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_knl_cstates;
  56 unsigned int do_pc2;
  57 unsigned int do_pc3;
  58 unsigned int do_pc6;
  59 unsigned int do_pc7;
  60 unsigned int do_c8_c9_c10;
  61 unsigned int do_skl_residency;
  62 unsigned int do_slm_cstates;
  63 unsigned int use_c1_residency_msr;
  64 unsigned int has_aperf;
  65 unsigned int has_epb;
  66 unsigned int units = 1000000;   /* MHz etc */
  67 unsigned int genuine_intel;
  68 unsigned int has_invariant_tsc;
  69 unsigned int do_nhm_platform_info;
  70 unsigned int extra_msr_offset32;
  71 unsigned int extra_msr_offset64;
  72 unsigned int extra_delta_offset32;
  73 unsigned int extra_delta_offset64;
  74 unsigned int aperf_mperf_multiplier = 1;
  75 int do_smi;
  76 double bclk;
  77 double base_hz;
  78 double tsc_tweak = 1.0;
  79 unsigned int show_pkg;
  80 unsigned int show_core;
  81 unsigned int show_cpu;
  82 unsigned int show_pkg_only;
  83 unsigned int show_core_only;
  84 char *output_buffer, *outp;
  85 unsigned int do_rapl;
  86 unsigned int do_dts;
  87 unsigned int do_ptm;
  88 unsigned int tcc_activation_temp;
  89 unsigned int tcc_activation_temp_override;
  90 double rapl_power_units, rapl_time_units;
  91 double rapl_dram_energy_units, rapl_energy_units;
  92 double rapl_joule_counter_range;
  93 unsigned int do_core_perf_limit_reasons;
  94 unsigned int do_gfx_perf_limit_reasons;
  95 unsigned int do_ring_perf_limit_reasons;
  96 unsigned int crystal_hz;
  97 unsigned long long tsc_hz;
  98 int base_cpu;
  99
 100 #define RAPL_PKG                (1 << 0)
 101                                         /* 0x610 MSR_PKG_POWER_LIMIT */
 102                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
 103 #define RAPL_PKG_PERF_STATUS    (1 << 1)
 104                                         /* 0x613 MSR_PKG_PERF_STATUS */
 105 #define RAPL_PKG_POWER_INFO     (1 << 2)
 106                                         /* 0x614 MSR_PKG_POWER_INFO */
 107
 108 #define RAPL_DRAM               (1 << 3)
 109                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
 110                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
 111 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
 112                                         /* 0x61b MSR_DRAM_PERF_STATUS */
 113 #define RAPL_DRAM_POWER_INFO    (1 << 5)
 114                                         /* 0x61c MSR_DRAM_POWER_INFO */
 115
 116 #define RAPL_CORES              (1 << 6)
 117                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 118                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 119 #define RAPL_CORE_POLICY        (1 << 7)
 120                                         /* 0x63a MSR_PP0_POLICY */
 121
 122 #define RAPL_GFX                (1 << 8)
 123                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 124                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 125                                         /* 0x642 MSR_PP1_POLICY */
 126 #define TJMAX_DEFAULT   100
 127
 128 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 129
 130 int aperf_mperf_unstable;
 131 int backwards_count;
 132 char *progname;
 133
 134 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 135 size_t cpu_present_setsize, cpu_affinity_setsize;
 136
 137 struct thread_data {
 138         unsigned long long tsc;
 139         unsigned long long aperf;
 140         unsigned long long mperf;
 141         unsigned long long c1;
 142         unsigned long long extra_msr64;
 143         unsigned long long extra_delta64;
 144         unsigned long long extra_msr32;
 145         unsigned long long extra_delta32;
 146         unsigned int smi_count;
 147         unsigned int cpu_id;
 148         unsigned int flags;
 149 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 150 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 151 } *thread_even, *thread_odd;
 152
 153 struct core_data {
 154         unsigned long long c3;
 155         unsigned long long c6;
 156         unsigned long long c7;
 157         unsigned int core_temp_c;
 158         unsigned int core_id;
 159 } *core_even, *core_odd;
 160
 161 struct pkg_data {
 162         unsigned long long pc2;
 163         unsigned long long pc3;
 164         unsigned long long pc6;
 165         unsigned long long pc7;
 166         unsigned long long pc8;
 167         unsigned long long pc9;
 168         unsigned long long pc10;
 169         unsigned long long pkg_wtd_core_c0;
 170         unsigned long long pkg_any_core_c0;
 171         unsigned long long pkg_any_gfxe_c0;
 172         unsigned long long pkg_both_core_gfxe_c0;
 173         unsigned int package_id;
 174         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 175         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 176         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 177         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 178         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 179         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 180         unsigned int pkg_temp_c;
 181
 182 } *package_even, *package_odd;
 183
 184 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 185 #define EVEN_COUNTERS thread_even, core_even, package_even
 186
 187 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 188         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 189                 topo.num_threads_per_core + \
 190                 (core_no) * topo.num_threads_per_core + (thread_no))
 191 #define GET_CORE(core_base, core_no, pkg_no) \
 192         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 193 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 194
 195 struct system_summary {
 196         struct thread_data threads;
 197         struct core_data cores;
 198         struct pkg_data packages;
 199 } sum, average;
 200
 201
 202 struct topo_params {
 203         int num_packages;
 204         int num_cpus;
 205         int num_cores;
 206         int max_cpu_num;
 207         int num_cores_per_pkg;
 208         int num_threads_per_core;
 209 } topo;
 210
 211 struct timeval tv_even, tv_odd, tv_delta;
 212
 213 void setup_all_buffers(void);
 214
 215 int cpu_is_not_present(int cpu)
 216 {
 217         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 218 }
 219 /*
 220  * run func(thread, core, package) in topology order
 221  * skip non-present cpus
 222  */
 223
 224 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 225         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 226 {
 227         int retval, pkg_no, core_no, thread_no;
 228
 229         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 230                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 231                         for (thread_no = 0; thread_no <
 232                                 topo.num_threads_per_core; ++thread_no) {
 233                                 struct thread_data *t;
 234                                 struct core_data *c;
 235                                 struct pkg_data *p;
 236
 237                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 238
 239                                 if (cpu_is_not_present(t->cpu_id))
 240                                         continue;
 241
 242                                 c = GET_CORE(core_base, core_no, pkg_no);
 243                                 p = GET_PKG(pkg_base, pkg_no);
 244
 245                                 retval = func(t, c, p);
 246                                 if (retval)
 247                                         return retval;
 248                         }
 249                 }
 250         }
 251         return 0;
 252 }
 253
 254 int cpu_migrate(int cpu)
 255 {
 256         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 257         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 258         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 259                 return -1;
 260         else
 261                 return 0;
 262 }
 263
 264 int get_msr(int cpu, off_t offset, unsigned long long *msr)
 265 {
 266         ssize_t retval;
 267         char pathname[32];
 268         int fd;
 269
 270         sprintf(pathname, "/dev/cpu/%d/msr", cpu);
 271         fd = open(pathname, O_RDONLY);
 272         if (fd < 0)
 273                 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
 274
 275         retval = pread(fd, msr, sizeof *msr, offset);
 276         close(fd);
 277
 278         if (retval != sizeof *msr)
 279                 err(-1, "%s offset 0x%llx read failed", pathname, (unsigned long long)offset);
 280
 281         return 0;
 282 }
 283
 284 /*
 285  * Example Format w/ field column widths:
 286  *
 287  *  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
 288  * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
 289  */
 290
 291 void print_header(void)
 292 {
 293         if (show_pkg)
 294                 outp += sprintf(outp, " Package");
 295         if (show_core)
 296                 outp += sprintf(outp, "    Core");
 297         if (show_cpu)
 298                 outp += sprintf(outp, "     CPU");
 299         if (has_aperf)
 300                 outp += sprintf(outp, " Avg_MHz");
 301         if (has_aperf)
 302                 outp += sprintf(outp, "   %%Busy");
 303         if (has_aperf)
 304                 outp += sprintf(outp, " Bzy_MHz");
 305         outp += sprintf(outp, " TSC_MHz");
 306
 307         if (extra_delta_offset32)
 308                 outp += sprintf(outp, "  count 0x%03X", extra_delta_offset32);
 309         if (extra_delta_offset64)
 310                 outp += sprintf(outp, "  COUNT 0x%03X", extra_delta_offset64);
 311         if (extra_msr_offset32)
 312                 outp += sprintf(outp, "   MSR 0x%03X", extra_msr_offset32);
 313         if (extra_msr_offset64)
 314                 outp += sprintf(outp, "           MSR 0x%03X", extra_msr_offset64);
 315
 316         if (!debug)
 317                 goto done;
 318
 319         if (do_smi)
 320                 outp += sprintf(outp, "     SMI");
 321
 322         if (do_nhm_cstates)
 323                 outp += sprintf(outp, "  CPU%%c1");
 324         if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
 325                 outp += sprintf(outp, "  CPU%%c3");
 326         if (do_nhm_cstates)
 327                 outp += sprintf(outp, "  CPU%%c6");
 328         if (do_snb_cstates)
 329                 outp += sprintf(outp, "  CPU%%c7");
 330
 331         if (do_dts)
 332                 outp += sprintf(outp, " CoreTmp");
 333         if (do_ptm)
 334                 outp += sprintf(outp, "  PkgTmp");
 335
 336         if (do_skl_residency) {
 337                 outp += sprintf(outp, " Totl%%C0");
 338                 outp += sprintf(outp, "  Any%%C0");
 339                 outp += sprintf(outp, "  GFX%%C0");
 340                 outp += sprintf(outp, " CPUGFX%%");
 341         }
 342
 343         if (do_pc2)
 344                 outp += sprintf(outp, " Pkg%%pc2");
 345         if (do_pc3)
 346                 outp += sprintf(outp, " Pkg%%pc3");
 347         if (do_pc6)
 348                 outp += sprintf(outp, " Pkg%%pc6");
 349         if (do_pc7)
 350                 outp += sprintf(outp, " Pkg%%pc7");
 351         if (do_c8_c9_c10) {
 352                 outp += sprintf(outp, " Pkg%%pc8");
 353                 outp += sprintf(outp, " Pkg%%pc9");
 354                 outp += sprintf(outp, " Pk%%pc10");
 355         }
 356
 357         if (do_rapl && !rapl_joules) {
 358                 if (do_rapl & RAPL_PKG)
 359                         outp += sprintf(outp, " PkgWatt");
 360                 if (do_rapl & RAPL_CORES)
 361                         outp += sprintf(outp, " CorWatt");
 362                 if (do_rapl & RAPL_GFX)
 363                         outp += sprintf(outp, " GFXWatt");
 364                 if (do_rapl & RAPL_DRAM)
 365                         outp += sprintf(outp, " RAMWatt");
 366                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 367                         outp += sprintf(outp, "   PKG_%%");
 368                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 369                         outp += sprintf(outp, "   RAM_%%");
 370         } else if (do_rapl && rapl_joules) {
 371                 if (do_rapl & RAPL_PKG)
 372                         outp += sprintf(outp, "   Pkg_J");
 373                 if (do_rapl & RAPL_CORES)
 374                         outp += sprintf(outp, "   Cor_J");
 375                 if (do_rapl & RAPL_GFX)
 376                         outp += sprintf(outp, "   GFX_J");
 377                 if (do_rapl & RAPL_DRAM)
 378                         outp += sprintf(outp, "   RAM_J");
 379                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 380                         outp += sprintf(outp, "   PKG_%%");
 381                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 382                         outp += sprintf(outp, "   RAM_%%");
 383                 outp += sprintf(outp, "   time");
 384
 385         }
 386     done:
 387         outp += sprintf(outp, "\n");
 388 }
 389
 390 int dump_counters(struct thread_data *t, struct core_data *c,
 391         struct pkg_data *p)
 392 {
 393         outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
 394
 395         if (t) {
 396                 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
 397                         t->cpu_id, t->flags);
 398                 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
 399                 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
 400                 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
 401                 outp += sprintf(outp, "c1: %016llX\n", t->c1);
 402                 outp += sprintf(outp, "msr0x%x: %08llX\n",
 403                         extra_delta_offset32, t->extra_delta32);
 404                 outp += sprintf(outp, "msr0x%x: %016llX\n",
 405                         extra_delta_offset64, t->extra_delta64);
 406                 outp += sprintf(outp, "msr0x%x: %08llX\n",
 407                         extra_msr_offset32, t->extra_msr32);
 408                 outp += sprintf(outp, "msr0x%x: %016llX\n",
 409                         extra_msr_offset64, t->extra_msr64);
 410                 if (do_smi)
 411                         outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
 412         }
 413
 414         if (c) {
 415                 outp += sprintf(outp, "core: %d\n", c->core_id);
 416                 outp += sprintf(outp, "c3: %016llX\n", c->c3);
 417                 outp += sprintf(outp, "c6: %016llX\n", c->c6);
 418                 outp += sprintf(outp, "c7: %016llX\n", c->c7);
 419                 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
 420         }
 421
 422         if (p) {
 423                 outp += sprintf(outp, "package: %d\n", p->package_id);
 424
 425                 outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
 426                 outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
 427                 outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
 428                 outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
 429
 430                 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
 431                 if (do_pc3)
 432                         outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
 433                 if (do_pc6)
 434                         outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
 435                 if (do_pc7)
 436                         outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
 437                 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
 438                 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
 439                 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
 440                 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
 441                 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
 442                 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
 443                 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
 444                 outp += sprintf(outp, "Throttle PKG: %0X\n",
 445                         p->rapl_pkg_perf_status);
 446                 outp += sprintf(outp, "Throttle RAM: %0X\n",
 447                         p->rapl_dram_perf_status);
 448                 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
 449         }
 450
 451         outp += sprintf(outp, "\n");
 452
 453         return 0;
 454 }
 455
 456 /*
 457  * column formatting convention & formats
 458  */
 459 int format_counters(struct thread_data *t, struct core_data *c,
 460         struct pkg_data *p)
 461 {
 462         double interval_float;
 463         char *fmt8;
 464
 465          /* if showing only 1st thread in core and this isn't one, bail out */
 466         if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 467                 return 0;
 468
 469          /* if showing only 1st thread in pkg and this isn't one, bail out */
 470         if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 471                 return 0;
 472
 473         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 474
 475         /* topo columns, print blanks on 1st (average) line */
 476         if (t == &average.threads) {
 477                 if (show_pkg)
 478                         outp += sprintf(outp, "       -");
 479                 if (show_core)
 480                         outp += sprintf(outp, "       -");
 481                 if (show_cpu)
 482                         outp += sprintf(outp, "       -");
 483         } else {
 484                 if (show_pkg) {
 485                         if (p)
 486                                 outp += sprintf(outp, "%8d", p->package_id);
 487                         else
 488                                 outp += sprintf(outp, "       -");
 489                 }
 490                 if (show_core) {
 491                         if (c)
 492                                 outp += sprintf(outp, "%8d", c->core_id);
 493                         else
 494                                 outp += sprintf(outp, "       -");
 495                 }
 496                 if (show_cpu)
 497                         outp += sprintf(outp, "%8d", t->cpu_id);
 498         }
 499
 500         /* Avg_MHz */
 501         if (has_aperf)
 502                 outp += sprintf(outp, "%8.0f",
 503                         1.0 / units * t->aperf / interval_float);
 504
 505         /* %Busy */
 506         if (has_aperf) {
 507                 if (!skip_c0)
 508                         outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc/tsc_tweak);
 509                 else
 510                         outp += sprintf(outp, "********");
 511         }
 512
 513         /* Bzy_MHz */
 514         if (has_aperf)
 515                 outp += sprintf(outp, "%8.0f",
 516                         1.0 * t->tsc * tsc_tweak / units * t->aperf / t->mperf / interval_float);
 517
 518         /* TSC_MHz */
 519         outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
 520
 521         /* delta */
 522         if (extra_delta_offset32)
 523                 outp += sprintf(outp, "  %11llu", t->extra_delta32);
 524
 525         /* DELTA */
 526         if (extra_delta_offset64)
 527                 outp += sprintf(outp, "  %11llu", t->extra_delta64);
 528         /* msr */
 529         if (extra_msr_offset32)
 530                 outp += sprintf(outp, "  0x%08llx", t->extra_msr32);
 531
 532         /* MSR */
 533         if (extra_msr_offset64)
 534                 outp += sprintf(outp, "  0x%016llx", t->extra_msr64);
 535
 536         if (!debug)
 537                 goto done;
 538
 539         /* SMI */
 540         if (do_smi)
 541                 outp += sprintf(outp, "%8d", t->smi_count);
 542
 543         if (do_nhm_cstates) {
 544                 if (!skip_c1)
 545                         outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
 546                 else
 547                         outp += sprintf(outp, "********");
 548         }
 549
 550         /* print per-core data only for 1st thread in core */
 551         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 552                 goto done;
 553
 554         if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
 555                 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
 556         if (do_nhm_cstates)
 557                 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
 558         if (do_snb_cstates)
 559                 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
 560
 561         if (do_dts)
 562                 outp += sprintf(outp, "%8d", c->core_temp_c);
 563
 564         /* print per-package data only for 1st core in package */
 565         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 566                 goto done;
 567
 568         /* PkgTmp */
 569         if (do_ptm)
 570                 outp += sprintf(outp, "%8d", p->pkg_temp_c);
 571
 572         /* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
 573         if (do_skl_residency) {
 574                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_wtd_core_c0/t->tsc);
 575                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_any_core_c0/t->tsc);
 576                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_any_gfxe_c0/t->tsc);
 577                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_both_core_gfxe_c0/t->tsc);
 578         }
 579
 580         if (do_pc2)
 581                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
 582         if (do_pc3)
 583                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
 584         if (do_pc6)
 585                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
 586         if (do_pc7)
 587                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
 588         if (do_c8_c9_c10) {
 589                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
 590                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
 591                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
 592         }
 593
 594         /*
 595          * If measurement interval exceeds minimum RAPL Joule Counter range,
 596          * indicate that results are suspect by printing "**" in fraction place.
 597          */
 598         if (interval_float < rapl_joule_counter_range)
 599                 fmt8 = "%8.2f";
 600         else
 601                 fmt8 = " %6.0f**";
 602
 603         if (do_rapl && !rapl_joules) {
 604                 if (do_rapl & RAPL_PKG)
 605                         outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
 606                 if (do_rapl & RAPL_CORES)
 607                         outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
 608                 if (do_rapl & RAPL_GFX)
 609                         outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
 610                 if (do_rapl & RAPL_DRAM)
 611                         outp += sprintf(outp, fmt8, p->energy_dram * rapl_dram_energy_units / interval_float);
 612                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 613                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
 614                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 615                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
 616         } else if (do_rapl && rapl_joules) {
 617                 if (do_rapl & RAPL_PKG)
 618                         outp += sprintf(outp, fmt8,
 619                                         p->energy_pkg * rapl_energy_units);
 620                 if (do_rapl & RAPL_CORES)
 621                         outp += sprintf(outp, fmt8,
 622                                         p->energy_cores * rapl_energy_units);
 623                 if (do_rapl & RAPL_GFX)
 624                         outp += sprintf(outp, fmt8,
 625                                         p->energy_gfx * rapl_energy_units);
 626                 if (do_rapl & RAPL_DRAM)
 627                         outp += sprintf(outp, fmt8,
 628                                         p->energy_dram * rapl_dram_energy_units);
 629                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 630                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
 631                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 632                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
 633
 634                 outp += sprintf(outp, fmt8, interval_float);
 635         }
 636 done:
 637         outp += sprintf(outp, "\n");
 638
 639         return 0;
 640 }
 641
 642 void flush_stdout()
 643 {
 644         fputs(output_buffer, stdout);
 645         fflush(stdout);
 646         outp = output_buffer;
 647 }
 648 void flush_stderr()
 649 {
 650         fputs(output_buffer, stderr);
 651         outp = output_buffer;
 652 }
 653 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 654 {
 655         static int printed;
 656
 657         if (!printed || !summary_only)
 658                 print_header();
 659
 660         if (topo.num_cpus > 1)
 661                 format_counters(&average.threads, &average.cores,
 662                         &average.packages);
 663
 664         printed = 1;
 665
 666         if (summary_only)
 667                 return;
 668
 669         for_all_cpus(format_counters, t, c, p);
 670 }
 671
 672 #define DELTA_WRAP32(new, old)                  \
 673         if (new > old) {                        \
 674                 old = new - old;                \
 675         } else {                                \
 676                 old = 0x100000000 + new - old;  \
 677         }
 678
 679 void
 680 delta_package(struct pkg_data *new, struct pkg_data *old)
 681 {
 682
 683         if (do_skl_residency) {
 684                 old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
 685                 old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
 686                 old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
 687                 old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
 688         }
 689         old->pc2 = new->pc2 - old->pc2;
 690         if (do_pc3)
 691                 old->pc3 = new->pc3 - old->pc3;
 692         if (do_pc6)
 693                 old->pc6 = new->pc6 - old->pc6;
 694         if (do_pc7)
 695                 old->pc7 = new->pc7 - old->pc7;
 696         old->pc8 = new->pc8 - old->pc8;
 697         old->pc9 = new->pc9 - old->pc9;
 698         old->pc10 = new->pc10 - old->pc10;
 699         old->pkg_temp_c = new->pkg_temp_c;
 700
 701         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 702         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 703         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 704         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 705         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 706         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 707 }
 708
 709 void
 710 delta_core(struct core_data *new, struct core_data *old)
 711 {
 712         old->c3 = new->c3 - old->c3;
 713         old->c6 = new->c6 - old->c6;
 714         old->c7 = new->c7 - old->c7;
 715         old->core_temp_c = new->core_temp_c;
 716 }
 717
 718 /*
 719  * old = new - old
 720  */
 721 void
 722 delta_thread(struct thread_data *new, struct thread_data *old,
 723         struct core_data *core_delta)
 724 {
 725         old->tsc = new->tsc - old->tsc;
 726
 727         /* check for TSC < 1 Mcycles over interval */
 728         if (old->tsc < (1000 * 1000))
 729                 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
 730                      "You can disable all c-states by booting with \"idle=poll\"\n"
 731                      "or just the deep ones with \"processor.max_cstate=1\"");
 732
 733         old->c1 = new->c1 - old->c1;
 734
 735         if (has_aperf) {
 736                 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 737                         old->aperf = new->aperf - old->aperf;
 738                         old->mperf = new->mperf - old->mperf;
 739                 } else {
 740
 741                         if (!aperf_mperf_unstable) {
 742                                 fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
 743                                 fprintf(stderr, "* Frequency results do not cover entire interval *\n");
 744                                 fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
 745
 746                                 aperf_mperf_unstable = 1;
 747                         }
 748                         /*
 749                          * mperf delta is likely a huge "positive" number
 750                          * can not use it for calculating c0 time
 751                          */
 752                         skip_c0 = 1;
 753                         skip_c1 = 1;
 754                 }
 755         }
 756
 757
 758         if (use_c1_residency_msr) {
 759                 /*
 760                  * Some models have a dedicated C1 residency MSR,
 761                  * which should be more accurate than the derivation below.
 762                  */
 763         } else {
 764                 /*
 765                  * As counter collection is not atomic,
 766                  * it is possible for mperf's non-halted cycles + idle states
 767                  * to exceed TSC's all cycles: show c1 = 0% in that case.
 768                  */
 769                 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 770                         old->c1 = 0;
 771                 else {
 772                         /* normal case, derive c1 */
 773                         old->c1 = old->tsc - old->mperf - core_delta->c3
 774                                 - core_delta->c6 - core_delta->c7;
 775                 }
 776         }
 777
 778         if (old->mperf == 0) {
 779                 if (debug > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
 780                 old->mperf = 1; /* divide by 0 protection */
 781         }
 782
 783         old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
 784         old->extra_delta32 &= 0xFFFFFFFF;
 785
 786         old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
 787
 788         /*
 789          * Extra MSR is just a snapshot, simply copy latest w/o subtracting
 790          */
 791         old->extra_msr32 = new->extra_msr32;
 792         old->extra_msr64 = new->extra_msr64;
 793
 794         if (do_smi)
 795                 old->smi_count = new->smi_count - old->smi_count;
 796 }
 797
 798 int delta_cpu(struct thread_data *t, struct core_data *c,
 799         struct pkg_data *p, struct thread_data *t2,
 800         struct core_data *c2, struct pkg_data *p2)
 801 {
 802         /* calculate core delta only for 1st thread in core */
 803         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 804                 delta_core(c, c2);
 805
 806         /* always calculate thread delta */
 807         delta_thread(t, t2, c2);        /* c2 is core delta */
 808
 809         /* calculate package delta only for 1st core in package */
 810         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 811                 delta_package(p, p2);
 812
 813         return 0;
 814 }
 815
 816 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 817 {
 818         t->tsc = 0;
 819         t->aperf = 0;
 820         t->mperf = 0;
 821         t->c1 = 0;
 822
 823         t->smi_count = 0;
 824         t->extra_delta32 = 0;
 825         t->extra_delta64 = 0;
 826
 827         /* tells format_counters to dump all fields from this set */
 828         t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
 829
 830         c->c3 = 0;
 831         c->c6 = 0;
 832         c->c7 = 0;
 833         c->core_temp_c = 0;
 834
 835         p->pkg_wtd_core_c0 = 0;
 836         p->pkg_any_core_c0 = 0;
 837         p->pkg_any_gfxe_c0 = 0;
 838         p->pkg_both_core_gfxe_c0 = 0;
 839
 840         p->pc2 = 0;
 841         if (do_pc3)
 842                 p->pc3 = 0;
 843         if (do_pc6)
 844                 p->pc6 = 0;
 845         if (do_pc7)
 846                 p->pc7 = 0;
 847         p->pc8 = 0;
 848         p->pc9 = 0;
 849         p->pc10 = 0;
 850
 851         p->energy_pkg = 0;
 852         p->energy_dram = 0;
 853         p->energy_cores = 0;
 854         p->energy_gfx = 0;
 855         p->rapl_pkg_perf_status = 0;
 856         p->rapl_dram_perf_status = 0;
 857         p->pkg_temp_c = 0;
 858 }
 859 int sum_counters(struct thread_data *t, struct core_data *c,
 860         struct pkg_data *p)
 861 {
 862         average.threads.tsc += t->tsc;
 863         average.threads.aperf += t->aperf;
 864         average.threads.mperf += t->mperf;
 865         average.threads.c1 += t->c1;
 866
 867         average.threads.extra_delta32 += t->extra_delta32;
 868         average.threads.extra_delta64 += t->extra_delta64;
 869
 870         /* sum per-core values only for 1st thread in core */
 871         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 872                 return 0;
 873
 874         average.cores.c3 += c->c3;
 875         average.cores.c6 += c->c6;
 876         average.cores.c7 += c->c7;
 877
 878         average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
 879
 880         /* sum per-pkg values only for 1st core in pkg */
 881         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 882                 return 0;
 883
 884         if (do_skl_residency) {
 885                 average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
 886                 average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
 887                 average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
 888                 average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
 889         }
 890
 891         average.packages.pc2 += p->pc2;
 892         if (do_pc3)
 893                 average.packages.pc3 += p->pc3;
 894         if (do_pc6)
 895                 average.packages.pc6 += p->pc6;
 896         if (do_pc7)
 897                 average.packages.pc7 += p->pc7;
 898         average.packages.pc8 += p->pc8;
 899         average.packages.pc9 += p->pc9;
 900         average.packages.pc10 += p->pc10;
 901
 902         average.packages.energy_pkg += p->energy_pkg;
 903         average.packages.energy_dram += p->energy_dram;
 904         average.packages.energy_cores += p->energy_cores;
 905         average.packages.energy_gfx += p->energy_gfx;
 906
 907         average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
 908
 909         average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
 910         average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
 911         return 0;
 912 }
 913 /*
 914  * sum the counters for all cpus in the system
 915  * compute the weighted average
 916  */
 917 void compute_average(struct thread_data *t, struct core_data *c,
 918         struct pkg_data *p)
 919 {
 920         clear_counters(&average.threads, &average.cores, &average.packages);
 921
 922         for_all_cpus(sum_counters, t, c, p);
 923
 924         average.threads.tsc /= topo.num_cpus;
 925         average.threads.aperf /= topo.num_cpus;
 926         average.threads.mperf /= topo.num_cpus;
 927         average.threads.c1 /= topo.num_cpus;
 928
 929         average.threads.extra_delta32 /= topo.num_cpus;
 930         average.threads.extra_delta32 &= 0xFFFFFFFF;
 931
 932         average.threads.extra_delta64 /= topo.num_cpus;
 933
 934         average.cores.c3 /= topo.num_cores;
 935         average.cores.c6 /= topo.num_cores;
 936         average.cores.c7 /= topo.num_cores;
 937
 938         if (do_skl_residency) {
 939                 average.packages.pkg_wtd_core_c0 /= topo.num_packages;
 940                 average.packages.pkg_any_core_c0 /= topo.num_packages;
 941                 average.packages.pkg_any_gfxe_c0 /= topo.num_packages;
 942                 average.packages.pkg_both_core_gfxe_c0 /= topo.num_packages;
 943         }
 944
 945         average.packages.pc2 /= topo.num_packages;
 946         if (do_pc3)
 947                 average.packages.pc3 /= topo.num_packages;
 948         if (do_pc6)
 949                 average.packages.pc6 /= topo.num_packages;
 950         if (do_pc7)
 951                 average.packages.pc7 /= topo.num_packages;
 952
 953         average.packages.pc8 /= topo.num_packages;
 954         average.packages.pc9 /= topo.num_packages;
 955         average.packages.pc10 /= topo.num_packages;
 956 }
 957
 958 static unsigned long long rdtsc(void)
 959 {
 960         unsigned int low, high;
 961
 962         asm volatile("rdtsc" : "=a" (low), "=d" (high));
 963
 964         return low | ((unsigned long long)high) << 32;
 965 }
 966
 967
 968 /*
 969  * get_counters(...)
 970  * migrate to cpu
 971  * acquire and record local counters for that cpu
 972  */
 973 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 974 {
 975         int cpu = t->cpu_id;
 976         unsigned long long msr;
 977
 978         if (cpu_migrate(cpu)) {
 979                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
 980                 return -1;
 981         }
 982
 983         t->tsc = rdtsc();       /* we are running on local CPU of interest */
 984
 985         if (has_aperf) {
 986                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
 987                         return -3;
 988                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
 989                         return -4;
 990                 t->aperf = t->aperf * aperf_mperf_multiplier;
 991                 t->mperf = t->mperf * aperf_mperf_multiplier;
 992         }
 993
 994         if (do_smi) {
 995                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
 996                         return -5;
 997                 t->smi_count = msr & 0xFFFFFFFF;
 998         }
 999         if (extra_delta_offset32) {
1000                 if (get_msr(cpu, extra_delta_offset32, &msr))
1001                         return -5;
1002                 t->extra_delta32 = msr & 0xFFFFFFFF;
1003         }
1004
1005         if (extra_delta_offset64)
1006                 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
1007                         return -5;
1008
1009         if (extra_msr_offset32) {
1010                 if (get_msr(cpu, extra_msr_offset32, &msr))
1011                         return -5;
1012                 t->extra_msr32 = msr & 0xFFFFFFFF;
1013         }
1014
1015         if (extra_msr_offset64)
1016                 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
1017                         return -5;
1018
1019         if (use_c1_residency_msr) {
1020                 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
1021                         return -6;
1022         }
1023
1024         /* collect core counters only for 1st thread in core */
1025         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1026                 return 0;
1027
1028         if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates) {
1029                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
1030                         return -6;
1031         }
1032
1033         if (do_nhm_cstates && !do_knl_cstates) {
1034                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
1035                         return -7;
1036         } else if (do_knl_cstates) {
1037                 if (get_msr(cpu, MSR_KNL_CORE_C6_RESIDENCY, &c->c6))
1038                         return -7;
1039         }
1040
1041         if (do_snb_cstates)
1042                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
1043                         return -8;
1044
1045         if (do_dts) {
1046                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1047                         return -9;
1048                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1049         }
1050
1051
1052         /* collect package counters only for 1st core in package */
1053         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1054                 return 0;
1055
1056         if (do_skl_residency) {
1057                 if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
1058                         return -10;
1059                 if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
1060                         return -11;
1061                 if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
1062                         return -12;
1063                 if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
1064                         return -13;
1065         }
1066         if (do_pc3)
1067                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
1068                         return -9;
1069         if (do_pc6)
1070                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
1071                         return -10;
1072         if (do_pc2)
1073                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
1074                         return -11;
1075         if (do_pc7)
1076                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1077                         return -12;
1078         if (do_c8_c9_c10) {
1079                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1080                         return -13;
1081                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1082                         return -13;
1083                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1084                         return -13;
1085         }
1086         if (do_rapl & RAPL_PKG) {
1087                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1088                         return -13;
1089                 p->energy_pkg = msr & 0xFFFFFFFF;
1090         }
1091         if (do_rapl & RAPL_CORES) {
1092                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1093                         return -14;
1094                 p->energy_cores = msr & 0xFFFFFFFF;
1095         }
1096         if (do_rapl & RAPL_DRAM) {
1097                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1098                         return -15;
1099                 p->energy_dram = msr & 0xFFFFFFFF;
1100         }
1101         if (do_rapl & RAPL_GFX) {
1102                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1103                         return -16;
1104                 p->energy_gfx = msr & 0xFFFFFFFF;
1105         }
1106         if (do_rapl & RAPL_PKG_PERF_STATUS) {
1107                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1108                         return -16;
1109                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1110         }
1111         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1112                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1113                         return -16;
1114                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1115         }
1116         if (do_ptm) {
1117                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1118                         return -17;
1119                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1120         }
1121         return 0;
1122 }
1123
1124 /*
1125  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1126  * If you change the values, note they are used both in comparisons
1127  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1128  */
1129
1130 #define PCLUKN 0 /* Unknown */
1131 #define PCLRSV 1 /* Reserved */
1132 #define PCL__0 2 /* PC0 */
1133 #define PCL__1 3 /* PC1 */
1134 #define PCL__2 4 /* PC2 */
1135 #define PCL__3 5 /* PC3 */
1136 #define PCL__4 6 /* PC4 */
1137 #define PCL__6 7 /* PC6 */
1138 #define PCL_6N 8 /* PC6 No Retention */
1139 #define PCL_6R 9 /* PC6 Retention */
1140 #define PCL__7 10 /* PC7 */
1141 #define PCL_7S 11 /* PC7 Shrink */
1142 #define PCL__8 12 /* PC8 */
1143 #define PCL__9 13 /* PC9 */
1144 #define PCLUNL 14 /* Unlimited */
1145
1146 int pkg_cstate_limit = PCLUKN;
1147 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1148         "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "unlimited"};
1149
1150 int nhm_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1151 int snb_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1152 int hsw_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1153 int slv_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1154 int amt_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1155 int phi_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1156
1157
1158 static void
1159 calculate_tsc_tweak()
1160 {
1161         unsigned long long msr;
1162         unsigned int base_ratio;
1163
1164         get_msr(base_cpu, MSR_NHM_PLATFORM_INFO, &msr);
1165         base_ratio = (msr >> 8) & 0xFF;
1166         base_hz = base_ratio * bclk * 1000000;
1167         tsc_tweak = base_hz / tsc_hz;
1168 }
1169
1170 static void
1171 dump_nhm_platform_info(void)
1172 {
1173         unsigned long long msr;
1174         unsigned int ratio;
1175
1176         get_msr(base_cpu, MSR_NHM_PLATFORM_INFO, &msr);
1177
1178         fprintf(stderr, "cpu%d: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
1179
1180         ratio = (msr >> 40) & 0xFF;
1181         fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency frequency\n",
1182                 ratio, bclk, ratio * bclk);
1183
1184         ratio = (msr >> 8) & 0xFF;
1185         fprintf(stderr, "%d * %.0f = %.0f MHz base frequency\n",
1186                 ratio, bclk, ratio * bclk);
1187
1188         get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
1189         fprintf(stderr, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1190                 base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
1191
1192         return;
1193 }
1194
1195 static void
1196 dump_hsw_turbo_ratio_limits(void)
1197 {
1198         unsigned long long msr;
1199         unsigned int ratio;
1200
1201         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
1202
1203         fprintf(stderr, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
1204
1205         ratio = (msr >> 8) & 0xFF;
1206         if (ratio)
1207                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 18 active cores\n",
1208                         ratio, bclk, ratio * bclk);
1209
1210         ratio = (msr >> 0) & 0xFF;
1211         if (ratio)
1212                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 17 active cores\n",
1213                         ratio, bclk, ratio * bclk);
1214         return;
1215 }
1216
1217 static void
1218 dump_ivt_turbo_ratio_limits(void)
1219 {
1220         unsigned long long msr;
1221         unsigned int ratio;
1222
1223         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
1224
1225         fprintf(stderr, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
1226
1227         ratio = (msr >> 56) & 0xFF;
1228         if (ratio)
1229                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1230                         ratio, bclk, ratio * bclk);
1231
1232         ratio = (msr >> 48) & 0xFF;
1233         if (ratio)
1234                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1235                         ratio, bclk, ratio * bclk);
1236
1237         ratio = (msr >> 40) & 0xFF;
1238         if (ratio)
1239                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1240                         ratio, bclk, ratio * bclk);
1241
1242         ratio = (msr >> 32) & 0xFF;
1243         if (ratio)
1244                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1245                         ratio, bclk, ratio * bclk);
1246
1247         ratio = (msr >> 24) & 0xFF;
1248         if (ratio)
1249                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1250                         ratio, bclk, ratio * bclk);
1251
1252         ratio = (msr >> 16) & 0xFF;
1253         if (ratio)
1254                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1255                         ratio, bclk, ratio * bclk);
1256
1257         ratio = (msr >> 8) & 0xFF;
1258         if (ratio)
1259                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1260                         ratio, bclk, ratio * bclk);
1261
1262         ratio = (msr >> 0) & 0xFF;
1263         if (ratio)
1264                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1265                         ratio, bclk, ratio * bclk);
1266         return;
1267 }
1268
1269 static void
1270 dump_nhm_turbo_ratio_limits(void)
1271 {
1272         unsigned long long msr;
1273         unsigned int ratio;
1274
1275         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
1276
1277         fprintf(stderr, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
1278
1279         ratio = (msr >> 56) & 0xFF;
1280         if (ratio)
1281                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1282                         ratio, bclk, ratio * bclk);
1283
1284         ratio = (msr >> 48) & 0xFF;
1285         if (ratio)
1286                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1287                         ratio, bclk, ratio * bclk);
1288
1289         ratio = (msr >> 40) & 0xFF;
1290         if (ratio)
1291                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1292                         ratio, bclk, ratio * bclk);
1293
1294         ratio = (msr >> 32) & 0xFF;
1295         if (ratio)
1296                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1297                         ratio, bclk, ratio * bclk);
1298
1299         ratio = (msr >> 24) & 0xFF;
1300         if (ratio)
1301                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1302                         ratio, bclk, ratio * bclk);
1303
1304         ratio = (msr >> 16) & 0xFF;
1305         if (ratio)
1306                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1307                         ratio, bclk, ratio * bclk);
1308
1309         ratio = (msr >> 8) & 0xFF;
1310         if (ratio)
1311                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1312                         ratio, bclk, ratio * bclk);
1313
1314         ratio = (msr >> 0) & 0xFF;
1315         if (ratio)
1316                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1317                         ratio, bclk, ratio * bclk);
1318         return;
1319 }
1320
1321 static void
1322 dump_knl_turbo_ratio_limits(void)
1323 {
1324         int cores;
1325         unsigned int ratio;
1326         unsigned long long msr;
1327         int delta_cores;
1328         int delta_ratio;
1329         int i;
1330
1331         get_msr(base_cpu, MSR_NHM_TURBO_RATIO_LIMIT, &msr);
1332
1333         fprintf(stderr, "cpu%d: MSR_NHM_TURBO_RATIO_LIMIT: 0x%08llx\n",
1334                 base_cpu, msr);
1335
1336         /**
1337          * Turbo encoding in KNL is as follows:
1338          * [7:0] -- Base value of number of active cores of bucket 1.
1339          * [15:8] -- Base value of freq ratio of bucket 1.
1340          * [20:16] -- +ve delta of number of active cores of bucket 2.
1341          * i.e. active cores of bucket 2 =
1342          * active cores of bucket 1 + delta
1343          * [23:21] -- Negative delta of freq ratio of bucket 2.
1344          * i.e. freq ratio of bucket 2 =
1345          * freq ratio of bucket 1 - delta
1346          * [28:24]-- +ve delta of number of active cores of bucket 3.
1347          * [31:29]-- -ve delta of freq ratio of bucket 3.
1348          * [36:32]-- +ve delta of number of active cores of bucket 4.
1349          * [39:37]-- -ve delta of freq ratio of bucket 4.
1350          * [44:40]-- +ve delta of number of active cores of bucket 5.
1351          * [47:45]-- -ve delta of freq ratio of bucket 5.
1352          * [52:48]-- +ve delta of number of active cores of bucket 6.
1353          * [55:53]-- -ve delta of freq ratio of bucket 6.
1354          * [60:56]-- +ve delta of number of active cores of bucket 7.
1355          * [63:61]-- -ve delta of freq ratio of bucket 7.
1356          */
1357         cores = msr & 0xFF;
1358         ratio = (msr >> 8) && 0xFF;
1359         if (ratio > 0)
1360                 fprintf(stderr,
1361                         "%d * %.0f = %.0f MHz max turbo %d active cores\n",
1362                         ratio, bclk, ratio * bclk, cores);
1363
1364         for (i = 16; i < 64; i = i + 8) {
1365                 delta_cores = (msr >> i) & 0x1F;
1366                 delta_ratio = (msr >> (i + 5)) && 0x7;
1367                 if (!delta_cores || !delta_ratio)
1368                         return;
1369                 cores = cores + delta_cores;
1370                 ratio = ratio - delta_ratio;
1371
1372                 /** -ve ratios will make successive ratio calculations
1373                  * negative. Hence return instead of carrying on.
1374                  */
1375                 if (ratio > 0)
1376                         fprintf(stderr,
1377                                 "%d * %.0f = %.0f MHz max turbo %d active cores\n",
1378                                 ratio, bclk, ratio * bclk, cores);
1379         }
1380 }
1381
1382 static void
1383 dump_nhm_cst_cfg(void)
1384 {
1385         unsigned long long msr;
1386
1387         get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1388
1389 #define SNB_C1_AUTO_UNDEMOTE              (1UL << 27)
1390 #define SNB_C3_AUTO_UNDEMOTE              (1UL << 28)
1391
1392         fprintf(stderr, "cpu%d: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", base_cpu, msr);
1393
1394         fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: %s)\n",
1395                 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1396                 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1397                 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1398                 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1399                 (msr & (1 << 15)) ? "" : "UN",
1400                 (unsigned int)msr & 7,
1401                 pkg_cstate_limit_strings[pkg_cstate_limit]);
1402         return;
1403 }
1404
1405 static void
1406 dump_config_tdp(void)
1407 {
1408         unsigned long long msr;
1409
1410         get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
1411         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
1412         fprintf(stderr, " (base_ratio=%d)\n", (unsigned int)msr & 0xEF);
1413
1414         get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
1415         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
1416         if (msr) {
1417                 fprintf(stderr, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0xEFFF);
1418                 fprintf(stderr, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0xEFFF);
1419                 fprintf(stderr, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xEF);
1420                 fprintf(stderr, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0xEFFF);
1421         }
1422         fprintf(stderr, ")\n");
1423
1424         get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
1425         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
1426         if (msr) {
1427                 fprintf(stderr, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0xEFFF);
1428                 fprintf(stderr, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0xEFFF);
1429                 fprintf(stderr, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xEF);
1430                 fprintf(stderr, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0xEFFF);
1431         }
1432         fprintf(stderr, ")\n");
1433
1434         get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
1435         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
1436         if ((msr) & 0x3)
1437                 fprintf(stderr, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
1438         fprintf(stderr, " lock=%d", (unsigned int)(msr >> 31) & 1);
1439         fprintf(stderr, ")\n");
1440
1441         get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
1442         fprintf(stderr, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
1443         fprintf(stderr, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xEF);
1444         fprintf(stderr, " lock=%d", (unsigned int)(msr >> 31) & 1);
1445         fprintf(stderr, ")\n");
1446 }
1447
1448 void free_all_buffers(void)
1449 {
1450         CPU_FREE(cpu_present_set);
1451         cpu_present_set = NULL;
1452         cpu_present_set = 0;
1453
1454         CPU_FREE(cpu_affinity_set);
1455         cpu_affinity_set = NULL;
1456         cpu_affinity_setsize = 0;
1457
1458         free(thread_even);
1459         free(core_even);
1460         free(package_even);
1461
1462         thread_even = NULL;
1463         core_even = NULL;
1464         package_even = NULL;
1465
1466         free(thread_odd);
1467         free(core_odd);
1468         free(package_odd);
1469
1470         thread_odd = NULL;
1471         core_odd = NULL;
1472         package_odd = NULL;
1473
1474         free(output_buffer);
1475         output_buffer = NULL;
1476         outp = NULL;
1477 }
1478
1479 /*
1480  * Open a file, and exit on failure
1481  */
1482 FILE *fopen_or_die(const char *path, const char *mode)
1483 {
1484         FILE *filep = fopen(path, "r");
1485         if (!filep)
1486                 err(1, "%s: open failed", path);
1487         return filep;
1488 }
1489
1490 /*
1491  * Parse a file containing a single int.
1492  */
1493 int parse_int_file(const char *fmt, ...)
1494 {
1495         va_list args;
1496         char path[PATH_MAX];
1497         FILE *filep;
1498         int value;
1499
1500         va_start(args, fmt);
1501         vsnprintf(path, sizeof(path), fmt, args);
1502         va_end(args);
1503         filep = fopen_or_die(path, "r");
1504         if (fscanf(filep, "%d", &value) != 1)
1505                 err(1, "%s: failed to parse number from file", path);
1506         fclose(filep);
1507         return value;
1508 }
1509
1510 /*
1511  * get_cpu_position_in_core(cpu)
1512  * return the position of the CPU among its HT siblings in the core
1513  * return -1 if the sibling is not in list
1514  */
1515 int get_cpu_position_in_core(int cpu)
1516 {
1517         char path[64];
1518         FILE *filep;
1519         int this_cpu;
1520         char character;
1521         int i;
1522
1523         sprintf(path,
1524                 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list",
1525                 cpu);
1526         filep = fopen(path, "r");
1527         if (filep == NULL) {
1528                 perror(path);
1529                 exit(1);
1530         }
1531
1532         for (i = 0; i < topo.num_threads_per_core; i++) {
1533                 fscanf(filep, "%d", &this_cpu);
1534                 if (this_cpu == cpu) {
1535                         fclose(filep);
1536                         return i;
1537                 }
1538
1539                 /* Account for no separator after last thread*/
1540                 if (i != (topo.num_threads_per_core - 1))
1541                         fscanf(filep, "%c", &character);
1542         }
1543
1544         fclose(filep);
1545         return -1;
1546 }
1547
1548 /*
1549  * cpu_is_first_core_in_package(cpu)
1550  * return 1 if given CPU is 1st core in package
1551  */
1552 int cpu_is_first_core_in_package(int cpu)
1553 {
1554         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1555 }
1556
1557 int get_physical_package_id(int cpu)
1558 {
1559         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1560 }
1561
1562 int get_core_id(int cpu)
1563 {
1564         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1565 }
1566
1567 int get_num_ht_siblings(int cpu)
1568 {
1569         char path[80];
1570         FILE *filep;
1571         int sib1;
1572         int matches = 0;
1573         char character;
1574         char str[100];
1575         char *ch;
1576
1577         sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1578         filep = fopen_or_die(path, "r");
1579
1580         /*
1581          * file format:
1582          * A ',' separated or '-' separated set of numbers
1583          * (eg 1-2 or 1,3,4,5)
1584          */
1585         fscanf(filep, "%d%c\n", &sib1, &character);
1586         fseek(filep, 0, SEEK_SET);
1587         fgets(str, 100, filep);
1588         ch = strchr(str, character);
1589         while (ch != NULL) {
1590                 matches++;
1591                 ch = strchr(ch+1, character);
1592         }
1593
1594         fclose(filep);
1595         return matches+1;
1596 }
1597
1598 /*
1599  * run func(thread, core, package) in topology order
1600  * skip non-present cpus
1601  */
1602
1603 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1604         struct pkg_data *, struct thread_data *, struct core_data *,
1605         struct pkg_data *), struct thread_data *thread_base,
1606         struct core_data *core_base, struct pkg_data *pkg_base,
1607         struct thread_data *thread_base2, struct core_data *core_base2,
1608         struct pkg_data *pkg_base2)
1609 {
1610         int retval, pkg_no, core_no, thread_no;
1611
1612         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1613                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1614                         for (thread_no = 0; thread_no <
1615                                 topo.num_threads_per_core; ++thread_no) {
1616                                 struct thread_data *t, *t2;
1617                                 struct core_data *c, *c2;
1618                                 struct pkg_data *p, *p2;
1619
1620                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1621
1622                                 if (cpu_is_not_present(t->cpu_id))
1623                                         continue;
1624
1625                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1626
1627                                 c = GET_CORE(core_base, core_no, pkg_no);
1628                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
1629
1630                                 p = GET_PKG(pkg_base, pkg_no);
1631                                 p2 = GET_PKG(pkg_base2, pkg_no);
1632
1633                                 retval = func(t, c, p, t2, c2, p2);
1634                                 if (retval)
1635                                         return retval;
1636                         }
1637                 }
1638         }
1639         return 0;
1640 }
1641
1642 /*
1643  * run func(cpu) on every cpu in /proc/stat
1644  * return max_cpu number
1645  */
1646 int for_all_proc_cpus(int (func)(int))
1647 {
1648         FILE *fp;
1649         int cpu_num;
1650         int retval;
1651
1652         fp = fopen_or_die(proc_stat, "r");
1653
1654         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1655         if (retval != 0)
1656                 err(1, "%s: failed to parse format", proc_stat);
1657
1658         while (1) {
1659                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1660                 if (retval != 1)
1661                         break;
1662
1663                 retval = func(cpu_num);
1664                 if (retval) {
1665                         fclose(fp);
1666                         return(retval);
1667                 }
1668         }
1669         fclose(fp);
1670         return 0;
1671 }
1672
1673 void re_initialize(void)
1674 {
1675         free_all_buffers();
1676         setup_all_buffers();
1677         printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1678 }
1679
1680
1681 /*
1682  * count_cpus()
1683  * remember the last one seen, it will be the max
1684  */
1685 int count_cpus(int cpu)
1686 {
1687         if (topo.max_cpu_num < cpu)
1688                 topo.max_cpu_num = cpu;
1689
1690         topo.num_cpus += 1;
1691         return 0;
1692 }
1693 int mark_cpu_present(int cpu)
1694 {
1695         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1696         return 0;
1697 }
1698
1699 void turbostat_loop()
1700 {
1701         int retval;
1702         int restarted = 0;
1703
1704 restart:
1705         restarted++;
1706
1707         retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1708         if (retval < -1) {
1709                 exit(retval);
1710         } else if (retval == -1) {
1711                 if (restarted > 1) {
1712                         exit(retval);
1713                 }
1714                 re_initialize();
1715                 goto restart;
1716         }
1717         restarted = 0;
1718         gettimeofday(&tv_even, (struct timezone *)NULL);
1719
1720         while (1) {
1721                 if (for_all_proc_cpus(cpu_is_not_present)) {
1722                         re_initialize();
1723                         goto restart;
1724                 }
1725                 sleep(interval_sec);
1726                 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1727                 if (retval < -1) {
1728                         exit(retval);
1729                 } else if (retval == -1) {
1730                         re_initialize();
1731                         goto restart;
1732                 }
1733                 gettimeofday(&tv_odd, (struct timezone *)NULL);
1734                 timersub(&tv_odd, &tv_even, &tv_delta);
1735                 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1736                 compute_average(EVEN_COUNTERS);
1737                 format_all_counters(EVEN_COUNTERS);
1738                 flush_stdout();
1739                 sleep(interval_sec);
1740                 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1741                 if (retval < -1) {
1742                         exit(retval);
1743                 } else if (retval == -1) {
1744                         re_initialize();
1745                         goto restart;
1746                 }
1747                 gettimeofday(&tv_even, (struct timezone *)NULL);
1748                 timersub(&tv_even, &tv_odd, &tv_delta);
1749                 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1750                 compute_average(ODD_COUNTERS);
1751                 format_all_counters(ODD_COUNTERS);
1752                 flush_stdout();
1753         }
1754 }
1755
1756 void check_dev_msr()
1757 {
1758         struct stat sb;
1759         char pathname[32];
1760
1761         sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
1762         if (stat(pathname, &sb))
1763                 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
1764                         err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
1765 }
1766
1767 void check_permissions()
1768 {
1769         struct __user_cap_header_struct cap_header_data;
1770         cap_user_header_t cap_header = &cap_header_data;
1771         struct __user_cap_data_struct cap_data_data;
1772         cap_user_data_t cap_data = &cap_data_data;
1773         extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
1774         int do_exit = 0;
1775         char pathname[32];
1776
1777         /* check for CAP_SYS_RAWIO */
1778         cap_header->pid = getpid();
1779         cap_header->version = _LINUX_CAPABILITY_VERSION;
1780         if (capget(cap_header, cap_data) < 0)
1781                 err(-6, "capget(2) failed");
1782
1783         if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1784                 do_exit++;
1785                 warnx("capget(CAP_SYS_RAWIO) failed,"
1786                         " try \"# setcap cap_sys_rawio=ep %s\"", progname);
1787         }
1788
1789         /* test file permissions */
1790         sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
1791         if (euidaccess(pathname, R_OK)) {
1792                 do_exit++;
1793                 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1794         }
1795
1796         /* if all else fails, thell them to be root */
1797         if (do_exit)
1798                 if (getuid() != 0)
1799                         warnx("... or simply run as root");
1800
1801         if (do_exit)
1802                 exit(-6);
1803 }
1804
1805 /*
1806  * NHM adds support for additional MSRs:
1807  *
1808  * MSR_SMI_COUNT                   0x00000034
1809  *
1810  * MSR_NHM_PLATFORM_INFO           0x000000ce
1811  * MSR_NHM_SNB_PKG_CST_CFG_CTL     0x000000e2
1812  *
1813  * MSR_PKG_C3_RESIDENCY            0x000003f8
1814  * MSR_PKG_C6_RESIDENCY            0x000003f9
1815  * MSR_CORE_C3_RESIDENCY           0x000003fc
1816  * MSR_CORE_C6_RESIDENCY           0x000003fd
1817  *
1818  * Side effect:
1819  * sets global pkg_cstate_limit to decode MSR_NHM_SNB_PKG_CST_CFG_CTL
1820  */
1821 int probe_nhm_msrs(unsigned int family, unsigned int model)
1822 {
1823         unsigned long long msr;
1824         int *pkg_cstate_limits;
1825
1826         if (!genuine_intel)
1827                 return 0;
1828
1829         if (family != 6)
1830                 return 0;
1831
1832         switch (model) {
1833         case 0x1A:      /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1834         case 0x1E:      /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1835         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
1836         case 0x25:      /* Westmere Client - Clarkdale, Arrandale */
1837         case 0x2C:      /* Westmere EP - Gulftown */
1838         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1839         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1840                 pkg_cstate_limits = nhm_pkg_cstate_limits;
1841                 break;
1842         case 0x2A:      /* SNB */
1843         case 0x2D:      /* SNB Xeon */
1844         case 0x3A:      /* IVB */
1845         case 0x3E:      /* IVB Xeon */
1846                 pkg_cstate_limits = snb_pkg_cstate_limits;
1847                 break;
1848         case 0x3C:      /* HSW */
1849         case 0x3F:      /* HSX */
1850         case 0x45:      /* HSW */
1851         case 0x46:      /* HSW */
1852         case 0x3D:      /* BDW */
1853         case 0x47:      /* BDW */
1854         case 0x4F:      /* BDX */
1855         case 0x56:      /* BDX-DE */
1856         case 0x4E:      /* SKL */
1857         case 0x5E:      /* SKL */
1858                 pkg_cstate_limits = hsw_pkg_cstate_limits;
1859                 break;
1860         case 0x37:      /* BYT */
1861         case 0x4D:      /* AVN */
1862                 pkg_cstate_limits = slv_pkg_cstate_limits;
1863                 break;
1864         case 0x4C:      /* AMT */
1865                 pkg_cstate_limits = amt_pkg_cstate_limits;
1866                 break;
1867         case 0x57:      /* PHI */
1868                 pkg_cstate_limits = phi_pkg_cstate_limits;
1869                 break;
1870         default:
1871                 return 0;
1872         }
1873         get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1874
1875         pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
1876
1877         return 1;
1878 }
1879 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
1880 {
1881         switch (model) {
1882         /* Nehalem compatible, but do not include turbo-ratio limit support */
1883         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1884         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1885                 return 0;
1886         default:
1887                 return 1;
1888         }
1889 }
1890 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1891 {
1892         if (!genuine_intel)
1893                 return 0;
1894
1895         if (family != 6)
1896                 return 0;
1897
1898         switch (model) {
1899         case 0x3E:      /* IVB Xeon */
1900         case 0x3F:      /* HSW Xeon */
1901                 return 1;
1902         default:
1903                 return 0;
1904         }
1905 }
1906 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
1907 {
1908         if (!genuine_intel)
1909                 return 0;
1910
1911         if (family != 6)
1912                 return 0;
1913
1914         switch (model) {
1915         case 0x3F:      /* HSW Xeon */
1916                 return 1;
1917         default:
1918                 return 0;
1919         }
1920 }
1921
1922 int has_knl_turbo_ratio_limit(unsigned int family, unsigned int model)
1923 {
1924         if (!genuine_intel)
1925                 return 0;
1926
1927         if (family != 6)
1928                 return 0;
1929
1930         switch (model) {
1931         case 0x57:      /* Knights Landing */
1932                 return 1;
1933         default:
1934                 return 0;
1935         }
1936 }
1937 int has_config_tdp(unsigned int family, unsigned int model)
1938 {
1939         if (!genuine_intel)
1940                 return 0;
1941
1942         if (family != 6)
1943                 return 0;
1944
1945         switch (model) {
1946         case 0x3A:      /* IVB */
1947         case 0x3C:      /* HSW */
1948         case 0x3F:      /* HSX */
1949         case 0x45:      /* HSW */
1950         case 0x46:      /* HSW */
1951         case 0x3D:      /* BDW */
1952         case 0x47:      /* BDW */
1953         case 0x4F:      /* BDX */
1954         case 0x56:      /* BDX-DE */
1955         case 0x4E:      /* SKL */
1956         case 0x5E:      /* SKL */
1957
1958         case 0x57:      /* Knights Landing */
1959                 return 1;
1960         default:
1961                 return 0;
1962         }
1963 }
1964
1965 static void
1966 dump_cstate_pstate_config_info(family, model)
1967 {
1968         if (!do_nhm_platform_info)
1969                 return;
1970
1971         dump_nhm_platform_info();
1972
1973         if (has_hsw_turbo_ratio_limit(family, model))
1974                 dump_hsw_turbo_ratio_limits();
1975
1976         if (has_ivt_turbo_ratio_limit(family, model))
1977                 dump_ivt_turbo_ratio_limits();
1978
1979         if (has_nhm_turbo_ratio_limit(family, model))
1980                 dump_nhm_turbo_ratio_limits();
1981
1982         if (has_knl_turbo_ratio_limit(family, model))
1983                 dump_knl_turbo_ratio_limits();
1984
1985         if (has_config_tdp(family, model))
1986                 dump_config_tdp();
1987
1988         dump_nhm_cst_cfg();
1989 }
1990
1991
1992 /*
1993  * print_epb()
1994  * Decode the ENERGY_PERF_BIAS MSR
1995  */
1996 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1997 {
1998         unsigned long long msr;
1999         char *epb_string;
2000         int cpu;
2001
2002         if (!has_epb)
2003                 return 0;
2004
2005         cpu = t->cpu_id;
2006
2007         /* EPB is per-package */
2008         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2009                 return 0;
2010
2011         if (cpu_migrate(cpu)) {
2012                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2013                 return -1;
2014         }
2015
2016         if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
2017                 return 0;
2018
2019         switch (msr & 0xF) {
2020         case ENERGY_PERF_BIAS_PERFORMANCE:
2021                 epb_string = "performance";
2022                 break;
2023         case ENERGY_PERF_BIAS_NORMAL:
2024                 epb_string = "balanced";
2025                 break;
2026         case ENERGY_PERF_BIAS_POWERSAVE:
2027                 epb_string = "powersave";
2028                 break;
2029         default:
2030                 epb_string = "custom";
2031                 break;
2032         }
2033         fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
2034
2035         return 0;
2036 }
2037
2038 /*
2039  * print_perf_limit()
2040  */
2041 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2042 {
2043         unsigned long long msr;
2044         int cpu;
2045
2046         cpu = t->cpu_id;
2047
2048         /* per-package */
2049         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2050                 return 0;
2051
2052         if (cpu_migrate(cpu)) {
2053                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2054                 return -1;
2055         }
2056
2057         if (do_core_perf_limit_reasons) {
2058                 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
2059                 fprintf(stderr, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2060                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
2061                         (msr & 1 << 15) ? "bit15, " : "",
2062                         (msr & 1 << 14) ? "bit14, " : "",
2063                         (msr & 1 << 13) ? "Transitions, " : "",
2064                         (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
2065                         (msr & 1 << 11) ? "PkgPwrL2, " : "",
2066                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
2067                         (msr & 1 << 9) ? "CorePwr, " : "",
2068                         (msr & 1 << 8) ? "Amps, " : "",
2069                         (msr & 1 << 6) ? "VR-Therm, " : "",
2070                         (msr & 1 << 5) ? "Auto-HWP, " : "",
2071                         (msr & 1 << 4) ? "Graphics, " : "",
2072                         (msr & 1 << 2) ? "bit2, " : "",
2073                         (msr & 1 << 1) ? "ThermStatus, " : "",
2074                         (msr & 1 << 0) ? "PROCHOT, " : "");
2075                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
2076                         (msr & 1 << 31) ? "bit31, " : "",
2077                         (msr & 1 << 30) ? "bit30, " : "",
2078                         (msr & 1 << 29) ? "Transitions, " : "",
2079                         (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
2080                         (msr & 1 << 27) ? "PkgPwrL2, " : "",
2081                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
2082                         (msr & 1 << 25) ? "CorePwr, " : "",
2083                         (msr & 1 << 24) ? "Amps, " : "",
2084                         (msr & 1 << 22) ? "VR-Therm, " : "",
2085                         (msr & 1 << 21) ? "Auto-HWP, " : "",
2086                         (msr & 1 << 20) ? "Graphics, " : "",
2087                         (msr & 1 << 18) ? "bit18, " : "",
2088                         (msr & 1 << 17) ? "ThermStatus, " : "",
2089                         (msr & 1 << 16) ? "PROCHOT, " : "");
2090
2091         }
2092         if (do_gfx_perf_limit_reasons) {
2093                 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
2094                 fprintf(stderr, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2095                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s)",
2096                         (msr & 1 << 0) ? "PROCHOT, " : "",
2097                         (msr & 1 << 1) ? "ThermStatus, " : "",
2098                         (msr & 1 << 4) ? "Graphics, " : "",
2099                         (msr & 1 << 6) ? "VR-Therm, " : "",
2100                         (msr & 1 << 8) ? "Amps, " : "",
2101                         (msr & 1 << 9) ? "GFXPwr, " : "",
2102                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
2103                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
2104                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s)\n",
2105                         (msr & 1 << 16) ? "PROCHOT, " : "",
2106                         (msr & 1 << 17) ? "ThermStatus, " : "",
2107                         (msr & 1 << 20) ? "Graphics, " : "",
2108                         (msr & 1 << 22) ? "VR-Therm, " : "",
2109                         (msr & 1 << 24) ? "Amps, " : "",
2110                         (msr & 1 << 25) ? "GFXPwr, " : "",
2111                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
2112                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
2113         }
2114         if (do_ring_perf_limit_reasons) {
2115                 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
2116                 fprintf(stderr, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2117                 fprintf(stderr, " (Active: %s%s%s%s%s%s)",
2118                         (msr & 1 << 0) ? "PROCHOT, " : "",
2119                         (msr & 1 << 1) ? "ThermStatus, " : "",
2120                         (msr & 1 << 6) ? "VR-Therm, " : "",
2121                         (msr & 1 << 8) ? "Amps, " : "",
2122                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
2123                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
2124                 fprintf(stderr, " (Logged: %s%s%s%s%s%s)\n",
2125                         (msr & 1 << 16) ? "PROCHOT, " : "",
2126                         (msr & 1 << 17) ? "ThermStatus, " : "",
2127                         (msr & 1 << 22) ? "VR-Therm, " : "",
2128                         (msr & 1 << 24) ? "Amps, " : "",
2129                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
2130                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
2131         }
2132         return 0;
2133 }
2134
2135 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
2136 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
2137
2138 double get_tdp(model)
2139 {
2140         unsigned long long msr;
2141
2142         if (do_rapl & RAPL_PKG_POWER_INFO)
2143                 if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
2144                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
2145
2146         switch (model) {
2147         case 0x37:
2148         case 0x4D:
2149                 return 30.0;
2150         default:
2151                 return 135.0;
2152         }
2153 }
2154
2155 /*
2156  * rapl_dram_energy_units_probe()
2157  * Energy units are either hard-coded, or come from RAPL Energy Unit MSR.
2158  */
2159 static double
2160 rapl_dram_energy_units_probe(int  model, double rapl_energy_units)
2161 {
2162         /* only called for genuine_intel, family 6 */
2163
2164         switch (model) {
2165         case 0x3F:      /* HSX */
2166         case 0x4F:      /* BDX */
2167         case 0x56:      /* BDX-DE */
2168         case 0x57:      /* KNL */
2169                 return (rapl_dram_energy_units = 15.3 / 1000000);
2170         default:
2171                 return (rapl_energy_units);
2172         }
2173 }
2174
2175
2176 /*
2177  * rapl_probe()
2178  *
2179  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
2180  */
2181 void rapl_probe(unsigned int family, unsigned int model)
2182 {
2183         unsigned long long msr;
2184         unsigned int time_unit;
2185         double tdp;
2186
2187         if (!genuine_intel)
2188                 return;
2189
2190         if (family != 6)
2191                 return;
2192
2193         switch (model) {
2194         case 0x2A:
2195         case 0x3A:
2196         case 0x3C:      /* HSW */
2197         case 0x45:      /* HSW */
2198         case 0x46:      /* HSW */
2199         case 0x3D:      /* BDW */
2200         case 0x47:      /* BDW */
2201                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
2202                 break;
2203         case 0x4E:      /* SKL */
2204         case 0x5E:      /* SKL */
2205                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2206                 break;
2207         case 0x3F:      /* HSX */
2208         case 0x4F:      /* BDX */
2209         case 0x56:      /* BDX-DE */
2210         case 0x57:      /* KNL */
2211                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2212                 break;
2213         case 0x2D:
2214         case 0x3E:
2215                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
2216                 break;
2217         case 0x37:      /* BYT */
2218         case 0x4D:      /* AVN */
2219                 do_rapl = RAPL_PKG | RAPL_CORES ;
2220                 break;
2221         default:
2222                 return;
2223         }
2224
2225         /* units on package 0, verify later other packages match */
2226         if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
2227                 return;
2228
2229         rapl_power_units = 1.0 / (1 << (msr & 0xF));
2230         if (model == 0x37)
2231                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
2232         else
2233                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
2234
2235         rapl_dram_energy_units = rapl_dram_energy_units_probe(model, rapl_energy_units);
2236
2237         time_unit = msr >> 16 & 0xF;
2238         if (time_unit == 0)
2239                 time_unit = 0xA;
2240
2241         rapl_time_units = 1.0 / (1 << (time_unit));
2242
2243         tdp = get_tdp(model);
2244
2245         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
2246         if (debug)
2247                 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
2248
2249         return;
2250 }
2251
2252 void perf_limit_reasons_probe(family, model)
2253 {
2254         if (!genuine_intel)
2255                 return;
2256
2257         if (family != 6)
2258                 return;
2259
2260         switch (model) {
2261         case 0x3C:      /* HSW */
2262         case 0x45:      /* HSW */
2263         case 0x46:      /* HSW */
2264                 do_gfx_perf_limit_reasons = 1;
2265         case 0x3F:      /* HSX */
2266                 do_core_perf_limit_reasons = 1;
2267                 do_ring_perf_limit_reasons = 1;
2268         default:
2269                 return;
2270         }
2271 }
2272
2273 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2274 {
2275         unsigned long long msr;
2276         unsigned int dts;
2277         int cpu;
2278
2279         if (!(do_dts || do_ptm))
2280                 return 0;
2281
2282         cpu = t->cpu_id;
2283
2284         /* DTS is per-core, no need to print for each thread */
2285         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
2286                 return 0;
2287
2288         if (cpu_migrate(cpu)) {
2289                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2290                 return -1;
2291         }
2292
2293         if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
2294                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
2295                         return 0;
2296
2297                 dts = (msr >> 16) & 0x7F;
2298                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
2299                         cpu, msr, tcc_activation_temp - dts);
2300
2301 #ifdef  THERM_DEBUG
2302                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
2303                         return 0;
2304
2305                 dts = (msr >> 16) & 0x7F;
2306                 dts2 = (msr >> 8) & 0x7F;
2307                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2308                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2309 #endif
2310         }
2311
2312
2313         if (do_dts) {
2314                 unsigned int resolution;
2315
2316                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
2317                         return 0;
2318
2319                 dts = (msr >> 16) & 0x7F;
2320                 resolution = (msr >> 27) & 0xF;
2321                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
2322                         cpu, msr, tcc_activation_temp - dts, resolution);
2323
2324 #ifdef THERM_DEBUG
2325                 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
2326                         return 0;
2327
2328                 dts = (msr >> 16) & 0x7F;
2329                 dts2 = (msr >> 8) & 0x7F;
2330                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2331                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2332 #endif
2333         }
2334
2335         return 0;
2336 }
2337
2338 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
2339 {
2340         fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
2341                 cpu, label,
2342                 ((msr >> 15) & 1) ? "EN" : "DIS",
2343                 ((msr >> 0) & 0x7FFF) * rapl_power_units,
2344                 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
2345                 (((msr >> 16) & 1) ? "EN" : "DIS"));
2346
2347         return;
2348 }
2349
2350 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2351 {
2352         unsigned long long msr;
2353         int cpu;
2354
2355         if (!do_rapl)
2356                 return 0;
2357
2358         /* RAPL counters are per package, so print only for 1st thread/package */
2359         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2360                 return 0;
2361
2362         cpu = t->cpu_id;
2363         if (cpu_migrate(cpu)) {
2364                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2365                 return -1;
2366         }
2367
2368         if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
2369                 return -1;
2370
2371         if (debug) {
2372                 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
2373                         "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
2374                         rapl_power_units, rapl_energy_units, rapl_time_units);
2375         }
2376         if (do_rapl & RAPL_PKG_POWER_INFO) {
2377
2378                 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
2379                         return -5;
2380
2381
2382                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2383                         cpu, msr,
2384                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2385                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2386                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2387                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2388
2389         }
2390         if (do_rapl & RAPL_PKG) {
2391
2392                 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
2393                         return -9;
2394
2395                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
2396                         cpu, msr, (msr >> 63) & 1 ? "": "UN");
2397
2398                 print_power_limit_msr(cpu, msr, "PKG Limit #1");
2399                 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
2400                         cpu,
2401                         ((msr >> 47) & 1) ? "EN" : "DIS",
2402                         ((msr >> 32) & 0x7FFF) * rapl_power_units,
2403                         (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
2404                         ((msr >> 48) & 1) ? "EN" : "DIS");
2405         }
2406
2407         if (do_rapl & RAPL_DRAM_POWER_INFO) {
2408                 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
2409                         return -6;
2410
2411                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2412                         cpu, msr,
2413                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2414                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2415                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2416                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2417         }
2418         if (do_rapl & RAPL_DRAM) {
2419                 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
2420                         return -9;
2421                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
2422                                 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2423
2424                 print_power_limit_msr(cpu, msr, "DRAM Limit");
2425         }
2426         if (do_rapl & RAPL_CORE_POLICY) {
2427                 if (debug) {
2428                         if (get_msr(cpu, MSR_PP0_POLICY, &msr))
2429                                 return -7;
2430
2431                         fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
2432                 }
2433         }
2434         if (do_rapl & RAPL_CORES) {
2435                 if (debug) {
2436
2437                         if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
2438                                 return -9;
2439                         fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2440                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2441                         print_power_limit_msr(cpu, msr, "Cores Limit");
2442                 }
2443         }
2444         if (do_rapl & RAPL_GFX) {
2445                 if (debug) {
2446                         if (get_msr(cpu, MSR_PP1_POLICY, &msr))
2447                                 return -8;
2448
2449                         fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
2450
2451                         if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
2452                                 return -9;
2453                         fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2454                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2455                         print_power_limit_msr(cpu, msr, "GFX Limit");
2456                 }
2457         }
2458         return 0;
2459 }
2460
2461 /*
2462  * SNB adds support for additional MSRs:
2463  *
2464  * MSR_PKG_C7_RESIDENCY            0x000003fa
2465  * MSR_CORE_C7_RESIDENCY           0x000003fe
2466  * MSR_PKG_C2_RESIDENCY            0x0000060d
2467  */
2468
2469 int has_snb_msrs(unsigned int family, unsigned int model)
2470 {
2471         if (!genuine_intel)
2472                 return 0;
2473
2474         switch (model) {
2475         case 0x2A:
2476         case 0x2D:
2477         case 0x3A:      /* IVB */
2478         case 0x3E:      /* IVB Xeon */
2479         case 0x3C:      /* HSW */
2480         case 0x3F:      /* HSW */
2481         case 0x45:      /* HSW */
2482         case 0x46:      /* HSW */
2483         case 0x3D:      /* BDW */
2484         case 0x47:      /* BDW */
2485         case 0x4F:      /* BDX */
2486         case 0x56:      /* BDX-DE */
2487         case 0x4E:      /* SKL */
2488         case 0x5E:      /* SKL */
2489                 return 1;
2490         }
2491         return 0;
2492 }
2493
2494 /*
2495  * HSW adds support for additional MSRs:
2496  *
2497  * MSR_PKG_C8_RESIDENCY            0x00000630
2498  * MSR_PKG_C9_RESIDENCY            0x00000631
2499  * MSR_PKG_C10_RESIDENCY           0x00000632
2500  */
2501 int has_hsw_msrs(unsigned int family, unsigned int model)
2502 {
2503         if (!genuine_intel)
2504                 return 0;
2505
2506         switch (model) {
2507         case 0x45:      /* HSW */
2508         case 0x3D:      /* BDW */
2509         case 0x4E:      /* SKL */
2510         case 0x5E:      /* SKL */
2511                 return 1;
2512         }
2513         return 0;
2514 }
2515
2516 /*
2517  * SKL adds support for additional MSRS:
2518  *
2519  * MSR_PKG_WEIGHTED_CORE_C0_RES    0x00000658
2520  * MSR_PKG_ANY_CORE_C0_RES         0x00000659
2521  * MSR_PKG_ANY_GFXE_C0_RES         0x0000065A
2522  * MSR_PKG_BOTH_CORE_GFXE_C0_RES   0x0000065B
2523  */
2524 int has_skl_msrs(unsigned int family, unsigned int model)
2525 {
2526         if (!genuine_intel)
2527                 return 0;
2528
2529         switch (model) {
2530         case 0x4E:      /* SKL */
2531         case 0x5E:      /* SKL */
2532                 return 1;
2533         }
2534         return 0;
2535 }
2536
2537
2538
2539 int is_slm(unsigned int family, unsigned int model)
2540 {
2541         if (!genuine_intel)
2542                 return 0;
2543         switch (model) {
2544         case 0x37:      /* BYT */
2545         case 0x4D:      /* AVN */
2546                 return 1;
2547         }
2548         return 0;
2549 }
2550
2551 int is_knl(unsigned int family, unsigned int model)
2552 {
2553         if (!genuine_intel)
2554                 return 0;
2555         switch (model) {
2556         case 0x57:      /* KNL */
2557                 return 1;
2558         }
2559         return 0;
2560 }
2561
2562 unsigned int get_aperf_mperf_multiplier(unsigned int family, unsigned int model)
2563 {
2564         if (is_knl(family, model))
2565                 return 1024;
2566         return 1;
2567 }
2568
2569 #define SLM_BCLK_FREQS 5
2570 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
2571
2572 double slm_bclk(void)
2573 {
2574         unsigned long long msr = 3;
2575         unsigned int i;
2576         double freq;
2577
2578         if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
2579                 fprintf(stderr, "SLM BCLK: unknown\n");
2580
2581         i = msr & 0xf;
2582         if (i >= SLM_BCLK_FREQS) {
2583                 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
2584                 msr = 3;
2585         }
2586         freq = slm_freq_table[i];
2587
2588         fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
2589
2590         return freq;
2591 }
2592
2593 double discover_bclk(unsigned int family, unsigned int model)
2594 {
2595         if (has_snb_msrs(family, model))
2596                 return 100.00;
2597         else if (is_slm(family, model))
2598                 return slm_bclk();
2599         else
2600                 return 133.33;
2601 }
2602
2603 /*
2604  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
2605  * the Thermal Control Circuit (TCC) activates.
2606  * This is usually equal to tjMax.
2607  *
2608  * Older processors do not have this MSR, so there we guess,
2609  * but also allow cmdline over-ride with -T.
2610  *
2611  * Several MSR temperature values are in units of degrees-C
2612  * below this value, including the Digital Thermal Sensor (DTS),
2613  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
2614  */
2615 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2616 {
2617         unsigned long long msr;
2618         unsigned int target_c_local;
2619         int cpu;
2620
2621         /* tcc_activation_temp is used only for dts or ptm */
2622         if (!(do_dts || do_ptm))
2623                 return 0;
2624
2625         /* this is a per-package concept */
2626         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2627                 return 0;
2628
2629         cpu = t->cpu_id;
2630         if (cpu_migrate(cpu)) {
2631                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2632                 return -1;
2633         }
2634
2635         if (tcc_activation_temp_override != 0) {
2636                 tcc_activation_temp = tcc_activation_temp_override;
2637                 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
2638                         cpu, tcc_activation_temp);
2639                 return 0;
2640         }
2641
2642         /* Temperature Target MSR is Nehalem and newer only */
2643         if (!do_nhm_platform_info)
2644                 goto guess;
2645
2646         if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
2647                 goto guess;
2648
2649         target_c_local = (msr >> 16) & 0xFF;
2650
2651         if (debug)
2652                 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2653                         cpu, msr, target_c_local);
2654
2655         if (!target_c_local)
2656                 goto guess;
2657
2658         tcc_activation_temp = target_c_local;
2659
2660         return 0;
2661
2662 guess:
2663         tcc_activation_temp = TJMAX_DEFAULT;
2664         fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2665                 cpu, tcc_activation_temp);
2666
2667         return 0;
2668 }
2669 void process_cpuid()
2670 {
2671         unsigned int eax, ebx, ecx, edx, max_level;
2672         unsigned int fms, family, model, stepping;
2673
2674         eax = ebx = ecx = edx = 0;
2675
2676         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2677
2678         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2679                 genuine_intel = 1;
2680
2681         if (debug)
2682                 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2683                         (char *)&ebx, (char *)&edx, (char *)&ecx);
2684
2685         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2686         family = (fms >> 8) & 0xf;
2687         model = (fms >> 4) & 0xf;
2688         stepping = fms & 0xf;
2689         if (family == 6 || family == 0xf)
2690                 model += ((fms >> 16) & 0xf) << 4;
2691
2692         if (debug)
2693                 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2694                         max_level, family, model, stepping, family, model, stepping);
2695
2696         if (!(edx & (1 << 5)))
2697                 errx(1, "CPUID: no MSR");
2698
2699         /*
2700          * check max extended function levels of CPUID.
2701          * This is needed to check for invariant TSC.
2702          * This check is valid for both Intel and AMD.
2703          */
2704         ebx = ecx = edx = 0;
2705         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2706
2707         if (max_level >= 0x80000007) {
2708
2709                 /*
2710                  * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2711                  * this check is valid for both Intel and AMD
2712                  */
2713                 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2714                 has_invariant_tsc = edx & (1 << 8);
2715         }
2716
2717         /*
2718          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2719          * this check is valid for both Intel and AMD
2720          */
2721
2722         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2723         has_aperf = ecx & (1 << 0);
2724         do_dts = eax & (1 << 0);
2725         do_ptm = eax & (1 << 6);
2726         has_epb = ecx & (1 << 3);
2727
2728         if (debug)
2729                 fprintf(stderr, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sEPB\n",
2730                         has_aperf ? "" : "No ",
2731                         do_dts ? "" : "No ",
2732                         do_ptm ? "" : "No ",
2733                         has_epb ? "" : "No ");
2734
2735         if (max_level > 0x15) {
2736                 unsigned int eax_crystal;
2737                 unsigned int ebx_tsc;
2738
2739                 /*
2740                  * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
2741                  */
2742                 eax_crystal = ebx_tsc = crystal_hz = edx = 0;
2743                 __get_cpuid(0x15, &eax_crystal, &ebx_tsc, &crystal_hz, &edx);
2744
2745                 if (ebx_tsc != 0) {
2746
2747                         if (debug && (ebx != 0))
2748                                 fprintf(stderr, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
2749                                         eax_crystal, ebx_tsc, crystal_hz);
2750
2751                         if (crystal_hz == 0)
2752                                 switch(model) {
2753                                 case 0x4E:      /* SKL */
2754                                 case 0x5E:      /* SKL */
2755                                         crystal_hz = 24000000;  /* 24 MHz */
2756                                         break;
2757                                 default:
2758                                         crystal_hz = 0;
2759                         }
2760
2761                         if (crystal_hz) {
2762                                 tsc_hz =  (unsigned long long) crystal_hz * ebx_tsc / eax_crystal;
2763                                 if (debug)
2764                                         fprintf(stderr, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
2765                                                 tsc_hz / 1000000, crystal_hz, ebx_tsc,  eax_crystal);
2766                         }
2767                 }
2768         }
2769
2770         if (has_aperf)
2771                 aperf_mperf_multiplier = get_aperf_mperf_multiplier(family, model);
2772
2773         do_nhm_platform_info = do_nhm_cstates = do_smi = probe_nhm_msrs(family, model);
2774         do_snb_cstates = has_snb_msrs(family, model);
2775         do_pc2 = do_snb_cstates && (pkg_cstate_limit >= PCL__2);
2776         do_pc3 = (pkg_cstate_limit >= PCL__3);
2777         do_pc6 = (pkg_cstate_limit >= PCL__6);
2778         do_pc7 = do_snb_cstates && (pkg_cstate_limit >= PCL__7);
2779         do_c8_c9_c10 = has_hsw_msrs(family, model);
2780         do_skl_residency = has_skl_msrs(family, model);
2781         do_slm_cstates = is_slm(family, model);
2782         do_knl_cstates  = is_knl(family, model);
2783         bclk = discover_bclk(family, model);
2784
2785         rapl_probe(family, model);
2786         perf_limit_reasons_probe(family, model);
2787
2788         if (debug)
2789                 dump_cstate_pstate_config_info();
2790
2791         if (has_skl_msrs(family, model))
2792                 calculate_tsc_tweak();
2793
2794         return;
2795 }
2796
2797 void help()
2798 {
2799         fprintf(stderr,
2800         "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2801         "\n"
2802         "Turbostat forks the specified COMMAND and prints statistics\n"
2803         "when COMMAND completes.\n"
2804         "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2805         "to print statistics, until interrupted.\n"
2806         "--debug        run in \"debug\" mode\n"
2807         "--interval sec Override default 5-second measurement interval\n"
2808         "--help         print this help message\n"
2809         "--counter msr  print 32-bit counter at address \"msr\"\n"
2810         "--Counter msr  print 64-bit Counter at address \"msr\"\n"
2811         "--msr msr      print 32-bit value at address \"msr\"\n"
2812         "--MSR msr      print 64-bit Value at address \"msr\"\n"
2813         "--version      print version information\n"
2814         "\n"
2815         "For more help, run \"man turbostat\"\n");
2816 }
2817
2818
2819 /*
2820  * in /dev/cpu/ return success for names that are numbers
2821  * ie. filter out ".", "..", "microcode".
2822  */
2823 int dir_filter(const struct dirent *dirp)
2824 {
2825         if (isdigit(dirp->d_name[0]))
2826                 return 1;
2827         else
2828                 return 0;
2829 }
2830
2831 int open_dev_cpu_msr(int dummy1)
2832 {
2833         return 0;
2834 }
2835
2836 void topology_probe()
2837 {
2838         int i;
2839         int max_core_id = 0;
2840         int max_package_id = 0;
2841         int max_siblings = 0;
2842         struct cpu_topology {
2843                 int core_id;
2844                 int physical_package_id;
2845         } *cpus;
2846
2847         /* Initialize num_cpus, max_cpu_num */
2848         topo.num_cpus = 0;
2849         topo.max_cpu_num = 0;
2850         for_all_proc_cpus(count_cpus);
2851         if (!summary_only && topo.num_cpus > 1)
2852                 show_cpu = 1;
2853
2854         if (debug > 1)
2855                 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2856
2857         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
2858         if (cpus == NULL)
2859                 err(1, "calloc cpus");
2860
2861         /*
2862          * Allocate and initialize cpu_present_set
2863          */
2864         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2865         if (cpu_present_set == NULL)
2866                 err(3, "CPU_ALLOC");
2867         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2868         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2869         for_all_proc_cpus(mark_cpu_present);
2870
2871         /*
2872          * Allocate and initialize cpu_affinity_set
2873          */
2874         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2875         if (cpu_affinity_set == NULL)
2876                 err(3, "CPU_ALLOC");
2877         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2878         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2879
2880
2881         /*
2882          * For online cpus
2883          * find max_core_id, max_package_id
2884          */
2885         for (i = 0; i <= topo.max_cpu_num; ++i) {
2886                 int siblings;
2887
2888                 if (cpu_is_not_present(i)) {
2889                         if (debug > 1)
2890                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2891                         continue;
2892                 }
2893                 cpus[i].core_id = get_core_id(i);
2894                 if (cpus[i].core_id > max_core_id)
2895                         max_core_id = cpus[i].core_id;
2896
2897                 cpus[i].physical_package_id = get_physical_package_id(i);
2898                 if (cpus[i].physical_package_id > max_package_id)
2899                         max_package_id = cpus[i].physical_package_id;
2900
2901                 siblings = get_num_ht_siblings(i);
2902                 if (siblings > max_siblings)
2903                         max_siblings = siblings;
2904                 if (debug > 1)
2905                         fprintf(stderr, "cpu %d pkg %d core %d\n",
2906                                 i, cpus[i].physical_package_id, cpus[i].core_id);
2907         }
2908         topo.num_cores_per_pkg = max_core_id + 1;
2909         if (debug > 1)
2910                 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2911                         max_core_id, topo.num_cores_per_pkg);
2912         if (debug && !summary_only && topo.num_cores_per_pkg > 1)
2913                 show_core = 1;
2914
2915         topo.num_packages = max_package_id + 1;
2916         if (debug > 1)
2917                 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2918                         max_package_id, topo.num_packages);
2919         if (debug && !summary_only && topo.num_packages > 1)
2920                 show_pkg = 1;
2921
2922         topo.num_threads_per_core = max_siblings;
2923         if (debug > 1)
2924                 fprintf(stderr, "max_siblings %d\n", max_siblings);
2925
2926         free(cpus);
2927 }
2928
2929 void
2930 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2931 {
2932         int i;
2933
2934         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2935                 topo.num_packages, sizeof(struct thread_data));
2936         if (*t == NULL)
2937                 goto error;
2938
2939         for (i = 0; i < topo.num_threads_per_core *
2940                 topo.num_cores_per_pkg * topo.num_packages; i++)
2941                 (*t)[i].cpu_id = -1;
2942
2943         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2944                 sizeof(struct core_data));
2945         if (*c == NULL)
2946                 goto error;
2947
2948         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2949                 (*c)[i].core_id = -1;
2950
2951         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2952         if (*p == NULL)
2953                 goto error;
2954
2955         for (i = 0; i < topo.num_packages; i++)
2956                 (*p)[i].package_id = i;
2957
2958         return;
2959 error:
2960         err(1, "calloc counters");
2961 }
2962 /*
2963  * init_counter()
2964  *
2965  * set cpu_id, core_num, pkg_num
2966  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2967  *
2968  * increment topo.num_cores when 1st core in pkg seen
2969  */
2970 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2971         struct pkg_data *pkg_base, int thread_num, int core_num,
2972         int pkg_num, int cpu_id)
2973 {
2974         struct thread_data *t;
2975         struct core_data *c;
2976         struct pkg_data *p;
2977
2978         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2979         c = GET_CORE(core_base, core_num, pkg_num);
2980         p = GET_PKG(pkg_base, pkg_num);
2981
2982         t->cpu_id = cpu_id;
2983         if (thread_num == 0) {
2984                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2985                 if (cpu_is_first_core_in_package(cpu_id))
2986                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2987         }
2988
2989         c->core_id = core_num;
2990         p->package_id = pkg_num;
2991 }
2992
2993
2994 int initialize_counters(int cpu_id)
2995 {
2996         int my_thread_id, my_core_id, my_package_id;
2997
2998         my_package_id = get_physical_package_id(cpu_id);
2999         my_core_id = get_core_id(cpu_id);
3000         my_thread_id = get_cpu_position_in_core(cpu_id);
3001         if (!my_thread_id)
3002                 topo.num_cores++;
3003
3004         init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
3005         init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
3006         return 0;
3007 }
3008
3009 void allocate_output_buffer()
3010 {
3011         output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
3012         outp = output_buffer;
3013         if (outp == NULL)
3014                 err(-1, "calloc output buffer");
3015 }
3016
3017 void setup_all_buffers(void)
3018 {
3019         topology_probe();
3020         allocate_counters(&thread_even, &core_even, &package_even);
3021         allocate_counters(&thread_odd, &core_odd, &package_odd);
3022         allocate_output_buffer();
3023         for_all_proc_cpus(initialize_counters);
3024 }
3025
3026 void set_base_cpu(void)
3027 {
3028         base_cpu = sched_getcpu();
3029         if (base_cpu < 0)
3030                 err(-ENODEV, "No valid cpus found");
3031
3032         if (debug > 1)
3033                 fprintf(stderr, "base_cpu = %d\n", base_cpu);
3034 }
3035
3036 void turbostat_init()
3037 {
3038         setup_all_buffers();
3039         set_base_cpu();
3040         check_dev_msr();
3041         check_permissions();
3042         process_cpuid();
3043
3044
3045         if (debug)
3046                 for_all_cpus(print_epb, ODD_COUNTERS);
3047
3048         if (debug)
3049                 for_all_cpus(print_perf_limit, ODD_COUNTERS);
3050
3051         if (debug)
3052                 for_all_cpus(print_rapl, ODD_COUNTERS);
3053
3054         for_all_cpus(set_temperature_target, ODD_COUNTERS);
3055
3056         if (debug)
3057                 for_all_cpus(print_thermal, ODD_COUNTERS);
3058 }
3059
3060 int fork_it(char **argv)
3061 {
3062         pid_t child_pid;
3063         int status;
3064
3065         status = for_all_cpus(get_counters, EVEN_COUNTERS);
3066         if (status)
3067                 exit(status);
3068         /* clear affinity side-effect of get_counters() */
3069         sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
3070         gettimeofday(&tv_even, (struct timezone *)NULL);
3071
3072         child_pid = fork();
3073         if (!child_pid) {
3074                 /* child */
3075                 execvp(argv[0], argv);
3076         } else {
3077
3078                 /* parent */
3079                 if (child_pid == -1)
3080                         err(1, "fork");
3081
3082                 signal(SIGINT, SIG_IGN);
3083                 signal(SIGQUIT, SIG_IGN);
3084                 if (waitpid(child_pid, &status, 0) == -1)
3085                         err(status, "waitpid");
3086         }
3087         /*
3088          * n.b. fork_it() does not check for errors from for_all_cpus()
3089          * because re-starting is problematic when forking
3090          */
3091         for_all_cpus(get_counters, ODD_COUNTERS);
3092         gettimeofday(&tv_odd, (struct timezone *)NULL);
3093         timersub(&tv_odd, &tv_even, &tv_delta);
3094         for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
3095         compute_average(EVEN_COUNTERS);
3096         format_all_counters(EVEN_COUNTERS);
3097         flush_stderr();
3098
3099         fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
3100
3101         return status;
3102 }
3103
3104 int get_and_dump_counters(void)
3105 {
3106         int status;
3107
3108         status = for_all_cpus(get_counters, ODD_COUNTERS);
3109         if (status)
3110                 return status;
3111
3112         status = for_all_cpus(dump_counters, ODD_COUNTERS);
3113         if (status)
3114                 return status;
3115
3116         flush_stdout();
3117
3118         return status;
3119 }
3120
3121 void print_version() {
3122         fprintf(stderr, "turbostat version 4.8 26-Sep, 2015"
3123                 " - Len Brown <lenb@kernel.org>\n");
3124 }
3125
3126 void cmdline(int argc, char **argv)
3127 {
3128         int opt;
3129         int option_index = 0;
3130         static struct option long_options[] = {
3131                 {"Counter",     required_argument,      0, 'C'},
3132                 {"counter",     required_argument,      0, 'c'},
3133                 {"Dump",        no_argument,            0, 'D'},
3134                 {"debug",       no_argument,            0, 'd'},
3135                 {"interval",    required_argument,      0, 'i'},
3136                 {"help",        no_argument,            0, 'h'},
3137                 {"Joules",      no_argument,            0, 'J'},
3138                 {"MSR",         required_argument,      0, 'M'},
3139                 {"msr",         required_argument,      0, 'm'},
3140                 {"Package",     no_argument,            0, 'p'},
3141                 {"processor",   no_argument,            0, 'p'},
3142                 {"Summary",     no_argument,            0, 'S'},
3143                 {"TCC",         required_argument,      0, 'T'},
3144                 {"version",     no_argument,            0, 'v' },
3145                 {0,             0,                      0,  0 }
3146         };
3147
3148         progname = argv[0];
3149
3150         while ((opt = getopt_long_only(argc, argv, "+C:c:Ddhi:JM:m:PpST:v",
3151                                 long_options, &option_index)) != -1) {
3152                 switch (opt) {
3153                 case 'C':
3154                         sscanf(optarg, "%x", &extra_delta_offset64);
3155                         break;
3156                 case 'c':
3157                         sscanf(optarg, "%x", &extra_delta_offset32);
3158                         break;
3159                 case 'D':
3160                         dump_only++;
3161                         break;
3162                 case 'd':
3163                         debug++;
3164                         break;
3165                 case 'h':
3166                 default:
3167                         help();
3168                         exit(1);
3169                 case 'i':
3170                         interval_sec = atoi(optarg);
3171                         break;
3172                 case 'J':
3173                         rapl_joules++;
3174                         break;
3175                 case 'M':
3176                         sscanf(optarg, "%x", &extra_msr_offset64);
3177                         break;
3178                 case 'm':
3179                         sscanf(optarg, "%x", &extra_msr_offset32);
3180                         break;
3181                 case 'P':
3182                         show_pkg_only++;
3183                         break;
3184                 case 'p':
3185                         show_core_only++;
3186                         break;
3187                 case 'S':
3188                         summary_only++;
3189                         break;
3190                 case 'T':
3191                         tcc_activation_temp_override = atoi(optarg);
3192                         break;
3193                 case 'v':
3194                         print_version();
3195                         exit(0);
3196                         break;
3197                 }
3198         }
3199 }
3200
3201 int main(int argc, char **argv)
3202 {
3203         cmdline(argc, argv);
3204
3205         if (debug)
3206                 print_version();
3207
3208         turbostat_init();
3209
3210         /* dump counters and exit */
3211         if (dump_only)
3212                 return get_and_dump_counters();
3213
3214         /*
3215          * if any params left, it must be a command to fork
3216          */
3217         if (argc - optind)
3218                 return fork_it(argv + optind);
3219         else
3220                 turbostat_loop();
3221
3222         return 0;
3223 }