kernel/rcu/rcuscale.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Read-Copy Update module-based scalability-test facility
   4  *
   5  * Copyright (C) IBM Corporation, 2015
   6  *
   7  * Authors: Paul E. McKenney <[email protected]>
   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 #include <linux/rcupdate_trace.h>
  42
  43 #include "rcu.h"
  44
  45 MODULE_LICENSE("GPL");
  46 MODULE_AUTHOR("Paul E. McKenney <[email protected]>");
  47
  48 #define SCALE_FLAG "-scale:"
  49 #define SCALEOUT_STRING(s) \
  50         pr_alert("%s" SCALE_FLAG " %s\n", scale_type, s)
  51 #define VERBOSE_SCALEOUT_STRING(s) \
  52         do { if (verbose) pr_alert("%s" SCALE_FLAG " %s\n", scale_type, s); } while (0)
  53 #define SCALEOUT_ERRSTRING(s) \
  54         pr_alert("%s" SCALE_FLAG "!!! %s\n", scale_type, s)
  55
  56 /*
  57  * The intended use cases for the nreaders and nwriters module parameters
  58  * are as follows:
  59  *
  60  * 1.   Specify only the nr_cpus kernel boot parameter.  This will
  61  *      set both nreaders and nwriters to the value specified by
  62  *      nr_cpus for a mixed reader/writer test.
  63  *
  64  * 2.   Specify the nr_cpus kernel boot parameter, but set
  65  *      rcuscale.nreaders to zero.  This will set nwriters to the
  66  *      value specified by nr_cpus for an update-only test.
  67  *
  68  * 3.   Specify the nr_cpus kernel boot parameter, but set
  69  *      rcuscale.nwriters to zero.  This will set nreaders to the
  70  *      value specified by nr_cpus for a read-only test.
  71  *
  72  * Various other use cases may of course be specified.
  73  *
  74  * Note that this test's readers are intended only as a test load for
  75  * the writers.  The reader scalability statistics will be overly
  76  * pessimistic due to the per-critical-section interrupt disabling,
  77  * test-end checks, and the pair of calls through pointers.
  78  */
  79
  80 #ifdef MODULE
  81 # define RCUSCALE_SHUTDOWN 0
  82 #else
  83 # define RCUSCALE_SHUTDOWN 1
  84 #endif
  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, RCUSCALE_SHUTDOWN,
  93               "Shutdown at end of scalability 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 torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() scale test?");
  97 torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate.");
  98
  99 static char *scale_type = "rcu";
 100 module_param(scale_type, charp, 0444);
 101 MODULE_PARM_DESC(scale_type, "Type of RCU to scalability-test (rcu, srcu, ...)");
 102
 103 static int nrealreaders;
 104 static int nrealwriters;
 105 static struct task_struct **writer_tasks;
 106 static struct task_struct **reader_tasks;
 107 static struct task_struct *shutdown_task;
 108
 109 static u64 **writer_durations;
 110 static int *writer_n_durations;
 111 static atomic_t n_rcu_scale_reader_started;
 112 static atomic_t n_rcu_scale_writer_started;
 113 static atomic_t n_rcu_scale_writer_finished;
 114 static wait_queue_head_t shutdown_wq;
 115 static u64 t_rcu_scale_writer_started;
 116 static u64 t_rcu_scale_writer_finished;
 117 static unsigned long b_rcu_gp_test_started;
 118 static unsigned long b_rcu_gp_test_finished;
 119 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
 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_scale_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_scale_ops *cur_ops;
 145
 146 /*
 147  * Definitions for rcu scalability testing.
 148  */
 149
 150 static int rcu_scale_read_lock(void) __acquires(RCU)
 151 {
 152         rcu_read_lock();
 153         return 0;
 154 }
 155
 156 static void rcu_scale_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_scale_init(void)
 167 {
 168 }
 169
 170 static struct rcu_scale_ops rcu_ops = {
 171         .ptype          = RCU_FLAVOR,
 172         .init           = rcu_sync_scale_init,
 173         .readlock       = rcu_scale_read_lock,
 174         .readunlock     = rcu_scale_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 scalability testing.
 187  */
 188
 189 DEFINE_STATIC_SRCU(srcu_ctl_scale);
 190 static struct srcu_struct *srcu_ctlp = &srcu_ctl_scale;
 191
 192 static int srcu_scale_read_lock(void) __acquires(srcu_ctlp)
 193 {
 194         return srcu_read_lock(srcu_ctlp);
 195 }
 196
 197 static void srcu_scale_read_unlock(int idx) __releases(srcu_ctlp)
 198 {
 199         srcu_read_unlock(srcu_ctlp, idx);
 200 }
 201
 202 static unsigned long srcu_scale_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_scale_synchronize(void)
 218 {
 219         synchronize_srcu(srcu_ctlp);
 220 }
 221
 222 static void srcu_scale_synchronize_expedited(void)
 223 {
 224         synchronize_srcu_expedited(srcu_ctlp);
 225 }
 226
 227 static struct rcu_scale_ops srcu_ops = {
 228         .ptype          = SRCU_FLAVOR,
 229         .init           = rcu_sync_scale_init,
 230         .readlock       = srcu_scale_read_lock,
 231         .readunlock     = srcu_scale_read_unlock,
 232         .get_gp_seq     = srcu_scale_completed,
 233         .gp_diff        = rcu_seq_diff,
 234         .exp_completed  = srcu_scale_completed,
 235         .async          = srcu_call_rcu,
 236         .gp_barrier     = srcu_rcu_barrier,
 237         .sync           = srcu_scale_synchronize,
 238         .exp_sync       = srcu_scale_synchronize_expedited,
 239         .name           = "srcu"
 240 };
 241
 242 static struct srcu_struct srcud;
 243
 244 static void srcu_sync_scale_init(void)
 245 {
 246         srcu_ctlp = &srcud;
 247         init_srcu_struct(srcu_ctlp);
 248 }
 249
 250 static void srcu_sync_scale_cleanup(void)
 251 {
 252         cleanup_srcu_struct(srcu_ctlp);
 253 }
 254
 255 static struct rcu_scale_ops srcud_ops = {
 256         .ptype          = SRCU_FLAVOR,
 257         .init           = srcu_sync_scale_init,
 258         .cleanup        = srcu_sync_scale_cleanup,
 259         .readlock       = srcu_scale_read_lock,
 260         .readunlock     = srcu_scale_read_unlock,
 261         .get_gp_seq     = srcu_scale_completed,
 262         .gp_diff        = rcu_seq_diff,
 263         .exp_completed  = srcu_scale_completed,
 264         .async          = srcu_call_rcu,
 265         .gp_barrier     = srcu_rcu_barrier,
 266         .sync           = srcu_scale_synchronize,
 267         .exp_sync       = srcu_scale_synchronize_expedited,
 268         .name           = "srcud"
 269 };
 270
 271 #ifdef CONFIG_TASKS_RCU
 272
 273 /*
 274  * Definitions for RCU-tasks scalability testing.
 275  */
 276
 277 static int tasks_scale_read_lock(void)
 278 {
 279         return 0;
 280 }
 281
 282 static void tasks_scale_read_unlock(int idx)
 283 {
 284 }
 285
 286 static struct rcu_scale_ops tasks_ops = {
 287         .ptype          = RCU_TASKS_FLAVOR,
 288         .init           = rcu_sync_scale_init,
 289         .readlock       = tasks_scale_read_lock,
 290         .readunlock     = tasks_scale_read_unlock,
 291         .get_gp_seq     = rcu_no_completed,
 292         .gp_diff        = rcu_seq_diff,
 293         .async          = call_rcu_tasks,
 294         .gp_barrier     = rcu_barrier_tasks,
 295         .sync           = synchronize_rcu_tasks,
 296         .exp_sync       = synchronize_rcu_tasks,
 297         .name           = "tasks"
 298 };
 299
 300 #define TASKS_OPS &tasks_ops,
 301
 302 #else // #ifdef CONFIG_TASKS_RCU
 303
 304 #define TASKS_OPS
 305
 306 #endif // #else // #ifdef CONFIG_TASKS_RCU
 307
 308 #ifdef CONFIG_TASKS_TRACE_RCU
 309
 310 /*
 311  * Definitions for RCU-tasks-trace scalability testing.
 312  */
 313
 314 static int tasks_trace_scale_read_lock(void)
 315 {
 316         rcu_read_lock_trace();
 317         return 0;
 318 }
 319
 320 static void tasks_trace_scale_read_unlock(int idx)
 321 {
 322         rcu_read_unlock_trace();
 323 }
 324
 325 static struct rcu_scale_ops tasks_tracing_ops = {
 326         .ptype          = RCU_TASKS_FLAVOR,
 327         .init           = rcu_sync_scale_init,
 328         .readlock       = tasks_trace_scale_read_lock,
 329         .readunlock     = tasks_trace_scale_read_unlock,
 330         .get_gp_seq     = rcu_no_completed,
 331         .gp_diff        = rcu_seq_diff,
 332         .async          = call_rcu_tasks_trace,
 333         .gp_barrier     = rcu_barrier_tasks_trace,
 334         .sync           = synchronize_rcu_tasks_trace,
 335         .exp_sync       = synchronize_rcu_tasks_trace,
 336         .name           = "tasks-tracing"
 337 };
 338
 339 #define TASKS_TRACING_OPS &tasks_tracing_ops,
 340
 341 #else // #ifdef CONFIG_TASKS_TRACE_RCU
 342
 343 #define TASKS_TRACING_OPS
 344
 345 #endif // #else // #ifdef CONFIG_TASKS_TRACE_RCU
 346
 347 static unsigned long rcuscale_seq_diff(unsigned long new, unsigned long old)
 348 {
 349         if (!cur_ops->gp_diff)
 350                 return new - old;
 351         return cur_ops->gp_diff(new, old);
 352 }
 353
 354 /*
 355  * If scalability tests complete, wait for shutdown to commence.
 356  */
 357 static void rcu_scale_wait_shutdown(void)
 358 {
 359         cond_resched_tasks_rcu_qs();
 360         if (atomic_read(&n_rcu_scale_writer_finished) < nrealwriters)
 361                 return;
 362         while (!torture_must_stop())
 363                 schedule_timeout_uninterruptible(1);
 364 }
 365
 366 /*
 367  * RCU scalability reader kthread.  Repeatedly does empty RCU read-side
 368  * critical section, minimizing update-side interference.  However, the
 369  * point of this test is not to evaluate reader scalability, but instead
 370  * to serve as a test load for update-side scalability testing.
 371  */
 372 static int
 373 rcu_scale_reader(void *arg)
 374 {
 375         unsigned long flags;
 376         int idx;
 377         long me = (long)arg;
 378
 379         VERBOSE_SCALEOUT_STRING("rcu_scale_reader task started");
 380         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 381         set_user_nice(current, MAX_NICE);
 382         atomic_inc(&n_rcu_scale_reader_started);
 383
 384         do {
 385                 local_irq_save(flags);
 386                 idx = cur_ops->readlock();
 387                 cur_ops->readunlock(idx);
 388                 local_irq_restore(flags);
 389                 rcu_scale_wait_shutdown();
 390         } while (!torture_must_stop());
 391         torture_kthread_stopping("rcu_scale_reader");
 392         return 0;
 393 }
 394
 395 /*
 396  * Callback function for asynchronous grace periods from rcu_scale_writer().
 397  */
 398 static void rcu_scale_async_cb(struct rcu_head *rhp)
 399 {
 400         atomic_dec(this_cpu_ptr(&n_async_inflight));
 401         kfree(rhp);
 402 }
 403
 404 /*
 405  * RCU scale writer kthread.  Repeatedly does a grace period.
 406  */
 407 static int
 408 rcu_scale_writer(void *arg)
 409 {
 410         int i = 0;
 411         int i_max;
 412         long me = (long)arg;
 413         struct rcu_head *rhp = NULL;
 414         bool started = false, done = false, alldone = false;
 415         u64 t;
 416         u64 *wdp;
 417         u64 *wdpp = writer_durations[me];
 418
 419         VERBOSE_SCALEOUT_STRING("rcu_scale_writer task started");
 420         WARN_ON(!wdpp);
 421         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 422         current->flags |= PF_NO_SETAFFINITY;
 423         sched_set_fifo_low(current);
 424
 425         if (holdoff)
 426                 schedule_timeout_uninterruptible(holdoff * HZ);
 427
 428         /*
 429          * Wait until rcu_end_inkernel_boot() is called for normal GP tests
 430          * so that RCU is not always expedited for normal GP tests.
 431          * The system_state test is approximate, but works well in practice.
 432          */
 433         while (!gp_exp && system_state != SYSTEM_RUNNING)
 434                 schedule_timeout_uninterruptible(1);
 435
 436         t = ktime_get_mono_fast_ns();
 437         if (atomic_inc_return(&n_rcu_scale_writer_started) >= nrealwriters) {
 438                 t_rcu_scale_writer_started = t;
 439                 if (gp_exp) {
 440                         b_rcu_gp_test_started =
 441                                 cur_ops->exp_completed() / 2;
 442                 } else {
 443                         b_rcu_gp_test_started = cur_ops->get_gp_seq();
 444                 }
 445         }
 446
 447         do {
 448                 if (writer_holdoff)
 449                         udelay(writer_holdoff);
 450                 wdp = &wdpp[i];
 451                 *wdp = ktime_get_mono_fast_ns();
 452                 if (gp_async) {
 453 retry:
 454                         if (!rhp)
 455                                 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
 456                         if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
 457                                 atomic_inc(this_cpu_ptr(&n_async_inflight));
 458                                 cur_ops->async(rhp, rcu_scale_async_cb);
 459                                 rhp = NULL;
 460                         } else if (!kthread_should_stop()) {
 461                                 cur_ops->gp_barrier();
 462                                 goto retry;
 463                         } else {
 464                                 kfree(rhp); /* Because we are stopping. */
 465                         }
 466                 } else if (gp_exp) {
 467                         cur_ops->exp_sync();
 468                 } else {
 469                         cur_ops->sync();
 470                 }
 471                 t = ktime_get_mono_fast_ns();
 472                 *wdp = t - *wdp;
 473                 i_max = i;
 474                 if (!started &&
 475                     atomic_read(&n_rcu_scale_writer_started) >= nrealwriters)
 476                         started = true;
 477                 if (!done && i >= MIN_MEAS) {
 478                         done = true;
 479                         sched_set_normal(current, 0);
 480                         pr_alert("%s%s rcu_scale_writer %ld has %d measurements\n",
 481                                  scale_type, SCALE_FLAG, me, MIN_MEAS);
 482                         if (atomic_inc_return(&n_rcu_scale_writer_finished) >=
 483                             nrealwriters) {
 484                                 schedule_timeout_interruptible(10);
 485                                 rcu_ftrace_dump(DUMP_ALL);
 486                                 SCALEOUT_STRING("Test complete");
 487                                 t_rcu_scale_writer_finished = t;
 488                                 if (gp_exp) {
 489                                         b_rcu_gp_test_finished =
 490                                                 cur_ops->exp_completed() / 2;
 491                                 } else {
 492                                         b_rcu_gp_test_finished =
 493                                                 cur_ops->get_gp_seq();
 494                                 }
 495                                 if (shutdown) {
 496                                         smp_mb(); /* Assign before wake. */
 497                                         wake_up(&shutdown_wq);
 498                                 }
 499                         }
 500                 }
 501                 if (done && !alldone &&
 502                     atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters)
 503                         alldone = true;
 504                 if (started && !alldone && i < MAX_MEAS - 1)
 505                         i++;
 506                 rcu_scale_wait_shutdown();
 507         } while (!torture_must_stop());
 508         if (gp_async) {
 509                 cur_ops->gp_barrier();
 510         }
 511         writer_n_durations[me] = i_max + 1;
 512         torture_kthread_stopping("rcu_scale_writer");
 513         return 0;
 514 }
 515
 516 static void
 517 rcu_scale_print_module_parms(struct rcu_scale_ops *cur_ops, const char *tag)
 518 {
 519         pr_alert("%s" SCALE_FLAG
 520                  "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n",
 521                  scale_type, tag, nrealreaders, nrealwriters, verbose, shutdown);
 522 }
 523
 524 static void
 525 rcu_scale_cleanup(void)
 526 {
 527         int i;
 528         int j;
 529         int ngps = 0;
 530         u64 *wdp;
 531         u64 *wdpp;
 532
 533         /*
 534          * Would like warning at start, but everything is expedited
 535          * during the mid-boot phase, so have to wait till the end.
 536          */
 537         if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
 538                 SCALEOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
 539         if (rcu_gp_is_normal() && gp_exp)
 540                 SCALEOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
 541         if (gp_exp && gp_async)
 542                 SCALEOUT_ERRSTRING("No expedited async GPs, so went with async!");
 543
 544         if (torture_cleanup_begin())
 545                 return;
 546         if (!cur_ops) {
 547                 torture_cleanup_end();
 548                 return;
 549         }
 550
 551         if (reader_tasks) {
 552                 for (i = 0; i < nrealreaders; i++)
 553                         torture_stop_kthread(rcu_scale_reader,
 554                                              reader_tasks[i]);
 555                 kfree(reader_tasks);
 556         }
 557
 558         if (writer_tasks) {
 559                 for (i = 0; i < nrealwriters; i++) {
 560                         torture_stop_kthread(rcu_scale_writer,
 561                                              writer_tasks[i]);
 562                         if (!writer_n_durations)
 563                                 continue;
 564                         j = writer_n_durations[i];
 565                         pr_alert("%s%s writer %d gps: %d\n",
 566                                  scale_type, SCALE_FLAG, i, j);
 567                         ngps += j;
 568                 }
 569                 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
 570                          scale_type, SCALE_FLAG,
 571                          t_rcu_scale_writer_started, t_rcu_scale_writer_finished,
 572                          t_rcu_scale_writer_finished -
 573                          t_rcu_scale_writer_started,
 574                          ngps,
 575                          rcuscale_seq_diff(b_rcu_gp_test_finished,
 576                                            b_rcu_gp_test_started));
 577                 for (i = 0; i < nrealwriters; i++) {
 578                         if (!writer_durations)
 579                                 break;
 580                         if (!writer_n_durations)
 581                                 continue;
 582                         wdpp = writer_durations[i];
 583                         if (!wdpp)
 584                                 continue;
 585                         for (j = 0; j < writer_n_durations[i]; j++) {
 586                                 wdp = &wdpp[j];
 587                                 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
 588                                         scale_type, SCALE_FLAG,
 589                                         i, j, *wdp);
 590                                 if (j % 100 == 0)
 591                                         schedule_timeout_uninterruptible(1);
 592                         }
 593                         kfree(writer_durations[i]);
 594                 }
 595                 kfree(writer_tasks);
 596                 kfree(writer_durations);
 597                 kfree(writer_n_durations);
 598         }
 599
 600         /* Do torture-type-specific cleanup operations.  */
 601         if (cur_ops->cleanup != NULL)
 602                 cur_ops->cleanup();
 603
 604         torture_cleanup_end();
 605 }
 606
 607 /*
 608  * Return the number if non-negative.  If -1, the number of CPUs.
 609  * If less than -1, that much less than the number of CPUs, but
 610  * at least one.
 611  */
 612 static int compute_real(int n)
 613 {
 614         int nr;
 615
 616         if (n >= 0) {
 617                 nr = n;
 618         } else {
 619                 nr = num_online_cpus() + 1 + n;
 620                 if (nr <= 0)
 621                         nr = 1;
 622         }
 623         return nr;
 624 }
 625
 626 /*
 627  * RCU scalability shutdown kthread.  Just waits to be awakened, then shuts
 628  * down system.
 629  */
 630 static int
 631 rcu_scale_shutdown(void *arg)
 632 {
 633         wait_event(shutdown_wq,
 634                    atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters);
 635         smp_mb(); /* Wake before output. */
 636         rcu_scale_cleanup();
 637         kernel_power_off();
 638         return -EINVAL;
 639 }
 640
 641 /*
 642  * kfree_rcu() scalability tests: Start a kfree_rcu() loop on all CPUs for number
 643  * of iterations and measure total time and number of GP for all iterations to complete.
 644  */
 645
 646 torture_param(int, kfree_nthreads, -1, "Number of threads running loops of kfree_rcu().");
 647 torture_param(int, kfree_alloc_num, 8000, "Number of allocations and frees done in an iteration.");
 648 torture_param(int, kfree_loops, 10, "Number of loops doing kfree_alloc_num allocations and frees.");
 649 torture_param(bool, kfree_rcu_test_double, false, "Do we run a kfree_rcu() double-argument scale test?");
 650 torture_param(bool, kfree_rcu_test_single, false, "Do we run a kfree_rcu() single-argument scale test?");
 651
 652 static struct task_struct **kfree_reader_tasks;
 653 static int kfree_nrealthreads;
 654 static atomic_t n_kfree_scale_thread_started;
 655 static atomic_t n_kfree_scale_thread_ended;
 656
 657 struct kfree_obj {
 658         char kfree_obj[8];
 659         struct rcu_head rh;
 660 };
 661
 662 static int
 663 kfree_scale_thread(void *arg)
 664 {
 665         int i, loop = 0;
 666         long me = (long)arg;
 667         struct kfree_obj *alloc_ptr;
 668         u64 start_time, end_time;
 669         long long mem_begin, mem_during = 0;
 670         bool kfree_rcu_test_both;
 671         DEFINE_TORTURE_RANDOM(tr);
 672
 673         VERBOSE_SCALEOUT_STRING("kfree_scale_thread task started");
 674         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 675         set_user_nice(current, MAX_NICE);
 676         kfree_rcu_test_both = (kfree_rcu_test_single == kfree_rcu_test_double);
 677
 678         start_time = ktime_get_mono_fast_ns();
 679
 680         if (atomic_inc_return(&n_kfree_scale_thread_started) >= kfree_nrealthreads) {
 681                 if (gp_exp)
 682                         b_rcu_gp_test_started = cur_ops->exp_completed() / 2;
 683                 else
 684                         b_rcu_gp_test_started = cur_ops->get_gp_seq();
 685         }
 686
 687         do {
 688                 if (!mem_during) {
 689                         mem_during = mem_begin = si_mem_available();
 690                 } else if (loop % (kfree_loops / 4) == 0) {
 691                         mem_during = (mem_during + si_mem_available()) / 2;
 692                 }
 693
 694                 for (i = 0; i < kfree_alloc_num; i++) {
 695                         alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL);
 696                         if (!alloc_ptr)
 697                                 return -ENOMEM;
 698
 699                         // By default kfree_rcu_test_single and kfree_rcu_test_double are
 700                         // initialized to false. If both have the same value (false or true)
 701                         // both are randomly tested, otherwise only the one with value true
 702                         // is tested.
 703                         if ((kfree_rcu_test_single && !kfree_rcu_test_double) ||
 704                                         (kfree_rcu_test_both && torture_random(&tr) & 0x800))
 705                                 kfree_rcu(alloc_ptr);
 706                         else
 707                                 kfree_rcu(alloc_ptr, rh);
 708                 }
 709
 710                 cond_resched();
 711         } while (!torture_must_stop() && ++loop < kfree_loops);
 712
 713         if (atomic_inc_return(&n_kfree_scale_thread_ended) >= kfree_nrealthreads) {
 714                 end_time = ktime_get_mono_fast_ns();
 715
 716                 if (gp_exp)
 717                         b_rcu_gp_test_finished = cur_ops->exp_completed() / 2;
 718                 else
 719                         b_rcu_gp_test_finished = cur_ops->get_gp_seq();
 720
 721                 pr_alert("Total time taken by all kfree'ers: %llu ns, loops: %d, batches: %ld, memory footprint: %lldMB\n",
 722                        (unsigned long long)(end_time - start_time), kfree_loops,
 723                        rcuscale_seq_diff(b_rcu_gp_test_finished, b_rcu_gp_test_started),
 724                        (mem_begin - mem_during) >> (20 - PAGE_SHIFT));
 725
 726                 if (shutdown) {
 727                         smp_mb(); /* Assign before wake. */
 728                         wake_up(&shutdown_wq);
 729                 }
 730         }
 731
 732         torture_kthread_stopping("kfree_scale_thread");
 733         return 0;
 734 }
 735
 736 static void
 737 kfree_scale_cleanup(void)
 738 {
 739         int i;
 740
 741         if (torture_cleanup_begin())
 742                 return;
 743
 744         if (kfree_reader_tasks) {
 745                 for (i = 0; i < kfree_nrealthreads; i++)
 746                         torture_stop_kthread(kfree_scale_thread,
 747                                              kfree_reader_tasks[i]);
 748                 kfree(kfree_reader_tasks);
 749         }
 750
 751         torture_cleanup_end();
 752 }
 753
 754 /*
 755  * shutdown kthread.  Just waits to be awakened, then shuts down system.
 756  */
 757 static int
 758 kfree_scale_shutdown(void *arg)
 759 {
 760         wait_event(shutdown_wq,
 761                    atomic_read(&n_kfree_scale_thread_ended) >= kfree_nrealthreads);
 762
 763         smp_mb(); /* Wake before output. */
 764
 765         kfree_scale_cleanup();
 766         kernel_power_off();
 767         return -EINVAL;
 768 }
 769
 770 static int __init
 771 kfree_scale_init(void)
 772 {
 773         long i;
 774         int firsterr = 0;
 775
 776         kfree_nrealthreads = compute_real(kfree_nthreads);
 777         /* Start up the kthreads. */
 778         if (shutdown) {
 779                 init_waitqueue_head(&shutdown_wq);
 780                 firsterr = torture_create_kthread(kfree_scale_shutdown, NULL,
 781                                                   shutdown_task);
 782                 if (torture_init_error(firsterr))
 783                         goto unwind;
 784                 schedule_timeout_uninterruptible(1);
 785         }
 786
 787         pr_alert("kfree object size=%zu\n", kfree_mult * sizeof(struct kfree_obj));
 788
 789         kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]),
 790                                GFP_KERNEL);
 791         if (kfree_reader_tasks == NULL) {
 792                 firsterr = -ENOMEM;
 793                 goto unwind;
 794         }
 795
 796         for (i = 0; i < kfree_nrealthreads; i++) {
 797                 firsterr = torture_create_kthread(kfree_scale_thread, (void *)i,
 798                                                   kfree_reader_tasks[i]);
 799                 if (torture_init_error(firsterr))
 800                         goto unwind;
 801         }
 802
 803         while (atomic_read(&n_kfree_scale_thread_started) < kfree_nrealthreads)
 804                 schedule_timeout_uninterruptible(1);
 805
 806         torture_init_end();
 807         return 0;
 808
 809 unwind:
 810         torture_init_end();
 811         kfree_scale_cleanup();
 812         return firsterr;
 813 }
 814
 815 static int __init
 816 rcu_scale_init(void)
 817 {
 818         long i;
 819         int firsterr = 0;
 820         static struct rcu_scale_ops *scale_ops[] = {
 821                 &rcu_ops, &srcu_ops, &srcud_ops, TASKS_OPS TASKS_TRACING_OPS
 822         };
 823
 824         if (!torture_init_begin(scale_type, verbose))
 825                 return -EBUSY;
 826
 827         /* Process args and announce that the scalability'er is on the job. */
 828         for (i = 0; i < ARRAY_SIZE(scale_ops); i++) {
 829                 cur_ops = scale_ops[i];
 830                 if (strcmp(scale_type, cur_ops->name) == 0)
 831                         break;
 832         }
 833         if (i == ARRAY_SIZE(scale_ops)) {
 834                 pr_alert("rcu-scale: invalid scale type: \"%s\"\n", scale_type);
 835                 pr_alert("rcu-scale types:");
 836                 for (i = 0; i < ARRAY_SIZE(scale_ops); i++)
 837                         pr_cont(" %s", scale_ops[i]->name);
 838                 pr_cont("\n");
 839                 firsterr = -EINVAL;
 840                 cur_ops = NULL;
 841                 goto unwind;
 842         }
 843         if (cur_ops->init)
 844                 cur_ops->init();
 845
 846         if (kfree_rcu_test)
 847                 return kfree_scale_init();
 848
 849         nrealwriters = compute_real(nwriters);
 850         nrealreaders = compute_real(nreaders);
 851         atomic_set(&n_rcu_scale_reader_started, 0);
 852         atomic_set(&n_rcu_scale_writer_started, 0);
 853         atomic_set(&n_rcu_scale_writer_finished, 0);
 854         rcu_scale_print_module_parms(cur_ops, "Start of test");
 855
 856         /* Start up the kthreads. */
 857
 858         if (shutdown) {
 859                 init_waitqueue_head(&shutdown_wq);
 860                 firsterr = torture_create_kthread(rcu_scale_shutdown, NULL,
 861                                                   shutdown_task);
 862                 if (torture_init_error(firsterr))
 863                         goto unwind;
 864                 schedule_timeout_uninterruptible(1);
 865         }
 866         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
 867                                GFP_KERNEL);
 868         if (reader_tasks == NULL) {
 869                 SCALEOUT_ERRSTRING("out of memory");
 870                 firsterr = -ENOMEM;
 871                 goto unwind;
 872         }
 873         for (i = 0; i < nrealreaders; i++) {
 874                 firsterr = torture_create_kthread(rcu_scale_reader, (void *)i,
 875                                                   reader_tasks[i]);
 876                 if (torture_init_error(firsterr))
 877                         goto unwind;
 878         }
 879         while (atomic_read(&n_rcu_scale_reader_started) < nrealreaders)
 880                 schedule_timeout_uninterruptible(1);
 881         writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
 882                                GFP_KERNEL);
 883         writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
 884                                    GFP_KERNEL);
 885         writer_n_durations =
 886                 kcalloc(nrealwriters, sizeof(*writer_n_durations),
 887                         GFP_KERNEL);
 888         if (!writer_tasks || !writer_durations || !writer_n_durations) {
 889                 SCALEOUT_ERRSTRING("out of memory");
 890                 firsterr = -ENOMEM;
 891                 goto unwind;
 892         }
 893         for (i = 0; i < nrealwriters; i++) {
 894                 writer_durations[i] =
 895                         kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
 896                                 GFP_KERNEL);
 897                 if (!writer_durations[i]) {
 898                         firsterr = -ENOMEM;
 899                         goto unwind;
 900                 }
 901                 firsterr = torture_create_kthread(rcu_scale_writer, (void *)i,
 902                                                   writer_tasks[i]);
 903                 if (torture_init_error(firsterr))
 904                         goto unwind;
 905         }
 906         torture_init_end();
 907         return 0;
 908
 909 unwind:
 910         torture_init_end();
 911         rcu_scale_cleanup();
 912         if (shutdown) {
 913                 WARN_ON(!IS_MODULE(CONFIG_RCU_SCALE_TEST));
 914                 kernel_power_off();
 915         }
 916         return firsterr;
 917 }
 918
 919 module_init(rcu_scale_init);
 920 module_exit(rcu_scale_cleanup);