uatomic/x86: Remove redundant memory barriers
[urcu.git] / tests / benchmark / test_urcu_lfs.c
1 // SPDX-FileCopyrightText: 2010-2012 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 lock-free stack
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 "thread-id.h"
26
27 /* hardcoded number of CPUs */
28 #define NR_CPUS 16384
29
30 #ifndef DYNAMIC_LINK_TEST
31 #define _LGPL_SOURCE
32 #endif
33 #include <urcu.h>
34 #include <urcu/cds.h>
35
36 #define POISON_PTR ((void *) 0x42UL)
37
38 /*
39 * External synchronization used.
40 */
41 enum test_sync {
42 TEST_SYNC_NONE = 0,
43 TEST_SYNC_RCU,
44 };
45
46 static enum test_sync test_sync;
47
48 static unsigned long rduration;
49
50 static unsigned long duration;
51
52 /* read-side C.S. duration, in loops */
53 static unsigned long wdelay;
54
55 static inline void loop_sleep(unsigned long loops)
56 {
57 while (loops-- != 0)
58 caa_cpu_relax();
59 }
60
61 static int verbose_mode;
62
63 static int test_pop, test_pop_all;
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 test_duration_read();
112 }
113
114 static int test_duration_enqueue(void)
115 {
116 return test_duration_write();
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
125 static unsigned int nr_enqueuers;
126 static unsigned int nr_dequeuers;
127
128 struct test {
129 struct cds_lfs_node list;
130 struct rcu_head rcu;
131 };
132
133 static struct cds_lfs_stack s;
134
135 static void *thr_enqueuer(void *_count)
136 {
137 unsigned long long *count = _count;
138
139 printf_verbose("thread_begin %s, tid %lu\n",
140 "enqueuer", urcu_get_thread_id());
141
142 set_affinity();
143
144 rcu_register_thread();
145
146 wait_until_go();
147
148 for (;;) {
149 struct test *node = malloc(sizeof(*node));
150 if (!node)
151 goto fail;
152 cds_lfs_node_init(&node->list);
153 cds_lfs_push(&s, &node->list);
154 URCU_TLS(nr_successful_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 rcu_unregister_thread();
165
166 count[0] = URCU_TLS(nr_enqueues);
167 count[1] = URCU_TLS(nr_successful_enqueues);
168 printf_verbose("enqueuer thread_end, tid %lu, "
169 "enqueues %llu successful_enqueues %llu\n",
170 urcu_get_thread_id(),
171 URCU_TLS(nr_enqueues),
172 URCU_TLS(nr_successful_enqueues));
173 return ((void*)1);
174
175 }
176
177 static
178 void free_node_cb(struct rcu_head *head)
179 {
180 struct test *node =
181 caa_container_of(head, struct test, rcu);
182 free(node);
183 }
184
185 static
186 void do_test_pop(enum test_sync sync)
187 {
188 struct cds_lfs_node *snode;
189
190 if (sync == TEST_SYNC_RCU)
191 rcu_read_lock();
192 snode = __cds_lfs_pop(&s);
193 if (sync == TEST_SYNC_RCU)
194 rcu_read_unlock();
195 if (snode) {
196 struct test *node;
197
198 snode->next = POISON_PTR;
199 node = caa_container_of(snode,
200 struct test, list);
201 if (sync == TEST_SYNC_RCU)
202 call_rcu(&node->rcu, free_node_cb);
203 else
204 free(node);
205 URCU_TLS(nr_successful_dequeues)++;
206 }
207 URCU_TLS(nr_dequeues)++;
208 }
209
210 static
211 void do_test_pop_all(enum test_sync sync)
212 {
213 struct cds_lfs_node *snode;
214 struct cds_lfs_head *head;
215 struct cds_lfs_node *n;
216
217 head = __cds_lfs_pop_all(&s);
218 cds_lfs_for_each_safe(head, snode, n) {
219 struct test *node;
220
221 snode->next = POISON_PTR;
222 node = caa_container_of(snode, struct test, list);
223 if (sync == TEST_SYNC_RCU)
224 call_rcu(&node->rcu, free_node_cb);
225 else
226 free(node);
227 URCU_TLS(nr_successful_dequeues)++;
228 URCU_TLS(nr_dequeues)++;
229 }
230
231 }
232
233 static void *thr_dequeuer(void *_count)
234 {
235 unsigned long long *count = _count;
236 unsigned int counter = 0;
237
238 printf_verbose("thread_begin %s, tid %lu\n",
239 "dequeuer", urcu_get_thread_id());
240
241 set_affinity();
242
243 rcu_register_thread();
244
245 wait_until_go();
246
247 urcu_posix_assert(test_pop || test_pop_all);
248
249 for (;;) {
250 if (test_pop && test_pop_all) {
251 /* both pop and pop all */
252 if (counter & 1)
253 do_test_pop(test_sync);
254 else
255 do_test_pop_all(test_sync);
256 counter++;
257 } else {
258 if (test_pop)
259 do_test_pop(test_sync);
260 else
261 do_test_pop_all(test_sync);
262 }
263
264 if (caa_unlikely(!test_duration_dequeue()))
265 break;
266 if (caa_unlikely(rduration))
267 loop_sleep(rduration);
268 }
269
270 rcu_unregister_thread();
271
272 printf_verbose("dequeuer thread_end, tid %lu, "
273 "dequeues %llu, successful_dequeues %llu\n",
274 urcu_get_thread_id(),
275 URCU_TLS(nr_dequeues),
276 URCU_TLS(nr_successful_dequeues));
277 count[0] = URCU_TLS(nr_dequeues);
278 count[1] = URCU_TLS(nr_successful_dequeues);
279 return ((void*)2);
280 }
281
282 static void test_end(unsigned long long *nr_dequeues_l)
283 {
284 struct cds_lfs_node *snode;
285
286 do {
287 snode = __cds_lfs_pop(&s);
288 if (snode) {
289 struct test *node;
290
291 node = caa_container_of(snode, struct test, list);
292 free(node);
293 (*nr_dequeues_l)++;
294 }
295 } while (snode);
296 }
297
298 static void show_usage(char **argv)
299 {
300 printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
301 argv[0]);
302 printf("OPTIONS:\n");
303 printf(" [-d delay] (enqueuer period (in loops))\n");
304 printf(" [-c duration] (dequeuer period (in loops))\n");
305 printf(" [-v] (verbose output)\n");
306 printf(" [-a cpu#] [-a cpu#]... (affinity)\n");
307 printf(" [-p] (test pop)\n");
308 printf(" [-P] (test pop_all, enabled by default)\n");
309 printf(" [-R] (use RCU external synchronization)\n");
310 printf(" Note: default: no external synchronization used.\n");
311 printf("\n");
312 }
313
314 int main(int argc, char **argv)
315 {
316 int err;
317 pthread_t *tid_enqueuer, *tid_dequeuer;
318 void *tret;
319 unsigned long long *count_enqueuer, *count_dequeuer;
320 unsigned long long tot_enqueues = 0, tot_dequeues = 0;
321 unsigned long long tot_successful_enqueues = 0,
322 tot_successful_dequeues = 0;
323 unsigned long long end_dequeues = 0;
324 int i, a;
325 unsigned int i_thr;
326
327 if (argc < 4) {
328 show_usage(argv);
329 return -1;
330 }
331
332 err = sscanf(argv[1], "%u", &nr_dequeuers);
333 if (err != 1) {
334 show_usage(argv);
335 return -1;
336 }
337
338 err = sscanf(argv[2], "%u", &nr_enqueuers);
339 if (err != 1) {
340 show_usage(argv);
341 return -1;
342 }
343
344 err = sscanf(argv[3], "%lu", &duration);
345 if (err != 1) {
346 show_usage(argv);
347 return -1;
348 }
349
350 for (i = 4; i < argc; i++) {
351 if (argv[i][0] != '-')
352 continue;
353 switch (argv[i][1]) {
354 case 'a':
355 if (argc < i + 2) {
356 show_usage(argv);
357 return -1;
358 }
359 a = atoi(argv[++i]);
360 cpu_affinities[next_aff++] = a;
361 use_affinity = 1;
362 printf_verbose("Adding CPU %d affinity\n", a);
363 break;
364 case 'c':
365 if (argc < i + 2) {
366 show_usage(argv);
367 return -1;
368 }
369 rduration = atol(argv[++i]);
370 break;
371 case 'd':
372 if (argc < i + 2) {
373 show_usage(argv);
374 return -1;
375 }
376 wdelay = atol(argv[++i]);
377 break;
378 case 'v':
379 verbose_mode = 1;
380 break;
381 case 'p':
382 test_pop = 1;
383 break;
384 case 'P':
385 test_pop_all = 1;
386 break;
387 case 'R':
388 test_sync = TEST_SYNC_RCU;
389 break;
390 }
391 }
392
393 /* activate pop_all test by default */
394 if (!test_pop && !test_pop_all)
395 test_pop_all = 1;
396
397 printf_verbose("running test for %lu seconds, %u enqueuers, "
398 "%u dequeuers.\n",
399 duration, nr_enqueuers, nr_dequeuers);
400 if (test_pop)
401 printf_verbose("pop test activated.\n");
402 if (test_pop_all)
403 printf_verbose("pop_all test activated.\n");
404 if (test_sync == TEST_SYNC_RCU)
405 printf_verbose("External sync: RCU.\n");
406 else
407 printf_verbose("External sync: none.\n");
408 printf_verbose("Writer delay : %lu loops.\n", rduration);
409 printf_verbose("Reader duration : %lu loops.\n", wdelay);
410 printf_verbose("thread %-6s, tid %lu\n",
411 "main", urcu_get_thread_id());
412
413 tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
414 tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
415 count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
416 count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
417 cds_lfs_init(&s);
418 err = create_all_cpu_call_rcu_data(0);
419 if (err) {
420 printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n");
421 }
422
423 next_aff = 0;
424
425 for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
426 err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
427 &count_enqueuer[2 * i_thr]);
428 if (err != 0)
429 exit(1);
430 }
431 for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
432 err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
433 &count_dequeuer[2 * i_thr]);
434 if (err != 0)
435 exit(1);
436 }
437
438 cmm_smp_mb();
439
440 begin_test();
441
442 for (i_thr = 0; i_thr < duration; i_thr++) {
443 sleep(1);
444 if (verbose_mode) {
445 fwrite(".", sizeof(char), 1, stdout);
446 fflush(stdout);
447 }
448 }
449
450 end_test();
451
452 for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
453 err = pthread_join(tid_enqueuer[i_thr], &tret);
454 if (err != 0)
455 exit(1);
456 tot_enqueues += count_enqueuer[2 * i_thr];
457 tot_successful_enqueues += count_enqueuer[2 * i_thr + 1];
458 }
459 for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
460 err = pthread_join(tid_dequeuer[i_thr], &tret);
461 if (err != 0)
462 exit(1);
463 tot_dequeues += count_dequeuer[2 * i_thr];
464 tot_successful_dequeues += count_dequeuer[2 * i_thr + 1];
465 }
466
467 test_end(&end_dequeues);
468
469 printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
470 tot_enqueues, tot_dequeues);
471 printf_verbose("total number of successful enqueues : %llu, "
472 "successful dequeues %llu\n",
473 tot_successful_enqueues, tot_successful_dequeues);
474 printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
475 "nr_dequeuers %3u "
476 "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
477 "successful enqueues %12llu successful dequeues %12llu "
478 "end_dequeues %llu nr_ops %12llu\n",
479 argv[0], duration, nr_enqueuers, wdelay,
480 nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
481 tot_successful_enqueues,
482 tot_successful_dequeues, end_dequeues,
483 tot_enqueues + tot_dequeues);
484 if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
485 printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
486 "succ. dequeues + end dequeues %llu.\n",
487 tot_successful_enqueues,
488 tot_successful_dequeues + end_dequeues);
489
490 free_all_cpu_call_rcu_data();
491 cds_lfs_destroy(&s);
492 free(count_enqueuer);
493 free(count_dequeuer);
494 free(tid_enqueuer);
495 free(tid_dequeuer);
496
497 return 0;
498 }
This page took 0.038801 seconds and 5 git commands to generate.