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