]>
Commit | Line | Data |
---|---|---|
00b7ec5c ME |
1 | /* |
2 | * Context switch microbenchmark. | |
3 | * | |
4 | * Copyright (C) 2015 Anton Blanchard <anton@au.ibm.com>, IBM | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public License | |
8 | * as published by the Free Software Foundation; either version | |
9 | * 2 of the License, or (at your option) any later version. | |
10 | */ | |
11 | ||
12 | #define _GNU_SOURCE | |
13 | #include <sched.h> | |
14 | #include <string.h> | |
15 | #include <stdio.h> | |
16 | #include <unistd.h> | |
17 | #include <stdlib.h> | |
18 | #include <getopt.h> | |
19 | #include <signal.h> | |
20 | #include <assert.h> | |
21 | #include <pthread.h> | |
22 | #include <limits.h> | |
23 | #include <sys/time.h> | |
24 | #include <sys/syscall.h> | |
25 | #include <sys/types.h> | |
26 | #include <sys/shm.h> | |
27 | #include <linux/futex.h> | |
28 | ||
ea0c3217 ME |
29 | #include "../utils.h" |
30 | ||
31 | static unsigned int timeout = 30; | |
00b7ec5c ME |
32 | |
33 | static int touch_vdso; | |
34 | struct timeval tv; | |
35 | ||
36 | static int touch_fp; | |
37 | double fp; | |
38 | ||
39 | static int touch_vector; | |
40 | typedef int v4si __attribute__ ((vector_size (16))); | |
41 | v4si a, b, c; | |
42 | ||
43 | #ifdef __powerpc__ | |
44 | static int touch_altivec; | |
45 | ||
46 | static void __attribute__((__target__("no-vsx"))) altivec_touch_fn(void) | |
47 | { | |
48 | c = a + b; | |
49 | } | |
50 | #endif | |
51 | ||
52 | static void touch(void) | |
53 | { | |
54 | if (touch_vdso) | |
55 | gettimeofday(&tv, NULL); | |
56 | ||
57 | if (touch_fp) | |
58 | fp += 0.1; | |
59 | ||
60 | #ifdef __powerpc__ | |
61 | if (touch_altivec) | |
62 | altivec_touch_fn(); | |
63 | #endif | |
64 | ||
65 | if (touch_vector) | |
66 | c = a + b; | |
67 | ||
68 | asm volatile("# %0 %1 %2": : "r"(&tv), "r"(&fp), "r"(&c)); | |
69 | } | |
70 | ||
71 | static void start_thread_on(void *(*fn)(void *), void *arg, unsigned long cpu) | |
72 | { | |
73 | pthread_t tid; | |
74 | cpu_set_t cpuset; | |
75 | pthread_attr_t attr; | |
76 | ||
77 | CPU_ZERO(&cpuset); | |
78 | CPU_SET(cpu, &cpuset); | |
79 | ||
80 | pthread_attr_init(&attr); | |
81 | ||
82 | if (pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpuset)) { | |
83 | perror("pthread_attr_setaffinity_np"); | |
84 | exit(1); | |
85 | } | |
86 | ||
87 | if (pthread_create(&tid, &attr, fn, arg)) { | |
88 | perror("pthread_create"); | |
89 | exit(1); | |
90 | } | |
91 | } | |
92 | ||
93 | static void start_process_on(void *(*fn)(void *), void *arg, unsigned long cpu) | |
94 | { | |
95 | int pid; | |
96 | cpu_set_t cpuset; | |
97 | ||
98 | pid = fork(); | |
99 | if (pid == -1) { | |
100 | perror("fork"); | |
101 | exit(1); | |
102 | } | |
103 | ||
104 | if (pid) | |
105 | return; | |
106 | ||
107 | CPU_ZERO(&cpuset); | |
108 | CPU_SET(cpu, &cpuset); | |
109 | ||
110 | if (sched_setaffinity(0, sizeof(cpuset), &cpuset)) { | |
111 | perror("sched_setaffinity"); | |
112 | exit(1); | |
113 | } | |
114 | ||
115 | fn(arg); | |
116 | ||
117 | exit(0); | |
118 | } | |
119 | ||
120 | static unsigned long iterations; | |
121 | static unsigned long iterations_prev; | |
122 | ||
123 | static void sigalrm_handler(int junk) | |
124 | { | |
125 | unsigned long i = iterations; | |
126 | ||
127 | printf("%ld\n", i - iterations_prev); | |
128 | iterations_prev = i; | |
129 | ||
130 | if (--timeout == 0) | |
131 | kill(0, SIGUSR1); | |
132 | ||
133 | alarm(1); | |
134 | } | |
135 | ||
136 | static void sigusr1_handler(int junk) | |
137 | { | |
138 | exit(0); | |
139 | } | |
140 | ||
141 | struct actions { | |
142 | void (*setup)(int, int); | |
143 | void *(*thread1)(void *); | |
144 | void *(*thread2)(void *); | |
145 | }; | |
146 | ||
147 | #define READ 0 | |
148 | #define WRITE 1 | |
149 | ||
150 | static int pipe_fd1[2]; | |
151 | static int pipe_fd2[2]; | |
152 | ||
153 | static void pipe_setup(int cpu1, int cpu2) | |
154 | { | |
155 | if (pipe(pipe_fd1) || pipe(pipe_fd2)) | |
156 | exit(1); | |
157 | } | |
158 | ||
159 | static void *pipe_thread1(void *arg) | |
160 | { | |
161 | signal(SIGALRM, sigalrm_handler); | |
162 | alarm(1); | |
163 | ||
164 | while (1) { | |
165 | assert(read(pipe_fd1[READ], &c, 1) == 1); | |
166 | touch(); | |
167 | ||
168 | assert(write(pipe_fd2[WRITE], &c, 1) == 1); | |
169 | touch(); | |
170 | ||
171 | iterations += 2; | |
172 | } | |
173 | ||
174 | return NULL; | |
175 | } | |
176 | ||
177 | static void *pipe_thread2(void *arg) | |
178 | { | |
179 | while (1) { | |
180 | assert(write(pipe_fd1[WRITE], &c, 1) == 1); | |
181 | touch(); | |
182 | ||
183 | assert(read(pipe_fd2[READ], &c, 1) == 1); | |
184 | touch(); | |
185 | } | |
186 | ||
187 | return NULL; | |
188 | } | |
189 | ||
190 | static struct actions pipe_actions = { | |
191 | .setup = pipe_setup, | |
192 | .thread1 = pipe_thread1, | |
193 | .thread2 = pipe_thread2, | |
194 | }; | |
195 | ||
196 | static void yield_setup(int cpu1, int cpu2) | |
197 | { | |
198 | if (cpu1 != cpu2) { | |
199 | fprintf(stderr, "Both threads must be on the same CPU for yield test\n"); | |
200 | exit(1); | |
201 | } | |
202 | } | |
203 | ||
204 | static void *yield_thread1(void *arg) | |
205 | { | |
206 | signal(SIGALRM, sigalrm_handler); | |
207 | alarm(1); | |
208 | ||
209 | while (1) { | |
210 | sched_yield(); | |
211 | touch(); | |
212 | ||
213 | iterations += 2; | |
214 | } | |
215 | ||
216 | return NULL; | |
217 | } | |
218 | ||
219 | static void *yield_thread2(void *arg) | |
220 | { | |
221 | while (1) { | |
222 | sched_yield(); | |
223 | touch(); | |
224 | } | |
225 | ||
226 | return NULL; | |
227 | } | |
228 | ||
229 | static struct actions yield_actions = { | |
230 | .setup = yield_setup, | |
231 | .thread1 = yield_thread1, | |
232 | .thread2 = yield_thread2, | |
233 | }; | |
234 | ||
235 | static long sys_futex(void *addr1, int op, int val1, struct timespec *timeout, | |
236 | void *addr2, int val3) | |
237 | { | |
238 | return syscall(SYS_futex, addr1, op, val1, timeout, addr2, val3); | |
239 | } | |
240 | ||
241 | static unsigned long cmpxchg(unsigned long *p, unsigned long expected, | |
242 | unsigned long desired) | |
243 | { | |
244 | unsigned long exp = expected; | |
245 | ||
246 | __atomic_compare_exchange_n(p, &exp, desired, 0, | |
247 | __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST); | |
248 | return exp; | |
249 | } | |
250 | ||
251 | static unsigned long xchg(unsigned long *p, unsigned long val) | |
252 | { | |
253 | return __atomic_exchange_n(p, val, __ATOMIC_SEQ_CST); | |
254 | } | |
255 | ||
256 | static int mutex_lock(unsigned long *m) | |
257 | { | |
258 | int c; | |
259 | ||
260 | c = cmpxchg(m, 0, 1); | |
261 | if (!c) | |
262 | return 0; | |
263 | ||
264 | if (c == 1) | |
265 | c = xchg(m, 2); | |
266 | ||
267 | while (c) { | |
268 | sys_futex(m, FUTEX_WAIT, 2, NULL, NULL, 0); | |
269 | c = xchg(m, 2); | |
270 | } | |
271 | ||
272 | return 0; | |
273 | } | |
274 | ||
275 | static int mutex_unlock(unsigned long *m) | |
276 | { | |
277 | if (*m == 2) | |
278 | *m = 0; | |
279 | else if (xchg(m, 0) == 1) | |
280 | return 0; | |
281 | ||
282 | sys_futex(m, FUTEX_WAKE, 1, NULL, NULL, 0); | |
283 | ||
284 | return 0; | |
285 | } | |
286 | ||
287 | static unsigned long *m1, *m2; | |
288 | ||
289 | static void futex_setup(int cpu1, int cpu2) | |
290 | { | |
291 | int shmid; | |
292 | void *shmaddr; | |
293 | ||
294 | shmid = shmget(IPC_PRIVATE, getpagesize(), SHM_R | SHM_W); | |
295 | if (shmid < 0) { | |
296 | perror("shmget"); | |
297 | exit(1); | |
298 | } | |
299 | ||
300 | shmaddr = shmat(shmid, NULL, 0); | |
301 | if (shmaddr == (char *)-1) { | |
302 | perror("shmat"); | |
303 | shmctl(shmid, IPC_RMID, NULL); | |
304 | exit(1); | |
305 | } | |
306 | ||
307 | shmctl(shmid, IPC_RMID, NULL); | |
308 | ||
309 | m1 = shmaddr; | |
310 | m2 = shmaddr + sizeof(*m1); | |
311 | ||
312 | *m1 = 0; | |
313 | *m2 = 0; | |
314 | ||
315 | mutex_lock(m1); | |
316 | mutex_lock(m2); | |
317 | } | |
318 | ||
319 | static void *futex_thread1(void *arg) | |
320 | { | |
321 | signal(SIGALRM, sigalrm_handler); | |
322 | alarm(1); | |
323 | ||
324 | while (1) { | |
325 | mutex_lock(m2); | |
326 | mutex_unlock(m1); | |
327 | ||
328 | iterations += 2; | |
329 | } | |
330 | ||
331 | return NULL; | |
332 | } | |
333 | ||
334 | static void *futex_thread2(void *arg) | |
335 | { | |
336 | while (1) { | |
337 | mutex_unlock(m2); | |
338 | mutex_lock(m1); | |
339 | } | |
340 | ||
341 | return NULL; | |
342 | } | |
343 | ||
344 | static struct actions futex_actions = { | |
345 | .setup = futex_setup, | |
346 | .thread1 = futex_thread1, | |
347 | .thread2 = futex_thread2, | |
348 | }; | |
349 | ||
350 | static int processes; | |
351 | ||
352 | static struct option options[] = { | |
353 | { "test", required_argument, 0, 't' }, | |
354 | { "process", no_argument, &processes, 1 }, | |
355 | { "timeout", required_argument, 0, 's' }, | |
356 | { "vdso", no_argument, &touch_vdso, 1 }, | |
357 | { "fp", no_argument, &touch_fp, 1 }, | |
358 | #ifdef __powerpc__ | |
359 | { "altivec", no_argument, &touch_altivec, 1 }, | |
360 | #endif | |
361 | { "vector", no_argument, &touch_vector, 1 }, | |
362 | { 0, }, | |
363 | }; | |
364 | ||
365 | static void usage(void) | |
366 | { | |
367 | fprintf(stderr, "Usage: context_switch2 <options> CPU1 CPU2\n\n"); | |
ea0c3217 | 368 | fprintf(stderr, "\t\t--test=X\tpipe, futex or yield (default)\n"); |
00b7ec5c | 369 | fprintf(stderr, "\t\t--process\tUse processes (default threads)\n"); |
ea0c3217 | 370 | fprintf(stderr, "\t\t--timeout=X\tDuration in seconds to run (default 30)\n"); |
00b7ec5c ME |
371 | fprintf(stderr, "\t\t--vdso\t\ttouch VDSO\n"); |
372 | fprintf(stderr, "\t\t--fp\t\ttouch FP\n"); | |
373 | #ifdef __powerpc__ | |
374 | fprintf(stderr, "\t\t--altivec\ttouch altivec\n"); | |
375 | #endif | |
376 | fprintf(stderr, "\t\t--vector\ttouch vector\n"); | |
377 | } | |
378 | ||
379 | int main(int argc, char *argv[]) | |
380 | { | |
381 | signed char c; | |
ea0c3217 | 382 | struct actions *actions = &yield_actions; |
00b7ec5c ME |
383 | int cpu1; |
384 | int cpu2; | |
385 | static void (*start_fn)(void *(*fn)(void *), void *arg, unsigned long cpu); | |
386 | ||
387 | while (1) { | |
388 | int option_index = 0; | |
389 | ||
390 | c = getopt_long(argc, argv, "", options, &option_index); | |
391 | ||
392 | if (c == -1) | |
393 | break; | |
394 | ||
395 | switch (c) { | |
396 | case 0: | |
397 | if (options[option_index].flag != 0) | |
398 | break; | |
399 | ||
400 | usage(); | |
401 | exit(1); | |
402 | break; | |
403 | ||
404 | case 't': | |
405 | if (!strcmp(optarg, "pipe")) { | |
406 | actions = &pipe_actions; | |
407 | } else if (!strcmp(optarg, "yield")) { | |
408 | actions = &yield_actions; | |
409 | } else if (!strcmp(optarg, "futex")) { | |
410 | actions = &futex_actions; | |
411 | } else { | |
412 | usage(); | |
413 | exit(1); | |
414 | } | |
415 | break; | |
416 | ||
417 | case 's': | |
418 | timeout = atoi(optarg); | |
419 | break; | |
420 | ||
421 | default: | |
422 | usage(); | |
423 | exit(1); | |
424 | } | |
425 | } | |
426 | ||
427 | if (processes) | |
428 | start_fn = start_process_on; | |
429 | else | |
430 | start_fn = start_thread_on; | |
431 | ||
432 | if (((argc - optind) != 2)) { | |
ea0c3217 ME |
433 | cpu1 = cpu2 = pick_online_cpu(); |
434 | } else { | |
435 | cpu1 = atoi(argv[optind++]); | |
436 | cpu2 = atoi(argv[optind++]); | |
00b7ec5c ME |
437 | } |
438 | ||
ea0c3217 ME |
439 | printf("Using %s with ", processes ? "processes" : "threads"); |
440 | ||
441 | if (actions == &pipe_actions) | |
442 | printf("pipe"); | |
443 | else if (actions == &yield_actions) | |
444 | printf("yield"); | |
445 | else | |
446 | printf("futex"); | |
447 | ||
448 | printf(" on cpus %d/%d touching FP:%s altivec:%s vector:%s vdso:%s\n", | |
449 | cpu1, cpu2, touch_fp ? "yes" : "no", touch_altivec ? "yes" : "no", | |
450 | touch_vector ? "yes" : "no", touch_vdso ? "yes" : "no"); | |
451 | ||
00b7ec5c ME |
452 | /* Create a new process group so we can signal everyone for exit */ |
453 | setpgid(getpid(), getpid()); | |
454 | ||
455 | signal(SIGUSR1, sigusr1_handler); | |
456 | ||
00b7ec5c ME |
457 | actions->setup(cpu1, cpu2); |
458 | ||
459 | start_fn(actions->thread1, NULL, cpu1); | |
460 | start_fn(actions->thread2, NULL, cpu2); | |
461 | ||
462 | while (1) | |
463 | sleep(3600); | |
464 | ||
465 | return 0; | |
466 | } |