Commit | Line | Data |
---|---|---|
ce29b371 MJ |
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 | ||
453629a9 | 6 | /* |
cb8818f0 | 7 | * Userspace RCU library - example lock-free stack |
453629a9 MD |
8 | */ |
9 | ||
453629a9 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> | |
453629a9 MD |
20 | #include <errno.h> |
21 | ||
22 | #include <urcu/arch.h> | |
01477510 | 23 | #include <urcu/assert.h> |
bd252a04 | 24 | #include <urcu/tls-compat.h> |
94df6318 | 25 | #include "thread-id.h" |
65fcc7e9 | 26 | |
453629a9 MD |
27 | /* hardcoded number of CPUs */ |
28 | #define NR_CPUS 16384 | |
29 | ||
453629a9 MD |
30 | #ifndef DYNAMIC_LINK_TEST |
31 | #define _LGPL_SOURCE | |
32 | #endif | |
33 | #include <urcu.h> | |
e5a7d709 | 34 | #include <urcu/cds.h> |
453629a9 | 35 | |
f9b5c2b6 MD |
36 | #define POISON_PTR ((void *) 0x42UL) |
37 | ||
111ce0c3 MD |
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 | ||
453629a9 MD |
48 | static volatile int test_go, test_stop; |
49 | ||
50 | static unsigned long rduration; | |
51 | ||
52 | static unsigned long duration; | |
53 | ||
54 | /* read-side C.S. duration, in loops */ | |
55 | static unsigned long wdelay; | |
56 | ||
ab0aacbe | 57 | static inline void loop_sleep(unsigned long loops) |
453629a9 | 58 | { |
ab0aacbe | 59 | while (loops-- != 0) |
06f22bdb | 60 | caa_cpu_relax(); |
453629a9 MD |
61 | } |
62 | ||
63 | static int verbose_mode; | |
64 | ||
111ce0c3 MD |
65 | static int test_pop, test_pop_all; |
66 | ||
453629a9 MD |
67 | #define printf_verbose(fmt, args...) \ |
68 | do { \ | |
69 | if (verbose_mode) \ | |
111ce0c3 | 70 | printf(fmt, ## args); \ |
453629a9 MD |
71 | } while (0) |
72 | ||
73 | static unsigned int cpu_affinities[NR_CPUS]; | |
74 | static unsigned int next_aff = 0; | |
75 | static int use_affinity = 0; | |
76 | ||
77 | pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER; | |
78 | ||
453629a9 MD |
79 | static void set_affinity(void) |
80 | { | |
2388c075 | 81 | #ifdef HAVE_SCHED_SETAFFINITY |
453629a9 | 82 | cpu_set_t mask; |
95bc7fb9 MD |
83 | int cpu, ret; |
84 | #endif /* HAVE_SCHED_SETAFFINITY */ | |
453629a9 MD |
85 | |
86 | if (!use_affinity) | |
87 | return; | |
88 | ||
2388c075 | 89 | #ifdef HAVE_SCHED_SETAFFINITY |
453629a9 MD |
90 | ret = pthread_mutex_lock(&affinity_mutex); |
91 | if (ret) { | |
92 | perror("Error in pthread mutex lock"); | |
93 | exit(-1); | |
94 | } | |
95 | cpu = cpu_affinities[next_aff++]; | |
96 | ret = pthread_mutex_unlock(&affinity_mutex); | |
97 | if (ret) { | |
98 | perror("Error in pthread mutex unlock"); | |
99 | exit(-1); | |
100 | } | |
101 | ||
102 | CPU_ZERO(&mask); | |
103 | CPU_SET(cpu, &mask); | |
453629a9 | 104 | sched_setaffinity(0, sizeof(mask), &mask); |
453629a9 MD |
105 | #endif /* HAVE_SCHED_SETAFFINITY */ |
106 | } | |
107 | ||
108 | /* | |
109 | * returns 0 if test should end. | |
110 | */ | |
111 | static int test_duration_dequeue(void) | |
112 | { | |
113 | return !test_stop; | |
114 | } | |
115 | ||
116 | static int test_duration_enqueue(void) | |
117 | { | |
118 | return !test_stop; | |
119 | } | |
120 | ||
bd252a04 MD |
121 | static DEFINE_URCU_TLS(unsigned long long, nr_dequeues); |
122 | static DEFINE_URCU_TLS(unsigned long long, nr_enqueues); | |
453629a9 | 123 | |
bd252a04 MD |
124 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues); |
125 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues); | |
453629a9 MD |
126 | |
127 | static unsigned int nr_enqueuers; | |
128 | static unsigned int nr_dequeuers; | |
129 | ||
94aec122 | 130 | struct test { |
cb8818f0 | 131 | struct cds_lfs_node list; |
94aec122 MD |
132 | struct rcu_head rcu; |
133 | }; | |
134 | ||
cb8818f0 | 135 | static struct cds_lfs_stack s; |
453629a9 | 136 | |
eb314468 | 137 | static void *thr_enqueuer(void *_count) |
453629a9 MD |
138 | { |
139 | unsigned long long *count = _count; | |
140 | ||
94df6318 MD |
141 | printf_verbose("thread_begin %s, tid %lu\n", |
142 | "enqueuer", urcu_get_thread_id()); | |
453629a9 MD |
143 | |
144 | set_affinity(); | |
145 | ||
146 | rcu_register_thread(); | |
147 | ||
148 | while (!test_go) | |
149 | { | |
150 | } | |
5481ddb3 | 151 | cmm_smp_mb(); |
453629a9 MD |
152 | |
153 | for (;;) { | |
94aec122 | 154 | struct test *node = malloc(sizeof(*node)); |
453629a9 MD |
155 | if (!node) |
156 | goto fail; | |
cb8818f0 MD |
157 | cds_lfs_node_init(&node->list); |
158 | cds_lfs_push(&s, &node->list); | |
bd252a04 | 159 | URCU_TLS(nr_successful_enqueues)++; |
453629a9 | 160 | |
a0b7f7ea | 161 | if (caa_unlikely(wdelay)) |
453629a9 MD |
162 | loop_sleep(wdelay); |
163 | fail: | |
bd252a04 | 164 | URCU_TLS(nr_enqueues)++; |
a0b7f7ea | 165 | if (caa_unlikely(!test_duration_enqueue())) |
453629a9 MD |
166 | break; |
167 | } | |
168 | ||
169 | rcu_unregister_thread(); | |
170 | ||
bd252a04 MD |
171 | count[0] = URCU_TLS(nr_enqueues); |
172 | count[1] = URCU_TLS(nr_successful_enqueues); | |
94df6318 MD |
173 | printf_verbose("enqueuer thread_end, tid %lu, " |
174 | "enqueues %llu successful_enqueues %llu\n", | |
175 | urcu_get_thread_id(), | |
176 | URCU_TLS(nr_enqueues), | |
177 | URCU_TLS(nr_successful_enqueues)); | |
453629a9 MD |
178 | return ((void*)1); |
179 | ||
180 | } | |
181 | ||
94aec122 MD |
182 | static |
183 | void free_node_cb(struct rcu_head *head) | |
184 | { | |
185 | struct test *node = | |
186 | caa_container_of(head, struct test, rcu); | |
187 | free(node); | |
188 | } | |
189 | ||
111ce0c3 MD |
190 | static |
191 | void do_test_pop(enum test_sync sync) | |
192 | { | |
193 | struct cds_lfs_node *snode; | |
194 | ||
195 | if (sync == TEST_SYNC_RCU) | |
196 | rcu_read_lock(); | |
197 | snode = __cds_lfs_pop(&s); | |
198 | if (sync == TEST_SYNC_RCU) | |
199 | rcu_read_unlock(); | |
200 | if (snode) { | |
201 | struct test *node; | |
202 | ||
f9b5c2b6 | 203 | snode->next = POISON_PTR; |
111ce0c3 MD |
204 | node = caa_container_of(snode, |
205 | struct test, list); | |
206 | if (sync == TEST_SYNC_RCU) | |
207 | call_rcu(&node->rcu, free_node_cb); | |
208 | else | |
209 | free(node); | |
210 | URCU_TLS(nr_successful_dequeues)++; | |
211 | } | |
212 | URCU_TLS(nr_dequeues)++; | |
213 | } | |
214 | ||
215 | static | |
216 | void do_test_pop_all(enum test_sync sync) | |
217 | { | |
218 | struct cds_lfs_node *snode; | |
219 | struct cds_lfs_head *head; | |
220 | struct cds_lfs_node *n; | |
221 | ||
222 | head = __cds_lfs_pop_all(&s); | |
223 | cds_lfs_for_each_safe(head, snode, n) { | |
224 | struct test *node; | |
225 | ||
f9b5c2b6 | 226 | snode->next = POISON_PTR; |
111ce0c3 MD |
227 | node = caa_container_of(snode, struct test, list); |
228 | if (sync == TEST_SYNC_RCU) | |
229 | call_rcu(&node->rcu, free_node_cb); | |
230 | else | |
231 | free(node); | |
232 | URCU_TLS(nr_successful_dequeues)++; | |
233 | URCU_TLS(nr_dequeues)++; | |
234 | } | |
235 | ||
236 | } | |
237 | ||
eb314468 | 238 | static void *thr_dequeuer(void *_count) |
453629a9 MD |
239 | { |
240 | unsigned long long *count = _count; | |
f42bc7ef | 241 | unsigned int counter = 0; |
453629a9 | 242 | |
94df6318 MD |
243 | printf_verbose("thread_begin %s, tid %lu\n", |
244 | "dequeuer", urcu_get_thread_id()); | |
453629a9 MD |
245 | |
246 | set_affinity(); | |
247 | ||
453629a9 MD |
248 | rcu_register_thread(); |
249 | ||
250 | while (!test_go) | |
251 | { | |
252 | } | |
5481ddb3 | 253 | cmm_smp_mb(); |
453629a9 | 254 | |
01477510 | 255 | urcu_posix_assert(test_pop || test_pop_all); |
f3eb6ef1 | 256 | |
111ce0c3 | 257 | for (;;) { |
111ce0c3 MD |
258 | if (test_pop && test_pop_all) { |
259 | /* both pop and pop all */ | |
260 | if (counter & 1) | |
261 | do_test_pop(test_sync); | |
262 | else | |
263 | do_test_pop_all(test_sync); | |
264 | counter++; | |
265 | } else { | |
266 | if (test_pop) | |
267 | do_test_pop(test_sync); | |
268 | else | |
269 | do_test_pop_all(test_sync); | |
453629a9 | 270 | } |
111ce0c3 | 271 | |
a0b7f7ea | 272 | if (caa_unlikely(!test_duration_dequeue())) |
453629a9 | 273 | break; |
a0b7f7ea | 274 | if (caa_unlikely(rduration)) |
453629a9 MD |
275 | loop_sleep(rduration); |
276 | } | |
277 | ||
278 | rcu_unregister_thread(); | |
453629a9 | 279 | |
94df6318 MD |
280 | printf_verbose("dequeuer thread_end, tid %lu, " |
281 | "dequeues %llu, successful_dequeues %llu\n", | |
282 | urcu_get_thread_id(), | |
283 | URCU_TLS(nr_dequeues), | |
284 | URCU_TLS(nr_successful_dequeues)); | |
bd252a04 MD |
285 | count[0] = URCU_TLS(nr_dequeues); |
286 | count[1] = URCU_TLS(nr_successful_dequeues); | |
453629a9 MD |
287 | return ((void*)2); |
288 | } | |
289 | ||
bd23a6c6 | 290 | static void test_end(unsigned long long *nr_dequeues_l) |
453629a9 | 291 | { |
cb8818f0 | 292 | struct cds_lfs_node *snode; |
453629a9 MD |
293 | |
294 | do { | |
bd23a6c6 | 295 | snode = __cds_lfs_pop(&s); |
94aec122 MD |
296 | if (snode) { |
297 | struct test *node; | |
298 | ||
299 | node = caa_container_of(snode, struct test, list); | |
453629a9 | 300 | free(node); |
bd23a6c6 | 301 | (*nr_dequeues_l)++; |
453629a9 | 302 | } |
94aec122 | 303 | } while (snode); |
453629a9 MD |
304 | } |
305 | ||
70469b43 | 306 | static void show_usage(char **argv) |
453629a9 | 307 | { |
06637138 MD |
308 | printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n", |
309 | argv[0]); | |
310 | printf("OPTIONS:\n"); | |
311 | printf(" [-d delay] (enqueuer period (in loops))\n"); | |
312 | printf(" [-c duration] (dequeuer period (in loops))\n"); | |
313 | printf(" [-v] (verbose output)\n"); | |
314 | printf(" [-a cpu#] [-a cpu#]... (affinity)\n"); | |
315 | printf(" [-p] (test pop)\n"); | |
316 | printf(" [-P] (test pop_all, enabled by default)\n"); | |
317 | printf(" [-R] (use RCU external synchronization)\n"); | |
318 | printf(" Note: default: no external synchronization used.\n"); | |
453629a9 MD |
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 | unsigned long long end_dequeues = 0; | |
332 | int i, a; | |
83e334d0 | 333 | unsigned int i_thr; |
453629a9 MD |
334 | |
335 | if (argc < 4) { | |
70469b43 | 336 | show_usage(argv); |
453629a9 MD |
337 | return -1; |
338 | } | |
339 | ||
340 | err = sscanf(argv[1], "%u", &nr_dequeuers); | |
341 | if (err != 1) { | |
70469b43 | 342 | show_usage(argv); |
453629a9 MD |
343 | return -1; |
344 | } | |
345 | ||
346 | err = sscanf(argv[2], "%u", &nr_enqueuers); | |
347 | if (err != 1) { | |
70469b43 | 348 | show_usage(argv); |
453629a9 MD |
349 | return -1; |
350 | } | |
83e334d0 | 351 | |
453629a9 MD |
352 | err = sscanf(argv[3], "%lu", &duration); |
353 | if (err != 1) { | |
70469b43 | 354 | show_usage(argv); |
453629a9 MD |
355 | return -1; |
356 | } | |
357 | ||
358 | for (i = 4; i < argc; i++) { | |
359 | if (argv[i][0] != '-') | |
360 | continue; | |
361 | switch (argv[i][1]) { | |
362 | case 'a': | |
363 | if (argc < i + 2) { | |
70469b43 | 364 | show_usage(argv); |
453629a9 MD |
365 | return -1; |
366 | } | |
367 | a = atoi(argv[++i]); | |
368 | cpu_affinities[next_aff++] = a; | |
369 | use_affinity = 1; | |
370 | printf_verbose("Adding CPU %d affinity\n", a); | |
371 | break; | |
372 | case 'c': | |
373 | if (argc < i + 2) { | |
70469b43 | 374 | show_usage(argv); |
453629a9 MD |
375 | return -1; |
376 | } | |
377 | rduration = atol(argv[++i]); | |
378 | break; | |
379 | case 'd': | |
380 | if (argc < i + 2) { | |
70469b43 | 381 | show_usage(argv); |
453629a9 MD |
382 | return -1; |
383 | } | |
384 | wdelay = atol(argv[++i]); | |
385 | break; | |
386 | case 'v': | |
387 | verbose_mode = 1; | |
388 | break; | |
111ce0c3 MD |
389 | case 'p': |
390 | test_pop = 1; | |
391 | break; | |
392 | case 'P': | |
393 | test_pop_all = 1; | |
394 | break; | |
395 | case 'R': | |
396 | test_sync = TEST_SYNC_RCU; | |
397 | break; | |
453629a9 MD |
398 | } |
399 | } | |
400 | ||
111ce0c3 MD |
401 | /* activate pop_all test by default */ |
402 | if (!test_pop && !test_pop_all) | |
403 | test_pop_all = 1; | |
404 | ||
453629a9 MD |
405 | printf_verbose("running test for %lu seconds, %u enqueuers, " |
406 | "%u dequeuers.\n", | |
407 | duration, nr_enqueuers, nr_dequeuers); | |
111ce0c3 MD |
408 | if (test_pop) |
409 | printf_verbose("pop test activated.\n"); | |
410 | if (test_pop_all) | |
411 | printf_verbose("pop_all test activated.\n"); | |
eb314468 MD |
412 | if (test_sync == TEST_SYNC_RCU) |
413 | printf_verbose("External sync: RCU.\n"); | |
414 | else | |
415 | printf_verbose("External sync: none.\n"); | |
453629a9 MD |
416 | printf_verbose("Writer delay : %lu loops.\n", rduration); |
417 | printf_verbose("Reader duration : %lu loops.\n", wdelay); | |
94df6318 MD |
418 | printf_verbose("thread %-6s, tid %lu\n", |
419 | "main", urcu_get_thread_id()); | |
453629a9 | 420 | |
9aa14175 MD |
421 | tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer)); |
422 | tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer)); | |
423 | count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer)); | |
424 | count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer)); | |
cb8818f0 | 425 | cds_lfs_init(&s); |
94aec122 | 426 | err = create_all_cpu_call_rcu_data(0); |
8bcbd94a MD |
427 | if (err) { |
428 | printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n"); | |
429 | } | |
453629a9 MD |
430 | |
431 | next_aff = 0; | |
432 | ||
83e334d0 MJ |
433 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
434 | err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer, | |
435 | &count_enqueuer[2 * i_thr]); | |
453629a9 MD |
436 | if (err != 0) |
437 | exit(1); | |
438 | } | |
83e334d0 MJ |
439 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
440 | err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer, | |
441 | &count_dequeuer[2 * i_thr]); | |
453629a9 MD |
442 | if (err != 0) |
443 | exit(1); | |
444 | } | |
445 | ||
5481ddb3 | 446 | cmm_smp_mb(); |
453629a9 MD |
447 | |
448 | test_go = 1; | |
449 | ||
83e334d0 | 450 | for (i_thr = 0; i_thr < duration; i_thr++) { |
453629a9 | 451 | sleep(1); |
4b665350 MD |
452 | if (verbose_mode) { |
453 | fwrite(".", sizeof(char), 1, stdout); | |
454 | fflush(stdout); | |
455 | } | |
453629a9 MD |
456 | } |
457 | ||
458 | test_stop = 1; | |
459 | ||
83e334d0 MJ |
460 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
461 | err = pthread_join(tid_enqueuer[i_thr], &tret); | |
453629a9 MD |
462 | if (err != 0) |
463 | exit(1); | |
83e334d0 MJ |
464 | tot_enqueues += count_enqueuer[2 * i_thr]; |
465 | tot_successful_enqueues += count_enqueuer[2 * i_thr + 1]; | |
453629a9 | 466 | } |
83e334d0 MJ |
467 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
468 | err = pthread_join(tid_dequeuer[i_thr], &tret); | |
453629a9 MD |
469 | if (err != 0) |
470 | exit(1); | |
83e334d0 MJ |
471 | tot_dequeues += count_dequeuer[2 * i_thr]; |
472 | tot_successful_dequeues += count_dequeuer[2 * i_thr + 1]; | |
453629a9 | 473 | } |
83e334d0 | 474 | |
bd23a6c6 | 475 | test_end(&end_dequeues); |
453629a9 MD |
476 | |
477 | printf_verbose("total number of enqueues : %llu, dequeues %llu\n", | |
478 | tot_enqueues, tot_dequeues); | |
479 | printf_verbose("total number of successful enqueues : %llu, " | |
480 | "successful dequeues %llu\n", | |
481 | tot_successful_enqueues, tot_successful_dequeues); | |
482 | printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu " | |
483 | "nr_dequeuers %3u " | |
484 | "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu " | |
485 | "successful enqueues %12llu successful dequeues %12llu " | |
486 | "end_dequeues %llu nr_ops %12llu\n", | |
487 | argv[0], duration, nr_enqueuers, wdelay, | |
488 | nr_dequeuers, rduration, tot_enqueues, tot_dequeues, | |
489 | tot_successful_enqueues, | |
490 | tot_successful_dequeues, end_dequeues, | |
491 | tot_enqueues + tot_dequeues); | |
492 | if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) | |
493 | printf("WARNING! Discrepancy between nr succ. enqueues %llu vs " | |
494 | "succ. dequeues + end dequeues %llu.\n", | |
495 | tot_successful_enqueues, | |
496 | tot_successful_dequeues + end_dequeues); | |
497 | ||
0938c541 | 498 | free_all_cpu_call_rcu_data(); |
2af1c19e | 499 | cds_lfs_destroy(&s); |
453629a9 MD |
500 | free(count_enqueuer); |
501 | free(count_dequeuer); | |
502 | free(tid_enqueuer); | |
503 | free(tid_dequeuer); | |
83e334d0 | 504 | |
453629a9 MD |
505 | return 0; |
506 | } |