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 <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 #include <linux/rcupdate_trace.h>
  42
  43 #include "rcu.h"
  44
  45 MODULE_LICENSE("GPL");
  46 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
  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 writer");
  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, minruntime, 0, "Minimum run time (s)");
  91 torture_param(int, nreaders, -1, "Number of RCU reader threads");
  92 torture_param(int, nwriters, -1, "Number of RCU updater threads");
  93 torture_param(bool, shutdown, RCUSCALE_SHUTDOWN,
  94               "Shutdown at end of scalability tests.");
  95 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
  96 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
  97 torture_param(int, writer_holdoff_jiffies, 0, "Holdoff (jiffies) between GPs, zero to disable");
  98 torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() scale test?");
  99 torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate.");
 100 torture_param(int, kfree_by_call_rcu, 0, "Use call_rcu() to emulate kfree_rcu()?");
 101
 102 static char *scale_type = "rcu";
 103 module_param(scale_type, charp, 0444);
 104 MODULE_PARM_DESC(scale_type, "Type of RCU to scalability-test (rcu, srcu, ...)");
 105
 106 static int nrealreaders;
 107 static int nrealwriters;
 108 static struct task_struct **writer_tasks;
 109 static struct task_struct **reader_tasks;
 110 static struct task_struct *shutdown_task;
 111
 112 static u64 **writer_durations;
 113 static int *writer_n_durations;
 114 static atomic_t n_rcu_scale_reader_started;
 115 static atomic_t n_rcu_scale_writer_started;
 116 static atomic_t n_rcu_scale_writer_finished;
 117 static wait_queue_head_t shutdown_wq;
 118 static u64 t_rcu_scale_writer_started;
 119 static u64 t_rcu_scale_writer_finished;
 120 static unsigned long b_rcu_gp_test_started;
 121 static unsigned long b_rcu_gp_test_finished;
 122 static DEFINE_PER_CPU(atomic_t, n_async_inflight);
 123
 124 #define MAX_MEAS 10000
 125 #define MIN_MEAS 100
 126
 127 /*
 128  * Operations vector for selecting different types of tests.
 129  */
 130
 131 struct rcu_scale_ops {
 132         int ptype;
 133         void (*init)(void);
 134         void (*cleanup)(void);
 135         int (*readlock)(void);
 136         void (*readunlock)(int idx);
 137         unsigned long (*get_gp_seq)(void);
 138         unsigned long (*gp_diff)(unsigned long new, unsigned long old);
 139         unsigned long (*exp_completed)(void);
 140         void (*async)(struct rcu_head *head, rcu_callback_t func);
 141         void (*gp_barrier)(void);
 142         void (*sync)(void);
 143         void (*exp_sync)(void);
 144         struct task_struct *(*rso_gp_kthread)(void);
 145         const char *name;
 146 };
 147
 148 static struct rcu_scale_ops *cur_ops;
 149
 150 /*
 151  * Definitions for rcu scalability testing.
 152  */
 153
 154 static int rcu_scale_read_lock(void) __acquires(RCU)
 155 {
 156         rcu_read_lock();
 157         return 0;
 158 }
 159
 160 static void rcu_scale_read_unlock(int idx) __releases(RCU)
 161 {
 162         rcu_read_unlock();
 163 }
 164
 165 static unsigned long __maybe_unused rcu_no_completed(void)
 166 {
 167         return 0;
 168 }
 169
 170 static void rcu_sync_scale_init(void)
 171 {
 172 }
 173
 174 static struct rcu_scale_ops rcu_ops = {
 175         .ptype          = RCU_FLAVOR,
 176         .init           = rcu_sync_scale_init,
 177         .readlock       = rcu_scale_read_lock,
 178         .readunlock     = rcu_scale_read_unlock,
 179         .get_gp_seq     = rcu_get_gp_seq,
 180         .gp_diff        = rcu_seq_diff,
 181         .exp_completed  = rcu_exp_batches_completed,
 182         .async          = call_rcu_hurry,
 183         .gp_barrier     = rcu_barrier,
 184         .sync           = synchronize_rcu,
 185         .exp_sync       = synchronize_rcu_expedited,
 186         .name           = "rcu"
 187 };
 188
 189 /*
 190  * Definitions for srcu scalability testing.
 191  */
 192
 193 DEFINE_STATIC_SRCU(srcu_ctl_scale);
 194 static struct srcu_struct *srcu_ctlp = &srcu_ctl_scale;
 195
 196 static int srcu_scale_read_lock(void) __acquires(srcu_ctlp)
 197 {
 198         return srcu_read_lock(srcu_ctlp);
 199 }
 200
 201 static void srcu_scale_read_unlock(int idx) __releases(srcu_ctlp)
 202 {
 203         srcu_read_unlock(srcu_ctlp, idx);
 204 }
 205
 206 static unsigned long srcu_scale_completed(void)
 207 {
 208         return srcu_batches_completed(srcu_ctlp);
 209 }
 210
 211 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
 212 {
 213         call_srcu(srcu_ctlp, head, func);
 214 }
 215
 216 static void srcu_rcu_barrier(void)
 217 {
 218         srcu_barrier(srcu_ctlp);
 219 }
 220
 221 static void srcu_scale_synchronize(void)
 222 {
 223         synchronize_srcu(srcu_ctlp);
 224 }
 225
 226 static void srcu_scale_synchronize_expedited(void)
 227 {
 228         synchronize_srcu_expedited(srcu_ctlp);
 229 }
 230
 231 static struct rcu_scale_ops srcu_ops = {
 232         .ptype          = SRCU_FLAVOR,
 233         .init           = rcu_sync_scale_init,
 234         .readlock       = srcu_scale_read_lock,
 235         .readunlock     = srcu_scale_read_unlock,
 236         .get_gp_seq     = srcu_scale_completed,
 237         .gp_diff        = rcu_seq_diff,
 238         .exp_completed  = srcu_scale_completed,
 239         .async          = srcu_call_rcu,
 240         .gp_barrier     = srcu_rcu_barrier,
 241         .sync           = srcu_scale_synchronize,
 242         .exp_sync       = srcu_scale_synchronize_expedited,
 243         .name           = "srcu"
 244 };
 245
 246 static struct srcu_struct srcud;
 247
 248 static void srcu_sync_scale_init(void)
 249 {
 250         srcu_ctlp = &srcud;
 251         init_srcu_struct(srcu_ctlp);
 252 }
 253
 254 static void srcu_sync_scale_cleanup(void)
 255 {
 256         cleanup_srcu_struct(srcu_ctlp);
 257 }
 258
 259 static struct rcu_scale_ops srcud_ops = {
 260         .ptype          = SRCU_FLAVOR,
 261         .init           = srcu_sync_scale_init,
 262         .cleanup        = srcu_sync_scale_cleanup,
 263         .readlock       = srcu_scale_read_lock,
 264         .readunlock     = srcu_scale_read_unlock,
 265         .get_gp_seq     = srcu_scale_completed,
 266         .gp_diff        = rcu_seq_diff,
 267         .exp_completed  = srcu_scale_completed,
 268         .async          = srcu_call_rcu,
 269         .gp_barrier     = srcu_rcu_barrier,
 270         .sync           = srcu_scale_synchronize,
 271         .exp_sync       = srcu_scale_synchronize_expedited,
 272         .name           = "srcud"
 273 };
 274
 275 #ifdef CONFIG_TASKS_RCU
 276
 277 /*
 278  * Definitions for RCU-tasks scalability testing.
 279  */
 280
 281 static int tasks_scale_read_lock(void)
 282 {
 283         return 0;
 284 }
 285
 286 static void tasks_scale_read_unlock(int idx)
 287 {
 288 }
 289
 290 static struct rcu_scale_ops tasks_ops = {
 291         .ptype          = RCU_TASKS_FLAVOR,
 292         .init           = rcu_sync_scale_init,
 293         .readlock       = tasks_scale_read_lock,
 294         .readunlock     = tasks_scale_read_unlock,
 295         .get_gp_seq     = rcu_no_completed,
 296         .gp_diff        = rcu_seq_diff,
 297         .async          = call_rcu_tasks,
 298         .gp_barrier     = rcu_barrier_tasks,
 299         .sync           = synchronize_rcu_tasks,
 300         .exp_sync       = synchronize_rcu_tasks,
 301         .rso_gp_kthread = get_rcu_tasks_gp_kthread,
 302         .name           = "tasks"
 303 };
 304
 305 #define TASKS_OPS &tasks_ops,
 306
 307 #else // #ifdef CONFIG_TASKS_RCU
 308
 309 #define TASKS_OPS
 310
 311 #endif // #else // #ifdef CONFIG_TASKS_RCU
 312
 313 #ifdef CONFIG_TASKS_RUDE_RCU
 314
 315 /*
 316  * Definitions for RCU-tasks-rude scalability testing.
 317  */
 318
 319 static int tasks_rude_scale_read_lock(void)
 320 {
 321         return 0;
 322 }
 323
 324 static void tasks_rude_scale_read_unlock(int idx)
 325 {
 326 }
 327
 328 static struct rcu_scale_ops tasks_rude_ops = {
 329         .ptype          = RCU_TASKS_RUDE_FLAVOR,
 330         .init           = rcu_sync_scale_init,
 331         .readlock       = tasks_rude_scale_read_lock,
 332         .readunlock     = tasks_rude_scale_read_unlock,
 333         .get_gp_seq     = rcu_no_completed,
 334         .gp_diff        = rcu_seq_diff,
 335         .async          = call_rcu_tasks_rude,
 336         .gp_barrier     = rcu_barrier_tasks_rude,
 337         .sync           = synchronize_rcu_tasks_rude,
 338         .exp_sync       = synchronize_rcu_tasks_rude,
 339         .rso_gp_kthread = get_rcu_tasks_rude_gp_kthread,
 340         .name           = "tasks-rude"
 341 };
 342
 343 #define TASKS_RUDE_OPS &tasks_rude_ops,
 344
 345 #else // #ifdef CONFIG_TASKS_RUDE_RCU
 346
 347 #define TASKS_RUDE_OPS
 348
 349 #endif // #else // #ifdef CONFIG_TASKS_RUDE_RCU
 350
 351 #ifdef CONFIG_TASKS_TRACE_RCU
 352
 353 /*
 354  * Definitions for RCU-tasks-trace scalability testing.
 355  */
 356
 357 static int tasks_trace_scale_read_lock(void)
 358 {
 359         rcu_read_lock_trace();
 360         return 0;
 361 }
 362
 363 static void tasks_trace_scale_read_unlock(int idx)
 364 {
 365         rcu_read_unlock_trace();
 366 }
 367
 368 static struct rcu_scale_ops tasks_tracing_ops = {
 369         .ptype          = RCU_TASKS_FLAVOR,
 370         .init           = rcu_sync_scale_init,
 371         .readlock       = tasks_trace_scale_read_lock,
 372         .readunlock     = tasks_trace_scale_read_unlock,
 373         .get_gp_seq     = rcu_no_completed,
 374         .gp_diff        = rcu_seq_diff,
 375         .async          = call_rcu_tasks_trace,
 376         .gp_barrier     = rcu_barrier_tasks_trace,
 377         .sync           = synchronize_rcu_tasks_trace,
 378         .exp_sync       = synchronize_rcu_tasks_trace,
 379         .rso_gp_kthread = get_rcu_tasks_trace_gp_kthread,
 380         .name           = "tasks-tracing"
 381 };
 382
 383 #define TASKS_TRACING_OPS &tasks_tracing_ops,
 384
 385 #else // #ifdef CONFIG_TASKS_TRACE_RCU
 386
 387 #define TASKS_TRACING_OPS
 388
 389 #endif // #else // #ifdef CONFIG_TASKS_TRACE_RCU
 390
 391 static unsigned long rcuscale_seq_diff(unsigned long new, unsigned long old)
 392 {
 393         if (!cur_ops->gp_diff)
 394                 return new - old;
 395         return cur_ops->gp_diff(new, old);
 396 }
 397
 398 /*
 399  * If scalability tests complete, wait for shutdown to commence.
 400  */
 401 static void rcu_scale_wait_shutdown(void)
 402 {
 403         cond_resched_tasks_rcu_qs();
 404         if (atomic_read(&n_rcu_scale_writer_finished) < nrealwriters)
 405                 return;
 406         while (!torture_must_stop())
 407                 schedule_timeout_uninterruptible(1);
 408 }
 409
 410 /*
 411  * RCU scalability reader kthread.  Repeatedly does empty RCU read-side
 412  * critical section, minimizing update-side interference.  However, the
 413  * point of this test is not to evaluate reader scalability, but instead
 414  * to serve as a test load for update-side scalability testing.
 415  */
 416 static int
 417 rcu_scale_reader(void *arg)
 418 {
 419         unsigned long flags;
 420         int idx;
 421         long me = (long)arg;
 422
 423         VERBOSE_SCALEOUT_STRING("rcu_scale_reader task started");
 424         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 425         set_user_nice(current, MAX_NICE);
 426         atomic_inc(&n_rcu_scale_reader_started);
 427
 428         do {
 429                 local_irq_save(flags);
 430                 idx = cur_ops->readlock();
 431                 cur_ops->readunlock(idx);
 432                 local_irq_restore(flags);
 433                 rcu_scale_wait_shutdown();
 434         } while (!torture_must_stop());
 435         torture_kthread_stopping("rcu_scale_reader");
 436         return 0;
 437 }
 438
 439 /*
 440  * Callback function for asynchronous grace periods from rcu_scale_writer().
 441  */
 442 static void rcu_scale_async_cb(struct rcu_head *rhp)
 443 {
 444         atomic_dec(this_cpu_ptr(&n_async_inflight));
 445         kfree(rhp);
 446 }
 447
 448 /*
 449  * RCU scale writer kthread.  Repeatedly does a grace period.
 450  */
 451 static int
 452 rcu_scale_writer(void *arg)
 453 {
 454         int i = 0;
 455         int i_max;
 456         unsigned long jdone;
 457         long me = (long)arg;
 458         struct rcu_head *rhp = NULL;
 459         bool started = false, done = false, alldone = false;
 460         u64 t;
 461         DEFINE_TORTURE_RANDOM(tr);
 462         u64 *wdp;
 463         u64 *wdpp = writer_durations[me];
 464
 465         VERBOSE_SCALEOUT_STRING("rcu_scale_writer task started");
 466         WARN_ON(!wdpp);
 467         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 468         current->flags |= PF_NO_SETAFFINITY;
 469         sched_set_fifo_low(current);
 470
 471         if (holdoff)
 472                 schedule_timeout_idle(holdoff * HZ);
 473
 474         /*
 475          * Wait until rcu_end_inkernel_boot() is called for normal GP tests
 476          * so that RCU is not always expedited for normal GP tests.
 477          * The system_state test is approximate, but works well in practice.
 478          */
 479         while (!gp_exp && system_state != SYSTEM_RUNNING)
 480                 schedule_timeout_uninterruptible(1);
 481
 482         t = ktime_get_mono_fast_ns();
 483         if (atomic_inc_return(&n_rcu_scale_writer_started) >= nrealwriters) {
 484                 t_rcu_scale_writer_started = t;
 485                 if (gp_exp) {
 486                         b_rcu_gp_test_started =
 487                                 cur_ops->exp_completed() / 2;
 488                 } else {
 489                         b_rcu_gp_test_started = cur_ops->get_gp_seq();
 490                 }
 491         }
 492
 493         jdone = jiffies + minruntime * HZ;
 494         do {
 495                 if (writer_holdoff)
 496                         udelay(writer_holdoff);
 497                 if (writer_holdoff_jiffies)
 498                         schedule_timeout_idle(torture_random(&tr) % writer_holdoff_jiffies + 1);
 499                 wdp = &wdpp[i];
 500                 *wdp = ktime_get_mono_fast_ns();
 501                 if (gp_async) {
 502 retry:
 503                         if (!rhp)
 504                                 rhp = kmalloc(sizeof(*rhp), GFP_KERNEL);
 505                         if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) {
 506                                 atomic_inc(this_cpu_ptr(&n_async_inflight));
 507                                 cur_ops->async(rhp, rcu_scale_async_cb);
 508                                 rhp = NULL;
 509                         } else if (!kthread_should_stop()) {
 510                                 cur_ops->gp_barrier();
 511                                 goto retry;
 512                         } else {
 513                                 kfree(rhp); /* Because we are stopping. */
 514                         }
 515                 } else if (gp_exp) {
 516                         cur_ops->exp_sync();
 517                 } else {
 518                         cur_ops->sync();
 519                 }
 520                 t = ktime_get_mono_fast_ns();
 521                 *wdp = t - *wdp;
 522                 i_max = i;
 523                 if (!started &&
 524                     atomic_read(&n_rcu_scale_writer_started) >= nrealwriters)
 525                         started = true;
 526                 if (!done && i >= MIN_MEAS && time_after(jiffies, jdone)) {
 527                         done = true;
 528                         sched_set_normal(current, 0);
 529                         pr_alert("%s%s rcu_scale_writer %ld has %d measurements\n",
 530                                  scale_type, SCALE_FLAG, me, MIN_MEAS);
 531                         if (atomic_inc_return(&n_rcu_scale_writer_finished) >=
 532                             nrealwriters) {
 533                                 schedule_timeout_interruptible(10);
 534                                 rcu_ftrace_dump(DUMP_ALL);
 535                                 SCALEOUT_STRING("Test complete");
 536                                 t_rcu_scale_writer_finished = t;
 537                                 if (gp_exp) {
 538                                         b_rcu_gp_test_finished =
 539                                                 cur_ops->exp_completed() / 2;
 540                                 } else {
 541                                         b_rcu_gp_test_finished =
 542                                                 cur_ops->get_gp_seq();
 543                                 }
 544                                 if (shutdown) {
 545                                         smp_mb(); /* Assign before wake. */
 546                                         wake_up(&shutdown_wq);
 547                                 }
 548                         }
 549                 }
 550                 if (done && !alldone &&
 551                     atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters)
 552                         alldone = true;
 553                 if (started && !alldone && i < MAX_MEAS - 1)
 554                         i++;
 555                 rcu_scale_wait_shutdown();
 556         } while (!torture_must_stop());
 557         if (gp_async) {
 558                 cur_ops->gp_barrier();
 559         }
 560         writer_n_durations[me] = i_max + 1;
 561         torture_kthread_stopping("rcu_scale_writer");
 562         return 0;
 563 }
 564
 565 static void
 566 rcu_scale_print_module_parms(struct rcu_scale_ops *cur_ops, const char *tag)
 567 {
 568         pr_alert("%s" SCALE_FLAG
 569                  "--- %s: gp_async=%d gp_async_max=%d gp_exp=%d holdoff=%d minruntime=%d nreaders=%d nwriters=%d writer_holdoff=%d writer_holdoff_jiffies=%d verbose=%d shutdown=%d\n",
 570                  scale_type, tag, gp_async, gp_async_max, gp_exp, holdoff, minruntime, nrealreaders, nrealwriters, writer_holdoff, writer_holdoff_jiffies, verbose, shutdown);
 571 }
 572
 573 /*
 574  * Return the number if non-negative.  If -1, the number of CPUs.
 575  * If less than -1, that much less than the number of CPUs, but
 576  * at least one.
 577  */
 578 static int compute_real(int n)
 579 {
 580         int nr;
 581
 582         if (n >= 0) {
 583                 nr = n;
 584         } else {
 585                 nr = num_online_cpus() + 1 + n;
 586                 if (nr <= 0)
 587                         nr = 1;
 588         }
 589         return nr;
 590 }
 591
 592 /*
 593  * kfree_rcu() scalability tests: Start a kfree_rcu() loop on all CPUs for number
 594  * of iterations and measure total time and number of GP for all iterations to complete.
 595  */
 596
 597 torture_param(int, kfree_nthreads, -1, "Number of threads running loops of kfree_rcu().");
 598 torture_param(int, kfree_alloc_num, 8000, "Number of allocations and frees done in an iteration.");
 599 torture_param(int, kfree_loops, 10, "Number of loops doing kfree_alloc_num allocations and frees.");
 600 torture_param(bool, kfree_rcu_test_double, false, "Do we run a kfree_rcu() double-argument scale test?");
 601 torture_param(bool, kfree_rcu_test_single, false, "Do we run a kfree_rcu() single-argument scale test?");
 602
 603 static struct task_struct **kfree_reader_tasks;
 604 static int kfree_nrealthreads;
 605 static atomic_t n_kfree_scale_thread_started;
 606 static atomic_t n_kfree_scale_thread_ended;
 607 static struct task_struct *kthread_tp;
 608 static u64 kthread_stime;
 609
 610 struct kfree_obj {
 611         char kfree_obj[8];
 612         struct rcu_head rh;
 613 };
 614
 615 /* Used if doing RCU-kfree'ing via call_rcu(). */
 616 static void kfree_call_rcu(struct rcu_head *rh)
 617 {
 618         struct kfree_obj *obj = container_of(rh, struct kfree_obj, rh);
 619
 620         kfree(obj);
 621 }
 622
 623 static int
 624 kfree_scale_thread(void *arg)
 625 {
 626         int i, loop = 0;
 627         long me = (long)arg;
 628         struct kfree_obj *alloc_ptr;
 629         u64 start_time, end_time;
 630         long long mem_begin, mem_during = 0;
 631         bool kfree_rcu_test_both;
 632         DEFINE_TORTURE_RANDOM(tr);
 633
 634         VERBOSE_SCALEOUT_STRING("kfree_scale_thread task started");
 635         set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
 636         set_user_nice(current, MAX_NICE);
 637         kfree_rcu_test_both = (kfree_rcu_test_single == kfree_rcu_test_double);
 638
 639         start_time = ktime_get_mono_fast_ns();
 640
 641         if (atomic_inc_return(&n_kfree_scale_thread_started) >= kfree_nrealthreads) {
 642                 if (gp_exp)
 643                         b_rcu_gp_test_started = cur_ops->exp_completed() / 2;
 644                 else
 645                         b_rcu_gp_test_started = cur_ops->get_gp_seq();
 646         }
 647
 648         do {
 649                 if (!mem_during) {
 650                         mem_during = mem_begin = si_mem_available();
 651                 } else if (loop % (kfree_loops / 4) == 0) {
 652                         mem_during = (mem_during + si_mem_available()) / 2;
 653                 }
 654
 655                 for (i = 0; i < kfree_alloc_num; i++) {
 656                         alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL);
 657                         if (!alloc_ptr)
 658                                 return -ENOMEM;
 659
 660                         if (kfree_by_call_rcu) {
 661                                 call_rcu(&(alloc_ptr->rh), kfree_call_rcu);
 662                                 continue;
 663                         }
 664
 665                         // By default kfree_rcu_test_single and kfree_rcu_test_double are
 666                         // initialized to false. If both have the same value (false or true)
 667                         // both are randomly tested, otherwise only the one with value true
 668                         // is tested.
 669                         if ((kfree_rcu_test_single && !kfree_rcu_test_double) ||
 670                                         (kfree_rcu_test_both && torture_random(&tr) & 0x800))
 671                                 kfree_rcu_mightsleep(alloc_ptr);
 672                         else
 673                                 kfree_rcu(alloc_ptr, rh);
 674                 }
 675
 676                 cond_resched();
 677         } while (!torture_must_stop() && ++loop < kfree_loops);
 678
 679         if (atomic_inc_return(&n_kfree_scale_thread_ended) >= kfree_nrealthreads) {
 680                 end_time = ktime_get_mono_fast_ns();
 681
 682                 if (gp_exp)
 683                         b_rcu_gp_test_finished = cur_ops->exp_completed() / 2;
 684                 else
 685                         b_rcu_gp_test_finished = cur_ops->get_gp_seq();
 686
 687                 pr_alert("Total time taken by all kfree'ers: %llu ns, loops: %d, batches: %ld, memory footprint: %lldMB\n",
 688                        (unsigned long long)(end_time - start_time), kfree_loops,
 689                        rcuscale_seq_diff(b_rcu_gp_test_finished, b_rcu_gp_test_started),
 690                        (mem_begin - mem_during) >> (20 - PAGE_SHIFT));
 691
 692                 if (shutdown) {
 693                         smp_mb(); /* Assign before wake. */
 694                         wake_up(&shutdown_wq);
 695                 }
 696         }
 697
 698         torture_kthread_stopping("kfree_scale_thread");
 699         return 0;
 700 }
 701
 702 static void
 703 kfree_scale_cleanup(void)
 704 {
 705         int i;
 706
 707         if (torture_cleanup_begin())
 708                 return;
 709
 710         if (kfree_reader_tasks) {
 711                 for (i = 0; i < kfree_nrealthreads; i++)
 712                         torture_stop_kthread(kfree_scale_thread,
 713                                              kfree_reader_tasks[i]);
 714                 kfree(kfree_reader_tasks);
 715         }
 716
 717         torture_cleanup_end();
 718 }
 719
 720 /*
 721  * shutdown kthread.  Just waits to be awakened, then shuts down system.
 722  */
 723 static int
 724 kfree_scale_shutdown(void *arg)
 725 {
 726         wait_event_idle(shutdown_wq,
 727                         atomic_read(&n_kfree_scale_thread_ended) >= kfree_nrealthreads);
 728
 729         smp_mb(); /* Wake before output. */
 730
 731         kfree_scale_cleanup();
 732         kernel_power_off();
 733         return -EINVAL;
 734 }
 735
 736 // Used if doing RCU-kfree'ing via call_rcu().
 737 static unsigned long jiffies_at_lazy_cb;
 738 static struct rcu_head lazy_test1_rh;
 739 static int rcu_lazy_test1_cb_called;
 740 static void call_rcu_lazy_test1(struct rcu_head *rh)
 741 {
 742         jiffies_at_lazy_cb = jiffies;
 743         WRITE_ONCE(rcu_lazy_test1_cb_called, 1);
 744 }
 745
 746 static int __init
 747 kfree_scale_init(void)
 748 {
 749         int firsterr = 0;
 750         long i;
 751         unsigned long jif_start;
 752         unsigned long orig_jif;
 753
 754         pr_alert("%s" SCALE_FLAG
 755                  "--- kfree_rcu_test: kfree_mult=%d kfree_by_call_rcu=%d kfree_nthreads=%d kfree_alloc_num=%d kfree_loops=%d kfree_rcu_test_double=%d kfree_rcu_test_single=%d\n",
 756                  scale_type, kfree_mult, kfree_by_call_rcu, kfree_nthreads, kfree_alloc_num, kfree_loops, kfree_rcu_test_double, kfree_rcu_test_single);
 757
 758         // Also, do a quick self-test to ensure laziness is as much as
 759         // expected.
 760         if (kfree_by_call_rcu && !IS_ENABLED(CONFIG_RCU_LAZY)) {
 761                 pr_alert("CONFIG_RCU_LAZY is disabled, falling back to kfree_rcu() for delayed RCU kfree'ing\n");
 762                 kfree_by_call_rcu = 0;
 763         }
 764
 765         if (kfree_by_call_rcu) {
 766                 /* do a test to check the timeout. */
 767                 orig_jif = rcu_lazy_get_jiffies_till_flush();
 768
 769                 rcu_lazy_set_jiffies_till_flush(2 * HZ);
 770                 rcu_barrier();
 771
 772                 jif_start = jiffies;
 773                 jiffies_at_lazy_cb = 0;
 774                 call_rcu(&lazy_test1_rh, call_rcu_lazy_test1);
 775
 776                 smp_cond_load_relaxed(&rcu_lazy_test1_cb_called, VAL == 1);
 777
 778                 rcu_lazy_set_jiffies_till_flush(orig_jif);
 779
 780                 if (WARN_ON_ONCE(jiffies_at_lazy_cb - jif_start < 2 * HZ)) {
 781                         pr_alert("ERROR: call_rcu() CBs are not being lazy as expected!\n");
 782                         WARN_ON_ONCE(1);
 783                         return -1;
 784                 }
 785
 786                 if (WARN_ON_ONCE(jiffies_at_lazy_cb - jif_start > 3 * HZ)) {
 787                         pr_alert("ERROR: call_rcu() CBs are being too lazy!\n");
 788                         WARN_ON_ONCE(1);
 789                         return -1;
 790                 }
 791         }
 792
 793         kfree_nrealthreads = compute_real(kfree_nthreads);
 794         /* Start up the kthreads. */
 795         if (shutdown) {
 796                 init_waitqueue_head(&shutdown_wq);
 797                 firsterr = torture_create_kthread(kfree_scale_shutdown, NULL,
 798                                                   shutdown_task);
 799                 if (torture_init_error(firsterr))
 800                         goto unwind;
 801                 schedule_timeout_uninterruptible(1);
 802         }
 803
 804         pr_alert("kfree object size=%zu, kfree_by_call_rcu=%d\n",
 805                         kfree_mult * sizeof(struct kfree_obj),
 806                         kfree_by_call_rcu);
 807
 808         kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]),
 809                                GFP_KERNEL);
 810         if (kfree_reader_tasks == NULL) {
 811                 firsterr = -ENOMEM;
 812                 goto unwind;
 813         }
 814
 815         for (i = 0; i < kfree_nrealthreads; i++) {
 816                 firsterr = torture_create_kthread(kfree_scale_thread, (void *)i,
 817                                                   kfree_reader_tasks[i]);
 818                 if (torture_init_error(firsterr))
 819                         goto unwind;
 820         }
 821
 822         while (atomic_read(&n_kfree_scale_thread_started) < kfree_nrealthreads)
 823                 schedule_timeout_uninterruptible(1);
 824
 825         torture_init_end();
 826         return 0;
 827
 828 unwind:
 829         torture_init_end();
 830         kfree_scale_cleanup();
 831         return firsterr;
 832 }
 833
 834 static void
 835 rcu_scale_cleanup(void)
 836 {
 837         int i;
 838         int j;
 839         int ngps = 0;
 840         u64 *wdp;
 841         u64 *wdpp;
 842
 843         /*
 844          * Would like warning at start, but everything is expedited
 845          * during the mid-boot phase, so have to wait till the end.
 846          */
 847         if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
 848                 SCALEOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
 849         if (rcu_gp_is_normal() && gp_exp)
 850                 SCALEOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
 851         if (gp_exp && gp_async)
 852                 SCALEOUT_ERRSTRING("No expedited async GPs, so went with async!");
 853
 854         // If built-in, just report all of the GP kthread's CPU time.
 855         if (IS_BUILTIN(CONFIG_RCU_SCALE_TEST) && !kthread_tp && cur_ops->rso_gp_kthread)
 856                 kthread_tp = cur_ops->rso_gp_kthread();
 857         if (kthread_tp) {
 858                 u32 ns;
 859                 u64 us;
 860
 861                 kthread_stime = kthread_tp->stime - kthread_stime;
 862                 us = div_u64_rem(kthread_stime, 1000, &ns);
 863                 pr_info("rcu_scale: Grace-period kthread CPU time: %llu.%03u us\n", us, ns);
 864                 show_rcu_gp_kthreads();
 865         }
 866         if (kfree_rcu_test) {
 867                 kfree_scale_cleanup();
 868                 return;
 869         }
 870
 871         if (torture_cleanup_begin())
 872                 return;
 873         if (!cur_ops) {
 874                 torture_cleanup_end();
 875                 return;
 876         }
 877
 878         if (reader_tasks) {
 879                 for (i = 0; i < nrealreaders; i++)
 880                         torture_stop_kthread(rcu_scale_reader,
 881                                              reader_tasks[i]);
 882                 kfree(reader_tasks);
 883         }
 884
 885         if (writer_tasks) {
 886                 for (i = 0; i < nrealwriters; i++) {
 887                         torture_stop_kthread(rcu_scale_writer,
 888                                              writer_tasks[i]);
 889                         if (!writer_n_durations)
 890                                 continue;
 891                         j = writer_n_durations[i];
 892                         pr_alert("%s%s writer %d gps: %d\n",
 893                                  scale_type, SCALE_FLAG, i, j);
 894                         ngps += j;
 895                 }
 896                 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
 897                          scale_type, SCALE_FLAG,
 898                          t_rcu_scale_writer_started, t_rcu_scale_writer_finished,
 899                          t_rcu_scale_writer_finished -
 900                          t_rcu_scale_writer_started,
 901                          ngps,
 902                          rcuscale_seq_diff(b_rcu_gp_test_finished,
 903                                            b_rcu_gp_test_started));
 904                 for (i = 0; i < nrealwriters; i++) {
 905                         if (!writer_durations)
 906                                 break;
 907                         if (!writer_n_durations)
 908                                 continue;
 909                         wdpp = writer_durations[i];
 910                         if (!wdpp)
 911                                 continue;
 912                         for (j = 0; j < writer_n_durations[i]; j++) {
 913                                 wdp = &wdpp[j];
 914                                 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
 915                                         scale_type, SCALE_FLAG,
 916                                         i, j, *wdp);
 917                                 if (j % 100 == 0)
 918                                         schedule_timeout_uninterruptible(1);
 919                         }
 920                         kfree(writer_durations[i]);
 921                 }
 922                 kfree(writer_tasks);
 923                 kfree(writer_durations);
 924                 kfree(writer_n_durations);
 925         }
 926
 927         /* Do torture-type-specific cleanup operations.  */
 928         if (cur_ops->cleanup != NULL)
 929                 cur_ops->cleanup();
 930
 931         torture_cleanup_end();
 932 }
 933
 934 /*
 935  * RCU scalability shutdown kthread.  Just waits to be awakened, then shuts
 936  * down system.
 937  */
 938 static int
 939 rcu_scale_shutdown(void *arg)
 940 {
 941         wait_event_idle(shutdown_wq, atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters);
 942         smp_mb(); /* Wake before output. */
 943         rcu_scale_cleanup();
 944         kernel_power_off();
 945         return -EINVAL;
 946 }
 947
 948 static int __init
 949 rcu_scale_init(void)
 950 {
 951         long i;
 952         int firsterr = 0;
 953         static struct rcu_scale_ops *scale_ops[] = {
 954                 &rcu_ops, &srcu_ops, &srcud_ops, TASKS_OPS TASKS_RUDE_OPS TASKS_TRACING_OPS
 955         };
 956
 957         if (!torture_init_begin(scale_type, verbose))
 958                 return -EBUSY;
 959
 960         /* Process args and announce that the scalability'er is on the job. */
 961         for (i = 0; i < ARRAY_SIZE(scale_ops); i++) {
 962                 cur_ops = scale_ops[i];
 963                 if (strcmp(scale_type, cur_ops->name) == 0)
 964                         break;
 965         }
 966         if (i == ARRAY_SIZE(scale_ops)) {
 967                 pr_alert("rcu-scale: invalid scale type: \"%s\"\n", scale_type);
 968                 pr_alert("rcu-scale types:");
 969                 for (i = 0; i < ARRAY_SIZE(scale_ops); i++)
 970                         pr_cont(" %s", scale_ops[i]->name);
 971                 pr_cont("\n");
 972                 firsterr = -EINVAL;
 973                 cur_ops = NULL;
 974                 goto unwind;
 975         }
 976         if (cur_ops->init)
 977                 cur_ops->init();
 978
 979         if (cur_ops->rso_gp_kthread) {
 980                 kthread_tp = cur_ops->rso_gp_kthread();
 981                 if (kthread_tp)
 982                         kthread_stime = kthread_tp->stime;
 983         }
 984         if (kfree_rcu_test)
 985                 return kfree_scale_init();
 986
 987         nrealwriters = compute_real(nwriters);
 988         nrealreaders = compute_real(nreaders);
 989         atomic_set(&n_rcu_scale_reader_started, 0);
 990         atomic_set(&n_rcu_scale_writer_started, 0);
 991         atomic_set(&n_rcu_scale_writer_finished, 0);
 992         rcu_scale_print_module_parms(cur_ops, "Start of test");
 993
 994         /* Start up the kthreads. */
 995
 996         if (shutdown) {
 997                 init_waitqueue_head(&shutdown_wq);
 998                 firsterr = torture_create_kthread(rcu_scale_shutdown, NULL,
 999                                                   shutdown_task);
1000                 if (torture_init_error(firsterr))
1001                         goto unwind;
1002                 schedule_timeout_uninterruptible(1);
1003         }
1004         reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
1005                                GFP_KERNEL);
1006         if (reader_tasks == NULL) {
1007                 SCALEOUT_ERRSTRING("out of memory");
1008                 firsterr = -ENOMEM;
1009                 goto unwind;
1010         }
1011         for (i = 0; i < nrealreaders; i++) {
1012                 firsterr = torture_create_kthread(rcu_scale_reader, (void *)i,
1013                                                   reader_tasks[i]);
1014                 if (torture_init_error(firsterr))
1015                         goto unwind;
1016         }
1017         while (atomic_read(&n_rcu_scale_reader_started) < nrealreaders)
1018                 schedule_timeout_uninterruptible(1);
1019         writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]),
1020                                GFP_KERNEL);
1021         writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations),
1022                                    GFP_KERNEL);
1023         writer_n_durations =
1024                 kcalloc(nrealwriters, sizeof(*writer_n_durations),
1025                         GFP_KERNEL);
1026         if (!writer_tasks || !writer_durations || !writer_n_durations) {
1027                 SCALEOUT_ERRSTRING("out of memory");
1028                 firsterr = -ENOMEM;
1029                 goto unwind;
1030         }
1031         for (i = 0; i < nrealwriters; i++) {
1032                 writer_durations[i] =
1033                         kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
1034                                 GFP_KERNEL);
1035                 if (!writer_durations[i]) {
1036                         firsterr = -ENOMEM;
1037                         goto unwind;
1038                 }
1039                 firsterr = torture_create_kthread(rcu_scale_writer, (void *)i,
1040                                                   writer_tasks[i]);
1041                 if (torture_init_error(firsterr))
1042                         goto unwind;
1043         }
1044         torture_init_end();
1045         return 0;
1046
1047 unwind:
1048         torture_init_end();
1049         rcu_scale_cleanup();
1050         if (shutdown) {
1051                 WARN_ON(!IS_MODULE(CONFIG_RCU_SCALE_TEST));
1052                 kernel_power_off();
1053         }
1054         return firsterr;
1055 }
1056
1057 module_init(rcu_scale_init);
1058 module_exit(rcu_scale_cleanup);