Allow _Qp_fgt in sparc64 localplt.data
[glibc.git] / nptl / perf.c
blobb1373ffabdb4fc8074c0e6d59c759b9b559e5e39
1 /* Copyright (C) 2002-2022 Free Software Foundation, Inc.
2 This file is part of the GNU C Library.
4 The GNU C Library is free software; you can redistribute it and/or
5 modify it under the terms of the GNU Lesser General Public
6 License as published by the Free Software Foundation; either
7 version 2.1 of the License, or (at your option) any later version.
9 The GNU C Library 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 GNU
12 Lesser General Public License for more details.
14 You should have received a copy of the GNU Lesser General Public
15 License along with the GNU C Library; if not, see
16 <https://www.gnu.org/licenses/>. */
18 #define _GNU_SOURCE 1
19 #include <argp.h>
20 #include <error.h>
21 #include <errno.h>
22 #include <fcntl.h>
23 #include <inttypes.h>
24 #include <limits.h>
25 #include <pthread.h>
26 #include <signal.h>
27 #include <stdbool.h>
28 #include <stdlib.h>
29 #include <string.h>
30 #include <time.h>
31 #include <unistd.h>
32 #include <sys/param.h>
33 #include <sys/types.h>
35 #ifndef MAX_THREADS
36 # define MAX_THREADS 100000
37 #endif
38 #ifndef DEFAULT_THREADS
39 # define DEFAULT_THREADS 50
40 #endif
43 #define OPT_TO_THREAD 300
44 #define OPT_TO_PROCESS 301
45 #define OPT_SYNC_SIGNAL 302
46 #define OPT_SYNC_JOIN 303
47 #define OPT_TOPLEVEL 304
50 static const struct argp_option options[] =
52 { NULL, 0, NULL, 0, "\
53 This is a test for threads so we allow ther user to selection the number of \
54 threads which are used at any one time. Independently the total number of \
55 rounds can be selected. This is the total number of threads which will have \
56 run when the process terminates:" },
57 { "threads", 't', "NUMBER", 0, "Number of threads used at once" },
58 { "starts", 's', "NUMBER", 0, "Total number of working threads" },
59 { "toplevel", OPT_TOPLEVEL, "NUMBER", 0,
60 "Number of toplevel threads which start the other threads; this \
61 implies --sync-join" },
63 { NULL, 0, NULL, 0, "\
64 Each thread can do one of two things: sleep or do work. The latter is 100% \
65 CPU bound. The work load is the probability a thread does work. All values \
66 from zero to 100 (inclusive) are valid. How often each thread repeats this \
67 can be determined by the number of rounds. The work cost determines how long \
68 each work session (not sleeping) takes. If it is zero a thread would \
69 effectively nothing. By setting the number of rounds to zero the thread \
70 does no work at all and pure thread creation times can be measured." },
71 { "workload", 'w', "PERCENT", 0, "Percentage of time spent working" },
72 { "workcost", 'c', "NUMBER", 0,
73 "Factor in the cost of each round of working" },
74 { "rounds", 'r', "NUMBER", 0, "Number of rounds each thread runs" },
76 { NULL, 0, NULL, 0, "\
77 There are a number of different methods how thread creation can be \
78 synchronized. Synchronization is necessary since the number of concurrently \
79 running threads is limited." },
80 { "sync-signal", OPT_SYNC_SIGNAL, NULL, 0,
81 "Synchronize using a signal (default)" },
82 { "sync-join", OPT_SYNC_JOIN, NULL, 0, "Synchronize using pthread_join" },
84 { NULL, 0, NULL, 0, "\
85 One parameter for each threads execution is the size of the stack. If this \
86 parameter is not used the system's default stack size is used. If many \
87 threads are used the stack size should be chosen quite small." },
88 { "stacksize", 'S', "BYTES", 0, "Size of threads stack" },
89 { "guardsize", 'g', "BYTES", 0,
90 "Size of stack guard area; must fit into the stack" },
92 { NULL, 0, NULL, 0, "Signal options:" },
93 { "to-thread", OPT_TO_THREAD, NULL, 0, "Send signal to main thread" },
94 { "to-process", OPT_TO_PROCESS, NULL, 0,
95 "Send signal to process (default)" },
97 { NULL, 0, NULL, 0, "Administrative options:" },
98 { "progress", 'p', NULL, 0, "Show signs of progress" },
99 { "timing", 'T', NULL, 0,
100 "Measure time from startup to the last thread finishing" },
101 { NULL, 0, NULL, 0, NULL }
104 /* Prototype for option handler. */
105 static error_t parse_opt (int key, char *arg, struct argp_state *state);
107 /* Data structure to communicate with argp functions. */
108 static struct argp argp =
110 options, parse_opt
114 static unsigned long int threads = DEFAULT_THREADS;
115 static unsigned long int workload = 75;
116 static unsigned long int workcost = 20;
117 static unsigned long int rounds = 10;
118 static long int starts = 5000;
119 static unsigned long int stacksize;
120 static long int guardsize = -1;
121 static bool progress;
122 static bool timing;
123 static bool to_thread;
124 static unsigned long int toplevel = 1;
127 static long int running;
128 static pthread_mutex_t running_mutex = PTHREAD_MUTEX_INITIALIZER;
130 static pid_t pid;
131 static pthread_t tmain;
133 static clockid_t cl;
134 static struct timespec start_time;
137 static pthread_mutex_t sum_mutex = PTHREAD_MUTEX_INITIALIZER;
138 unsigned int sum;
140 static enum
142 sync_signal,
143 sync_join
145 sync_method;
148 /* We use 64bit values for the times. */
149 typedef unsigned long long int hp_timing_t;
152 /* Attributes for all created threads. */
153 static pthread_attr_t attr;
156 static void *
157 work (void *arg)
159 unsigned long int i;
160 unsigned int state = (unsigned long int) arg;
162 for (i = 0; i < rounds; ++i)
164 /* Determine what to do. */
165 unsigned int rnum;
167 /* Uniform distribution. */
169 rnum = rand_r (&state);
170 while (rnum >= UINT_MAX - (UINT_MAX % 100));
172 rnum %= 100;
174 if (rnum < workload)
176 int j;
177 int a[4] = { i, rnum, i + rnum, rnum - i };
179 if (progress)
180 write (STDERR_FILENO, "c", 1);
182 for (j = 0; j < workcost; ++j)
184 a[0] += a[3] >> 12;
185 a[1] += a[2] >> 20;
186 a[2] += a[1] ^ 0x3423423;
187 a[3] += a[0] - a[1];
190 pthread_mutex_lock (&sum_mutex);
191 sum += a[0] + a[1] + a[2] + a[3];
192 pthread_mutex_unlock (&sum_mutex);
194 else
196 /* Just sleep. */
197 struct timespec tv;
199 tv.tv_sec = 0;
200 tv.tv_nsec = 10000000;
202 if (progress)
203 write (STDERR_FILENO, "w", 1);
205 nanosleep (&tv, NULL);
209 return NULL;
213 static void *
214 thread_function (void *arg)
216 work (arg);
218 pthread_mutex_lock (&running_mutex);
219 if (--running <= 0 && starts <= 0)
221 /* We are done. */
222 if (progress)
223 write (STDERR_FILENO, "\n", 1);
225 if (timing)
227 struct timespec end_time;
229 if (clock_gettime (cl, &end_time) == 0)
231 end_time.tv_sec -= start_time.tv_sec;
232 end_time.tv_nsec -= start_time.tv_nsec;
233 if (end_time.tv_nsec < 0)
235 end_time.tv_nsec += 1000000000;
236 --end_time.tv_sec;
239 printf ("\nRuntime: %lu.%09lu seconds\n",
240 (unsigned long int) end_time.tv_sec,
241 (unsigned long int) end_time.tv_nsec);
245 printf ("Result: %08x\n", sum);
247 exit (0);
249 pthread_mutex_unlock (&running_mutex);
251 if (sync_method == sync_signal)
253 if (to_thread)
254 /* This code sends a signal to the main thread. */
255 pthread_kill (tmain, SIGUSR1);
256 else
257 /* Use this code to test sending a signal to the process. */
258 kill (pid, SIGUSR1);
261 if (progress)
262 write (STDERR_FILENO, "f", 1);
264 return NULL;
268 struct start_info
270 unsigned int starts;
271 unsigned int threads;
275 static void *
276 start_threads (void *arg)
278 struct start_info *si = arg;
279 unsigned int starts = si->starts;
280 pthread_t ths[si->threads];
281 unsigned int state = starts;
282 unsigned int n;
283 unsigned int i = 0;
284 int err;
286 if (progress)
287 write (STDERR_FILENO, "T", 1);
289 memset (ths, '\0', sizeof (pthread_t) * si->threads);
291 while (starts-- > 0)
293 if (ths[i] != 0)
295 /* Wait for the threads in the order they were created. */
296 err = pthread_join (ths[i], NULL);
297 if (err != 0)
298 error (EXIT_FAILURE, err, "cannot join thread");
300 if (progress)
301 write (STDERR_FILENO, "f", 1);
304 err = pthread_create (&ths[i], &attr, work,
305 (void *) (long) (rand_r (&state) + starts + i));
307 if (err != 0)
308 error (EXIT_FAILURE, err, "cannot start thread");
310 if (progress)
311 write (STDERR_FILENO, "t", 1);
313 if (++i == si->threads)
314 i = 0;
317 n = i;
320 if (ths[i] != 0)
322 err = pthread_join (ths[i], NULL);
323 if (err != 0)
324 error (EXIT_FAILURE, err, "cannot join thread");
326 if (progress)
327 write (STDERR_FILENO, "f", 1);
330 if (++i == si->threads)
331 i = 0;
333 while (i != n);
335 if (progress)
336 write (STDERR_FILENO, "F", 1);
338 return NULL;
343 main (int argc, char *argv[])
345 int remaining;
346 sigset_t ss;
347 pthread_t th;
348 pthread_t *ths = NULL;
349 int empty = 0;
350 int last;
351 bool cont = true;
353 /* Parse and process arguments. */
354 argp_parse (&argp, argc, argv, 0, &remaining, NULL);
356 if (sync_method == sync_join)
358 ths = (pthread_t *) calloc (threads, sizeof (pthread_t));
359 if (ths == NULL)
360 error (EXIT_FAILURE, errno,
361 "cannot allocate memory for thread descriptor array");
363 last = threads;
365 else
367 ths = &th;
368 last = 1;
371 if (toplevel > threads)
373 printf ("resetting number of toplevel threads to %lu to not surpass number to concurrent threads\n",
374 threads);
375 toplevel = threads;
378 if (timing)
380 if (clock_getcpuclockid (0, &cl) != 0
381 || clock_gettime (cl, &start_time) != 0)
382 timing = false;
385 /* We need this later. */
386 pid = getpid ();
387 tmain = pthread_self ();
389 /* We use signal SIGUSR1 for communication between the threads and
390 the main thread. We only want sychronous notification. */
391 if (sync_method == sync_signal)
393 sigemptyset (&ss);
394 sigaddset (&ss, SIGUSR1);
395 if (sigprocmask (SIG_BLOCK, &ss, NULL) != 0)
396 error (EXIT_FAILURE, errno, "cannot set signal mask");
399 /* Create the thread attributes. */
400 pthread_attr_init (&attr);
402 /* If the user provided a stack size use it. */
403 if (stacksize != 0
404 && pthread_attr_setstacksize (&attr, stacksize) != 0)
405 puts ("could not set stack size; will use default");
406 /* And stack guard size. */
407 if (guardsize != -1
408 && pthread_attr_setguardsize (&attr, guardsize) != 0)
409 puts ("invalid stack guard size; will use default");
411 /* All threads are created detached if we are not using pthread_join
412 to synchronize. */
413 if (sync_method != sync_join)
414 pthread_attr_setdetachstate (&attr, PTHREAD_CREATE_DETACHED);
416 if (sync_method == sync_signal)
418 while (1)
420 int err;
421 bool do_wait = false;
423 pthread_mutex_lock (&running_mutex);
424 if (starts-- < 0)
425 cont = false;
426 else
427 do_wait = ++running >= threads && starts > 0;
429 pthread_mutex_unlock (&running_mutex);
431 if (! cont)
432 break;
434 if (progress)
435 write (STDERR_FILENO, "t", 1);
437 err = pthread_create (&ths[empty], &attr, thread_function,
438 (void *) starts);
439 if (err != 0)
440 error (EXIT_FAILURE, err, "cannot start thread %lu", starts);
442 if (++empty == last)
443 empty = 0;
445 if (do_wait)
446 sigwaitinfo (&ss, NULL);
449 /* Do nothing anymore. On of the threads will terminate the program. */
450 sigfillset (&ss);
451 sigdelset (&ss, SIGINT);
452 while (1)
453 sigsuspend (&ss);
455 else
457 pthread_t ths[toplevel];
458 struct start_info si[toplevel];
459 unsigned int i;
461 for (i = 0; i < toplevel; ++i)
463 unsigned int child_starts = starts / (toplevel - i);
464 unsigned int child_threads = threads / (toplevel - i);
465 int err;
467 si[i].starts = child_starts;
468 si[i].threads = child_threads;
470 err = pthread_create (&ths[i], &attr, start_threads, &si[i]);
471 if (err != 0)
472 error (EXIT_FAILURE, err, "cannot start thread");
474 starts -= child_starts;
475 threads -= child_threads;
478 for (i = 0; i < toplevel; ++i)
480 int err = pthread_join (ths[i], NULL);
482 if (err != 0)
483 error (EXIT_FAILURE, err, "cannot join thread");
486 /* We are done. */
487 if (progress)
488 write (STDERR_FILENO, "\n", 1);
490 if (timing)
492 struct timespec end_time;
494 if (clock_gettime (cl, &end_time) == 0)
496 end_time.tv_sec -= start_time.tv_sec;
497 end_time.tv_nsec -= start_time.tv_nsec;
498 if (end_time.tv_nsec < 0)
500 end_time.tv_nsec += 1000000000;
501 --end_time.tv_sec;
504 printf ("\nRuntime: %lu.%09lu seconds\n",
505 (unsigned long int) end_time.tv_sec,
506 (unsigned long int) end_time.tv_nsec);
510 printf ("Result: %08x\n", sum);
512 exit (0);
515 /* NOTREACHED */
516 return 0;
520 /* Handle program arguments. */
521 static error_t
522 parse_opt (int key, char *arg, struct argp_state *state)
524 unsigned long int num;
525 long int snum;
527 switch (key)
529 case 't':
530 num = strtoul (arg, NULL, 0);
531 if (num <= MAX_THREADS)
532 threads = num;
533 else
534 printf ("\
535 number of threads limited to %u; recompile with a higher limit if necessary",
536 MAX_THREADS);
537 break;
539 case 'w':
540 num = strtoul (arg, NULL, 0);
541 if (num <= 100)
542 workload = num;
543 else
544 puts ("workload must be between 0 and 100 percent");
545 break;
547 case 'c':
548 workcost = strtoul (arg, NULL, 0);
549 break;
551 case 'r':
552 rounds = strtoul (arg, NULL, 0);
553 break;
555 case 's':
556 starts = strtoul (arg, NULL, 0);
557 break;
559 case 'S':
560 num = strtoul (arg, NULL, 0);
561 if (num >= PTHREAD_STACK_MIN)
562 stacksize = num;
563 else
564 printf ("minimum stack size is %d\n", PTHREAD_STACK_MIN);
565 break;
567 case 'g':
568 snum = strtol (arg, NULL, 0);
569 if (snum < 0)
570 printf ("invalid guard size %s\n", arg);
571 else
572 guardsize = snum;
573 break;
575 case 'p':
576 progress = true;
577 break;
579 case 'T':
580 timing = true;
581 break;
583 case OPT_TO_THREAD:
584 to_thread = true;
585 break;
587 case OPT_TO_PROCESS:
588 to_thread = false;
589 break;
591 case OPT_SYNC_SIGNAL:
592 sync_method = sync_signal;
593 break;
595 case OPT_SYNC_JOIN:
596 sync_method = sync_join;
597 break;
599 case OPT_TOPLEVEL:
600 num = strtoul (arg, NULL, 0);
601 if (num < MAX_THREADS)
602 toplevel = num;
603 else
604 printf ("\
605 number of threads limited to %u; recompile with a higher limit if necessary",
606 MAX_THREADS);
607 sync_method = sync_join;
608 break;
610 default:
611 return ARGP_ERR_UNKNOWN;
614 return 0;
618 static hp_timing_t
619 get_clockfreq (void)
621 /* We read the information from the /proc filesystem. It contains at
622 least one line like
623 cpu MHz : 497.840237
624 or also
625 cpu MHz : 497.841
626 We search for this line and convert the number in an integer. */
627 static hp_timing_t result;
628 int fd;
630 /* If this function was called before, we know the result. */
631 if (result != 0)
632 return result;
634 fd = open ("/proc/cpuinfo", O_RDONLY);
635 if (__glibc_likely (fd != -1))
637 /* XXX AFAIK the /proc filesystem can generate "files" only up
638 to a size of 4096 bytes. */
639 char buf[4096];
640 ssize_t n;
642 n = read (fd, buf, sizeof buf);
643 if (__builtin_expect (n, 1) > 0)
645 char *mhz = memmem (buf, n, "cpu MHz", 7);
647 if (__glibc_likely (mhz != NULL))
649 char *endp = buf + n;
650 int seen_decpoint = 0;
651 int ndigits = 0;
653 /* Search for the beginning of the string. */
654 while (mhz < endp && (*mhz < '0' || *mhz > '9') && *mhz != '\n')
655 ++mhz;
657 while (mhz < endp && *mhz != '\n')
659 if (*mhz >= '0' && *mhz <= '9')
661 result *= 10;
662 result += *mhz - '0';
663 if (seen_decpoint)
664 ++ndigits;
666 else if (*mhz == '.')
667 seen_decpoint = 1;
669 ++mhz;
672 /* Compensate for missing digits at the end. */
673 while (ndigits++ < 6)
674 result *= 10;
678 close (fd);
681 return result;
686 clock_getcpuclockid (pid_t pid, clockid_t *clock_id)
688 /* We don't allow any process ID but our own. */
689 if (pid != 0 && pid != getpid ())
690 return EPERM;
692 #ifdef CLOCK_PROCESS_CPUTIME_ID
693 /* Store the number. */
694 *clock_id = CLOCK_PROCESS_CPUTIME_ID;
696 return 0;
697 #else
698 /* We don't have a timer for that. */
699 return ENOENT;
700 #endif
704 #ifdef i386
705 #define HP_TIMING_NOW(Var) __asm__ __volatile__ ("rdtsc" : "=A" (Var))
706 #elif defined __x86_64__
707 # define HP_TIMING_NOW(Var) \
708 ({ unsigned int _hi, _lo; \
709 asm volatile ("rdtsc" : "=a" (_lo), "=d" (_hi)); \
710 (Var) = ((unsigned long long int) _hi << 32) | _lo; })
711 #elif defined __ia64__
712 #define HP_TIMING_NOW(Var) __asm__ __volatile__ ("mov %0=ar.itc" : "=r" (Var) : : "memory")
713 #else
714 #error "HP_TIMING_NOW missing"
715 #endif
717 /* Get current value of CLOCK and store it in TP. */
719 clock_gettime (clockid_t clock_id, struct timespec *tp)
721 int retval = -1;
723 switch (clock_id)
725 case CLOCK_PROCESS_CPUTIME_ID:
728 static hp_timing_t freq;
729 hp_timing_t tsc;
731 /* Get the current counter. */
732 HP_TIMING_NOW (tsc);
734 if (freq == 0)
736 freq = get_clockfreq ();
737 if (freq == 0)
738 return EINVAL;
741 /* Compute the seconds. */
742 tp->tv_sec = tsc / freq;
744 /* And the nanoseconds. This computation should be stable until
745 we get machines with about 16GHz frequency. */
746 tp->tv_nsec = ((tsc % freq) * UINT64_C (1000000000)) / freq;
748 retval = 0;
750 break;
752 default:
753 errno = EINVAL;
754 break;
757 return retval;