tests/regression/rcutorture: Add wait state
[userspace-rcu.git] / tests / regression / 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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
60 *
61 * Copyright (c) 2008 Paul E. McKenney, IBM Corporation.
62 */
63
64 /*
65 * Test variables.
66 */
67
68 #include <stdlib.h>
69 #include "tap.h"
70
71 #define NR_TESTS 1
72
73 DEFINE_PER_THREAD(long long, n_reads_pt);
74 DEFINE_PER_THREAD(long long, n_updates_pt);
75
76 enum callrcu_type {
77 CALLRCU_GLOBAL,
78 CALLRCU_PERCPU,
79 CALLRCU_PERTHREAD,
80 };
81
82 enum writer_state {
83 WRITER_STATE_SYNC_RCU,
84 WRITER_STATE_CALL_RCU,
85 WRITER_STATE_POLL_RCU,
86 };
87
88 static enum callrcu_type callrcu_type = CALLRCU_GLOBAL;
89
90 long long n_reads = 0LL;
91 long n_updates = 0L;
92 int nthreadsrunning;
93 char argsbuf[64];
94
95 #define GOFLAG_INIT 0
96 #define GOFLAG_RUN 1
97 #define GOFLAG_STOP 2
98
99 volatile int goflag __attribute__((__aligned__(CAA_CACHE_LINE_SIZE)))
100 = GOFLAG_INIT;
101
102 #define RCU_READ_RUN 1000
103
104 //MD
105 #define RCU_READ_NESTABLE
106
107 #ifdef RCU_READ_NESTABLE
108 #define rcu_read_lock_nest() rcu_read_lock()
109 #define rcu_read_unlock_nest() rcu_read_unlock()
110 #else /* #ifdef RCU_READ_NESTABLE */
111 #define rcu_read_lock_nest()
112 #define rcu_read_unlock_nest()
113 #endif /* #else #ifdef RCU_READ_NESTABLE */
114
115 #ifdef TORTURE_QSBR
116 #define mark_rcu_quiescent_state rcu_quiescent_state
117 #define put_thread_offline rcu_thread_offline
118 #define put_thread_online rcu_thread_online
119 #endif
120
121 #ifndef mark_rcu_quiescent_state
122 #define mark_rcu_quiescent_state() do {} while (0)
123 #endif /* #ifdef mark_rcu_quiescent_state */
124
125 #ifndef put_thread_offline
126 #define put_thread_offline() do {} while (0)
127 #define put_thread_online() do {} while (0)
128 #define put_thread_online_delay() do {} while (0)
129 #else /* #ifndef put_thread_offline */
130 #define put_thread_online_delay() synchronize_rcu()
131 #endif /* #else #ifndef put_thread_offline */
132
133 /*
134 * Performance test.
135 */
136
137 static
138 void *rcu_read_perf_test(void *arg)
139 {
140 int i;
141 int me = (long)arg;
142 long long n_reads_local = 0;
143
144 rcu_register_thread();
145 run_on(me);
146 uatomic_inc(&nthreadsrunning);
147 put_thread_offline();
148 while (goflag == GOFLAG_INIT)
149 (void) poll(NULL, 0, 1);
150 put_thread_online();
151 while (goflag == GOFLAG_RUN) {
152 for (i = 0; i < RCU_READ_RUN; i++) {
153 rcu_read_lock();
154 /* rcu_read_lock_nest(); */
155 /* rcu_read_unlock_nest(); */
156 rcu_read_unlock();
157 }
158 n_reads_local += RCU_READ_RUN;
159 mark_rcu_quiescent_state();
160 }
161 __get_thread_var(n_reads_pt) += n_reads_local;
162 put_thread_offline();
163 rcu_unregister_thread();
164
165 return (NULL);
166 }
167
168 static
169 void *rcu_update_perf_test(void *arg __attribute__((unused)))
170 {
171 long long n_updates_local = 0;
172
173 if (callrcu_type == CALLRCU_PERTHREAD) {
174 struct call_rcu_data *crdp;
175
176 crdp = create_call_rcu_data(0, -1);
177 if (crdp != NULL) {
178 diag("Successfully using per-thread call_rcu() worker.");
179 set_thread_call_rcu_data(crdp);
180 }
181 }
182 uatomic_inc(&nthreadsrunning);
183 while (goflag == GOFLAG_INIT)
184 (void) poll(NULL, 0, 1);
185 while (goflag == GOFLAG_RUN) {
186 synchronize_rcu();
187 n_updates_local++;
188 }
189 __get_thread_var(n_updates_pt) += n_updates_local;
190 if (callrcu_type == CALLRCU_PERTHREAD) {
191 struct call_rcu_data *crdp;
192
193 crdp = get_thread_call_rcu_data();
194 set_thread_call_rcu_data(NULL);
195 call_rcu_data_free(crdp);
196 }
197 return NULL;
198 }
199
200 static
201 void perftestinit(void)
202 {
203 init_per_thread(n_reads_pt, 0LL);
204 init_per_thread(n_updates_pt, 0LL);
205 uatomic_set(&nthreadsrunning, 0);
206 }
207
208 static
209 int perftestrun(int nthreads, int nreaders, int nupdaters)
210 {
211 int t;
212 int duration = 1;
213
214 cmm_smp_mb();
215 while (uatomic_read(&nthreadsrunning) < nthreads)
216 (void) poll(NULL, 0, 1);
217 goflag = GOFLAG_RUN;
218 cmm_smp_mb();
219 sleep(duration);
220 cmm_smp_mb();
221 goflag = GOFLAG_STOP;
222 cmm_smp_mb();
223 wait_all_threads();
224 for_each_thread(t) {
225 n_reads += per_thread(n_reads_pt, t);
226 n_updates += per_thread(n_updates_pt, t);
227 }
228 diag("n_reads: %lld n_updates: %ld nreaders: %d nupdaters: %d duration: %d",
229 n_reads, n_updates, nreaders, nupdaters, duration);
230 diag("ns/read: %g ns/update: %g",
231 ((duration * 1000*1000*1000.*(double)nreaders) /
232 (double)n_reads),
233 ((duration * 1000*1000*1000.*(double)nupdaters) /
234 (double)n_updates));
235 if (get_cpu_call_rcu_data(0)) {
236 diag("Deallocating per-CPU call_rcu threads.\n");
237 free_all_cpu_call_rcu_data();
238 }
239 return 0;
240 }
241
242 static
243 int perftest(int nreaders, int cpustride)
244 {
245 int i;
246 long arg;
247
248 perftestinit();
249 for (i = 0; i < nreaders; i++) {
250 arg = (long)(i * cpustride);
251 create_thread(rcu_read_perf_test, (void *)arg);
252 }
253 arg = (long)(i * cpustride);
254 create_thread(rcu_update_perf_test, (void *)arg);
255 return perftestrun(i + 1, nreaders, 1);
256 }
257
258 static
259 int rperftest(int nreaders, int cpustride)
260 {
261 int i;
262 long arg;
263
264 perftestinit();
265 init_per_thread(n_reads_pt, 0LL);
266 for (i = 0; i < nreaders; i++) {
267 arg = (long)(i * cpustride);
268 create_thread(rcu_read_perf_test, (void *)arg);
269 }
270 return perftestrun(i, nreaders, 0);
271 }
272
273 static
274 int uperftest(int nupdaters, int cpustride)
275 {
276 int i;
277 long arg;
278
279 perftestinit();
280 init_per_thread(n_reads_pt, 0LL);
281 for (i = 0; i < nupdaters; i++) {
282 arg = (long)(i * cpustride);
283 create_thread(rcu_update_perf_test, (void *)arg);
284 }
285 return perftestrun(i, 0, nupdaters);
286 }
287
288 /*
289 * Stress test.
290 */
291
292 #define RCU_STRESS_PIPE_LEN 10
293
294 struct rcu_stress {
295 int pipe_count;
296 int mbtest;
297 };
298
299 struct rcu_stress rcu_stress_array[RCU_STRESS_PIPE_LEN] = { { 0, 0 } };
300 struct rcu_stress *rcu_stress_current;
301 int rcu_stress_idx = 0;
302
303 int n_mberror = 0;
304 DEFINE_PER_THREAD(long long [RCU_STRESS_PIPE_LEN + 1], rcu_stress_count);
305
306 int garbage = 0;
307
308 static
309 void *rcu_read_stress_test(void *arg __attribute__((unused)))
310 {
311 int i;
312 int itercnt = 0;
313 struct rcu_stress *p;
314 int pc;
315
316 rcu_register_thread();
317 put_thread_offline();
318 while (goflag == GOFLAG_INIT)
319 (void) poll(NULL, 0, 1);
320 put_thread_online();
321 while (goflag == GOFLAG_RUN) {
322 rcu_read_lock();
323 p = rcu_dereference(rcu_stress_current);
324 if (p->mbtest == 0)
325 n_mberror++;
326 rcu_read_lock_nest();
327 for (i = 0; i < 100; i++)
328 garbage++;
329 rcu_read_unlock_nest();
330 pc = p->pipe_count;
331 rcu_read_unlock();
332 if ((pc > RCU_STRESS_PIPE_LEN) || (pc < 0))
333 pc = RCU_STRESS_PIPE_LEN;
334 __get_thread_var(rcu_stress_count)[pc]++;
335 __get_thread_var(n_reads_pt)++;
336 mark_rcu_quiescent_state();
337 if ((++itercnt % 0x1000) == 0) {
338 put_thread_offline();
339 put_thread_online_delay();
340 put_thread_online();
341 }
342 }
343 put_thread_offline();
344 rcu_unregister_thread();
345
346 return (NULL);
347 }
348
349 static pthread_mutex_t call_rcu_test_mutex = PTHREAD_MUTEX_INITIALIZER;
350 static pthread_cond_t call_rcu_test_cond = PTHREAD_COND_INITIALIZER;
351 static bool call_rcu_wait;
352
353 static
354 void rcu_update_stress_test_rcu(struct rcu_head *head __attribute__((unused)))
355 {
356 int ret;
357
358 ret = pthread_mutex_lock(&call_rcu_test_mutex);
359 if (ret) {
360 errno = ret;
361 diag("pthread_mutex_lock: %s",
362 strerror(errno));
363 abort();
364 }
365 ret = pthread_cond_signal(&call_rcu_test_cond);
366 if (ret) {
367 errno = ret;
368 diag("pthread_cond_signal: %s",
369 strerror(errno));
370 abort();
371 }
372 call_rcu_wait = false;
373 ret = pthread_mutex_unlock(&call_rcu_test_mutex);
374 if (ret) {
375 errno = ret;
376 diag("pthread_mutex_unlock: %s",
377 strerror(errno));
378 abort();
379 }
380 }
381
382 static
383 void advance_writer_state(enum writer_state *state)
384 {
385 switch (*state) {
386 case WRITER_STATE_SYNC_RCU:
387 *state = WRITER_STATE_CALL_RCU;
388 break;
389 case WRITER_STATE_CALL_RCU:
390 *state = WRITER_STATE_POLL_RCU;
391 break;
392 case WRITER_STATE_POLL_RCU:
393 *state = WRITER_STATE_SYNC_RCU;
394 break;
395 }
396 }
397
398 static
399 void *rcu_update_stress_test(void *arg __attribute__((unused)))
400 {
401 int i;
402 struct rcu_stress *p;
403 struct rcu_head rh;
404 enum writer_state writer_state = WRITER_STATE_SYNC_RCU;
405
406 while (goflag == GOFLAG_INIT)
407 (void) poll(NULL, 0, 1);
408 while (goflag == GOFLAG_RUN) {
409 i = rcu_stress_idx + 1;
410 if (i >= RCU_STRESS_PIPE_LEN)
411 i = 0;
412 p = &rcu_stress_array[i];
413 p->mbtest = 0;
414 cmm_smp_mb();
415 p->pipe_count = 0;
416 p->mbtest = 1;
417 rcu_assign_pointer(rcu_stress_current, p);
418 rcu_stress_idx = i;
419 for (i = 0; i < RCU_STRESS_PIPE_LEN; i++)
420 if (i != rcu_stress_idx)
421 rcu_stress_array[i].pipe_count++;
422 switch (writer_state) {
423 case WRITER_STATE_SYNC_RCU:
424 synchronize_rcu();
425 break;
426 case WRITER_STATE_CALL_RCU:
427 {
428 int ret;
429
430 ret = pthread_mutex_lock(&call_rcu_test_mutex);
431 if (ret) {
432 errno = ret;
433 diag("pthread_mutex_lock: %s",
434 strerror(errno));
435 abort();
436 }
437 rcu_register_thread();
438 call_rcu(&rh, rcu_update_stress_test_rcu);
439 rcu_unregister_thread();
440 /*
441 * Our MacOS X test machine with the following
442 * config:
443 * 15.6.0 Darwin Kernel Version 15.6.0
444 * root:xnu-3248.60.10~1/RELEASE_X86_64
445 * appears to have issues with liburcu-signal
446 * signal being delivered on top of
447 * pthread_cond_wait. It seems to make the
448 * thread continue, and therefore corrupt the
449 * rcu_head. Work around this issue by
450 * unregistering the RCU read-side thread
451 * immediately after call_rcu (call_rcu needs
452 * us to be registered RCU readers).
453 */
454 call_rcu_wait = true;
455 do {
456 ret = pthread_cond_wait(&call_rcu_test_cond,
457 &call_rcu_test_mutex);
458 } while (call_rcu_wait);
459 if (ret) {
460 errno = ret;
461 diag("pthread_cond_signal: %s",
462 strerror(errno));
463 abort();
464 }
465 ret = pthread_mutex_unlock(&call_rcu_test_mutex);
466 if (ret) {
467 errno = ret;
468 diag("pthread_mutex_unlock: %s",
469 strerror(errno));
470 abort();
471 }
472 break;
473 }
474 case WRITER_STATE_POLL_RCU:
475 {
476 struct urcu_gp_poll_state poll_state;
477
478 rcu_register_thread();
479 poll_state = start_poll_synchronize_rcu();
480 rcu_unregister_thread();
481 while (!poll_state_synchronize_rcu(poll_state))
482 (void) poll(NULL, 0, 1); /* Wait for 1ms */
483 break;
484 }
485 }
486 n_updates++;
487 advance_writer_state(&writer_state);
488 }
489
490 return NULL;
491 }
492
493 static
494 void *rcu_fake_update_stress_test(void *arg __attribute__((unused)))
495 {
496 if (callrcu_type == CALLRCU_PERTHREAD) {
497 struct call_rcu_data *crdp;
498
499 crdp = create_call_rcu_data(0, -1);
500 if (crdp != NULL) {
501 diag("Successfully using per-thread call_rcu() worker.");
502 set_thread_call_rcu_data(crdp);
503 }
504 }
505 while (goflag == GOFLAG_INIT)
506 (void) poll(NULL, 0, 1);
507 while (goflag == GOFLAG_RUN) {
508 synchronize_rcu();
509 (void) poll(NULL, 0, 1);
510 }
511 if (callrcu_type == CALLRCU_PERTHREAD) {
512 struct call_rcu_data *crdp;
513
514 crdp = get_thread_call_rcu_data();
515 set_thread_call_rcu_data(NULL);
516 call_rcu_data_free(crdp);
517 }
518 return NULL;
519 }
520
521 static
522 int stresstest(int nreaders)
523 {
524 int i;
525 int t;
526 long long *p;
527 long long sum;
528
529 init_per_thread(n_reads_pt, 0LL);
530 for_each_thread(t) {
531 p = &per_thread(rcu_stress_count,t)[0];
532 for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++)
533 p[i] = 0LL;
534 }
535 rcu_stress_current = &rcu_stress_array[0];
536 rcu_stress_current->pipe_count = 0;
537 rcu_stress_current->mbtest = 1;
538 for (i = 0; i < nreaders; i++)
539 create_thread(rcu_read_stress_test, NULL);
540 create_thread(rcu_update_stress_test, NULL);
541 for (i = 0; i < 5; i++)
542 create_thread(rcu_fake_update_stress_test, NULL);
543 cmm_smp_mb();
544 goflag = GOFLAG_RUN;
545 cmm_smp_mb();
546 sleep(10);
547 cmm_smp_mb();
548 goflag = GOFLAG_STOP;
549 cmm_smp_mb();
550 wait_all_threads();
551 for_each_thread(t)
552 n_reads += per_thread(n_reads_pt, t);
553 diag("n_reads: %lld n_updates: %ld n_mberror: %d",
554 n_reads, n_updates, n_mberror);
555 rdiag_start();
556 rdiag("rcu_stress_count:");
557 for (i = 0; i <= RCU_STRESS_PIPE_LEN; i++) {
558 sum = 0LL;
559 for_each_thread(t) {
560 sum += per_thread(rcu_stress_count, t)[i];
561 }
562 rdiag(" %lld", sum);
563 }
564 rdiag_end();
565 if (get_cpu_call_rcu_data(0)) {
566 diag("Deallocating per-CPU call_rcu threads.");
567 free_all_cpu_call_rcu_data();
568 }
569 if (!n_mberror)
570 return 0;
571 else
572 return -1;
573 }
574
575 /*
576 * Mainprogram.
577 */
578
579 static
580 void usage(char *argv[]) __attribute__((__noreturn__));
581
582 static
583 void usage(char *argv[])
584 {
585 diag("Usage: %s nreaders [ perf | rperf | uperf | stress ] [ stride ] [ callrcu_global | callrcu_percpu | callrcu_perthread ]\n", argv[0]);
586 exit(-1);
587 }
588
589 int main(int argc, char *argv[])
590 {
591 int nreaders = 1;
592 int cpustride = 1;
593
594 plan_tests(NR_TESTS);
595
596 smp_init();
597 //rcu_init();
598 if (argc > 4) {
599 const char *callrcu_str = argv[4];;
600
601 if (strcmp(callrcu_str, "callrcu_global") == 0) {
602 callrcu_type = CALLRCU_GLOBAL;
603 } else if (strcmp(callrcu_str, "callrcu_percpu") == 0) {
604 callrcu_type = CALLRCU_PERCPU;
605 } else if (strcmp(callrcu_str, "callrcu_perthread") == 0) {
606 callrcu_type = CALLRCU_PERTHREAD;
607 } else {
608 usage(argv);
609 goto end;
610 }
611 }
612
613 switch (callrcu_type) {
614 case CALLRCU_GLOBAL:
615 diag("Using global per-process call_rcu thread.");
616 break;
617 case CALLRCU_PERCPU:
618 diag("Using per-CPU call_rcu threads.");
619 if (create_all_cpu_call_rcu_data(0))
620 diag("create_all_cpu_call_rcu_data: %s",
621 strerror(errno));
622 break;
623 case CALLRCU_PERTHREAD:
624 diag("Using per-thread call_rcu() worker.");
625 break;
626 default:
627 abort();
628 }
629
630 #ifdef DEBUG_YIELD
631 yield_active |= YIELD_READ;
632 yield_active |= YIELD_WRITE;
633 #endif
634
635 if (argc > 1) {
636 if (strcmp(argv[1], "-h") == 0
637 || strcmp(argv[1], "--help") == 0) {
638 usage(argv);
639 goto end;
640 }
641 nreaders = strtoul(argv[1], NULL, 0);
642 if (argc == 2) {
643 ok(!perftest(nreaders, cpustride),
644 "perftest readers: %d, stride: %d",
645 nreaders, cpustride);
646 goto end;
647 }
648 if (argc > 3)
649 cpustride = strtoul(argv[3], NULL, 0);
650 if (strcmp(argv[2], "perf") == 0)
651 ok(!perftest(nreaders, cpustride),
652 "perftest readers: %d, stride: %d",
653 nreaders, cpustride);
654 else if (strcmp(argv[2], "rperf") == 0)
655 ok(!rperftest(nreaders, cpustride),
656 "rperftest readers: %d, stride: %d",
657 nreaders, cpustride);
658 else if (strcmp(argv[2], "uperf") == 0)
659 ok(!uperftest(nreaders, cpustride),
660 "uperftest readers: %d, stride: %d",
661 nreaders, cpustride);
662 else if (strcmp(argv[2], "stress") == 0)
663 ok(!stresstest(nreaders),
664 "stresstest readers: %d, stride: %d",
665 nreaders, cpustride);
666 else
667 usage(argv);
668 } else {
669 usage(argv);
670 }
671 end:
672 return exit_status();
673 }
This page took 0.068824 seconds and 4 git commands to generate.