tests/regression/rcutorture.h

   1 // SPDX-FileCopyrightText: 2008 Paul E. McKenney, IBM Corporation.
   2 //
   3 // SPDX-License-Identifier: GPL-2.0-or-later
   4
   5 /*
   6  * rcutorture.h: simple user-level performance/stress test of RCU.
   7  *
   8  * Usage:
   9  *      ./rcu <nreaders> rperf [ <cpustride> ]
  10  *              Run a read-side performance test with the specified
  11  *              number of readers spaced by <cpustride>.
  12  *              Thus "./rcu 16 rperf 2" would run 16 readers on even-numbered
  13  *              CPUs from 0 to 30.
  14  *      ./rcu <nupdaters> uperf [ <cpustride> ]
  15  *              Run an update-side performance test with the specified
  16  *              number of updaters and specified CPU spacing.
  17  *      ./rcu <nreaders> perf [ <cpustride> ]
  18  *              Run a combined read/update performance test with the specified
  19  *              number of readers and one updater and specified CPU spacing.
  20  *              The readers run on the low-numbered CPUs and the updater
  21  *              of the highest-numbered CPU.
  22  *
  23  * The above tests produce output as follows:
  24  *
  25  * n_reads: 46008000  n_updates: 146026  nreaders: 2  nupdaters: 1 duration: 1
  26  * ns/read: 43.4707  ns/update: 6848.1
  27  *
  28  * The first line lists the total number of RCU reads and updates executed
  29  * during the test, the number of reader threads, the number of updater
  30  * threads, and the duration of the test in seconds.  The second line
  31  * lists the average duration of each type of operation in nanoseconds,
  32  * or "nan" if the corresponding type of operation was not performed.
  33  *
  34  *      ./rcu <nreaders> stress
  35  *              Run a stress test with the specified number of readers and
  36  *              one updater.  None of the threads are affinitied to any
  37  *              particular CPU.
  38  *
  39  * This test produces output as follows:
  40  *
  41  * n_reads: 114633217  n_updates: 3903415  n_mberror: 0
  42  * rcu_stress_count: 114618391 14826 0 0 0 0 0 0 0 0 0
  43  *
  44  * The first line lists the number of RCU read and update operations
  45  * executed, followed by the number of memory-ordering violations
  46  * (which will be zero in a correct RCU implementation).  The second
  47  * line lists the number of readers observing progressively more stale
  48  * data.  A correct RCU implementation will have all but the first two
  49  * numbers non-zero.
  50  */
  51
  52 /*
  53  * Test variables.
  54  */
  55
  56 #include <stdlib.h>
  57 #include "tap.h"
  58
  59 #include "urcu-wait.h"
  60
  61 #define NR_TESTS        1
  62
  63 DEFINE_PER_THREAD(long long, n_reads_pt);
  64 DEFINE_PER_THREAD(long long, n_updates_pt);
  65
  66 enum callrcu_type {
  67         CALLRCU_GLOBAL,
  68         CALLRCU_PERCPU,
  69         CALLRCU_PERTHREAD,
  70 };
  71
  72 enum writer_state {
  73         WRITER_STATE_SYNC_RCU,
  74         WRITER_STATE_CALL_RCU,
  75         WRITER_STATE_POLL_RCU,
  76 };
  77
  78 static enum callrcu_type callrcu_type = CALLRCU_GLOBAL;
  79
  80 long long n_reads = 0LL;
  81 long n_updates = 0L;
  82 int nthreadsrunning;
  83 char argsbuf[64];
  84
  85 #define GOFLAG_INIT 0
  86 #define GOFLAG_RUN  1
  87 #define GOFLAG_STOP 2
  88
  89 volatile int goflag __attribute__((__aligned__(CAA_CACHE_LINE_SIZE)))
  90         = GOFLAG_INIT;
  91
  92 #define RCU_READ_RUN 1000
  93
  94 //MD
  95 #define RCU_READ_NESTABLE
  96
  97 #ifdef RCU_READ_NESTABLE
  98 #define rcu_read_lock_nest() rcu_read_lock()
  99 #define rcu_read_unlock_nest() rcu_read_unlock()
 100 #else /* #ifdef RCU_READ_NESTABLE */
 101 #define rcu_read_lock_nest()
 102 #define rcu_read_unlock_nest()
 103 #endif /* #else #ifdef RCU_READ_NESTABLE */
 104
 105 #ifdef TORTURE_QSBR
 106 #define mark_rcu_quiescent_state        rcu_quiescent_state
 107 #define put_thread_offline              rcu_thread_offline
 108 #define put_thread_online               rcu_thread_online
 109 #endif
 110
 111 #ifndef mark_rcu_quiescent_state
 112 #define mark_rcu_quiescent_state() do {} while (0)
 113 #endif /* #ifdef mark_rcu_quiescent_state */
 114
 115 #ifndef put_thread_offline
 116 #define put_thread_offline()            do {} while (0)
 117 #define put_thread_online()             do {} while (0)
 118 #define put_thread_online_delay()       do {} while (0)
 119 #else /* #ifndef put_thread_offline */
 120 #define put_thread_online_delay()       synchronize_rcu()
 121 #endif /* #else #ifndef put_thread_offline */
 122
 123 /*
 124  * Performance test.
 125  */
 126
 127 static
 128 void *rcu_read_perf_test(void *arg)
 129 {
 130         int i;
 131         int me = (long)arg;
 132         long long n_reads_local = 0;
 133
 134         rcu_register_thread();
 135         run_on(me);
 136         uatomic_inc(&nthreadsrunning);
 137         put_thread_offline();
 138         while (goflag == GOFLAG_INIT)
 139                 (void) poll(NULL, 0, 1);
 140         put_thread_online();
 141         while (goflag == GOFLAG_RUN) {
 142                 for (i = 0; i < RCU_READ_RUN; i++) {
 143                         rcu_read_lock();
 144                         /* rcu_read_lock_nest(); */
 145                         /* rcu_read_unlock_nest(); */
 146                         rcu_read_unlock();
 147                 }
 148                 n_reads_local += RCU_READ_RUN;
 149                 mark_rcu_quiescent_state();
 150         }
 151         __get_thread_var(n_reads_pt) += n_reads_local;
 152         put_thread_offline();
 153         rcu_unregister_thread();
 154
 155         return (NULL);
 156 }
 157
 158 static
 159 void *rcu_update_perf_test(void *arg __attribute__((unused)))
 160 {
 161         long long n_updates_local = 0;
 162
 163         if (callrcu_type == CALLRCU_PERTHREAD) {
 164                 struct call_rcu_data *crdp;
 165
 166                 crdp = create_call_rcu_data(0, -1);
 167                 if (crdp != NULL) {
 168                         diag("Successfully using per-thread call_rcu() worker.");
 169                         set_thread_call_rcu_data(crdp);
 170                 }
 171         }
 172         uatomic_inc(&nthreadsrunning);
 173         while (goflag == GOFLAG_INIT)
 174                 (void) poll(NULL, 0, 1);
 175         while (goflag == GOFLAG_RUN) {
 176                 synchronize_rcu();
 177                 n_updates_local++;
 178         }
 179         __get_thread_var(n_updates_pt) += n_updates_local;
 180         if (callrcu_type == CALLRCU_PERTHREAD) {
 181                 struct call_rcu_data *crdp;
 182
 183                 crdp = get_thread_call_rcu_data();
 184                 set_thread_call_rcu_data(NULL);
 185                 call_rcu_data_free(crdp);
 186         }
 187         return NULL;
 188 }
 189
 190 static
 191 void perftestinit(void)
 192 {
 193         init_per_thread(n_reads_pt, 0LL);
 194         init_per_thread(n_updates_pt, 0LL);
 195         uatomic_set(&nthreadsrunning, 0);
 196 }
 197
 198 static
 199 int perftestrun(int nthreads, int nreaders, int nupdaters)
 200 {
 201         int t;
 202         int duration = 1;
 203
 204         cmm_smp_mb();
 205         while (uatomic_read(&nthreadsrunning) < nthreads)
 206                 (void) poll(NULL, 0, 1);
 207         goflag = GOFLAG_RUN;
 208         cmm_smp_mb();
 209         sleep(duration);
 210         cmm_smp_mb();
 211         goflag = GOFLAG_STOP;
 212         cmm_smp_mb();
 213         wait_all_threads();
 214         for_each_thread(t) {
 215                 n_reads += per_thread(n_reads_pt, t);
 216                 n_updates += per_thread(n_updates_pt, t);
 217         }
 218         diag("n_reads: %lld  n_updates: %ld  nreaders: %d  nupdaters: %d duration: %d",
 219                n_reads, n_updates, nreaders, nupdaters, duration);
 220         diag("ns/read: %g  ns/update: %g",
 221                ((duration * 1000*1000*1000.*(double)nreaders) /
 222                 (double)n_reads),
 223                ((duration * 1000*1000*1000.*(double)nupdaters) /
 224                 (double)n_updates));
 225         if (get_cpu_call_rcu_data(0)) {
 226                 diag("Deallocating per-CPU call_rcu threads.\n");
 227                 free_all_cpu_call_rcu_data();
 228         }
 229         return 0;
 230 }
 231
 232 static
 233 int perftest(int nreaders, int cpustride)
 234 {
 235         int i;
 236         long arg;
 237
 238         perftestinit();
 239         for (i = 0; i < nreaders; i++) {
 240                 arg = (long)(i * cpustride);
 241                 create_thread(rcu_read_perf_test, (void *)arg);
 242         }
 243         arg = (long)(i * cpustride);
 244         create_thread(rcu_update_perf_test, (void *)arg);
 245         return perftestrun(i + 1, nreaders, 1);
 246 }
 247
 248 static
 249 int rperftest(int nreaders, int cpustride)
 250 {
 251         int i;
 252         long arg;
 253
 254         perftestinit();
 255         init_per_thread(n_reads_pt, 0LL);
 256         for (i = 0; i < nreaders; i++) {
 257                 arg = (long)(i * cpustride);
 258                 create_thread(rcu_read_perf_test, (void *)arg);
 259         }
 260         return perftestrun(i, nreaders, 0);
 261 }
 262
 263 static
 264 int uperftest(int nupdaters, int cpustride)
 265 {
 266         int i;
 267         long arg;
 268
 269         perftestinit();
 270         init_per_thread(n_reads_pt, 0LL);
 271         for (i = 0; i < nupdaters; i++) {
 272                 arg = (long)(i * cpustride);
 273                 create_thread(rcu_update_perf_test, (void *)arg);
 274         }
 275         return perftestrun(i, 0, nupdaters);
 276 }
 277
 278 /*
 279  * Stress test.
 280  */
 281
 282 #define RCU_STRESS_PIPE_LEN 10
 283
 284 struct rcu_stress {
 285         int pipe_count;
 286         int mbtest;
 287 };
 288
 289 struct rcu_stress rcu_stress_array[RCU_STRESS_PIPE_LEN] = { { 0, 0 } };
 290 struct rcu_stress *rcu_stress_current;
 291 int rcu_stress_idx = 0;
 292
 293 int n_mberror = 0;
 294 DEFINE_PER_THREAD(long long [RCU_STRESS_PIPE_LEN + 1], rcu_stress_count);
 295
 296 int garbage = 0;
 297
 298 static
 299 void *rcu_read_stress_test(void *arg __attribute__((unused)))
 300 {
 301         int i;
 302         int itercnt = 0;
 303         struct rcu_stress *p;
 304         int pc;
 305
 306         rcu_register_thread();
 307         put_thread_offline();
 308         while (goflag == GOFLAG_INIT)
 309                 (void) poll(NULL, 0, 1);
 310         put_thread_online();
 311         while (goflag == GOFLAG_RUN) {
 312                 rcu_read_lock();
 313                 p = rcu_dereference(rcu_stress_current);
 314                 if (p->mbtest == 0)
 315                         n_mberror++;
 316                 rcu_read_lock_nest();
 317                 for (i = 0; i < 100; i++)
 318                         garbage++;
 319                 rcu_read_unlock_nest();
 320                 pc = p->pipe_count;
 321                 rcu_read_unlock();
 322                 if ((pc > RCU_STRESS_PIPE_LEN) || (pc < 0))
 323                         pc = RCU_STRESS_PIPE_LEN;
 324                 __get_thread_var(rcu_stress_count)[pc]++;
 325                 __get_thread_var(n_reads_pt)++;
 326                 mark_rcu_quiescent_state();
 327                 if ((++itercnt % 0x1000) == 0) {
 328                         put_thread_offline();
 329                         put_thread_online_delay();
 330                         put_thread_online();
 331                 }
 332         }
 333         put_thread_offline();
 334         rcu_unregister_thread();
 335
 336         return (NULL);
 337 }
 338
 339 static DEFINE_URCU_WAIT_QUEUE(call_rcu_waiters);
 340
 341 static
 342 void rcu_update_stress_test_rcu(struct rcu_head *head __attribute__((unused)))
 343 {
 344         struct urcu_waiters waiters;
 345
 346         urcu_move_waiters(&waiters, &call_rcu_waiters);
 347         urcu_wake_all_waiters(&waiters);
 348 }
 349
 350 static
 351 void advance_writer_state(enum writer_state *state)
 352 {
 353         switch (*state) {
 354         case WRITER_STATE_SYNC_RCU:
 355                 *state = WRITER_STATE_CALL_RCU;
 356                 break;
 357         case WRITER_STATE_CALL_RCU:
 358                 *state = WRITER_STATE_POLL_RCU;
 359                 break;
 360         case WRITER_STATE_POLL_RCU:
 361                 *state = WRITER_STATE_SYNC_RCU;
 362                 break;
 363         }
 364 }
 365
 366 static
 367 void *rcu_update_stress_test(void *arg __attribute__((unused)))
 368 {
 369         int i;
 370         struct rcu_stress *p;
 371         struct rcu_head rh;
 372         enum writer_state writer_state = WRITER_STATE_SYNC_RCU;
 373
 374         rcu_register_thread();
 375
 376         /* Offline for poll. */
 377         put_thread_offline();
 378         while (goflag == GOFLAG_INIT)
 379                 (void) poll(NULL, 0, 1);
 380         put_thread_online();
 381
 382         while (goflag == GOFLAG_RUN) {
 383                 i = rcu_stress_idx + 1;
 384                 if (i >= RCU_STRESS_PIPE_LEN)
 385                         i = 0;
 386                 p = &rcu_stress_array[i];
 387                 p->mbtest = 0;
 388                 cmm_smp_mb();
 389                 p->pipe_count = 0;
 390                 p->mbtest = 1;
 391                 rcu_assign_pointer(rcu_stress_current, p);
 392                 rcu_stress_idx = i;
 393                 for (i = 0; i < RCU_STRESS_PIPE_LEN; i++)
 394                         if (i != rcu_stress_idx)
 395                                 rcu_stress_array[i].pipe_count++;
 396                 switch (writer_state) {
 397                 case WRITER_STATE_SYNC_RCU:
 398                         synchronize_rcu();
 399                         break;
 400                 case WRITER_STATE_CALL_RCU:
 401                 {
 402                         DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING);
 403
 404                         urcu_wait_add(&call_rcu_waiters, &wait);
 405
 406                         call_rcu(&rh, rcu_update_stress_test_rcu);
 407
 408                         urcu_adaptative_busy_wait(&wait);
 409                         break;
 410                 }
 411                 case WRITER_STATE_POLL_RCU:
 412                 {
 413                         struct urcu_gp_poll_state poll_state;
 414
 415                         poll_state = start_poll_synchronize_rcu();
 416
 417                         /* Offline for poll. */
 418                         put_thread_offline();
 419                         while (!poll_state_synchronize_rcu(poll_state))
 420                                 (void) poll(NULL, 0, 1);        /* Wait for 1ms */
 421                         put_thread_online();
 422                         break;
 423                 }
 424                 }
 425                 n_updates++;
 426                 advance_writer_state(&writer_state);
 427         }
 428
 429         rcu_unregister_thread();
 430
 431         return NULL;
 432 }
 433
 434 static
 435 void *rcu_fake_update_stress_test(void *arg __attribute__((unused)))
 436 {
 437         if (callrcu_type == CALLRCU_PERTHREAD) {
 438                 struct call_rcu_data *crdp;
 439
 440                 crdp = create_call_rcu_data(0, -1);
 441                 if (crdp != NULL) {
 442                         diag("Successfully using per-thread call_rcu() worker.");
 443                         set_thread_call_rcu_data(crdp);
 444                 }
 445         }
 446         while (goflag == GOFLAG_INIT)
 447                 (void) poll(NULL, 0, 1);
 448         while (goflag == GOFLAG_RUN) {
 449                 synchronize_rcu();
 450                 (void) poll(NULL, 0, 1);
 451         }
 452         if (callrcu_type == CALLRCU_PERTHREAD) {
 453                 struct call_rcu_data *crdp;
 454
 455                 crdp = get_thread_call_rcu_data();
 456                 set_thread_call_rcu_data(NULL);
 457                 call_rcu_data_free(crdp);
 458         }
 459         return NULL;
 460 }
 461
 462 static
 463 int stresstest(int nreaders)
 464 {
 465         int i;
 466         int t;
 467         long long *p;
 468         long long sum;
 469
 470         init_per_thread(n_reads_pt, 0LL);
 471         for_each_thread(t) {
 472                 p = &per_thread(rcu_stress_count,t)[0];
 473                 for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++)
 474                         p[i] = 0LL;
 475         }
 476         rcu_stress_current = &rcu_stress_array[0];
 477         rcu_stress_current->pipe_count = 0;
 478         rcu_stress_current->mbtest = 1;
 479         for (i = 0; i < nreaders; i++)
 480                 create_thread(rcu_read_stress_test, NULL);
 481         create_thread(rcu_update_stress_test, NULL);
 482         for (i = 0; i < 5; i++)
 483                 create_thread(rcu_fake_update_stress_test, NULL);
 484         cmm_smp_mb();
 485         goflag = GOFLAG_RUN;
 486         cmm_smp_mb();
 487         sleep(10);
 488         cmm_smp_mb();
 489         goflag = GOFLAG_STOP;
 490         cmm_smp_mb();
 491         wait_all_threads();
 492         for_each_thread(t)
 493                 n_reads += per_thread(n_reads_pt, t);
 494         diag("n_reads: %lld  n_updates: %ld  n_mberror: %d",
 495                n_reads, n_updates, n_mberror);
 496         rdiag_start();
 497         rdiag("rcu_stress_count:");
 498         for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++) {
 499                 sum = 0LL;
 500                 for_each_thread(t) {
 501                         sum += per_thread(rcu_stress_count, t)[i];
 502                 }
 503                 rdiag(" %lld", sum);
 504         }
 505         rdiag_end();
 506         if (get_cpu_call_rcu_data(0)) {
 507                 diag("Deallocating per-CPU call_rcu threads.");
 508                 free_all_cpu_call_rcu_data();
 509         }
 510         if (!n_mberror)
 511                 return 0;
 512         else
 513                 return -1;
 514 }
 515
 516 /*
 517  * Mainprogram.
 518  */
 519
 520 static
 521 void usage(char *argv[]) __attribute__((__noreturn__));
 522
 523 static
 524 void usage(char *argv[])
 525 {
 526         diag("Usage: %s nreaders [ perf | rperf | uperf | stress ] [ stride ] [ callrcu_global | callrcu_percpu | callrcu_perthread ]\n", argv[0]);
 527         exit(-1);
 528 }
 529
 530 int main(int argc, char *argv[])
 531 {
 532         int nreaders = 1;
 533         int cpustride = 1;
 534
 535         plan_tests(NR_TESTS);
 536
 537         smp_init();
 538         //rcu_init();
 539         if (argc > 4) {
 540                 const char *callrcu_str = argv[4];;
 541
 542                 if (strcmp(callrcu_str, "callrcu_global") == 0) {
 543                         callrcu_type = CALLRCU_GLOBAL;
 544                 } else if (strcmp(callrcu_str, "callrcu_percpu") == 0) {
 545                         callrcu_type = CALLRCU_PERCPU;
 546                 } else if (strcmp(callrcu_str, "callrcu_perthread") == 0) {
 547                         callrcu_type = CALLRCU_PERTHREAD;
 548                 } else {
 549                         usage(argv);
 550                         goto end;
 551                 }
 552         }
 553
 554         switch (callrcu_type) {
 555         case CALLRCU_GLOBAL:
 556                 diag("Using global per-process call_rcu thread.");
 557                 break;
 558         case CALLRCU_PERCPU:
 559                 diag("Using per-CPU call_rcu threads.");
 560                 if (create_all_cpu_call_rcu_data(0))
 561                         diag("create_all_cpu_call_rcu_data: %s",
 562                                 strerror(errno));
 563                 break;
 564         case CALLRCU_PERTHREAD:
 565                 diag("Using per-thread call_rcu() worker.");
 566                 break;
 567         default:
 568                 abort();
 569         }
 570
 571 #ifdef DEBUG_YIELD
 572         yield_active |= YIELD_READ;
 573         yield_active |= YIELD_WRITE;
 574 #endif
 575
 576         if (argc > 1) {
 577                 if (strcmp(argv[1], "-h") == 0
 578                                 || strcmp(argv[1], "--help") == 0) {
 579                         usage(argv);
 580                         goto end;
 581                 }
 582                 nreaders = strtoul(argv[1], NULL, 0);
 583                 if (argc == 2) {
 584                         ok(!perftest(nreaders, cpustride),
 585                                 "perftest readers: %d, stride: %d",
 586                                 nreaders, cpustride);
 587                         goto end;
 588                 }
 589                 if (argc > 3)
 590                         cpustride = strtoul(argv[3], NULL, 0);
 591                 if (strcmp(argv[2], "perf") == 0)
 592                         ok(!perftest(nreaders, cpustride),
 593                                 "perftest readers: %d, stride: %d",
 594                                 nreaders, cpustride);
 595                 else if (strcmp(argv[2], "rperf") == 0)
 596                         ok(!rperftest(nreaders, cpustride),
 597                                 "rperftest readers: %d, stride: %d",
 598                                 nreaders, cpustride);
 599                 else if (strcmp(argv[2], "uperf") == 0)
 600                         ok(!uperftest(nreaders, cpustride),
 601                                 "uperftest readers: %d, stride: %d",
 602                                 nreaders, cpustride);
 603                 else if (strcmp(argv[2], "stress") == 0)
 604                         ok(!stresstest(nreaders),
 605                                 "stresstest readers: %d, stride: %d",
 606                                 nreaders, cpustride);
 607                 else
 608                         usage(argv);
 609         } else {
 610                 usage(argv);
 611         }
 612 end:
 613         return exit_status();
 614 }