Merge branch 'master' into urcu/ht-shrink
[urcu.git] / tests / rcutorture.h
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 #include <stdlib.h>
69
70 DEFINE_PER_THREAD(long long, n_reads_pt);
71 DEFINE_PER_THREAD(long long, n_updates_pt);
72
73 long long n_reads = 0LL;
74 long n_updates = 0L;
75 int nthreadsrunning;
76 char argsbuf[64];
77
78 #define GOFLAG_INIT 0
79 #define GOFLAG_RUN 1
80 #define GOFLAG_STOP 2
81
82 volatile int goflag __attribute__((__aligned__(CAA_CACHE_LINE_SIZE)))
83 = GOFLAG_INIT;
84
85 #define RCU_READ_RUN 1000
86
87 //MD
88 #define RCU_READ_NESTABLE
89
90 #ifdef RCU_READ_NESTABLE
91 #define rcu_read_lock_nest() rcu_read_lock()
92 #define rcu_read_unlock_nest() rcu_read_unlock()
93 #else /* #ifdef RCU_READ_NESTABLE */
94 #define rcu_read_lock_nest()
95 #define rcu_read_unlock_nest()
96 #endif /* #else #ifdef RCU_READ_NESTABLE */
97
98 #ifdef TORTURE_QSBR
99 #define mark_rcu_quiescent_state rcu_quiescent_state
100 #define put_thread_offline rcu_thread_offline
101 #define put_thread_online rcu_thread_online
102 #endif
103
104 #ifndef mark_rcu_quiescent_state
105 #define mark_rcu_quiescent_state() do ; while (0)
106 #endif /* #ifdef mark_rcu_quiescent_state */
107
108 #ifndef put_thread_offline
109 #define put_thread_offline() do ; while (0)
110 #define put_thread_online() do ; while (0)
111 #define put_thread_online_delay() do ; while (0)
112 #else /* #ifndef put_thread_offline */
113 #define put_thread_online_delay() synchronize_rcu()
114 #endif /* #else #ifndef put_thread_offline */
115
116 /*
117 * Performance test.
118 */
119
120 void *rcu_read_perf_test(void *arg)
121 {
122 struct call_rcu_data *crdp;
123 int i;
124 int me = (long)arg;
125 long long n_reads_local = 0;
126
127 rcu_register_thread();
128 run_on(me);
129 uatomic_inc(&nthreadsrunning);
130 put_thread_offline();
131 while (goflag == GOFLAG_INIT)
132 poll(NULL, 0, 1);
133 put_thread_online();
134 while (goflag == GOFLAG_RUN) {
135 for (i = 0; i < RCU_READ_RUN; i++) {
136 rcu_read_lock();
137 /* rcu_read_lock_nest(); */
138 /* rcu_read_unlock_nest(); */
139 rcu_read_unlock();
140 }
141 n_reads_local += RCU_READ_RUN;
142 mark_rcu_quiescent_state();
143 }
144 __get_thread_var(n_reads_pt) += n_reads_local;
145 put_thread_offline();
146 crdp = get_thread_call_rcu_data();
147 set_thread_call_rcu_data(NULL);
148 call_rcu_data_free(crdp);
149 rcu_unregister_thread();
150
151 return (NULL);
152 }
153
154 void *rcu_update_perf_test(void *arg)
155 {
156 long long n_updates_local = 0;
157
158 if ((random() & 0xf00) == 0) {
159 struct call_rcu_data *crdp;
160
161 crdp = create_call_rcu_data(0, -1);
162 if (crdp != NULL) {
163 fprintf(stderr,
164 "Using per-thread call_rcu() worker.\n");
165 set_thread_call_rcu_data(crdp);
166 }
167 }
168 uatomic_inc(&nthreadsrunning);
169 while (goflag == GOFLAG_INIT)
170 poll(NULL, 0, 1);
171 while (goflag == GOFLAG_RUN) {
172 synchronize_rcu();
173 n_updates_local++;
174 }
175 __get_thread_var(n_updates_pt) += n_updates_local;
176 return NULL;
177 }
178
179 void perftestinit(void)
180 {
181 init_per_thread(n_reads_pt, 0LL);
182 init_per_thread(n_updates_pt, 0LL);
183 uatomic_set(&nthreadsrunning, 0);
184 }
185
186 void perftestrun(int nthreads, int nreaders, int nupdaters)
187 {
188 int t;
189 int duration = 1;
190
191 cmm_smp_mb();
192 while (uatomic_read(&nthreadsrunning) < nthreads)
193 poll(NULL, 0, 1);
194 goflag = GOFLAG_RUN;
195 cmm_smp_mb();
196 sleep(duration);
197 cmm_smp_mb();
198 goflag = GOFLAG_STOP;
199 cmm_smp_mb();
200 wait_all_threads();
201 for_each_thread(t) {
202 n_reads += per_thread(n_reads_pt, t);
203 n_updates += per_thread(n_updates_pt, t);
204 }
205 printf("n_reads: %lld n_updates: %ld nreaders: %d nupdaters: %d duration: %d\n",
206 n_reads, n_updates, nreaders, nupdaters, duration);
207 printf("ns/read: %g ns/update: %g\n",
208 ((duration * 1000*1000*1000.*(double)nreaders) /
209 (double)n_reads),
210 ((duration * 1000*1000*1000.*(double)nupdaters) /
211 (double)n_updates));
212 if (get_cpu_call_rcu_data(0)) {
213 fprintf(stderr, "Deallocating per-CPU call_rcu threads.\n");
214 free_all_cpu_call_rcu_data();
215 }
216 exit(0);
217 }
218
219 void perftest(int nreaders, int cpustride)
220 {
221 int i;
222 long arg;
223
224 perftestinit();
225 for (i = 0; i < nreaders; i++) {
226 arg = (long)(i * cpustride);
227 create_thread(rcu_read_perf_test, (void *)arg);
228 }
229 arg = (long)(i * cpustride);
230 create_thread(rcu_update_perf_test, (void *)arg);
231 perftestrun(i + 1, nreaders, 1);
232 }
233
234 void rperftest(int nreaders, int cpustride)
235 {
236 int i;
237 long arg;
238
239 perftestinit();
240 init_per_thread(n_reads_pt, 0LL);
241 for (i = 0; i < nreaders; i++) {
242 arg = (long)(i * cpustride);
243 create_thread(rcu_read_perf_test, (void *)arg);
244 }
245 perftestrun(i, nreaders, 0);
246 }
247
248 void uperftest(int nupdaters, int cpustride)
249 {
250 int i;
251 long arg;
252
253 perftestinit();
254 init_per_thread(n_reads_pt, 0LL);
255 for (i = 0; i < nupdaters; i++) {
256 arg = (long)(i * cpustride);
257 create_thread(rcu_update_perf_test, (void *)arg);
258 }
259 perftestrun(i, 0, nupdaters);
260 }
261
262 /*
263 * Stress test.
264 */
265
266 #define RCU_STRESS_PIPE_LEN 10
267
268 struct rcu_stress {
269 int pipe_count;
270 int mbtest;
271 };
272
273 struct rcu_stress rcu_stress_array[RCU_STRESS_PIPE_LEN] = { { 0 } };
274 struct rcu_stress *rcu_stress_current;
275 int rcu_stress_idx = 0;
276
277 int n_mberror = 0;
278 DEFINE_PER_THREAD(long long [RCU_STRESS_PIPE_LEN + 1], rcu_stress_count);
279
280 int garbage = 0;
281
282 void *rcu_read_stress_test(void *arg)
283 {
284 int i;
285 int itercnt = 0;
286 struct rcu_stress *p;
287 int pc;
288
289 rcu_register_thread();
290 put_thread_offline();
291 while (goflag == GOFLAG_INIT)
292 poll(NULL, 0, 1);
293 put_thread_online();
294 while (goflag == GOFLAG_RUN) {
295 rcu_read_lock();
296 p = rcu_dereference(rcu_stress_current);
297 if (p->mbtest == 0)
298 n_mberror++;
299 rcu_read_lock_nest();
300 for (i = 0; i < 100; i++)
301 garbage++;
302 rcu_read_unlock_nest();
303 pc = p->pipe_count;
304 rcu_read_unlock();
305 if ((pc > RCU_STRESS_PIPE_LEN) || (pc < 0))
306 pc = RCU_STRESS_PIPE_LEN;
307 __get_thread_var(rcu_stress_count)[pc]++;
308 __get_thread_var(n_reads_pt)++;
309 mark_rcu_quiescent_state();
310 if ((++itercnt % 0x1000) == 0) {
311 put_thread_offline();
312 put_thread_online_delay();
313 put_thread_online();
314 }
315 }
316 put_thread_offline();
317 rcu_unregister_thread();
318
319 return (NULL);
320 }
321
322 static pthread_mutex_t call_rcu_test_mutex = PTHREAD_MUTEX_INITIALIZER;
323 static pthread_cond_t call_rcu_test_cond = PTHREAD_COND_INITIALIZER;
324
325 void rcu_update_stress_test_rcu(struct rcu_head *head)
326 {
327 if (pthread_mutex_lock(&call_rcu_test_mutex) != 0) {
328 perror("pthread_mutex_lock");
329 exit(-1);
330 }
331 if (pthread_cond_signal(&call_rcu_test_cond) != 0) {
332 perror("pthread_cond_signal");
333 exit(-1);
334 }
335 if (pthread_mutex_unlock(&call_rcu_test_mutex) != 0) {
336 perror("pthread_mutex_unlock");
337 exit(-1);
338 }
339 }
340
341 void *rcu_update_stress_test(void *arg)
342 {
343 int i;
344 struct rcu_stress *p;
345 struct rcu_head rh;
346
347 while (goflag == GOFLAG_INIT)
348 poll(NULL, 0, 1);
349 while (goflag == GOFLAG_RUN) {
350 i = rcu_stress_idx + 1;
351 if (i >= RCU_STRESS_PIPE_LEN)
352 i = 0;
353 p = &rcu_stress_array[i];
354 p->mbtest = 0;
355 cmm_smp_mb();
356 p->pipe_count = 0;
357 p->mbtest = 1;
358 rcu_assign_pointer(rcu_stress_current, p);
359 rcu_stress_idx = i;
360 for (i = 0; i < RCU_STRESS_PIPE_LEN; i++)
361 if (i != rcu_stress_idx)
362 rcu_stress_array[i].pipe_count++;
363 if (n_updates & 0x1)
364 synchronize_rcu();
365 else {
366 if (pthread_mutex_lock(&call_rcu_test_mutex) != 0) {
367 perror("pthread_mutex_lock");
368 exit(-1);
369 }
370 call_rcu(&rh, rcu_update_stress_test_rcu);
371 if (pthread_cond_wait(&call_rcu_test_cond,
372 &call_rcu_test_mutex) != 0) {
373 perror("pthread_cond_wait");
374 exit(-1);
375 }
376 if (pthread_mutex_unlock(&call_rcu_test_mutex) != 0) {
377 perror("pthread_mutex_unlock");
378 exit(-1);
379 }
380 }
381 n_updates++;
382 }
383 return NULL;
384 }
385
386 void *rcu_fake_update_stress_test(void *arg)
387 {
388 if ((random() & 0xf00) == 0) {
389 struct call_rcu_data *crdp;
390
391 crdp = create_call_rcu_data(0, -1);
392 if (crdp != NULL) {
393 fprintf(stderr,
394 "Using per-thread call_rcu() worker.\n");
395 set_thread_call_rcu_data(crdp);
396 }
397 }
398 while (goflag == GOFLAG_INIT)
399 poll(NULL, 0, 1);
400 while (goflag == GOFLAG_RUN) {
401 synchronize_rcu();
402 poll(NULL, 0, 1);
403 }
404 return NULL;
405 }
406
407 void stresstest(int nreaders)
408 {
409 int i;
410 int t;
411 long long *p;
412 long long sum;
413
414 init_per_thread(n_reads_pt, 0LL);
415 for_each_thread(t) {
416 p = &per_thread(rcu_stress_count,t)[0];
417 for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++)
418 p[i] = 0LL;
419 }
420 rcu_stress_current = &rcu_stress_array[0];
421 rcu_stress_current->pipe_count = 0;
422 rcu_stress_current->mbtest = 1;
423 for (i = 0; i < nreaders; i++)
424 create_thread(rcu_read_stress_test, NULL);
425 create_thread(rcu_update_stress_test, NULL);
426 for (i = 0; i < 5; i++)
427 create_thread(rcu_fake_update_stress_test, NULL);
428 cmm_smp_mb();
429 goflag = GOFLAG_RUN;
430 cmm_smp_mb();
431 sleep(10);
432 cmm_smp_mb();
433 goflag = GOFLAG_STOP;
434 cmm_smp_mb();
435 wait_all_threads();
436 for_each_thread(t)
437 n_reads += per_thread(n_reads_pt, t);
438 printf("n_reads: %lld n_updates: %ld n_mberror: %d\n",
439 n_reads, n_updates, n_mberror);
440 printf("rcu_stress_count:");
441 for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++) {
442 sum = 0LL;
443 for_each_thread(t) {
444 sum += per_thread(rcu_stress_count, t)[i];
445 }
446 printf(" %lld", sum);
447 }
448 printf("\n");
449 if (get_cpu_call_rcu_data(0)) {
450 fprintf(stderr, "Deallocating per-CPU call_rcu threads.\n");
451 free_all_cpu_call_rcu_data();
452 }
453 exit(0);
454 }
455
456 /*
457 * Mainprogram.
458 */
459
460 void usage(int argc, char *argv[])
461 {
462 fprintf(stderr, "Usage: %s [nreaders [ perf | stress ] ]\n", argv[0]);
463 exit(-1);
464 }
465
466 int main(int argc, char *argv[])
467 {
468 int nreaders = 1;
469 int cpustride = 1;
470
471 smp_init();
472 //rcu_init();
473 srandom(time(NULL));
474 if (random() & 0x100) {
475 fprintf(stderr, "Allocating per-CPU call_rcu threads.\n");
476 if (create_all_cpu_call_rcu_data(0))
477 perror("create_all_cpu_call_rcu_data");
478 }
479
480 #ifdef DEBUG_YIELD
481 yield_active |= YIELD_READ;
482 yield_active |= YIELD_WRITE;
483 #endif
484
485 if (argc > 1) {
486 nreaders = strtoul(argv[1], NULL, 0);
487 if (argc == 2)
488 perftest(nreaders, cpustride);
489 if (argc > 3)
490 cpustride = strtoul(argv[3], NULL, 0);
491 if (strcmp(argv[2], "perf") == 0)
492 perftest(nreaders, cpustride);
493 else if (strcmp(argv[2], "rperf") == 0)
494 rperftest(nreaders, cpustride);
495 else if (strcmp(argv[2], "uperf") == 0)
496 uperftest(nreaders, cpustride);
497 else if (strcmp(argv[2], "stress") == 0)
498 stresstest(nreaders);
499 usage(argc, argv);
500 }
501 perftest(nreaders, cpustride);
502 return 0;
503 }
This page took 0.041262 seconds and 4 git commands to generate.