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 | ||
cb8818f0 MD |
40 | static unsigned long rduration; |
41 | ||
42 | static unsigned long duration; | |
43 | ||
44 | /* read-side C.S. duration, in loops */ | |
45 | static unsigned long wdelay; | |
46 | ||
47 | static inline void loop_sleep(unsigned long loops) | |
48 | { | |
49 | while (loops-- != 0) | |
50 | caa_cpu_relax(); | |
51 | } | |
52 | ||
53 | static int verbose_mode; | |
54 | ||
55 | #define printf_verbose(fmt, args...) \ | |
56 | do { \ | |
57 | if (verbose_mode) \ | |
58 | printf(fmt, args); \ | |
59 | } while (0) | |
60 | ||
61 | static unsigned int cpu_affinities[NR_CPUS]; | |
62 | static unsigned int next_aff = 0; | |
63 | static int use_affinity = 0; | |
64 | ||
65 | pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER; | |
66 | ||
cb8818f0 MD |
67 | static void set_affinity(void) |
68 | { | |
2388c075 | 69 | #ifdef HAVE_SCHED_SETAFFINITY |
cb8818f0 | 70 | cpu_set_t mask; |
ed0d6f2b MJ |
71 | int cpu, ret; |
72 | #endif /* HAVE_SCHED_SETAFFINITY */ | |
cb8818f0 MD |
73 | |
74 | if (!use_affinity) | |
75 | return; | |
76 | ||
2388c075 | 77 | #ifdef HAVE_SCHED_SETAFFINITY |
cb8818f0 MD |
78 | ret = pthread_mutex_lock(&affinity_mutex); |
79 | if (ret) { | |
80 | perror("Error in pthread mutex lock"); | |
81 | exit(-1); | |
82 | } | |
83 | cpu = cpu_affinities[next_aff++]; | |
84 | ret = pthread_mutex_unlock(&affinity_mutex); | |
85 | if (ret) { | |
86 | perror("Error in pthread mutex unlock"); | |
87 | exit(-1); | |
88 | } | |
89 | ||
90 | CPU_ZERO(&mask); | |
91 | CPU_SET(cpu, &mask); | |
cb8818f0 | 92 | sched_setaffinity(0, sizeof(mask), &mask); |
cb8818f0 MD |
93 | #endif /* HAVE_SCHED_SETAFFINITY */ |
94 | } | |
95 | ||
96 | /* | |
97 | * returns 0 if test should end. | |
98 | */ | |
99 | static int test_duration_dequeue(void) | |
100 | { | |
9e4e7ad1 | 101 | return test_duration_read(); |
cb8818f0 MD |
102 | } |
103 | ||
104 | static int test_duration_enqueue(void) | |
105 | { | |
9e4e7ad1 | 106 | return test_duration_write(); |
cb8818f0 MD |
107 | } |
108 | ||
109 | static DEFINE_URCU_TLS(unsigned long long, nr_dequeues); | |
110 | static DEFINE_URCU_TLS(unsigned long long, nr_enqueues); | |
111 | ||
112 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues); | |
113 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues); | |
114 | ||
115 | static unsigned int nr_enqueuers; | |
116 | static unsigned int nr_dequeuers; | |
117 | ||
118 | struct test { | |
119 | struct cds_lfs_node_rcu list; | |
120 | struct rcu_head rcu; | |
121 | }; | |
122 | ||
123 | static struct cds_lfs_stack_rcu s; | |
124 | ||
61c3fb60 | 125 | static |
cb8818f0 MD |
126 | void *thr_enqueuer(void *_count) |
127 | { | |
128 | unsigned long long *count = _count; | |
129 | ||
94df6318 MD |
130 | printf_verbose("thread_begin %s, tid %lu\n", |
131 | "enqueuer", urcu_get_thread_id()); | |
cb8818f0 MD |
132 | |
133 | set_affinity(); | |
134 | ||
135 | rcu_register_thread(); | |
136 | ||
9e4e7ad1 | 137 | wait_until_go(); |
cb8818f0 MD |
138 | |
139 | for (;;) { | |
140 | struct test *node = malloc(sizeof(*node)); | |
141 | if (!node) | |
142 | goto fail; | |
143 | cds_lfs_node_init_rcu(&node->list); | |
144 | /* No rcu read-side is needed for push */ | |
145 | cds_lfs_push_rcu(&s, &node->list); | |
146 | URCU_TLS(nr_successful_enqueues)++; | |
147 | ||
148 | if (caa_unlikely(wdelay)) | |
149 | loop_sleep(wdelay); | |
150 | fail: | |
151 | URCU_TLS(nr_enqueues)++; | |
152 | if (caa_unlikely(!test_duration_enqueue())) | |
153 | break; | |
154 | } | |
155 | ||
156 | rcu_unregister_thread(); | |
157 | ||
158 | count[0] = URCU_TLS(nr_enqueues); | |
159 | count[1] = URCU_TLS(nr_successful_enqueues); | |
94df6318 MD |
160 | printf_verbose("enqueuer thread_end, tid %lu, " |
161 | "enqueues %llu successful_enqueues %llu\n", | |
162 | urcu_get_thread_id(), | |
163 | URCU_TLS(nr_enqueues), | |
164 | URCU_TLS(nr_successful_enqueues)); | |
cb8818f0 MD |
165 | return ((void*)1); |
166 | ||
167 | } | |
168 | ||
169 | static | |
170 | void free_node_cb(struct rcu_head *head) | |
171 | { | |
172 | struct test *node = | |
173 | caa_container_of(head, struct test, rcu); | |
174 | free(node); | |
175 | } | |
176 | ||
61c3fb60 | 177 | static |
cb8818f0 MD |
178 | void *thr_dequeuer(void *_count) |
179 | { | |
180 | unsigned long long *count = _count; | |
181 | ||
94df6318 MD |
182 | printf_verbose("thread_begin %s, tid %lu\n", |
183 | "dequeuer", urcu_get_thread_id()); | |
cb8818f0 MD |
184 | |
185 | set_affinity(); | |
186 | ||
187 | rcu_register_thread(); | |
188 | ||
9e4e7ad1 | 189 | wait_until_go(); |
cb8818f0 MD |
190 | |
191 | for (;;) { | |
192 | struct cds_lfs_node_rcu *snode; | |
193 | ||
194 | rcu_read_lock(); | |
195 | snode = cds_lfs_pop_rcu(&s); | |
196 | rcu_read_unlock(); | |
197 | if (snode) { | |
198 | struct test *node; | |
199 | ||
200 | node = caa_container_of(snode, struct test, list); | |
201 | call_rcu(&node->rcu, free_node_cb); | |
202 | URCU_TLS(nr_successful_dequeues)++; | |
203 | } | |
204 | URCU_TLS(nr_dequeues)++; | |
205 | if (caa_unlikely(!test_duration_dequeue())) | |
206 | break; | |
207 | if (caa_unlikely(rduration)) | |
208 | loop_sleep(rduration); | |
209 | } | |
210 | ||
211 | rcu_unregister_thread(); | |
212 | ||
94df6318 MD |
213 | printf_verbose("dequeuer thread_end, tid %lu, " |
214 | "dequeues %llu, successful_dequeues %llu\n", | |
215 | urcu_get_thread_id(), | |
216 | URCU_TLS(nr_dequeues), | |
217 | URCU_TLS(nr_successful_dequeues)); | |
cb8818f0 MD |
218 | count[0] = URCU_TLS(nr_dequeues); |
219 | count[1] = URCU_TLS(nr_successful_dequeues); | |
220 | return ((void*)2); | |
221 | } | |
222 | ||
61c3fb60 | 223 | static |
bd23a6c6 | 224 | void test_end(unsigned long long *nr_dequeues_l) |
cb8818f0 MD |
225 | { |
226 | struct cds_lfs_node_rcu *snode; | |
227 | ||
228 | do { | |
bd23a6c6 | 229 | snode = cds_lfs_pop_rcu(&s); |
cb8818f0 MD |
230 | if (snode) { |
231 | struct test *node; | |
232 | ||
233 | node = caa_container_of(snode, struct test, list); | |
234 | free(node); | |
bd23a6c6 | 235 | (*nr_dequeues_l)++; |
cb8818f0 MD |
236 | } |
237 | } while (snode); | |
238 | } | |
239 | ||
61c3fb60 | 240 | static |
70469b43 | 241 | void show_usage(char **argv) |
cb8818f0 | 242 | { |
06637138 MD |
243 | printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n", |
244 | argv[0]); | |
245 | printf("OPTIONS:\n"); | |
246 | printf(" [-d delay] (enqueuer period (in loops))\n"); | |
247 | printf(" [-c duration] (dequeuer period (in loops))\n"); | |
248 | printf(" [-v] (verbose output)\n"); | |
249 | printf(" [-a cpu#] [-a cpu#]... (affinity)\n"); | |
cb8818f0 MD |
250 | printf("\n"); |
251 | } | |
252 | ||
253 | int main(int argc, char **argv) | |
254 | { | |
255 | int err; | |
256 | pthread_t *tid_enqueuer, *tid_dequeuer; | |
257 | void *tret; | |
258 | unsigned long long *count_enqueuer, *count_dequeuer; | |
259 | unsigned long long tot_enqueues = 0, tot_dequeues = 0; | |
260 | unsigned long long tot_successful_enqueues = 0, | |
261 | tot_successful_dequeues = 0; | |
262 | unsigned long long end_dequeues = 0; | |
263 | int i, a; | |
83e334d0 | 264 | unsigned int i_thr; |
cb8818f0 MD |
265 | |
266 | if (argc < 4) { | |
70469b43 | 267 | show_usage(argv); |
cb8818f0 MD |
268 | return -1; |
269 | } | |
270 | ||
271 | err = sscanf(argv[1], "%u", &nr_dequeuers); | |
272 | if (err != 1) { | |
70469b43 | 273 | show_usage(argv); |
cb8818f0 MD |
274 | return -1; |
275 | } | |
276 | ||
277 | err = sscanf(argv[2], "%u", &nr_enqueuers); | |
278 | if (err != 1) { | |
70469b43 | 279 | show_usage(argv); |
cb8818f0 MD |
280 | return -1; |
281 | } | |
83e334d0 | 282 | |
cb8818f0 MD |
283 | err = sscanf(argv[3], "%lu", &duration); |
284 | if (err != 1) { | |
70469b43 | 285 | show_usage(argv); |
cb8818f0 MD |
286 | return -1; |
287 | } | |
288 | ||
289 | for (i = 4; i < argc; i++) { | |
290 | if (argv[i][0] != '-') | |
291 | continue; | |
292 | switch (argv[i][1]) { | |
293 | case 'a': | |
294 | if (argc < i + 2) { | |
70469b43 | 295 | show_usage(argv); |
cb8818f0 MD |
296 | return -1; |
297 | } | |
298 | a = atoi(argv[++i]); | |
299 | cpu_affinities[next_aff++] = a; | |
300 | use_affinity = 1; | |
301 | printf_verbose("Adding CPU %d affinity\n", a); | |
302 | break; | |
303 | case 'c': | |
304 | if (argc < i + 2) { | |
70469b43 | 305 | show_usage(argv); |
cb8818f0 MD |
306 | return -1; |
307 | } | |
308 | rduration = atol(argv[++i]); | |
309 | break; | |
310 | case 'd': | |
311 | if (argc < i + 2) { | |
70469b43 | 312 | show_usage(argv); |
cb8818f0 MD |
313 | return -1; |
314 | } | |
315 | wdelay = atol(argv[++i]); | |
316 | break; | |
317 | case 'v': | |
318 | verbose_mode = 1; | |
319 | break; | |
320 | } | |
321 | } | |
322 | ||
323 | printf_verbose("running test for %lu seconds, %u enqueuers, " | |
324 | "%u dequeuers.\n", | |
325 | duration, nr_enqueuers, nr_dequeuers); | |
326 | printf_verbose("Writer delay : %lu loops.\n", rduration); | |
327 | printf_verbose("Reader duration : %lu loops.\n", wdelay); | |
94df6318 MD |
328 | printf_verbose("thread %-6s, tid %lu\n", |
329 | "main", urcu_get_thread_id()); | |
cb8818f0 | 330 | |
9aa14175 MD |
331 | tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer)); |
332 | tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer)); | |
333 | count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer)); | |
334 | count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer)); | |
cb8818f0 MD |
335 | cds_lfs_init_rcu(&s); |
336 | err = create_all_cpu_call_rcu_data(0); | |
337 | if (err) { | |
338 | printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n"); | |
339 | } | |
340 | ||
341 | next_aff = 0; | |
342 | ||
83e334d0 MJ |
343 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
344 | err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer, | |
345 | &count_enqueuer[2 * i_thr]); | |
cb8818f0 MD |
346 | if (err != 0) |
347 | exit(1); | |
348 | } | |
83e334d0 MJ |
349 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
350 | err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer, | |
351 | &count_dequeuer[2 * i_thr]); | |
cb8818f0 MD |
352 | if (err != 0) |
353 | exit(1); | |
354 | } | |
355 | ||
356 | cmm_smp_mb(); | |
357 | ||
9e4e7ad1 | 358 | begin_test(); |
cb8818f0 | 359 | |
83e334d0 | 360 | for (i_thr = 0; i_thr < duration; i_thr++) { |
cb8818f0 | 361 | sleep(1); |
4b665350 MD |
362 | if (verbose_mode) { |
363 | fwrite(".", sizeof(char), 1, stdout); | |
364 | fflush(stdout); | |
365 | } | |
cb8818f0 MD |
366 | } |
367 | ||
9e4e7ad1 | 368 | end_test(); |
cb8818f0 | 369 | |
83e334d0 MJ |
370 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
371 | err = pthread_join(tid_enqueuer[i_thr], &tret); | |
cb8818f0 MD |
372 | if (err != 0) |
373 | exit(1); | |
83e334d0 MJ |
374 | tot_enqueues += count_enqueuer[2 * i_thr]; |
375 | tot_successful_enqueues += count_enqueuer[2 * i_thr + 1]; | |
cb8818f0 | 376 | } |
83e334d0 MJ |
377 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
378 | err = pthread_join(tid_dequeuer[i_thr], &tret); | |
cb8818f0 MD |
379 | if (err != 0) |
380 | exit(1); | |
83e334d0 MJ |
381 | tot_dequeues += count_dequeuer[2 * i_thr]; |
382 | tot_successful_dequeues += count_dequeuer[2 * i_thr + 1]; | |
cb8818f0 | 383 | } |
83e334d0 | 384 | |
bd23a6c6 | 385 | test_end(&end_dequeues); |
cb8818f0 MD |
386 | |
387 | printf_verbose("total number of enqueues : %llu, dequeues %llu\n", | |
388 | tot_enqueues, tot_dequeues); | |
389 | printf_verbose("total number of successful enqueues : %llu, " | |
390 | "successful dequeues %llu\n", | |
391 | tot_successful_enqueues, tot_successful_dequeues); | |
392 | printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu " | |
393 | "nr_dequeuers %3u " | |
394 | "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu " | |
395 | "successful enqueues %12llu successful dequeues %12llu " | |
396 | "end_dequeues %llu nr_ops %12llu\n", | |
397 | argv[0], duration, nr_enqueuers, wdelay, | |
398 | nr_dequeuers, rduration, tot_enqueues, tot_dequeues, | |
399 | tot_successful_enqueues, | |
400 | tot_successful_dequeues, end_dequeues, | |
401 | tot_enqueues + tot_dequeues); | |
402 | if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) | |
403 | printf("WARNING! Discrepancy between nr succ. enqueues %llu vs " | |
404 | "succ. dequeues + end dequeues %llu.\n", | |
405 | tot_successful_enqueues, | |
406 | tot_successful_dequeues + end_dequeues); | |
407 | ||
408 | free_all_cpu_call_rcu_data(); | |
409 | free(count_enqueuer); | |
410 | free(count_dequeuer); | |
411 | free(tid_enqueuer); | |
412 | free(tid_dequeuer); | |
413 | return 0; | |
414 | } |