Commit | Line | Data |
---|---|---|
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 | ||
453629a9 | 6 | /* |
453629a9 | 7 | * Userspace RCU library - example RCU-based lock-free queue |
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 MD |
35 | |
36 | static volatile int test_go, test_stop; | |
37 | ||
38 | static unsigned long rduration; | |
39 | ||
40 | static unsigned long duration; | |
41 | ||
42 | /* read-side C.S. duration, in loops */ | |
43 | static unsigned long wdelay; | |
44 | ||
ab0aacbe | 45 | static inline void loop_sleep(unsigned long loops) |
453629a9 | 46 | { |
ab0aacbe | 47 | while (loops-- != 0) |
06f22bdb | 48 | caa_cpu_relax(); |
453629a9 MD |
49 | } |
50 | ||
51 | static int verbose_mode; | |
52 | ||
53 | #define printf_verbose(fmt, args...) \ | |
54 | do { \ | |
55 | if (verbose_mode) \ | |
56 | printf(fmt, args); \ | |
57 | } while (0) | |
58 | ||
59 | static unsigned int cpu_affinities[NR_CPUS]; | |
60 | static unsigned int next_aff = 0; | |
61 | static int use_affinity = 0; | |
62 | ||
63 | pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER; | |
64 | ||
453629a9 MD |
65 | static void set_affinity(void) |
66 | { | |
2388c075 | 67 | #ifdef HAVE_SCHED_SETAFFINITY |
453629a9 | 68 | cpu_set_t mask; |
95bc7fb9 MD |
69 | int cpu, ret; |
70 | #endif /* HAVE_SCHED_SETAFFINITY */ | |
453629a9 MD |
71 | |
72 | if (!use_affinity) | |
73 | return; | |
74 | ||
2388c075 | 75 | #ifdef HAVE_SCHED_SETAFFINITY |
453629a9 MD |
76 | ret = pthread_mutex_lock(&affinity_mutex); |
77 | if (ret) { | |
78 | perror("Error in pthread mutex lock"); | |
79 | exit(-1); | |
80 | } | |
81 | cpu = cpu_affinities[next_aff++]; | |
82 | ret = pthread_mutex_unlock(&affinity_mutex); | |
83 | if (ret) { | |
84 | perror("Error in pthread mutex unlock"); | |
85 | exit(-1); | |
86 | } | |
87 | ||
88 | CPU_ZERO(&mask); | |
89 | CPU_SET(cpu, &mask); | |
453629a9 | 90 | sched_setaffinity(0, sizeof(mask), &mask); |
453629a9 MD |
91 | #endif /* HAVE_SCHED_SETAFFINITY */ |
92 | } | |
93 | ||
94 | /* | |
95 | * returns 0 if test should end. | |
96 | */ | |
97 | static int test_duration_dequeue(void) | |
98 | { | |
99 | return !test_stop; | |
100 | } | |
101 | ||
102 | static int test_duration_enqueue(void) | |
103 | { | |
104 | return !test_stop; | |
105 | } | |
106 | ||
bd252a04 MD |
107 | static DEFINE_URCU_TLS(unsigned long long, nr_dequeues); |
108 | static DEFINE_URCU_TLS(unsigned long long, nr_enqueues); | |
453629a9 | 109 | |
bd252a04 MD |
110 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues); |
111 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues); | |
453629a9 MD |
112 | |
113 | static unsigned int nr_enqueuers; | |
114 | static unsigned int nr_dequeuers; | |
115 | ||
94aec122 MD |
116 | struct test { |
117 | struct cds_lfq_node_rcu list; | |
118 | struct rcu_head rcu; | |
119 | }; | |
120 | ||
16aa9ee8 | 121 | static struct cds_lfq_queue_rcu q; |
453629a9 | 122 | |
61c3fb60 | 123 | static |
453629a9 MD |
124 | void *thr_enqueuer(void *_count) |
125 | { | |
126 | unsigned long long *count = _count; | |
127 | ||
94df6318 MD |
128 | printf_verbose("thread_begin %s, tid %lu\n", |
129 | "enqueuer", urcu_get_thread_id()); | |
453629a9 MD |
130 | |
131 | set_affinity(); | |
132 | ||
133 | rcu_register_thread(); | |
134 | ||
135 | while (!test_go) | |
136 | { | |
137 | } | |
5481ddb3 | 138 | cmm_smp_mb(); |
453629a9 MD |
139 | |
140 | for (;;) { | |
94aec122 | 141 | struct test *node = malloc(sizeof(*node)); |
453629a9 MD |
142 | if (!node) |
143 | goto fail; | |
94aec122 | 144 | cds_lfq_node_init_rcu(&node->list); |
6e5f88cf | 145 | rcu_read_lock(); |
94aec122 | 146 | cds_lfq_enqueue_rcu(&q, &node->list); |
6e5f88cf | 147 | rcu_read_unlock(); |
bd252a04 | 148 | URCU_TLS(nr_successful_enqueues)++; |
453629a9 | 149 | |
a0b7f7ea | 150 | if (caa_unlikely(wdelay)) |
453629a9 MD |
151 | loop_sleep(wdelay); |
152 | fail: | |
bd252a04 | 153 | URCU_TLS(nr_enqueues)++; |
a0b7f7ea | 154 | if (caa_unlikely(!test_duration_enqueue())) |
453629a9 MD |
155 | break; |
156 | } | |
157 | ||
158 | rcu_unregister_thread(); | |
159 | ||
bd252a04 MD |
160 | count[0] = URCU_TLS(nr_enqueues); |
161 | count[1] = URCU_TLS(nr_successful_enqueues); | |
94df6318 MD |
162 | printf_verbose("enqueuer thread_end, tid %lu, " |
163 | "enqueues %llu successful_enqueues %llu\n", | |
164 | urcu_get_thread_id(), | |
165 | URCU_TLS(nr_enqueues), | |
166 | URCU_TLS(nr_successful_enqueues)); | |
453629a9 MD |
167 | return ((void*)1); |
168 | ||
169 | } | |
170 | ||
94aec122 MD |
171 | static |
172 | void free_node_cb(struct rcu_head *head) | |
173 | { | |
174 | struct test *node = | |
175 | caa_container_of(head, struct test, rcu); | |
176 | free(node); | |
177 | } | |
178 | ||
61c3fb60 | 179 | static |
453629a9 MD |
180 | void *thr_dequeuer(void *_count) |
181 | { | |
182 | unsigned long long *count = _count; | |
183 | ||
94df6318 MD |
184 | printf_verbose("thread_begin %s, tid %lu\n", |
185 | "dequeuer", urcu_get_thread_id()); | |
453629a9 MD |
186 | |
187 | set_affinity(); | |
188 | ||
453629a9 MD |
189 | rcu_register_thread(); |
190 | ||
191 | while (!test_go) | |
192 | { | |
193 | } | |
5481ddb3 | 194 | cmm_smp_mb(); |
453629a9 MD |
195 | |
196 | for (;;) { | |
94aec122 | 197 | struct cds_lfq_node_rcu *qnode; |
d9b52143 | 198 | |
6e5f88cf | 199 | rcu_read_lock(); |
94aec122 | 200 | qnode = cds_lfq_dequeue_rcu(&q); |
6e5f88cf MD |
201 | rcu_read_unlock(); |
202 | ||
747b5930 LJ |
203 | if (qnode) { |
204 | struct test *node; | |
205 | ||
206 | node = caa_container_of(qnode, struct test, list); | |
94aec122 | 207 | call_rcu(&node->rcu, free_node_cb); |
bd252a04 | 208 | URCU_TLS(nr_successful_dequeues)++; |
453629a9 MD |
209 | } |
210 | ||
bd252a04 | 211 | URCU_TLS(nr_dequeues)++; |
a0b7f7ea | 212 | if (caa_unlikely(!test_duration_dequeue())) |
453629a9 | 213 | break; |
a0b7f7ea | 214 | if (caa_unlikely(rduration)) |
453629a9 MD |
215 | loop_sleep(rduration); |
216 | } | |
217 | ||
218 | rcu_unregister_thread(); | |
94df6318 MD |
219 | printf_verbose("dequeuer thread_end, tid %lu, " |
220 | "dequeues %llu, successful_dequeues %llu\n", | |
221 | urcu_get_thread_id(), | |
222 | URCU_TLS(nr_dequeues), | |
223 | URCU_TLS(nr_successful_dequeues)); | |
bd252a04 MD |
224 | count[0] = URCU_TLS(nr_dequeues); |
225 | count[1] = URCU_TLS(nr_successful_dequeues); | |
453629a9 MD |
226 | return ((void*)2); |
227 | } | |
228 | ||
61c3fb60 | 229 | static |
bd23a6c6 | 230 | void test_end(unsigned long long *nr_dequeues_l) |
453629a9 | 231 | { |
94aec122 | 232 | struct cds_lfq_node_rcu *snode; |
453629a9 MD |
233 | |
234 | do { | |
bd23a6c6 | 235 | snode = cds_lfq_dequeue_rcu(&q); |
94aec122 MD |
236 | if (snode) { |
237 | struct test *node; | |
238 | ||
239 | node = caa_container_of(snode, struct test, list); | |
e17d9985 | 240 | free(node); /* no more concurrent access */ |
bd23a6c6 | 241 | (*nr_dequeues_l)++; |
453629a9 | 242 | } |
94aec122 | 243 | } while (snode); |
453629a9 MD |
244 | } |
245 | ||
61c3fb60 | 246 | static |
70469b43 | 247 | void show_usage(char **argv) |
453629a9 | 248 | { |
06637138 MD |
249 | printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n", |
250 | argv[0]); | |
251 | printf("OPTIONS:\n"); | |
252 | printf(" [-d delay] (enqueuer period (in loops))\n"); | |
253 | printf(" [-c duration] (dequeuer period (in loops))\n"); | |
254 | printf(" [-v] (verbose output)\n"); | |
255 | printf(" [-a cpu#] [-a cpu#]... (affinity)\n"); | |
453629a9 MD |
256 | printf("\n"); |
257 | } | |
258 | ||
259 | int main(int argc, char **argv) | |
260 | { | |
261 | int err; | |
262 | pthread_t *tid_enqueuer, *tid_dequeuer; | |
263 | void *tret; | |
264 | unsigned long long *count_enqueuer, *count_dequeuer; | |
265 | unsigned long long tot_enqueues = 0, tot_dequeues = 0; | |
266 | unsigned long long tot_successful_enqueues = 0, | |
267 | tot_successful_dequeues = 0; | |
268 | unsigned long long end_dequeues = 0; | |
269 | int i, a; | |
83e334d0 | 270 | unsigned int i_thr; |
453629a9 MD |
271 | |
272 | if (argc < 4) { | |
70469b43 | 273 | show_usage(argv); |
453629a9 MD |
274 | return -1; |
275 | } | |
276 | ||
277 | err = sscanf(argv[1], "%u", &nr_dequeuers); | |
278 | if (err != 1) { | |
70469b43 | 279 | show_usage(argv); |
453629a9 MD |
280 | return -1; |
281 | } | |
282 | ||
283 | err = sscanf(argv[2], "%u", &nr_enqueuers); | |
284 | if (err != 1) { | |
70469b43 | 285 | show_usage(argv); |
453629a9 MD |
286 | return -1; |
287 | } | |
83e334d0 | 288 | |
453629a9 MD |
289 | err = sscanf(argv[3], "%lu", &duration); |
290 | if (err != 1) { | |
70469b43 | 291 | show_usage(argv); |
453629a9 MD |
292 | return -1; |
293 | } | |
294 | ||
295 | for (i = 4; i < argc; i++) { | |
296 | if (argv[i][0] != '-') | |
297 | continue; | |
298 | switch (argv[i][1]) { | |
299 | case 'a': | |
300 | if (argc < i + 2) { | |
70469b43 | 301 | show_usage(argv); |
453629a9 MD |
302 | return -1; |
303 | } | |
304 | a = atoi(argv[++i]); | |
305 | cpu_affinities[next_aff++] = a; | |
306 | use_affinity = 1; | |
307 | printf_verbose("Adding CPU %d affinity\n", a); | |
308 | break; | |
309 | case 'c': | |
310 | if (argc < i + 2) { | |
70469b43 | 311 | show_usage(argv); |
453629a9 MD |
312 | return -1; |
313 | } | |
314 | rduration = atol(argv[++i]); | |
315 | break; | |
316 | case 'd': | |
317 | if (argc < i + 2) { | |
70469b43 | 318 | show_usage(argv); |
453629a9 MD |
319 | return -1; |
320 | } | |
321 | wdelay = atol(argv[++i]); | |
322 | break; | |
323 | case 'v': | |
324 | verbose_mode = 1; | |
325 | break; | |
326 | } | |
327 | } | |
328 | ||
329 | printf_verbose("running test for %lu seconds, %u enqueuers, " | |
330 | "%u dequeuers.\n", | |
331 | duration, nr_enqueuers, nr_dequeuers); | |
332 | printf_verbose("Writer delay : %lu loops.\n", rduration); | |
333 | printf_verbose("Reader duration : %lu loops.\n", wdelay); | |
94df6318 MD |
334 | printf_verbose("thread %-6s, tid %lu\n", |
335 | "main", urcu_get_thread_id()); | |
453629a9 | 336 | |
9aa14175 MD |
337 | tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer)); |
338 | tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer)); | |
339 | count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer)); | |
340 | count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer)); | |
6e5f88cf | 341 | cds_lfq_init_rcu(&q, call_rcu); |
94aec122 | 342 | err = create_all_cpu_call_rcu_data(0); |
8bcbd94a MD |
343 | if (err) { |
344 | printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n"); | |
345 | } | |
453629a9 MD |
346 | |
347 | next_aff = 0; | |
348 | ||
83e334d0 MJ |
349 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
350 | err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer, | |
351 | &count_enqueuer[2 * i_thr]); | |
453629a9 MD |
352 | if (err != 0) |
353 | exit(1); | |
354 | } | |
83e334d0 MJ |
355 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
356 | err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer, | |
357 | &count_dequeuer[2 * i_thr]); | |
453629a9 MD |
358 | if (err != 0) |
359 | exit(1); | |
360 | } | |
361 | ||
5481ddb3 | 362 | cmm_smp_mb(); |
453629a9 MD |
363 | |
364 | test_go = 1; | |
365 | ||
83e334d0 | 366 | for (i_thr = 0; i_thr < duration; i_thr++) { |
453629a9 | 367 | sleep(1); |
4b665350 MD |
368 | if (verbose_mode) { |
369 | fwrite(".", sizeof(char), 1, stdout); | |
370 | fflush(stdout); | |
371 | } | |
453629a9 MD |
372 | } |
373 | ||
374 | test_stop = 1; | |
375 | ||
83e334d0 MJ |
376 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
377 | err = pthread_join(tid_enqueuer[i_thr], &tret); | |
453629a9 MD |
378 | if (err != 0) |
379 | exit(1); | |
83e334d0 MJ |
380 | tot_enqueues += count_enqueuer[2 * i_thr]; |
381 | tot_successful_enqueues += count_enqueuer[2 * i_thr + 1]; | |
453629a9 | 382 | } |
83e334d0 MJ |
383 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
384 | err = pthread_join(tid_dequeuer[i_thr], &tret); | |
453629a9 MD |
385 | if (err != 0) |
386 | exit(1); | |
83e334d0 MJ |
387 | tot_dequeues += count_dequeuer[2 * i_thr]; |
388 | tot_successful_dequeues += count_dequeuer[2 * i_thr + 1]; | |
453629a9 | 389 | } |
83e334d0 | 390 | |
bd23a6c6 | 391 | test_end(&end_dequeues); |
e17d9985 | 392 | err = cds_lfq_destroy_rcu(&q); |
01477510 | 393 | urcu_posix_assert(!err); |
453629a9 MD |
394 | |
395 | printf_verbose("total number of enqueues : %llu, dequeues %llu\n", | |
396 | tot_enqueues, tot_dequeues); | |
397 | printf_verbose("total number of successful enqueues : %llu, " | |
398 | "successful dequeues %llu\n", | |
399 | tot_successful_enqueues, tot_successful_dequeues); | |
400 | printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu " | |
401 | "nr_dequeuers %3u " | |
402 | "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu " | |
403 | "successful enqueues %12llu successful dequeues %12llu " | |
404 | "end_dequeues %llu nr_ops %12llu\n", | |
405 | argv[0], duration, nr_enqueuers, wdelay, | |
406 | nr_dequeuers, rduration, tot_enqueues, tot_dequeues, | |
407 | tot_successful_enqueues, | |
408 | tot_successful_dequeues, end_dequeues, | |
409 | tot_enqueues + tot_dequeues); | |
410 | if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) | |
411 | printf("WARNING! Discrepancy between nr succ. enqueues %llu vs " | |
412 | "succ. dequeues + end dequeues %llu.\n", | |
413 | tot_successful_enqueues, | |
414 | tot_successful_dequeues + end_dequeues); | |
415 | ||
0938c541 | 416 | free_all_cpu_call_rcu_data(); |
453629a9 MD |
417 | free(count_enqueuer); |
418 | free(count_dequeuer); | |
419 | free(tid_enqueuer); | |
420 | free(tid_dequeuer); | |
83e334d0 | 421 | |
453629a9 MD |
422 | return 0; |
423 | } |