| 1 | /* |
| 2 | * rcutorture.h: simple user-level performance/stress test of RCU. |
| 3 | * |
| 4 | * Usage: |
| 5 | * ./rcu <nreaders> rperf [ <cpustride> ] |
| 6 | * Run a read-side performance test with the specified |
| 7 | * number of readers spaced by <cpustride>. |
| 8 | * Thus "./rcu 16 rperf 2" would run 16 readers on even-numbered |
| 9 | * CPUs from 0 to 30. |
| 10 | * ./rcu <nupdaters> uperf [ <cpustride> ] |
| 11 | * Run an update-side performance test with the specified |
| 12 | * number of updaters and specified CPU spacing. |
| 13 | * ./rcu <nreaders> perf [ <cpustride> ] |
| 14 | * Run a combined read/update performance test with the specified |
| 15 | * number of readers and one updater and specified CPU spacing. |
| 16 | * The readers run on the low-numbered CPUs and the updater |
| 17 | * of the highest-numbered CPU. |
| 18 | * |
| 19 | * The above tests produce output as follows: |
| 20 | * |
| 21 | * n_reads: 46008000 n_updates: 146026 nreaders: 2 nupdaters: 1 duration: 1 |
| 22 | * ns/read: 43.4707 ns/update: 6848.1 |
| 23 | * |
| 24 | * The first line lists the total number of RCU reads and updates executed |
| 25 | * during the test, the number of reader threads, the number of updater |
| 26 | * threads, and the duration of the test in seconds. The second line |
| 27 | * lists the average duration of each type of operation in nanoseconds, |
| 28 | * or "nan" if the corresponding type of operation was not performed. |
| 29 | * |
| 30 | * ./rcu <nreaders> stress |
| 31 | * Run a stress test with the specified number of readers and |
| 32 | * one updater. None of the threads are affinitied to any |
| 33 | * particular CPU. |
| 34 | * |
| 35 | * This test produces output as follows: |
| 36 | * |
| 37 | * n_reads: 114633217 n_updates: 3903415 n_mberror: 0 |
| 38 | * rcu_stress_count: 114618391 14826 0 0 0 0 0 0 0 0 0 |
| 39 | * |
| 40 | * The first line lists the number of RCU read and update operations |
| 41 | * executed, followed by the number of memory-ordering violations |
| 42 | * (which will be zero in a correct RCU implementation). The second |
| 43 | * line lists the number of readers observing progressively more stale |
| 44 | * data. A correct RCU implementation will have all but the first two |
| 45 | * numbers non-zero. |
| 46 | * |
| 47 | * This program is free software; you can redistribute it and/or modify |
| 48 | * it under the terms of the GNU General Public License as published by |
| 49 | * the Free Software Foundation; either version 2 of the License, or |
| 50 | * (at your option) any later version. |
| 51 | * |
| 52 | * This program is distributed in the hope that it will be useful, |
| 53 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 54 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 55 | * GNU General Public License for more details. |
| 56 | * |
| 57 | * You should have received a copy of the GNU General Public License |
| 58 | * along with this program; if not, write to the Free Software |
| 59 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. |
| 60 | * |
| 61 | * Copyright (c) 2008 Paul E. McKenney, IBM Corporation. |
| 62 | */ |
| 63 | |
| 64 | /* |
| 65 | * Test variables. |
| 66 | */ |
| 67 | |
| 68 | DEFINE_PER_THREAD(long long, n_reads_pt); |
| 69 | DEFINE_PER_THREAD(long long, n_updates_pt); |
| 70 | |
| 71 | long long n_reads = 0LL; |
| 72 | long n_updates = 0L; |
| 73 | atomic_t nthreadsrunning; |
| 74 | char argsbuf[64]; |
| 75 | |
| 76 | #define GOFLAG_INIT 0 |
| 77 | #define GOFLAG_RUN 1 |
| 78 | #define GOFLAG_STOP 2 |
| 79 | |
| 80 | int goflag __attribute__((__aligned__(CACHE_LINE_SIZE))) = GOFLAG_INIT; |
| 81 | |
| 82 | #define RCU_READ_RUN 1000 |
| 83 | |
| 84 | //MD |
| 85 | #define RCU_READ_NESTABLE |
| 86 | |
| 87 | #ifdef RCU_READ_NESTABLE |
| 88 | #define rcu_read_lock_nest() rcu_read_lock() |
| 89 | #define rcu_read_unlock_nest() rcu_read_unlock() |
| 90 | #else /* #ifdef RCU_READ_NESTABLE */ |
| 91 | #define rcu_read_lock_nest() |
| 92 | #define rcu_read_unlock_nest() |
| 93 | #endif /* #else #ifdef RCU_READ_NESTABLE */ |
| 94 | |
| 95 | #ifndef mark_rcu_quiescent_state |
| 96 | #define mark_rcu_quiescent_state() do ; while (0) |
| 97 | #endif /* #ifdef mark_rcu_quiescent_state */ |
| 98 | |
| 99 | #ifndef put_thread_offline |
| 100 | #define put_thread_offline() do ; while (0) |
| 101 | #define put_thread_online() do ; while (0) |
| 102 | #define put_thread_online_delay() do ; while (0) |
| 103 | #else /* #ifndef put_thread_offline */ |
| 104 | #define put_thread_online_delay() synchronize_rcu() |
| 105 | #endif /* #else #ifndef put_thread_offline */ |
| 106 | |
| 107 | /* |
| 108 | * Performance test. |
| 109 | */ |
| 110 | |
| 111 | void *rcu_read_perf_test(void *arg) |
| 112 | { |
| 113 | int i; |
| 114 | int me = (long)arg; |
| 115 | long long n_reads_local = 0; |
| 116 | |
| 117 | rcu_register_thread(); |
| 118 | run_on(me); |
| 119 | atomic_inc(&nthreadsrunning); |
| 120 | while (goflag == GOFLAG_INIT) |
| 121 | poll(NULL, 0, 1); |
| 122 | mark_rcu_quiescent_state(); |
| 123 | while (goflag == GOFLAG_RUN) { |
| 124 | for (i = 0; i < RCU_READ_RUN; i++) { |
| 125 | rcu_read_lock(); |
| 126 | /* rcu_read_lock_nest(); */ |
| 127 | /* rcu_read_unlock_nest(); */ |
| 128 | rcu_read_unlock(); |
| 129 | } |
| 130 | n_reads_local += RCU_READ_RUN; |
| 131 | mark_rcu_quiescent_state(); |
| 132 | } |
| 133 | __get_thread_var(n_reads_pt) += n_reads_local; |
| 134 | put_thread_offline(); |
| 135 | rcu_unregister_thread(); |
| 136 | |
| 137 | return (NULL); |
| 138 | } |
| 139 | |
| 140 | void *rcu_update_perf_test(void *arg) |
| 141 | { |
| 142 | long long n_updates_local = 0; |
| 143 | |
| 144 | atomic_inc(&nthreadsrunning); |
| 145 | while (goflag == GOFLAG_INIT) |
| 146 | poll(NULL, 0, 1); |
| 147 | while (goflag == GOFLAG_RUN) { |
| 148 | synchronize_rcu(); |
| 149 | n_updates_local++; |
| 150 | } |
| 151 | __get_thread_var(n_updates_pt) += n_updates_local; |
| 152 | return NULL; |
| 153 | } |
| 154 | |
| 155 | void perftestinit(void) |
| 156 | { |
| 157 | init_per_thread(n_reads_pt, 0LL); |
| 158 | init_per_thread(n_updates_pt, 0LL); |
| 159 | atomic_set(&nthreadsrunning, 0); |
| 160 | } |
| 161 | |
| 162 | void perftestrun(int nthreads, int nreaders, int nupdaters) |
| 163 | { |
| 164 | int t; |
| 165 | int duration = 1; |
| 166 | |
| 167 | smp_mb(); |
| 168 | while (atomic_read(&nthreadsrunning) < nthreads) |
| 169 | poll(NULL, 0, 1); |
| 170 | goflag = GOFLAG_RUN; |
| 171 | smp_mb(); |
| 172 | sleep(duration); |
| 173 | smp_mb(); |
| 174 | goflag = GOFLAG_STOP; |
| 175 | smp_mb(); |
| 176 | wait_all_threads(); |
| 177 | for_each_thread(t) { |
| 178 | n_reads += per_thread(n_reads_pt, t); |
| 179 | n_updates += per_thread(n_updates_pt, t); |
| 180 | } |
| 181 | printf("n_reads: %lld n_updates: %ld nreaders: %d nupdaters: %d duration: %d\n", |
| 182 | n_reads, n_updates, nreaders, nupdaters, duration); |
| 183 | printf("ns/read: %g ns/update: %g\n", |
| 184 | ((duration * 1000*1000*1000.*(double)nreaders) / |
| 185 | (double)n_reads), |
| 186 | ((duration * 1000*1000*1000.*(double)nupdaters) / |
| 187 | (double)n_updates)); |
| 188 | exit(0); |
| 189 | } |
| 190 | |
| 191 | void perftest(int nreaders, int cpustride) |
| 192 | { |
| 193 | int i; |
| 194 | long arg; |
| 195 | |
| 196 | perftestinit(); |
| 197 | for (i = 0; i < nreaders; i++) { |
| 198 | arg = (long)(i * cpustride); |
| 199 | create_thread(rcu_read_perf_test, (void *)arg); |
| 200 | } |
| 201 | arg = (long)(i * cpustride); |
| 202 | create_thread(rcu_update_perf_test, (void *)arg); |
| 203 | perftestrun(i + 1, nreaders, 1); |
| 204 | } |
| 205 | |
| 206 | void rperftest(int nreaders, int cpustride) |
| 207 | { |
| 208 | int i; |
| 209 | long arg; |
| 210 | |
| 211 | perftestinit(); |
| 212 | init_per_thread(n_reads_pt, 0LL); |
| 213 | for (i = 0; i < nreaders; i++) { |
| 214 | arg = (long)(i * cpustride); |
| 215 | create_thread(rcu_read_perf_test, (void *)arg); |
| 216 | } |
| 217 | perftestrun(i, nreaders, 0); |
| 218 | } |
| 219 | |
| 220 | void uperftest(int nupdaters, int cpustride) |
| 221 | { |
| 222 | int i; |
| 223 | long arg; |
| 224 | |
| 225 | perftestinit(); |
| 226 | init_per_thread(n_reads_pt, 0LL); |
| 227 | for (i = 0; i < nupdaters; i++) { |
| 228 | arg = (long)(i * cpustride); |
| 229 | create_thread(rcu_update_perf_test, (void *)arg); |
| 230 | } |
| 231 | perftestrun(i, 0, nupdaters); |
| 232 | } |
| 233 | |
| 234 | /* |
| 235 | * Stress test. |
| 236 | */ |
| 237 | |
| 238 | #define RCU_STRESS_PIPE_LEN 10 |
| 239 | |
| 240 | struct rcu_stress { |
| 241 | int pipe_count; |
| 242 | int mbtest; |
| 243 | }; |
| 244 | |
| 245 | struct rcu_stress rcu_stress_array[RCU_STRESS_PIPE_LEN] = { { 0 } }; |
| 246 | struct rcu_stress *rcu_stress_current; |
| 247 | int rcu_stress_idx = 0; |
| 248 | |
| 249 | int n_mberror = 0; |
| 250 | DEFINE_PER_THREAD(long long [RCU_STRESS_PIPE_LEN + 1], rcu_stress_count); |
| 251 | |
| 252 | int garbage = 0; |
| 253 | |
| 254 | void *rcu_read_stress_test(void *arg) |
| 255 | { |
| 256 | int i; |
| 257 | int itercnt = 0; |
| 258 | struct rcu_stress *p; |
| 259 | int pc; |
| 260 | |
| 261 | rcu_register_thread(); |
| 262 | while (goflag == GOFLAG_INIT) |
| 263 | poll(NULL, 0, 1); |
| 264 | mark_rcu_quiescent_state(); |
| 265 | while (goflag == GOFLAG_RUN) { |
| 266 | rcu_read_lock(); |
| 267 | p = rcu_dereference(rcu_stress_current); |
| 268 | if (p->mbtest == 0) |
| 269 | n_mberror++; |
| 270 | rcu_read_lock_nest(); |
| 271 | for (i = 0; i < 100; i++) |
| 272 | garbage++; |
| 273 | rcu_read_unlock_nest(); |
| 274 | pc = p->pipe_count; |
| 275 | rcu_read_unlock(); |
| 276 | if ((pc > RCU_STRESS_PIPE_LEN) || (pc < 0)) |
| 277 | pc = RCU_STRESS_PIPE_LEN; |
| 278 | __get_thread_var(rcu_stress_count)[pc]++; |
| 279 | __get_thread_var(n_reads_pt)++; |
| 280 | mark_rcu_quiescent_state(); |
| 281 | if ((++itercnt % 0x1000) == 0) { |
| 282 | put_thread_offline(); |
| 283 | put_thread_online_delay(); |
| 284 | put_thread_online(); |
| 285 | } |
| 286 | } |
| 287 | put_thread_offline(); |
| 288 | rcu_unregister_thread(); |
| 289 | |
| 290 | return (NULL); |
| 291 | } |
| 292 | |
| 293 | void *rcu_update_stress_test(void *arg) |
| 294 | { |
| 295 | int i; |
| 296 | struct rcu_stress *p; |
| 297 | |
| 298 | while (goflag == GOFLAG_INIT) |
| 299 | poll(NULL, 0, 1); |
| 300 | while (goflag == GOFLAG_RUN) { |
| 301 | i = rcu_stress_idx + 1; |
| 302 | if (i >= RCU_STRESS_PIPE_LEN) |
| 303 | i = 0; |
| 304 | p = &rcu_stress_array[i]; |
| 305 | p->mbtest = 0; |
| 306 | smp_mb(); |
| 307 | p->pipe_count = 0; |
| 308 | p->mbtest = 1; |
| 309 | rcu_assign_pointer(rcu_stress_current, p); |
| 310 | rcu_stress_idx = i; |
| 311 | for (i = 0; i < RCU_STRESS_PIPE_LEN; i++) |
| 312 | if (i != rcu_stress_idx) |
| 313 | rcu_stress_array[i].pipe_count++; |
| 314 | synchronize_rcu(); |
| 315 | n_updates++; |
| 316 | } |
| 317 | return NULL; |
| 318 | } |
| 319 | |
| 320 | void *rcu_fake_update_stress_test(void *arg) |
| 321 | { |
| 322 | while (goflag == GOFLAG_INIT) |
| 323 | poll(NULL, 0, 1); |
| 324 | while (goflag == GOFLAG_RUN) { |
| 325 | synchronize_rcu(); |
| 326 | poll(NULL, 0, 1); |
| 327 | } |
| 328 | return NULL; |
| 329 | } |
| 330 | |
| 331 | void stresstest(int nreaders) |
| 332 | { |
| 333 | int i; |
| 334 | int t; |
| 335 | long long *p; |
| 336 | long long sum; |
| 337 | |
| 338 | init_per_thread(n_reads_pt, 0LL); |
| 339 | for_each_thread(t) { |
| 340 | p = &per_thread(rcu_stress_count,t)[0]; |
| 341 | for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++) |
| 342 | p[i] = 0LL; |
| 343 | } |
| 344 | rcu_stress_current = &rcu_stress_array[0]; |
| 345 | rcu_stress_current->pipe_count = 0; |
| 346 | rcu_stress_current->mbtest = 1; |
| 347 | for (i = 0; i < nreaders; i++) |
| 348 | create_thread(rcu_read_stress_test, NULL); |
| 349 | create_thread(rcu_update_stress_test, NULL); |
| 350 | for (i = 0; i < 5; i++) |
| 351 | create_thread(rcu_fake_update_stress_test, NULL); |
| 352 | smp_mb(); |
| 353 | goflag = GOFLAG_RUN; |
| 354 | smp_mb(); |
| 355 | sleep(10); |
| 356 | smp_mb(); |
| 357 | goflag = GOFLAG_STOP; |
| 358 | smp_mb(); |
| 359 | wait_all_threads(); |
| 360 | for_each_thread(t) |
| 361 | n_reads += per_thread(n_reads_pt, t); |
| 362 | printf("n_reads: %lld n_updates: %ld n_mberror: %d\n", |
| 363 | n_reads, n_updates, n_mberror); |
| 364 | printf("rcu_stress_count:"); |
| 365 | for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++) { |
| 366 | sum = 0LL; |
| 367 | for_each_thread(t) { |
| 368 | sum += per_thread(rcu_stress_count, t)[i]; |
| 369 | } |
| 370 | printf(" %lld", sum); |
| 371 | } |
| 372 | printf("\n"); |
| 373 | exit(0); |
| 374 | } |
| 375 | |
| 376 | /* |
| 377 | * Mainprogram. |
| 378 | */ |
| 379 | |
| 380 | void usage(int argc, char *argv[]) |
| 381 | { |
| 382 | fprintf(stderr, "Usage: %s [nreaders [ perf | stress ] ]\n", argv[0]); |
| 383 | exit(-1); |
| 384 | } |
| 385 | |
| 386 | int main(int argc, char *argv[]) |
| 387 | { |
| 388 | int nreaders = 1; |
| 389 | int cpustride = 1; |
| 390 | |
| 391 | smp_init(); |
| 392 | //rcu_init(); |
| 393 | |
| 394 | #ifdef DEBUG_YIELD |
| 395 | yield_active |= YIELD_READ; |
| 396 | yield_active |= YIELD_WRITE; |
| 397 | #endif |
| 398 | |
| 399 | if (argc > 1) { |
| 400 | nreaders = strtoul(argv[1], NULL, 0); |
| 401 | if (argc == 2) |
| 402 | perftest(nreaders, cpustride); |
| 403 | if (argc > 3) |
| 404 | cpustride = strtoul(argv[3], NULL, 0); |
| 405 | if (strcmp(argv[2], "perf") == 0) |
| 406 | perftest(nreaders, cpustride); |
| 407 | else if (strcmp(argv[2], "rperf") == 0) |
| 408 | rperftest(nreaders, cpustride); |
| 409 | else if (strcmp(argv[2], "uperf") == 0) |
| 410 | uperftest(nreaders, cpustride); |
| 411 | else if (strcmp(argv[2], "stress") == 0) |
| 412 | stresstest(nreaders); |
| 413 | usage(argc, argv); |
| 414 | } |
| 415 | perftest(nreaders, cpustride); |
| 416 | return 0; |
| 417 | } |