1 /* 2 * turbostat -- show CPU frequency and C-state residency 3 * on modern Intel turbo-capable processors. 4 * 5 * Copyright (c) 2013 Intel Corporation. 6 * Len Brown <len.brown@intel.com> 7 * 8 * This program is free software; you can redistribute it and/or modify it 9 * under the terms and conditions of the GNU General Public License, 10 * version 2, as published by the Free Software Foundation. 11 * 12 * This program is distributed in the hope it will be useful, but WITHOUT 13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 15 * more details. 16 * 17 * You should have received a copy of the GNU General Public License along with 18 * this program; if not, write to the Free Software Foundation, Inc., 19 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. 20 */ 21 22 #define _GNU_SOURCE 23 #include MSRHEADER 24 #include <stdarg.h> 25 #include <stdio.h> 26 #include <err.h> 27 #include <unistd.h> 28 #include <sys/types.h> 29 #include <sys/wait.h> 30 #include <sys/stat.h> 31 #include <sys/resource.h> 32 #include <fcntl.h> 33 #include <signal.h> 34 #include <sys/time.h> 35 #include <stdlib.h> 36 #include <dirent.h> 37 #include <string.h> 38 #include <ctype.h> 39 #include <sched.h> 40 #include <cpuid.h> 41 42 char *proc_stat = "/proc/stat"; 43 unsigned int interval_sec = 5; /* set with -i interval_sec */ 44 unsigned int verbose; /* set with -v */ 45 unsigned int rapl_verbose; /* set with -R */ 46 unsigned int thermal_verbose; /* set with -T */ 47 unsigned int summary_only; /* set with -s */ 48 unsigned int skip_c0; 49 unsigned int skip_c1; 50 unsigned int do_nhm_cstates; 51 unsigned int do_snb_cstates; 52 unsigned int do_c8_c9_c10; 53 unsigned int do_slm_cstates; 54 unsigned int use_c1_residency_msr; 55 unsigned int has_aperf; 56 unsigned int has_epb; 57 unsigned int units = 1000000000; /* Ghz etc */ 58 unsigned int genuine_intel; 59 unsigned int has_invariant_tsc; 60 unsigned int do_nehalem_platform_info; 61 unsigned int do_nehalem_turbo_ratio_limit; 62 unsigned int do_ivt_turbo_ratio_limit; 63 unsigned int extra_msr_offset32; 64 unsigned int extra_msr_offset64; 65 unsigned int extra_delta_offset32; 66 unsigned int extra_delta_offset64; 67 int do_smi; 68 double bclk; 69 unsigned int show_pkg; 70 unsigned int show_core; 71 unsigned int show_cpu; 72 unsigned int show_pkg_only; 73 unsigned int show_core_only; 74 char *output_buffer, *outp; 75 unsigned int do_rapl; 76 unsigned int do_dts; 77 unsigned int do_ptm; 78 unsigned int tcc_activation_temp; 79 unsigned int tcc_activation_temp_override; 80 double rapl_power_units, rapl_energy_units, rapl_time_units; 81 double rapl_joule_counter_range; 82 83 #define RAPL_PKG (1 << 0) 84 /* 0x610 MSR_PKG_POWER_LIMIT */ 85 /* 0x611 MSR_PKG_ENERGY_STATUS */ 86 #define RAPL_PKG_PERF_STATUS (1 << 1) 87 /* 0x613 MSR_PKG_PERF_STATUS */ 88 #define RAPL_PKG_POWER_INFO (1 << 2) 89 /* 0x614 MSR_PKG_POWER_INFO */ 90 91 #define RAPL_DRAM (1 << 3) 92 /* 0x618 MSR_DRAM_POWER_LIMIT */ 93 /* 0x619 MSR_DRAM_ENERGY_STATUS */ 94 /* 0x61c MSR_DRAM_POWER_INFO */ 95 #define RAPL_DRAM_PERF_STATUS (1 << 4) 96 /* 0x61b MSR_DRAM_PERF_STATUS */ 97 98 #define RAPL_CORES (1 << 5) 99 /* 0x638 MSR_PP0_POWER_LIMIT */ 100 /* 0x639 MSR_PP0_ENERGY_STATUS */ 101 #define RAPL_CORE_POLICY (1 << 6) 102 /* 0x63a MSR_PP0_POLICY */ 103 104 105 #define RAPL_GFX (1 << 7) 106 /* 0x640 MSR_PP1_POWER_LIMIT */ 107 /* 0x641 MSR_PP1_ENERGY_STATUS */ 108 /* 0x642 MSR_PP1_POLICY */ 109 #define TJMAX_DEFAULT 100 110 111 #define MAX(a, b) ((a) > (b) ? (a) : (b)) 112 113 int aperf_mperf_unstable; 114 int backwards_count; 115 char *progname; 116 117 cpu_set_t *cpu_present_set, *cpu_affinity_set; 118 size_t cpu_present_setsize, cpu_affinity_setsize; 119 120 struct thread_data { 121 unsigned long long tsc; 122 unsigned long long aperf; 123 unsigned long long mperf; 124 unsigned long long c1; 125 unsigned long long extra_msr64; 126 unsigned long long extra_delta64; 127 unsigned long long extra_msr32; 128 unsigned long long extra_delta32; 129 unsigned int smi_count; 130 unsigned int cpu_id; 131 unsigned int flags; 132 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2 133 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4 134 } *thread_even, *thread_odd; 135 136 struct core_data { 137 unsigned long long c3; 138 unsigned long long c6; 139 unsigned long long c7; 140 unsigned int core_temp_c; 141 unsigned int core_id; 142 } *core_even, *core_odd; 143 144 struct pkg_data { 145 unsigned long long pc2; 146 unsigned long long pc3; 147 unsigned long long pc6; 148 unsigned long long pc7; 149 unsigned long long pc8; 150 unsigned long long pc9; 151 unsigned long long pc10; 152 unsigned int package_id; 153 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */ 154 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */ 155 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */ 156 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */ 157 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */ 158 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */ 159 unsigned int pkg_temp_c; 160 161 } *package_even, *package_odd; 162 163 #define ODD_COUNTERS thread_odd, core_odd, package_odd 164 #define EVEN_COUNTERS thread_even, core_even, package_even 165 166 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \ 167 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \ 168 topo.num_threads_per_core + \ 169 (core_no) * topo.num_threads_per_core + (thread_no)) 170 #define GET_CORE(core_base, core_no, pkg_no) \ 171 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no)) 172 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no) 173 174 struct system_summary { 175 struct thread_data threads; 176 struct core_data cores; 177 struct pkg_data packages; 178 } sum, average; 179 180 181 struct topo_params { 182 int num_packages; 183 int num_cpus; 184 int num_cores; 185 int max_cpu_num; 186 int num_cores_per_pkg; 187 int num_threads_per_core; 188 } topo; 189 190 struct timeval tv_even, tv_odd, tv_delta; 191 192 void setup_all_buffers(void); 193 194 int cpu_is_not_present(int cpu) 195 { 196 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set); 197 } 198 /* 199 * run func(thread, core, package) in topology order 200 * skip non-present cpus 201 */ 202 203 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *), 204 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base) 205 { 206 int retval, pkg_no, core_no, thread_no; 207 208 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) { 209 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) { 210 for (thread_no = 0; thread_no < 211 topo.num_threads_per_core; ++thread_no) { 212 struct thread_data *t; 213 struct core_data *c; 214 struct pkg_data *p; 215 216 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no); 217 218 if (cpu_is_not_present(t->cpu_id)) 219 continue; 220 221 c = GET_CORE(core_base, core_no, pkg_no); 222 p = GET_PKG(pkg_base, pkg_no); 223 224 retval = func(t, c, p); 225 if (retval) 226 return retval; 227 } 228 } 229 } 230 return 0; 231 } 232 233 int cpu_migrate(int cpu) 234 { 235 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set); 236 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set); 237 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) 238 return -1; 239 else 240 return 0; 241 } 242 243 int get_msr(int cpu, off_t offset, unsigned long long *msr) 244 { 245 ssize_t retval; 246 char pathname[32]; 247 int fd; 248 249 sprintf(pathname, "/dev/cpu/%d/msr", cpu); 250 fd = open(pathname, O_RDONLY); 251 if (fd < 0) 252 return -1; 253 254 retval = pread(fd, msr, sizeof *msr, offset); 255 close(fd); 256 257 if (retval != sizeof *msr) { 258 fprintf(stderr, "%s offset 0x%llx read failed\n", pathname, (unsigned long long)offset); 259 return -1; 260 } 261 262 return 0; 263 } 264 265 void print_header(void) 266 { 267 if (show_pkg) 268 outp += sprintf(outp, "pk"); 269 if (show_pkg) 270 outp += sprintf(outp, " "); 271 if (show_core) 272 outp += sprintf(outp, "cor"); 273 if (show_cpu) 274 outp += sprintf(outp, " CPU"); 275 if (show_pkg || show_core || show_cpu) 276 outp += sprintf(outp, " "); 277 if (do_nhm_cstates) 278 outp += sprintf(outp, " %%c0"); 279 if (has_aperf) 280 outp += sprintf(outp, " GHz"); 281 outp += sprintf(outp, " TSC"); 282 if (do_smi) 283 outp += sprintf(outp, " SMI"); 284 if (extra_delta_offset32) 285 outp += sprintf(outp, " count 0x%03X", extra_delta_offset32); 286 if (extra_delta_offset64) 287 outp += sprintf(outp, " COUNT 0x%03X", extra_delta_offset64); 288 if (extra_msr_offset32) 289 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset32); 290 if (extra_msr_offset64) 291 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset64); 292 if (do_nhm_cstates) 293 outp += sprintf(outp, " %%c1"); 294 if (do_nhm_cstates && !do_slm_cstates) 295 outp += sprintf(outp, " %%c3"); 296 if (do_nhm_cstates) 297 outp += sprintf(outp, " %%c6"); 298 if (do_snb_cstates) 299 outp += sprintf(outp, " %%c7"); 300 301 if (do_dts) 302 outp += sprintf(outp, " CTMP"); 303 if (do_ptm) 304 outp += sprintf(outp, " PTMP"); 305 306 if (do_snb_cstates) 307 outp += sprintf(outp, " %%pc2"); 308 if (do_nhm_cstates && !do_slm_cstates) 309 outp += sprintf(outp, " %%pc3"); 310 if (do_nhm_cstates && !do_slm_cstates) 311 outp += sprintf(outp, " %%pc6"); 312 if (do_snb_cstates) 313 outp += sprintf(outp, " %%pc7"); 314 if (do_c8_c9_c10) { 315 outp += sprintf(outp, " %%pc8"); 316 outp += sprintf(outp, " %%pc9"); 317 outp += sprintf(outp, " %%pc10"); 318 } 319 320 if (do_rapl & RAPL_PKG) 321 outp += sprintf(outp, " Pkg_W"); 322 if (do_rapl & RAPL_CORES) 323 outp += sprintf(outp, " Cor_W"); 324 if (do_rapl & RAPL_GFX) 325 outp += sprintf(outp, " GFX_W"); 326 if (do_rapl & RAPL_DRAM) 327 outp += sprintf(outp, " RAM_W"); 328 if (do_rapl & RAPL_PKG_PERF_STATUS) 329 outp += sprintf(outp, " PKG_%%"); 330 if (do_rapl & RAPL_DRAM_PERF_STATUS) 331 outp += sprintf(outp, " RAM_%%"); 332 333 outp += sprintf(outp, "\n"); 334 } 335 336 int dump_counters(struct thread_data *t, struct core_data *c, 337 struct pkg_data *p) 338 { 339 fprintf(stderr, "t %p, c %p, p %p\n", t, c, p); 340 341 if (t) { 342 fprintf(stderr, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags); 343 fprintf(stderr, "TSC: %016llX\n", t->tsc); 344 fprintf(stderr, "aperf: %016llX\n", t->aperf); 345 fprintf(stderr, "mperf: %016llX\n", t->mperf); 346 fprintf(stderr, "c1: %016llX\n", t->c1); 347 fprintf(stderr, "msr0x%x: %08llX\n", 348 extra_delta_offset32, t->extra_delta32); 349 fprintf(stderr, "msr0x%x: %016llX\n", 350 extra_delta_offset64, t->extra_delta64); 351 fprintf(stderr, "msr0x%x: %08llX\n", 352 extra_msr_offset32, t->extra_msr32); 353 fprintf(stderr, "msr0x%x: %016llX\n", 354 extra_msr_offset64, t->extra_msr64); 355 if (do_smi) 356 fprintf(stderr, "SMI: %08X\n", t->smi_count); 357 } 358 359 if (c) { 360 fprintf(stderr, "core: %d\n", c->core_id); 361 fprintf(stderr, "c3: %016llX\n", c->c3); 362 fprintf(stderr, "c6: %016llX\n", c->c6); 363 fprintf(stderr, "c7: %016llX\n", c->c7); 364 fprintf(stderr, "DTS: %dC\n", c->core_temp_c); 365 } 366 367 if (p) { 368 fprintf(stderr, "package: %d\n", p->package_id); 369 fprintf(stderr, "pc2: %016llX\n", p->pc2); 370 fprintf(stderr, "pc3: %016llX\n", p->pc3); 371 fprintf(stderr, "pc6: %016llX\n", p->pc6); 372 fprintf(stderr, "pc7: %016llX\n", p->pc7); 373 fprintf(stderr, "pc8: %016llX\n", p->pc8); 374 fprintf(stderr, "pc9: %016llX\n", p->pc9); 375 fprintf(stderr, "pc10: %016llX\n", p->pc10); 376 fprintf(stderr, "Joules PKG: %0X\n", p->energy_pkg); 377 fprintf(stderr, "Joules COR: %0X\n", p->energy_cores); 378 fprintf(stderr, "Joules GFX: %0X\n", p->energy_gfx); 379 fprintf(stderr, "Joules RAM: %0X\n", p->energy_dram); 380 fprintf(stderr, "Throttle PKG: %0X\n", p->rapl_pkg_perf_status); 381 fprintf(stderr, "Throttle RAM: %0X\n", p->rapl_dram_perf_status); 382 fprintf(stderr, "PTM: %dC\n", p->pkg_temp_c); 383 } 384 return 0; 385 } 386 387 /* 388 * column formatting convention & formats 389 * package: "pk" 2 columns %2d 390 * core: "cor" 3 columns %3d 391 * CPU: "CPU" 3 columns %3d 392 * Pkg_W: %6.2 393 * Cor_W: %6.2 394 * GFX_W: %5.2 395 * RAM_W: %5.2 396 * GHz: "GHz" 3 columns %3.2 397 * TSC: "TSC" 3 columns %3.2 398 * SMI: "SMI" 4 columns %4d 399 * percentage " %pc3" %6.2 400 * Perf Status percentage: %5.2 401 * "CTMP" 4 columns %4d 402 */ 403 int format_counters(struct thread_data *t, struct core_data *c, 404 struct pkg_data *p) 405 { 406 double interval_float; 407 char *fmt5, *fmt6; 408 409 /* if showing only 1st thread in core and this isn't one, bail out */ 410 if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 411 return 0; 412 413 /* if showing only 1st thread in pkg and this isn't one, bail out */ 414 if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 415 return 0; 416 417 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0; 418 419 /* topo columns, print blanks on 1st (average) line */ 420 if (t == &average.threads) { 421 if (show_pkg) 422 outp += sprintf(outp, " "); 423 if (show_pkg && show_core) 424 outp += sprintf(outp, " "); 425 if (show_core) 426 outp += sprintf(outp, " "); 427 if (show_cpu) 428 outp += sprintf(outp, " " " "); 429 } else { 430 if (show_pkg) { 431 if (p) 432 outp += sprintf(outp, "%2d", p->package_id); 433 else 434 outp += sprintf(outp, " "); 435 } 436 if (show_pkg && show_core) 437 outp += sprintf(outp, " "); 438 if (show_core) { 439 if (c) 440 outp += sprintf(outp, "%3d", c->core_id); 441 else 442 outp += sprintf(outp, " "); 443 } 444 if (show_cpu) 445 outp += sprintf(outp, " %3d", t->cpu_id); 446 } 447 /* %c0 */ 448 if (do_nhm_cstates) { 449 if (show_pkg || show_core || show_cpu) 450 outp += sprintf(outp, " "); 451 if (!skip_c0) 452 outp += sprintf(outp, "%6.2f", 100.0 * t->mperf/t->tsc); 453 else 454 outp += sprintf(outp, " ****"); 455 } 456 457 /* GHz */ 458 if (has_aperf) { 459 if (!aperf_mperf_unstable) { 460 outp += sprintf(outp, " %3.2f", 461 1.0 * t->tsc / units * t->aperf / 462 t->mperf / interval_float); 463 } else { 464 if (t->aperf > t->tsc || t->mperf > t->tsc) { 465 outp += sprintf(outp, " ***"); 466 } else { 467 outp += sprintf(outp, "%3.1f*", 468 1.0 * t->tsc / 469 units * t->aperf / 470 t->mperf / interval_float); 471 } 472 } 473 } 474 475 /* TSC */ 476 outp += sprintf(outp, "%5.2f", 1.0 * t->tsc/units/interval_float); 477 478 /* SMI */ 479 if (do_smi) 480 outp += sprintf(outp, "%4d", t->smi_count); 481 482 /* delta */ 483 if (extra_delta_offset32) 484 outp += sprintf(outp, " %11llu", t->extra_delta32); 485 486 /* DELTA */ 487 if (extra_delta_offset64) 488 outp += sprintf(outp, " %11llu", t->extra_delta64); 489 /* msr */ 490 if (extra_msr_offset32) 491 outp += sprintf(outp, " 0x%08llx", t->extra_msr32); 492 493 /* MSR */ 494 if (extra_msr_offset64) 495 outp += sprintf(outp, " 0x%016llx", t->extra_msr64); 496 497 if (do_nhm_cstates) { 498 if (!skip_c1) 499 outp += sprintf(outp, " %6.2f", 100.0 * t->c1/t->tsc); 500 else 501 outp += sprintf(outp, " ****"); 502 } 503 504 /* print per-core data only for 1st thread in core */ 505 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 506 goto done; 507 508 if (do_nhm_cstates && !do_slm_cstates) 509 outp += sprintf(outp, " %6.2f", 100.0 * c->c3/t->tsc); 510 if (do_nhm_cstates) 511 outp += sprintf(outp, " %6.2f", 100.0 * c->c6/t->tsc); 512 if (do_snb_cstates) 513 outp += sprintf(outp, " %6.2f", 100.0 * c->c7/t->tsc); 514 515 if (do_dts) 516 outp += sprintf(outp, " %4d", c->core_temp_c); 517 518 /* print per-package data only for 1st core in package */ 519 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 520 goto done; 521 522 if (do_ptm) 523 outp += sprintf(outp, " %4d", p->pkg_temp_c); 524 525 if (do_snb_cstates) 526 outp += sprintf(outp, " %6.2f", 100.0 * p->pc2/t->tsc); 527 if (do_nhm_cstates && !do_slm_cstates) 528 outp += sprintf(outp, " %6.2f", 100.0 * p->pc3/t->tsc); 529 if (do_nhm_cstates && !do_slm_cstates) 530 outp += sprintf(outp, " %6.2f", 100.0 * p->pc6/t->tsc); 531 if (do_snb_cstates) 532 outp += sprintf(outp, " %6.2f", 100.0 * p->pc7/t->tsc); 533 if (do_c8_c9_c10) { 534 outp += sprintf(outp, " %6.2f", 100.0 * p->pc8/t->tsc); 535 outp += sprintf(outp, " %6.2f", 100.0 * p->pc9/t->tsc); 536 outp += sprintf(outp, " %6.2f", 100.0 * p->pc10/t->tsc); 537 } 538 539 /* 540 * If measurement interval exceeds minimum RAPL Joule Counter range, 541 * indicate that results are suspect by printing "**" in fraction place. 542 */ 543 if (interval_float < rapl_joule_counter_range) { 544 fmt5 = " %5.2f"; 545 fmt6 = " %6.2f"; 546 } else { 547 fmt5 = " %3.0f**"; 548 fmt6 = " %4.0f**"; 549 } 550 551 if (do_rapl & RAPL_PKG) 552 outp += sprintf(outp, fmt6, p->energy_pkg * rapl_energy_units / interval_float); 553 if (do_rapl & RAPL_CORES) 554 outp += sprintf(outp, fmt6, p->energy_cores * rapl_energy_units / interval_float); 555 if (do_rapl & RAPL_GFX) 556 outp += sprintf(outp, fmt5, p->energy_gfx * rapl_energy_units / interval_float); 557 if (do_rapl & RAPL_DRAM) 558 outp += sprintf(outp, fmt5, p->energy_dram * rapl_energy_units / interval_float); 559 if (do_rapl & RAPL_PKG_PERF_STATUS ) 560 outp += sprintf(outp, fmt5, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float); 561 if (do_rapl & RAPL_DRAM_PERF_STATUS ) 562 outp += sprintf(outp, fmt5, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float); 563 564 done: 565 outp += sprintf(outp, "\n"); 566 567 return 0; 568 } 569 570 void flush_stdout() 571 { 572 fputs(output_buffer, stdout); 573 fflush(stdout); 574 outp = output_buffer; 575 } 576 void flush_stderr() 577 { 578 fputs(output_buffer, stderr); 579 outp = output_buffer; 580 } 581 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 582 { 583 static int printed; 584 585 if (!printed || !summary_only) 586 print_header(); 587 588 if (topo.num_cpus > 1) 589 format_counters(&average.threads, &average.cores, 590 &average.packages); 591 592 printed = 1; 593 594 if (summary_only) 595 return; 596 597 for_all_cpus(format_counters, t, c, p); 598 } 599 600 #define DELTA_WRAP32(new, old) \ 601 if (new > old) { \ 602 old = new - old; \ 603 } else { \ 604 old = 0x100000000 + new - old; \ 605 } 606 607 void 608 delta_package(struct pkg_data *new, struct pkg_data *old) 609 { 610 old->pc2 = new->pc2 - old->pc2; 611 old->pc3 = new->pc3 - old->pc3; 612 old->pc6 = new->pc6 - old->pc6; 613 old->pc7 = new->pc7 - old->pc7; 614 old->pc8 = new->pc8 - old->pc8; 615 old->pc9 = new->pc9 - old->pc9; 616 old->pc10 = new->pc10 - old->pc10; 617 old->pkg_temp_c = new->pkg_temp_c; 618 619 DELTA_WRAP32(new->energy_pkg, old->energy_pkg); 620 DELTA_WRAP32(new->energy_cores, old->energy_cores); 621 DELTA_WRAP32(new->energy_gfx, old->energy_gfx); 622 DELTA_WRAP32(new->energy_dram, old->energy_dram); 623 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status); 624 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status); 625 } 626 627 void 628 delta_core(struct core_data *new, struct core_data *old) 629 { 630 old->c3 = new->c3 - old->c3; 631 old->c6 = new->c6 - old->c6; 632 old->c7 = new->c7 - old->c7; 633 old->core_temp_c = new->core_temp_c; 634 } 635 636 /* 637 * old = new - old 638 */ 639 void 640 delta_thread(struct thread_data *new, struct thread_data *old, 641 struct core_data *core_delta) 642 { 643 old->tsc = new->tsc - old->tsc; 644 645 /* check for TSC < 1 Mcycles over interval */ 646 if (old->tsc < (1000 * 1000)) 647 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n" 648 "You can disable all c-states by booting with \"idle=poll\"\n" 649 "or just the deep ones with \"processor.max_cstate=1\""); 650 651 old->c1 = new->c1 - old->c1; 652 653 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) { 654 old->aperf = new->aperf - old->aperf; 655 old->mperf = new->mperf - old->mperf; 656 } else { 657 658 if (!aperf_mperf_unstable) { 659 fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname); 660 fprintf(stderr, "* Frequency results do not cover entire interval *\n"); 661 fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n"); 662 663 aperf_mperf_unstable = 1; 664 } 665 /* 666 * mperf delta is likely a huge "positive" number 667 * can not use it for calculating c0 time 668 */ 669 skip_c0 = 1; 670 skip_c1 = 1; 671 } 672 673 674 if (use_c1_residency_msr) { 675 /* 676 * Some models have a dedicated C1 residency MSR, 677 * which should be more accurate than the derivation below. 678 */ 679 } else { 680 /* 681 * As counter collection is not atomic, 682 * it is possible for mperf's non-halted cycles + idle states 683 * to exceed TSC's all cycles: show c1 = 0% in that case. 684 */ 685 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc) 686 old->c1 = 0; 687 else { 688 /* normal case, derive c1 */ 689 old->c1 = old->tsc - old->mperf - core_delta->c3 690 - core_delta->c6 - core_delta->c7; 691 } 692 } 693 694 if (old->mperf == 0) { 695 if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id); 696 old->mperf = 1; /* divide by 0 protection */ 697 } 698 699 old->extra_delta32 = new->extra_delta32 - old->extra_delta32; 700 old->extra_delta32 &= 0xFFFFFFFF; 701 702 old->extra_delta64 = new->extra_delta64 - old->extra_delta64; 703 704 /* 705 * Extra MSR is just a snapshot, simply copy latest w/o subtracting 706 */ 707 old->extra_msr32 = new->extra_msr32; 708 old->extra_msr64 = new->extra_msr64; 709 710 if (do_smi) 711 old->smi_count = new->smi_count - old->smi_count; 712 } 713 714 int delta_cpu(struct thread_data *t, struct core_data *c, 715 struct pkg_data *p, struct thread_data *t2, 716 struct core_data *c2, struct pkg_data *p2) 717 { 718 /* calculate core delta only for 1st thread in core */ 719 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE) 720 delta_core(c, c2); 721 722 /* always calculate thread delta */ 723 delta_thread(t, t2, c2); /* c2 is core delta */ 724 725 /* calculate package delta only for 1st core in package */ 726 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) 727 delta_package(p, p2); 728 729 return 0; 730 } 731 732 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 733 { 734 t->tsc = 0; 735 t->aperf = 0; 736 t->mperf = 0; 737 t->c1 = 0; 738 739 t->smi_count = 0; 740 t->extra_delta32 = 0; 741 t->extra_delta64 = 0; 742 743 /* tells format_counters to dump all fields from this set */ 744 t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE; 745 746 c->c3 = 0; 747 c->c6 = 0; 748 c->c7 = 0; 749 c->core_temp_c = 0; 750 751 p->pc2 = 0; 752 p->pc3 = 0; 753 p->pc6 = 0; 754 p->pc7 = 0; 755 p->pc8 = 0; 756 p->pc9 = 0; 757 p->pc10 = 0; 758 759 p->energy_pkg = 0; 760 p->energy_dram = 0; 761 p->energy_cores = 0; 762 p->energy_gfx = 0; 763 p->rapl_pkg_perf_status = 0; 764 p->rapl_dram_perf_status = 0; 765 p->pkg_temp_c = 0; 766 } 767 int sum_counters(struct thread_data *t, struct core_data *c, 768 struct pkg_data *p) 769 { 770 average.threads.tsc += t->tsc; 771 average.threads.aperf += t->aperf; 772 average.threads.mperf += t->mperf; 773 average.threads.c1 += t->c1; 774 775 average.threads.extra_delta32 += t->extra_delta32; 776 average.threads.extra_delta64 += t->extra_delta64; 777 778 /* sum per-core values only for 1st thread in core */ 779 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 780 return 0; 781 782 average.cores.c3 += c->c3; 783 average.cores.c6 += c->c6; 784 average.cores.c7 += c->c7; 785 786 average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c); 787 788 /* sum per-pkg values only for 1st core in pkg */ 789 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 790 return 0; 791 792 average.packages.pc2 += p->pc2; 793 average.packages.pc3 += p->pc3; 794 average.packages.pc6 += p->pc6; 795 average.packages.pc7 += p->pc7; 796 average.packages.pc8 += p->pc8; 797 average.packages.pc9 += p->pc9; 798 average.packages.pc10 += p->pc10; 799 800 average.packages.energy_pkg += p->energy_pkg; 801 average.packages.energy_dram += p->energy_dram; 802 average.packages.energy_cores += p->energy_cores; 803 average.packages.energy_gfx += p->energy_gfx; 804 805 average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c); 806 807 average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status; 808 average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status; 809 return 0; 810 } 811 /* 812 * sum the counters for all cpus in the system 813 * compute the weighted average 814 */ 815 void compute_average(struct thread_data *t, struct core_data *c, 816 struct pkg_data *p) 817 { 818 clear_counters(&average.threads, &average.cores, &average.packages); 819 820 for_all_cpus(sum_counters, t, c, p); 821 822 average.threads.tsc /= topo.num_cpus; 823 average.threads.aperf /= topo.num_cpus; 824 average.threads.mperf /= topo.num_cpus; 825 average.threads.c1 /= topo.num_cpus; 826 827 average.threads.extra_delta32 /= topo.num_cpus; 828 average.threads.extra_delta32 &= 0xFFFFFFFF; 829 830 average.threads.extra_delta64 /= topo.num_cpus; 831 832 average.cores.c3 /= topo.num_cores; 833 average.cores.c6 /= topo.num_cores; 834 average.cores.c7 /= topo.num_cores; 835 836 average.packages.pc2 /= topo.num_packages; 837 average.packages.pc3 /= topo.num_packages; 838 average.packages.pc6 /= topo.num_packages; 839 average.packages.pc7 /= topo.num_packages; 840 841 average.packages.pc8 /= topo.num_packages; 842 average.packages.pc9 /= topo.num_packages; 843 average.packages.pc10 /= topo.num_packages; 844 } 845 846 static unsigned long long rdtsc(void) 847 { 848 unsigned int low, high; 849 850 asm volatile("rdtsc" : "=a" (low), "=d" (high)); 851 852 return low | ((unsigned long long)high) << 32; 853 } 854 855 856 /* 857 * get_counters(...) 858 * migrate to cpu 859 * acquire and record local counters for that cpu 860 */ 861 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 862 { 863 int cpu = t->cpu_id; 864 unsigned long long msr; 865 866 if (cpu_migrate(cpu)) { 867 fprintf(stderr, "Could not migrate to CPU %d\n", cpu); 868 return -1; 869 } 870 871 t->tsc = rdtsc(); /* we are running on local CPU of interest */ 872 873 if (has_aperf) { 874 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf)) 875 return -3; 876 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf)) 877 return -4; 878 } 879 880 if (do_smi) { 881 if (get_msr(cpu, MSR_SMI_COUNT, &msr)) 882 return -5; 883 t->smi_count = msr & 0xFFFFFFFF; 884 } 885 if (extra_delta_offset32) { 886 if (get_msr(cpu, extra_delta_offset32, &msr)) 887 return -5; 888 t->extra_delta32 = msr & 0xFFFFFFFF; 889 } 890 891 if (extra_delta_offset64) 892 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64)) 893 return -5; 894 895 if (extra_msr_offset32) { 896 if (get_msr(cpu, extra_msr_offset32, &msr)) 897 return -5; 898 t->extra_msr32 = msr & 0xFFFFFFFF; 899 } 900 901 if (extra_msr_offset64) 902 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64)) 903 return -5; 904 905 if (use_c1_residency_msr) { 906 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1)) 907 return -6; 908 } 909 910 /* collect core counters only for 1st thread in core */ 911 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 912 return 0; 913 914 if (do_nhm_cstates && !do_slm_cstates) { 915 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3)) 916 return -6; 917 } 918 919 if (do_nhm_cstates) { 920 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6)) 921 return -7; 922 } 923 924 if (do_snb_cstates) 925 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7)) 926 return -8; 927 928 if (do_dts) { 929 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr)) 930 return -9; 931 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F); 932 } 933 934 935 /* collect package counters only for 1st core in package */ 936 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 937 return 0; 938 939 if (do_nhm_cstates && !do_slm_cstates) { 940 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3)) 941 return -9; 942 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6)) 943 return -10; 944 } 945 if (do_snb_cstates) { 946 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2)) 947 return -11; 948 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7)) 949 return -12; 950 } 951 if (do_c8_c9_c10) { 952 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8)) 953 return -13; 954 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9)) 955 return -13; 956 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10)) 957 return -13; 958 } 959 if (do_rapl & RAPL_PKG) { 960 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr)) 961 return -13; 962 p->energy_pkg = msr & 0xFFFFFFFF; 963 } 964 if (do_rapl & RAPL_CORES) { 965 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr)) 966 return -14; 967 p->energy_cores = msr & 0xFFFFFFFF; 968 } 969 if (do_rapl & RAPL_DRAM) { 970 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr)) 971 return -15; 972 p->energy_dram = msr & 0xFFFFFFFF; 973 } 974 if (do_rapl & RAPL_GFX) { 975 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr)) 976 return -16; 977 p->energy_gfx = msr & 0xFFFFFFFF; 978 } 979 if (do_rapl & RAPL_PKG_PERF_STATUS) { 980 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr)) 981 return -16; 982 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF; 983 } 984 if (do_rapl & RAPL_DRAM_PERF_STATUS) { 985 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr)) 986 return -16; 987 p->rapl_dram_perf_status = msr & 0xFFFFFFFF; 988 } 989 if (do_ptm) { 990 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr)) 991 return -17; 992 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F); 993 } 994 return 0; 995 } 996 997 void print_verbose_header(void) 998 { 999 unsigned long long msr; 1000 unsigned int ratio; 1001 1002 if (!do_nehalem_platform_info) 1003 return; 1004 1005 get_msr(0, MSR_NHM_PLATFORM_INFO, &msr); 1006 1007 fprintf(stderr, "cpu0: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", msr); 1008 1009 ratio = (msr >> 40) & 0xFF; 1010 fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n", 1011 ratio, bclk, ratio * bclk); 1012 1013 ratio = (msr >> 8) & 0xFF; 1014 fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n", 1015 ratio, bclk, ratio * bclk); 1016 1017 get_msr(0, MSR_IA32_POWER_CTL, &msr); 1018 fprintf(stderr, "cpu0: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n", 1019 msr, msr & 0x2 ? "EN" : "DIS"); 1020 1021 if (!do_ivt_turbo_ratio_limit) 1022 goto print_nhm_turbo_ratio_limits; 1023 1024 get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr); 1025 1026 fprintf(stderr, "cpu0: MSR_IVT_TURBO_RATIO_LIMIT: 0x%08llx\n", msr); 1027 1028 ratio = (msr >> 56) & 0xFF; 1029 if (ratio) 1030 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n", 1031 ratio, bclk, ratio * bclk); 1032 1033 ratio = (msr >> 48) & 0xFF; 1034 if (ratio) 1035 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n", 1036 ratio, bclk, ratio * bclk); 1037 1038 ratio = (msr >> 40) & 0xFF; 1039 if (ratio) 1040 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n", 1041 ratio, bclk, ratio * bclk); 1042 1043 ratio = (msr >> 32) & 0xFF; 1044 if (ratio) 1045 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n", 1046 ratio, bclk, ratio * bclk); 1047 1048 ratio = (msr >> 24) & 0xFF; 1049 if (ratio) 1050 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n", 1051 ratio, bclk, ratio * bclk); 1052 1053 ratio = (msr >> 16) & 0xFF; 1054 if (ratio) 1055 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n", 1056 ratio, bclk, ratio * bclk); 1057 1058 ratio = (msr >> 8) & 0xFF; 1059 if (ratio) 1060 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n", 1061 ratio, bclk, ratio * bclk); 1062 1063 ratio = (msr >> 0) & 0xFF; 1064 if (ratio) 1065 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n", 1066 ratio, bclk, ratio * bclk); 1067 1068 print_nhm_turbo_ratio_limits: 1069 get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr); 1070 1071 #define SNB_C1_AUTO_UNDEMOTE (1UL << 27) 1072 #define SNB_C3_AUTO_UNDEMOTE (1UL << 28) 1073 1074 fprintf(stderr, "cpu0: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", msr); 1075 1076 fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: ", 1077 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "", 1078 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "", 1079 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "", 1080 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "", 1081 (msr & (1 << 15)) ? "" : "UN", 1082 (unsigned int)msr & 7); 1083 1084 1085 switch(msr & 0x7) { 1086 case 0: 1087 fprintf(stderr, do_slm_cstates ? "no pkg states" : "pc0"); 1088 break; 1089 case 1: 1090 fprintf(stderr, do_slm_cstates ? "no pkg states" : do_snb_cstates ? "pc2" : "pc0"); 1091 break; 1092 case 2: 1093 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc6-noret" : "pc3"); 1094 break; 1095 case 3: 1096 fprintf(stderr, do_slm_cstates ? "invalid" : "pc6"); 1097 break; 1098 case 4: 1099 fprintf(stderr, do_slm_cstates ? "pc4" : "pc7"); 1100 break; 1101 case 5: 1102 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc7s" : "invalid"); 1103 break; 1104 case 6: 1105 fprintf(stderr, do_slm_cstates ? "pc6" : "invalid"); 1106 break; 1107 case 7: 1108 fprintf(stderr, do_slm_cstates ? "pc7" : "unlimited"); 1109 break; 1110 default: 1111 fprintf(stderr, "invalid"); 1112 } 1113 fprintf(stderr, ")\n"); 1114 1115 if (!do_nehalem_turbo_ratio_limit) 1116 return; 1117 1118 get_msr(0, MSR_NHM_TURBO_RATIO_LIMIT, &msr); 1119 1120 fprintf(stderr, "cpu0: MSR_NHM_TURBO_RATIO_LIMIT: 0x%08llx\n", msr); 1121 1122 ratio = (msr >> 56) & 0xFF; 1123 if (ratio) 1124 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n", 1125 ratio, bclk, ratio * bclk); 1126 1127 ratio = (msr >> 48) & 0xFF; 1128 if (ratio) 1129 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n", 1130 ratio, bclk, ratio * bclk); 1131 1132 ratio = (msr >> 40) & 0xFF; 1133 if (ratio) 1134 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n", 1135 ratio, bclk, ratio * bclk); 1136 1137 ratio = (msr >> 32) & 0xFF; 1138 if (ratio) 1139 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n", 1140 ratio, bclk, ratio * bclk); 1141 1142 ratio = (msr >> 24) & 0xFF; 1143 if (ratio) 1144 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n", 1145 ratio, bclk, ratio * bclk); 1146 1147 ratio = (msr >> 16) & 0xFF; 1148 if (ratio) 1149 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n", 1150 ratio, bclk, ratio * bclk); 1151 1152 ratio = (msr >> 8) & 0xFF; 1153 if (ratio) 1154 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n", 1155 ratio, bclk, ratio * bclk); 1156 1157 ratio = (msr >> 0) & 0xFF; 1158 if (ratio) 1159 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n", 1160 ratio, bclk, ratio * bclk); 1161 } 1162 1163 void free_all_buffers(void) 1164 { 1165 CPU_FREE(cpu_present_set); 1166 cpu_present_set = NULL; 1167 cpu_present_set = 0; 1168 1169 CPU_FREE(cpu_affinity_set); 1170 cpu_affinity_set = NULL; 1171 cpu_affinity_setsize = 0; 1172 1173 free(thread_even); 1174 free(core_even); 1175 free(package_even); 1176 1177 thread_even = NULL; 1178 core_even = NULL; 1179 package_even = NULL; 1180 1181 free(thread_odd); 1182 free(core_odd); 1183 free(package_odd); 1184 1185 thread_odd = NULL; 1186 core_odd = NULL; 1187 package_odd = NULL; 1188 1189 free(output_buffer); 1190 output_buffer = NULL; 1191 outp = NULL; 1192 } 1193 1194 /* 1195 * Open a file, and exit on failure 1196 */ 1197 FILE *fopen_or_die(const char *path, const char *mode) 1198 { 1199 FILE *filep = fopen(path, "r"); 1200 if (!filep) 1201 err(1, "%s: open failed", path); 1202 return filep; 1203 } 1204 1205 /* 1206 * Parse a file containing a single int. 1207 */ 1208 int parse_int_file(const char *fmt, ...) 1209 { 1210 va_list args; 1211 char path[PATH_MAX]; 1212 FILE *filep; 1213 int value; 1214 1215 va_start(args, fmt); 1216 vsnprintf(path, sizeof(path), fmt, args); 1217 va_end(args); 1218 filep = fopen_or_die(path, "r"); 1219 if (fscanf(filep, "%d", &value) != 1) 1220 err(1, "%s: failed to parse number from file", path); 1221 fclose(filep); 1222 return value; 1223 } 1224 1225 /* 1226 * cpu_is_first_sibling_in_core(cpu) 1227 * return 1 if given CPU is 1st HT sibling in the core 1228 */ 1229 int cpu_is_first_sibling_in_core(int cpu) 1230 { 1231 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu); 1232 } 1233 1234 /* 1235 * cpu_is_first_core_in_package(cpu) 1236 * return 1 if given CPU is 1st core in package 1237 */ 1238 int cpu_is_first_core_in_package(int cpu) 1239 { 1240 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu); 1241 } 1242 1243 int get_physical_package_id(int cpu) 1244 { 1245 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu); 1246 } 1247 1248 int get_core_id(int cpu) 1249 { 1250 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu); 1251 } 1252 1253 int get_num_ht_siblings(int cpu) 1254 { 1255 char path[80]; 1256 FILE *filep; 1257 int sib1, sib2; 1258 int matches; 1259 char character; 1260 1261 sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu); 1262 filep = fopen_or_die(path, "r"); 1263 /* 1264 * file format: 1265 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4) 1266 * otherwinse 1 sibling (self). 1267 */ 1268 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2); 1269 1270 fclose(filep); 1271 1272 if (matches == 3) 1273 return 2; 1274 else 1275 return 1; 1276 } 1277 1278 /* 1279 * run func(thread, core, package) in topology order 1280 * skip non-present cpus 1281 */ 1282 1283 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *, 1284 struct pkg_data *, struct thread_data *, struct core_data *, 1285 struct pkg_data *), struct thread_data *thread_base, 1286 struct core_data *core_base, struct pkg_data *pkg_base, 1287 struct thread_data *thread_base2, struct core_data *core_base2, 1288 struct pkg_data *pkg_base2) 1289 { 1290 int retval, pkg_no, core_no, thread_no; 1291 1292 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) { 1293 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) { 1294 for (thread_no = 0; thread_no < 1295 topo.num_threads_per_core; ++thread_no) { 1296 struct thread_data *t, *t2; 1297 struct core_data *c, *c2; 1298 struct pkg_data *p, *p2; 1299 1300 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no); 1301 1302 if (cpu_is_not_present(t->cpu_id)) 1303 continue; 1304 1305 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no); 1306 1307 c = GET_CORE(core_base, core_no, pkg_no); 1308 c2 = GET_CORE(core_base2, core_no, pkg_no); 1309 1310 p = GET_PKG(pkg_base, pkg_no); 1311 p2 = GET_PKG(pkg_base2, pkg_no); 1312 1313 retval = func(t, c, p, t2, c2, p2); 1314 if (retval) 1315 return retval; 1316 } 1317 } 1318 } 1319 return 0; 1320 } 1321 1322 /* 1323 * run func(cpu) on every cpu in /proc/stat 1324 * return max_cpu number 1325 */ 1326 int for_all_proc_cpus(int (func)(int)) 1327 { 1328 FILE *fp; 1329 int cpu_num; 1330 int retval; 1331 1332 fp = fopen_or_die(proc_stat, "r"); 1333 1334 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n"); 1335 if (retval != 0) 1336 err(1, "%s: failed to parse format", proc_stat); 1337 1338 while (1) { 1339 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num); 1340 if (retval != 1) 1341 break; 1342 1343 retval = func(cpu_num); 1344 if (retval) { 1345 fclose(fp); 1346 return(retval); 1347 } 1348 } 1349 fclose(fp); 1350 return 0; 1351 } 1352 1353 void re_initialize(void) 1354 { 1355 free_all_buffers(); 1356 setup_all_buffers(); 1357 printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus); 1358 } 1359 1360 1361 /* 1362 * count_cpus() 1363 * remember the last one seen, it will be the max 1364 */ 1365 int count_cpus(int cpu) 1366 { 1367 if (topo.max_cpu_num < cpu) 1368 topo.max_cpu_num = cpu; 1369 1370 topo.num_cpus += 1; 1371 return 0; 1372 } 1373 int mark_cpu_present(int cpu) 1374 { 1375 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set); 1376 return 0; 1377 } 1378 1379 void turbostat_loop() 1380 { 1381 int retval; 1382 int restarted = 0; 1383 1384 restart: 1385 restarted++; 1386 1387 retval = for_all_cpus(get_counters, EVEN_COUNTERS); 1388 if (retval < -1) { 1389 exit(retval); 1390 } else if (retval == -1) { 1391 if (restarted > 1) { 1392 exit(retval); 1393 } 1394 re_initialize(); 1395 goto restart; 1396 } 1397 restarted = 0; 1398 gettimeofday(&tv_even, (struct timezone *)NULL); 1399 1400 while (1) { 1401 if (for_all_proc_cpus(cpu_is_not_present)) { 1402 re_initialize(); 1403 goto restart; 1404 } 1405 sleep(interval_sec); 1406 retval = for_all_cpus(get_counters, ODD_COUNTERS); 1407 if (retval < -1) { 1408 exit(retval); 1409 } else if (retval == -1) { 1410 re_initialize(); 1411 goto restart; 1412 } 1413 gettimeofday(&tv_odd, (struct timezone *)NULL); 1414 timersub(&tv_odd, &tv_even, &tv_delta); 1415 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS); 1416 compute_average(EVEN_COUNTERS); 1417 format_all_counters(EVEN_COUNTERS); 1418 flush_stdout(); 1419 sleep(interval_sec); 1420 retval = for_all_cpus(get_counters, EVEN_COUNTERS); 1421 if (retval < -1) { 1422 exit(retval); 1423 } else if (retval == -1) { 1424 re_initialize(); 1425 goto restart; 1426 } 1427 gettimeofday(&tv_even, (struct timezone *)NULL); 1428 timersub(&tv_even, &tv_odd, &tv_delta); 1429 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS); 1430 compute_average(ODD_COUNTERS); 1431 format_all_counters(ODD_COUNTERS); 1432 flush_stdout(); 1433 } 1434 } 1435 1436 void check_dev_msr() 1437 { 1438 struct stat sb; 1439 1440 if (stat("/dev/cpu/0/msr", &sb)) 1441 err(-5, "no /dev/cpu/0/msr\n" 1442 "Try \"# modprobe msr\""); 1443 } 1444 1445 void check_super_user() 1446 { 1447 if (getuid() != 0) 1448 errx(-6, "must be root"); 1449 } 1450 1451 int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model) 1452 { 1453 if (!genuine_intel) 1454 return 0; 1455 1456 if (family != 6) 1457 return 0; 1458 1459 switch (model) { 1460 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */ 1461 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */ 1462 case 0x1F: /* Core i7 and i5 Processor - Nehalem */ 1463 case 0x25: /* Westmere Client - Clarkdale, Arrandale */ 1464 case 0x2C: /* Westmere EP - Gulftown */ 1465 case 0x2A: /* SNB */ 1466 case 0x2D: /* SNB Xeon */ 1467 case 0x3A: /* IVB */ 1468 case 0x3E: /* IVB Xeon */ 1469 case 0x3C: /* HSW */ 1470 case 0x3F: /* HSX */ 1471 case 0x45: /* HSW */ 1472 case 0x46: /* HSW */ 1473 case 0x37: /* BYT */ 1474 case 0x4D: /* AVN */ 1475 return 1; 1476 case 0x2E: /* Nehalem-EX Xeon - Beckton */ 1477 case 0x2F: /* Westmere-EX Xeon - Eagleton */ 1478 default: 1479 return 0; 1480 } 1481 } 1482 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model) 1483 { 1484 if (!genuine_intel) 1485 return 0; 1486 1487 if (family != 6) 1488 return 0; 1489 1490 switch (model) { 1491 case 0x3E: /* IVB Xeon */ 1492 return 1; 1493 default: 1494 return 0; 1495 } 1496 } 1497 1498 /* 1499 * print_epb() 1500 * Decode the ENERGY_PERF_BIAS MSR 1501 */ 1502 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p) 1503 { 1504 unsigned long long msr; 1505 char *epb_string; 1506 int cpu; 1507 1508 if (!has_epb) 1509 return 0; 1510 1511 cpu = t->cpu_id; 1512 1513 /* EPB is per-package */ 1514 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 1515 return 0; 1516 1517 if (cpu_migrate(cpu)) { 1518 fprintf(stderr, "Could not migrate to CPU %d\n", cpu); 1519 return -1; 1520 } 1521 1522 if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr)) 1523 return 0; 1524 1525 switch (msr & 0x7) { 1526 case ENERGY_PERF_BIAS_PERFORMANCE: 1527 epb_string = "performance"; 1528 break; 1529 case ENERGY_PERF_BIAS_NORMAL: 1530 epb_string = "balanced"; 1531 break; 1532 case ENERGY_PERF_BIAS_POWERSAVE: 1533 epb_string = "powersave"; 1534 break; 1535 default: 1536 epb_string = "custom"; 1537 break; 1538 } 1539 fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string); 1540 1541 return 0; 1542 } 1543 1544 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */ 1545 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */ 1546 1547 double get_tdp(model) 1548 { 1549 unsigned long long msr; 1550 1551 if (do_rapl & RAPL_PKG_POWER_INFO) 1552 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr)) 1553 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units; 1554 1555 switch (model) { 1556 case 0x37: 1557 case 0x4D: 1558 return 30.0; 1559 default: 1560 return 135.0; 1561 } 1562 } 1563 1564 1565 /* 1566 * rapl_probe() 1567 * 1568 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units 1569 */ 1570 void rapl_probe(unsigned int family, unsigned int model) 1571 { 1572 unsigned long long msr; 1573 unsigned int time_unit; 1574 double tdp; 1575 1576 if (!genuine_intel) 1577 return; 1578 1579 if (family != 6) 1580 return; 1581 1582 switch (model) { 1583 case 0x2A: 1584 case 0x3A: 1585 case 0x3C: /* HSW */ 1586 case 0x45: /* HSW */ 1587 case 0x46: /* HSW */ 1588 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO; 1589 break; 1590 case 0x3F: /* HSX */ 1591 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO; 1592 break; 1593 case 0x2D: 1594 case 0x3E: 1595 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO; 1596 break; 1597 case 0x37: /* BYT */ 1598 case 0x4D: /* AVN */ 1599 do_rapl = RAPL_PKG | RAPL_CORES ; 1600 break; 1601 default: 1602 return; 1603 } 1604 1605 /* units on package 0, verify later other packages match */ 1606 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr)) 1607 return; 1608 1609 rapl_power_units = 1.0 / (1 << (msr & 0xF)); 1610 if (model == 0x37) 1611 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000; 1612 else 1613 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F)); 1614 1615 time_unit = msr >> 16 & 0xF; 1616 if (time_unit == 0) 1617 time_unit = 0xA; 1618 1619 rapl_time_units = 1.0 / (1 << (time_unit)); 1620 1621 tdp = get_tdp(model); 1622 1623 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp; 1624 if (verbose) 1625 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp); 1626 1627 return; 1628 } 1629 1630 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p) 1631 { 1632 unsigned long long msr; 1633 unsigned int dts; 1634 int cpu; 1635 1636 if (!(do_dts || do_ptm)) 1637 return 0; 1638 1639 cpu = t->cpu_id; 1640 1641 /* DTS is per-core, no need to print for each thread */ 1642 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 1643 return 0; 1644 1645 if (cpu_migrate(cpu)) { 1646 fprintf(stderr, "Could not migrate to CPU %d\n", cpu); 1647 return -1; 1648 } 1649 1650 if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) { 1651 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr)) 1652 return 0; 1653 1654 dts = (msr >> 16) & 0x7F; 1655 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n", 1656 cpu, msr, tcc_activation_temp - dts); 1657 1658 #ifdef THERM_DEBUG 1659 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr)) 1660 return 0; 1661 1662 dts = (msr >> 16) & 0x7F; 1663 dts2 = (msr >> 8) & 0x7F; 1664 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n", 1665 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2); 1666 #endif 1667 } 1668 1669 1670 if (do_dts) { 1671 unsigned int resolution; 1672 1673 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr)) 1674 return 0; 1675 1676 dts = (msr >> 16) & 0x7F; 1677 resolution = (msr >> 27) & 0xF; 1678 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n", 1679 cpu, msr, tcc_activation_temp - dts, resolution); 1680 1681 #ifdef THERM_DEBUG 1682 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr)) 1683 return 0; 1684 1685 dts = (msr >> 16) & 0x7F; 1686 dts2 = (msr >> 8) & 0x7F; 1687 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n", 1688 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2); 1689 #endif 1690 } 1691 1692 return 0; 1693 } 1694 1695 void print_power_limit_msr(int cpu, unsigned long long msr, char *label) 1696 { 1697 fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n", 1698 cpu, label, 1699 ((msr >> 15) & 1) ? "EN" : "DIS", 1700 ((msr >> 0) & 0x7FFF) * rapl_power_units, 1701 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units, 1702 (((msr >> 16) & 1) ? "EN" : "DIS")); 1703 1704 return; 1705 } 1706 1707 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p) 1708 { 1709 unsigned long long msr; 1710 int cpu; 1711 1712 if (!do_rapl) 1713 return 0; 1714 1715 /* RAPL counters are per package, so print only for 1st thread/package */ 1716 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 1717 return 0; 1718 1719 cpu = t->cpu_id; 1720 if (cpu_migrate(cpu)) { 1721 fprintf(stderr, "Could not migrate to CPU %d\n", cpu); 1722 return -1; 1723 } 1724 1725 if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr)) 1726 return -1; 1727 1728 if (verbose) { 1729 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx " 1730 "(%f Watts, %f Joules, %f sec.)\n", cpu, msr, 1731 rapl_power_units, rapl_energy_units, rapl_time_units); 1732 } 1733 if (do_rapl & RAPL_PKG_POWER_INFO) { 1734 1735 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr)) 1736 return -5; 1737 1738 1739 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n", 1740 cpu, msr, 1741 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units, 1742 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units, 1743 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units, 1744 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units); 1745 1746 } 1747 if (do_rapl & RAPL_PKG) { 1748 1749 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr)) 1750 return -9; 1751 1752 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n", 1753 cpu, msr, (msr >> 63) & 1 ? "": "UN"); 1754 1755 print_power_limit_msr(cpu, msr, "PKG Limit #1"); 1756 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n", 1757 cpu, 1758 ((msr >> 47) & 1) ? "EN" : "DIS", 1759 ((msr >> 32) & 0x7FFF) * rapl_power_units, 1760 (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units, 1761 ((msr >> 48) & 1) ? "EN" : "DIS"); 1762 } 1763 1764 if (do_rapl & RAPL_DRAM) { 1765 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr)) 1766 return -6; 1767 1768 1769 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n", 1770 cpu, msr, 1771 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units, 1772 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units, 1773 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units, 1774 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units); 1775 1776 1777 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr)) 1778 return -9; 1779 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n", 1780 cpu, msr, (msr >> 31) & 1 ? "": "UN"); 1781 1782 print_power_limit_msr(cpu, msr, "DRAM Limit"); 1783 } 1784 if (do_rapl & RAPL_CORE_POLICY) { 1785 if (verbose) { 1786 if (get_msr(cpu, MSR_PP0_POLICY, &msr)) 1787 return -7; 1788 1789 fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF); 1790 } 1791 } 1792 if (do_rapl & RAPL_CORES) { 1793 if (verbose) { 1794 1795 if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr)) 1796 return -9; 1797 fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n", 1798 cpu, msr, (msr >> 31) & 1 ? "": "UN"); 1799 print_power_limit_msr(cpu, msr, "Cores Limit"); 1800 } 1801 } 1802 if (do_rapl & RAPL_GFX) { 1803 if (verbose) { 1804 if (get_msr(cpu, MSR_PP1_POLICY, &msr)) 1805 return -8; 1806 1807 fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF); 1808 1809 if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr)) 1810 return -9; 1811 fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n", 1812 cpu, msr, (msr >> 31) & 1 ? "": "UN"); 1813 print_power_limit_msr(cpu, msr, "GFX Limit"); 1814 } 1815 } 1816 return 0; 1817 } 1818 1819 1820 int is_snb(unsigned int family, unsigned int model) 1821 { 1822 if (!genuine_intel) 1823 return 0; 1824 1825 switch (model) { 1826 case 0x2A: 1827 case 0x2D: 1828 case 0x3A: /* IVB */ 1829 case 0x3E: /* IVB Xeon */ 1830 case 0x3C: /* HSW */ 1831 case 0x3F: /* HSW */ 1832 case 0x45: /* HSW */ 1833 case 0x46: /* HSW */ 1834 return 1; 1835 } 1836 return 0; 1837 } 1838 1839 int has_c8_c9_c10(unsigned int family, unsigned int model) 1840 { 1841 if (!genuine_intel) 1842 return 0; 1843 1844 switch (model) { 1845 case 0x45: 1846 return 1; 1847 } 1848 return 0; 1849 } 1850 1851 1852 int is_slm(unsigned int family, unsigned int model) 1853 { 1854 if (!genuine_intel) 1855 return 0; 1856 switch (model) { 1857 case 0x37: /* BYT */ 1858 case 0x4D: /* AVN */ 1859 return 1; 1860 } 1861 return 0; 1862 } 1863 1864 #define SLM_BCLK_FREQS 5 1865 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0}; 1866 1867 double slm_bclk(void) 1868 { 1869 unsigned long long msr = 3; 1870 unsigned int i; 1871 double freq; 1872 1873 if (get_msr(0, MSR_FSB_FREQ, &msr)) 1874 fprintf(stderr, "SLM BCLK: unknown\n"); 1875 1876 i = msr & 0xf; 1877 if (i >= SLM_BCLK_FREQS) { 1878 fprintf(stderr, "SLM BCLK[%d] invalid\n", i); 1879 msr = 3; 1880 } 1881 freq = slm_freq_table[i]; 1882 1883 fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq); 1884 1885 return freq; 1886 } 1887 1888 double discover_bclk(unsigned int family, unsigned int model) 1889 { 1890 if (is_snb(family, model)) 1891 return 100.00; 1892 else if (is_slm(family, model)) 1893 return slm_bclk(); 1894 else 1895 return 133.33; 1896 } 1897 1898 /* 1899 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where 1900 * the Thermal Control Circuit (TCC) activates. 1901 * This is usually equal to tjMax. 1902 * 1903 * Older processors do not have this MSR, so there we guess, 1904 * but also allow cmdline over-ride with -T. 1905 * 1906 * Several MSR temperature values are in units of degrees-C 1907 * below this value, including the Digital Thermal Sensor (DTS), 1908 * Package Thermal Management Sensor (PTM), and thermal event thresholds. 1909 */ 1910 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p) 1911 { 1912 unsigned long long msr; 1913 unsigned int target_c_local; 1914 int cpu; 1915 1916 /* tcc_activation_temp is used only for dts or ptm */ 1917 if (!(do_dts || do_ptm)) 1918 return 0; 1919 1920 /* this is a per-package concept */ 1921 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 1922 return 0; 1923 1924 cpu = t->cpu_id; 1925 if (cpu_migrate(cpu)) { 1926 fprintf(stderr, "Could not migrate to CPU %d\n", cpu); 1927 return -1; 1928 } 1929 1930 if (tcc_activation_temp_override != 0) { 1931 tcc_activation_temp = tcc_activation_temp_override; 1932 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n", 1933 cpu, tcc_activation_temp); 1934 return 0; 1935 } 1936 1937 /* Temperature Target MSR is Nehalem and newer only */ 1938 if (!do_nehalem_platform_info) 1939 goto guess; 1940 1941 if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr)) 1942 goto guess; 1943 1944 target_c_local = (msr >> 16) & 0x7F; 1945 1946 if (verbose) 1947 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n", 1948 cpu, msr, target_c_local); 1949 1950 if (target_c_local < 85 || target_c_local > 127) 1951 goto guess; 1952 1953 tcc_activation_temp = target_c_local; 1954 1955 return 0; 1956 1957 guess: 1958 tcc_activation_temp = TJMAX_DEFAULT; 1959 fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n", 1960 cpu, tcc_activation_temp); 1961 1962 return 0; 1963 } 1964 void check_cpuid() 1965 { 1966 unsigned int eax, ebx, ecx, edx, max_level; 1967 unsigned int fms, family, model, stepping; 1968 1969 eax = ebx = ecx = edx = 0; 1970 1971 __get_cpuid(0, &max_level, &ebx, &ecx, &edx); 1972 1973 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e) 1974 genuine_intel = 1; 1975 1976 if (verbose) 1977 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ", 1978 (char *)&ebx, (char *)&edx, (char *)&ecx); 1979 1980 __get_cpuid(1, &fms, &ebx, &ecx, &edx); 1981 family = (fms >> 8) & 0xf; 1982 model = (fms >> 4) & 0xf; 1983 stepping = fms & 0xf; 1984 if (family == 6 || family == 0xf) 1985 model += ((fms >> 16) & 0xf) << 4; 1986 1987 if (verbose) 1988 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n", 1989 max_level, family, model, stepping, family, model, stepping); 1990 1991 if (!(edx & (1 << 5))) 1992 errx(1, "CPUID: no MSR"); 1993 1994 /* 1995 * check max extended function levels of CPUID. 1996 * This is needed to check for invariant TSC. 1997 * This check is valid for both Intel and AMD. 1998 */ 1999 ebx = ecx = edx = 0; 2000 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx); 2001 2002 if (max_level < 0x80000007) 2003 errx(1, "CPUID: no invariant TSC (max_level 0x%x)", max_level); 2004 2005 /* 2006 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8 2007 * this check is valid for both Intel and AMD 2008 */ 2009 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx); 2010 has_invariant_tsc = edx & (1 << 8); 2011 2012 if (!has_invariant_tsc) 2013 errx(1, "No invariant TSC"); 2014 2015 /* 2016 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0 2017 * this check is valid for both Intel and AMD 2018 */ 2019 2020 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx); 2021 has_aperf = ecx & (1 << 0); 2022 do_dts = eax & (1 << 0); 2023 do_ptm = eax & (1 << 6); 2024 has_epb = ecx & (1 << 3); 2025 2026 if (verbose) 2027 fprintf(stderr, "CPUID(6): %s%s%s%s\n", 2028 has_aperf ? "APERF" : "No APERF!", 2029 do_dts ? ", DTS" : "", 2030 do_ptm ? ", PTM": "", 2031 has_epb ? ", EPB": ""); 2032 2033 if (!has_aperf) 2034 errx(-1, "No APERF"); 2035 2036 do_nehalem_platform_info = genuine_intel && has_invariant_tsc; 2037 do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */ 2038 do_smi = do_nhm_cstates; 2039 do_snb_cstates = is_snb(family, model); 2040 do_c8_c9_c10 = has_c8_c9_c10(family, model); 2041 do_slm_cstates = is_slm(family, model); 2042 bclk = discover_bclk(family, model); 2043 2044 do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model); 2045 do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model); 2046 rapl_probe(family, model); 2047 2048 return; 2049 } 2050 2051 2052 void usage() 2053 { 2054 errx(1, "%s: [-v][-R][-T][-p|-P|-S][-c MSR# | -s]][-C MSR#][-m MSR#][-M MSR#][-i interval_sec | command ...]\n", 2055 progname); 2056 } 2057 2058 2059 /* 2060 * in /dev/cpu/ return success for names that are numbers 2061 * ie. filter out ".", "..", "microcode". 2062 */ 2063 int dir_filter(const struct dirent *dirp) 2064 { 2065 if (isdigit(dirp->d_name[0])) 2066 return 1; 2067 else 2068 return 0; 2069 } 2070 2071 int open_dev_cpu_msr(int dummy1) 2072 { 2073 return 0; 2074 } 2075 2076 void topology_probe() 2077 { 2078 int i; 2079 int max_core_id = 0; 2080 int max_package_id = 0; 2081 int max_siblings = 0; 2082 struct cpu_topology { 2083 int core_id; 2084 int physical_package_id; 2085 } *cpus; 2086 2087 /* Initialize num_cpus, max_cpu_num */ 2088 topo.num_cpus = 0; 2089 topo.max_cpu_num = 0; 2090 for_all_proc_cpus(count_cpus); 2091 if (!summary_only && topo.num_cpus > 1) 2092 show_cpu = 1; 2093 2094 if (verbose > 1) 2095 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num); 2096 2097 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology)); 2098 if (cpus == NULL) 2099 err(1, "calloc cpus"); 2100 2101 /* 2102 * Allocate and initialize cpu_present_set 2103 */ 2104 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1)); 2105 if (cpu_present_set == NULL) 2106 err(3, "CPU_ALLOC"); 2107 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1)); 2108 CPU_ZERO_S(cpu_present_setsize, cpu_present_set); 2109 for_all_proc_cpus(mark_cpu_present); 2110 2111 /* 2112 * Allocate and initialize cpu_affinity_set 2113 */ 2114 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1)); 2115 if (cpu_affinity_set == NULL) 2116 err(3, "CPU_ALLOC"); 2117 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1)); 2118 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set); 2119 2120 2121 /* 2122 * For online cpus 2123 * find max_core_id, max_package_id 2124 */ 2125 for (i = 0; i <= topo.max_cpu_num; ++i) { 2126 int siblings; 2127 2128 if (cpu_is_not_present(i)) { 2129 if (verbose > 1) 2130 fprintf(stderr, "cpu%d NOT PRESENT\n", i); 2131 continue; 2132 } 2133 cpus[i].core_id = get_core_id(i); 2134 if (cpus[i].core_id > max_core_id) 2135 max_core_id = cpus[i].core_id; 2136 2137 cpus[i].physical_package_id = get_physical_package_id(i); 2138 if (cpus[i].physical_package_id > max_package_id) 2139 max_package_id = cpus[i].physical_package_id; 2140 2141 siblings = get_num_ht_siblings(i); 2142 if (siblings > max_siblings) 2143 max_siblings = siblings; 2144 if (verbose > 1) 2145 fprintf(stderr, "cpu %d pkg %d core %d\n", 2146 i, cpus[i].physical_package_id, cpus[i].core_id); 2147 } 2148 topo.num_cores_per_pkg = max_core_id + 1; 2149 if (verbose > 1) 2150 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n", 2151 max_core_id, topo.num_cores_per_pkg); 2152 if (!summary_only && topo.num_cores_per_pkg > 1) 2153 show_core = 1; 2154 2155 topo.num_packages = max_package_id + 1; 2156 if (verbose > 1) 2157 fprintf(stderr, "max_package_id %d, sizing for %d packages\n", 2158 max_package_id, topo.num_packages); 2159 if (!summary_only && topo.num_packages > 1) 2160 show_pkg = 1; 2161 2162 topo.num_threads_per_core = max_siblings; 2163 if (verbose > 1) 2164 fprintf(stderr, "max_siblings %d\n", max_siblings); 2165 2166 free(cpus); 2167 } 2168 2169 void 2170 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p) 2171 { 2172 int i; 2173 2174 *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg * 2175 topo.num_packages, sizeof(struct thread_data)); 2176 if (*t == NULL) 2177 goto error; 2178 2179 for (i = 0; i < topo.num_threads_per_core * 2180 topo.num_cores_per_pkg * topo.num_packages; i++) 2181 (*t)[i].cpu_id = -1; 2182 2183 *c = calloc(topo.num_cores_per_pkg * topo.num_packages, 2184 sizeof(struct core_data)); 2185 if (*c == NULL) 2186 goto error; 2187 2188 for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++) 2189 (*c)[i].core_id = -1; 2190 2191 *p = calloc(topo.num_packages, sizeof(struct pkg_data)); 2192 if (*p == NULL) 2193 goto error; 2194 2195 for (i = 0; i < topo.num_packages; i++) 2196 (*p)[i].package_id = i; 2197 2198 return; 2199 error: 2200 err(1, "calloc counters"); 2201 } 2202 /* 2203 * init_counter() 2204 * 2205 * set cpu_id, core_num, pkg_num 2206 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE 2207 * 2208 * increment topo.num_cores when 1st core in pkg seen 2209 */ 2210 void init_counter(struct thread_data *thread_base, struct core_data *core_base, 2211 struct pkg_data *pkg_base, int thread_num, int core_num, 2212 int pkg_num, int cpu_id) 2213 { 2214 struct thread_data *t; 2215 struct core_data *c; 2216 struct pkg_data *p; 2217 2218 t = GET_THREAD(thread_base, thread_num, core_num, pkg_num); 2219 c = GET_CORE(core_base, core_num, pkg_num); 2220 p = GET_PKG(pkg_base, pkg_num); 2221 2222 t->cpu_id = cpu_id; 2223 if (thread_num == 0) { 2224 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE; 2225 if (cpu_is_first_core_in_package(cpu_id)) 2226 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE; 2227 } 2228 2229 c->core_id = core_num; 2230 p->package_id = pkg_num; 2231 } 2232 2233 2234 int initialize_counters(int cpu_id) 2235 { 2236 int my_thread_id, my_core_id, my_package_id; 2237 2238 my_package_id = get_physical_package_id(cpu_id); 2239 my_core_id = get_core_id(cpu_id); 2240 2241 if (cpu_is_first_sibling_in_core(cpu_id)) { 2242 my_thread_id = 0; 2243 topo.num_cores++; 2244 } else { 2245 my_thread_id = 1; 2246 } 2247 2248 init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id); 2249 init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id); 2250 return 0; 2251 } 2252 2253 void allocate_output_buffer() 2254 { 2255 output_buffer = calloc(1, (1 + topo.num_cpus) * 256); 2256 outp = output_buffer; 2257 if (outp == NULL) 2258 err(-1, "calloc output buffer"); 2259 } 2260 2261 void setup_all_buffers(void) 2262 { 2263 topology_probe(); 2264 allocate_counters(&thread_even, &core_even, &package_even); 2265 allocate_counters(&thread_odd, &core_odd, &package_odd); 2266 allocate_output_buffer(); 2267 for_all_proc_cpus(initialize_counters); 2268 } 2269 void turbostat_init() 2270 { 2271 check_cpuid(); 2272 2273 check_dev_msr(); 2274 check_super_user(); 2275 2276 setup_all_buffers(); 2277 2278 if (verbose) 2279 print_verbose_header(); 2280 2281 if (verbose) 2282 for_all_cpus(print_epb, ODD_COUNTERS); 2283 2284 if (verbose) 2285 for_all_cpus(print_rapl, ODD_COUNTERS); 2286 2287 for_all_cpus(set_temperature_target, ODD_COUNTERS); 2288 2289 if (verbose) 2290 for_all_cpus(print_thermal, ODD_COUNTERS); 2291 } 2292 2293 int fork_it(char **argv) 2294 { 2295 pid_t child_pid; 2296 int status; 2297 2298 status = for_all_cpus(get_counters, EVEN_COUNTERS); 2299 if (status) 2300 exit(status); 2301 /* clear affinity side-effect of get_counters() */ 2302 sched_setaffinity(0, cpu_present_setsize, cpu_present_set); 2303 gettimeofday(&tv_even, (struct timezone *)NULL); 2304 2305 child_pid = fork(); 2306 if (!child_pid) { 2307 /* child */ 2308 execvp(argv[0], argv); 2309 } else { 2310 2311 /* parent */ 2312 if (child_pid == -1) 2313 err(1, "fork"); 2314 2315 signal(SIGINT, SIG_IGN); 2316 signal(SIGQUIT, SIG_IGN); 2317 if (waitpid(child_pid, &status, 0) == -1) 2318 err(status, "waitpid"); 2319 } 2320 /* 2321 * n.b. fork_it() does not check for errors from for_all_cpus() 2322 * because re-starting is problematic when forking 2323 */ 2324 for_all_cpus(get_counters, ODD_COUNTERS); 2325 gettimeofday(&tv_odd, (struct timezone *)NULL); 2326 timersub(&tv_odd, &tv_even, &tv_delta); 2327 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS); 2328 compute_average(EVEN_COUNTERS); 2329 format_all_counters(EVEN_COUNTERS); 2330 flush_stderr(); 2331 2332 fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0); 2333 2334 return status; 2335 } 2336 2337 void cmdline(int argc, char **argv) 2338 { 2339 int opt; 2340 2341 progname = argv[0]; 2342 2343 while ((opt = getopt(argc, argv, "+pPSvi:sc:sC:m:M:RT:")) != -1) { 2344 switch (opt) { 2345 case 'p': 2346 show_core_only++; 2347 break; 2348 case 'P': 2349 show_pkg_only++; 2350 break; 2351 case 'S': 2352 summary_only++; 2353 break; 2354 case 'v': 2355 verbose++; 2356 break; 2357 case 'i': 2358 interval_sec = atoi(optarg); 2359 break; 2360 case 'c': 2361 sscanf(optarg, "%x", &extra_delta_offset32); 2362 break; 2363 case 'C': 2364 sscanf(optarg, "%x", &extra_delta_offset64); 2365 break; 2366 case 'm': 2367 sscanf(optarg, "%x", &extra_msr_offset32); 2368 break; 2369 case 'M': 2370 sscanf(optarg, "%x", &extra_msr_offset64); 2371 break; 2372 case 'R': 2373 rapl_verbose++; 2374 break; 2375 case 'T': 2376 tcc_activation_temp_override = atoi(optarg); 2377 break; 2378 default: 2379 usage(); 2380 } 2381 } 2382 } 2383 2384 int main(int argc, char **argv) 2385 { 2386 cmdline(argc, argv); 2387 2388 if (verbose) 2389 fprintf(stderr, "turbostat v3.6 Dec 2, 2013" 2390 " - Len Brown <lenb@kernel.org>\n"); 2391 2392 turbostat_init(); 2393 2394 /* 2395 * if any params left, it must be a command to fork 2396 */ 2397 if (argc - optind) 2398 return fork_it(argv + optind); 2399 else 2400 turbostat_loop(); 2401 2402 return 0; 2403 } 2404