1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2018, Matthew Macy 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 * 27 */ 28 29 #include <sys/cdefs.h> 30 __FBSDID("$FreeBSD$"); 31 32 #include <sys/param.h> 33 #include <sys/cpuset.h> 34 #include <sys/event.h> 35 #include <sys/queue.h> 36 #include <sys/socket.h> 37 #include <sys/stat.h> 38 #include <sys/sysctl.h> 39 #include <sys/time.h> 40 #include <sys/ttycom.h> 41 #include <sys/user.h> 42 #include <sys/wait.h> 43 44 #include <assert.h> 45 #include <curses.h> 46 #include <err.h> 47 #include <errno.h> 48 #include <fcntl.h> 49 #include <getopt.h> 50 #include <kvm.h> 51 #include <libgen.h> 52 #include <limits.h> 53 #include <locale.h> 54 #include <math.h> 55 #include <pmc.h> 56 #include <pmclog.h> 57 #include <regex.h> 58 #include <signal.h> 59 #include <stdarg.h> 60 #include <stdint.h> 61 #include <stdio.h> 62 #include <stdlib.h> 63 #include <string.h> 64 #include <sysexits.h> 65 #include <unistd.h> 66 67 #include <libpmcstat.h> 68 #include "cmd_pmc.h" 69 70 /* 71 * Return the frequency of the kernel's statistics clock. 72 */ 73 static int 74 getstathz(void) 75 { 76 int mib[2]; 77 size_t size; 78 struct clockinfo clockrate; 79 80 mib[0] = CTL_KERN; 81 mib[1] = KERN_CLOCKRATE; 82 size = sizeof clockrate; 83 if (sysctl(mib, 2, &clockrate, &size, NULL, 0) == -1) 84 err(1, "sysctl kern.clockrate"); 85 return clockrate.stathz; 86 } 87 88 #define STAT_MODE_NPMCS 6 89 #define FIXED_MODE_NPMCS 2 90 static struct timespec before_ts; 91 #define CYCLES 0 92 #define INST 1 93 #define BR 2 94 #define IAP_START BR 95 #define BR_MISS 3 96 #define CACHE 4 97 #define CACHE_MISS 5 98 static const char *pmc_stat_mode_names[] = { 99 "cycles", 100 "instructions", 101 "branches", 102 "branch-misses", 103 "cache-references", 104 "cache-misses", 105 }; 106 107 static int pmcstat_sockpair[NSOCKPAIRFD]; 108 109 static void __dead2 110 usage(void) 111 { 112 errx(EX_USAGE, 113 "\t get basic stats from command line program\n" 114 "\t -j <eventlist>, --events <eventlist> comma-delimited list of event specifiers\n" 115 ); 116 } 117 118 static void 119 showtime(FILE *out, struct timespec *before, struct timespec *after, 120 struct rusage *ru) 121 { 122 char decimal_point; 123 uint64_t real, user, sys; 124 125 (void)setlocale(LC_NUMERIC, ""); 126 decimal_point = localeconv()->decimal_point[0]; 127 128 after->tv_sec -= before->tv_sec; 129 after->tv_nsec -= before->tv_nsec; 130 if (after->tv_nsec < 0) { 131 after->tv_sec--; 132 after->tv_nsec += 1000000000; 133 } 134 135 real = (after->tv_sec * 1000000000 + after->tv_nsec) / 1000; 136 user = ru->ru_utime.tv_sec * 1000000 + ru->ru_utime.tv_usec; 137 sys = ru->ru_stime.tv_sec * 1000000 + ru->ru_stime.tv_usec; 138 fprintf(out, "%13jd%c%02ld real\t\t\t#\t%2.02f%% cpu\n", 139 (intmax_t)after->tv_sec, decimal_point, 140 after->tv_nsec / 10000000, 100 * (double)(sys + user + 1) / (double)(real + 1)); 141 fprintf(out, "%13jd%c%02ld user\t\t\t#\t%2.2f%% cpu\n", 142 (intmax_t)ru->ru_utime.tv_sec, decimal_point, 143 ru->ru_utime.tv_usec / 10000, 100 * (double)(user + 1) / (double)(real + 1)); 144 fprintf(out, "%13jd%c%02ld sys\t\t\t#\t%2.02f%% cpu\n", 145 (intmax_t)ru->ru_stime.tv_sec, decimal_point, 146 ru->ru_stime.tv_usec / 10000, 100 * (double)(sys + 1) / (double)(real + 1)); 147 } 148 149 static const char *stat_mode_cntrs[STAT_MODE_NPMCS]; 150 static const char *stat_mode_names[STAT_MODE_NPMCS]; 151 152 static void 153 pmc_stat_setup_stat(int system_mode, const char *arg) 154 { 155 const char *new_cntrs[STAT_MODE_NPMCS]; 156 static const char **pmc_stat_mode_cntrs; 157 struct pmcstat_ev *ev; 158 char *counters, *counter; 159 int i, c, start, newcnt; 160 cpuset_t cpumask, rootmask; 161 162 if (cpuset_getaffinity(CPU_LEVEL_ROOT, CPU_WHICH_PID, -1, 163 sizeof(rootmask), &rootmask) == -1) 164 err(EX_OSERR, "ERROR: Cannot determine the root set of CPUs"); 165 CPU_COPY(&rootmask, &cpumask); 166 167 if (pmc_pmu_stat_mode(&pmc_stat_mode_cntrs) != 0) 168 errx(EX_USAGE, "ERROR: hwmpc.ko not loaded or stat not supported on host."); 169 if (system_mode && geteuid() != 0) 170 errx(EX_USAGE, "ERROR: system mode counters can only be used as root"); 171 counters = NULL; 172 for (i = 0; i < STAT_MODE_NPMCS; i++) { 173 stat_mode_cntrs[i] = pmc_stat_mode_cntrs[i]; 174 stat_mode_names[i] = pmc_stat_mode_names[i]; 175 } 176 if (arg) { 177 counters = strdup(arg); 178 newcnt = 0; 179 while ((counter = strsep(&counters, ",")) != NULL && 180 newcnt < STAT_MODE_NPMCS - IAP_START) { 181 new_cntrs[newcnt++] = counter; 182 if (pmc_pmu_sample_rate_get(counter) == DEFAULT_SAMPLE_COUNT) 183 errx(EX_USAGE, "ERROR: %s not recognized on host", counter); 184 } 185 start = IAP_START + STAT_MODE_NPMCS - FIXED_MODE_NPMCS - newcnt; 186 for (i = 0; i < newcnt; i++) { 187 stat_mode_cntrs[start + i] = new_cntrs[i]; 188 stat_mode_names[start + i] = new_cntrs[i]; 189 } 190 } 191 if (system_mode) 192 pmc_args.pa_flags |= FLAG_HAS_SYSTEM_PMCS; 193 else 194 pmc_args.pa_flags |= FLAG_HAS_PROCESS_PMCS; 195 pmc_args.pa_flags |= FLAG_HAS_COUNTING_PMCS; 196 pmc_args.pa_flags |= FLAG_HAS_COMMANDLINE | FLAG_HAS_TARGET; 197 pmc_args.pa_flags |= FLAG_HAS_PIPE; 198 pmc_args.pa_required |= FLAG_HAS_COMMANDLINE | FLAG_HAS_TARGET | FLAG_HAS_OUTPUT_LOGFILE; 199 pmc_args.pa_outputpath = strdup("/dev/null"); 200 pmc_args.pa_logfd = pmcstat_open_log(pmc_args.pa_outputpath, 201 PMCSTAT_OPEN_FOR_WRITE); 202 for (i = 0; i < STAT_MODE_NPMCS; i++) { 203 if ((ev = malloc(sizeof(*ev))) == NULL) 204 errx(EX_SOFTWARE, "ERROR: Out of memory."); 205 if (system_mode) 206 ev->ev_mode = PMC_MODE_SC; 207 else 208 ev->ev_mode = PMC_MODE_TC; 209 ev->ev_spec = strdup(stat_mode_cntrs[i]); 210 if (ev->ev_spec == NULL) 211 errx(EX_SOFTWARE, "ERROR: Out of memory."); 212 c = strcspn(strdup(stat_mode_cntrs[i]), ", \t"); 213 ev->ev_name = malloc(c + 1); 214 if (ev->ev_name == NULL) 215 errx(EX_SOFTWARE, "ERROR: Out of memory."); 216 (void)strncpy(ev->ev_name, stat_mode_cntrs[i], c); 217 *(ev->ev_name + c) = '\0'; 218 219 ev->ev_count = -1; 220 ev->ev_flags = 0; 221 ev->ev_flags |= PMC_F_DESCENDANTS; 222 ev->ev_cumulative = 1; 223 224 ev->ev_saved = 0LL; 225 ev->ev_pmcid = PMC_ID_INVALID; 226 STAILQ_INSERT_TAIL(&pmc_args.pa_events, ev, ev_next); 227 if (system_mode) { 228 ev->ev_cpu = CPU_FFS(&cpumask) - 1; 229 CPU_CLR(ev->ev_cpu, &cpumask); 230 pmcstat_clone_event_descriptor(ev, &cpumask, &pmc_args); 231 CPU_SET(ev->ev_cpu, &cpumask); 232 } else 233 ev->ev_cpu = PMC_CPU_ANY; 234 235 } 236 if (clock_gettime(CLOCK_MONOTONIC, &before_ts)) 237 err(1, "clock_gettime"); 238 } 239 240 static void 241 pmc_stat_print_stat(struct rusage *ru) 242 { 243 struct pmcstat_ev *ev; 244 struct timespec after; 245 uint64_t cvals[STAT_MODE_NPMCS]; 246 uint64_t ticks, value; 247 int hz, i; 248 249 if (ru) { 250 hz = getstathz(); 251 ticks = hz * (ru->ru_utime.tv_sec + ru->ru_stime.tv_sec) + 252 hz * (ru->ru_utime.tv_usec + ru->ru_stime.tv_usec) / 1000000; 253 if (clock_gettime(CLOCK_MONOTONIC, &after)) 254 err(1, "clock_gettime"); 255 /* 256 * If our round-off on the tick calculation still puts us at 0, 257 * then always assume at least one tick. 258 */ 259 if (ticks == 0) 260 ticks = 1; 261 fprintf(pmc_args.pa_printfile, "%16ld %s\t\t#\t%02.03f M/sec\n", 262 ru->ru_minflt, "page faults", ((double)ru->ru_minflt / (double)ticks) / hz); 263 fprintf(pmc_args.pa_printfile, "%16ld %s\t\t#\t%02.03f M/sec\n", 264 ru->ru_nvcsw, "voluntary csw", ((double)ru->ru_nvcsw / (double)ticks) / hz); 265 fprintf(pmc_args.pa_printfile, "%16ld %s\t#\t%02.03f M/sec\n", 266 ru->ru_nivcsw, "involuntary csw", ((double)ru->ru_nivcsw / (double)ticks) / hz); 267 } 268 269 bzero(&cvals, sizeof(cvals)); 270 STAILQ_FOREACH(ev, &pmc_args.pa_events, ev_next) { 271 if (pmc_read(ev->ev_pmcid, &value) < 0) 272 err(EX_OSERR, "ERROR: Cannot read pmc \"%s\"", 273 ev->ev_name); 274 for (i = 0; i < STAT_MODE_NPMCS; i++) 275 if (strcmp(ev->ev_name, stat_mode_cntrs[i]) == 0) 276 cvals[i] += value; 277 } 278 279 fprintf(pmc_args.pa_printfile, "%16jd %s\n", (uintmax_t)cvals[CYCLES], stat_mode_names[CYCLES]); 280 fprintf(pmc_args.pa_printfile, "%16jd %s\t\t#\t%01.03f inst/cycle\n", (uintmax_t)cvals[INST], stat_mode_names[INST], 281 (double)cvals[INST] / cvals[CYCLES]); 282 fprintf(pmc_args.pa_printfile, "%16jd %s\n", (uintmax_t)cvals[BR], stat_mode_names[BR]); 283 if (stat_mode_names[BR_MISS] == pmc_stat_mode_names[BR_MISS]) 284 fprintf(pmc_args.pa_printfile, "%16jd %s\t\t#\t%.03f%%\n", 285 (uintmax_t)cvals[BR_MISS], stat_mode_names[BR_MISS], 286 100 * ((double)cvals[BR_MISS] / cvals[BR])); 287 else 288 fprintf(pmc_args.pa_printfile, "%16jd %s\n", 289 (uintmax_t)cvals[BR_MISS], stat_mode_names[BR_MISS]); 290 fprintf(pmc_args.pa_printfile, "%16jd %s%s", (uintmax_t)cvals[CACHE], stat_mode_names[CACHE], 291 stat_mode_names[CACHE] != pmc_stat_mode_names[CACHE] ? "\n" : ""); 292 if (stat_mode_names[CACHE] == pmc_stat_mode_names[CACHE]) 293 fprintf(pmc_args.pa_printfile, "\t#\t%.03f refs/inst\n", 294 ((double)cvals[CACHE] / cvals[INST])); 295 fprintf(pmc_args.pa_printfile, "%16jd %s%s", (uintmax_t)cvals[CACHE_MISS], stat_mode_names[CACHE_MISS], 296 stat_mode_names[CACHE_MISS] != pmc_stat_mode_names[CACHE_MISS] ? "\n" : ""); 297 if (stat_mode_names[CACHE_MISS] == pmc_stat_mode_names[CACHE_MISS]) 298 fprintf(pmc_args.pa_printfile, "\t\t#\t%.03f%%\n", 299 100 * ((double)cvals[CACHE_MISS] / cvals[CACHE])); 300 301 if (ru) 302 showtime(pmc_args.pa_printfile, &before_ts, &after, ru); 303 } 304 305 static struct option longopts[] = { 306 {"events", required_argument, NULL, 'j'}, 307 {NULL, 0, NULL, 0} 308 }; 309 310 static int 311 pmc_stat_internal(int argc, char **argv, int system_mode) 312 { 313 char *event, *r; 314 struct sigaction sa; 315 struct kevent kev; 316 struct rusage ru; 317 struct winsize ws; 318 struct pmcstat_ev *ev; 319 int c, option, runstate; 320 int waitstatus, ru_valid, do_debug; 321 322 do_debug = ru_valid = 0; 323 r = event = NULL; 324 while ((option = getopt_long(argc, argv, "dj:", longopts, NULL)) != -1) { 325 switch (option) { 326 case 'j': 327 r = event = strdup(optarg); 328 break; 329 case 'd': 330 do_debug = 1; 331 break; 332 case '?': 333 default: 334 usage(); 335 } 336 } 337 pmc_args.pa_argc = (argc -= optind); 338 pmc_args.pa_argv = (argv += optind); 339 if (argc == 0) 340 usage(); 341 pmc_args.pa_flags |= FLAG_HAS_COMMANDLINE; 342 pmc_stat_setup_stat(system_mode, event); 343 free(r); 344 bzero(&ru, sizeof(ru)); 345 EV_SET(&kev, SIGINT, EVFILT_SIGNAL, EV_ADD, 0, 0, NULL); 346 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 347 err(EX_OSERR, "ERROR: Cannot register kevent for SIGINT"); 348 349 EV_SET(&kev, SIGIO, EVFILT_SIGNAL, EV_ADD, 0, 0, NULL); 350 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 351 err(EX_OSERR, "ERROR: Cannot register kevent for SIGIO"); 352 EV_SET(&kev, 0, EVFILT_TIMER, EV_ADD, 0, 1000, NULL); 353 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 354 err(EX_OSERR, 355 "ERROR: Cannot register kevent for timer"); 356 357 STAILQ_FOREACH(ev, &pmc_args.pa_events, ev_next) { 358 if (pmc_allocate(ev->ev_spec, ev->ev_mode, 359 ev->ev_flags, ev->ev_cpu, &ev->ev_pmcid, ev->ev_count) < 0) 360 err(EX_OSERR, 361 "ERROR: Cannot allocate %s-mode pmc with specification \"%s\"", 362 PMC_IS_SYSTEM_MODE(ev->ev_mode) ? 363 "system" : "process", ev->ev_spec); 364 365 if (PMC_IS_SAMPLING_MODE(ev->ev_mode) && 366 pmc_set(ev->ev_pmcid, ev->ev_count) < 0) 367 err(EX_OSERR, 368 "ERROR: Cannot set sampling count for PMC \"%s\"", 369 ev->ev_name); 370 } 371 372 /* 373 * An exec() failure of a forked child is signalled by the 374 * child sending the parent a SIGCHLD. We don't register an 375 * actual signal handler for SIGCHLD, but instead use our 376 * kqueue to pick up the signal. 377 */ 378 EV_SET(&kev, SIGCHLD, EVFILT_SIGNAL, EV_ADD, 0, 0, NULL); 379 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 380 err(EX_OSERR, "ERROR: Cannot register kevent for SIGCHLD"); 381 382 pmcstat_create_process(pmcstat_sockpair, &pmc_args, pmc_kq); 383 384 if (SLIST_EMPTY(&pmc_args.pa_targets)) 385 errx(EX_DATAERR, 386 "ERROR: No matching target processes."); 387 if (pmc_args.pa_flags & FLAG_HAS_PROCESS_PMCS) 388 pmcstat_attach_pmcs(&pmc_args); 389 390 /* start the pmcs */ 391 pmc_util_start_pmcs(&pmc_args); 392 393 /* start the (commandline) process if needed */ 394 pmcstat_start_process(pmcstat_sockpair); 395 396 /* Handle SIGINT using the kqueue loop */ 397 sa.sa_handler = SIG_IGN; 398 sa.sa_flags = 0; 399 (void)sigemptyset(&sa.sa_mask); 400 401 if (sigaction(SIGINT, &sa, NULL) < 0) 402 err(EX_OSERR, "ERROR: Cannot install signal handler"); 403 404 /* 405 * loop till either the target process (if any) exits, or we 406 * are killed by a SIGINT or we reached the time duration. 407 */ 408 runstate = PMCSTAT_RUNNING; 409 do { 410 if ((c = kevent(pmc_kq, NULL, 0, &kev, 1, NULL)) <= 0) { 411 if (errno != EINTR) 412 err(EX_OSERR, "ERROR: kevent failed"); 413 else 414 continue; 415 } 416 if (kev.flags & EV_ERROR) 417 errc(EX_OSERR, kev.data, "ERROR: kevent failed"); 418 419 switch (kev.filter) { 420 case EVFILT_PROC: /* target has exited */ 421 if (wait4(pmc_util_get_pid(&pmc_args), &waitstatus, 0, &ru) > 0) { 422 getrusage(RUSAGE_CHILDREN, &ru); 423 ru_valid = 1; 424 } 425 break; 426 427 case EVFILT_READ: /* log file data is present */ 428 break; 429 case EVFILT_TIMER: 430 if (do_debug) 431 pmc_stat_print_stat(NULL); 432 break; 433 case EVFILT_SIGNAL: 434 if (kev.ident == SIGCHLD) { 435 /* 436 * The child process sends us a 437 * SIGCHLD if its exec() failed. We 438 * wait for it to exit and then exit 439 * ourselves. 440 */ 441 (void)wait(&c); 442 runstate = PMCSTAT_FINISHED; 443 } else if (kev.ident == SIGIO) { 444 /* 445 * We get a SIGIO if a PMC loses all 446 * of its targets, or if logfile 447 * writes encounter an error. 448 */ 449 if (wait4(pmc_util_get_pid(&pmc_args), &waitstatus, 0, &ru) > 0) { 450 getrusage(RUSAGE_CHILDREN, &ru); 451 ru_valid = 1; 452 } 453 runstate = pmcstat_close_log(&pmc_args); 454 } else if (kev.ident == SIGINT) { 455 /* Kill the child process if we started it */ 456 if (pmc_args.pa_flags & FLAG_HAS_COMMANDLINE) 457 pmc_util_kill_process(&pmc_args); 458 runstate = pmcstat_close_log(&pmc_args); 459 } else if (kev.ident == SIGWINCH) { 460 if (ioctl(fileno(pmc_args.pa_printfile), 461 TIOCGWINSZ, &ws) < 0) 462 err(EX_OSERR, 463 "ERROR: Cannot determine window size"); 464 pmc_displayheight = ws.ws_row - 1; 465 pmc_displaywidth = ws.ws_col - 1; 466 } else 467 assert(0); 468 469 break; 470 } 471 } while (runstate != PMCSTAT_FINISHED); 472 if (!ru_valid) 473 warnx("couldn't get rusage"); 474 pmc_stat_print_stat(&ru); 475 pmc_util_cleanup(&pmc_args); 476 return (0); 477 } 478 479 int 480 cmd_pmc_stat(int argc, char **argv) 481 { 482 return (pmc_stat_internal(argc, argv, 0)); 483 } 484 485 int 486 cmd_pmc_stat_system(int argc, char **argv) 487 { 488 return (pmc_stat_internal(argc, argv, 1)); 489 } 490