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) 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 <getopt.h>
  37 #include <dirent.h>
  38 #include <string.h>
  39 #include <ctype.h>
  40 #include <sched.h>
  41 #include <cpuid.h>
  42 #include <linux/capability.h>
  43 #include <errno.h>
  44
  45 char *proc_stat = "/proc/stat";
  46 unsigned int interval_sec = 5;
  47 unsigned int debug;
  48 unsigned int rapl_joules;
  49 unsigned int summary_only;
  50 unsigned int dump_only;
  51 unsigned int skip_c0;
  52 unsigned int skip_c1;
  53 unsigned int do_nhm_cstates;
  54 unsigned int do_snb_cstates;
  55 unsigned int do_pc2;
  56 unsigned int do_pc3;
  57 unsigned int do_pc6;
  58 unsigned int do_pc7;
  59 unsigned int do_c8_c9_c10;
  60 unsigned int do_slm_cstates;
  61 unsigned int use_c1_residency_msr;
  62 unsigned int has_aperf;
  63 unsigned int has_epb;
  64 unsigned int units = 1000000;   /* MHz etc */
  65 unsigned int genuine_intel;
  66 unsigned int has_invariant_tsc;
  67 unsigned int do_nhm_platform_info;
  68 unsigned int extra_msr_offset32;
  69 unsigned int extra_msr_offset64;
  70 unsigned int extra_delta_offset32;
  71 unsigned int extra_delta_offset64;
  72 int do_smi;
  73 double bclk;
  74 unsigned int show_pkg;
  75 unsigned int show_core;
  76 unsigned int show_cpu;
  77 unsigned int show_pkg_only;
  78 unsigned int show_core_only;
  79 char *output_buffer, *outp;
  80 unsigned int do_rapl;
  81 unsigned int do_dts;
  82 unsigned int do_ptm;
  83 unsigned int tcc_activation_temp;
  84 unsigned int tcc_activation_temp_override;
  85 double rapl_power_units, rapl_energy_units, rapl_time_units;
  86 double rapl_joule_counter_range;
  87 unsigned int do_core_perf_limit_reasons;
  88 unsigned int do_gfx_perf_limit_reasons;
  89 unsigned int do_ring_perf_limit_reasons;
  90
  91 #define RAPL_PKG                (1 << 0)
  92                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  93                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  94 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  95                                         /* 0x613 MSR_PKG_PERF_STATUS */
  96 #define RAPL_PKG_POWER_INFO     (1 << 2)
  97                                         /* 0x614 MSR_PKG_POWER_INFO */
  98
  99 #define RAPL_DRAM               (1 << 3)
 100                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
 101                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
 102                                         /* 0x61c MSR_DRAM_POWER_INFO */
 103 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
 104                                         /* 0x61b MSR_DRAM_PERF_STATUS */
 105
 106 #define RAPL_CORES              (1 << 5)
 107                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 108                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 109 #define RAPL_CORE_POLICY        (1 << 6)
 110                                         /* 0x63a MSR_PP0_POLICY */
 111
 112
 113 #define RAPL_GFX                (1 << 7)
 114                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 115                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 116                                         /* 0x642 MSR_PP1_POLICY */
 117 #define TJMAX_DEFAULT   100
 118
 119 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 120
 121 int aperf_mperf_unstable;
 122 int backwards_count;
 123 char *progname;
 124
 125 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 126 size_t cpu_present_setsize, cpu_affinity_setsize;
 127
 128 struct thread_data {
 129         unsigned long long tsc;
 130         unsigned long long aperf;
 131         unsigned long long mperf;
 132         unsigned long long c1;
 133         unsigned long long extra_msr64;
 134         unsigned long long extra_delta64;
 135         unsigned long long extra_msr32;
 136         unsigned long long extra_delta32;
 137         unsigned int smi_count;
 138         unsigned int cpu_id;
 139         unsigned int flags;
 140 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 141 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 142 } *thread_even, *thread_odd;
 143
 144 struct core_data {
 145         unsigned long long c3;
 146         unsigned long long c6;
 147         unsigned long long c7;
 148         unsigned int core_temp_c;
 149         unsigned int core_id;
 150 } *core_even, *core_odd;
 151
 152 struct pkg_data {
 153         unsigned long long pc2;
 154         unsigned long long pc3;
 155         unsigned long long pc6;
 156         unsigned long long pc7;
 157         unsigned long long pc8;
 158         unsigned long long pc9;
 159         unsigned long long pc10;
 160         unsigned int package_id;
 161         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 162         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 163         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 164         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 165         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 166         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 167         unsigned int pkg_temp_c;
 168
 169 } *package_even, *package_odd;
 170
 171 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 172 #define EVEN_COUNTERS thread_even, core_even, package_even
 173
 174 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 175         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 176                 topo.num_threads_per_core + \
 177                 (core_no) * topo.num_threads_per_core + (thread_no))
 178 #define GET_CORE(core_base, core_no, pkg_no) \
 179         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 180 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 181
 182 struct system_summary {
 183         struct thread_data threads;
 184         struct core_data cores;
 185         struct pkg_data packages;
 186 } sum, average;
 187
 188
 189 struct topo_params {
 190         int num_packages;
 191         int num_cpus;
 192         int num_cores;
 193         int max_cpu_num;
 194         int num_cores_per_pkg;
 195         int num_threads_per_core;
 196 } topo;
 197
 198 struct timeval tv_even, tv_odd, tv_delta;
 199
 200 void setup_all_buffers(void);
 201
 202 int cpu_is_not_present(int cpu)
 203 {
 204         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 205 }
 206 /*
 207  * run func(thread, core, package) in topology order
 208  * skip non-present cpus
 209  */
 210
 211 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 212         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 213 {
 214         int retval, pkg_no, core_no, thread_no;
 215
 216         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 217                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 218                         for (thread_no = 0; thread_no <
 219                                 topo.num_threads_per_core; ++thread_no) {
 220                                 struct thread_data *t;
 221                                 struct core_data *c;
 222                                 struct pkg_data *p;
 223
 224                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 225
 226                                 if (cpu_is_not_present(t->cpu_id))
 227                                         continue;
 228
 229                                 c = GET_CORE(core_base, core_no, pkg_no);
 230                                 p = GET_PKG(pkg_base, pkg_no);
 231
 232                                 retval = func(t, c, p);
 233                                 if (retval)
 234                                         return retval;
 235                         }
 236                 }
 237         }
 238         return 0;
 239 }
 240
 241 int cpu_migrate(int cpu)
 242 {
 243         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 244         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 245         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 246                 return -1;
 247         else
 248                 return 0;
 249 }
 250
 251 int get_msr(int cpu, off_t offset, unsigned long long *msr)
 252 {
 253         ssize_t retval;
 254         char pathname[32];
 255         int fd;
 256
 257         sprintf(pathname, "/dev/cpu/%d/msr", cpu);
 258         fd = open(pathname, O_RDONLY);
 259         if (fd < 0)
 260                 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
 261
 262         retval = pread(fd, msr, sizeof *msr, offset);
 263         close(fd);
 264
 265         if (retval != sizeof *msr)
 266                 err(-1, "%s offset 0x%llx read failed", pathname, (unsigned long long)offset);
 267
 268         return 0;
 269 }
 270
 271 /*
 272  * Example Format w/ field column widths:
 273  *
 274  *  Package    Core     CPU Avg_MHz Bzy_MHz TSC_MHz     SMI   %Busy CPU_%c1 CPU_%c3 CPU_%c6 CPU_%c7 CoreTmp  PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 PkgWatt CorWatt GFXWatt
 275  * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
 276  */
 277
 278 void print_header(void)
 279 {
 280         if (show_pkg)
 281                 outp += sprintf(outp, " Package");
 282         if (show_core)
 283                 outp += sprintf(outp, "    Core");
 284         if (show_cpu)
 285                 outp += sprintf(outp, "     CPU");
 286         if (has_aperf)
 287                 outp += sprintf(outp, " Avg_MHz");
 288         if (has_aperf)
 289                 outp += sprintf(outp, "   %%Busy");
 290         if (has_aperf)
 291                 outp += sprintf(outp, " Bzy_MHz");
 292         outp += sprintf(outp, " TSC_MHz");
 293
 294         if (extra_delta_offset32)
 295                 outp += sprintf(outp, "  count 0x%03X", extra_delta_offset32);
 296         if (extra_delta_offset64)
 297                 outp += sprintf(outp, "  COUNT 0x%03X", extra_delta_offset64);
 298         if (extra_msr_offset32)
 299                 outp += sprintf(outp, "   MSR 0x%03X", extra_msr_offset32);
 300         if (extra_msr_offset64)
 301                 outp += sprintf(outp, "           MSR 0x%03X", extra_msr_offset64);
 302
 303         if (!debug)
 304                 goto done;
 305
 306         if (do_smi)
 307                 outp += sprintf(outp, "     SMI");
 308
 309         if (do_nhm_cstates)
 310                 outp += sprintf(outp, "  CPU%%c1");
 311         if (do_nhm_cstates && !do_slm_cstates)
 312                 outp += sprintf(outp, "  CPU%%c3");
 313         if (do_nhm_cstates)
 314                 outp += sprintf(outp, "  CPU%%c6");
 315         if (do_snb_cstates)
 316                 outp += sprintf(outp, "  CPU%%c7");
 317
 318         if (do_dts)
 319                 outp += sprintf(outp, " CoreTmp");
 320         if (do_ptm)
 321                 outp += sprintf(outp, "  PkgTmp");
 322
 323         if (do_pc2)
 324                 outp += sprintf(outp, " Pkg%%pc2");
 325         if (do_pc3)
 326                 outp += sprintf(outp, " Pkg%%pc3");
 327         if (do_pc6)
 328                 outp += sprintf(outp, " Pkg%%pc6");
 329         if (do_pc7)
 330                 outp += sprintf(outp, " Pkg%%pc7");
 331         if (do_c8_c9_c10) {
 332                 outp += sprintf(outp, " Pkg%%pc8");
 333                 outp += sprintf(outp, " Pkg%%pc9");
 334                 outp += sprintf(outp, " Pk%%pc10");
 335         }
 336
 337         if (do_rapl && !rapl_joules) {
 338                 if (do_rapl & RAPL_PKG)
 339                         outp += sprintf(outp, " PkgWatt");
 340                 if (do_rapl & RAPL_CORES)
 341                         outp += sprintf(outp, " CorWatt");
 342                 if (do_rapl & RAPL_GFX)
 343                         outp += sprintf(outp, " GFXWatt");
 344                 if (do_rapl & RAPL_DRAM)
 345                         outp += sprintf(outp, " RAMWatt");
 346                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 347                         outp += sprintf(outp, "   PKG_%%");
 348                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 349                         outp += sprintf(outp, "   RAM_%%");
 350         } else if (do_rapl && rapl_joules) {
 351                 if (do_rapl & RAPL_PKG)
 352                         outp += sprintf(outp, "   Pkg_J");
 353                 if (do_rapl & RAPL_CORES)
 354                         outp += sprintf(outp, "   Cor_J");
 355                 if (do_rapl & RAPL_GFX)
 356                         outp += sprintf(outp, "   GFX_J");
 357                 if (do_rapl & RAPL_DRAM)
 358                         outp += sprintf(outp, "   RAM_W");
 359                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 360                         outp += sprintf(outp, "   PKG_%%");
 361                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 362                         outp += sprintf(outp, "   RAM_%%");
 363                 outp += sprintf(outp, "   time");
 364
 365         }
 366     done:
 367         outp += sprintf(outp, "\n");
 368 }
 369
 370 int dump_counters(struct thread_data *t, struct core_data *c,
 371         struct pkg_data *p)
 372 {
 373         outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
 374
 375         if (t) {
 376                 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
 377                         t->cpu_id, t->flags);
 378                 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
 379                 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
 380                 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
 381                 outp += sprintf(outp, "c1: %016llX\n", t->c1);
 382                 outp += sprintf(outp, "msr0x%x: %08llX\n",
 383                         extra_delta_offset32, t->extra_delta32);
 384                 outp += sprintf(outp, "msr0x%x: %016llX\n",
 385                         extra_delta_offset64, t->extra_delta64);
 386                 outp += sprintf(outp, "msr0x%x: %08llX\n",
 387                         extra_msr_offset32, t->extra_msr32);
 388                 outp += sprintf(outp, "msr0x%x: %016llX\n",
 389                         extra_msr_offset64, t->extra_msr64);
 390                 if (do_smi)
 391                         outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
 392         }
 393
 394         if (c) {
 395                 outp += sprintf(outp, "core: %d\n", c->core_id);
 396                 outp += sprintf(outp, "c3: %016llX\n", c->c3);
 397                 outp += sprintf(outp, "c6: %016llX\n", c->c6);
 398                 outp += sprintf(outp, "c7: %016llX\n", c->c7);
 399                 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
 400         }
 401
 402         if (p) {
 403                 outp += sprintf(outp, "package: %d\n", p->package_id);
 404                 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
 405                 if (do_pc3)
 406                         outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
 407                 if (do_pc6)
 408                         outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
 409                 if (do_pc7)
 410                         outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
 411                 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
 412                 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
 413                 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
 414                 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
 415                 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
 416                 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
 417                 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
 418                 outp += sprintf(outp, "Throttle PKG: %0X\n",
 419                         p->rapl_pkg_perf_status);
 420                 outp += sprintf(outp, "Throttle RAM: %0X\n",
 421                         p->rapl_dram_perf_status);
 422                 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
 423         }
 424
 425         outp += sprintf(outp, "\n");
 426
 427         return 0;
 428 }
 429
 430 /*
 431  * column formatting convention & formats
 432  */
 433 int format_counters(struct thread_data *t, struct core_data *c,
 434         struct pkg_data *p)
 435 {
 436         double interval_float;
 437         char *fmt8;
 438
 439          /* if showing only 1st thread in core and this isn't one, bail out */
 440         if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 441                 return 0;
 442
 443          /* if showing only 1st thread in pkg and this isn't one, bail out */
 444         if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 445                 return 0;
 446
 447         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 448
 449         /* topo columns, print blanks on 1st (average) line */
 450         if (t == &average.threads) {
 451                 if (show_pkg)
 452                         outp += sprintf(outp, "       -");
 453                 if (show_core)
 454                         outp += sprintf(outp, "       -");
 455                 if (show_cpu)
 456                         outp += sprintf(outp, "       -");
 457         } else {
 458                 if (show_pkg) {
 459                         if (p)
 460                                 outp += sprintf(outp, "%8d", p->package_id);
 461                         else
 462                                 outp += sprintf(outp, "       -");
 463                 }
 464                 if (show_core) {
 465                         if (c)
 466                                 outp += sprintf(outp, "%8d", c->core_id);
 467                         else
 468                                 outp += sprintf(outp, "       -");
 469                 }
 470                 if (show_cpu)
 471                         outp += sprintf(outp, "%8d", t->cpu_id);
 472         }
 473
 474         /* Avg_MHz */
 475         if (has_aperf)
 476                 outp += sprintf(outp, "%8.0f",
 477                         1.0 / units * t->aperf / interval_float);
 478
 479         /* %Busy */
 480         if (has_aperf) {
 481                 if (!skip_c0)
 482                         outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc);
 483                 else
 484                         outp += sprintf(outp, "********");
 485         }
 486
 487         /* Bzy_MHz */
 488         if (has_aperf)
 489                 outp += sprintf(outp, "%8.0f",
 490                         1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
 491
 492         /* TSC_MHz */
 493         outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
 494
 495         /* delta */
 496         if (extra_delta_offset32)
 497                 outp += sprintf(outp, "  %11llu", t->extra_delta32);
 498
 499         /* DELTA */
 500         if (extra_delta_offset64)
 501                 outp += sprintf(outp, "  %11llu", t->extra_delta64);
 502         /* msr */
 503         if (extra_msr_offset32)
 504                 outp += sprintf(outp, "  0x%08llx", t->extra_msr32);
 505
 506         /* MSR */
 507         if (extra_msr_offset64)
 508                 outp += sprintf(outp, "  0x%016llx", t->extra_msr64);
 509
 510         if (!debug)
 511                 goto done;
 512
 513         /* SMI */
 514         if (do_smi)
 515                 outp += sprintf(outp, "%8d", t->smi_count);
 516
 517         if (do_nhm_cstates) {
 518                 if (!skip_c1)
 519                         outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
 520                 else
 521                         outp += sprintf(outp, "********");
 522         }
 523
 524         /* print per-core data only for 1st thread in core */
 525         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 526                 goto done;
 527
 528         if (do_nhm_cstates && !do_slm_cstates)
 529                 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
 530         if (do_nhm_cstates)
 531                 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
 532         if (do_snb_cstates)
 533                 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
 534
 535         if (do_dts)
 536                 outp += sprintf(outp, "%8d", c->core_temp_c);
 537
 538         /* print per-package data only for 1st core in package */
 539         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 540                 goto done;
 541
 542         if (do_ptm)
 543                 outp += sprintf(outp, "%8d", p->pkg_temp_c);
 544
 545         if (do_pc2)
 546                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
 547         if (do_pc3)
 548                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
 549         if (do_pc6)
 550                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
 551         if (do_pc7)
 552                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
 553         if (do_c8_c9_c10) {
 554                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
 555                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
 556                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
 557         }
 558
 559         /*
 560          * If measurement interval exceeds minimum RAPL Joule Counter range,
 561          * indicate that results are suspect by printing "**" in fraction place.
 562          */
 563         if (interval_float < rapl_joule_counter_range)
 564                 fmt8 = "%8.2f";
 565         else
 566                 fmt8 = " %6.0f**";
 567
 568         if (do_rapl && !rapl_joules) {
 569                 if (do_rapl & RAPL_PKG)
 570                         outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
 571                 if (do_rapl & RAPL_CORES)
 572                         outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
 573                 if (do_rapl & RAPL_GFX)
 574                         outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
 575                 if (do_rapl & RAPL_DRAM)
 576                         outp += sprintf(outp, fmt8, p->energy_dram * rapl_energy_units / interval_float);
 577                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 578                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
 579                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 580                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
 581         } else if (do_rapl && rapl_joules) {
 582                 if (do_rapl & RAPL_PKG)
 583                         outp += sprintf(outp, fmt8,
 584                                         p->energy_pkg * rapl_energy_units);
 585                 if (do_rapl & RAPL_CORES)
 586                         outp += sprintf(outp, fmt8,
 587                                         p->energy_cores * rapl_energy_units);
 588                 if (do_rapl & RAPL_GFX)
 589                         outp += sprintf(outp, fmt8,
 590                                         p->energy_gfx * rapl_energy_units);
 591                 if (do_rapl & RAPL_DRAM)
 592                         outp += sprintf(outp, fmt8,
 593                                         p->energy_dram * rapl_energy_units);
 594                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 595                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
 596                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 597                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
 598
 599                 outp += sprintf(outp, fmt8, interval_float);
 600         }
 601 done:
 602         outp += sprintf(outp, "\n");
 603
 604         return 0;
 605 }
 606
 607 void flush_stdout()
 608 {
 609         fputs(output_buffer, stdout);
 610         fflush(stdout);
 611         outp = output_buffer;
 612 }
 613 void flush_stderr()
 614 {
 615         fputs(output_buffer, stderr);
 616         outp = output_buffer;
 617 }
 618 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 619 {
 620         static int printed;
 621
 622         if (!printed || !summary_only)
 623                 print_header();
 624
 625         if (topo.num_cpus > 1)
 626                 format_counters(&average.threads, &average.cores,
 627                         &average.packages);
 628
 629         printed = 1;
 630
 631         if (summary_only)
 632                 return;
 633
 634         for_all_cpus(format_counters, t, c, p);
 635 }
 636
 637 #define DELTA_WRAP32(new, old)                  \
 638         if (new > old) {                        \
 639                 old = new - old;                \
 640         } else {                                \
 641                 old = 0x100000000 + new - old;  \
 642         }
 643
 644 void
 645 delta_package(struct pkg_data *new, struct pkg_data *old)
 646 {
 647         old->pc2 = new->pc2 - old->pc2;
 648         if (do_pc3)
 649                 old->pc3 = new->pc3 - old->pc3;
 650         if (do_pc6)
 651                 old->pc6 = new->pc6 - old->pc6;
 652         if (do_pc7)
 653                 old->pc7 = new->pc7 - old->pc7;
 654         old->pc8 = new->pc8 - old->pc8;
 655         old->pc9 = new->pc9 - old->pc9;
 656         old->pc10 = new->pc10 - old->pc10;
 657         old->pkg_temp_c = new->pkg_temp_c;
 658
 659         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 660         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 661         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 662         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 663         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 664         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 665 }
 666
 667 void
 668 delta_core(struct core_data *new, struct core_data *old)
 669 {
 670         old->c3 = new->c3 - old->c3;
 671         old->c6 = new->c6 - old->c6;
 672         old->c7 = new->c7 - old->c7;
 673         old->core_temp_c = new->core_temp_c;
 674 }
 675
 676 /*
 677  * old = new - old
 678  */
 679 void
 680 delta_thread(struct thread_data *new, struct thread_data *old,
 681         struct core_data *core_delta)
 682 {
 683         old->tsc = new->tsc - old->tsc;
 684
 685         /* check for TSC < 1 Mcycles over interval */
 686         if (old->tsc < (1000 * 1000))
 687                 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
 688                      "You can disable all c-states by booting with \"idle=poll\"\n"
 689                      "or just the deep ones with \"processor.max_cstate=1\"");
 690
 691         old->c1 = new->c1 - old->c1;
 692
 693         if (has_aperf) {
 694                 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 695                         old->aperf = new->aperf - old->aperf;
 696                         old->mperf = new->mperf - old->mperf;
 697                 } else {
 698
 699                         if (!aperf_mperf_unstable) {
 700                                 fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
 701                                 fprintf(stderr, "* Frequency results do not cover entire interval *\n");
 702                                 fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
 703
 704                                 aperf_mperf_unstable = 1;
 705                         }
 706                         /*
 707                          * mperf delta is likely a huge "positive" number
 708                          * can not use it for calculating c0 time
 709                          */
 710                         skip_c0 = 1;
 711                         skip_c1 = 1;
 712                 }
 713         }
 714
 715
 716         if (use_c1_residency_msr) {
 717                 /*
 718                  * Some models have a dedicated C1 residency MSR,
 719                  * which should be more accurate than the derivation below.
 720                  */
 721         } else {
 722                 /*
 723                  * As counter collection is not atomic,
 724                  * it is possible for mperf's non-halted cycles + idle states
 725                  * to exceed TSC's all cycles: show c1 = 0% in that case.
 726                  */
 727                 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 728                         old->c1 = 0;
 729                 else {
 730                         /* normal case, derive c1 */
 731                         old->c1 = old->tsc - old->mperf - core_delta->c3
 732                                 - core_delta->c6 - core_delta->c7;
 733                 }
 734         }
 735
 736         if (old->mperf == 0) {
 737                 if (debug > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
 738                 old->mperf = 1; /* divide by 0 protection */
 739         }
 740
 741         old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
 742         old->extra_delta32 &= 0xFFFFFFFF;
 743
 744         old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
 745
 746         /*
 747          * Extra MSR is just a snapshot, simply copy latest w/o subtracting
 748          */
 749         old->extra_msr32 = new->extra_msr32;
 750         old->extra_msr64 = new->extra_msr64;
 751
 752         if (do_smi)
 753                 old->smi_count = new->smi_count - old->smi_count;
 754 }
 755
 756 int delta_cpu(struct thread_data *t, struct core_data *c,
 757         struct pkg_data *p, struct thread_data *t2,
 758         struct core_data *c2, struct pkg_data *p2)
 759 {
 760         /* calculate core delta only for 1st thread in core */
 761         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 762                 delta_core(c, c2);
 763
 764         /* always calculate thread delta */
 765         delta_thread(t, t2, c2);        /* c2 is core delta */
 766
 767         /* calculate package delta only for 1st core in package */
 768         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 769                 delta_package(p, p2);
 770
 771         return 0;
 772 }
 773
 774 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 775 {
 776         t->tsc = 0;
 777         t->aperf = 0;
 778         t->mperf = 0;
 779         t->c1 = 0;
 780
 781         t->smi_count = 0;
 782         t->extra_delta32 = 0;
 783         t->extra_delta64 = 0;
 784
 785         /* tells format_counters to dump all fields from this set */
 786         t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
 787
 788         c->c3 = 0;
 789         c->c6 = 0;
 790         c->c7 = 0;
 791         c->core_temp_c = 0;
 792
 793         p->pc2 = 0;
 794         if (do_pc3)
 795                 p->pc3 = 0;
 796         if (do_pc6)
 797                 p->pc6 = 0;
 798         if (do_pc7)
 799                 p->pc7 = 0;
 800         p->pc8 = 0;
 801         p->pc9 = 0;
 802         p->pc10 = 0;
 803
 804         p->energy_pkg = 0;
 805         p->energy_dram = 0;
 806         p->energy_cores = 0;
 807         p->energy_gfx = 0;
 808         p->rapl_pkg_perf_status = 0;
 809         p->rapl_dram_perf_status = 0;
 810         p->pkg_temp_c = 0;
 811 }
 812 int sum_counters(struct thread_data *t, struct core_data *c,
 813         struct pkg_data *p)
 814 {
 815         average.threads.tsc += t->tsc;
 816         average.threads.aperf += t->aperf;
 817         average.threads.mperf += t->mperf;
 818         average.threads.c1 += t->c1;
 819
 820         average.threads.extra_delta32 += t->extra_delta32;
 821         average.threads.extra_delta64 += t->extra_delta64;
 822
 823         /* sum per-core values only for 1st thread in core */
 824         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 825                 return 0;
 826
 827         average.cores.c3 += c->c3;
 828         average.cores.c6 += c->c6;
 829         average.cores.c7 += c->c7;
 830
 831         average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
 832
 833         /* sum per-pkg values only for 1st core in pkg */
 834         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 835                 return 0;
 836
 837         average.packages.pc2 += p->pc2;
 838         if (do_pc3)
 839                 average.packages.pc3 += p->pc3;
 840         if (do_pc6)
 841                 average.packages.pc6 += p->pc6;
 842         if (do_pc7)
 843                 average.packages.pc7 += p->pc7;
 844         average.packages.pc8 += p->pc8;
 845         average.packages.pc9 += p->pc9;
 846         average.packages.pc10 += p->pc10;
 847
 848         average.packages.energy_pkg += p->energy_pkg;
 849         average.packages.energy_dram += p->energy_dram;
 850         average.packages.energy_cores += p->energy_cores;
 851         average.packages.energy_gfx += p->energy_gfx;
 852
 853         average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
 854
 855         average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
 856         average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
 857         return 0;
 858 }
 859 /*
 860  * sum the counters for all cpus in the system
 861  * compute the weighted average
 862  */
 863 void compute_average(struct thread_data *t, struct core_data *c,
 864         struct pkg_data *p)
 865 {
 866         clear_counters(&average.threads, &average.cores, &average.packages);
 867
 868         for_all_cpus(sum_counters, t, c, p);
 869
 870         average.threads.tsc /= topo.num_cpus;
 871         average.threads.aperf /= topo.num_cpus;
 872         average.threads.mperf /= topo.num_cpus;
 873         average.threads.c1 /= topo.num_cpus;
 874
 875         average.threads.extra_delta32 /= topo.num_cpus;
 876         average.threads.extra_delta32 &= 0xFFFFFFFF;
 877
 878         average.threads.extra_delta64 /= topo.num_cpus;
 879
 880         average.cores.c3 /= topo.num_cores;
 881         average.cores.c6 /= topo.num_cores;
 882         average.cores.c7 /= topo.num_cores;
 883
 884         average.packages.pc2 /= topo.num_packages;
 885         if (do_pc3)
 886                 average.packages.pc3 /= topo.num_packages;
 887         if (do_pc6)
 888                 average.packages.pc6 /= topo.num_packages;
 889         if (do_pc7)
 890                 average.packages.pc7 /= topo.num_packages;
 891
 892         average.packages.pc8 /= topo.num_packages;
 893         average.packages.pc9 /= topo.num_packages;
 894         average.packages.pc10 /= topo.num_packages;
 895 }
 896
 897 static unsigned long long rdtsc(void)
 898 {
 899         unsigned int low, high;
 900
 901         asm volatile("rdtsc" : "=a" (low), "=d" (high));
 902
 903         return low | ((unsigned long long)high) << 32;
 904 }
 905
 906
 907 /*
 908  * get_counters(...)
 909  * migrate to cpu
 910  * acquire and record local counters for that cpu
 911  */
 912 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 913 {
 914         int cpu = t->cpu_id;
 915         unsigned long long msr;
 916
 917         if (cpu_migrate(cpu)) {
 918                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
 919                 return -1;
 920         }
 921
 922         t->tsc = rdtsc();       /* we are running on local CPU of interest */
 923
 924         if (has_aperf) {
 925                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
 926                         return -3;
 927                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
 928                         return -4;
 929         }
 930
 931         if (do_smi) {
 932                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
 933                         return -5;
 934                 t->smi_count = msr & 0xFFFFFFFF;
 935         }
 936         if (extra_delta_offset32) {
 937                 if (get_msr(cpu, extra_delta_offset32, &msr))
 938                         return -5;
 939                 t->extra_delta32 = msr & 0xFFFFFFFF;
 940         }
 941
 942         if (extra_delta_offset64)
 943                 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
 944                         return -5;
 945
 946         if (extra_msr_offset32) {
 947                 if (get_msr(cpu, extra_msr_offset32, &msr))
 948                         return -5;
 949                 t->extra_msr32 = msr & 0xFFFFFFFF;
 950         }
 951
 952         if (extra_msr_offset64)
 953                 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
 954                         return -5;
 955
 956         if (use_c1_residency_msr) {
 957                 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
 958                         return -6;
 959         }
 960
 961         /* collect core counters only for 1st thread in core */
 962         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 963                 return 0;
 964
 965         if (do_nhm_cstates && !do_slm_cstates) {
 966                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
 967                         return -6;
 968         }
 969
 970         if (do_nhm_cstates) {
 971                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
 972                         return -7;
 973         }
 974
 975         if (do_snb_cstates)
 976                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
 977                         return -8;
 978
 979         if (do_dts) {
 980                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
 981                         return -9;
 982                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 983         }
 984
 985
 986         /* collect package counters only for 1st core in package */
 987         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 988                 return 0;
 989
 990         if (do_pc3)
 991                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
 992                         return -9;
 993         if (do_pc6)
 994                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
 995                         return -10;
 996         if (do_pc2)
 997                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
 998                         return -11;
 999         if (do_pc7)
1000                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1001                         return -12;
1002         if (do_c8_c9_c10) {
1003                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1004                         return -13;
1005                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1006                         return -13;
1007                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1008                         return -13;
1009         }
1010         if (do_rapl & RAPL_PKG) {
1011                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1012                         return -13;
1013                 p->energy_pkg = msr & 0xFFFFFFFF;
1014         }
1015         if (do_rapl & RAPL_CORES) {
1016                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1017                         return -14;
1018                 p->energy_cores = msr & 0xFFFFFFFF;
1019         }
1020         if (do_rapl & RAPL_DRAM) {
1021                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1022                         return -15;
1023                 p->energy_dram = msr & 0xFFFFFFFF;
1024         }
1025         if (do_rapl & RAPL_GFX) {
1026                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1027                         return -16;
1028                 p->energy_gfx = msr & 0xFFFFFFFF;
1029         }
1030         if (do_rapl & RAPL_PKG_PERF_STATUS) {
1031                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1032                         return -16;
1033                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1034         }
1035         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1036                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1037                         return -16;
1038                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1039         }
1040         if (do_ptm) {
1041                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1042                         return -17;
1043                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1044         }
1045         return 0;
1046 }
1047
1048 /*
1049  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1050  * If you change the values, note they are used both in comparisons
1051  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1052  */
1053
1054 #define PCLUKN 0 /* Unknown */
1055 #define PCLRSV 1 /* Reserved */
1056 #define PCL__0 2 /* PC0 */
1057 #define PCL__1 3 /* PC1 */
1058 #define PCL__2 4 /* PC2 */
1059 #define PCL__3 5 /* PC3 */
1060 #define PCL__4 6 /* PC4 */
1061 #define PCL__6 7 /* PC6 */
1062 #define PCL_6N 8 /* PC6 No Retention */
1063 #define PCL_6R 9 /* PC6 Retention */
1064 #define PCL__7 10 /* PC7 */
1065 #define PCL_7S 11 /* PC7 Shrink */
1066 #define PCLUNL 12 /* Unlimited */
1067
1068 int pkg_cstate_limit = PCLUKN;
1069 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1070         "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "unlimited"};
1071
1072 int nhm_pkg_cstate_limits[8] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL};
1073 int snb_pkg_cstate_limits[8] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL};
1074 int hsw_pkg_cstate_limits[8] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCLRSV, PCLUNL};
1075 int slv_pkg_cstate_limits[8] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7};
1076 int amt_pkg_cstate_limits[8] = {PCL__0, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7};
1077 int phi_pkg_cstate_limits[8] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL};
1078
1079 static void
1080 dump_nhm_platform_info(void)
1081 {
1082         unsigned long long msr;
1083         unsigned int ratio;
1084
1085         get_msr(0, MSR_NHM_PLATFORM_INFO, &msr);
1086
1087         fprintf(stderr, "cpu0: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", msr);
1088
1089         ratio = (msr >> 40) & 0xFF;
1090         fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency frequency\n",
1091                 ratio, bclk, ratio * bclk);
1092
1093         ratio = (msr >> 8) & 0xFF;
1094         fprintf(stderr, "%d * %.0f = %.0f MHz base frequency\n",
1095                 ratio, bclk, ratio * bclk);
1096
1097         get_msr(0, MSR_IA32_POWER_CTL, &msr);
1098         fprintf(stderr, "cpu0: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1099                 msr, msr & 0x2 ? "EN" : "DIS");
1100
1101         return;
1102 }
1103
1104 static void
1105 dump_hsw_turbo_ratio_limits(void)
1106 {
1107         unsigned long long msr;
1108         unsigned int ratio;
1109
1110         get_msr(0, MSR_TURBO_RATIO_LIMIT2, &msr);
1111
1112         fprintf(stderr, "cpu0: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", msr);
1113
1114         ratio = (msr >> 8) & 0xFF;
1115         if (ratio)
1116                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 18 active cores\n",
1117                         ratio, bclk, ratio * bclk);
1118
1119         ratio = (msr >> 0) & 0xFF;
1120         if (ratio)
1121                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 17 active cores\n",
1122                         ratio, bclk, ratio * bclk);
1123         return;
1124 }
1125
1126 static void
1127 dump_ivt_turbo_ratio_limits(void)
1128 {
1129         unsigned long long msr;
1130         unsigned int ratio;
1131
1132         get_msr(0, MSR_TURBO_RATIO_LIMIT1, &msr);
1133
1134         fprintf(stderr, "cpu0: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", msr);
1135
1136         ratio = (msr >> 56) & 0xFF;
1137         if (ratio)
1138                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1139                         ratio, bclk, ratio * bclk);
1140
1141         ratio = (msr >> 48) & 0xFF;
1142         if (ratio)
1143                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1144                         ratio, bclk, ratio * bclk);
1145
1146         ratio = (msr >> 40) & 0xFF;
1147         if (ratio)
1148                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1149                         ratio, bclk, ratio * bclk);
1150
1151         ratio = (msr >> 32) & 0xFF;
1152         if (ratio)
1153                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1154                         ratio, bclk, ratio * bclk);
1155
1156         ratio = (msr >> 24) & 0xFF;
1157         if (ratio)
1158                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1159                         ratio, bclk, ratio * bclk);
1160
1161         ratio = (msr >> 16) & 0xFF;
1162         if (ratio)
1163                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1164                         ratio, bclk, ratio * bclk);
1165
1166         ratio = (msr >> 8) & 0xFF;
1167         if (ratio)
1168                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1169                         ratio, bclk, ratio * bclk);
1170
1171         ratio = (msr >> 0) & 0xFF;
1172         if (ratio)
1173                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1174                         ratio, bclk, ratio * bclk);
1175         return;
1176 }
1177
1178 static void
1179 dump_nhm_turbo_ratio_limits(void)
1180 {
1181         unsigned long long msr;
1182         unsigned int ratio;
1183
1184         get_msr(0, MSR_TURBO_RATIO_LIMIT, &msr);
1185
1186         fprintf(stderr, "cpu0: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1187
1188         ratio = (msr >> 56) & 0xFF;
1189         if (ratio)
1190                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1191                         ratio, bclk, ratio * bclk);
1192
1193         ratio = (msr >> 48) & 0xFF;
1194         if (ratio)
1195                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1196                         ratio, bclk, ratio * bclk);
1197
1198         ratio = (msr >> 40) & 0xFF;
1199         if (ratio)
1200                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1201                         ratio, bclk, ratio * bclk);
1202
1203         ratio = (msr >> 32) & 0xFF;
1204         if (ratio)
1205                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1206                         ratio, bclk, ratio * bclk);
1207
1208         ratio = (msr >> 24) & 0xFF;
1209         if (ratio)
1210                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1211                         ratio, bclk, ratio * bclk);
1212
1213         ratio = (msr >> 16) & 0xFF;
1214         if (ratio)
1215                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1216                         ratio, bclk, ratio * bclk);
1217
1218         ratio = (msr >> 8) & 0xFF;
1219         if (ratio)
1220                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1221                         ratio, bclk, ratio * bclk);
1222
1223         ratio = (msr >> 0) & 0xFF;
1224         if (ratio)
1225                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1226                         ratio, bclk, ratio * bclk);
1227         return;
1228 }
1229
1230 static void
1231 dump_nhm_cst_cfg(void)
1232 {
1233         unsigned long long msr;
1234
1235         get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1236
1237 #define SNB_C1_AUTO_UNDEMOTE              (1UL << 27)
1238 #define SNB_C3_AUTO_UNDEMOTE              (1UL << 28)
1239
1240         fprintf(stderr, "cpu0: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", msr);
1241
1242         fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: %s)\n",
1243                 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1244                 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1245                 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1246                 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1247                 (msr & (1 << 15)) ? "" : "UN",
1248                 (unsigned int)msr & 7,
1249                 pkg_cstate_limit_strings[pkg_cstate_limit]);
1250         return;
1251 }
1252
1253 void free_all_buffers(void)
1254 {
1255         CPU_FREE(cpu_present_set);
1256         cpu_present_set = NULL;
1257         cpu_present_set = 0;
1258
1259         CPU_FREE(cpu_affinity_set);
1260         cpu_affinity_set = NULL;
1261         cpu_affinity_setsize = 0;
1262
1263         free(thread_even);
1264         free(core_even);
1265         free(package_even);
1266
1267         thread_even = NULL;
1268         core_even = NULL;
1269         package_even = NULL;
1270
1271         free(thread_odd);
1272         free(core_odd);
1273         free(package_odd);
1274
1275         thread_odd = NULL;
1276         core_odd = NULL;
1277         package_odd = NULL;
1278
1279         free(output_buffer);
1280         output_buffer = NULL;
1281         outp = NULL;
1282 }
1283
1284 /*
1285  * Open a file, and exit on failure
1286  */
1287 FILE *fopen_or_die(const char *path, const char *mode)
1288 {
1289         FILE *filep = fopen(path, "r");
1290         if (!filep)
1291                 err(1, "%s: open failed", path);
1292         return filep;
1293 }
1294
1295 /*
1296  * Parse a file containing a single int.
1297  */
1298 int parse_int_file(const char *fmt, ...)
1299 {
1300         va_list args;
1301         char path[PATH_MAX];
1302         FILE *filep;
1303         int value;
1304
1305         va_start(args, fmt);
1306         vsnprintf(path, sizeof(path), fmt, args);
1307         va_end(args);
1308         filep = fopen_or_die(path, "r");
1309         if (fscanf(filep, "%d", &value) != 1)
1310                 err(1, "%s: failed to parse number from file", path);
1311         fclose(filep);
1312         return value;
1313 }
1314
1315 /*
1316  * cpu_is_first_sibling_in_core(cpu)
1317  * return 1 if given CPU is 1st HT sibling in the core
1318  */
1319 int cpu_is_first_sibling_in_core(int cpu)
1320 {
1321         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1322 }
1323
1324 /*
1325  * cpu_is_first_core_in_package(cpu)
1326  * return 1 if given CPU is 1st core in package
1327  */
1328 int cpu_is_first_core_in_package(int cpu)
1329 {
1330         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1331 }
1332
1333 int get_physical_package_id(int cpu)
1334 {
1335         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1336 }
1337
1338 int get_core_id(int cpu)
1339 {
1340         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1341 }
1342
1343 int get_num_ht_siblings(int cpu)
1344 {
1345         char path[80];
1346         FILE *filep;
1347         int sib1, sib2;
1348         int matches;
1349         char character;
1350
1351         sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1352         filep = fopen_or_die(path, "r");
1353         /*
1354          * file format:
1355          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1356          * otherwinse 1 sibling (self).
1357          */
1358         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1359
1360         fclose(filep);
1361
1362         if (matches == 3)
1363                 return 2;
1364         else
1365                 return 1;
1366 }
1367
1368 /*
1369  * run func(thread, core, package) in topology order
1370  * skip non-present cpus
1371  */
1372
1373 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1374         struct pkg_data *, struct thread_data *, struct core_data *,
1375         struct pkg_data *), struct thread_data *thread_base,
1376         struct core_data *core_base, struct pkg_data *pkg_base,
1377         struct thread_data *thread_base2, struct core_data *core_base2,
1378         struct pkg_data *pkg_base2)
1379 {
1380         int retval, pkg_no, core_no, thread_no;
1381
1382         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1383                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1384                         for (thread_no = 0; thread_no <
1385                                 topo.num_threads_per_core; ++thread_no) {
1386                                 struct thread_data *t, *t2;
1387                                 struct core_data *c, *c2;
1388                                 struct pkg_data *p, *p2;
1389
1390                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1391
1392                                 if (cpu_is_not_present(t->cpu_id))
1393                                         continue;
1394
1395                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1396
1397                                 c = GET_CORE(core_base, core_no, pkg_no);
1398                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
1399
1400                                 p = GET_PKG(pkg_base, pkg_no);
1401                                 p2 = GET_PKG(pkg_base2, pkg_no);
1402
1403                                 retval = func(t, c, p, t2, c2, p2);
1404                                 if (retval)
1405                                         return retval;
1406                         }
1407                 }
1408         }
1409         return 0;
1410 }
1411
1412 /*
1413  * run func(cpu) on every cpu in /proc/stat
1414  * return max_cpu number
1415  */
1416 int for_all_proc_cpus(int (func)(int))
1417 {
1418         FILE *fp;
1419         int cpu_num;
1420         int retval;
1421
1422         fp = fopen_or_die(proc_stat, "r");
1423
1424         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1425         if (retval != 0)
1426                 err(1, "%s: failed to parse format", proc_stat);
1427
1428         while (1) {
1429                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1430                 if (retval != 1)
1431                         break;
1432
1433                 retval = func(cpu_num);
1434                 if (retval) {
1435                         fclose(fp);
1436                         return(retval);
1437                 }
1438         }
1439         fclose(fp);
1440         return 0;
1441 }
1442
1443 void re_initialize(void)
1444 {
1445         free_all_buffers();
1446         setup_all_buffers();
1447         printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1448 }
1449
1450
1451 /*
1452  * count_cpus()
1453  * remember the last one seen, it will be the max
1454  */
1455 int count_cpus(int cpu)
1456 {
1457         if (topo.max_cpu_num < cpu)
1458                 topo.max_cpu_num = cpu;
1459
1460         topo.num_cpus += 1;
1461         return 0;
1462 }
1463 int mark_cpu_present(int cpu)
1464 {
1465         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1466         return 0;
1467 }
1468
1469 void turbostat_loop()
1470 {
1471         int retval;
1472         int restarted = 0;
1473
1474 restart:
1475         restarted++;
1476
1477         retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1478         if (retval < -1) {
1479                 exit(retval);
1480         } else if (retval == -1) {
1481                 if (restarted > 1) {
1482                         exit(retval);
1483                 }
1484                 re_initialize();
1485                 goto restart;
1486         }
1487         restarted = 0;
1488         gettimeofday(&tv_even, (struct timezone *)NULL);
1489
1490         while (1) {
1491                 if (for_all_proc_cpus(cpu_is_not_present)) {
1492                         re_initialize();
1493                         goto restart;
1494                 }
1495                 sleep(interval_sec);
1496                 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1497                 if (retval < -1) {
1498                         exit(retval);
1499                 } else if (retval == -1) {
1500                         re_initialize();
1501                         goto restart;
1502                 }
1503                 gettimeofday(&tv_odd, (struct timezone *)NULL);
1504                 timersub(&tv_odd, &tv_even, &tv_delta);
1505                 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1506                 compute_average(EVEN_COUNTERS);
1507                 format_all_counters(EVEN_COUNTERS);
1508                 flush_stdout();
1509                 sleep(interval_sec);
1510                 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1511                 if (retval < -1) {
1512                         exit(retval);
1513                 } else if (retval == -1) {
1514                         re_initialize();
1515                         goto restart;
1516                 }
1517                 gettimeofday(&tv_even, (struct timezone *)NULL);
1518                 timersub(&tv_even, &tv_odd, &tv_delta);
1519                 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1520                 compute_average(ODD_COUNTERS);
1521                 format_all_counters(ODD_COUNTERS);
1522                 flush_stdout();
1523         }
1524 }
1525
1526 void check_dev_msr()
1527 {
1528         struct stat sb;
1529
1530         if (stat("/dev/cpu/0/msr", &sb))
1531                 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
1532 }
1533
1534 void check_permissions()
1535 {
1536         struct __user_cap_header_struct cap_header_data;
1537         cap_user_header_t cap_header = &cap_header_data;
1538         struct __user_cap_data_struct cap_data_data;
1539         cap_user_data_t cap_data = &cap_data_data;
1540         extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
1541         int do_exit = 0;
1542
1543         /* check for CAP_SYS_RAWIO */
1544         cap_header->pid = getpid();
1545         cap_header->version = _LINUX_CAPABILITY_VERSION;
1546         if (capget(cap_header, cap_data) < 0)
1547                 err(-6, "capget(2) failed");
1548
1549         if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1550                 do_exit++;
1551                 warnx("capget(CAP_SYS_RAWIO) failed,"
1552                         " try \"# setcap cap_sys_rawio=ep %s\"", progname);
1553         }
1554
1555         /* test file permissions */
1556         if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1557                 do_exit++;
1558                 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1559         }
1560
1561         /* if all else fails, thell them to be root */
1562         if (do_exit)
1563                 if (getuid() != 0)
1564                         warnx("... or simply run as root");
1565
1566         if (do_exit)
1567                 exit(-6);
1568 }
1569
1570 /*
1571  * NHM adds support for additional MSRs:
1572  *
1573  * MSR_SMI_COUNT                   0x00000034
1574  *
1575  * MSR_NHM_PLATFORM_INFO           0x000000ce
1576  * MSR_NHM_SNB_PKG_CST_CFG_CTL     0x000000e2
1577  *
1578  * MSR_PKG_C3_RESIDENCY            0x000003f8
1579  * MSR_PKG_C6_RESIDENCY            0x000003f9
1580  * MSR_CORE_C3_RESIDENCY           0x000003fc
1581  * MSR_CORE_C6_RESIDENCY           0x000003fd
1582  *
1583  * Side effect:
1584  * sets global pkg_cstate_limit to decode MSR_NHM_SNB_PKG_CST_CFG_CTL
1585  */
1586 int probe_nhm_msrs(unsigned int family, unsigned int model)
1587 {
1588         unsigned long long msr;
1589         int *pkg_cstate_limits;
1590
1591         if (!genuine_intel)
1592                 return 0;
1593
1594         if (family != 6)
1595                 return 0;
1596
1597         switch (model) {
1598         case 0x1A:      /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1599         case 0x1E:      /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1600         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
1601         case 0x25:      /* Westmere Client - Clarkdale, Arrandale */
1602         case 0x2C:      /* Westmere EP - Gulftown */
1603         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1604         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1605                 pkg_cstate_limits = nhm_pkg_cstate_limits;
1606                 break;
1607         case 0x2A:      /* SNB */
1608         case 0x2D:      /* SNB Xeon */
1609         case 0x3A:      /* IVB */
1610         case 0x3E:      /* IVB Xeon */
1611                 pkg_cstate_limits = snb_pkg_cstate_limits;
1612                 break;
1613         case 0x3C:      /* HSW */
1614         case 0x3F:      /* HSX */
1615         case 0x45:      /* HSW */
1616         case 0x46:      /* HSW */
1617         case 0x3D:      /* BDW */
1618         case 0x47:      /* BDW */
1619         case 0x4F:      /* BDX */
1620         case 0x56:      /* BDX-DE */
1621                 pkg_cstate_limits = hsw_pkg_cstate_limits;
1622                 break;
1623         case 0x37:      /* BYT */
1624         case 0x4D:      /* AVN */
1625                 pkg_cstate_limits = slv_pkg_cstate_limits;
1626                 break;
1627         case 0x4C:      /* AMT */
1628                 pkg_cstate_limits = amt_pkg_cstate_limits;
1629                 break;
1630         case 0x57:      /* PHI */
1631                 pkg_cstate_limits = phi_pkg_cstate_limits;
1632                 break;
1633         default:
1634                 return 0;
1635         }
1636         get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1637
1638         pkg_cstate_limit = pkg_cstate_limits[msr & 0x7];
1639
1640         return 1;
1641 }
1642 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
1643 {
1644         switch (model) {
1645         /* Nehalem compatible, but do not include turbo-ratio limit support */
1646         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1647         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1648                 return 0;
1649         default:
1650                 return 1;
1651         }
1652 }
1653 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1654 {
1655         if (!genuine_intel)
1656                 return 0;
1657
1658         if (family != 6)
1659                 return 0;
1660
1661         switch (model) {
1662         case 0x3E:      /* IVB Xeon */
1663         case 0x3F:      /* HSW Xeon */
1664                 return 1;
1665         default:
1666                 return 0;
1667         }
1668 }
1669 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
1670 {
1671         if (!genuine_intel)
1672                 return 0;
1673
1674         if (family != 6)
1675                 return 0;
1676
1677         switch (model) {
1678         case 0x3F:      /* HSW Xeon */
1679                 return 1;
1680         default:
1681                 return 0;
1682         }
1683 }
1684
1685 static void
1686 dump_cstate_pstate_config_info(family, model)
1687 {
1688         if (!do_nhm_platform_info)
1689                 return;
1690
1691         dump_nhm_platform_info();
1692
1693         if (has_hsw_turbo_ratio_limit(family, model))
1694                 dump_hsw_turbo_ratio_limits();
1695
1696         if (has_ivt_turbo_ratio_limit(family, model))
1697                 dump_ivt_turbo_ratio_limits();
1698
1699         if (has_nhm_turbo_ratio_limit(family, model))
1700                 dump_nhm_turbo_ratio_limits();
1701
1702         dump_nhm_cst_cfg();
1703 }
1704
1705
1706 /*
1707  * print_epb()
1708  * Decode the ENERGY_PERF_BIAS MSR
1709  */
1710 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1711 {
1712         unsigned long long msr;
1713         char *epb_string;
1714         int cpu;
1715
1716         if (!has_epb)
1717                 return 0;
1718
1719         cpu = t->cpu_id;
1720
1721         /* EPB is per-package */
1722         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1723                 return 0;
1724
1725         if (cpu_migrate(cpu)) {
1726                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1727                 return -1;
1728         }
1729
1730         if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
1731                 return 0;
1732
1733         switch (msr & 0x7) {
1734         case ENERGY_PERF_BIAS_PERFORMANCE:
1735                 epb_string = "performance";
1736                 break;
1737         case ENERGY_PERF_BIAS_NORMAL:
1738                 epb_string = "balanced";
1739                 break;
1740         case ENERGY_PERF_BIAS_POWERSAVE:
1741                 epb_string = "powersave";
1742                 break;
1743         default:
1744                 epb_string = "custom";
1745                 break;
1746         }
1747         fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
1748
1749         return 0;
1750 }
1751
1752 /*
1753  * print_perf_limit()
1754  */
1755 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1756 {
1757         unsigned long long msr;
1758         int cpu;
1759
1760         cpu = t->cpu_id;
1761
1762         /* per-package */
1763         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1764                 return 0;
1765
1766         if (cpu_migrate(cpu)) {
1767                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1768                 return -1;
1769         }
1770
1771         if (do_core_perf_limit_reasons) {
1772                 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
1773                 fprintf(stderr, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1774                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
1775                         (msr & 1 << 15) ? "bit15, " : "",
1776                         (msr & 1 << 14) ? "bit14, " : "",
1777                         (msr & 1 << 13) ? "Transitions, " : "",
1778                         (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
1779                         (msr & 1 << 11) ? "PkgPwrL2, " : "",
1780                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1781                         (msr & 1 << 9) ? "CorePwr, " : "",
1782                         (msr & 1 << 8) ? "Amps, " : "",
1783                         (msr & 1 << 6) ? "VR-Therm, " : "",
1784                         (msr & 1 << 5) ? "Auto-HWP, " : "",
1785                         (msr & 1 << 4) ? "Graphics, " : "",
1786                         (msr & 1 << 2) ? "bit2, " : "",
1787                         (msr & 1 << 1) ? "ThermStatus, " : "",
1788                         (msr & 1 << 0) ? "PROCHOT, " : "");
1789                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
1790                         (msr & 1 << 31) ? "bit31, " : "",
1791                         (msr & 1 << 30) ? "bit30, " : "",
1792                         (msr & 1 << 29) ? "Transitions, " : "",
1793                         (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
1794                         (msr & 1 << 27) ? "PkgPwrL2, " : "",
1795                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1796                         (msr & 1 << 25) ? "CorePwr, " : "",
1797                         (msr & 1 << 24) ? "Amps, " : "",
1798                         (msr & 1 << 22) ? "VR-Therm, " : "",
1799                         (msr & 1 << 21) ? "Auto-HWP, " : "",
1800                         (msr & 1 << 20) ? "Graphics, " : "",
1801                         (msr & 1 << 18) ? "bit18, " : "",
1802                         (msr & 1 << 17) ? "ThermStatus, " : "",
1803                         (msr & 1 << 16) ? "PROCHOT, " : "");
1804
1805         }
1806         if (do_gfx_perf_limit_reasons) {
1807                 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
1808                 fprintf(stderr, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1809                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s)",
1810                         (msr & 1 << 0) ? "PROCHOT, " : "",
1811                         (msr & 1 << 1) ? "ThermStatus, " : "",
1812                         (msr & 1 << 4) ? "Graphics, " : "",
1813                         (msr & 1 << 6) ? "VR-Therm, " : "",
1814                         (msr & 1 << 8) ? "Amps, " : "",
1815                         (msr & 1 << 9) ? "GFXPwr, " : "",
1816                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1817                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
1818                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s)\n",
1819                         (msr & 1 << 16) ? "PROCHOT, " : "",
1820                         (msr & 1 << 17) ? "ThermStatus, " : "",
1821                         (msr & 1 << 20) ? "Graphics, " : "",
1822                         (msr & 1 << 22) ? "VR-Therm, " : "",
1823                         (msr & 1 << 24) ? "Amps, " : "",
1824                         (msr & 1 << 25) ? "GFXPwr, " : "",
1825                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1826                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
1827         }
1828         if (do_ring_perf_limit_reasons) {
1829                 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
1830                 fprintf(stderr, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1831                 fprintf(stderr, " (Active: %s%s%s%s%s%s)",
1832                         (msr & 1 << 0) ? "PROCHOT, " : "",
1833                         (msr & 1 << 1) ? "ThermStatus, " : "",
1834                         (msr & 1 << 6) ? "VR-Therm, " : "",
1835                         (msr & 1 << 8) ? "Amps, " : "",
1836                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1837                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
1838                 fprintf(stderr, " (Logged: %s%s%s%s%s%s)\n",
1839                         (msr & 1 << 16) ? "PROCHOT, " : "",
1840                         (msr & 1 << 17) ? "ThermStatus, " : "",
1841                         (msr & 1 << 22) ? "VR-Therm, " : "",
1842                         (msr & 1 << 24) ? "Amps, " : "",
1843                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1844                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
1845         }
1846         return 0;
1847 }
1848
1849 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
1850 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
1851
1852 double get_tdp(model)
1853 {
1854         unsigned long long msr;
1855
1856         if (do_rapl & RAPL_PKG_POWER_INFO)
1857                 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
1858                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
1859
1860         switch (model) {
1861         case 0x37:
1862         case 0x4D:
1863                 return 30.0;
1864         default:
1865                 return 135.0;
1866         }
1867 }
1868
1869
1870 /*
1871  * rapl_probe()
1872  *
1873  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
1874  */
1875 void rapl_probe(unsigned int family, unsigned int model)
1876 {
1877         unsigned long long msr;
1878         unsigned int time_unit;
1879         double tdp;
1880
1881         if (!genuine_intel)
1882                 return;
1883
1884         if (family != 6)
1885                 return;
1886
1887         switch (model) {
1888         case 0x2A:
1889         case 0x3A:
1890         case 0x3C:      /* HSW */
1891         case 0x45:      /* HSW */
1892         case 0x46:      /* HSW */
1893         case 0x3D:      /* BDW */
1894         case 0x47:      /* BDW */
1895                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1896                 break;
1897         case 0x3F:      /* HSX */
1898         case 0x4F:      /* BDX */
1899         case 0x56:      /* BDX-DE */
1900                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1901                 break;
1902         case 0x2D:
1903         case 0x3E:
1904                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1905                 break;
1906         case 0x37:      /* BYT */
1907         case 0x4D:      /* AVN */
1908                 do_rapl = RAPL_PKG | RAPL_CORES ;
1909                 break;
1910         default:
1911                 return;
1912         }
1913
1914         /* units on package 0, verify later other packages match */
1915         if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1916                 return;
1917
1918         rapl_power_units = 1.0 / (1 << (msr & 0xF));
1919         if (model == 0x37)
1920                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1921         else
1922                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1923
1924         time_unit = msr >> 16 & 0xF;
1925         if (time_unit == 0)
1926                 time_unit = 0xA;
1927
1928         rapl_time_units = 1.0 / (1 << (time_unit));
1929
1930         tdp = get_tdp(model);
1931
1932         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1933         if (debug)
1934                 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1935
1936         return;
1937 }
1938
1939 void perf_limit_reasons_probe(family, model)
1940 {
1941         if (!genuine_intel)
1942                 return;
1943
1944         if (family != 6)
1945                 return;
1946
1947         switch (model) {
1948         case 0x3C:      /* HSW */
1949         case 0x45:      /* HSW */
1950         case 0x46:      /* HSW */
1951                 do_gfx_perf_limit_reasons = 1;
1952         case 0x3F:      /* HSX */
1953                 do_core_perf_limit_reasons = 1;
1954                 do_ring_perf_limit_reasons = 1;
1955         default:
1956                 return;
1957         }
1958 }
1959
1960 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1961 {
1962         unsigned long long msr;
1963         unsigned int dts;
1964         int cpu;
1965
1966         if (!(do_dts || do_ptm))
1967                 return 0;
1968
1969         cpu = t->cpu_id;
1970
1971         /* DTS is per-core, no need to print for each thread */
1972         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1973                 return 0;
1974
1975         if (cpu_migrate(cpu)) {
1976                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1977                 return -1;
1978         }
1979
1980         if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
1981                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1982                         return 0;
1983
1984                 dts = (msr >> 16) & 0x7F;
1985                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
1986                         cpu, msr, tcc_activation_temp - dts);
1987
1988 #ifdef  THERM_DEBUG
1989                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
1990                         return 0;
1991
1992                 dts = (msr >> 16) & 0x7F;
1993                 dts2 = (msr >> 8) & 0x7F;
1994                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1995                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1996 #endif
1997         }
1998
1999
2000         if (do_dts) {
2001                 unsigned int resolution;
2002
2003                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
2004                         return 0;
2005
2006                 dts = (msr >> 16) & 0x7F;
2007                 resolution = (msr >> 27) & 0xF;
2008                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
2009                         cpu, msr, tcc_activation_temp - dts, resolution);
2010
2011 #ifdef THERM_DEBUG
2012                 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
2013                         return 0;
2014
2015                 dts = (msr >> 16) & 0x7F;
2016                 dts2 = (msr >> 8) & 0x7F;
2017                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2018                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2019 #endif
2020         }
2021
2022         return 0;
2023 }
2024
2025 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
2026 {
2027         fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
2028                 cpu, label,
2029                 ((msr >> 15) & 1) ? "EN" : "DIS",
2030                 ((msr >> 0) & 0x7FFF) * rapl_power_units,
2031                 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
2032                 (((msr >> 16) & 1) ? "EN" : "DIS"));
2033
2034         return;
2035 }
2036
2037 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2038 {
2039         unsigned long long msr;
2040         int cpu;
2041
2042         if (!do_rapl)
2043                 return 0;
2044
2045         /* RAPL counters are per package, so print only for 1st thread/package */
2046         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2047                 return 0;
2048
2049         cpu = t->cpu_id;
2050         if (cpu_migrate(cpu)) {
2051                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2052                 return -1;
2053         }
2054
2055         if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
2056                 return -1;
2057
2058         if (debug) {
2059                 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
2060                         "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
2061                         rapl_power_units, rapl_energy_units, rapl_time_units);
2062         }
2063         if (do_rapl & RAPL_PKG_POWER_INFO) {
2064
2065                 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
2066                         return -5;
2067
2068
2069                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2070                         cpu, msr,
2071                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2072                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2073                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2074                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2075
2076         }
2077         if (do_rapl & RAPL_PKG) {
2078
2079                 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
2080                         return -9;
2081
2082                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
2083                         cpu, msr, (msr >> 63) & 1 ? "": "UN");
2084
2085                 print_power_limit_msr(cpu, msr, "PKG Limit #1");
2086                 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
2087                         cpu,
2088                         ((msr >> 47) & 1) ? "EN" : "DIS",
2089                         ((msr >> 32) & 0x7FFF) * rapl_power_units,
2090                         (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
2091                         ((msr >> 48) & 1) ? "EN" : "DIS");
2092         }
2093
2094         if (do_rapl & RAPL_DRAM) {
2095                 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
2096                         return -6;
2097
2098
2099                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2100                         cpu, msr,
2101                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2102                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2103                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2104                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2105
2106
2107                 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
2108                         return -9;
2109                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
2110                                 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2111
2112                 print_power_limit_msr(cpu, msr, "DRAM Limit");
2113         }
2114         if (do_rapl & RAPL_CORE_POLICY) {
2115                 if (debug) {
2116                         if (get_msr(cpu, MSR_PP0_POLICY, &msr))
2117                                 return -7;
2118
2119                         fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
2120                 }
2121         }
2122         if (do_rapl & RAPL_CORES) {
2123                 if (debug) {
2124
2125                         if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
2126                                 return -9;
2127                         fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2128                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2129                         print_power_limit_msr(cpu, msr, "Cores Limit");
2130                 }
2131         }
2132         if (do_rapl & RAPL_GFX) {
2133                 if (debug) {
2134                         if (get_msr(cpu, MSR_PP1_POLICY, &msr))
2135                                 return -8;
2136
2137                         fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
2138
2139                         if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
2140                                 return -9;
2141                         fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2142                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2143                         print_power_limit_msr(cpu, msr, "GFX Limit");
2144                 }
2145         }
2146         return 0;
2147 }
2148
2149 /*
2150  * SNB adds support for additional MSRs:
2151  *
2152  * MSR_PKG_C7_RESIDENCY            0x000003fa
2153  * MSR_CORE_C7_RESIDENCY           0x000003fe
2154  * MSR_PKG_C2_RESIDENCY            0x0000060d
2155  */
2156
2157 int has_snb_msrs(unsigned int family, unsigned int model)
2158 {
2159         if (!genuine_intel)
2160                 return 0;
2161
2162         switch (model) {
2163         case 0x2A:
2164         case 0x2D:
2165         case 0x3A:      /* IVB */
2166         case 0x3E:      /* IVB Xeon */
2167         case 0x3C:      /* HSW */
2168         case 0x3F:      /* HSW */
2169         case 0x45:      /* HSW */
2170         case 0x46:      /* HSW */
2171         case 0x3D:      /* BDW */
2172         case 0x47:      /* BDW */
2173         case 0x4F:      /* BDX */
2174         case 0x56:      /* BDX-DE */
2175                 return 1;
2176         }
2177         return 0;
2178 }
2179
2180 /*
2181  * HSW adds support for additional MSRs:
2182  *
2183  * MSR_PKG_C8_RESIDENCY            0x00000630
2184  * MSR_PKG_C9_RESIDENCY            0x00000631
2185  * MSR_PKG_C10_RESIDENCY           0x00000632
2186  */
2187 int has_hsw_msrs(unsigned int family, unsigned int model)
2188 {
2189         if (!genuine_intel)
2190                 return 0;
2191
2192         switch (model) {
2193         case 0x45:      /* HSW */
2194         case 0x3D:      /* BDW */
2195                 return 1;
2196         }
2197         return 0;
2198 }
2199
2200
2201 int is_slm(unsigned int family, unsigned int model)
2202 {
2203         if (!genuine_intel)
2204                 return 0;
2205         switch (model) {
2206         case 0x37:      /* BYT */
2207         case 0x4D:      /* AVN */
2208                 return 1;
2209         }
2210         return 0;
2211 }
2212
2213 #define SLM_BCLK_FREQS 5
2214 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
2215
2216 double slm_bclk(void)
2217 {
2218         unsigned long long msr = 3;
2219         unsigned int i;
2220         double freq;
2221
2222         if (get_msr(0, MSR_FSB_FREQ, &msr))
2223                 fprintf(stderr, "SLM BCLK: unknown\n");
2224
2225         i = msr & 0xf;
2226         if (i >= SLM_BCLK_FREQS) {
2227                 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
2228                 msr = 3;
2229         }
2230         freq = slm_freq_table[i];
2231
2232         fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
2233
2234         return freq;
2235 }
2236
2237 double discover_bclk(unsigned int family, unsigned int model)
2238 {
2239         if (has_snb_msrs(family, model))
2240                 return 100.00;
2241         else if (is_slm(family, model))
2242                 return slm_bclk();
2243         else
2244                 return 133.33;
2245 }
2246
2247 /*
2248  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
2249  * the Thermal Control Circuit (TCC) activates.
2250  * This is usually equal to tjMax.
2251  *
2252  * Older processors do not have this MSR, so there we guess,
2253  * but also allow cmdline over-ride with -T.
2254  *
2255  * Several MSR temperature values are in units of degrees-C
2256  * below this value, including the Digital Thermal Sensor (DTS),
2257  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
2258  */
2259 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2260 {
2261         unsigned long long msr;
2262         unsigned int target_c_local;
2263         int cpu;
2264
2265         /* tcc_activation_temp is used only for dts or ptm */
2266         if (!(do_dts || do_ptm))
2267                 return 0;
2268
2269         /* this is a per-package concept */
2270         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2271                 return 0;
2272
2273         cpu = t->cpu_id;
2274         if (cpu_migrate(cpu)) {
2275                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2276                 return -1;
2277         }
2278
2279         if (tcc_activation_temp_override != 0) {
2280                 tcc_activation_temp = tcc_activation_temp_override;
2281                 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
2282                         cpu, tcc_activation_temp);
2283                 return 0;
2284         }
2285
2286         /* Temperature Target MSR is Nehalem and newer only */
2287         if (!do_nhm_platform_info)
2288                 goto guess;
2289
2290         if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
2291                 goto guess;
2292
2293         target_c_local = (msr >> 16) & 0xFF;
2294
2295         if (debug)
2296                 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2297                         cpu, msr, target_c_local);
2298
2299         if (!target_c_local)
2300                 goto guess;
2301
2302         tcc_activation_temp = target_c_local;
2303
2304         return 0;
2305
2306 guess:
2307         tcc_activation_temp = TJMAX_DEFAULT;
2308         fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2309                 cpu, tcc_activation_temp);
2310
2311         return 0;
2312 }
2313 void process_cpuid()
2314 {
2315         unsigned int eax, ebx, ecx, edx, max_level;
2316         unsigned int fms, family, model, stepping;
2317
2318         eax = ebx = ecx = edx = 0;
2319
2320         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2321
2322         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2323                 genuine_intel = 1;
2324
2325         if (debug)
2326                 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2327                         (char *)&ebx, (char *)&edx, (char *)&ecx);
2328
2329         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2330         family = (fms >> 8) & 0xf;
2331         model = (fms >> 4) & 0xf;
2332         stepping = fms & 0xf;
2333         if (family == 6 || family == 0xf)
2334                 model += ((fms >> 16) & 0xf) << 4;
2335
2336         if (debug)
2337                 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2338                         max_level, family, model, stepping, family, model, stepping);
2339
2340         if (!(edx & (1 << 5)))
2341                 errx(1, "CPUID: no MSR");
2342
2343         /*
2344          * check max extended function levels of CPUID.
2345          * This is needed to check for invariant TSC.
2346          * This check is valid for both Intel and AMD.
2347          */
2348         ebx = ecx = edx = 0;
2349         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2350
2351         if (max_level >= 0x80000007) {
2352
2353                 /*
2354                  * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2355                  * this check is valid for both Intel and AMD
2356                  */
2357                 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2358                 has_invariant_tsc = edx & (1 << 8);
2359         }
2360
2361         /*
2362          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2363          * this check is valid for both Intel and AMD
2364          */
2365
2366         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2367         has_aperf = ecx & (1 << 0);
2368         do_dts = eax & (1 << 0);
2369         do_ptm = eax & (1 << 6);
2370         has_epb = ecx & (1 << 3);
2371
2372         if (debug)
2373                 fprintf(stderr, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sEPB\n",
2374                         has_aperf ? "" : "No ",
2375                         do_dts ? "" : "No ",
2376                         do_ptm ? "" : "No ",
2377                         has_epb ? "" : "No ");
2378
2379         do_nhm_platform_info = do_nhm_cstates = do_smi = probe_nhm_msrs(family, model);
2380         do_snb_cstates = has_snb_msrs(family, model);
2381         do_pc2 = do_snb_cstates && (pkg_cstate_limit >= PCL__2);
2382         do_pc3 = (pkg_cstate_limit >= PCL__3);
2383         do_pc6 = (pkg_cstate_limit >= PCL__6);
2384         do_pc7 = do_snb_cstates && (pkg_cstate_limit >= PCL__7);
2385         do_c8_c9_c10 = has_hsw_msrs(family, model);
2386         do_slm_cstates = is_slm(family, model);
2387         bclk = discover_bclk(family, model);
2388
2389         rapl_probe(family, model);
2390         perf_limit_reasons_probe(family, model);
2391
2392         if (debug)
2393                 dump_cstate_pstate_config_info();
2394
2395         return;
2396 }
2397
2398 void help()
2399 {
2400         fprintf(stderr,
2401         "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2402         "\n"
2403         "Turbostat forks the specified COMMAND and prints statistics\n"
2404         "when COMMAND completes.\n"
2405         "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2406         "to print statistics, until interrupted.\n"
2407         "--debug        run in \"debug\" mode\n"
2408         "--interval sec Override default 5-second measurement interval\n"
2409         "--help         print this help message\n"
2410         "--counter msr  print 32-bit counter at address \"msr\"\n"
2411         "--Counter msr  print 64-bit Counter at address \"msr\"\n"
2412         "--msr msr      print 32-bit value at address \"msr\"\n"
2413         "--MSR msr      print 64-bit Value at address \"msr\"\n"
2414         "--version      print version information\n"
2415         "\n"
2416         "For more help, run \"man turbostat\"\n");
2417 }
2418
2419
2420 /*
2421  * in /dev/cpu/ return success for names that are numbers
2422  * ie. filter out ".", "..", "microcode".
2423  */
2424 int dir_filter(const struct dirent *dirp)
2425 {
2426         if (isdigit(dirp->d_name[0]))
2427                 return 1;
2428         else
2429                 return 0;
2430 }
2431
2432 int open_dev_cpu_msr(int dummy1)
2433 {
2434         return 0;
2435 }
2436
2437 void topology_probe()
2438 {
2439         int i;
2440         int max_core_id = 0;
2441         int max_package_id = 0;
2442         int max_siblings = 0;
2443         struct cpu_topology {
2444                 int core_id;
2445                 int physical_package_id;
2446         } *cpus;
2447
2448         /* Initialize num_cpus, max_cpu_num */
2449         topo.num_cpus = 0;
2450         topo.max_cpu_num = 0;
2451         for_all_proc_cpus(count_cpus);
2452         if (!summary_only && topo.num_cpus > 1)
2453                 show_cpu = 1;
2454
2455         if (debug > 1)
2456                 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2457
2458         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
2459         if (cpus == NULL)
2460                 err(1, "calloc cpus");
2461
2462         /*
2463          * Allocate and initialize cpu_present_set
2464          */
2465         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2466         if (cpu_present_set == NULL)
2467                 err(3, "CPU_ALLOC");
2468         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2469         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2470         for_all_proc_cpus(mark_cpu_present);
2471
2472         /*
2473          * Allocate and initialize cpu_affinity_set
2474          */
2475         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2476         if (cpu_affinity_set == NULL)
2477                 err(3, "CPU_ALLOC");
2478         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2479         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2480
2481
2482         /*
2483          * For online cpus
2484          * find max_core_id, max_package_id
2485          */
2486         for (i = 0; i <= topo.max_cpu_num; ++i) {
2487                 int siblings;
2488
2489                 if (cpu_is_not_present(i)) {
2490                         if (debug > 1)
2491                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2492                         continue;
2493                 }
2494                 cpus[i].core_id = get_core_id(i);
2495                 if (cpus[i].core_id > max_core_id)
2496                         max_core_id = cpus[i].core_id;
2497
2498                 cpus[i].physical_package_id = get_physical_package_id(i);
2499                 if (cpus[i].physical_package_id > max_package_id)
2500                         max_package_id = cpus[i].physical_package_id;
2501
2502                 siblings = get_num_ht_siblings(i);
2503                 if (siblings > max_siblings)
2504                         max_siblings = siblings;
2505                 if (debug > 1)
2506                         fprintf(stderr, "cpu %d pkg %d core %d\n",
2507                                 i, cpus[i].physical_package_id, cpus[i].core_id);
2508         }
2509         topo.num_cores_per_pkg = max_core_id + 1;
2510         if (debug > 1)
2511                 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2512                         max_core_id, topo.num_cores_per_pkg);
2513         if (debug && !summary_only && topo.num_cores_per_pkg > 1)
2514                 show_core = 1;
2515
2516         topo.num_packages = max_package_id + 1;
2517         if (debug > 1)
2518                 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2519                         max_package_id, topo.num_packages);
2520         if (debug && !summary_only && topo.num_packages > 1)
2521                 show_pkg = 1;
2522
2523         topo.num_threads_per_core = max_siblings;
2524         if (debug > 1)
2525                 fprintf(stderr, "max_siblings %d\n", max_siblings);
2526
2527         free(cpus);
2528 }
2529
2530 void
2531 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2532 {
2533         int i;
2534
2535         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2536                 topo.num_packages, sizeof(struct thread_data));
2537         if (*t == NULL)
2538                 goto error;
2539
2540         for (i = 0; i < topo.num_threads_per_core *
2541                 topo.num_cores_per_pkg * topo.num_packages; i++)
2542                 (*t)[i].cpu_id = -1;
2543
2544         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2545                 sizeof(struct core_data));
2546         if (*c == NULL)
2547                 goto error;
2548
2549         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2550                 (*c)[i].core_id = -1;
2551
2552         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2553         if (*p == NULL)
2554                 goto error;
2555
2556         for (i = 0; i < topo.num_packages; i++)
2557                 (*p)[i].package_id = i;
2558
2559         return;
2560 error:
2561         err(1, "calloc counters");
2562 }
2563 /*
2564  * init_counter()
2565  *
2566  * set cpu_id, core_num, pkg_num
2567  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2568  *
2569  * increment topo.num_cores when 1st core in pkg seen
2570  */
2571 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2572         struct pkg_data *pkg_base, int thread_num, int core_num,
2573         int pkg_num, int cpu_id)
2574 {
2575         struct thread_data *t;
2576         struct core_data *c;
2577         struct pkg_data *p;
2578
2579         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2580         c = GET_CORE(core_base, core_num, pkg_num);
2581         p = GET_PKG(pkg_base, pkg_num);
2582
2583         t->cpu_id = cpu_id;
2584         if (thread_num == 0) {
2585                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2586                 if (cpu_is_first_core_in_package(cpu_id))
2587                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2588         }
2589
2590         c->core_id = core_num;
2591         p->package_id = pkg_num;
2592 }
2593
2594
2595 int initialize_counters(int cpu_id)
2596 {
2597         int my_thread_id, my_core_id, my_package_id;
2598
2599         my_package_id = get_physical_package_id(cpu_id);
2600         my_core_id = get_core_id(cpu_id);
2601
2602         if (cpu_is_first_sibling_in_core(cpu_id)) {
2603                 my_thread_id = 0;
2604                 topo.num_cores++;
2605         } else {
2606                 my_thread_id = 1;
2607         }
2608
2609         init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2610         init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2611         return 0;
2612 }
2613
2614 void allocate_output_buffer()
2615 {
2616         output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
2617         outp = output_buffer;
2618         if (outp == NULL)
2619                 err(-1, "calloc output buffer");
2620 }
2621
2622 void setup_all_buffers(void)
2623 {
2624         topology_probe();
2625         allocate_counters(&thread_even, &core_even, &package_even);
2626         allocate_counters(&thread_odd, &core_odd, &package_odd);
2627         allocate_output_buffer();
2628         for_all_proc_cpus(initialize_counters);
2629 }
2630
2631 void turbostat_init()
2632 {
2633         check_dev_msr();
2634         check_permissions();
2635         process_cpuid();
2636
2637         setup_all_buffers();
2638
2639         if (debug)
2640                 for_all_cpus(print_epb, ODD_COUNTERS);
2641
2642         if (debug)
2643                 for_all_cpus(print_perf_limit, ODD_COUNTERS);
2644
2645         if (debug)
2646                 for_all_cpus(print_rapl, ODD_COUNTERS);
2647
2648         for_all_cpus(set_temperature_target, ODD_COUNTERS);
2649
2650         if (debug)
2651                 for_all_cpus(print_thermal, ODD_COUNTERS);
2652 }
2653
2654 int fork_it(char **argv)
2655 {
2656         pid_t child_pid;
2657         int status;
2658
2659         status = for_all_cpus(get_counters, EVEN_COUNTERS);
2660         if (status)
2661                 exit(status);
2662         /* clear affinity side-effect of get_counters() */
2663         sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
2664         gettimeofday(&tv_even, (struct timezone *)NULL);
2665
2666         child_pid = fork();
2667         if (!child_pid) {
2668                 /* child */
2669                 execvp(argv[0], argv);
2670         } else {
2671
2672                 /* parent */
2673                 if (child_pid == -1)
2674                         err(1, "fork");
2675
2676                 signal(SIGINT, SIG_IGN);
2677                 signal(SIGQUIT, SIG_IGN);
2678                 if (waitpid(child_pid, &status, 0) == -1)
2679                         err(status, "waitpid");
2680         }
2681         /*
2682          * n.b. fork_it() does not check for errors from for_all_cpus()
2683          * because re-starting is problematic when forking
2684          */
2685         for_all_cpus(get_counters, ODD_COUNTERS);
2686         gettimeofday(&tv_odd, (struct timezone *)NULL);
2687         timersub(&tv_odd, &tv_even, &tv_delta);
2688         for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
2689         compute_average(EVEN_COUNTERS);
2690         format_all_counters(EVEN_COUNTERS);
2691         flush_stderr();
2692
2693         fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
2694
2695         return status;
2696 }
2697
2698 int get_and_dump_counters(void)
2699 {
2700         int status;
2701
2702         status = for_all_cpus(get_counters, ODD_COUNTERS);
2703         if (status)
2704                 return status;
2705
2706         status = for_all_cpus(dump_counters, ODD_COUNTERS);
2707         if (status)
2708                 return status;
2709
2710         flush_stdout();
2711
2712         return status;
2713 }
2714
2715 void print_version() {
2716         fprintf(stderr, "turbostat version 4.2 23 Mar, 2015"
2717                 " - Len Brown <lenb@kernel.org>\n");
2718 }
2719
2720 void cmdline(int argc, char **argv)
2721 {
2722         int opt;
2723         int option_index = 0;
2724         static struct option long_options[] = {
2725                 {"Counter",     required_argument,      0, 'C'},
2726                 {"counter",     required_argument,      0, 'c'},
2727                 {"Dump",        no_argument,            0, 'D'},
2728                 {"debug",       no_argument,            0, 'd'},
2729                 {"interval",    required_argument,      0, 'i'},
2730                 {"help",        no_argument,            0, 'h'},
2731                 {"Joules",      no_argument,            0, 'J'},
2732                 {"MSR",         required_argument,      0, 'M'},
2733                 {"msr",         required_argument,      0, 'm'},
2734                 {"Package",     no_argument,            0, 'p'},
2735                 {"processor",   no_argument,            0, 'p'},
2736                 {"Summary",     no_argument,            0, 'S'},
2737                 {"TCC",         required_argument,      0, 'T'},
2738                 {"version",     no_argument,            0, 'v' },
2739                 {0,             0,                      0,  0 }
2740         };
2741
2742         progname = argv[0];
2743
2744         while ((opt = getopt_long_only(argc, argv, "C:c:Ddhi:JM:m:PpST:v",
2745                                 long_options, &option_index)) != -1) {
2746                 switch (opt) {
2747                 case 'C':
2748                         sscanf(optarg, "%x", &extra_delta_offset64);
2749                         break;
2750                 case 'c':
2751                         sscanf(optarg, "%x", &extra_delta_offset32);
2752                         break;
2753                 case 'D':
2754                         dump_only++;
2755                         break;
2756                 case 'd':
2757                         debug++;
2758                         break;
2759                 case 'h':
2760                 default:
2761                         help();
2762                         exit(1);
2763                 case 'i':
2764                         interval_sec = atoi(optarg);
2765                         break;
2766                 case 'J':
2767                         rapl_joules++;
2768                         break;
2769                 case 'M':
2770                         sscanf(optarg, "%x", &extra_msr_offset64);
2771                         break;
2772                 case 'm':
2773                         sscanf(optarg, "%x", &extra_msr_offset32);
2774                         break;
2775                 case 'P':
2776                         show_pkg_only++;
2777                         break;
2778                 case 'p':
2779                         show_core_only++;
2780                         break;
2781                 case 'S':
2782                         summary_only++;
2783                         break;
2784                 case 'T':
2785                         tcc_activation_temp_override = atoi(optarg);
2786                         break;
2787                 case 'v':
2788                         print_version();
2789                         exit(0);
2790                         break;
2791                 }
2792         }
2793 }
2794
2795 int main(int argc, char **argv)
2796 {
2797         cmdline(argc, argv);
2798
2799         if (debug)
2800                 print_version();
2801
2802         turbostat_init();
2803
2804         /* dump counters and exit */
2805         if (dump_only)
2806                 return get_and_dump_counters();
2807
2808         /*
2809          * if any params left, it must be a command to fork
2810          */
2811         if (argc - optind)
2812                 return fork_it(argv + optind);
2813         else
2814                 turbostat_loop();
2815
2816         return 0;
2817 }