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 
41 static 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 
70 char *MAX_MSGS = "/proc/sys/fs/mqueue/msg_max";
71 char *MAX_MSGSIZE = "/proc/sys/fs/mqueue/msgsize_max";
72 
73 #define min(a, b) ((a) < (b) ? (a) : (b))
74 #define MAX_CPUS 64
75 char *cpu_option_string;
76 int cpus_to_pin[MAX_CPUS];
77 int num_cpus_to_pin;
78 pthread_t cpu_threads[MAX_CPUS];
79 pthread_t main_thread;
80 cpu_set_t *cpu_set;
81 int cpu_set_size;
82 int cpus_online;
83 
84 #define MSG_SIZE 16
85 #define TEST1_LOOPS 10000000
86 #define TEST2_LOOPS 100000
87 int continuous_mode;
88 int continuous_mode_fake;
89 
90 struct rlimit saved_limits, cur_limits;
91 int saved_max_msgs, saved_max_msgsize;
92 int cur_max_msgs, cur_max_msgsize;
93 FILE *max_msgs, *max_msgsize;
94 int cur_nice;
95 char *queue_path = "/mq_perf_tests";
96 mqd_t queue = -1;
97 struct mq_attr result;
98 int mq_prio_max;
99 
100 const 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 
150 static inline void __set(FILE *stream, int value, char *err_msg);
151 void shutdown(int exit_val, char *err_cause, int line_no);
152 void sig_action_SIGUSR1(int signum, siginfo_t *info, void *context);
153 void sig_action(int signum, siginfo_t *info, void *context);
154 static inline int get(FILE *stream);
155 static inline void set(FILE *stream, int value);
156 static inline int try_set(FILE *stream, int value);
157 static inline void getr(int type, struct rlimit *rlim);
158 static inline void setr(int type, struct rlimit *rlim);
159 static inline void open_queue(struct mq_attr *attr);
160 void increase_limits(void);
161 
162 static 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 
170 void 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 
207 void 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 
220 void 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 
232 static 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 
241 static 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 
254 static 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 
264 static inline void getr(int type, struct rlimit *rlim)
265 {
266 	if (getrlimit(type, rlim))
267 		shutdown(6, "getrlimit()", __LINE__);
268 }
269 
270 static 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  */
286 static 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%lu\n", result.mq_maxmsg);
300 	printf("\t\tmq_msgsize:\t\t\t%lu\n", result.mq_msgsize);
301 	printf("\t\tmq_curmsgs:\t\t\t%lu\n", result.mq_curmsgs);
302 }
303 
304 void *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 
317 void *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 
368 struct test {
369 	char *desc;
370 	void (*func)(int *);
371 };
372 
373 void const_prio(int *prio)
374 {
375 	return;
376 }
377 
378 void inc_prio(int *prio)
379 {
380 	if (++*prio == mq_prio_max)
381 		*prio = 0;
382 }
383 
384 void dec_prio(int *prio)
385 {
386 	if (--*prio < 0)
387 		*prio = mq_prio_max - 1;
388 }
389 
390 void random_prio(int *prio)
391 {
392 	*prio = random() % mq_prio_max;
393 }
394 
395 struct 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  */
419 void *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%lu 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%ld.%lus 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%lld nsec/msg\n", nsec);
462 	printf("\t\tRecv msg:\t\t\t%ld.%lus 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%lld nsec/msg\n", nsec);
467 
468 
469 	for (cur_test = test2; cur_test->desc != NULL; cur_test++) {
470 		printf("%s:\n", 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%ld.%lus 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%lld nsec/msg\n", nsec);
501 		printf("\t\tRecv msg:\t\t\t%ld.%lus 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%lld 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 
519 void 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 
535 int main(int argc, char *argv[])
536 {
537 	struct mq_attr attr;
538 	char *option, *next_option;
539 	int i, cpu, rc;
540 	struct sigaction sa;
541 	poptContext popt_context;
542 	void *retval;
543 
544 	main_thread = pthread_self();
545 	num_cpus_to_pin = 0;
546 
547 	if (sysconf(_SC_NPROCESSORS_ONLN) == -1) {
548 		perror("sysconf(_SC_NPROCESSORS_ONLN)");
549 		exit(1);
550 	}
551 	cpus_online = min(MAX_CPUS, sysconf(_SC_NPROCESSORS_ONLN));
552 	cpu_set = CPU_ALLOC(cpus_online);
553 	if (cpu_set == NULL) {
554 		perror("CPU_ALLOC()");
555 		exit(1);
556 	}
557 	cpu_set_size = CPU_ALLOC_SIZE(cpus_online);
558 	CPU_ZERO_S(cpu_set_size, cpu_set);
559 
560 	popt_context = poptGetContext(NULL, argc, (const char **)argv,
561 				      options, 0);
562 
563 	while ((rc = poptGetNextOpt(popt_context)) > 0) {
564 		switch (rc) {
565 		case 'c':
566 			continuous_mode = 1;
567 			option = cpu_option_string;
568 			do {
569 				next_option = strchr(option, ',');
570 				if (next_option)
571 					*next_option = '\0';
572 				cpu = atoi(option);
573 				if (cpu >= cpus_online)
574 					fprintf(stderr, "CPU %d exceeds "
575 						"cpus online, ignoring.\n",
576 						cpu);
577 				else
578 					cpus_to_pin[num_cpus_to_pin++] = cpu;
579 				if (next_option)
580 					option = ++next_option;
581 			} while (next_option && num_cpus_to_pin < MAX_CPUS);
582 			/* Double check that they didn't give us the same CPU
583 			 * more than once */
584 			for (cpu = 0; cpu < num_cpus_to_pin; cpu++) {
585 				if (CPU_ISSET_S(cpus_to_pin[cpu], cpu_set_size,
586 						cpu_set)) {
587 					fprintf(stderr, "Any given CPU may "
588 						"only be given once.\n");
589 					exit(1);
590 				} else
591 					CPU_SET_S(cpus_to_pin[cpu],
592 						  cpu_set_size, cpu_set);
593 			}
594 			break;
595 		case 'p':
596 			/*
597 			 * Although we can create a msg queue with a
598 			 * non-absolute path name, unlink will fail.  So,
599 			 * if the name doesn't start with a /, add one
600 			 * when we save it.
601 			 */
602 			option = queue_path;
603 			if (*option != '/') {
604 				queue_path = malloc(strlen(option) + 2);
605 				if (!queue_path) {
606 					perror("malloc()");
607 					exit(1);
608 				}
609 				queue_path[0] = '/';
610 				queue_path[1] = 0;
611 				strcat(queue_path, option);
612 				free(option);
613 			}
614 			break;
615 		}
616 	}
617 
618 	if (continuous_mode && num_cpus_to_pin == 0) {
619 		fprintf(stderr, "Must pass at least one CPU to continuous "
620 			"mode.\n");
621 		poptPrintUsage(popt_context, stderr, 0);
622 		exit(1);
623 	} else if (!continuous_mode) {
624 		num_cpus_to_pin = 1;
625 		cpus_to_pin[0] = cpus_online - 1;
626 	}
627 
628 	if (getuid() != 0) {
629 		fprintf(stderr, "Not running as root, but almost all tests "
630 			"require root in order to modify\nsystem settings.  "
631 			"Exiting.\n");
632 		exit(1);
633 	}
634 
635 	max_msgs = fopen(MAX_MSGS, "r+");
636 	max_msgsize = fopen(MAX_MSGSIZE, "r+");
637 	if (!max_msgs)
638 		shutdown(2, "Failed to open msg_max", __LINE__);
639 	if (!max_msgsize)
640 		shutdown(2, "Failed to open msgsize_max", __LINE__);
641 
642 	/* Load up the current system values for everything we can */
643 	getr(RLIMIT_MSGQUEUE, &saved_limits);
644 	cur_limits = saved_limits;
645 	saved_max_msgs = cur_max_msgs = get(max_msgs);
646 	saved_max_msgsize = cur_max_msgsize = get(max_msgsize);
647 	errno = 0;
648 	cur_nice = getpriority(PRIO_PROCESS, 0);
649 	if (errno)
650 		shutdown(2, "getpriority()", __LINE__);
651 
652 	/* Tell the user our initial state */
653 	printf("\nInitial system state:\n");
654 	printf("\tUsing queue path:\t\t\t%s\n", queue_path);
655 	printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
656 		(long) saved_limits.rlim_cur);
657 	printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
658 		(long) saved_limits.rlim_max);
659 	printf("\tMaximum Message Size:\t\t\t%d\n", saved_max_msgsize);
660 	printf("\tMaximum Queue Size:\t\t\t%d\n", saved_max_msgs);
661 	printf("\tNice value:\t\t\t\t%d\n", cur_nice);
662 	printf("\n");
663 
664 	increase_limits();
665 
666 	printf("Adjusted system state for testing:\n");
667 	if (cur_limits.rlim_cur == RLIM_INFINITY) {
668 		printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t(unlimited)\n");
669 		printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t(unlimited)\n");
670 	} else {
671 		printf("\tRLIMIT_MSGQUEUE(soft):\t\t\t%ld\n",
672 		       (long) cur_limits.rlim_cur);
673 		printf("\tRLIMIT_MSGQUEUE(hard):\t\t\t%ld\n",
674 		       (long) cur_limits.rlim_max);
675 	}
676 	printf("\tMaximum Message Size:\t\t\t%d\n", cur_max_msgsize);
677 	printf("\tMaximum Queue Size:\t\t\t%d\n", cur_max_msgs);
678 	printf("\tNice value:\t\t\t\t%d\n", cur_nice);
679 	printf("\tContinuous mode:\t\t\t(%s)\n", continuous_mode ?
680 	       (continuous_mode_fake ? "fake mode" : "enabled") :
681 	       "disabled");
682 	printf("\tCPUs to pin:\t\t\t\t%d", cpus_to_pin[0]);
683 	for (cpu = 1; cpu < num_cpus_to_pin; cpu++)
684 			printf(",%d", cpus_to_pin[cpu]);
685 	printf("\n");
686 
687 	sa.sa_sigaction = sig_action_SIGUSR1;
688 	sigemptyset(&sa.sa_mask);
689 	sigaddset(&sa.sa_mask, SIGHUP);
690 	sigaddset(&sa.sa_mask, SIGINT);
691 	sigaddset(&sa.sa_mask, SIGQUIT);
692 	sigaddset(&sa.sa_mask, SIGTERM);
693 	sa.sa_flags = SA_SIGINFO;
694 	if (sigaction(SIGUSR1, &sa, NULL) == -1)
695 		shutdown(1, "sigaction(SIGUSR1)", __LINE__);
696 	sa.sa_sigaction = sig_action;
697 	if (sigaction(SIGHUP, &sa, NULL) == -1)
698 		shutdown(1, "sigaction(SIGHUP)", __LINE__);
699 	if (sigaction(SIGINT, &sa, NULL) == -1)
700 		shutdown(1, "sigaction(SIGINT)", __LINE__);
701 	if (sigaction(SIGQUIT, &sa, NULL) == -1)
702 		shutdown(1, "sigaction(SIGQUIT)", __LINE__);
703 	if (sigaction(SIGTERM, &sa, NULL) == -1)
704 		shutdown(1, "sigaction(SIGTERM)", __LINE__);
705 
706 	if (!continuous_mode_fake) {
707 		attr.mq_flags = O_NONBLOCK;
708 		attr.mq_maxmsg = cur_max_msgs;
709 		attr.mq_msgsize = MSG_SIZE;
710 		open_queue(&attr);
711 	}
712 	for (i = 0; i < num_cpus_to_pin; i++) {
713 		pthread_attr_t thread_attr;
714 		void *thread_func;
715 
716 		if (continuous_mode_fake)
717 			thread_func = &fake_cont_thread;
718 		else if (continuous_mode)
719 			thread_func = &cont_thread;
720 		else
721 			thread_func = &perf_test_thread;
722 
723 		CPU_ZERO_S(cpu_set_size, cpu_set);
724 		CPU_SET_S(cpus_to_pin[i], cpu_set_size, cpu_set);
725 		pthread_attr_init(&thread_attr);
726 		pthread_attr_setaffinity_np(&thread_attr, cpu_set_size,
727 					    cpu_set);
728 		if (pthread_create(&cpu_threads[i], &thread_attr, thread_func,
729 				   NULL))
730 			shutdown(1, "pthread_create()", __LINE__);
731 		pthread_attr_destroy(&thread_attr);
732 	}
733 
734 	if (!continuous_mode) {
735 		pthread_join(cpu_threads[0], &retval);
736 		shutdown((long)retval, "perf_test_thread()", __LINE__);
737 	} else {
738 		while (1)
739 			sleep(1);
740 	}
741 	shutdown(0, "", 0);
742 }
743