tools/power/x86/turbostat/turbostat.c

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