tests: Use uatomic for accessing global states
[urcu.git] / tests / benchmark / test_urcu_wfcq.c
CommitLineData
ce29b371
MJ
1// SPDX-FileCopyrightText: 2010 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
2// SPDX-FileCopyrightText: 2010 Paolo Bonzini <pbonzini@redhat.com>
3//
4// SPDX-License-Identifier: GPL-2.0-or-later
5
5e70d572 6/*
5e70d572 7 * Userspace RCU library - example RCU-based lock-free concurrent queue
5e70d572
MD
8 */
9
5e70d572
MD
10#include <stdio.h>
11#include <pthread.h>
12#include <stdlib.h>
13#include <stdint.h>
14#include <stdbool.h>
15#include <string.h>
16#include <sys/types.h>
17#include <sys/wait.h>
18#include <unistd.h>
19#include <stdio.h>
5e70d572
MD
20#include <errno.h>
21
22#include <urcu/arch.h>
01477510 23#include <urcu/assert.h>
5e70d572 24#include <urcu/tls-compat.h>
985c3168 25#include <urcu/uatomic.h>
94df6318 26#include "thread-id.h"
5e70d572
MD
27
28/* hardcoded number of CPUs */
29#define NR_CPUS 16384
30
5e70d572
MD
31#ifndef DYNAMIC_LINK_TEST
32#define _LGPL_SOURCE
33#endif
5e70d572
MD
34#include <urcu/wfcqueue.h>
35
f5d36fed 36enum test_sync {
372b9ad7
MD
37 TEST_SYNC_NONE = 0,
38 TEST_SYNC_MUTEX,
f5d36fed
MD
39};
40
41static enum test_sync test_sync;
42
372b9ad7
MD
43static int test_force_sync;
44
cf9451dc 45static volatile int test_go, test_stop_enqueue, test_stop_dequeue;
5e70d572
MD
46
47static unsigned long rduration;
48
49static unsigned long duration;
50
51/* read-side C.S. duration, in loops */
52static unsigned long wdelay;
53
ab0aacbe 54static inline void loop_sleep(unsigned long loops)
5e70d572 55{
ab0aacbe 56 while (loops-- != 0)
5e70d572
MD
57 caa_cpu_relax();
58}
59
60static int verbose_mode;
61
cf9451dc 62static int test_dequeue, test_splice, test_wait_empty;
83e334d0 63static unsigned int test_enqueue_stopped;
f5d36fed 64
5e70d572
MD
65#define printf_verbose(fmt, args...) \
66 do { \
67 if (verbose_mode) \
f5d36fed 68 printf(fmt, ## args); \
5e70d572
MD
69 } while (0)
70
71static unsigned int cpu_affinities[NR_CPUS];
72static unsigned int next_aff = 0;
73static int use_affinity = 0;
74
75pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
76
5e70d572
MD
77static void set_affinity(void)
78{
2388c075 79#ifdef HAVE_SCHED_SETAFFINITY
5e70d572 80 cpu_set_t mask;
95bc7fb9
MD
81 int cpu, ret;
82#endif /* HAVE_SCHED_SETAFFINITY */
5e70d572
MD
83
84 if (!use_affinity)
85 return;
86
2388c075 87#ifdef HAVE_SCHED_SETAFFINITY
5e70d572
MD
88 ret = pthread_mutex_lock(&affinity_mutex);
89 if (ret) {
90 perror("Error in pthread mutex lock");
91 exit(-1);
92 }
93 cpu = cpu_affinities[next_aff++];
94 ret = pthread_mutex_unlock(&affinity_mutex);
95 if (ret) {
96 perror("Error in pthread mutex unlock");
97 exit(-1);
98 }
99
100 CPU_ZERO(&mask);
101 CPU_SET(cpu, &mask);
5e70d572 102 sched_setaffinity(0, sizeof(mask), &mask);
5e70d572
MD
103#endif /* HAVE_SCHED_SETAFFINITY */
104}
105
106/*
107 * returns 0 if test should end.
108 */
109static int test_duration_dequeue(void)
110{
cf9451dc 111 return !test_stop_dequeue;
5e70d572
MD
112}
113
114static int test_duration_enqueue(void)
115{
cf9451dc 116 return !test_stop_enqueue;
5e70d572
MD
117}
118
119static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
120static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
121
122static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
123static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
cf9451dc
MD
124static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
125static DEFINE_URCU_TLS(unsigned long long, nr_splice);
db182e31 126static DEFINE_URCU_TLS(unsigned long long, nr_dequeue_last);
5e70d572
MD
127
128static unsigned int nr_enqueuers;
129static unsigned int nr_dequeuers;
130
131static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
132static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
133
f5d36fed 134static void *thr_enqueuer(void *_count)
5e70d572
MD
135{
136 unsigned long long *count = _count;
cf9451dc 137 bool was_nonempty;
5e70d572 138
94df6318
MD
139 printf_verbose("thread_begin %s, tid %lu\n",
140 "enqueuer", urcu_get_thread_id());
5e70d572
MD
141
142 set_affinity();
143
144 while (!test_go)
145 {
146 }
147 cmm_smp_mb();
148
149 for (;;) {
150 struct cds_wfcq_node *node = malloc(sizeof(*node));
151 if (!node)
152 goto fail;
153 cds_wfcq_node_init(node);
cf9451dc 154 was_nonempty = cds_wfcq_enqueue(&head, &tail, node);
5e70d572 155 URCU_TLS(nr_successful_enqueues)++;
cf9451dc
MD
156 if (!was_nonempty)
157 URCU_TLS(nr_empty_dest_enqueues)++;
5e70d572
MD
158
159 if (caa_unlikely(wdelay))
160 loop_sleep(wdelay);
161fail:
162 URCU_TLS(nr_enqueues)++;
163 if (caa_unlikely(!test_duration_enqueue()))
164 break;
165 }
166
cf9451dc 167 uatomic_inc(&test_enqueue_stopped);
5e70d572
MD
168 count[0] = URCU_TLS(nr_enqueues);
169 count[1] = URCU_TLS(nr_successful_enqueues);
cf9451dc 170 count[2] = URCU_TLS(nr_empty_dest_enqueues);
94df6318
MD
171 printf_verbose("enqueuer thread_end, tid %lu, "
172 "enqueues %llu successful_enqueues %llu, "
173 "empty_dest_enqueues %llu\n",
174 urcu_get_thread_id(),
175 URCU_TLS(nr_enqueues),
176 URCU_TLS(nr_successful_enqueues),
177 URCU_TLS(nr_empty_dest_enqueues));
5e70d572
MD
178 return ((void*)1);
179
180}
181
f5d36fed
MD
182static void do_test_dequeue(enum test_sync sync)
183{
184 struct cds_wfcq_node *node;
db182e31 185 int state;
f5d36fed
MD
186
187 if (sync == TEST_SYNC_MUTEX)
db182e31
MD
188 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
189 &state);
f5d36fed 190 else
db182e31
MD
191 node = __cds_wfcq_dequeue_with_state_blocking(&head, &tail,
192 &state);
193
194 if (state & CDS_WFCQ_STATE_LAST)
195 URCU_TLS(nr_dequeue_last)++;
f5d36fed
MD
196
197 if (node) {
198 free(node);
199 URCU_TLS(nr_successful_dequeues)++;
200 }
201 URCU_TLS(nr_dequeues)++;
202}
203
204static void do_test_splice(enum test_sync sync)
205{
206 struct cds_wfcq_head tmp_head;
207 struct cds_wfcq_tail tmp_tail;
208 struct cds_wfcq_node *node, *n;
cf9451dc 209 enum cds_wfcq_ret ret;
f5d36fed
MD
210
211 cds_wfcq_init(&tmp_head, &tmp_tail);
212
213 if (sync == TEST_SYNC_MUTEX)
cf9451dc 214 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
f5d36fed
MD
215 &head, &tail);
216 else
cf9451dc 217 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
f5d36fed
MD
218 &head, &tail);
219
cf9451dc
MD
220 switch (ret) {
221 case CDS_WFCQ_RET_WOULDBLOCK:
01477510 222 urcu_posix_assert(0); /* blocking call */
cf9451dc
MD
223 break;
224 case CDS_WFCQ_RET_DEST_EMPTY:
225 URCU_TLS(nr_splice)++;
db182e31 226 URCU_TLS(nr_dequeue_last)++;
cf9451dc
MD
227 /* ok */
228 break;
229 case CDS_WFCQ_RET_DEST_NON_EMPTY:
01477510 230 urcu_posix_assert(0); /* entirely unexpected */
cf9451dc
MD
231 break;
232 case CDS_WFCQ_RET_SRC_EMPTY:
233 /* ok, we could even skip iteration on dest if we wanted */
234 break;
235 }
236
f5d36fed
MD
237 __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
238 free(node);
239 URCU_TLS(nr_successful_dequeues)++;
240 URCU_TLS(nr_dequeues)++;
241 }
2af1c19e 242 cds_wfcq_destroy(&tmp_head, &tmp_tail);
f5d36fed
MD
243}
244
245static void *thr_dequeuer(void *_count)
5e70d572
MD
246{
247 unsigned long long *count = _count;
f42bc7ef 248 unsigned int counter = 0;
5e70d572 249
94df6318
MD
250 printf_verbose("thread_begin %s, tid %lu\n",
251 "dequeuer", urcu_get_thread_id());
5e70d572
MD
252
253 set_affinity();
254
255 while (!test_go)
256 {
257 }
258 cmm_smp_mb();
259
260 for (;;) {
f5d36fed
MD
261 if (test_dequeue && test_splice) {
262 if (counter & 1)
263 do_test_dequeue(test_sync);
264 else
265 do_test_splice(test_sync);
266 counter++;
267 } else {
268 if (test_dequeue)
269 do_test_dequeue(test_sync);
270 else
271 do_test_splice(test_sync);
5e70d572 272 }
5e70d572
MD
273 if (caa_unlikely(!test_duration_dequeue()))
274 break;
275 if (caa_unlikely(rduration))
276 loop_sleep(rduration);
277 }
278
94df6318
MD
279 printf_verbose("dequeuer thread_end, tid %lu, "
280 "dequeues %llu, successful_dequeues %llu, "
281 "nr_splice %llu\n",
282 urcu_get_thread_id(),
283 URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
284 URCU_TLS(nr_splice));
5e70d572
MD
285 count[0] = URCU_TLS(nr_dequeues);
286 count[1] = URCU_TLS(nr_successful_dequeues);
cf9451dc 287 count[2] = URCU_TLS(nr_splice);
db182e31 288 count[3] = URCU_TLS(nr_dequeue_last);
5e70d572
MD
289 return ((void*)2);
290}
291
bd23a6c6
MJ
292static void test_end(unsigned long long *nr_dequeues_l,
293 unsigned long long *nr_dequeue_last_l)
5e70d572
MD
294{
295 struct cds_wfcq_node *node;
db182e31 296 int state;
5e70d572
MD
297
298 do {
db182e31
MD
299 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
300 &state);
5e70d572 301 if (node) {
db182e31 302 if (state & CDS_WFCQ_STATE_LAST)
bd23a6c6 303 (*nr_dequeue_last_l)++;
5e70d572 304 free(node);
bd23a6c6 305 (*nr_dequeues_l)++;
5e70d572
MD
306 }
307 } while (node);
308}
309
70469b43 310static void show_usage(char **argv)
5e70d572 311{
06637138
MD
312 printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
313 argv[0]);
314 printf("OPTIONS:\n");
315 printf(" [-d delay] (enqueuer period (in loops))\n");
316 printf(" [-c duration] (dequeuer period (in loops))\n");
317 printf(" [-v] (verbose output)\n");
318 printf(" [-a cpu#] [-a cpu#]... (affinity)\n");
319 printf(" [-q] (test dequeue)\n");
320 printf(" [-s] (test splice, enabled by default)\n");
321 printf(" [-M] (use mutex external synchronization)\n");
322 printf(" Note: default: no external synchronization used.\n");
323 printf(" [-f] (force user-provided synchronization)\n");
324 printf(" [-w] Wait for dequeuer to empty queue\n");
5e70d572
MD
325 printf("\n");
326}
327
328int main(int argc, char **argv)
329{
330 int err;
331 pthread_t *tid_enqueuer, *tid_dequeuer;
332 void *tret;
333 unsigned long long *count_enqueuer, *count_dequeuer;
334 unsigned long long tot_enqueues = 0, tot_dequeues = 0;
335 unsigned long long tot_successful_enqueues = 0,
cf9451dc
MD
336 tot_successful_dequeues = 0,
337 tot_empty_dest_enqueues = 0,
db182e31 338 tot_splice = 0, tot_dequeue_last = 0;
5e70d572 339 unsigned long long end_dequeues = 0;
cf9451dc 340 int i, a, retval = 0;
83e334d0 341 unsigned int i_thr;
5e70d572
MD
342
343 if (argc < 4) {
70469b43 344 show_usage(argv);
5e70d572
MD
345 return -1;
346 }
347
348 err = sscanf(argv[1], "%u", &nr_dequeuers);
349 if (err != 1) {
70469b43 350 show_usage(argv);
5e70d572
MD
351 return -1;
352 }
353
354 err = sscanf(argv[2], "%u", &nr_enqueuers);
355 if (err != 1) {
70469b43 356 show_usage(argv);
5e70d572
MD
357 return -1;
358 }
83e334d0 359
5e70d572
MD
360 err = sscanf(argv[3], "%lu", &duration);
361 if (err != 1) {
70469b43 362 show_usage(argv);
5e70d572
MD
363 return -1;
364 }
365
366 for (i = 4; i < argc; i++) {
367 if (argv[i][0] != '-')
368 continue;
369 switch (argv[i][1]) {
370 case 'a':
371 if (argc < i + 2) {
70469b43 372 show_usage(argv);
5e70d572
MD
373 return -1;
374 }
375 a = atoi(argv[++i]);
376 cpu_affinities[next_aff++] = a;
377 use_affinity = 1;
378 printf_verbose("Adding CPU %d affinity\n", a);
379 break;
380 case 'c':
381 if (argc < i + 2) {
70469b43 382 show_usage(argv);
5e70d572
MD
383 return -1;
384 }
385 rduration = atol(argv[++i]);
386 break;
387 case 'd':
388 if (argc < i + 2) {
70469b43 389 show_usage(argv);
5e70d572
MD
390 return -1;
391 }
392 wdelay = atol(argv[++i]);
393 break;
394 case 'v':
395 verbose_mode = 1;
396 break;
f5d36fed
MD
397 case 'q':
398 test_dequeue = 1;
399 break;
400 case 's':
401 test_splice = 1;
402 break;
403 case 'M':
404 test_sync = TEST_SYNC_MUTEX;
405 break;
cf9451dc
MD
406 case 'w':
407 test_wait_empty = 1;
408 break;
372b9ad7
MD
409 case 'f':
410 test_force_sync = 1;
411 break;
5e70d572
MD
412 }
413 }
414
f5d36fed
MD
415 /* activate splice test by default */
416 if (!test_dequeue && !test_splice)
417 test_splice = 1;
418
372b9ad7
MD
419 if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_dequeue) {
420 if (test_force_sync) {
421 fprintf(stderr, "[WARNING] Using dequeue concurrently "
422 "with other dequeue or splice without external "
423 "synchronization. Expect run-time failure.\n");
424 } else {
425 printf("Enforcing mutex synchronization\n");
426 test_sync = TEST_SYNC_MUTEX;
427 }
428 }
429
5e70d572
MD
430 printf_verbose("running test for %lu seconds, %u enqueuers, "
431 "%u dequeuers.\n",
432 duration, nr_enqueuers, nr_dequeuers);
f5d36fed
MD
433 if (test_dequeue)
434 printf_verbose("dequeue test activated.\n");
435 else
436 printf_verbose("splice test activated.\n");
437 if (test_sync == TEST_SYNC_MUTEX)
438 printf_verbose("External sync: mutex.\n");
439 else
440 printf_verbose("External sync: none.\n");
cf9451dc
MD
441 if (test_wait_empty)
442 printf_verbose("Wait for dequeuers to empty queue.\n");
5e70d572
MD
443 printf_verbose("Writer delay : %lu loops.\n", rduration);
444 printf_verbose("Reader duration : %lu loops.\n", wdelay);
94df6318
MD
445 printf_verbose("thread %-6s, tid %lu\n",
446 "main", urcu_get_thread_id());
5e70d572 447
9aa14175
MD
448 tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
449 tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
450 count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
451 count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
5e70d572
MD
452 cds_wfcq_init(&head, &tail);
453
454 next_aff = 0;
455
83e334d0
MJ
456 for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
457 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
458 &count_enqueuer[3 * i_thr]);
5e70d572
MD
459 if (err != 0)
460 exit(1);
461 }
83e334d0
MJ
462 for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
463 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
464 &count_dequeuer[4 * i_thr]);
5e70d572
MD
465 if (err != 0)
466 exit(1);
467 }
468
469 cmm_smp_mb();
470
471 test_go = 1;
472
83e334d0 473 for (i_thr = 0; i_thr < duration; i_thr++) {
5e70d572 474 sleep(1);
4b665350
MD
475 if (verbose_mode) {
476 fwrite(".", sizeof(char), 1, stdout);
477 fflush(stdout);
478 }
5e70d572
MD
479 }
480
cf9451dc
MD
481 test_stop_enqueue = 1;
482
483 if (test_wait_empty) {
484 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
485 sleep(1);
486 }
487 while (!cds_wfcq_empty(&head, &tail)) {
488 sleep(1);
489 }
490 }
491
492 test_stop_dequeue = 1;
5e70d572 493
83e334d0
MJ
494 for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
495 err = pthread_join(tid_enqueuer[i_thr], &tret);
5e70d572
MD
496 if (err != 0)
497 exit(1);
83e334d0
MJ
498 tot_enqueues += count_enqueuer[3 * i_thr];
499 tot_successful_enqueues += count_enqueuer[3 * i_thr + 1];
500 tot_empty_dest_enqueues += count_enqueuer[3 * i_thr + 2];
5e70d572 501 }
83e334d0
MJ
502 for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
503 err = pthread_join(tid_dequeuer[i_thr], &tret);
5e70d572
MD
504 if (err != 0)
505 exit(1);
83e334d0
MJ
506 tot_dequeues += count_dequeuer[4 * i_thr];
507 tot_successful_dequeues += count_dequeuer[4 * i_thr + 1];
508 tot_splice += count_dequeuer[4 * i_thr + 2];
509 tot_dequeue_last += count_dequeuer[4 * i_thr + 3];
5e70d572 510 }
83e334d0 511
db182e31 512 test_end(&end_dequeues, &tot_dequeue_last);
5e70d572
MD
513
514 printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
515 tot_enqueues, tot_dequeues);
516 printf_verbose("total number of successful enqueues : %llu, "
cf9451dc 517 "enqueues to empty dest : %llu, "
23ae48ef 518 "successful dequeues %llu, "
db182e31 519 "splice : %llu, dequeue_last : %llu\n",
cf9451dc
MD
520 tot_successful_enqueues,
521 tot_empty_dest_enqueues,
522 tot_successful_dequeues,
db182e31 523 tot_splice, tot_dequeue_last);
5e70d572
MD
524 printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
525 "nr_dequeuers %3u "
526 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
cf9451dc
MD
527 "successful enqueues %12llu enqueues to empty dest %12llu "
528 "successful dequeues %12llu splice %12llu "
db182e31 529 "dequeue_last %llu "
5e70d572
MD
530 "end_dequeues %llu nr_ops %12llu\n",
531 argv[0], duration, nr_enqueuers, wdelay,
532 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
533 tot_successful_enqueues,
cf9451dc 534 tot_empty_dest_enqueues,
db182e31
MD
535 tot_successful_dequeues, tot_splice, tot_dequeue_last,
536 end_dequeues,
5e70d572 537 tot_enqueues + tot_dequeues);
cf9451dc
MD
538
539 if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
5e70d572
MD
540 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
541 "succ. dequeues + end dequeues %llu.\n",
542 tot_successful_enqueues,
543 tot_successful_dequeues + end_dequeues);
cf9451dc
MD
544 retval = 1;
545 }
5e70d572 546
cf9451dc
MD
547 /*
548 * If only using splice to dequeue, the enqueuer should see
549 * exactly as many empty queues than the number of non-empty
550 * src splice.
551 */
db182e31 552 if (tot_empty_dest_enqueues != tot_dequeue_last) {
cf9451dc 553 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
db182e31 554 "number of dequeue of last element (%llu)\n",
cf9451dc 555 tot_empty_dest_enqueues,
db182e31 556 tot_dequeue_last);
cf9451dc
MD
557 retval = 1;
558 }
2af1c19e 559 cds_wfcq_destroy(&head, &tail);
5e70d572
MD
560 free(count_enqueuer);
561 free(count_dequeuer);
562 free(tid_enqueuer);
563 free(tid_dequeuer);
83e334d0 564
cf9451dc 565 return retval;
5e70d572 566}
This page took 0.065895 seconds and 4 git commands to generate.