Loading...
1/*
2 * This application is Copyright 2012 Red Hat, Inc.
3 * Doug Ledford <dledford@redhat.com>
4 *
5 * mq_perf_tests is free software: you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, version 3.
8 *
9 * mq_perf_tests is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * For the full text of the license, see <http://www.gnu.org/licenses/>.
15 *
16 * mq_perf_tests.c
17 * Tests various types of message queue workloads, concentrating on those
18 * situations that invole large message sizes, large message queue depths,
19 * or both, and reports back useful metrics about kernel message queue
20 * performance.
21 *
22 */
23#define _GNU_SOURCE
24#include <stdio.h>
25#include <stdlib.h>
26#include <unistd.h>
27#include <fcntl.h>
28#include <string.h>
29#include <limits.h>
30#include <errno.h>
31#include <signal.h>
32#include <pthread.h>
33#include <sched.h>
34#include <sys/types.h>
35#include <sys/time.h>
36#include <sys/resource.h>
37#include <sys/stat.h>
38#include <mqueue.h>
39#include <popt.h>
40#include <error.h>
41
42static char *usage =
43"Usage:\n"
44" %s [-c #[,#..] -f] path\n"
45"\n"
46" -c # Skip most tests and go straight to a high queue depth test\n"
47" and then run that test continuously (useful for running at\n"
48" the same time as some other workload to see how much the\n"
49" cache thrashing caused by adding messages to a very deep\n"
50" queue impacts the performance of other programs). The number\n"
51" indicates which CPU core we should bind the process to during\n"
52" the run. If you have more than one physical CPU, then you\n"
53" will need one copy per physical CPU package, and you should\n"
54" specify the CPU cores to pin ourself to via a comma separated\n"
55" list of CPU values.\n"
56" -f Only usable with continuous mode. Pin ourself to the CPUs\n"
57" as requested, then instead of looping doing a high mq\n"
58" workload, just busy loop. This will allow us to lock up a\n"
59" single CPU just like we normally would, but without actually\n"
60" thrashing the CPU cache. This is to make it easier to get\n"
61" comparable numbers from some other workload running on the\n"
62" other CPUs. One set of numbers with # CPUs locked up running\n"
63" an mq workload, and another set of numbers with those same\n"
64" CPUs locked away from the test workload, but not doing\n"
65" anything to trash the cache like the mq workload might.\n"
66" path Path name of the message queue to create\n"
67"\n"
68" Note: this program must be run as root in order to enable all tests\n"
69"\n";
70
71char *MAX_MSGS = "/proc/sys/fs/mqueue/msg_max";
72char *MAX_MSGSIZE = "/proc/sys/fs/mqueue/msgsize_max";
73
74#define min(a, b) ((a) < (b) ? (a) : (b))
75#define MAX_CPUS 64
76char *cpu_option_string;
77int cpus_to_pin[MAX_CPUS];
78int num_cpus_to_pin;
79pthread_t cpu_threads[MAX_CPUS];
80pthread_t main_thread;
81cpu_set_t *cpu_set;
82int cpu_set_size;
83int cpus_online;
84
85#define MSG_SIZE 16
86#define TEST1_LOOPS 10000000
87#define TEST2_LOOPS 100000
88int continuous_mode;
89int continuous_mode_fake;
90
91struct rlimit saved_limits, cur_limits;
92int saved_max_msgs, saved_max_msgsize;
93int cur_max_msgs, cur_max_msgsize;
94FILE *max_msgs, *max_msgsize;
95int cur_nice;
96char *queue_path = "/mq_perf_tests";
97mqd_t queue = -1;
98struct mq_attr result;
99int mq_prio_max;
100
101const struct poptOption options[] = {
102 {
103 .longName = "continuous",
104 .shortName = 'c',
105 .argInfo = POPT_ARG_STRING,
106 .arg = &cpu_option_string,
107 .val = 'c',
108 .descrip = "Run continuous tests at a high queue depth in "
109 "order to test the effects of cache thrashing on "
110 "other tasks on the system. This test is intended "
111 "to be run on one core of each physical CPU while "
112 "some other CPU intensive task is run on all the other "
113 "cores of that same physical CPU and the other task "
114 "is timed. It is assumed that the process of adding "
115 "messages to the message queue in a tight loop will "
116 "impact that other task to some degree. Once the "
117 "tests are performed in this way, you should then "
118 "re-run the tests using fake mode in order to check "
119 "the difference in time required to perform the CPU "
120 "intensive task",
121 .argDescrip = "cpu[,cpu]",
122 },
123 {
124 .longName = "fake",
125 .shortName = 'f',
126 .argInfo = POPT_ARG_NONE,
127 .arg = &continuous_mode_fake,
128 .val = 0,
129 .descrip = "Tie up the CPUs that we would normally tie up in"
130 "continuous mode, but don't actually do any mq stuff, "
131 "just keep the CPU busy so it can't be used to process "
132 "system level tasks as this would free up resources on "
133 "the other CPU cores and skew the comparison between "
134 "the no-mqueue work and mqueue work tests",
135 .argDescrip = NULL,
136 },
137 {
138 .longName = "path",
139 .shortName = 'p',
140 .argInfo = POPT_ARG_STRING | POPT_ARGFLAG_SHOW_DEFAULT,
141 .arg = &queue_path,
142 .val = 'p',
143 .descrip = "The name of the path to use in the mqueue "
144 "filesystem for our tests",
145 .argDescrip = "pathname",
146 },
147 POPT_AUTOHELP
148 POPT_TABLEEND
149};
150
151static inline void __set(FILE *stream, int value, char *err_msg);
152void shutdown(int exit_val, char *err_cause, int line_no);
153void sig_action_SIGUSR1(int signum, siginfo_t *info, void *context);
154void sig_action(int signum, siginfo_t *info, void *context);
155static inline int get(FILE *stream);
156static inline void set(FILE *stream, int value);
157static inline int try_set(FILE *stream, int value);
158static inline void getr(int type, struct rlimit *rlim);
159static inline void setr(int type, struct rlimit *rlim);
160static inline void open_queue(struct mq_attr *attr);
161void increase_limits(void);
162
163static inline void __set(FILE *stream, int value, char *err_msg)
164{
165 rewind(stream);
166 if (fprintf(stream, "%d", value) < 0)
167 perror(err_msg);
168}
169
170
171void shutdown(int exit_val, char *err_cause, int line_no)
172{
173 static int in_shutdown = 0;
174 int errno_at_shutdown = errno;
175 int i;
176
177 /* In case we get called by multiple threads or from an sighandler */
178 if (in_shutdown++)
179 return;
180
181 for (i = 0; i < num_cpus_to_pin; i++)
182 if (cpu_threads[i]) {
183 pthread_kill(cpu_threads[i], SIGUSR1);
184 pthread_join(cpu_threads[i], NULL);
185 }
186
187 if (queue != -1)
188 if (mq_close(queue))
189 perror("mq_close() during shutdown");
190 if (queue_path)
191 /*
192 * Be silent if this fails, if we cleaned up already it's
193 * expected to fail
194 */
195 mq_unlink(queue_path);
196 if (saved_max_msgs)
197 __set(max_msgs, saved_max_msgs,
198 "failed to restore saved_max_msgs");
199 if (saved_max_msgsize)
200 __set(max_msgsize, saved_max_msgsize,
201 "failed to restore saved_max_msgsize");
202 if (exit_val)
203 error(exit_val, errno_at_shutdown, "%s at %d",
204 err_cause, line_no);
205 exit(0);
206}
207
208void sig_action_SIGUSR1(int signum, siginfo_t *info, void *context)
209{
210 if (pthread_self() != main_thread)
211 pthread_exit(0);
212 else {
213 fprintf(stderr, "Caught signal %d in SIGUSR1 handler, "
214 "exiting\n", signum);
215 shutdown(0, "", 0);
216 fprintf(stderr, "\n\nReturned from shutdown?!?!\n\n");
217 exit(0);
218 }
219}
220
221void sig_action(int signum, siginfo_t *info, void *context)
222{
223 if (pthread_self() != main_thread)
224 pthread_kill(main_thread, signum);
225 else {
226 fprintf(stderr, "Caught signal %d, exiting\n", signum);
227 shutdown(0, "", 0);
228 fprintf(stderr, "\n\nReturned from shutdown?!?!\n\n");
229 exit(0);
230 }
231}
232
233static inline int get(FILE *stream)
234{
235 int value;
236 rewind(stream);
237 if (fscanf(stream, "%d", &value) != 1)
238 shutdown(4, "Error reading /proc entry", __LINE__);
239 return value;
240}
241
242static inline void set(FILE *stream, int value)
243{
244 int new_value;
245
246 rewind(stream);
247 if (fprintf(stream, "%d", value) < 0)
248 return shutdown(5, "Failed writing to /proc file", __LINE__);
249 new_value = get(stream);
250 if (new_value != value)
251 return shutdown(5, "We didn't get what we wrote to /proc back",
252 __LINE__);
253}
254
255static inline int try_set(FILE *stream, int value)
256{
257 int new_value;
258
259 rewind(stream);
260 fprintf(stream, "%d", value);
261 new_value = get(stream);
262 return new_value == value;
263}
264
265static inline void getr(int type, struct rlimit *rlim)
266{
267 if (getrlimit(type, rlim))
268 shutdown(6, "getrlimit()", __LINE__);
269}
270
271static inline void setr(int type, struct rlimit *rlim)
272{
273 if (setrlimit(type, rlim))
274 shutdown(7, "setrlimit()", __LINE__);
275}
276
277/**
278 * open_queue - open the global queue for testing
279 * @attr - An attr struct specifying the desired queue traits
280 * @result - An attr struct that lists the actual traits the queue has
281 *
282 * This open is not allowed to fail, failure will result in an orderly
283 * shutdown of the program. The global queue_path is used to set what
284 * queue to open, the queue descriptor is saved in the global queue
285 * variable.
286 */
287static inline void open_queue(struct mq_attr *attr)
288{
289 int flags = O_RDWR | O_EXCL | O_CREAT | O_NONBLOCK;
290 int perms = DEFFILEMODE;
291
292 queue = mq_open(queue_path, flags, perms, attr);
293 if (queue == -1)
294 shutdown(1, "mq_open()", __LINE__);
295 if (mq_getattr(queue, &result))
296 shutdown(1, "mq_getattr()", __LINE__);
297 printf("\n\tQueue %s created:\n", queue_path);
298 printf("\t\tmq_flags:\t\t\t%s\n", result.mq_flags & O_NONBLOCK ?
299 "O_NONBLOCK" : "(null)");
300 printf("\t\tmq_maxmsg:\t\t\t%lu\n", result.mq_maxmsg);
301 printf("\t\tmq_msgsize:\t\t\t%lu\n", result.mq_msgsize);
302 printf("\t\tmq_curmsgs:\t\t\t%lu\n", result.mq_curmsgs);
303}
304
305void *fake_cont_thread(void *arg)
306{
307 int i;
308
309 for (i = 0; i < num_cpus_to_pin; i++)
310 if (cpu_threads[i] == pthread_self())
311 break;
312 printf("\tStarted fake continuous mode thread %d on CPU %d\n", i,
313 cpus_to_pin[i]);
314 while (1)
315 ;
316}
317
318void *cont_thread(void *arg)
319{
320 char buff[MSG_SIZE];
321 int i, priority;
322
323 for (i = 0; i < num_cpus_to_pin; i++)
324 if (cpu_threads[i] == pthread_self())
325 break;
326 printf("\tStarted continuous mode thread %d on CPU %d\n", i,
327 cpus_to_pin[i]);
328 while (1) {
329 while (mq_send(queue, buff, sizeof(buff), 0) == 0)
330 ;
331 mq_receive(queue, buff, sizeof(buff), &priority);
332 }
333}
334
335#define drain_queue() \
336 while (mq_receive(queue, buff, MSG_SIZE, &prio_in) == MSG_SIZE)
337
338#define do_untimed_send() \
339 do { \
340 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
341 shutdown(3, "Test send failure", __LINE__); \
342 } while (0)
343
344#define do_send_recv() \
345 do { \
346 clock_gettime(clock, &start); \
347 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
348 shutdown(3, "Test send failure", __LINE__); \
349 clock_gettime(clock, &middle); \
350 if (mq_receive(queue, buff, MSG_SIZE, &prio_in) != MSG_SIZE) \
351 shutdown(3, "Test receive failure", __LINE__); \
352 clock_gettime(clock, &end); \
353 nsec = ((middle.tv_sec - start.tv_sec) * 1000000000) + \
354 (middle.tv_nsec - start.tv_nsec); \
355 send_total.tv_nsec += nsec; \
356 if (send_total.tv_nsec >= 1000000000) { \
357 send_total.tv_sec++; \
358 send_total.tv_nsec -= 1000000000; \
359 } \
360 nsec = ((end.tv_sec - middle.tv_sec) * 1000000000) + \
361 (end.tv_nsec - middle.tv_nsec); \
362 recv_total.tv_nsec += nsec; \
363 if (recv_total.tv_nsec >= 1000000000) { \
364 recv_total.tv_sec++; \
365 recv_total.tv_nsec -= 1000000000; \
366 } \
367 } while (0)
368
369struct test {
370 char *desc;
371 void (*func)(int *);
372};
373
374void const_prio(int *prio)
375{
376 return;
377}
378
379void inc_prio(int *prio)
380{
381 if (++*prio == mq_prio_max)
382 *prio = 0;
383}
384
385void dec_prio(int *prio)
386{
387 if (--*prio < 0)
388 *prio = mq_prio_max - 1;
389}
390
391void random_prio(int *prio)
392{
393 *prio = random() % mq_prio_max;
394}
395
396struct test test2[] = {
397 {"\n\tTest #2a: Time send/recv message, queue full, constant prio\n",
398 const_prio},
399 {"\n\tTest #2b: Time send/recv message, queue full, increasing prio\n",
400 inc_prio},
401 {"\n\tTest #2c: Time send/recv message, queue full, decreasing prio\n",
402 dec_prio},
403 {"\n\tTest #2d: Time send/recv message, queue full, random prio\n",
404 random_prio},
405 {NULL, NULL}
406};
407
408/**
409 * Tests to perform (all done with MSG_SIZE messages):
410 *
411 * 1) Time to add/remove message with 0 messages on queue
412 * 1a) with constant prio
413 * 2) Time to add/remove message when queue close to capacity:
414 * 2a) with constant prio
415 * 2b) with increasing prio
416 * 2c) with decreasing prio
417 * 2d) with random prio
418 * 3) Test limits of priorities honored (double check _SC_MQ_PRIO_MAX)
419 */
420void *perf_test_thread(void *arg)
421{
422 char buff[MSG_SIZE];
423 int prio_out, prio_in;
424 int i;
425 clockid_t clock;
426 pthread_t *t;
427 struct timespec res, start, middle, end, send_total, recv_total;
428 unsigned long long nsec;
429 struct test *cur_test;
430
431 t = &cpu_threads[0];
432 printf("\n\tStarted mqueue performance test thread on CPU %d\n",
433 cpus_to_pin[0]);
434 mq_prio_max = sysconf(_SC_MQ_PRIO_MAX);
435 if (mq_prio_max == -1)
436 shutdown(2, "sysconf(_SC_MQ_PRIO_MAX)", __LINE__);
437 if (pthread_getcpuclockid(cpu_threads[0], &clock) != 0)
438 shutdown(2, "pthread_getcpuclockid", __LINE__);
439
440 if (clock_getres(clock, &res))
441 shutdown(2, "clock_getres()", __LINE__);
442
443 printf("\t\tMax priorities:\t\t\t%d\n", mq_prio_max);
444 printf("\t\tClock resolution:\t\t%lu nsec%s\n", res.tv_nsec,
445 res.tv_nsec > 1 ? "s" : "");
446
447
448
449 printf("\n\tTest #1: Time send/recv message, queue empty\n");
450 printf("\t\t(%d iterations)\n", TEST1_LOOPS);
451 prio_out = 0;
452 send_total.tv_sec = 0;
453 send_total.tv_nsec = 0;
454 recv_total.tv_sec = 0;
455 recv_total.tv_nsec = 0;
456 for (i = 0; i < TEST1_LOOPS; i++)
457 do_send_recv();
458 printf("\t\tSend msg:\t\t\t%ld.%lus total time\n",
459 send_total.tv_sec, send_total.tv_nsec);
460 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
461 send_total.tv_nsec) / TEST1_LOOPS;
462 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
463 printf("\t\tRecv msg:\t\t\t%ld.%lus total time\n",
464 recv_total.tv_sec, recv_total.tv_nsec);
465 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
466 recv_total.tv_nsec) / TEST1_LOOPS;
467 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
468
469
470 for (cur_test = test2; cur_test->desc != NULL; cur_test++) {
471 printf("%s:\n", cur_test->desc);
472 printf("\t\t(%d iterations)\n", TEST2_LOOPS);
473 prio_out = 0;
474 send_total.tv_sec = 0;
475 send_total.tv_nsec = 0;
476 recv_total.tv_sec = 0;
477 recv_total.tv_nsec = 0;
478 printf("\t\tFilling queue...");
479 fflush(stdout);
480 clock_gettime(clock, &start);
481 for (i = 0; i < result.mq_maxmsg - 1; i++) {
482 do_untimed_send();
483 cur_test->func(&prio_out);
484 }
485 clock_gettime(clock, &end);
486 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
487 1000000000) + (end.tv_nsec - start.tv_nsec);
488 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
489 nsec % 1000000000);
490 printf("\t\tTesting...");
491 fflush(stdout);
492 for (i = 0; i < TEST2_LOOPS; i++) {
493 do_send_recv();
494 cur_test->func(&prio_out);
495 }
496 printf("done.\n");
497 printf("\t\tSend msg:\t\t\t%ld.%lus total time\n",
498 send_total.tv_sec, send_total.tv_nsec);
499 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
500 send_total.tv_nsec) / TEST2_LOOPS;
501 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
502 printf("\t\tRecv msg:\t\t\t%ld.%lus total time\n",
503 recv_total.tv_sec, recv_total.tv_nsec);
504 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
505 recv_total.tv_nsec) / TEST2_LOOPS;
506 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
507 printf("\t\tDraining queue...");
508 fflush(stdout);
509 clock_gettime(clock, &start);
510 drain_queue();
511 clock_gettime(clock, &end);
512 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
513 1000000000) + (end.tv_nsec - start.tv_nsec);
514 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
515 nsec % 1000000000);
516 }
517 return 0;
518}
519
520void increase_limits(void)
521{
522 cur_limits.rlim_cur = RLIM_INFINITY;
523 cur_limits.rlim_max = RLIM_INFINITY;
524 setr(RLIMIT_MSGQUEUE, &cur_limits);
525 while (try_set(max_msgs, cur_max_msgs += 10))
526 ;
527 cur_max_msgs = get(max_msgs);
528 while (try_set(max_msgsize, cur_max_msgsize += 1024))
529 ;
530 cur_max_msgsize = get(max_msgsize);
531 if (setpriority(PRIO_PROCESS, 0, -20) != 0)
532 shutdown(2, "setpriority()", __LINE__);
533 cur_nice = -20;
534}
535
536int main(int argc, char *argv[])
537{
538 struct mq_attr attr;
539 char *option, *next_option;
540 int i, cpu, rc;
541 struct sigaction sa;
542 poptContext popt_context;
543 void *retval;
544
545 main_thread = pthread_self();
546 num_cpus_to_pin = 0;
547
548 if (sysconf(_SC_NPROCESSORS_ONLN) == -1) {
549 perror("sysconf(_SC_NPROCESSORS_ONLN)");
550 exit(1);
551 }
552 cpus_online = min(MAX_CPUS, sysconf(_SC_NPROCESSORS_ONLN));
553 cpu_set = CPU_ALLOC(cpus_online);
554 if (cpu_set == NULL) {
555 perror("CPU_ALLOC()");
556 exit(1);
557 }
558 cpu_set_size = CPU_ALLOC_SIZE(cpus_online);
559 CPU_ZERO_S(cpu_set_size, cpu_set);
560
561 popt_context = poptGetContext(NULL, argc, (const char **)argv,
562 options, 0);
563
564 while ((rc = poptGetNextOpt(popt_context)) > 0) {
565 switch (rc) {
566 case 'c':
567 continuous_mode = 1;
568 option = cpu_option_string;
569 do {
570 next_option = strchr(option, ',');
571 if (next_option)
572 *next_option = '\0';
573 cpu = atoi(option);
574 if (cpu >= cpus_online)
575 fprintf(stderr, "CPU %d exceeds "
576 "cpus online, ignoring.\n",
577 cpu);
578 else
579 cpus_to_pin[num_cpus_to_pin++] = cpu;
580 if (next_option)
581 option = ++next_option;
582 } while (next_option && num_cpus_to_pin < MAX_CPUS);
583 /* Double check that they didn't give us the same CPU
584 * more than once */
585 for (cpu = 0; cpu < num_cpus_to_pin; cpu++) {
586 if (CPU_ISSET_S(cpus_to_pin[cpu], cpu_set_size,
587 cpu_set)) {
588 fprintf(stderr, "Any given CPU may "
589 "only be given once.\n");
590 exit(1);
591 } else
592 CPU_SET_S(cpus_to_pin[cpu],
593 cpu_set_size, cpu_set);
594 }
595 break;
596 case 'p':
597 /*
598 * Although we can create a msg queue with a
599 * non-absolute path name, unlink will fail. So,
600 * if the name doesn't start with a /, add one
601 * when we save it.
602 */
603 option = queue_path;
604 if (*option != '/') {
605 queue_path = malloc(strlen(option) + 2);
606 if (!queue_path) {
607 perror("malloc()");
608 exit(1);
609 }
610 queue_path[0] = '/';
611 queue_path[1] = 0;
612 strcat(queue_path, option);
613 free(option);
614 }
615 break;
616 }
617 }
618
619 if (continuous_mode && num_cpus_to_pin == 0) {
620 fprintf(stderr, "Must pass at least one CPU to continuous "
621 "mode.\n");
622 poptPrintUsage(popt_context, stderr, 0);
623 exit(1);
624 } else if (!continuous_mode) {
625 num_cpus_to_pin = 1;
626 cpus_to_pin[0] = cpus_online - 1;
627 }
628
629 if (getuid() != 0) {
630 fprintf(stderr, "Not running as root, but almost all tests "
631 "require root in order to modify\nsystem settings. "
632 "Exiting.\n");
633 exit(1);
634 }
635
636 max_msgs = fopen(MAX_MSGS, "r+");
637 max_msgsize = fopen(MAX_MSGSIZE, "r+");
638 if (!max_msgs)
639 shutdown(2, "Failed to open msg_max", __LINE__);
640 if (!max_msgsize)
641 shutdown(2, "Failed to open msgsize_max", __LINE__);
642
643 /* Load up the current system values for everything we can */
644 getr(RLIMIT_MSGQUEUE, &saved_limits);
645 cur_limits = saved_limits;
646 saved_max_msgs = cur_max_msgs = get(max_msgs);
647 saved_max_msgsize = cur_max_msgsize = get(max_msgsize);
648 errno = 0;
649 cur_nice = getpriority(PRIO_PROCESS, 0);
650 if (errno)
651 shutdown(2, "getpriority()", __LINE__);
652
653 /* Tell the user our initial state */
654 printf("\nInitial system state:\n");
655 printf("\tUsing queue path:\t\t\t%s\n", queue_path);
656 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
657 (long) saved_limits.rlim_cur);
658 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
659 (long) saved_limits.rlim_max);
660 printf("\tMaximum Message Size:\t\t\t%d\n", saved_max_msgsize);
661 printf("\tMaximum Queue Size:\t\t\t%d\n", saved_max_msgs);
662 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
663 printf("\n");
664
665 increase_limits();
666
667 printf("Adjusted system state for testing:\n");
668 if (cur_limits.rlim_cur == RLIM_INFINITY) {
669 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t(unlimited)\n");
670 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t(unlimited)\n");
671 } else {
672 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
673 (long) cur_limits.rlim_cur);
674 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
675 (long) cur_limits.rlim_max);
676 }
677 printf("\tMaximum Message Size:\t\t\t%d\n", cur_max_msgsize);
678 printf("\tMaximum Queue Size:\t\t\t%d\n", cur_max_msgs);
679 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
680 printf("\tContinuous mode:\t\t\t(%s)\n", continuous_mode ?
681 (continuous_mode_fake ? "fake mode" : "enabled") :
682 "disabled");
683 printf("\tCPUs to pin:\t\t\t\t%d", cpus_to_pin[0]);
684 for (cpu = 1; cpu < num_cpus_to_pin; cpu++)
685 printf(",%d", cpus_to_pin[cpu]);
686 printf("\n");
687
688 sa.sa_sigaction = sig_action_SIGUSR1;
689 sigemptyset(&sa.sa_mask);
690 sigaddset(&sa.sa_mask, SIGHUP);
691 sigaddset(&sa.sa_mask, SIGINT);
692 sigaddset(&sa.sa_mask, SIGQUIT);
693 sigaddset(&sa.sa_mask, SIGTERM);
694 sa.sa_flags = SA_SIGINFO;
695 if (sigaction(SIGUSR1, &sa, NULL) == -1)
696 shutdown(1, "sigaction(SIGUSR1)", __LINE__);
697 sa.sa_sigaction = sig_action;
698 if (sigaction(SIGHUP, &sa, NULL) == -1)
699 shutdown(1, "sigaction(SIGHUP)", __LINE__);
700 if (sigaction(SIGINT, &sa, NULL) == -1)
701 shutdown(1, "sigaction(SIGINT)", __LINE__);
702 if (sigaction(SIGQUIT, &sa, NULL) == -1)
703 shutdown(1, "sigaction(SIGQUIT)", __LINE__);
704 if (sigaction(SIGTERM, &sa, NULL) == -1)
705 shutdown(1, "sigaction(SIGTERM)", __LINE__);
706
707 if (!continuous_mode_fake) {
708 attr.mq_flags = O_NONBLOCK;
709 attr.mq_maxmsg = cur_max_msgs;
710 attr.mq_msgsize = MSG_SIZE;
711 open_queue(&attr);
712 }
713 for (i = 0; i < num_cpus_to_pin; i++) {
714 pthread_attr_t thread_attr;
715 void *thread_func;
716
717 if (continuous_mode_fake)
718 thread_func = &fake_cont_thread;
719 else if (continuous_mode)
720 thread_func = &cont_thread;
721 else
722 thread_func = &perf_test_thread;
723
724 CPU_ZERO_S(cpu_set_size, cpu_set);
725 CPU_SET_S(cpus_to_pin[i], cpu_set_size, cpu_set);
726 pthread_attr_init(&thread_attr);
727 pthread_attr_setaffinity_np(&thread_attr, cpu_set_size,
728 cpu_set);
729 if (pthread_create(&cpu_threads[i], &thread_attr, thread_func,
730 NULL))
731 shutdown(1, "pthread_create()", __LINE__);
732 pthread_attr_destroy(&thread_attr);
733 }
734
735 if (!continuous_mode) {
736 pthread_join(cpu_threads[0], &retval);
737 shutdown((long)retval, "perf_test_thread()", __LINE__);
738 } else {
739 while (1)
740 sleep(1);
741 }
742 shutdown(0, "", 0);
743}
1/*
2 * This application is Copyright 2012 Red Hat, Inc.
3 * Doug Ledford <dledford@redhat.com>
4 *
5 * mq_perf_tests is free software: you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation, version 3.
8 *
9 * mq_perf_tests is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * For the full text of the license, see <http://www.gnu.org/licenses/>.
15 *
16 * mq_perf_tests.c
17 * Tests various types of message queue workloads, concentrating on those
18 * situations that invole large message sizes, large message queue depths,
19 * or both, and reports back useful metrics about kernel message queue
20 * performance.
21 *
22 */
23#define _GNU_SOURCE
24#include <stdio.h>
25#include <stdlib.h>
26#include <unistd.h>
27#include <fcntl.h>
28#include <string.h>
29#include <limits.h>
30#include <errno.h>
31#include <signal.h>
32#include <pthread.h>
33#include <sched.h>
34#include <sys/types.h>
35#include <sys/time.h>
36#include <sys/resource.h>
37#include <sys/stat.h>
38#include <sys/param.h>
39#include <mqueue.h>
40#include <popt.h>
41#include <error.h>
42
43#include "../kselftest.h"
44
45static char *usage =
46"Usage:\n"
47" %s [-c #[,#..] -f] path\n"
48"\n"
49" -c # Skip most tests and go straight to a high queue depth test\n"
50" and then run that test continuously (useful for running at\n"
51" the same time as some other workload to see how much the\n"
52" cache thrashing caused by adding messages to a very deep\n"
53" queue impacts the performance of other programs). The number\n"
54" indicates which CPU core we should bind the process to during\n"
55" the run. If you have more than one physical CPU, then you\n"
56" will need one copy per physical CPU package, and you should\n"
57" specify the CPU cores to pin ourself to via a comma separated\n"
58" list of CPU values.\n"
59" -f Only usable with continuous mode. Pin ourself to the CPUs\n"
60" as requested, then instead of looping doing a high mq\n"
61" workload, just busy loop. This will allow us to lock up a\n"
62" single CPU just like we normally would, but without actually\n"
63" thrashing the CPU cache. This is to make it easier to get\n"
64" comparable numbers from some other workload running on the\n"
65" other CPUs. One set of numbers with # CPUs locked up running\n"
66" an mq workload, and another set of numbers with those same\n"
67" CPUs locked away from the test workload, but not doing\n"
68" anything to trash the cache like the mq workload might.\n"
69" path Path name of the message queue to create\n"
70"\n"
71" Note: this program must be run as root in order to enable all tests\n"
72"\n";
73
74char *MAX_MSGS = "/proc/sys/fs/mqueue/msg_max";
75char *MAX_MSGSIZE = "/proc/sys/fs/mqueue/msgsize_max";
76
77#define MAX_CPUS 64
78char *cpu_option_string;
79int cpus_to_pin[MAX_CPUS];
80int num_cpus_to_pin;
81pthread_t cpu_threads[MAX_CPUS];
82pthread_t main_thread;
83cpu_set_t *cpu_set;
84int cpu_set_size;
85int cpus_online;
86
87#define MSG_SIZE 16
88#define TEST1_LOOPS 10000000
89#define TEST2_LOOPS 100000
90int continuous_mode;
91int continuous_mode_fake;
92
93struct rlimit saved_limits, cur_limits;
94int saved_max_msgs, saved_max_msgsize;
95int cur_max_msgs, cur_max_msgsize;
96FILE *max_msgs, *max_msgsize;
97int cur_nice;
98char *queue_path = "/mq_perf_tests";
99mqd_t queue = -1;
100struct mq_attr result;
101int mq_prio_max;
102
103const struct poptOption options[] = {
104 {
105 .longName = "continuous",
106 .shortName = 'c',
107 .argInfo = POPT_ARG_STRING,
108 .arg = &cpu_option_string,
109 .val = 'c',
110 .descrip = "Run continuous tests at a high queue depth in "
111 "order to test the effects of cache thrashing on "
112 "other tasks on the system. This test is intended "
113 "to be run on one core of each physical CPU while "
114 "some other CPU intensive task is run on all the other "
115 "cores of that same physical CPU and the other task "
116 "is timed. It is assumed that the process of adding "
117 "messages to the message queue in a tight loop will "
118 "impact that other task to some degree. Once the "
119 "tests are performed in this way, you should then "
120 "re-run the tests using fake mode in order to check "
121 "the difference in time required to perform the CPU "
122 "intensive task",
123 .argDescrip = "cpu[,cpu]",
124 },
125 {
126 .longName = "fake",
127 .shortName = 'f',
128 .argInfo = POPT_ARG_NONE,
129 .arg = &continuous_mode_fake,
130 .val = 0,
131 .descrip = "Tie up the CPUs that we would normally tie up in"
132 "continuous mode, but don't actually do any mq stuff, "
133 "just keep the CPU busy so it can't be used to process "
134 "system level tasks as this would free up resources on "
135 "the other CPU cores and skew the comparison between "
136 "the no-mqueue work and mqueue work tests",
137 .argDescrip = NULL,
138 },
139 {
140 .longName = "path",
141 .shortName = 'p',
142 .argInfo = POPT_ARG_STRING | POPT_ARGFLAG_SHOW_DEFAULT,
143 .arg = &queue_path,
144 .val = 'p',
145 .descrip = "The name of the path to use in the mqueue "
146 "filesystem for our tests",
147 .argDescrip = "pathname",
148 },
149 POPT_AUTOHELP
150 POPT_TABLEEND
151};
152
153static inline void __set(FILE *stream, int value, char *err_msg);
154void shutdown(int exit_val, char *err_cause, int line_no);
155void sig_action_SIGUSR1(int signum, siginfo_t *info, void *context);
156void sig_action(int signum, siginfo_t *info, void *context);
157static inline int get(FILE *stream);
158static inline void set(FILE *stream, int value);
159static inline int try_set(FILE *stream, int value);
160static inline void getr(int type, struct rlimit *rlim);
161static inline void setr(int type, struct rlimit *rlim);
162static inline void open_queue(struct mq_attr *attr);
163void increase_limits(void);
164
165static inline void __set(FILE *stream, int value, char *err_msg)
166{
167 rewind(stream);
168 if (fprintf(stream, "%d", value) < 0)
169 perror(err_msg);
170}
171
172
173void shutdown(int exit_val, char *err_cause, int line_no)
174{
175 static int in_shutdown = 0;
176 int errno_at_shutdown = errno;
177 int i;
178
179 /* In case we get called by multiple threads or from an sighandler */
180 if (in_shutdown++)
181 return;
182
183 /* Free the cpu_set allocated using CPU_ALLOC in main function */
184 CPU_FREE(cpu_set);
185
186 for (i = 0; i < num_cpus_to_pin; i++)
187 if (cpu_threads[i]) {
188 pthread_kill(cpu_threads[i], SIGUSR1);
189 pthread_join(cpu_threads[i], NULL);
190 }
191
192 if (queue != -1)
193 if (mq_close(queue))
194 perror("mq_close() during shutdown");
195 if (queue_path)
196 /*
197 * Be silent if this fails, if we cleaned up already it's
198 * expected to fail
199 */
200 mq_unlink(queue_path);
201 if (saved_max_msgs)
202 __set(max_msgs, saved_max_msgs,
203 "failed to restore saved_max_msgs");
204 if (saved_max_msgsize)
205 __set(max_msgsize, saved_max_msgsize,
206 "failed to restore saved_max_msgsize");
207 if (exit_val)
208 error(exit_val, errno_at_shutdown, "%s at %d",
209 err_cause, line_no);
210 exit(0);
211}
212
213void sig_action_SIGUSR1(int signum, siginfo_t *info, void *context)
214{
215 if (pthread_self() != main_thread)
216 pthread_exit(0);
217 else {
218 fprintf(stderr, "Caught signal %d in SIGUSR1 handler, "
219 "exiting\n", signum);
220 shutdown(0, "", 0);
221 fprintf(stderr, "\n\nReturned from shutdown?!?!\n\n");
222 exit(0);
223 }
224}
225
226void sig_action(int signum, siginfo_t *info, void *context)
227{
228 if (pthread_self() != main_thread)
229 pthread_kill(main_thread, signum);
230 else {
231 fprintf(stderr, "Caught signal %d, exiting\n", signum);
232 shutdown(0, "", 0);
233 fprintf(stderr, "\n\nReturned from shutdown?!?!\n\n");
234 exit(0);
235 }
236}
237
238static inline int get(FILE *stream)
239{
240 int value;
241 rewind(stream);
242 if (fscanf(stream, "%d", &value) != 1)
243 shutdown(4, "Error reading /proc entry", __LINE__);
244 return value;
245}
246
247static inline void set(FILE *stream, int value)
248{
249 int new_value;
250
251 rewind(stream);
252 if (fprintf(stream, "%d", value) < 0)
253 return shutdown(5, "Failed writing to /proc file", __LINE__);
254 new_value = get(stream);
255 if (new_value != value)
256 return shutdown(5, "We didn't get what we wrote to /proc back",
257 __LINE__);
258}
259
260static inline int try_set(FILE *stream, int value)
261{
262 int new_value;
263
264 rewind(stream);
265 fprintf(stream, "%d", value);
266 new_value = get(stream);
267 return new_value == value;
268}
269
270static inline void getr(int type, struct rlimit *rlim)
271{
272 if (getrlimit(type, rlim))
273 shutdown(6, "getrlimit()", __LINE__);
274}
275
276static inline void setr(int type, struct rlimit *rlim)
277{
278 if (setrlimit(type, rlim))
279 shutdown(7, "setrlimit()", __LINE__);
280}
281
282/**
283 * open_queue - open the global queue for testing
284 * @attr - An attr struct specifying the desired queue traits
285 * @result - An attr struct that lists the actual traits the queue has
286 *
287 * This open is not allowed to fail, failure will result in an orderly
288 * shutdown of the program. The global queue_path is used to set what
289 * queue to open, the queue descriptor is saved in the global queue
290 * variable.
291 */
292static inline void open_queue(struct mq_attr *attr)
293{
294 int flags = O_RDWR | O_EXCL | O_CREAT | O_NONBLOCK;
295 int perms = DEFFILEMODE;
296
297 queue = mq_open(queue_path, flags, perms, attr);
298 if (queue == -1)
299 shutdown(1, "mq_open()", __LINE__);
300 if (mq_getattr(queue, &result))
301 shutdown(1, "mq_getattr()", __LINE__);
302 printf("\n\tQueue %s created:\n", queue_path);
303 printf("\t\tmq_flags:\t\t\t%s\n", result.mq_flags & O_NONBLOCK ?
304 "O_NONBLOCK" : "(null)");
305 printf("\t\tmq_maxmsg:\t\t\t%lu\n", result.mq_maxmsg);
306 printf("\t\tmq_msgsize:\t\t\t%lu\n", result.mq_msgsize);
307 printf("\t\tmq_curmsgs:\t\t\t%lu\n", result.mq_curmsgs);
308}
309
310void *fake_cont_thread(void *arg)
311{
312 int i;
313
314 for (i = 0; i < num_cpus_to_pin; i++)
315 if (cpu_threads[i] == pthread_self())
316 break;
317 printf("\tStarted fake continuous mode thread %d on CPU %d\n", i,
318 cpus_to_pin[i]);
319 while (1)
320 ;
321}
322
323void *cont_thread(void *arg)
324{
325 char buff[MSG_SIZE];
326 int i, priority;
327
328 for (i = 0; i < num_cpus_to_pin; i++)
329 if (cpu_threads[i] == pthread_self())
330 break;
331 printf("\tStarted continuous mode thread %d on CPU %d\n", i,
332 cpus_to_pin[i]);
333 while (1) {
334 while (mq_send(queue, buff, sizeof(buff), 0) == 0)
335 ;
336 mq_receive(queue, buff, sizeof(buff), &priority);
337 }
338}
339
340#define drain_queue() \
341 while (mq_receive(queue, buff, MSG_SIZE, &prio_in) == MSG_SIZE)
342
343#define do_untimed_send() \
344 do { \
345 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
346 shutdown(3, "Test send failure", __LINE__); \
347 } while (0)
348
349#define do_send_recv() \
350 do { \
351 clock_gettime(clock, &start); \
352 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
353 shutdown(3, "Test send failure", __LINE__); \
354 clock_gettime(clock, &middle); \
355 if (mq_receive(queue, buff, MSG_SIZE, &prio_in) != MSG_SIZE) \
356 shutdown(3, "Test receive failure", __LINE__); \
357 clock_gettime(clock, &end); \
358 nsec = ((middle.tv_sec - start.tv_sec) * 1000000000) + \
359 (middle.tv_nsec - start.tv_nsec); \
360 send_total.tv_nsec += nsec; \
361 if (send_total.tv_nsec >= 1000000000) { \
362 send_total.tv_sec++; \
363 send_total.tv_nsec -= 1000000000; \
364 } \
365 nsec = ((end.tv_sec - middle.tv_sec) * 1000000000) + \
366 (end.tv_nsec - middle.tv_nsec); \
367 recv_total.tv_nsec += nsec; \
368 if (recv_total.tv_nsec >= 1000000000) { \
369 recv_total.tv_sec++; \
370 recv_total.tv_nsec -= 1000000000; \
371 } \
372 } while (0)
373
374struct test {
375 char *desc;
376 void (*func)(int *);
377};
378
379void const_prio(int *prio)
380{
381 return;
382}
383
384void inc_prio(int *prio)
385{
386 if (++*prio == mq_prio_max)
387 *prio = 0;
388}
389
390void dec_prio(int *prio)
391{
392 if (--*prio < 0)
393 *prio = mq_prio_max - 1;
394}
395
396void random_prio(int *prio)
397{
398 *prio = random() % mq_prio_max;
399}
400
401struct test test2[] = {
402 {"\n\tTest #2a: Time send/recv message, queue full, constant prio\n",
403 const_prio},
404 {"\n\tTest #2b: Time send/recv message, queue full, increasing prio\n",
405 inc_prio},
406 {"\n\tTest #2c: Time send/recv message, queue full, decreasing prio\n",
407 dec_prio},
408 {"\n\tTest #2d: Time send/recv message, queue full, random prio\n",
409 random_prio},
410 {NULL, NULL}
411};
412
413/**
414 * Tests to perform (all done with MSG_SIZE messages):
415 *
416 * 1) Time to add/remove message with 0 messages on queue
417 * 1a) with constant prio
418 * 2) Time to add/remove message when queue close to capacity:
419 * 2a) with constant prio
420 * 2b) with increasing prio
421 * 2c) with decreasing prio
422 * 2d) with random prio
423 * 3) Test limits of priorities honored (double check _SC_MQ_PRIO_MAX)
424 */
425void *perf_test_thread(void *arg)
426{
427 char buff[MSG_SIZE];
428 int prio_out, prio_in;
429 int i;
430 clockid_t clock;
431 pthread_t *t;
432 struct timespec res, start, middle, end, send_total, recv_total;
433 unsigned long long nsec;
434 struct test *cur_test;
435
436 t = &cpu_threads[0];
437 printf("\n\tStarted mqueue performance test thread on CPU %d\n",
438 cpus_to_pin[0]);
439 mq_prio_max = sysconf(_SC_MQ_PRIO_MAX);
440 if (mq_prio_max == -1)
441 shutdown(2, "sysconf(_SC_MQ_PRIO_MAX)", __LINE__);
442 if (pthread_getcpuclockid(cpu_threads[0], &clock) != 0)
443 shutdown(2, "pthread_getcpuclockid", __LINE__);
444
445 if (clock_getres(clock, &res))
446 shutdown(2, "clock_getres()", __LINE__);
447
448 printf("\t\tMax priorities:\t\t\t%d\n", mq_prio_max);
449 printf("\t\tClock resolution:\t\t%lu nsec%s\n", res.tv_nsec,
450 res.tv_nsec > 1 ? "s" : "");
451
452
453
454 printf("\n\tTest #1: Time send/recv message, queue empty\n");
455 printf("\t\t(%d iterations)\n", TEST1_LOOPS);
456 prio_out = 0;
457 send_total.tv_sec = 0;
458 send_total.tv_nsec = 0;
459 recv_total.tv_sec = 0;
460 recv_total.tv_nsec = 0;
461 for (i = 0; i < TEST1_LOOPS; i++)
462 do_send_recv();
463 printf("\t\tSend msg:\t\t\t%ld.%lus total time\n",
464 send_total.tv_sec, send_total.tv_nsec);
465 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
466 send_total.tv_nsec) / TEST1_LOOPS;
467 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
468 printf("\t\tRecv msg:\t\t\t%ld.%lus total time\n",
469 recv_total.tv_sec, recv_total.tv_nsec);
470 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
471 recv_total.tv_nsec) / TEST1_LOOPS;
472 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
473
474
475 for (cur_test = test2; cur_test->desc != NULL; cur_test++) {
476 printf("%s:\n", cur_test->desc);
477 printf("\t\t(%d iterations)\n", TEST2_LOOPS);
478 prio_out = 0;
479 send_total.tv_sec = 0;
480 send_total.tv_nsec = 0;
481 recv_total.tv_sec = 0;
482 recv_total.tv_nsec = 0;
483 printf("\t\tFilling queue...");
484 fflush(stdout);
485 clock_gettime(clock, &start);
486 for (i = 0; i < result.mq_maxmsg - 1; i++) {
487 do_untimed_send();
488 cur_test->func(&prio_out);
489 }
490 clock_gettime(clock, &end);
491 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
492 1000000000) + (end.tv_nsec - start.tv_nsec);
493 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
494 nsec % 1000000000);
495 printf("\t\tTesting...");
496 fflush(stdout);
497 for (i = 0; i < TEST2_LOOPS; i++) {
498 do_send_recv();
499 cur_test->func(&prio_out);
500 }
501 printf("done.\n");
502 printf("\t\tSend msg:\t\t\t%ld.%lus total time\n",
503 send_total.tv_sec, send_total.tv_nsec);
504 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
505 send_total.tv_nsec) / TEST2_LOOPS;
506 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
507 printf("\t\tRecv msg:\t\t\t%ld.%lus total time\n",
508 recv_total.tv_sec, recv_total.tv_nsec);
509 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
510 recv_total.tv_nsec) / TEST2_LOOPS;
511 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
512 printf("\t\tDraining queue...");
513 fflush(stdout);
514 clock_gettime(clock, &start);
515 drain_queue();
516 clock_gettime(clock, &end);
517 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
518 1000000000) + (end.tv_nsec - start.tv_nsec);
519 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
520 nsec % 1000000000);
521 }
522 return 0;
523}
524
525void increase_limits(void)
526{
527 cur_limits.rlim_cur = RLIM_INFINITY;
528 cur_limits.rlim_max = RLIM_INFINITY;
529 setr(RLIMIT_MSGQUEUE, &cur_limits);
530 while (try_set(max_msgs, cur_max_msgs += 10))
531 ;
532 cur_max_msgs = get(max_msgs);
533 while (try_set(max_msgsize, cur_max_msgsize += 1024))
534 ;
535 cur_max_msgsize = get(max_msgsize);
536 if (setpriority(PRIO_PROCESS, 0, -20) != 0)
537 shutdown(2, "setpriority()", __LINE__);
538 cur_nice = -20;
539}
540
541int main(int argc, char *argv[])
542{
543 struct mq_attr attr;
544 char *option, *next_option;
545 int i, cpu, rc;
546 struct sigaction sa;
547 poptContext popt_context;
548 void *retval;
549
550 main_thread = pthread_self();
551 num_cpus_to_pin = 0;
552
553 if (sysconf(_SC_NPROCESSORS_ONLN) == -1) {
554 perror("sysconf(_SC_NPROCESSORS_ONLN)");
555 exit(1);
556 }
557
558 if (getuid() != 0)
559 ksft_exit_skip("Not running as root, but almost all tests "
560 "require root in order to modify\nsystem settings. "
561 "Exiting.\n");
562
563 cpus_online = MIN(MAX_CPUS, sysconf(_SC_NPROCESSORS_ONLN));
564 cpu_set = CPU_ALLOC(cpus_online);
565 if (cpu_set == NULL) {
566 perror("CPU_ALLOC()");
567 exit(1);
568 }
569 cpu_set_size = CPU_ALLOC_SIZE(cpus_online);
570 CPU_ZERO_S(cpu_set_size, cpu_set);
571
572 popt_context = poptGetContext(NULL, argc, (const char **)argv,
573 options, 0);
574
575 while ((rc = poptGetNextOpt(popt_context)) > 0) {
576 switch (rc) {
577 case 'c':
578 continuous_mode = 1;
579 option = cpu_option_string;
580 do {
581 next_option = strchr(option, ',');
582 if (next_option)
583 *next_option = '\0';
584 cpu = atoi(option);
585 if (cpu >= cpus_online)
586 fprintf(stderr, "CPU %d exceeds "
587 "cpus online, ignoring.\n",
588 cpu);
589 else
590 cpus_to_pin[num_cpus_to_pin++] = cpu;
591 if (next_option)
592 option = ++next_option;
593 } while (next_option && num_cpus_to_pin < MAX_CPUS);
594 /* Double check that they didn't give us the same CPU
595 * more than once */
596 for (cpu = 0; cpu < num_cpus_to_pin; cpu++) {
597 if (CPU_ISSET_S(cpus_to_pin[cpu], cpu_set_size,
598 cpu_set)) {
599 fprintf(stderr, "Any given CPU may "
600 "only be given once.\n");
601 goto err_code;
602 } else
603 CPU_SET_S(cpus_to_pin[cpu],
604 cpu_set_size, cpu_set);
605 }
606 break;
607 case 'p':
608 /*
609 * Although we can create a msg queue with a
610 * non-absolute path name, unlink will fail. So,
611 * if the name doesn't start with a /, add one
612 * when we save it.
613 */
614 option = queue_path;
615 if (*option != '/') {
616 queue_path = malloc(strlen(option) + 2);
617 if (!queue_path) {
618 perror("malloc()");
619 goto err_code;
620 }
621 queue_path[0] = '/';
622 queue_path[1] = 0;
623 strcat(queue_path, option);
624 free(option);
625 }
626 break;
627 }
628 }
629
630 if (continuous_mode && num_cpus_to_pin == 0) {
631 fprintf(stderr, "Must pass at least one CPU to continuous "
632 "mode.\n");
633 poptPrintUsage(popt_context, stderr, 0);
634 goto err_code;
635 } else if (!continuous_mode) {
636 num_cpus_to_pin = 1;
637 cpus_to_pin[0] = cpus_online - 1;
638 }
639
640 max_msgs = fopen(MAX_MSGS, "r+");
641 max_msgsize = fopen(MAX_MSGSIZE, "r+");
642 if (!max_msgs)
643 shutdown(2, "Failed to open msg_max", __LINE__);
644 if (!max_msgsize)
645 shutdown(2, "Failed to open msgsize_max", __LINE__);
646
647 /* Load up the current system values for everything we can */
648 getr(RLIMIT_MSGQUEUE, &saved_limits);
649 cur_limits = saved_limits;
650 saved_max_msgs = cur_max_msgs = get(max_msgs);
651 saved_max_msgsize = cur_max_msgsize = get(max_msgsize);
652 errno = 0;
653 cur_nice = getpriority(PRIO_PROCESS, 0);
654 if (errno)
655 shutdown(2, "getpriority()", __LINE__);
656
657 /* Tell the user our initial state */
658 printf("\nInitial system state:\n");
659 printf("\tUsing queue path:\t\t\t%s\n", queue_path);
660 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
661 (long) saved_limits.rlim_cur);
662 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
663 (long) saved_limits.rlim_max);
664 printf("\tMaximum Message Size:\t\t\t%d\n", saved_max_msgsize);
665 printf("\tMaximum Queue Size:\t\t\t%d\n", saved_max_msgs);
666 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
667 printf("\n");
668
669 increase_limits();
670
671 printf("Adjusted system state for testing:\n");
672 if (cur_limits.rlim_cur == RLIM_INFINITY) {
673 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t(unlimited)\n");
674 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t(unlimited)\n");
675 } else {
676 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
677 (long) cur_limits.rlim_cur);
678 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
679 (long) cur_limits.rlim_max);
680 }
681 printf("\tMaximum Message Size:\t\t\t%d\n", cur_max_msgsize);
682 printf("\tMaximum Queue Size:\t\t\t%d\n", cur_max_msgs);
683 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
684 printf("\tContinuous mode:\t\t\t(%s)\n", continuous_mode ?
685 (continuous_mode_fake ? "fake mode" : "enabled") :
686 "disabled");
687 printf("\tCPUs to pin:\t\t\t\t%d", cpus_to_pin[0]);
688 for (cpu = 1; cpu < num_cpus_to_pin; cpu++)
689 printf(",%d", cpus_to_pin[cpu]);
690 printf("\n");
691
692 sa.sa_sigaction = sig_action_SIGUSR1;
693 sigemptyset(&sa.sa_mask);
694 sigaddset(&sa.sa_mask, SIGHUP);
695 sigaddset(&sa.sa_mask, SIGINT);
696 sigaddset(&sa.sa_mask, SIGQUIT);
697 sigaddset(&sa.sa_mask, SIGTERM);
698 sa.sa_flags = SA_SIGINFO;
699 if (sigaction(SIGUSR1, &sa, NULL) == -1)
700 shutdown(1, "sigaction(SIGUSR1)", __LINE__);
701 sa.sa_sigaction = sig_action;
702 if (sigaction(SIGHUP, &sa, NULL) == -1)
703 shutdown(1, "sigaction(SIGHUP)", __LINE__);
704 if (sigaction(SIGINT, &sa, NULL) == -1)
705 shutdown(1, "sigaction(SIGINT)", __LINE__);
706 if (sigaction(SIGQUIT, &sa, NULL) == -1)
707 shutdown(1, "sigaction(SIGQUIT)", __LINE__);
708 if (sigaction(SIGTERM, &sa, NULL) == -1)
709 shutdown(1, "sigaction(SIGTERM)", __LINE__);
710
711 if (!continuous_mode_fake) {
712 attr.mq_flags = O_NONBLOCK;
713 attr.mq_maxmsg = cur_max_msgs;
714 attr.mq_msgsize = MSG_SIZE;
715 open_queue(&attr);
716 }
717 for (i = 0; i < num_cpus_to_pin; i++) {
718 pthread_attr_t thread_attr;
719 void *thread_func;
720
721 if (continuous_mode_fake)
722 thread_func = &fake_cont_thread;
723 else if (continuous_mode)
724 thread_func = &cont_thread;
725 else
726 thread_func = &perf_test_thread;
727
728 CPU_ZERO_S(cpu_set_size, cpu_set);
729 CPU_SET_S(cpus_to_pin[i], cpu_set_size, cpu_set);
730 pthread_attr_init(&thread_attr);
731 pthread_attr_setaffinity_np(&thread_attr, cpu_set_size,
732 cpu_set);
733 if (pthread_create(&cpu_threads[i], &thread_attr, thread_func,
734 NULL))
735 shutdown(1, "pthread_create()", __LINE__);
736 pthread_attr_destroy(&thread_attr);
737 }
738
739 if (!continuous_mode) {
740 pthread_join(cpu_threads[0], &retval);
741 shutdown((long)retval, "perf_test_thread()", __LINE__);
742 } else {
743 while (1)
744 sleep(1);
745 }
746 shutdown(0, "", 0);
747
748err_code:
749 CPU_FREE(cpu_set);
750 exit(1);
751
752}