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 | ||
ee566d46 | 6 | /* |
ee566d46 | 7 | * Userspace RCU library - example RCU-based lock-free queue |
ee566d46 MD |
8 | */ |
9 | ||
ee566d46 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> | |
ee566d46 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 | |
ee566d46 MD |
27 | /* hardcoded number of CPUs */ |
28 | #define NR_CPUS 16384 | |
29 | ||
ee566d46 MD |
30 | #ifndef DYNAMIC_LINK_TEST |
31 | #define _LGPL_SOURCE | |
32 | #endif | |
14748510 MD |
33 | |
34 | /* Remove deprecation warnings from test build. */ | |
35 | #define CDS_WFQ_DEPRECATED | |
36 | ||
ee566d46 | 37 | #include <urcu.h> |
4afee0a7 | 38 | #include <urcu/wfqueue.h> |
ee566d46 | 39 | |
ee566d46 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 | ||
ab0aacbe | 47 | static inline void loop_sleep(unsigned long loops) |
ee566d46 | 48 | { |
ab0aacbe | 49 | while (loops-- != 0) |
06f22bdb | 50 | caa_cpu_relax(); |
ee566d46 MD |
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 | ||
ee566d46 MD |
67 | static void set_affinity(void) |
68 | { | |
2388c075 | 69 | #ifdef HAVE_SCHED_SETAFFINITY |
ee566d46 | 70 | cpu_set_t mask; |
95bc7fb9 MD |
71 | int cpu, ret; |
72 | #endif /* HAVE_SCHED_SETAFFINITY */ | |
ee566d46 MD |
73 | |
74 | if (!use_affinity) | |
75 | return; | |
76 | ||
2388c075 | 77 | #ifdef HAVE_SCHED_SETAFFINITY |
ee566d46 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); | |
ee566d46 | 92 | sched_setaffinity(0, sizeof(mask), &mask); |
ee566d46 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(); |
ee566d46 MD |
102 | } |
103 | ||
104 | static int test_duration_enqueue(void) | |
105 | { | |
9e4e7ad1 | 106 | return test_duration_write(); |
ee566d46 MD |
107 | } |
108 | ||
bd252a04 MD |
109 | static DEFINE_URCU_TLS(unsigned long long, nr_dequeues); |
110 | static DEFINE_URCU_TLS(unsigned long long, nr_enqueues); | |
ee566d46 | 111 | |
bd252a04 MD |
112 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues); |
113 | static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues); | |
ee566d46 MD |
114 | |
115 | static unsigned int nr_enqueuers; | |
116 | static unsigned int nr_dequeuers; | |
117 | ||
16aa9ee8 | 118 | static struct cds_wfq_queue q; |
ee566d46 | 119 | |
61c3fb60 | 120 | static |
ee566d46 MD |
121 | void *thr_enqueuer(void *_count) |
122 | { | |
123 | unsigned long long *count = _count; | |
124 | ||
94df6318 MD |
125 | printf_verbose("thread_begin %s, tid %lu\n", |
126 | "enqueuer", urcu_get_thread_id()); | |
ee566d46 MD |
127 | |
128 | set_affinity(); | |
129 | ||
9e4e7ad1 | 130 | wait_until_go(); |
ee566d46 MD |
131 | |
132 | for (;;) { | |
16aa9ee8 | 133 | struct cds_wfq_node *node = malloc(sizeof(*node)); |
ee566d46 MD |
134 | if (!node) |
135 | goto fail; | |
16aa9ee8 DG |
136 | cds_wfq_node_init(node); |
137 | cds_wfq_enqueue(&q, node); | |
bd252a04 | 138 | URCU_TLS(nr_successful_enqueues)++; |
ee566d46 | 139 | |
a0b7f7ea | 140 | if (caa_unlikely(wdelay)) |
ee566d46 MD |
141 | loop_sleep(wdelay); |
142 | fail: | |
bd252a04 | 143 | URCU_TLS(nr_enqueues)++; |
a0b7f7ea | 144 | if (caa_unlikely(!test_duration_enqueue())) |
ee566d46 MD |
145 | break; |
146 | } | |
147 | ||
bd252a04 MD |
148 | count[0] = URCU_TLS(nr_enqueues); |
149 | count[1] = URCU_TLS(nr_successful_enqueues); | |
94df6318 MD |
150 | printf_verbose("enqueuer thread_end, tid %lu, " |
151 | "enqueues %llu successful_enqueues %llu\n", | |
152 | urcu_get_thread_id(), | |
153 | URCU_TLS(nr_enqueues), | |
154 | URCU_TLS(nr_successful_enqueues)); | |
ee566d46 MD |
155 | return ((void*)1); |
156 | ||
157 | } | |
158 | ||
61c3fb60 | 159 | static |
ee566d46 MD |
160 | void *thr_dequeuer(void *_count) |
161 | { | |
162 | unsigned long long *count = _count; | |
163 | ||
94df6318 MD |
164 | printf_verbose("thread_begin %s, tid %lu\n", |
165 | "dequeuer", urcu_get_thread_id()); | |
ee566d46 MD |
166 | |
167 | set_affinity(); | |
168 | ||
9e4e7ad1 | 169 | wait_until_go(); |
ee566d46 MD |
170 | |
171 | for (;;) { | |
16aa9ee8 | 172 | struct cds_wfq_node *node = cds_wfq_dequeue_blocking(&q); |
ee566d46 MD |
173 | |
174 | if (node) { | |
4afee0a7 | 175 | free(node); |
bd252a04 | 176 | URCU_TLS(nr_successful_dequeues)++; |
ee566d46 MD |
177 | } |
178 | ||
bd252a04 | 179 | URCU_TLS(nr_dequeues)++; |
a0b7f7ea | 180 | if (caa_unlikely(!test_duration_dequeue())) |
ee566d46 | 181 | break; |
a0b7f7ea | 182 | if (caa_unlikely(rduration)) |
ee566d46 MD |
183 | loop_sleep(rduration); |
184 | } | |
185 | ||
94df6318 MD |
186 | printf_verbose("dequeuer thread_end, tid %lu, " |
187 | "dequeues %llu, successful_dequeues %llu\n", | |
188 | urcu_get_thread_id(), | |
189 | URCU_TLS(nr_dequeues), | |
190 | URCU_TLS(nr_successful_dequeues)); | |
bd252a04 MD |
191 | count[0] = URCU_TLS(nr_dequeues); |
192 | count[1] = URCU_TLS(nr_successful_dequeues); | |
ee566d46 MD |
193 | return ((void*)2); |
194 | } | |
195 | ||
61c3fb60 | 196 | static |
bd23a6c6 | 197 | void test_end(unsigned long long *nr_dequeues_l) |
ee566d46 | 198 | { |
16aa9ee8 | 199 | struct cds_wfq_node *node; |
ee566d46 MD |
200 | |
201 | do { | |
bd23a6c6 | 202 | node = cds_wfq_dequeue_blocking(&q); |
ee566d46 | 203 | if (node) { |
4afee0a7 | 204 | free(node); |
bd23a6c6 | 205 | (*nr_dequeues_l)++; |
ee566d46 MD |
206 | } |
207 | } while (node); | |
208 | } | |
209 | ||
61c3fb60 | 210 | static |
70469b43 | 211 | void show_usage(char **argv) |
ee566d46 | 212 | { |
06637138 MD |
213 | printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n", |
214 | argv[0]); | |
215 | printf("OPTIONS:\n"); | |
216 | printf(" [-d delay] (enqueuer period (in loops))\n"); | |
217 | printf(" [-c duration] (dequeuer period (in loops))\n"); | |
218 | printf(" [-v] (verbose output)\n"); | |
219 | printf(" [-a cpu#] [-a cpu#]... (affinity)\n"); | |
ee566d46 MD |
220 | printf("\n"); |
221 | } | |
222 | ||
223 | int main(int argc, char **argv) | |
224 | { | |
225 | int err; | |
226 | pthread_t *tid_enqueuer, *tid_dequeuer; | |
227 | void *tret; | |
228 | unsigned long long *count_enqueuer, *count_dequeuer; | |
229 | unsigned long long tot_enqueues = 0, tot_dequeues = 0; | |
230 | unsigned long long tot_successful_enqueues = 0, | |
231 | tot_successful_dequeues = 0; | |
232 | unsigned long long end_dequeues = 0; | |
233 | int i, a; | |
83e334d0 | 234 | unsigned int i_thr; |
ee566d46 MD |
235 | |
236 | if (argc < 4) { | |
70469b43 | 237 | show_usage(argv); |
ee566d46 MD |
238 | return -1; |
239 | } | |
240 | ||
241 | err = sscanf(argv[1], "%u", &nr_dequeuers); | |
242 | if (err != 1) { | |
70469b43 | 243 | show_usage(argv); |
ee566d46 MD |
244 | return -1; |
245 | } | |
246 | ||
247 | err = sscanf(argv[2], "%u", &nr_enqueuers); | |
248 | if (err != 1) { | |
70469b43 | 249 | show_usage(argv); |
ee566d46 MD |
250 | return -1; |
251 | } | |
83e334d0 | 252 | |
ee566d46 MD |
253 | err = sscanf(argv[3], "%lu", &duration); |
254 | if (err != 1) { | |
70469b43 | 255 | show_usage(argv); |
ee566d46 MD |
256 | return -1; |
257 | } | |
258 | ||
259 | for (i = 4; i < argc; i++) { | |
260 | if (argv[i][0] != '-') | |
261 | continue; | |
262 | switch (argv[i][1]) { | |
263 | case 'a': | |
264 | if (argc < i + 2) { | |
70469b43 | 265 | show_usage(argv); |
ee566d46 MD |
266 | return -1; |
267 | } | |
268 | a = atoi(argv[++i]); | |
269 | cpu_affinities[next_aff++] = a; | |
270 | use_affinity = 1; | |
271 | printf_verbose("Adding CPU %d affinity\n", a); | |
272 | break; | |
273 | case 'c': | |
274 | if (argc < i + 2) { | |
70469b43 | 275 | show_usage(argv); |
ee566d46 MD |
276 | return -1; |
277 | } | |
278 | rduration = atol(argv[++i]); | |
279 | break; | |
280 | case 'd': | |
281 | if (argc < i + 2) { | |
70469b43 | 282 | show_usage(argv); |
ee566d46 MD |
283 | return -1; |
284 | } | |
285 | wdelay = atol(argv[++i]); | |
286 | break; | |
287 | case 'v': | |
288 | verbose_mode = 1; | |
289 | break; | |
290 | } | |
291 | } | |
292 | ||
293 | printf_verbose("running test for %lu seconds, %u enqueuers, " | |
294 | "%u dequeuers.\n", | |
295 | duration, nr_enqueuers, nr_dequeuers); | |
296 | printf_verbose("Writer delay : %lu loops.\n", rduration); | |
297 | printf_verbose("Reader duration : %lu loops.\n", wdelay); | |
94df6318 MD |
298 | printf_verbose("thread %-6s, tid %lu\n", |
299 | "main", urcu_get_thread_id()); | |
ee566d46 | 300 | |
9aa14175 MD |
301 | tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer)); |
302 | tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer)); | |
303 | count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer)); | |
304 | count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer)); | |
16aa9ee8 | 305 | cds_wfq_init(&q); |
ee566d46 MD |
306 | |
307 | next_aff = 0; | |
308 | ||
83e334d0 MJ |
309 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
310 | err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer, | |
311 | &count_enqueuer[2 * i_thr]); | |
ee566d46 MD |
312 | if (err != 0) |
313 | exit(1); | |
314 | } | |
83e334d0 MJ |
315 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
316 | err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer, | |
317 | &count_dequeuer[2 * i_thr]); | |
ee566d46 MD |
318 | if (err != 0) |
319 | exit(1); | |
320 | } | |
321 | ||
5481ddb3 | 322 | cmm_smp_mb(); |
ee566d46 | 323 | |
9e4e7ad1 | 324 | begin_test(); |
ee566d46 | 325 | |
83e334d0 | 326 | for (i_thr = 0; i_thr < duration; i_thr++) { |
ee566d46 | 327 | sleep(1); |
4b665350 MD |
328 | if (verbose_mode) { |
329 | fwrite(".", sizeof(char), 1, stdout); | |
330 | fflush(stdout); | |
331 | } | |
ee566d46 MD |
332 | } |
333 | ||
9e4e7ad1 | 334 | end_test(); |
ee566d46 | 335 | |
83e334d0 MJ |
336 | for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) { |
337 | err = pthread_join(tid_enqueuer[i_thr], &tret); | |
ee566d46 MD |
338 | if (err != 0) |
339 | exit(1); | |
83e334d0 MJ |
340 | tot_enqueues += count_enqueuer[2 * i_thr]; |
341 | tot_successful_enqueues += count_enqueuer[2 * i_thr + 1]; | |
ee566d46 | 342 | } |
83e334d0 MJ |
343 | for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) { |
344 | err = pthread_join(tid_dequeuer[i_thr], &tret); | |
ee566d46 MD |
345 | if (err != 0) |
346 | exit(1); | |
83e334d0 MJ |
347 | tot_dequeues += count_dequeuer[2 * i_thr]; |
348 | tot_successful_dequeues += count_dequeuer[2 * i_thr + 1]; | |
ee566d46 | 349 | } |
83e334d0 | 350 | |
bd23a6c6 | 351 | test_end(&end_dequeues); |
ee566d46 MD |
352 | |
353 | printf_verbose("total number of enqueues : %llu, dequeues %llu\n", | |
354 | tot_enqueues, tot_dequeues); | |
355 | printf_verbose("total number of successful enqueues : %llu, " | |
356 | "successful dequeues %llu\n", | |
357 | tot_successful_enqueues, tot_successful_dequeues); | |
358 | printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu " | |
359 | "nr_dequeuers %3u " | |
360 | "rdur %6lu nr_enqueues %12llu nr_dequeues %12llu " | |
361 | "successful enqueues %12llu successful dequeues %12llu " | |
362 | "end_dequeues %llu nr_ops %12llu\n", | |
363 | argv[0], duration, nr_enqueuers, wdelay, | |
364 | nr_dequeuers, rduration, tot_enqueues, tot_dequeues, | |
365 | tot_successful_enqueues, | |
366 | tot_successful_dequeues, end_dequeues, | |
367 | tot_enqueues + tot_dequeues); | |
368 | if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) | |
369 | printf("WARNING! Discrepancy between nr succ. enqueues %llu vs " | |
370 | "succ. dequeues + end dequeues %llu.\n", | |
371 | tot_successful_enqueues, | |
372 | tot_successful_dequeues + end_dequeues); | |
373 | ||
2af1c19e | 374 | cds_wfq_destroy(&q); |
ee566d46 MD |
375 | free(count_enqueuer); |
376 | free(count_dequeuer); | |
377 | free(tid_enqueuer); | |
378 | free(tid_dequeuer); | |
83e334d0 | 379 | |
ee566d46 MD |
380 | return 0; |
381 | } |