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 <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;
327 unsigned int priority;
328
329 for (i = 0; i < num_cpus_to_pin; i++)
330 if (cpu_threads[i] == pthread_self())
331 break;
332 printf("\tStarted continuous mode thread %d on CPU %d\n", i,
333 cpus_to_pin[i]);
334 while (1) {
335 while (mq_send(queue, buff, sizeof(buff), 0) == 0)
336 ;
337 mq_receive(queue, buff, sizeof(buff), &priority);
338 }
339}
340
341#define drain_queue() \
342 while (mq_receive(queue, buff, MSG_SIZE, &prio_in) == MSG_SIZE)
343
344#define do_untimed_send() \
345 do { \
346 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
347 shutdown(3, "Test send failure", __LINE__); \
348 } while (0)
349
350#define do_send_recv() \
351 do { \
352 clock_gettime(clock, &start); \
353 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
354 shutdown(3, "Test send failure", __LINE__); \
355 clock_gettime(clock, &middle); \
356 if (mq_receive(queue, buff, MSG_SIZE, &prio_in) != MSG_SIZE) \
357 shutdown(3, "Test receive failure", __LINE__); \
358 clock_gettime(clock, &end); \
359 nsec = ((middle.tv_sec - start.tv_sec) * 1000000000) + \
360 (middle.tv_nsec - start.tv_nsec); \
361 send_total.tv_nsec += nsec; \
362 if (send_total.tv_nsec >= 1000000000) { \
363 send_total.tv_sec++; \
364 send_total.tv_nsec -= 1000000000; \
365 } \
366 nsec = ((end.tv_sec - middle.tv_sec) * 1000000000) + \
367 (end.tv_nsec - middle.tv_nsec); \
368 recv_total.tv_nsec += nsec; \
369 if (recv_total.tv_nsec >= 1000000000) { \
370 recv_total.tv_sec++; \
371 recv_total.tv_nsec -= 1000000000; \
372 } \
373 } while (0)
374
375struct test {
376 char *desc;
377 void (*func)(int *);
378};
379
380void const_prio(int *prio)
381{
382 return;
383}
384
385void inc_prio(int *prio)
386{
387 if (++*prio == mq_prio_max)
388 *prio = 0;
389}
390
391void dec_prio(int *prio)
392{
393 if (--*prio < 0)
394 *prio = mq_prio_max - 1;
395}
396
397void random_prio(int *prio)
398{
399 *prio = random() % mq_prio_max;
400}
401
402struct test test2[] = {
403 {"\n\tTest #2a: Time send/recv message, queue full, constant prio\n",
404 const_prio},
405 {"\n\tTest #2b: Time send/recv message, queue full, increasing prio\n",
406 inc_prio},
407 {"\n\tTest #2c: Time send/recv message, queue full, decreasing prio\n",
408 dec_prio},
409 {"\n\tTest #2d: Time send/recv message, queue full, random prio\n",
410 random_prio},
411 {NULL, NULL}
412};
413
414/**
415 * Tests to perform (all done with MSG_SIZE messages):
416 *
417 * 1) Time to add/remove message with 0 messages on queue
418 * 1a) with constant prio
419 * 2) Time to add/remove message when queue close to capacity:
420 * 2a) with constant prio
421 * 2b) with increasing prio
422 * 2c) with decreasing prio
423 * 2d) with random prio
424 * 3) Test limits of priorities honored (double check _SC_MQ_PRIO_MAX)
425 */
426void *perf_test_thread(void *arg)
427{
428 char buff[MSG_SIZE];
429 int prio_out;
430 unsigned int prio_in;
431 int i;
432 clockid_t clock;
433 pthread_t *t;
434 struct timespec res, start, middle, end, send_total, recv_total;
435 unsigned long long nsec;
436 struct test *cur_test;
437
438 t = &cpu_threads[0];
439 printf("\n\tStarted mqueue performance test thread on CPU %d\n",
440 cpus_to_pin[0]);
441 mq_prio_max = sysconf(_SC_MQ_PRIO_MAX);
442 if (mq_prio_max == -1)
443 shutdown(2, "sysconf(_SC_MQ_PRIO_MAX)", __LINE__);
444 if (pthread_getcpuclockid(cpu_threads[0], &clock) != 0)
445 shutdown(2, "pthread_getcpuclockid", __LINE__);
446
447 if (clock_getres(clock, &res))
448 shutdown(2, "clock_getres()", __LINE__);
449
450 printf("\t\tMax priorities:\t\t\t%d\n", mq_prio_max);
451 printf("\t\tClock resolution:\t\t%lu nsec%s\n", res.tv_nsec,
452 res.tv_nsec > 1 ? "s" : "");
453
454
455
456 printf("\n\tTest #1: Time send/recv message, queue empty\n");
457 printf("\t\t(%d iterations)\n", TEST1_LOOPS);
458 prio_out = 0;
459 send_total.tv_sec = 0;
460 send_total.tv_nsec = 0;
461 recv_total.tv_sec = 0;
462 recv_total.tv_nsec = 0;
463 for (i = 0; i < TEST1_LOOPS; i++)
464 do_send_recv();
465 printf("\t\tSend msg:\t\t\t%ld.%lus total time\n",
466 send_total.tv_sec, send_total.tv_nsec);
467 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
468 send_total.tv_nsec) / TEST1_LOOPS;
469 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
470 printf("\t\tRecv msg:\t\t\t%ld.%lus total time\n",
471 recv_total.tv_sec, recv_total.tv_nsec);
472 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
473 recv_total.tv_nsec) / TEST1_LOOPS;
474 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
475
476
477 for (cur_test = test2; cur_test->desc != NULL; cur_test++) {
478 printf("%s:\n", cur_test->desc);
479 printf("\t\t(%d iterations)\n", TEST2_LOOPS);
480 prio_out = 0;
481 send_total.tv_sec = 0;
482 send_total.tv_nsec = 0;
483 recv_total.tv_sec = 0;
484 recv_total.tv_nsec = 0;
485 printf("\t\tFilling queue...");
486 fflush(stdout);
487 clock_gettime(clock, &start);
488 for (i = 0; i < result.mq_maxmsg - 1; i++) {
489 do_untimed_send();
490 cur_test->func(&prio_out);
491 }
492 clock_gettime(clock, &end);
493 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
494 1000000000) + (end.tv_nsec - start.tv_nsec);
495 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
496 nsec % 1000000000);
497 printf("\t\tTesting...");
498 fflush(stdout);
499 for (i = 0; i < TEST2_LOOPS; i++) {
500 do_send_recv();
501 cur_test->func(&prio_out);
502 }
503 printf("done.\n");
504 printf("\t\tSend msg:\t\t\t%ld.%lus total time\n",
505 send_total.tv_sec, send_total.tv_nsec);
506 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
507 send_total.tv_nsec) / TEST2_LOOPS;
508 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
509 printf("\t\tRecv msg:\t\t\t%ld.%lus total time\n",
510 recv_total.tv_sec, recv_total.tv_nsec);
511 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
512 recv_total.tv_nsec) / TEST2_LOOPS;
513 printf("\t\t\t\t\t\t%lld nsec/msg\n", nsec);
514 printf("\t\tDraining queue...");
515 fflush(stdout);
516 clock_gettime(clock, &start);
517 drain_queue();
518 clock_gettime(clock, &end);
519 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
520 1000000000) + (end.tv_nsec - start.tv_nsec);
521 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
522 nsec % 1000000000);
523 }
524 return 0;
525}
526
527void increase_limits(void)
528{
529 cur_limits.rlim_cur = RLIM_INFINITY;
530 cur_limits.rlim_max = RLIM_INFINITY;
531 setr(RLIMIT_MSGQUEUE, &cur_limits);
532 while (try_set(max_msgs, cur_max_msgs += 10))
533 ;
534 cur_max_msgs = get(max_msgs);
535 while (try_set(max_msgsize, cur_max_msgsize += 1024))
536 ;
537 cur_max_msgsize = get(max_msgsize);
538 if (setpriority(PRIO_PROCESS, 0, -20) != 0)
539 shutdown(2, "setpriority()", __LINE__);
540 cur_nice = -20;
541}
542
543int main(int argc, char *argv[])
544{
545 struct mq_attr attr;
546 char *option, *next_option;
547 int i, cpu, rc;
548 struct sigaction sa;
549 poptContext popt_context;
550 void *retval;
551
552 main_thread = pthread_self();
553 num_cpus_to_pin = 0;
554
555 if (sysconf(_SC_NPROCESSORS_ONLN) == -1) {
556 perror("sysconf(_SC_NPROCESSORS_ONLN)");
557 exit(1);
558 }
559
560 if (getuid() != 0)
561 ksft_exit_skip("Not running as root, but almost all tests "
562 "require root in order to modify\nsystem settings. "
563 "Exiting.\n");
564
565 cpus_online = MIN(MAX_CPUS, sysconf(_SC_NPROCESSORS_ONLN));
566 cpu_set = CPU_ALLOC(cpus_online);
567 if (cpu_set == NULL) {
568 perror("CPU_ALLOC()");
569 exit(1);
570 }
571 cpu_set_size = CPU_ALLOC_SIZE(cpus_online);
572 CPU_ZERO_S(cpu_set_size, cpu_set);
573
574 popt_context = poptGetContext(NULL, argc, (const char **)argv,
575 options, 0);
576
577 while ((rc = poptGetNextOpt(popt_context)) > 0) {
578 switch (rc) {
579 case 'c':
580 continuous_mode = 1;
581 option = cpu_option_string;
582 do {
583 next_option = strchr(option, ',');
584 if (next_option)
585 *next_option = '\0';
586 cpu = atoi(option);
587 if (cpu >= cpus_online)
588 fprintf(stderr, "CPU %d exceeds "
589 "cpus online, ignoring.\n",
590 cpu);
591 else
592 cpus_to_pin[num_cpus_to_pin++] = cpu;
593 if (next_option)
594 option = ++next_option;
595 } while (next_option && num_cpus_to_pin < MAX_CPUS);
596 /* Double check that they didn't give us the same CPU
597 * more than once */
598 for (cpu = 0; cpu < num_cpus_to_pin; cpu++) {
599 if (CPU_ISSET_S(cpus_to_pin[cpu], cpu_set_size,
600 cpu_set)) {
601 fprintf(stderr, "Any given CPU may "
602 "only be given once.\n");
603 goto err_code;
604 } else
605 CPU_SET_S(cpus_to_pin[cpu],
606 cpu_set_size, cpu_set);
607 }
608 break;
609 case 'p':
610 /*
611 * Although we can create a msg queue with a
612 * non-absolute path name, unlink will fail. So,
613 * if the name doesn't start with a /, add one
614 * when we save it.
615 */
616 option = queue_path;
617 if (*option != '/') {
618 queue_path = malloc(strlen(option) + 2);
619 if (!queue_path) {
620 perror("malloc()");
621 goto err_code;
622 }
623 queue_path[0] = '/';
624 queue_path[1] = 0;
625 strcat(queue_path, option);
626 free(option);
627 }
628 break;
629 }
630 }
631
632 if (continuous_mode && num_cpus_to_pin == 0) {
633 fprintf(stderr, "Must pass at least one CPU to continuous "
634 "mode.\n");
635 poptPrintUsage(popt_context, stderr, 0);
636 goto err_code;
637 } else if (!continuous_mode) {
638 num_cpus_to_pin = 1;
639 cpus_to_pin[0] = cpus_online - 1;
640 }
641
642 max_msgs = fopen(MAX_MSGS, "r+");
643 max_msgsize = fopen(MAX_MSGSIZE, "r+");
644 if (!max_msgs)
645 shutdown(2, "Failed to open msg_max", __LINE__);
646 if (!max_msgsize)
647 shutdown(2, "Failed to open msgsize_max", __LINE__);
648
649 /* Load up the current system values for everything we can */
650 getr(RLIMIT_MSGQUEUE, &saved_limits);
651 cur_limits = saved_limits;
652 saved_max_msgs = cur_max_msgs = get(max_msgs);
653 saved_max_msgsize = cur_max_msgsize = get(max_msgsize);
654 errno = 0;
655 cur_nice = getpriority(PRIO_PROCESS, 0);
656 if (errno)
657 shutdown(2, "getpriority()", __LINE__);
658
659 /* Tell the user our initial state */
660 printf("\nInitial system state:\n");
661 printf("\tUsing queue path:\t\t\t%s\n", queue_path);
662 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
663 (long) saved_limits.rlim_cur);
664 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
665 (long) saved_limits.rlim_max);
666 printf("\tMaximum Message Size:\t\t\t%d\n", saved_max_msgsize);
667 printf("\tMaximum Queue Size:\t\t\t%d\n", saved_max_msgs);
668 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
669 printf("\n");
670
671 increase_limits();
672
673 printf("Adjusted system state for testing:\n");
674 if (cur_limits.rlim_cur == RLIM_INFINITY) {
675 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t(unlimited)\n");
676 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t(unlimited)\n");
677 } else {
678 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
679 (long) cur_limits.rlim_cur);
680 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
681 (long) cur_limits.rlim_max);
682 }
683 printf("\tMaximum Message Size:\t\t\t%d\n", cur_max_msgsize);
684 printf("\tMaximum Queue Size:\t\t\t%d\n", cur_max_msgs);
685 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
686 printf("\tContinuous mode:\t\t\t(%s)\n", continuous_mode ?
687 (continuous_mode_fake ? "fake mode" : "enabled") :
688 "disabled");
689 printf("\tCPUs to pin:\t\t\t\t%d", cpus_to_pin[0]);
690 for (cpu = 1; cpu < num_cpus_to_pin; cpu++)
691 printf(",%d", cpus_to_pin[cpu]);
692 printf("\n");
693
694 sa.sa_sigaction = sig_action_SIGUSR1;
695 sigemptyset(&sa.sa_mask);
696 sigaddset(&sa.sa_mask, SIGHUP);
697 sigaddset(&sa.sa_mask, SIGINT);
698 sigaddset(&sa.sa_mask, SIGQUIT);
699 sigaddset(&sa.sa_mask, SIGTERM);
700 sa.sa_flags = SA_SIGINFO;
701 if (sigaction(SIGUSR1, &sa, NULL) == -1)
702 shutdown(1, "sigaction(SIGUSR1)", __LINE__);
703 sa.sa_sigaction = sig_action;
704 if (sigaction(SIGHUP, &sa, NULL) == -1)
705 shutdown(1, "sigaction(SIGHUP)", __LINE__);
706 if (sigaction(SIGINT, &sa, NULL) == -1)
707 shutdown(1, "sigaction(SIGINT)", __LINE__);
708 if (sigaction(SIGQUIT, &sa, NULL) == -1)
709 shutdown(1, "sigaction(SIGQUIT)", __LINE__);
710 if (sigaction(SIGTERM, &sa, NULL) == -1)
711 shutdown(1, "sigaction(SIGTERM)", __LINE__);
712
713 if (!continuous_mode_fake) {
714 attr.mq_flags = O_NONBLOCK;
715 attr.mq_maxmsg = cur_max_msgs;
716 attr.mq_msgsize = MSG_SIZE;
717 open_queue(&attr);
718 }
719 for (i = 0; i < num_cpus_to_pin; i++) {
720 pthread_attr_t thread_attr;
721 void *thread_func;
722
723 if (continuous_mode_fake)
724 thread_func = &fake_cont_thread;
725 else if (continuous_mode)
726 thread_func = &cont_thread;
727 else
728 thread_func = &perf_test_thread;
729
730 CPU_ZERO_S(cpu_set_size, cpu_set);
731 CPU_SET_S(cpus_to_pin[i], cpu_set_size, cpu_set);
732 pthread_attr_init(&thread_attr);
733 pthread_attr_setaffinity_np(&thread_attr, cpu_set_size,
734 cpu_set);
735 if (pthread_create(&cpu_threads[i], &thread_attr, thread_func,
736 NULL))
737 shutdown(1, "pthread_create()", __LINE__);
738 pthread_attr_destroy(&thread_attr);
739 }
740
741 if (!continuous_mode) {
742 pthread_join(cpu_threads[0], &retval);
743 shutdown((long)retval, "perf_test_thread()", __LINE__);
744 } else {
745 while (1)
746 sleep(1);
747 }
748 shutdown(0, "", 0);
749
750err_code:
751 CPU_FREE(cpu_set);
752 exit(1);
753
754}
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
41static char *usage =
42"Usage:\n"
43" %s [-c #[,#..] -f] path\n"
44"\n"
45" -c # Skip most tests and go straight to a high queue depth test\n"
46" and then run that test continuously (useful for running at\n"
47" the same time as some other workload to see how much the\n"
48" cache thrashing caused by adding messages to a very deep\n"
49" queue impacts the performance of other programs). The number\n"
50" indicates which CPU core we should bind the process to during\n"
51" the run. If you have more than one physical CPU, then you\n"
52" will need one copy per physical CPU package, and you should\n"
53" specify the CPU cores to pin ourself to via a comma separated\n"
54" list of CPU values.\n"
55" -f Only usable with continuous mode. Pin ourself to the CPUs\n"
56" as requested, then instead of looping doing a high mq\n"
57" workload, just busy loop. This will allow us to lock up a\n"
58" single CPU just like we normally would, but without actually\n"
59" thrashing the CPU cache. This is to make it easier to get\n"
60" comparable numbers from some other workload running on the\n"
61" other CPUs. One set of numbers with # CPUs locked up running\n"
62" an mq workload, and another set of numbers with those same\n"
63" CPUs locked away from the test workload, but not doing\n"
64" anything to trash the cache like the mq workload might.\n"
65" path Path name of the message queue to create\n"
66"\n"
67" Note: this program must be run as root in order to enable all tests\n"
68"\n";
69
70char *MAX_MSGS = "/proc/sys/fs/mqueue/msg_max";
71char *MAX_MSGSIZE = "/proc/sys/fs/mqueue/msgsize_max";
72
73#define min(a, b) ((a) < (b) ? (a) : (b))
74#define MAX_CPUS 64
75char *cpu_option_string;
76int cpus_to_pin[MAX_CPUS];
77int num_cpus_to_pin;
78pthread_t cpu_threads[MAX_CPUS];
79pthread_t main_thread;
80cpu_set_t *cpu_set;
81int cpu_set_size;
82int cpus_online;
83
84#define MSG_SIZE 16
85#define TEST1_LOOPS 10000000
86#define TEST2_LOOPS 100000
87int continuous_mode;
88int continuous_mode_fake;
89
90struct rlimit saved_limits, cur_limits;
91int saved_max_msgs, saved_max_msgsize;
92int cur_max_msgs, cur_max_msgsize;
93FILE *max_msgs, *max_msgsize;
94int cur_nice;
95char *queue_path = "/mq_perf_tests";
96mqd_t queue = -1;
97struct mq_attr result;
98int mq_prio_max;
99
100const struct poptOption options[] = {
101 {
102 .longName = "continuous",
103 .shortName = 'c',
104 .argInfo = POPT_ARG_STRING,
105 .arg = &cpu_option_string,
106 .val = 'c',
107 .descrip = "Run continuous tests at a high queue depth in "
108 "order to test the effects of cache thrashing on "
109 "other tasks on the system. This test is intended "
110 "to be run on one core of each physical CPU while "
111 "some other CPU intensive task is run on all the other "
112 "cores of that same physical CPU and the other task "
113 "is timed. It is assumed that the process of adding "
114 "messages to the message queue in a tight loop will "
115 "impact that other task to some degree. Once the "
116 "tests are performed in this way, you should then "
117 "re-run the tests using fake mode in order to check "
118 "the difference in time required to perform the CPU "
119 "intensive task",
120 .argDescrip = "cpu[,cpu]",
121 },
122 {
123 .longName = "fake",
124 .shortName = 'f',
125 .argInfo = POPT_ARG_NONE,
126 .arg = &continuous_mode_fake,
127 .val = 0,
128 .descrip = "Tie up the CPUs that we would normally tie up in"
129 "continuous mode, but don't actually do any mq stuff, "
130 "just keep the CPU busy so it can't be used to process "
131 "system level tasks as this would free up resources on "
132 "the other CPU cores and skew the comparison between "
133 "the no-mqueue work and mqueue work tests",
134 .argDescrip = NULL,
135 },
136 {
137 .longName = "path",
138 .shortName = 'p',
139 .argInfo = POPT_ARG_STRING | POPT_ARGFLAG_SHOW_DEFAULT,
140 .arg = &queue_path,
141 .val = 'p',
142 .descrip = "The name of the path to use in the mqueue "
143 "filesystem for our tests",
144 .argDescrip = "pathname",
145 },
146 POPT_AUTOHELP
147 POPT_TABLEEND
148};
149
150static inline void __set(FILE *stream, int value, char *err_msg);
151void shutdown(int exit_val, char *err_cause, int line_no);
152void sig_action_SIGUSR1(int signum, siginfo_t *info, void *context);
153void sig_action(int signum, siginfo_t *info, void *context);
154static inline int get(FILE *stream);
155static inline void set(FILE *stream, int value);
156static inline int try_set(FILE *stream, int value);
157static inline void getr(int type, struct rlimit *rlim);
158static inline void setr(int type, struct rlimit *rlim);
159static inline void open_queue(struct mq_attr *attr);
160void increase_limits(void);
161
162static inline void __set(FILE *stream, int value, char *err_msg)
163{
164 rewind(stream);
165 if (fprintf(stream, "%d", value) < 0)
166 perror(err_msg);
167}
168
169
170void shutdown(int exit_val, char *err_cause, int line_no)
171{
172 static int in_shutdown = 0;
173 int errno_at_shutdown = errno;
174 int i;
175
176 /* In case we get called by multiple threads or from an sighandler */
177 if (in_shutdown++)
178 return;
179
180 for (i = 0; i < num_cpus_to_pin; i++)
181 if (cpu_threads[i]) {
182 pthread_kill(cpu_threads[i], SIGUSR1);
183 pthread_join(cpu_threads[i], NULL);
184 }
185
186 if (queue != -1)
187 if (mq_close(queue))
188 perror("mq_close() during shutdown");
189 if (queue_path)
190 /*
191 * Be silent if this fails, if we cleaned up already it's
192 * expected to fail
193 */
194 mq_unlink(queue_path);
195 if (saved_max_msgs)
196 __set(max_msgs, saved_max_msgs,
197 "failed to restore saved_max_msgs");
198 if (saved_max_msgsize)
199 __set(max_msgsize, saved_max_msgsize,
200 "failed to restore saved_max_msgsize");
201 if (exit_val)
202 error(exit_val, errno_at_shutdown, "%s at %d",
203 err_cause, line_no);
204 exit(0);
205}
206
207void sig_action_SIGUSR1(int signum, siginfo_t *info, void *context)
208{
209 if (pthread_self() != main_thread)
210 pthread_exit(0);
211 else {
212 fprintf(stderr, "Caught signal %d in SIGUSR1 handler, "
213 "exiting\n", signum);
214 shutdown(0, "", 0);
215 fprintf(stderr, "\n\nReturned from shutdown?!?!\n\n");
216 exit(0);
217 }
218}
219
220void sig_action(int signum, siginfo_t *info, void *context)
221{
222 if (pthread_self() != main_thread)
223 pthread_kill(main_thread, signum);
224 else {
225 fprintf(stderr, "Caught signal %d, exiting\n", signum);
226 shutdown(0, "", 0);
227 fprintf(stderr, "\n\nReturned from shutdown?!?!\n\n");
228 exit(0);
229 }
230}
231
232static inline int get(FILE *stream)
233{
234 int value;
235 rewind(stream);
236 if (fscanf(stream, "%d", &value) != 1)
237 shutdown(4, "Error reading /proc entry", __LINE__);
238 return value;
239}
240
241static inline void set(FILE *stream, int value)
242{
243 int new_value;
244
245 rewind(stream);
246 if (fprintf(stream, "%d", value) < 0)
247 return shutdown(5, "Failed writing to /proc file", __LINE__);
248 new_value = get(stream);
249 if (new_value != value)
250 return shutdown(5, "We didn't get what we wrote to /proc back",
251 __LINE__);
252}
253
254static inline int try_set(FILE *stream, int value)
255{
256 int new_value;
257
258 rewind(stream);
259 fprintf(stream, "%d", value);
260 new_value = get(stream);
261 return new_value == value;
262}
263
264static inline void getr(int type, struct rlimit *rlim)
265{
266 if (getrlimit(type, rlim))
267 shutdown(6, "getrlimit()", __LINE__);
268}
269
270static inline void setr(int type, struct rlimit *rlim)
271{
272 if (setrlimit(type, rlim))
273 shutdown(7, "setrlimit()", __LINE__);
274}
275
276/**
277 * open_queue - open the global queue for testing
278 * @attr - An attr struct specifying the desired queue traits
279 * @result - An attr struct that lists the actual traits the queue has
280 *
281 * This open is not allowed to fail, failure will result in an orderly
282 * shutdown of the program. The global queue_path is used to set what
283 * queue to open, the queue descriptor is saved in the global queue
284 * variable.
285 */
286static inline void open_queue(struct mq_attr *attr)
287{
288 int flags = O_RDWR | O_EXCL | O_CREAT | O_NONBLOCK;
289 int perms = DEFFILEMODE;
290
291 queue = mq_open(queue_path, flags, perms, attr);
292 if (queue == -1)
293 shutdown(1, "mq_open()", __LINE__);
294 if (mq_getattr(queue, &result))
295 shutdown(1, "mq_getattr()", __LINE__);
296 printf("\n\tQueue %s created:\n", queue_path);
297 printf("\t\tmq_flags:\t\t\t%s\n", result.mq_flags & O_NONBLOCK ?
298 "O_NONBLOCK" : "(null)");
299 printf("\t\tmq_maxmsg:\t\t\t%d\n", result.mq_maxmsg);
300 printf("\t\tmq_msgsize:\t\t\t%d\n", result.mq_msgsize);
301 printf("\t\tmq_curmsgs:\t\t\t%d\n", result.mq_curmsgs);
302}
303
304void *fake_cont_thread(void *arg)
305{
306 int i;
307
308 for (i = 0; i < num_cpus_to_pin; i++)
309 if (cpu_threads[i] == pthread_self())
310 break;
311 printf("\tStarted fake continuous mode thread %d on CPU %d\n", i,
312 cpus_to_pin[i]);
313 while (1)
314 ;
315}
316
317void *cont_thread(void *arg)
318{
319 char buff[MSG_SIZE];
320 int i, priority;
321
322 for (i = 0; i < num_cpus_to_pin; i++)
323 if (cpu_threads[i] == pthread_self())
324 break;
325 printf("\tStarted continuous mode thread %d on CPU %d\n", i,
326 cpus_to_pin[i]);
327 while (1) {
328 while (mq_send(queue, buff, sizeof(buff), 0) == 0)
329 ;
330 mq_receive(queue, buff, sizeof(buff), &priority);
331 }
332}
333
334#define drain_queue() \
335 while (mq_receive(queue, buff, MSG_SIZE, &prio_in) == MSG_SIZE)
336
337#define do_untimed_send() \
338 do { \
339 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
340 shutdown(3, "Test send failure", __LINE__); \
341 } while (0)
342
343#define do_send_recv() \
344 do { \
345 clock_gettime(clock, &start); \
346 if (mq_send(queue, buff, MSG_SIZE, prio_out)) \
347 shutdown(3, "Test send failure", __LINE__); \
348 clock_gettime(clock, &middle); \
349 if (mq_receive(queue, buff, MSG_SIZE, &prio_in) != MSG_SIZE) \
350 shutdown(3, "Test receive failure", __LINE__); \
351 clock_gettime(clock, &end); \
352 nsec = ((middle.tv_sec - start.tv_sec) * 1000000000) + \
353 (middle.tv_nsec - start.tv_nsec); \
354 send_total.tv_nsec += nsec; \
355 if (send_total.tv_nsec >= 1000000000) { \
356 send_total.tv_sec++; \
357 send_total.tv_nsec -= 1000000000; \
358 } \
359 nsec = ((end.tv_sec - middle.tv_sec) * 1000000000) + \
360 (end.tv_nsec - middle.tv_nsec); \
361 recv_total.tv_nsec += nsec; \
362 if (recv_total.tv_nsec >= 1000000000) { \
363 recv_total.tv_sec++; \
364 recv_total.tv_nsec -= 1000000000; \
365 } \
366 } while (0)
367
368struct test {
369 char *desc;
370 void (*func)(int *);
371};
372
373void const_prio(int *prio)
374{
375 return;
376}
377
378void inc_prio(int *prio)
379{
380 if (++*prio == mq_prio_max)
381 *prio = 0;
382}
383
384void dec_prio(int *prio)
385{
386 if (--*prio < 0)
387 *prio = mq_prio_max - 1;
388}
389
390void random_prio(int *prio)
391{
392 *prio = random() % mq_prio_max;
393}
394
395struct test test2[] = {
396 {"\n\tTest #2a: Time send/recv message, queue full, constant prio\n",
397 const_prio},
398 {"\n\tTest #2b: Time send/recv message, queue full, increasing prio\n",
399 inc_prio},
400 {"\n\tTest #2c: Time send/recv message, queue full, decreasing prio\n",
401 dec_prio},
402 {"\n\tTest #2d: Time send/recv message, queue full, random prio\n",
403 random_prio},
404 {NULL, NULL}
405};
406
407/**
408 * Tests to perform (all done with MSG_SIZE messages):
409 *
410 * 1) Time to add/remove message with 0 messages on queue
411 * 1a) with constant prio
412 * 2) Time to add/remove message when queue close to capacity:
413 * 2a) with constant prio
414 * 2b) with increasing prio
415 * 2c) with decreasing prio
416 * 2d) with random prio
417 * 3) Test limits of priorities honored (double check _SC_MQ_PRIO_MAX)
418 */
419void *perf_test_thread(void *arg)
420{
421 char buff[MSG_SIZE];
422 int prio_out, prio_in;
423 int i;
424 clockid_t clock;
425 pthread_t *t;
426 struct timespec res, start, middle, end, send_total, recv_total;
427 unsigned long long nsec;
428 struct test *cur_test;
429
430 t = &cpu_threads[0];
431 printf("\n\tStarted mqueue performance test thread on CPU %d\n",
432 cpus_to_pin[0]);
433 mq_prio_max = sysconf(_SC_MQ_PRIO_MAX);
434 if (mq_prio_max == -1)
435 shutdown(2, "sysconf(_SC_MQ_PRIO_MAX)", __LINE__);
436 if (pthread_getcpuclockid(cpu_threads[0], &clock) != 0)
437 shutdown(2, "pthread_getcpuclockid", __LINE__);
438
439 if (clock_getres(clock, &res))
440 shutdown(2, "clock_getres()", __LINE__);
441
442 printf("\t\tMax priorities:\t\t\t%d\n", mq_prio_max);
443 printf("\t\tClock resolution:\t\t%d nsec%s\n", res.tv_nsec,
444 res.tv_nsec > 1 ? "s" : "");
445
446
447
448 printf("\n\tTest #1: Time send/recv message, queue empty\n");
449 printf("\t\t(%d iterations)\n", TEST1_LOOPS);
450 prio_out = 0;
451 send_total.tv_sec = 0;
452 send_total.tv_nsec = 0;
453 recv_total.tv_sec = 0;
454 recv_total.tv_nsec = 0;
455 for (i = 0; i < TEST1_LOOPS; i++)
456 do_send_recv();
457 printf("\t\tSend msg:\t\t\t%d.%ds total time\n",
458 send_total.tv_sec, send_total.tv_nsec);
459 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
460 send_total.tv_nsec) / TEST1_LOOPS;
461 printf("\t\t\t\t\t\t%d nsec/msg\n", nsec);
462 printf("\t\tRecv msg:\t\t\t%d.%ds total time\n",
463 recv_total.tv_sec, recv_total.tv_nsec);
464 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
465 recv_total.tv_nsec) / TEST1_LOOPS;
466 printf("\t\t\t\t\t\t%d nsec/msg\n", nsec);
467
468
469 for (cur_test = test2; cur_test->desc != NULL; cur_test++) {
470 printf(cur_test->desc);
471 printf("\t\t(%d iterations)\n", TEST2_LOOPS);
472 prio_out = 0;
473 send_total.tv_sec = 0;
474 send_total.tv_nsec = 0;
475 recv_total.tv_sec = 0;
476 recv_total.tv_nsec = 0;
477 printf("\t\tFilling queue...");
478 fflush(stdout);
479 clock_gettime(clock, &start);
480 for (i = 0; i < result.mq_maxmsg - 1; i++) {
481 do_untimed_send();
482 cur_test->func(&prio_out);
483 }
484 clock_gettime(clock, &end);
485 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
486 1000000000) + (end.tv_nsec - start.tv_nsec);
487 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
488 nsec % 1000000000);
489 printf("\t\tTesting...");
490 fflush(stdout);
491 for (i = 0; i < TEST2_LOOPS; i++) {
492 do_send_recv();
493 cur_test->func(&prio_out);
494 }
495 printf("done.\n");
496 printf("\t\tSend msg:\t\t\t%d.%ds total time\n",
497 send_total.tv_sec, send_total.tv_nsec);
498 nsec = ((unsigned long long)send_total.tv_sec * 1000000000 +
499 send_total.tv_nsec) / TEST2_LOOPS;
500 printf("\t\t\t\t\t\t%d nsec/msg\n", nsec);
501 printf("\t\tRecv msg:\t\t\t%d.%ds total time\n",
502 recv_total.tv_sec, recv_total.tv_nsec);
503 nsec = ((unsigned long long)recv_total.tv_sec * 1000000000 +
504 recv_total.tv_nsec) / TEST2_LOOPS;
505 printf("\t\t\t\t\t\t%d nsec/msg\n", nsec);
506 printf("\t\tDraining queue...");
507 fflush(stdout);
508 clock_gettime(clock, &start);
509 drain_queue();
510 clock_gettime(clock, &end);
511 nsec = ((unsigned long long)(end.tv_sec - start.tv_sec) *
512 1000000000) + (end.tv_nsec - start.tv_nsec);
513 printf("done.\t\t%lld.%llds\n", nsec / 1000000000,
514 nsec % 1000000000);
515 }
516 return 0;
517}
518
519void increase_limits(void)
520{
521 cur_limits.rlim_cur = RLIM_INFINITY;
522 cur_limits.rlim_max = RLIM_INFINITY;
523 setr(RLIMIT_MSGQUEUE, &cur_limits);
524 while (try_set(max_msgs, cur_max_msgs += 10))
525 ;
526 cur_max_msgs = get(max_msgs);
527 while (try_set(max_msgsize, cur_max_msgsize += 1024))
528 ;
529 cur_max_msgsize = get(max_msgsize);
530 if (setpriority(PRIO_PROCESS, 0, -20) != 0)
531 shutdown(2, "setpriority()", __LINE__);
532 cur_nice = -20;
533}
534
535int main(int argc, char *argv[])
536{
537 struct mq_attr attr;
538 char *option, *next_option;
539 int i, cpu;
540 struct sigaction sa;
541 poptContext popt_context;
542 char rc;
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%d\n", saved_limits.rlim_cur);
657 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%d\n", saved_limits.rlim_max);
658 printf("\tMaximum Message Size:\t\t\t%d\n", saved_max_msgsize);
659 printf("\tMaximum Queue Size:\t\t\t%d\n", saved_max_msgs);
660 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
661 printf("\n");
662
663 increase_limits();
664
665 printf("Adjusted system state for testing:\n");
666 if (cur_limits.rlim_cur == RLIM_INFINITY) {
667 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t(unlimited)\n");
668 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t(unlimited)\n");
669 } else {
670 printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%d\n",
671 cur_limits.rlim_cur);
672 printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%d\n",
673 cur_limits.rlim_max);
674 }
675 printf("\tMaximum Message Size:\t\t\t%d\n", cur_max_msgsize);
676 printf("\tMaximum Queue Size:\t\t\t%d\n", cur_max_msgs);
677 printf("\tNice value:\t\t\t\t%d\n", cur_nice);
678 printf("\tContinuous mode:\t\t\t(%s)\n", continuous_mode ?
679 (continuous_mode_fake ? "fake mode" : "enabled") :
680 "disabled");
681 printf("\tCPUs to pin:\t\t\t\t%d", cpus_to_pin[0]);
682 for (cpu = 1; cpu < num_cpus_to_pin; cpu++)
683 printf(",%d", cpus_to_pin[cpu]);
684 printf("\n");
685
686 sa.sa_sigaction = sig_action_SIGUSR1;
687 sigemptyset(&sa.sa_mask);
688 sigaddset(&sa.sa_mask, SIGHUP);
689 sigaddset(&sa.sa_mask, SIGINT);
690 sigaddset(&sa.sa_mask, SIGQUIT);
691 sigaddset(&sa.sa_mask, SIGTERM);
692 sa.sa_flags = SA_SIGINFO;
693 if (sigaction(SIGUSR1, &sa, NULL) == -1)
694 shutdown(1, "sigaction(SIGUSR1)", __LINE__);
695 sa.sa_sigaction = sig_action;
696 if (sigaction(SIGHUP, &sa, NULL) == -1)
697 shutdown(1, "sigaction(SIGHUP)", __LINE__);
698 if (sigaction(SIGINT, &sa, NULL) == -1)
699 shutdown(1, "sigaction(SIGINT)", __LINE__);
700 if (sigaction(SIGQUIT, &sa, NULL) == -1)
701 shutdown(1, "sigaction(SIGQUIT)", __LINE__);
702 if (sigaction(SIGTERM, &sa, NULL) == -1)
703 shutdown(1, "sigaction(SIGTERM)", __LINE__);
704
705 if (!continuous_mode_fake) {
706 attr.mq_flags = O_NONBLOCK;
707 attr.mq_maxmsg = cur_max_msgs;
708 attr.mq_msgsize = MSG_SIZE;
709 open_queue(&attr);
710 }
711 for (i = 0; i < num_cpus_to_pin; i++) {
712 pthread_attr_t thread_attr;
713 void *thread_func;
714
715 if (continuous_mode_fake)
716 thread_func = &fake_cont_thread;
717 else if (continuous_mode)
718 thread_func = &cont_thread;
719 else
720 thread_func = &perf_test_thread;
721
722 CPU_ZERO_S(cpu_set_size, cpu_set);
723 CPU_SET_S(cpus_to_pin[i], cpu_set_size, cpu_set);
724 pthread_attr_init(&thread_attr);
725 pthread_attr_setaffinity_np(&thread_attr, cpu_set_size,
726 cpu_set);
727 if (pthread_create(&cpu_threads[i], &thread_attr, thread_func,
728 NULL))
729 shutdown(1, "pthread_create()", __LINE__);
730 pthread_attr_destroy(&thread_attr);
731 }
732
733 if (!continuous_mode) {
734 pthread_join(cpu_threads[0], &retval);
735 shutdown((long)retval, "perf_test_thread()", __LINE__);
736 } else {
737 while (1)
738 sleep(1);
739 }
740 shutdown(0, "", 0);
741}