uatomic/x86: Remove redundant memory barriers
[urcu.git] / tests / benchmark / test_urcu_wfcq.c
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
6 /*
7 * Userspace RCU library - example RCU-based lock-free concurrent queue
8 */
9
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>
20 #include <errno.h>
21
22 #include <urcu/arch.h>
23 #include <urcu/assert.h>
24 #include <urcu/tls-compat.h>
25 #include <urcu/uatomic.h>
26 #include "thread-id.h"
27
28 /* hardcoded number of CPUs */
29 #define NR_CPUS 16384
30
31 #ifndef DYNAMIC_LINK_TEST
32 #define _LGPL_SOURCE
33 #endif
34 #include <urcu/wfcqueue.h>
35
36 enum test_sync {
37 TEST_SYNC_NONE = 0,
38 TEST_SYNC_MUTEX,
39 };
40
41 static enum test_sync test_sync;
42
43 static int test_force_sync;
44
45 static volatile int test_stop_enqueue, test_stop_dequeue;
46
47 static unsigned long rduration;
48
49 static unsigned long duration;
50
51 /* read-side C.S. duration, in loops */
52 static unsigned long wdelay;
53
54 static inline void loop_sleep(unsigned long loops)
55 {
56 while (loops-- != 0)
57 caa_cpu_relax();
58 }
59
60 static int verbose_mode;
61
62 static int test_dequeue, test_splice, test_wait_empty;
63 static unsigned int test_enqueue_stopped;
64
65 #define printf_verbose(fmt, args...) \
66 do { \
67 if (verbose_mode) \
68 printf(fmt, ## args); \
69 } while (0)
70
71 static unsigned int cpu_affinities[NR_CPUS];
72 static unsigned int next_aff = 0;
73 static int use_affinity = 0;
74
75 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
76
77 static void set_affinity(void)
78 {
79 #ifdef HAVE_SCHED_SETAFFINITY
80 cpu_set_t mask;
81 int cpu, ret;
82 #endif /* HAVE_SCHED_SETAFFINITY */
83
84 if (!use_affinity)
85 return;
86
87 #ifdef HAVE_SCHED_SETAFFINITY
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);
102 sched_setaffinity(0, sizeof(mask), &mask);
103 #endif /* HAVE_SCHED_SETAFFINITY */
104 }
105
106 /*
107 * returns 0 if test should end.
108 */
109 static int test_duration_dequeue(void)
110 {
111 return !uatomic_load(&test_stop_dequeue, CMM_RELAXED);
112 }
113
114 static int test_duration_enqueue(void)
115 {
116 return !uatomic_load(&test_stop_enqueue, CMM_RELAXED);
117 }
118
119 static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
120 static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
121
122 static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
123 static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
124 static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
125 static DEFINE_URCU_TLS(unsigned long long, nr_splice);
126 static DEFINE_URCU_TLS(unsigned long long, nr_dequeue_last);
127
128 static unsigned int nr_enqueuers;
129 static unsigned int nr_dequeuers;
130
131 static struct cds_wfcq_head __attribute__((aligned(CAA_CACHE_LINE_SIZE))) head;
132 static struct cds_wfcq_tail __attribute__((aligned(CAA_CACHE_LINE_SIZE))) tail;
133
134 static void *thr_enqueuer(void *_count)
135 {
136 unsigned long long *count = _count;
137 bool was_nonempty;
138
139 printf_verbose("thread_begin %s, tid %lu\n",
140 "enqueuer", urcu_get_thread_id());
141
142 set_affinity();
143
144 wait_until_go();
145
146 for (;;) {
147 struct cds_wfcq_node *node = malloc(sizeof(*node));
148 if (!node)
149 goto fail;
150 cds_wfcq_node_init(node);
151 was_nonempty = cds_wfcq_enqueue(&head, &tail, node);
152 URCU_TLS(nr_successful_enqueues)++;
153 if (!was_nonempty)
154 URCU_TLS(nr_empty_dest_enqueues)++;
155
156 if (caa_unlikely(wdelay))
157 loop_sleep(wdelay);
158 fail:
159 URCU_TLS(nr_enqueues)++;
160 if (caa_unlikely(!test_duration_enqueue()))
161 break;
162 }
163
164 uatomic_inc(&test_enqueue_stopped);
165 count[0] = URCU_TLS(nr_enqueues);
166 count[1] = URCU_TLS(nr_successful_enqueues);
167 count[2] = URCU_TLS(nr_empty_dest_enqueues);
168 printf_verbose("enqueuer thread_end, tid %lu, "
169 "enqueues %llu successful_enqueues %llu, "
170 "empty_dest_enqueues %llu\n",
171 urcu_get_thread_id(),
172 URCU_TLS(nr_enqueues),
173 URCU_TLS(nr_successful_enqueues),
174 URCU_TLS(nr_empty_dest_enqueues));
175 return ((void*)1);
176
177 }
178
179 static void do_test_dequeue(enum test_sync sync)
180 {
181 struct cds_wfcq_node *node;
182 int state;
183
184 if (sync == TEST_SYNC_MUTEX)
185 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
186 &state);
187 else
188 node = __cds_wfcq_dequeue_with_state_blocking(&head, &tail,
189 &state);
190
191 if (state & CDS_WFCQ_STATE_LAST)
192 URCU_TLS(nr_dequeue_last)++;
193
194 if (node) {
195 free(node);
196 URCU_TLS(nr_successful_dequeues)++;
197 }
198 URCU_TLS(nr_dequeues)++;
199 }
200
201 static void do_test_splice(enum test_sync sync)
202 {
203 struct cds_wfcq_head tmp_head;
204 struct cds_wfcq_tail tmp_tail;
205 struct cds_wfcq_node *node, *n;
206 enum cds_wfcq_ret ret;
207
208 cds_wfcq_init(&tmp_head, &tmp_tail);
209
210 if (sync == TEST_SYNC_MUTEX)
211 ret = cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
212 &head, &tail);
213 else
214 ret = __cds_wfcq_splice_blocking(&tmp_head, &tmp_tail,
215 &head, &tail);
216
217 switch (ret) {
218 case CDS_WFCQ_RET_WOULDBLOCK:
219 urcu_posix_assert(0); /* blocking call */
220 break;
221 case CDS_WFCQ_RET_DEST_EMPTY:
222 URCU_TLS(nr_splice)++;
223 URCU_TLS(nr_dequeue_last)++;
224 /* ok */
225 break;
226 case CDS_WFCQ_RET_DEST_NON_EMPTY:
227 urcu_posix_assert(0); /* entirely unexpected */
228 break;
229 case CDS_WFCQ_RET_SRC_EMPTY:
230 /* ok, we could even skip iteration on dest if we wanted */
231 break;
232 }
233
234 __cds_wfcq_for_each_blocking_safe(&tmp_head, &tmp_tail, node, n) {
235 free(node);
236 URCU_TLS(nr_successful_dequeues)++;
237 URCU_TLS(nr_dequeues)++;
238 }
239 cds_wfcq_destroy(&tmp_head, &tmp_tail);
240 }
241
242 static void *thr_dequeuer(void *_count)
243 {
244 unsigned long long *count = _count;
245 unsigned int counter = 0;
246
247 printf_verbose("thread_begin %s, tid %lu\n",
248 "dequeuer", urcu_get_thread_id());
249
250 set_affinity();
251
252 wait_until_go();
253
254 for (;;) {
255 if (test_dequeue && test_splice) {
256 if (counter & 1)
257 do_test_dequeue(test_sync);
258 else
259 do_test_splice(test_sync);
260 counter++;
261 } else {
262 if (test_dequeue)
263 do_test_dequeue(test_sync);
264 else
265 do_test_splice(test_sync);
266 }
267 if (caa_unlikely(!test_duration_dequeue()))
268 break;
269 if (caa_unlikely(rduration))
270 loop_sleep(rduration);
271 }
272
273 printf_verbose("dequeuer thread_end, tid %lu, "
274 "dequeues %llu, successful_dequeues %llu, "
275 "nr_splice %llu\n",
276 urcu_get_thread_id(),
277 URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
278 URCU_TLS(nr_splice));
279 count[0] = URCU_TLS(nr_dequeues);
280 count[1] = URCU_TLS(nr_successful_dequeues);
281 count[2] = URCU_TLS(nr_splice);
282 count[3] = URCU_TLS(nr_dequeue_last);
283 return ((void*)2);
284 }
285
286 static void test_end(unsigned long long *nr_dequeues_l,
287 unsigned long long *nr_dequeue_last_l)
288 {
289 struct cds_wfcq_node *node;
290 int state;
291
292 do {
293 node = cds_wfcq_dequeue_with_state_blocking(&head, &tail,
294 &state);
295 if (node) {
296 if (state & CDS_WFCQ_STATE_LAST)
297 (*nr_dequeue_last_l)++;
298 free(node);
299 (*nr_dequeues_l)++;
300 }
301 } while (node);
302 }
303
304 static void show_usage(char **argv)
305 {
306 printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
307 argv[0]);
308 printf("OPTIONS:\n");
309 printf(" [-d delay] (enqueuer period (in loops))\n");
310 printf(" [-c duration] (dequeuer period (in loops))\n");
311 printf(" [-v] (verbose output)\n");
312 printf(" [-a cpu#] [-a cpu#]... (affinity)\n");
313 printf(" [-q] (test dequeue)\n");
314 printf(" [-s] (test splice, enabled by default)\n");
315 printf(" [-M] (use mutex external synchronization)\n");
316 printf(" Note: default: no external synchronization used.\n");
317 printf(" [-f] (force user-provided synchronization)\n");
318 printf(" [-w] Wait for dequeuer to empty queue\n");
319 printf("\n");
320 }
321
322 int main(int argc, char **argv)
323 {
324 int err;
325 pthread_t *tid_enqueuer, *tid_dequeuer;
326 void *tret;
327 unsigned long long *count_enqueuer, *count_dequeuer;
328 unsigned long long tot_enqueues = 0, tot_dequeues = 0;
329 unsigned long long tot_successful_enqueues = 0,
330 tot_successful_dequeues = 0,
331 tot_empty_dest_enqueues = 0,
332 tot_splice = 0, tot_dequeue_last = 0;
333 unsigned long long end_dequeues = 0;
334 int i, a, retval = 0;
335 unsigned int i_thr;
336
337 if (argc < 4) {
338 show_usage(argv);
339 return -1;
340 }
341
342 err = sscanf(argv[1], "%u", &nr_dequeuers);
343 if (err != 1) {
344 show_usage(argv);
345 return -1;
346 }
347
348 err = sscanf(argv[2], "%u", &nr_enqueuers);
349 if (err != 1) {
350 show_usage(argv);
351 return -1;
352 }
353
354 err = sscanf(argv[3], "%lu", &duration);
355 if (err != 1) {
356 show_usage(argv);
357 return -1;
358 }
359
360 for (i = 4; i < argc; i++) {
361 if (argv[i][0] != '-')
362 continue;
363 switch (argv[i][1]) {
364 case 'a':
365 if (argc < i + 2) {
366 show_usage(argv);
367 return -1;
368 }
369 a = atoi(argv[++i]);
370 cpu_affinities[next_aff++] = a;
371 use_affinity = 1;
372 printf_verbose("Adding CPU %d affinity\n", a);
373 break;
374 case 'c':
375 if (argc < i + 2) {
376 show_usage(argv);
377 return -1;
378 }
379 rduration = atol(argv[++i]);
380 break;
381 case 'd':
382 if (argc < i + 2) {
383 show_usage(argv);
384 return -1;
385 }
386 wdelay = atol(argv[++i]);
387 break;
388 case 'v':
389 verbose_mode = 1;
390 break;
391 case 'q':
392 test_dequeue = 1;
393 break;
394 case 's':
395 test_splice = 1;
396 break;
397 case 'M':
398 test_sync = TEST_SYNC_MUTEX;
399 break;
400 case 'w':
401 test_wait_empty = 1;
402 break;
403 case 'f':
404 test_force_sync = 1;
405 break;
406 }
407 }
408
409 /* activate splice test by default */
410 if (!test_dequeue && !test_splice)
411 test_splice = 1;
412
413 if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_dequeue) {
414 if (test_force_sync) {
415 fprintf(stderr, "[WARNING] Using dequeue concurrently "
416 "with other dequeue or splice without external "
417 "synchronization. Expect run-time failure.\n");
418 } else {
419 printf("Enforcing mutex synchronization\n");
420 test_sync = TEST_SYNC_MUTEX;
421 }
422 }
423
424 printf_verbose("running test for %lu seconds, %u enqueuers, "
425 "%u dequeuers.\n",
426 duration, nr_enqueuers, nr_dequeuers);
427 if (test_dequeue)
428 printf_verbose("dequeue test activated.\n");
429 else
430 printf_verbose("splice test activated.\n");
431 if (test_sync == TEST_SYNC_MUTEX)
432 printf_verbose("External sync: mutex.\n");
433 else
434 printf_verbose("External sync: none.\n");
435 if (test_wait_empty)
436 printf_verbose("Wait for dequeuers to empty queue.\n");
437 printf_verbose("Writer delay : %lu loops.\n", rduration);
438 printf_verbose("Reader duration : %lu loops.\n", wdelay);
439 printf_verbose("thread %-6s, tid %lu\n",
440 "main", urcu_get_thread_id());
441
442 tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
443 tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
444 count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
445 count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
446 cds_wfcq_init(&head, &tail);
447
448 next_aff = 0;
449
450 for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
451 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
452 &count_enqueuer[3 * i_thr]);
453 if (err != 0)
454 exit(1);
455 }
456 for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
457 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
458 &count_dequeuer[4 * i_thr]);
459 if (err != 0)
460 exit(1);
461 }
462
463 cmm_smp_mb();
464
465 begin_test();
466
467 for (i_thr = 0; i_thr < duration; i_thr++) {
468 sleep(1);
469 if (verbose_mode) {
470 fwrite(".", sizeof(char), 1, stdout);
471 fflush(stdout);
472 }
473 }
474
475 uatomic_store(&test_stop_enqueue, 1, CMM_RELEASE);
476
477 if (test_wait_empty) {
478 while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
479 sleep(1);
480 }
481 while (!cds_wfcq_empty(&head, &tail)) {
482 sleep(1);
483 }
484 }
485
486 uatomic_store(&test_stop_dequeue, 1, CMM_RELAXED);
487
488 for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
489 err = pthread_join(tid_enqueuer[i_thr], &tret);
490 if (err != 0)
491 exit(1);
492 tot_enqueues += count_enqueuer[3 * i_thr];
493 tot_successful_enqueues += count_enqueuer[3 * i_thr + 1];
494 tot_empty_dest_enqueues += count_enqueuer[3 * i_thr + 2];
495 }
496 for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
497 err = pthread_join(tid_dequeuer[i_thr], &tret);
498 if (err != 0)
499 exit(1);
500 tot_dequeues += count_dequeuer[4 * i_thr];
501 tot_successful_dequeues += count_dequeuer[4 * i_thr + 1];
502 tot_splice += count_dequeuer[4 * i_thr + 2];
503 tot_dequeue_last += count_dequeuer[4 * i_thr + 3];
504 }
505
506 test_end(&end_dequeues, &tot_dequeue_last);
507
508 printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
509 tot_enqueues, tot_dequeues);
510 printf_verbose("total number of successful enqueues : %llu, "
511 "enqueues to empty dest : %llu, "
512 "successful dequeues %llu, "
513 "splice : %llu, dequeue_last : %llu\n",
514 tot_successful_enqueues,
515 tot_empty_dest_enqueues,
516 tot_successful_dequeues,
517 tot_splice, tot_dequeue_last);
518 printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
519 "nr_dequeuers %3u "
520 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
521 "successful enqueues %12llu enqueues to empty dest %12llu "
522 "successful dequeues %12llu splice %12llu "
523 "dequeue_last %llu "
524 "end_dequeues %llu nr_ops %12llu\n",
525 argv[0], duration, nr_enqueuers, wdelay,
526 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
527 tot_successful_enqueues,
528 tot_empty_dest_enqueues,
529 tot_successful_dequeues, tot_splice, tot_dequeue_last,
530 end_dequeues,
531 tot_enqueues + tot_dequeues);
532
533 if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
534 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
535 "succ. dequeues + end dequeues %llu.\n",
536 tot_successful_enqueues,
537 tot_successful_dequeues + end_dequeues);
538 retval = 1;
539 }
540
541 /*
542 * If only using splice to dequeue, the enqueuer should see
543 * exactly as many empty queues than the number of non-empty
544 * src splice.
545 */
546 if (tot_empty_dest_enqueues != tot_dequeue_last) {
547 printf("WARNING! Discrepancy between empty enqueue (%llu) and "
548 "number of dequeue of last element (%llu)\n",
549 tot_empty_dest_enqueues,
550 tot_dequeue_last);
551 retval = 1;
552 }
553 cds_wfcq_destroy(&head, &tail);
554 free(count_enqueuer);
555 free(count_dequeuer);
556 free(tid_enqueuer);
557 free(tid_dequeuer);
558
559 return retval;
560 }
This page took 0.040361 seconds and 5 git commands to generate.