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 | ||
cb8818f0 | 6 | /* |
cb8818f0 | 7 | * Userspace RCU library - example RCU-based lock-free stack |
cb8818f0 MD |
8 | */ |
9 | ||
cb8818f0 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> | |
cb8818f0 MD |
20 | #include <errno.h> |
21 | ||
22 | #include <urcu/arch.h> | |
01477510 | 23 | #include <urcu/assert.h> |
cb8818f0 | 24 | #include <urcu/tls-compat.h> |
94df6318 | 25 | #include "thread-id.h" |
cb8818f0 MD |
26 | |
27 | /* hardcoded number of CPUs */ | |
28 | #define NR_CPUS 16384 | |
29 | ||
cb8818f0 MD |
30 | #ifndef DYNAMIC_LINK_TEST |
31 | #define _LGPL_SOURCE | |
32 | #endif | |
33 | #include <urcu.h> | |
d89ec762 MD |
34 | |
35 | /* Remove deprecation warnings from test build. */ | |
36 | #define CDS_LFS_RCU_DEPRECATED | |
37 | ||
cb8818f0 MD |
38 | #include <urcu/cds.h> |
39 | ||
40 | static volatile int test_go, test_stop; | |
41 | ||
42 | static unsigned long rduration; | |
43 | ||
44 | static unsigned long duration; | |
45 | ||
46 | /* read-side C.S. duration, in loops */ | |
47 | static unsigned long wdelay; | |
48 | ||
49 | static inline void loop_sleep(unsigned long loops) | |
50 | { | |
51 | while (loops-- != 0) | |
52 | caa_cpu_relax(); | |
53 | } | |
54 | ||
55 | static int verbose_mode; | |
56 | ||
57 | #define printf_verbose(fmt, args...) \ | |
58 | do { \ | |
59 | if (verbose_mode) \ | |
60 | printf(fmt, args); \ | |
61 | } while (0) | |
62 | ||
63 | static unsigned int cpu_affinities[NR_CPUS]; | |
64 | static unsigned int next_aff = 0; | |
65 | static int use_affinity = 0; | |
66 | ||
67 | pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER; | |
68 | ||
cb8818f0 MD |
69 | static void set_affinity(void) |
70 | { | |
2388c075 | 71 | #ifdef HAVE_SCHED_SETAFFINITY |
cb8818f0 | 72 | cpu_set_t mask; |
ed0d6f2b MJ |
73 | int cpu, ret; |
74 | #endif /* HAVE_SCHED_SETAFFINITY */ | |
cb8818f0 MD |
75 | |
76 | if (!use_affinity) | |
77 | return; | |
78 | ||
2388c075 | 79 | #ifdef HAVE_SCHED_SETAFFINITY |
cb8818f0 MD |
80 | ret = pthread_mutex_lock(&affinity_mutex); |
81 | if (ret) { | |
82 | perror("Error in pthread mutex lock"); | |
83 | exit(-1); | |
84 | } | |
85 | cpu = cpu_affinities[next_aff++]; | |
86 | ret = pthread_mutex_unlock(&affinity_mutex); | |
87 | if (ret) { | |
88 | perror("Error in pthread mutex unlock"); | |
89 | exit(-1); | |
90 | } | |
91 | ||
92 | CPU_ZERO(&mask); | |
93 | CPU_SET(cpu, &mask); | |
cb8818f0 | 94 | sched_setaffinity(0, sizeof(mask), &mask); |
cb8818f0 MD |
95 | #endif /* HAVE_SCHED_SETAFFINITY */ |
96 | } | |
97 | ||
98 | /* | |
99 | * returns 0 if test should end. | |
100 | */ | |
101 | static int test_duration_dequeue(void) | |
102 | { | |
103 | return !test_stop; | |
104 | } | |
105 | ||
106 | static int test_duration_enqueue(void) | |
107 | { | |
108 | return !test_stop; | |
109 | } | |
110 | ||
111 | static DEFINE_URCU_TLS(unsigned long long, nr_dequeues); | |
112 | static DEFINE_URCU_TLS(unsigned long long, nr_enqueues); | |
113 | ||
114 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues); | |
115 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues); | |
116 | ||
117 | static unsigned int nr_enqueuers; | |
118 | static unsigned int nr_dequeuers; | |
119 | ||
120 | struct test { | |
121 | struct cds_lfs_node_rcu list; | |
122 | struct rcu_head rcu; | |
123 | }; | |
124 | ||
125 | static struct cds_lfs_stack_rcu s; | |
126 | ||
61c3fb60 | 127 | static |
cb8818f0 MD |
128 | void *thr_enqueuer(void *_count) |
129 | { | |
130 | unsigned long long *count = _count; | |
131 | ||
94df6318 MD |
132 | printf_verbose("thread_begin %s, tid %lu\n", |
133 | "enqueuer", urcu_get_thread_id()); | |
cb8818f0 MD |
134 | |
135 | set_affinity(); | |
136 | ||
137 | rcu_register_thread(); | |
138 | ||
139 | while (!test_go) | |
140 | { | |
141 | } | |
142 | cmm_smp_mb(); | |
143 | ||
144 | for (;;) { | |
145 | struct test *node = malloc(sizeof(*node)); | |
146 | if (!node) | |
147 | goto fail; | |
148 | cds_lfs_node_init_rcu(&node->list); | |
149 | /* No rcu read-side is needed for push */ | |
150 | cds_lfs_push_rcu(&s, &node->list); | |
151 | URCU_TLS(nr_successful_enqueues)++; | |
152 | ||
153 | if (caa_unlikely(wdelay)) | |
154 | loop_sleep(wdelay); | |
155 | fail: | |
156 | URCU_TLS(nr_enqueues)++; | |
157 | if (caa_unlikely(!test_duration_enqueue())) | |
158 | break; | |
159 | } | |
160 | ||
161 | rcu_unregister_thread(); | |
162 | ||
163 | count[0] = URCU_TLS(nr_enqueues); | |
164 | count[1] = URCU_TLS(nr_successful_enqueues); | |
94df6318 MD |
165 | printf_verbose("enqueuer thread_end, tid %lu, " |
166 | "enqueues %llu successful_enqueues %llu\n", | |
167 | urcu_get_thread_id(), | |
168 | URCU_TLS(nr_enqueues), | |
169 | URCU_TLS(nr_successful_enqueues)); | |
cb8818f0 MD |
170 | return ((void*)1); |
171 | ||
172 | } | |
173 | ||
174 | static | |
175 | void free_node_cb(struct rcu_head *head) | |
176 | { | |
177 | struct test *node = | |
178 | caa_container_of(head, struct test, rcu); | |
179 | free(node); | |
180 | } | |
181 | ||
61c3fb60 | 182 | static |
cb8818f0 MD |
183 | void *thr_dequeuer(void *_count) |
184 | { | |
185 | unsigned long long *count = _count; | |
186 | ||
94df6318 MD |
187 | printf_verbose("thread_begin %s, tid %lu\n", |
188 | "dequeuer", urcu_get_thread_id()); | |
cb8818f0 MD |
189 | |
190 | set_affinity(); | |
191 | ||
192 | rcu_register_thread(); | |
193 | ||
194 | while (!test_go) | |
195 | { | |
196 | } | |
197 | cmm_smp_mb(); | |
198 | ||
199 | for (;;) { | |
200 | struct cds_lfs_node_rcu *snode; | |
201 | ||
202 | rcu_read_lock(); | |
203 | snode = cds_lfs_pop_rcu(&s); | |
204 | rcu_read_unlock(); | |
205 | if (snode) { | |
206 | struct test *node; | |
207 | ||
208 | node = caa_container_of(snode, struct test, list); | |
209 | call_rcu(&node->rcu, free_node_cb); | |
210 | URCU_TLS(nr_successful_dequeues)++; | |
211 | } | |
212 | URCU_TLS(nr_dequeues)++; | |
213 | if (caa_unlikely(!test_duration_dequeue())) | |
214 | break; | |
215 | if (caa_unlikely(rduration)) | |
216 | loop_sleep(rduration); | |
217 | } | |
218 | ||
219 | rcu_unregister_thread(); | |
220 | ||
94df6318 MD |
221 | printf_verbose("dequeuer thread_end, tid %lu, " |
222 | "dequeues %llu, successful_dequeues %llu\n", | |
223 | urcu_get_thread_id(), | |
224 | URCU_TLS(nr_dequeues), | |
225 | URCU_TLS(nr_successful_dequeues)); | |
cb8818f0 MD |
226 | count[0] = URCU_TLS(nr_dequeues); |
227 | count[1] = URCU_TLS(nr_successful_dequeues); | |
228 | return ((void*)2); | |
229 | } | |
230 | ||
61c3fb60 | 231 | static |
bd23a6c6 | 232 | void test_end(unsigned long long *nr_dequeues_l) |
cb8818f0 MD |
233 | { |
234 | struct cds_lfs_node_rcu *snode; | |
235 | ||
236 | do { | |
bd23a6c6 | 237 | snode = cds_lfs_pop_rcu(&s); |
cb8818f0 MD |
238 | if (snode) { |
239 | struct test *node; | |
240 | ||
241 | node = caa_container_of(snode, struct test, list); | |
242 | free(node); | |
bd23a6c6 | 243 | (*nr_dequeues_l)++; |
cb8818f0 MD |
244 | } |
245 | } while (snode); | |
246 | } | |
247 | ||
61c3fb60 | 248 | static |
70469b43 | 249 | void show_usage(char **argv) |
cb8818f0 | 250 | { |
06637138 MD |
251 | printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n", |
252 | argv[0]); | |
253 | printf("OPTIONS:\n"); | |
254 | printf(" [-d delay] (enqueuer period (in loops))\n"); | |
255 | printf(" [-c duration] (dequeuer period (in loops))\n"); | |
256 | printf(" [-v] (verbose output)\n"); | |
257 | printf(" [-a cpu#] [-a cpu#]... (affinity)\n"); | |
cb8818f0 MD |
258 | printf("\n"); |
259 | } | |
260 | ||
261 | int main(int argc, char **argv) | |
262 | { | |
263 | int err; | |
264 | pthread_t *tid_enqueuer, *tid_dequeuer; | |
265 | void *tret; | |
266 | unsigned long long *count_enqueuer, *count_dequeuer; | |
267 | unsigned long long tot_enqueues = 0, tot_dequeues = 0; | |
268 | unsigned long long tot_successful_enqueues = 0, | |
269 | tot_successful_dequeues = 0; | |
270 | unsigned long long end_dequeues = 0; | |
271 | int i, a; | |
83e334d0 | 272 | unsigned int i_thr; |
cb8818f0 MD |
273 | |
274 | if (argc < 4) { | |
70469b43 | 275 | show_usage(argv); |
cb8818f0 MD |
276 | return -1; |
277 | } | |
278 | ||
279 | err = sscanf(argv[1], "%u", &nr_dequeuers); | |
280 | if (err != 1) { | |
70469b43 | 281 | show_usage(argv); |
cb8818f0 MD |
282 | return -1; |
283 | } | |
284 | ||
285 | err = sscanf(argv[2], "%u", &nr_enqueuers); | |
286 | if (err != 1) { | |
70469b43 | 287 | show_usage(argv); |
cb8818f0 MD |
288 | return -1; |
289 | } | |
83e334d0 | 290 | |
cb8818f0 MD |
291 | err = sscanf(argv[3], "%lu", &duration); |
292 | if (err != 1) { | |
70469b43 | 293 | show_usage(argv); |
cb8818f0 MD |
294 | return -1; |
295 | } | |
296 | ||
297 | for (i = 4; i < argc; i++) { | |
298 | if (argv[i][0] != '-') | |
299 | continue; | |
300 | switch (argv[i][1]) { | |
301 | case 'a': | |
302 | if (argc < i + 2) { | |
70469b43 | 303 | show_usage(argv); |
cb8818f0 MD |
304 | return -1; |
305 | } | |
306 | a = atoi(argv[++i]); | |
307 | cpu_affinities[next_aff++] = a; | |
308 | use_affinity = 1; | |
309 | printf_verbose("Adding CPU %d affinity\n", a); | |
310 | break; | |
311 | case 'c': | |
312 | if (argc < i + 2) { | |
70469b43 | 313 | show_usage(argv); |
cb8818f0 MD |
314 | return -1; |
315 | } | |
316 | rduration = atol(argv[++i]); | |
317 | break; | |
318 | case 'd': | |
319 | if (argc < i + 2) { | |
70469b43 | 320 | show_usage(argv); |
cb8818f0 MD |
321 | return -1; |
322 | } | |
323 | wdelay = atol(argv[++i]); | |
324 | break; | |
325 | case 'v': | |
326 | verbose_mode = 1; | |
327 | break; | |
328 | } | |
329 | } | |
330 | ||
331 | printf_verbose("running test for %lu seconds, %u enqueuers, " | |
332 | "%u dequeuers.\n", | |
333 | duration, nr_enqueuers, nr_dequeuers); | |
334 | printf_verbose("Writer delay : %lu loops.\n", rduration); | |
335 | printf_verbose("Reader duration : %lu loops.\n", wdelay); | |
94df6318 MD |
336 | printf_verbose("thread %-6s, tid %lu\n", |
337 | "main", urcu_get_thread_id()); | |
cb8818f0 | 338 | |
9aa14175 MD |
339 | tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer)); |
340 | tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer)); | |
341 | count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer)); | |
342 | count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer)); | |
cb8818f0 MD |
343 | cds_lfs_init_rcu(&s); |
344 | err = create_all_cpu_call_rcu_data(0); | |
345 | if (err) { | |
346 | printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n"); | |
347 | } | |
348 | ||
349 | next_aff = 0; | |
350 | ||
83e334d0 MJ |
351 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
352 | err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer, | |
353 | &count_enqueuer[2 * i_thr]); | |
cb8818f0 MD |
354 | if (err != 0) |
355 | exit(1); | |
356 | } | |
83e334d0 MJ |
357 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
358 | err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer, | |
359 | &count_dequeuer[2 * i_thr]); | |
cb8818f0 MD |
360 | if (err != 0) |
361 | exit(1); | |
362 | } | |
363 | ||
364 | cmm_smp_mb(); | |
365 | ||
366 | test_go = 1; | |
367 | ||
83e334d0 | 368 | for (i_thr = 0; i_thr < duration; i_thr++) { |
cb8818f0 | 369 | sleep(1); |
4b665350 MD |
370 | if (verbose_mode) { |
371 | fwrite(".", sizeof(char), 1, stdout); | |
372 | fflush(stdout); | |
373 | } | |
cb8818f0 MD |
374 | } |
375 | ||
376 | test_stop = 1; | |
377 | ||
83e334d0 MJ |
378 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
379 | err = pthread_join(tid_enqueuer[i_thr], &tret); | |
cb8818f0 MD |
380 | if (err != 0) |
381 | exit(1); | |
83e334d0 MJ |
382 | tot_enqueues += count_enqueuer[2 * i_thr]; |
383 | tot_successful_enqueues += count_enqueuer[2 * i_thr + 1]; | |
cb8818f0 | 384 | } |
83e334d0 MJ |
385 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
386 | err = pthread_join(tid_dequeuer[i_thr], &tret); | |
cb8818f0 MD |
387 | if (err != 0) |
388 | exit(1); | |
83e334d0 MJ |
389 | tot_dequeues += count_dequeuer[2 * i_thr]; |
390 | tot_successful_dequeues += count_dequeuer[2 * i_thr + 1]; | |
cb8818f0 | 391 | } |
83e334d0 | 392 | |
bd23a6c6 | 393 | test_end(&end_dequeues); |
cb8818f0 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 | ||
416 | free_all_cpu_call_rcu_data(); | |
417 | free(count_enqueuer); | |
418 | free(count_dequeuer); | |
419 | free(tid_enqueuer); | |
420 | free(tid_dequeuer); | |
421 | return 0; | |
422 | } |