1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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 #include <sys/param.h> 31 #include <sys/cpuset.h> 32 #include <sys/event.h> 33 #include <sys/queue.h> 34 #include <sys/socket.h> 35 #include <sys/stat.h> 36 #include <sys/sysctl.h> 37 #include <sys/time.h> 38 #include <sys/ttycom.h> 39 #include <sys/user.h> 40 #include <sys/wait.h> 41 42 #include <assert.h> 43 #include <curses.h> 44 #include <err.h> 45 #include <errno.h> 46 #include <fcntl.h> 47 #include <getopt.h> 48 #include <kvm.h> 49 #include <libgen.h> 50 #include <limits.h> 51 #include <locale.h> 52 #include <math.h> 53 #include <pmc.h> 54 #include <pmclog.h> 55 #include <regex.h> 56 #include <signal.h> 57 #include <stdarg.h> 58 #include <stdint.h> 59 #include <stdio.h> 60 #include <stdlib.h> 61 #include <string.h> 62 #include <sysexits.h> 63 #include <unistd.h> 64 65 #include <libpmcstat.h> 66 #include "cmd_pmc.h" 67 68 /* 69 * Return the frequency of the kernel's statistics clock. 70 */ 71 static int 72 getstathz(void) 73 { 74 int mib[2]; 75 size_t size; 76 struct clockinfo clockrate; 77 78 mib[0] = CTL_KERN; 79 mib[1] = KERN_CLOCKRATE; 80 size = sizeof clockrate; 81 if (sysctl(mib, 2, &clockrate, &size, NULL, 0) == -1) 82 err(1, "sysctl kern.clockrate"); 83 return clockrate.stathz; 84 } 85 86 #define STAT_MODE_NPMCS 6 87 #define FIXED_MODE_NPMCS 2 88 static struct timespec before_ts; 89 #define CYCLES 0 90 #define INST 1 91 #define BR 2 92 #define IAP_START BR 93 #define BR_MISS 3 94 #define CACHE 4 95 #define CACHE_MISS 5 96 static const char *pmc_stat_mode_names[] = { 97 "cycles", 98 "instructions", 99 "branches", 100 "branch-misses", 101 "cache-references", 102 "cache-misses", 103 }; 104 105 /* Common aliases for the desired stat counter */ 106 static const char *pmc_stat_mode_aliases[] = { 107 "unhalted-cycles", 108 "instructions", 109 "branches", 110 "branch-mispredicts", 111 "LLC-REFERENCE", 112 "LLC-MISSES", 113 }; 114 115 static int pmcstat_sockpair[NSOCKPAIRFD]; 116 117 static void __dead2 118 usage(void) 119 { 120 errx(EX_USAGE, 121 "\t get basic stats from command line program\n" 122 "\t -j <eventlist>, --events <eventlist> comma-delimited list of event specifiers\n" 123 ); 124 } 125 126 static void 127 showtime(FILE *out, struct timespec *before, struct timespec *after, 128 struct rusage *ru) 129 { 130 char decimal_point; 131 uint64_t real, user, sys; 132 133 (void)setlocale(LC_NUMERIC, ""); 134 decimal_point = localeconv()->decimal_point[0]; 135 136 after->tv_sec -= before->tv_sec; 137 after->tv_nsec -= before->tv_nsec; 138 if (after->tv_nsec < 0) { 139 after->tv_sec--; 140 after->tv_nsec += 1000000000; 141 } 142 143 real = (after->tv_sec * 1000000000 + after->tv_nsec) / 1000; 144 user = ru->ru_utime.tv_sec * 1000000 + ru->ru_utime.tv_usec; 145 sys = ru->ru_stime.tv_sec * 1000000 + ru->ru_stime.tv_usec; 146 fprintf(out, "%13jd%c%02ld real\t\t\t#\t%2.02f%% cpu\n", 147 (intmax_t)after->tv_sec, decimal_point, 148 after->tv_nsec / 10000000, 100 * (double)(sys + user + 1) / (double)(real + 1)); 149 fprintf(out, "%13jd%c%02ld user\t\t\t#\t%2.2f%% cpu\n", 150 (intmax_t)ru->ru_utime.tv_sec, decimal_point, 151 ru->ru_utime.tv_usec / 10000, 100 * (double)(user + 1) / (double)(real + 1)); 152 fprintf(out, "%13jd%c%02ld sys\t\t\t#\t%2.02f%% cpu\n", 153 (intmax_t)ru->ru_stime.tv_sec, decimal_point, 154 ru->ru_stime.tv_usec / 10000, 100 * (double)(sys + 1) / (double)(real + 1)); 155 } 156 157 static const char *stat_mode_cntrs[STAT_MODE_NPMCS]; 158 static const char *stat_mode_names[STAT_MODE_NPMCS]; 159 160 static void 161 pmc_stat_setup_stat(int system_mode, const char *arg) 162 { 163 const char *new_cntrs[STAT_MODE_NPMCS]; 164 struct pmcstat_ev *ev; 165 char *counters, *counter; 166 int i, c, start, newcnt; 167 cpuset_t cpumask, rootmask; 168 169 if (cpuset_getaffinity(CPU_LEVEL_ROOT, CPU_WHICH_PID, -1, 170 sizeof(rootmask), &rootmask) == -1) 171 err(EX_OSERR, "ERROR: Cannot determine the root set of CPUs"); 172 CPU_COPY(&rootmask, &cpumask); 173 174 if (system_mode && geteuid() != 0) 175 errx(EX_USAGE, "ERROR: system mode counters can only be used as root"); 176 counters = NULL; 177 for (i = 0; i < STAT_MODE_NPMCS; i++) { 178 stat_mode_cntrs[i] = pmc_stat_mode_aliases[i]; 179 stat_mode_names[i] = pmc_stat_mode_names[i]; 180 } 181 if (arg) { 182 counters = strdup(arg); 183 newcnt = 0; 184 while ((counter = strsep(&counters, ",")) != NULL && 185 newcnt < STAT_MODE_NPMCS - IAP_START) { 186 new_cntrs[newcnt++] = counter; 187 if (pmc_pmu_sample_rate_get(counter) == DEFAULT_SAMPLE_COUNT) 188 errx(EX_USAGE, "ERROR: %s not recognized on host", counter); 189 } 190 start = IAP_START + STAT_MODE_NPMCS - FIXED_MODE_NPMCS - newcnt; 191 for (i = 0; i < newcnt; i++) { 192 stat_mode_cntrs[start + i] = new_cntrs[i]; 193 stat_mode_names[start + i] = new_cntrs[i]; 194 } 195 } 196 if (system_mode) 197 pmc_args.pa_flags |= FLAG_HAS_SYSTEM_PMCS; 198 else 199 pmc_args.pa_flags |= FLAG_HAS_PROCESS_PMCS; 200 pmc_args.pa_flags |= FLAG_HAS_COUNTING_PMCS; 201 pmc_args.pa_flags |= FLAG_HAS_COMMANDLINE | FLAG_HAS_TARGET; 202 pmc_args.pa_flags |= FLAG_HAS_PIPE; 203 pmc_args.pa_required |= FLAG_HAS_COMMANDLINE | FLAG_HAS_TARGET | FLAG_HAS_OUTPUT_LOGFILE; 204 pmc_args.pa_outputpath = strdup("/dev/null"); 205 pmc_args.pa_logfd = pmcstat_open_log(pmc_args.pa_outputpath, 206 PMCSTAT_OPEN_FOR_WRITE); 207 for (i = 0; i < STAT_MODE_NPMCS; i++) { 208 if ((ev = malloc(sizeof(*ev))) == NULL) 209 errx(EX_SOFTWARE, "ERROR: Out of memory."); 210 if (system_mode) 211 ev->ev_mode = PMC_MODE_SC; 212 else 213 ev->ev_mode = PMC_MODE_TC; 214 ev->ev_spec = strdup(stat_mode_cntrs[i]); 215 if (ev->ev_spec == NULL) 216 errx(EX_SOFTWARE, "ERROR: Out of memory."); 217 c = strcspn(strdup(stat_mode_cntrs[i]), ", \t"); 218 ev->ev_name = malloc(c + 1); 219 if (ev->ev_name == NULL) 220 errx(EX_SOFTWARE, "ERROR: Out of memory."); 221 (void)strncpy(ev->ev_name, stat_mode_cntrs[i], c); 222 *(ev->ev_name + c) = '\0'; 223 224 ev->ev_count = -1; 225 ev->ev_flags = 0; 226 ev->ev_flags |= PMC_F_DESCENDANTS; 227 ev->ev_cumulative = 1; 228 229 ev->ev_saved = 0LL; 230 ev->ev_pmcid = PMC_ID_INVALID; 231 STAILQ_INSERT_TAIL(&pmc_args.pa_events, ev, ev_next); 232 if (system_mode) { 233 ev->ev_cpu = CPU_FFS(&cpumask) - 1; 234 CPU_CLR(ev->ev_cpu, &cpumask); 235 pmcstat_clone_event_descriptor(ev, &cpumask, &pmc_args); 236 CPU_SET(ev->ev_cpu, &cpumask); 237 } else 238 ev->ev_cpu = PMC_CPU_ANY; 239 240 } 241 if (clock_gettime(CLOCK_MONOTONIC, &before_ts)) 242 err(1, "clock_gettime"); 243 } 244 245 static void 246 pmc_stat_print_stat(struct rusage *ru) 247 { 248 struct pmcstat_ev *ev; 249 struct timespec after; 250 uint64_t cvals[STAT_MODE_NPMCS]; 251 uint64_t ticks, value; 252 int hz, i; 253 254 if (ru) { 255 hz = getstathz(); 256 ticks = hz * (ru->ru_utime.tv_sec + ru->ru_stime.tv_sec) + 257 hz * (ru->ru_utime.tv_usec + ru->ru_stime.tv_usec) / 1000000; 258 if (clock_gettime(CLOCK_MONOTONIC, &after)) 259 err(1, "clock_gettime"); 260 /* 261 * If our round-off on the tick calculation still puts us at 0, 262 * then always assume at least one tick. 263 */ 264 if (ticks == 0) 265 ticks = 1; 266 fprintf(pmc_args.pa_printfile, "%16ld %s\t\t#\t%02.03f M/sec\n", 267 ru->ru_minflt, "page faults", ((double)ru->ru_minflt / (double)ticks) / hz); 268 fprintf(pmc_args.pa_printfile, "%16ld %s\t\t#\t%02.03f M/sec\n", 269 ru->ru_nvcsw, "voluntary csw", ((double)ru->ru_nvcsw / (double)ticks) / hz); 270 fprintf(pmc_args.pa_printfile, "%16ld %s\t#\t%02.03f M/sec\n", 271 ru->ru_nivcsw, "involuntary csw", ((double)ru->ru_nivcsw / (double)ticks) / hz); 272 } 273 274 bzero(&cvals, sizeof(cvals)); 275 STAILQ_FOREACH(ev, &pmc_args.pa_events, ev_next) { 276 if (pmc_read(ev->ev_pmcid, &value) < 0) 277 err(EX_OSERR, "ERROR: Cannot read pmc \"%s\"", 278 ev->ev_name); 279 for (i = 0; i < STAT_MODE_NPMCS; i++) 280 if (strcmp(ev->ev_name, stat_mode_cntrs[i]) == 0) 281 cvals[i] += value; 282 } 283 284 fprintf(pmc_args.pa_printfile, "%16jd %s\n", (uintmax_t)cvals[CYCLES], stat_mode_names[CYCLES]); 285 fprintf(pmc_args.pa_printfile, "%16jd %s\t\t#\t%01.03f inst/cycle\n", (uintmax_t)cvals[INST], stat_mode_names[INST], 286 (double)cvals[INST] / cvals[CYCLES]); 287 fprintf(pmc_args.pa_printfile, "%16jd %s\n", (uintmax_t)cvals[BR], stat_mode_names[BR]); 288 if (stat_mode_names[BR_MISS] == pmc_stat_mode_names[BR_MISS]) 289 fprintf(pmc_args.pa_printfile, "%16jd %s\t\t#\t%.03f%%\n", 290 (uintmax_t)cvals[BR_MISS], stat_mode_names[BR_MISS], 291 100 * ((double)cvals[BR_MISS] / cvals[BR])); 292 else 293 fprintf(pmc_args.pa_printfile, "%16jd %s\n", 294 (uintmax_t)cvals[BR_MISS], stat_mode_names[BR_MISS]); 295 fprintf(pmc_args.pa_printfile, "%16jd %s%s", (uintmax_t)cvals[CACHE], stat_mode_names[CACHE], 296 stat_mode_names[CACHE] != pmc_stat_mode_names[CACHE] ? "\n" : ""); 297 if (stat_mode_names[CACHE] == pmc_stat_mode_names[CACHE]) 298 fprintf(pmc_args.pa_printfile, "\t#\t%.03f refs/inst\n", 299 ((double)cvals[CACHE] / cvals[INST])); 300 fprintf(pmc_args.pa_printfile, "%16jd %s%s", (uintmax_t)cvals[CACHE_MISS], stat_mode_names[CACHE_MISS], 301 stat_mode_names[CACHE_MISS] != pmc_stat_mode_names[CACHE_MISS] ? "\n" : ""); 302 if (stat_mode_names[CACHE_MISS] == pmc_stat_mode_names[CACHE_MISS]) 303 fprintf(pmc_args.pa_printfile, "\t\t#\t%.03f%%\n", 304 100 * ((double)cvals[CACHE_MISS] / cvals[CACHE])); 305 306 if (ru) 307 showtime(pmc_args.pa_printfile, &before_ts, &after, ru); 308 } 309 310 static struct option longopts[] = { 311 {"events", required_argument, NULL, 'j'}, 312 {NULL, 0, NULL, 0} 313 }; 314 315 static int 316 pmc_stat_internal(int argc, char **argv, int system_mode) 317 { 318 char *event, *r; 319 struct sigaction sa; 320 struct kevent kev; 321 struct rusage ru; 322 struct winsize ws; 323 struct pmcstat_ev *ev; 324 int c, option, runstate; 325 int waitstatus, ru_valid, do_debug; 326 327 do_debug = ru_valid = 0; 328 r = event = NULL; 329 while ((option = getopt_long(argc, argv, "dj:", longopts, NULL)) != -1) { 330 switch (option) { 331 case 'j': 332 r = event = strdup(optarg); 333 break; 334 case 'd': 335 do_debug = 1; 336 break; 337 case '?': 338 default: 339 usage(); 340 } 341 } 342 pmc_args.pa_argc = (argc -= optind); 343 pmc_args.pa_argv = (argv += optind); 344 if (argc == 0) 345 usage(); 346 pmc_args.pa_flags |= FLAG_HAS_COMMANDLINE; 347 pmc_stat_setup_stat(system_mode, event); 348 free(r); 349 bzero(&ru, sizeof(ru)); 350 EV_SET(&kev, SIGINT, EVFILT_SIGNAL, EV_ADD, 0, 0, NULL); 351 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 352 err(EX_OSERR, "ERROR: Cannot register kevent for SIGINT"); 353 354 EV_SET(&kev, SIGIO, EVFILT_SIGNAL, EV_ADD, 0, 0, NULL); 355 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 356 err(EX_OSERR, "ERROR: Cannot register kevent for SIGIO"); 357 EV_SET(&kev, 0, EVFILT_TIMER, EV_ADD, 0, 1000, NULL); 358 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 359 err(EX_OSERR, 360 "ERROR: Cannot register kevent for timer"); 361 362 STAILQ_FOREACH(ev, &pmc_args.pa_events, ev_next) { 363 if (pmc_allocate(ev->ev_spec, ev->ev_mode, 364 ev->ev_flags, ev->ev_cpu, &ev->ev_pmcid, ev->ev_count) < 0) 365 err(EX_OSERR, 366 "ERROR: Cannot allocate %s-mode pmc with specification \"%s\"", 367 PMC_IS_SYSTEM_MODE(ev->ev_mode) ? 368 "system" : "process", ev->ev_spec); 369 370 if (PMC_IS_SAMPLING_MODE(ev->ev_mode) && 371 pmc_set(ev->ev_pmcid, ev->ev_count) < 0) 372 err(EX_OSERR, 373 "ERROR: Cannot set sampling count for PMC \"%s\"", 374 ev->ev_name); 375 } 376 377 /* 378 * An exec() failure of a forked child is signalled by the 379 * child sending the parent a SIGCHLD. We don't register an 380 * actual signal handler for SIGCHLD, but instead use our 381 * kqueue to pick up the signal. 382 */ 383 EV_SET(&kev, SIGCHLD, EVFILT_SIGNAL, EV_ADD, 0, 0, NULL); 384 if (kevent(pmc_kq, &kev, 1, NULL, 0, NULL) < 0) 385 err(EX_OSERR, "ERROR: Cannot register kevent for SIGCHLD"); 386 387 pmcstat_create_process(pmcstat_sockpair, &pmc_args, pmc_kq); 388 389 if (SLIST_EMPTY(&pmc_args.pa_targets)) 390 errx(EX_DATAERR, 391 "ERROR: No matching target processes."); 392 if (pmc_args.pa_flags & FLAG_HAS_PROCESS_PMCS) 393 pmcstat_attach_pmcs(&pmc_args); 394 395 /* start the pmcs */ 396 pmc_util_start_pmcs(&pmc_args); 397 398 /* start the (commandline) process if needed */ 399 pmcstat_start_process(pmcstat_sockpair); 400 401 /* Handle SIGINT using the kqueue loop */ 402 sa.sa_handler = SIG_IGN; 403 sa.sa_flags = 0; 404 (void)sigemptyset(&sa.sa_mask); 405 406 if (sigaction(SIGINT, &sa, NULL) < 0) 407 err(EX_OSERR, "ERROR: Cannot install signal handler"); 408 409 /* 410 * loop till either the target process (if any) exits, or we 411 * are killed by a SIGINT or we reached the time duration. 412 */ 413 runstate = PMCSTAT_RUNNING; 414 do { 415 if ((c = kevent(pmc_kq, NULL, 0, &kev, 1, NULL)) <= 0) { 416 if (errno != EINTR) 417 err(EX_OSERR, "ERROR: kevent failed"); 418 else 419 continue; 420 } 421 if (kev.flags & EV_ERROR) 422 errc(EX_OSERR, kev.data, "ERROR: kevent failed"); 423 424 switch (kev.filter) { 425 case EVFILT_PROC: /* target has exited */ 426 if (wait4(pmc_util_get_pid(&pmc_args), &waitstatus, 0, &ru) > 0) { 427 getrusage(RUSAGE_CHILDREN, &ru); 428 ru_valid = 1; 429 } 430 break; 431 432 case EVFILT_READ: /* log file data is present */ 433 break; 434 case EVFILT_TIMER: 435 if (do_debug) 436 pmc_stat_print_stat(NULL); 437 break; 438 case EVFILT_SIGNAL: 439 if (kev.ident == SIGCHLD) { 440 /* 441 * The child process sends us a 442 * SIGCHLD if its exec() failed. We 443 * wait for it to exit and then exit 444 * ourselves. 445 */ 446 (void)wait(&c); 447 runstate = PMCSTAT_FINISHED; 448 } else if (kev.ident == SIGIO) { 449 /* 450 * We get a SIGIO if a PMC loses all 451 * of its targets, or if logfile 452 * writes encounter an error. 453 */ 454 if (wait4(pmc_util_get_pid(&pmc_args), &waitstatus, 0, &ru) > 0) { 455 getrusage(RUSAGE_CHILDREN, &ru); 456 ru_valid = 1; 457 } 458 runstate = pmcstat_close_log(&pmc_args); 459 } else if (kev.ident == SIGINT) { 460 /* Kill the child process if we started it */ 461 if (pmc_args.pa_flags & FLAG_HAS_COMMANDLINE) 462 pmc_util_kill_process(&pmc_args); 463 runstate = pmcstat_close_log(&pmc_args); 464 } else if (kev.ident == SIGWINCH) { 465 if (ioctl(fileno(pmc_args.pa_printfile), 466 TIOCGWINSZ, &ws) < 0) 467 err(EX_OSERR, 468 "ERROR: Cannot determine window size"); 469 pmc_displayheight = ws.ws_row - 1; 470 pmc_displaywidth = ws.ws_col - 1; 471 } else 472 assert(0); 473 474 break; 475 } 476 } while (runstate != PMCSTAT_FINISHED); 477 if (!ru_valid) 478 warnx("couldn't get rusage"); 479 pmc_stat_print_stat(&ru); 480 pmc_util_cleanup(&pmc_args); 481 return (0); 482 } 483 484 int 485 cmd_pmc_stat(int argc, char **argv) 486 { 487 return (pmc_stat_internal(argc, argv, 0)); 488 } 489 490 int 491 cmd_pmc_stat_system(int argc, char **argv) 492 { 493 return (pmc_stat_internal(argc, argv, 1)); 494 } 495