .exp_completed  = rcu_exp_batches_completed,
 177         .async          = call_rcu,
 178         .gp_barrier     = rcu_barrier,
 179         .sync           = synchronize_rcu,
 180         .exp_sync       = synchronize_rcu_expedited,
 181         .name           = "rcu"
 182 };
 183
 184 /*
 185  * Definitions for srcu perf testing.
 186  */
 187
 188 DEFINE_STATIC_SRCU(srcu_ctl_perf);
 189 static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf;
 190
 191 static int srcu_perf_read_lock(void) __acquires(srcu_ctlp)
 192 {
 193         return srcu_read_lock(srcu_ctlp);
 194 }
 195
 196 static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp)
 197 {
 198         srcu_read_unlock(srcu_ctlp, idx);
 199 }
 200
 201 static unsigned long srcu_perf_completed(void)
 202 {
 203         return srcu_batches_completed(srcu_ctlp);
 204 }
 205
 206 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
 207 {
 208         call_srcu(srcu_ctlp, head, func);
 209 }
 210
 211 static void srcu_rcu_barrier(void)
 212 {
 213         srcu_barrier(srcu_ctlp);
 214 }
 215
 216 static void srcu_perf_synchronize(void)
 217 {
 218         synchronize_srcu(srcu_ctlp);
 219 }
 220
 221 static void srcu_perf_synchronize_expedited(void)
 222 {
 223         synchronize_srcu_expedited(srcu_ctlp);
 224 }
 225
 226 static struct rcu_perf_ops srcu_ops = {
 227         .ptype          = SRCU_FLAVOR,
 228         .init           = rcu_sync_perf_init,
 229         .readlock       = srcu_perf_read_lock,
 230         .readunlock     = srcu_perf_read_unlock,
 231         .get_gp_seq     = srcu_perf_completed,
 232         .gp_diff        = rcu_seq_diff,
 233         .exp_completed  = srcu_perf_completed,
 234         .async          = srcu_call_rcu,
 235         .gp_barrier     = srcu_rcu_barrier,
 236         .sync           = srcu_perf_synchronize,
 237         .exp_sync       = srcu_perf_synchronize_expedited,
 238         .name           = "srcu"
 239 };
 240
 241 static struct srcu_struct srcud;
 242
 243 static void srcu_sync_perf_init(void)
 244 {
 245         srcu_ctlp = &srcud;
 246         init_srcu_struct(srcu_ctlp);
 247 }
 248
 249 static void srcu_sync_perf_cleanup(void)
 250 {
 251         cleanup_srcu_struct(srcu_ctlp);
 252 }
 253
 254 static struct rcu_perf_ops srcud_ops = {
 255         .ptype          = SRCU_FLAVOR,
 256         .init           = srcu_sync_perf_init,
 257         .cleanup        = srcu_sync_perf_cleanup,
 258         .readlock       = srcu_perf_read_lock,
 259         .readunlock     = srcu_perf_read_unlock,
 260         .get_gp_seq     = srcu_perf_completed,
 261         .gp_diff        = rcu_seq_diff,
 262         .exp_completed  = srcu_perf_completed,
 263         .async          = srcu_call_rcu,
 264         .gp_barrier     = srcu_rcu_barrier,
 265         .sync           = srcu_perf_synchronize,
 266         .exp_sync       = srcu_perf_synchronize_expedited,
 267         .name           = "srcud"
 268 };
 269
 270 /*
 271  * Definitions for RCU-tasks perf testing.
 272  */
 273
 274 static int tasks_perf_read_lock(void)
 275 {
 276         return 0;
 277 }
 278
 279 static void tasks_perf_read_unlock(int idx)
 280 {
 281 }
 282
 283 static struct rcu_perf_ops tasks_ops = {
 284         .ptype          = RCU_TASKS_FLAVOR,
 285         .init           = rcu_sync_perf_init,
 286         .readlock       = tasks_perf_read_lock,
 287         .readunlock     = tasks_perf_read_unlock,
 288         .get_gp_seq     = rcu_no_completed,
 289         .gp_diff        = rcu_seq_diff,
 290         .async          = call_rcu_tasks,
 291         .gp_barrier     = rcu_barrier_tasks,
 292         .sync           = synchronize_rcu_tasks,
 293         .exp_sync       = synchronize_rcu_tasks,
 294         .name           = "tasks"
 295 };
 296
 297 static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old)
 298 {
 299         if (!cur_ops->gp_diff)
 300                 return new - old;
 301         return cur_ops->gp_diff(new, old);
 302 }
 303
 304 /*
 305  * If performance tests complete, wait for shutdown to commence.
 306  */
 307 static void rcu_perf_wait_shutdown(void)
 308 {
 309         cond_resched_tasks_rcu_qs();
 310         if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters)
 311                 return;
 312         while (!torture_must_stop())
 313                 schedule_timeout_uninterruptible(1);
 314 }
 315
 316 /*
 317  * RCU perf reader kthread.  Repeatedly does empty RCU read-side critical
 318  * section, minimizing update-side interference.  However, the point of
 319  * this test is not to evaluate reader performance, but instead to serve
 320  * as a test load for update-side performance testing.
 321  */
 322 static int
 323 rcu_perf_reader(void *arg)
 324 {
 325         unsigned long flags;
 326         int idx;
 327         long me = (long)arg;
 328
 329         VERBOSE_PERFOUT_STRING("rcu_perf_reader task started");
 330         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 331         set_user_nice(current, MAX_NICE);
 332         atomic_inc(&n_rcu_perf_reader_started);
 333
 334         do {
 335                 local_irq_save(flags);
 336                 idx = cur_ops->readlock();
 337                 cur_ops->readunlock(idx);
 338                 local_irq_restore(flags);
 339                 rcu_perf_wait_shutdown();
 340         } while (!torture_must_stop());
 341         torture_kthread_stopping("rcu_perf_reader");
 342         return 0;
 343 }
 344
 345 /*
 346  * Callback function for asynchronous grace periods from rcu_perf_writer().
 347  */
 348 static void rcu_perf_async_cb(struct rcu_head *rhp)
 349 {
 350         atomic_dec(this_cpu_ptr(&n_async_inflight));
 351         kfree(rhp);
 352 }
 353
 354 /*
 355  * RCU perf writer kthread.  Repeatedly does a grace period.
 356  */
 357 static int
 358 rcu_perf_writer(void *arg)
 359 {
 360         int i = 0;
 361         int i_max;
 362         long me = (long)arg;
 363         struct rcu_head *rhp = NULL;
 364         struct sched_param sp;
 365         bool started = false, done = false, alldone = false;
 366         u64 t;
 367         u64 *wdp;
 368         u64 *wdpp = writer_durations[me];
 369
 370         VERBOSE_PERFOUT_STRING("rcu_perf_writer task started");
 371         WARN_ON(!wdpp);
 372         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 373         sp.sched_priority = 1;
 374         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
 375
 376         if (holdoff)
 377                 schedule_timeout_uninterruptible(holdoff * HZ);
 378
 379         /*
 380          * Wait until rcu_end_inkernel_boot() is called for normal GP tests
 381          * so that RCU is not always expedited for normal GP tests.
 382          * The system_state test is approximate, but works well in practice.
 383          */
 384         while (!gp_exp && system_state != SYSTEM_RUNNING)
 385                 schedule_timeout_uninterruptible(1);
 386
 387         t = ktime_get_mono_fast_ns();
 388         if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) {
 389                 t_rcu_perf_writer_started = t;
 390                 if (gp_exp) {
 391                         b_rcu_gp_test_started =
 392                                 cur_ops->exp_completed() / 2;
 393                 } else {
 394                         b_rcu_gp_test_started = cur_ops->get_gp_seq();
 395                 }
 396         }
 397
 398         do {
 399                 if (writer_holdoff)
 400                         udelay(writer_holdoff);
 401                 wdp = &wdpp[i];
 402                 *wdp = ktime_get_mono_fast_ns();
 403                 if (gp_async) {
 404 retry:
 405                         if (!rhp)
 406                                 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
 407                         if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
 408                                 atomic_inc(this_cpu_ptr(&n_async_inflight));
 409                                 cur_ops->async(rhp, rcu_perf_async_cb);
 410                                 rhp = NULL;
 411                         } else if (!kthread_should_stop()) {
 412                                 cur_ops->gp_barrier();
 413                                 goto retry;
 414                         } else {
 415                                 kfree(rhp); /* Because we are stopping. */
 416                         }
 417                 } else if (gp_exp) {
 418                         cur_ops->exp_sync();
 419                 } else {
 420                         cur_ops->sync();
 421                 }
 422                 t = ktime_get_mono_fast_ns();
 423                 *wdp = t - *wdp;
 424                 i_max = i;
 425                 if (!started &&
 426                     atomic_read(&n_rcu_perf_writer_started) >= nrealwriters)
 427                         started = true;
 428                 if (!done && i >= MIN_MEAS) {
 429                         done = true;
 430                         sp.sched_priority = 0;
 431                         sched_setscheduler_nocheck(current,
 432                                                    SCHED_NORMAL, &sp);
 433                         pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n",
 434                                  perf_type, PERF_FLAG, me, MIN_MEAS);
 435                         if (atomic_inc_return(&n_rcu_perf_writer_finished) >=
 436                             nrealwriters) {
 437                                 schedule_timeout_interruptible(10);
 438                                 rcu_ftrace_dump(DUMP_ALL);
 439                                 PERFOUT_STRING("Test complete");
 440                                 t_rcu_perf_writer_finished = t;
 441                                 if (gp_exp) {
 442                                         b_rcu_gp_test_finished =
 443                                                 cur_ops->exp_completed() / 2;
 444                                 } else {
 445                                         b_rcu_gp_test_finished =
 446                                                 cur_ops->get_gp_seq();
 447                                 }
 448                                 if (shutdown) {
 449                                         smp_mb(); /* Assign before wake. */
 450                                         wake_up(&shutdown_wq);
 451                                 }
 452                         }
 453                 }
 454                 if (done && !alldone &&
 455                     atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters)
 456                         alldone = true;
 457                 if (started && !alldone && i < MAX_MEAS - 1)
 458                         i++;
 459                 rcu_perf_wait_shutdown();
 460         } while (!torture_must_stop());
 461         if (gp_async) {
 462                 cur_ops->gp_barrier();
 463         }
 464         writer_n_durations[me] = i_max;
 465         torture_kthread_stopping("rcu_perf_writer");
 466         return 0;
 467 }
 468
 469 static void
 470 rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag)
 471 {
 472         pr_alert("%s" PERF_FLAG
 473                  "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
 474                  perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
 475 }
 476
 477 static void
 478 rcu_perf_cleanup(void)
 479 {
 480         int i;
 481         int j;
 482         int ngps = 0;
 483         u64 *wdp;
 484         u64 *wdpp;
 485
 486         /*
 487          * Would like warning at start, but everything is expedited
 488          * during the mid-boot phase, so have to wait till the end.
 489          */
 490         if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
 491                 VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
 492         if (rcu_gp_is_normal() && gp_exp)
 493                 VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
 494         if (gp_exp && gp_async)
 495                 VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!");
 496
 497         if (torture_cleanup_begin())
 498                 return;
 499         if (!cur_ops) {
 500                 torture_cleanup_end();
 501                 return;
 502         }
 503
 504         if (reader_tasks) {
 505                 for (i = 0; i < nrealreaders; i++)
 506                         torture_stop_kthread(rcu_perf_reader,
 507                                              reader_tasks[i]);
 508                 kfree(reader_tasks);
 509         }
 510
 511         if (writer_tasks) {
 512                 for (i = 0; i < nrealwriters; i++) {
 513                         torture_stop_kthread(rcu_perf_writer,
 514                                              writer_tasks[i]);
 515                         if (!writer_n_durations)
 516                                 continue;
 517                         j = writer_n_durations[i];
 518                         pr_alert("%s%s writer %d gps: %d\n",
 519                                  perf_type, PERF_FLAG, i, j);
 520                         ngps += j;
 521                 }
 522                 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
 523                          perf_type, PERF_FLAG,
 524                          t_rcu_perf_writer_started, t_rcu_perf_writer_finished,
 525                          t_rcu_perf_writer_finished -
 526                          t_rcu_perf_writer_started,
 527                          ngps,
 528                          rcuperf_seq_diff(b_rcu_gp_test_finished,
 529                                           b_rcu_gp_test_started));
 530                 for (i = 0; i < nrealwriters; i++) {
 531                         if (!writer_durations)
 532                                 break;
 533                         if (!writer_n_durations)
 534                                 continue;
 535                         wdpp = writer_durations[i];
 536                         if (!wdpp)
 537                                 continue;
 538                         for (j = 0; j <= writer_n_durations[i]; j++) {
 539                                 wdp = &wdpp[j];
 540                                 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
 541                                         perf_type, PERF_FLAG,
 542                                         i, j, *wdp);
 543                                 if (j % 100 == 0)
 544                                         schedule_timeout_uninterruptible(1);
 545                         }
 546                         kfree(writer_durations[i]);
 547                 }
 548                 kfree(writer_tasks);
 549                 kfree(writer_durations);
 550                 kfree(writer_n_durations);
 551         }
 552
 553         /* Do torture-type-specific cleanup operations.  */
 554         if (cur_ops->cleanup != NULL)
 555                 cur_ops->cleanup();
 556
 557         torture_cleanup_end();
 558 }
 559
 560 /*
 561  * Return the number if non-negative.  If -1, the number of CPUs.
 562  * If less than -1, that much less than the number of CPUs, but
 563  * at least one.
 564  */
 565 static int compute_real(int n)
 566 {
 567         int nr;
 568
 569         if (n >= 0) {
 570                 nr = n;
 571         } else {
 572                 nr = num_online_cpus() + 1 + n;
 573                 if (nr <= 0)
 574                         nr = 1;
 575         }
 576         return nr;
 577 }
 578
 579 /*
 580  * RCU perf shutdown kthread.  Just waits to be awakened, then shuts
 581  * down system.
 582  */
 583 static int
 584 rcu_perf_shutdown(void *arg)
 585 {
 586         wait_event(shutdown_wq,
 587                    atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters);
 588         smp_mb(); /* Wake before output. */
 589         rcu_perf_cleanup();
 590         kernel_power_off();
 591         return -EINVAL;
 592 }
 593
 594 /*
 595  * kfree_rcu() performance tests: Start a kfree_rcu() loop on all CPUs for number
 596  * of iterations and measure total time and number of GP for all iterations to complete.
 597  */
 598
 599 torture_param(int, kfree_nthreads, -1, "Number of threads running loops of kfree_rcu().");
 600 torture_param(int, kfree_alloc_num, 8000, "Number of allocations and frees done in an iteration.");
 601 torture_param(int, kfree_loops, 10, "Number of loops doing kfree_alloc_num allocations and frees.");
 602
 603 static struct task_struct **kfree_reader_tasks;
 604 static int kfree_nrealthreads;
 605 static atomic_t n_kfree_perf_thread_started;
 606 static atomic_t n_kfree_perf_thread_ended;
 607
 608 struct kfree_obj {
 609         char kfree_obj[8];
 610         struct rcu_head rh;
 611 };
 612
 613 static int
 614 kfree_perf_thread(void *arg)
 615 {
 616         int i, loop = 0;
 617         long me = (long)arg;
 618         struct kfree_obj *alloc_ptr;
 619         u64 start_time, end_time;
 620         long long mem_begin, mem_during = 0;
 621
 622         VERBOSE_PERFOUT_STRING("kfree_perf_thread task started");
 623         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 624         set_user_nice(current, MAX_NICE);
 625
 626         start_time = ktime_get_mono_fast_ns();
 627
 628         if (atomic_inc_return(&n_kfree_perf_thread_started) >= kfree_nrealthreads) {
 629                 if (gp_exp)
 630                         b_rcu_gp_test_started = cur_ops->exp_completed() / 2;
 631                 else
 632                         b_rcu_gp_test_started = cur_ops->get_gp_seq();
 633         }
 634
 635         do {
 636                 if (!mem_during) {
 637                         mem_during = mem_begin = si_mem_available();
 638                 } else if (loop % (kfree_loops / 4) == 0) {
 639                         mem_during = (mem_during + si_mem_available()) / 2;
 640                 }
 641
 642                 for (i = 0; i < kfree_alloc_num; i++) {
 643                         alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL);
 644                         if (!alloc_ptr)
 645                                 return -ENOMEM;
 646
 647                         kfree_rcu(alloc_ptr, rh);
 648                 }
 649
 650                 cond_resched();
 651         } while (!torture_must_stop() && ++loop < kfree_loops);
 652
 653         if (atomic_inc_return(&n_kfree_perf_thread_ended) >= kfree_nrealthreads) {
 654                 end_time = ktime_get_mono_fast_ns();
 655
 656                 if (gp_exp)
 657                         b_rcu_gp_test_finished = cur_ops->exp_completed() / 2;
 658                 else
 659                         b_rcu_gp_test_finished = cur_ops->get_gp_seq();
 660
 661                 pr_alert("Total time taken by all kfree'ers: %llu ns, loops: %d, batches: %ld, memory footprint: %lldMB\n",
 662                        (unsigned long long)(end_time - start_time), kfree_loops,
 663                        rcuperf_seq_diff(b_rcu_gp_test_finished, b_rcu_gp_test_started),
 664                        (mem_begin - mem_during) >> (20 - PAGE_SHIFT));
 665
 666                 if (shutdown) {
 667                         smp_mb(); /* Assign before wake. */
 668                         wake_up(&shutdown_wq);
 669                 }
 670         }
 671
 672         torture_kthread_stopping("kfree_perf_thread");
 673         return 0;
 674 }
 675
 676 static void
 677 kfree_perf_cleanup(void)
 678 {
 679         int i;
 680
 681         if (torture_cleanup_begin())
 682                 return;
 683
 684         if (kfree_reader_tasks) {
 685                 for (i = 0; i < kfree_nrealthreads; i++)
 686                         torture_stop_kthread(kfree_perf_thread,
 687                                              kfree_reader_tasks[i]);
 688                 kfree(kfree_reader_tasks);
 689         }
 690
 691         torture_cleanup_end();
 692 }
 693
 694 /*
 695  * shutdown kthread.  Just waits to be awakened, then shuts down system.
 696  */
 697 static int
 698 kfree_perf_shutdown(void *arg)
 699 {
 700         wait_event(shutdown_wq,
 701                    atomic_read(&n_kfree_perf_thread_ended) >= kfree_nrealthreads);
 702
 703         smp_mb(); /* Wake before output. */
 704
 705         kfree_perf_cleanup();
 706         kernel_power_off();
 707         return -EINVAL;
 708 }
 709
 710 static int __init
 711 kfree_perf_init(void)
 712 {
 713         long i;
 714         int firsterr = 0;
 715
 716         kfree_nrealthreads = compute_real(kfree_nthreads);
 717         /* Start up the kthreads. */
 718         if (shutdown) {
 719                 init_waitqueue_head(&shutdown_wq);
 720                 firsterr = torture_create_kthread(kfree_perf_shutdown, NULL,
 721                                                   shutdown_task);
 722                 if (firsterr)
 723                         goto unwind;
 724                 schedule_timeout_uninterruptible(1);
 725         }
 726
 727         pr_alert("kfree object size=%zu\n", kfree_mult * sizeof(struct kfree_obj));
 728
 729         kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]),
 730                                GFP_KERNEL);
 731         if (kfree_reader_tasks == NULL) {
 732                 firsterr = -ENOMEM;
 733                 goto unwind;
 734         }
 735
 736         for (i = 0; i < kfree_nrealthreads; i++) {
 737                 firsterr = torture_create_kthread(kfree_perf_thread, (void *)i,
 738                                                   kfree_reader_tasks[i]);
 739                 if (firsterr)
 740                         goto unwind;
 741         }
 742
 743         while (atomic_read(&n_kfree_perf_thread_started) < kfree_nrealthreads)
 744                 schedule_timeout_uninterruptible(1);
 745
 746         torture_init_end();
 747         return 0;
 748
 749 unwind:
 750         torture_init_end();
 751         kfree_perf_cleanup();
 752         return firsterr;
 753 }
 754
 755 static int __init
 756 rcu_perf_init(void)
 757 {
 758         long i;
 759         int firsterr = 0;
 760         static struct rcu_perf_ops *perf_ops[] = {
 761                 &rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops,
 762         };
 763
 764         if (!torture_init_begin(perf_type, verbose))
 765                 return -EBUSY;
 766
 767         /* Process args and tell the world that the perf'er is on the job. */
 768         for (i = 0; i < ARRAY_SIZE(perf_ops); i++) {
 769                 cur_ops = perf_ops[i];
 770                 if (strcmp(perf_type, cur_ops->name) == 0)
 771                         break;
 772         }
 773         if (i == ARRAY_SIZE(perf_ops)) {
 774                 pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type);
 775                 pr_alert("rcu-perf types:");
 776                 for (i = 0; i < ARRAY_SIZE(perf_ops); i++)
 777                         pr_cont(" %s", perf_ops[i]->name);
 778                 pr_cont("\n");
 779                 WARN_ON(!IS_MODULE(CONFIG_RCU_PERF_TEST));
 780                 firsterr = -EINVAL;
 781                 cur_ops = NULL;
 782                 goto unwind;
 783         }
 784         if (cur_ops->init)
 785                 cur_ops->init();
 786
 787         if (kfree_rcu_test)
 788                 return kfree_perf_init();
 789
 790         nrealwriters = compute_real(nwriters);
 791         nrealreaders = compute_real(nreaders);
 792         atomic_set(&n_rcu_perf_reader_started, 0);
 793         atomic_set(&n_rcu_perf_writer_started, 0);
 794         atomic_set(&n_rcu_perf_writer_finished, 0);
 795         rcu_perf_print_module_parms(cur_ops, "Start of test");
 796
 797         /* Start up the kthreads. */
 798
 799         if (shutdown) {
 800                 init_waitqueue_head(&shutdown_wq);
 801                 firsterr = torture_create_kthread(rcu_perf_shutdown, NULL,
 802                                                   shutdown_task);
 803                 if (firsterr)
 804                         goto unwind;
 805                 schedule_timeout_uninterruptible(1);
 806         }
 807         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
 808                                GFP_KERNEL);
 809         if (reader_tasks == NULL) {
 810                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 811                 firsterr = -ENOMEM;
 812                 goto unwind;
 813         }
 814         for (i = 0; i < nrealreaders; i++) {
 815                 firsterr = torture_create_kthread(rcu_perf_reader, (void *)i,
 816                                                   reader_tasks[i]);
 817                 if (firsterr)
 818                         goto unwind;
 819         }
 820         while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders)
 821                 schedule_timeout_uninterruptible(1);
 822         writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
 823                                GFP_KERNEL);
 824         writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
 825                                    GFP_KERNEL);
 826         writer_n_durations =
 827                 kcalloc(nrealwriters, sizeof(*writer_n_durations),
 828                         GFP_KERNEL);
 829         if (!writer_tasks || !writer_durations || !writer_n_durations) {
 830                 VERBOSE_PERFOUT_ERRSTRING("out of memory");
 831                 firsterr = -ENOMEM;
 832                 goto unwind;
 833         }
 834         for (i = 0; i < nrealwriters; i++) {
 835                 writer_durations[i] =
 836                         kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
 837                                 GFP_KERNEL);
 838                 if (!writer_durations[i]) {
 839                         firsterr = -ENOMEM;
 840                         goto unwind;
 841                 }
 842                 firsterr = torture_create_kthread(rcu_perf_writer, (void *)i,
 843                                                   writer_tasks[i]);
 844                 if (firsterr)
 845                         goto unwind;
 846         }
 847         torture_init_end();
 848         return 0;
 849
 850 unwind:
 851         torture_init_end();
 852         rcu_perf_cleanup();
 853         return firsterr;
 854 }
 855
 856 module_init(rcu_perf_init);
 857 module_exit(rcu_perf_cleanup);