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 <dirent.h>
  37 #include <string.h>
  38 #include <ctype.h>
  39 #include <sched.h>
  40 #include <cpuid.h>
  41 #include <linux/capability.h>
  42 #include <errno.h>
  43
  44 char *proc_stat = "/proc/stat";
  45 unsigned int interval_sec = 5;  /* set with -i interval_sec */
  46 unsigned int verbose;           /* set with -v */
  47 unsigned int rapl_verbose;      /* set with -R */
  48 unsigned int rapl_joules;       /* set with -J */
  49 unsigned int thermal_verbose;   /* set with -T */
  50 unsigned int summary_only;      /* set with -S */
  51 unsigned int dump_only;         /* set with -s */
  52 unsigned int skip_c0;
  53 unsigned int skip_c1;
  54 unsigned int do_nhm_cstates;
  55 unsigned int do_snb_cstates;
  56 unsigned int do_c8_c9_c10;
  57 unsigned int do_slm_cstates;
  58 unsigned int use_c1_residency_msr;
  59 unsigned int has_aperf;
  60 unsigned int has_epb;
  61 unsigned int units = 1000000;   /* MHz etc */
  62 unsigned int genuine_intel;
  63 unsigned int has_invariant_tsc;
  64 unsigned int do_nhm_platform_info;
  65 unsigned int do_nhm_turbo_ratio_limit;
  66 unsigned int do_ivt_turbo_ratio_limit;
  67 unsigned int extra_msr_offset32;
  68 unsigned int extra_msr_offset64;
  69 unsigned int extra_delta_offset32;
  70 unsigned int extra_delta_offset64;
  71 int do_smi;
  72 double bclk;
  73 unsigned int show_pkg;
  74 unsigned int show_core;
  75 unsigned int show_cpu;
  76 unsigned int show_pkg_only;
  77 unsigned int show_core_only;
  78 char *output_buffer, *outp;
  79 unsigned int do_rapl;
  80 unsigned int do_dts;
  81 unsigned int do_ptm;
  82 unsigned int tcc_activation_temp;
  83 unsigned int tcc_activation_temp_override;
  84 double rapl_power_units, rapl_energy_units, rapl_time_units;
  85 double rapl_joule_counter_range;
  86 unsigned int do_core_perf_limit_reasons;
  87 unsigned int do_gfx_perf_limit_reasons;
  88 unsigned int do_ring_perf_limit_reasons;
  89
  90 #define RAPL_PKG                (1 << 0)
  91                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  92                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  93 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  94                                         /* 0x613 MSR_PKG_PERF_STATUS */
  95 #define RAPL_PKG_POWER_INFO     (1 << 2)
  96                                         /* 0x614 MSR_PKG_POWER_INFO */
  97
  98 #define RAPL_DRAM               (1 << 3)
  99                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
 100                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
 101                                         /* 0x61c MSR_DRAM_POWER_INFO */
 102 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
 103                                         /* 0x61b MSR_DRAM_PERF_STATUS */
 104
 105 #define RAPL_CORES              (1 << 5)
 106                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 107                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 108 #define RAPL_CORE_POLICY        (1 << 6)
 109                                         /* 0x63a MSR_PP0_POLICY */
 110
 111
 112 #define RAPL_GFX                (1 << 7)
 113                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 114                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 115                                         /* 0x642 MSR_PP1_POLICY */
 116 #define TJMAX_DEFAULT   100
 117
 118 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 119
 120 int aperf_mperf_unstable;
 121 int backwards_count;
 122 char *progname;
 123
 124 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 125 size_t cpu_present_setsize, cpu_affinity_setsize;
 126
 127 struct thread_data {
 128         unsigned long long tsc;
 129         unsigned long long aperf;
 130         unsigned long long mperf;
 131         unsigned long long c1;
 132         unsigned long long extra_msr64;
 133         unsigned long long extra_delta64;
 134         unsigned long long extra_msr32;
 135         unsigned long long extra_delta32;
 136         unsigned int smi_count;
 137         unsigned int cpu_id;
 138         unsigned int flags;
 139 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 140 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 141 } *thread_even, *thread_odd;
 142
 143 struct core_data {
 144         unsigned long long c3;
 145         unsigned long long c6;
 146         unsigned long long c7;
 147         unsigned int core_temp_c;
 148         unsigned int core_id;
 149 } *core_even, *core_odd;
 150
 151 struct pkg_data {
 152         unsigned long long pc2;
 153         unsigned long long pc3;
 154         unsigned long long pc6;
 155         unsigned long long pc7;
 156         unsigned long long pc8;
 157         unsigned long long pc9;
 158         unsigned long long pc10;
 159         unsigned int package_id;
 160         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 161         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 162         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 163         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 164         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 165         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 166         unsigned int pkg_temp_c;
 167
 168 } *package_even, *package_odd;
 169
 170 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 171 #define EVEN_COUNTERS thread_even, core_even, package_even
 172
 173 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 174         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 175                 topo.num_threads_per_core + \
 176                 (core_no) * topo.num_threads_per_core + (thread_no))
 177 #define GET_CORE(core_base, core_no, pkg_no) \
 178         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 179 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 180
 181 struct system_summary {
 182         struct thread_data threads;
 183         struct core_data cores;
 184         struct pkg_data packages;
 185 } sum, average;
 186
 187
 188 struct topo_params {
 189         int num_packages;
 190         int num_cpus;
 191         int num_cores;
 192         int max_cpu_num;
 193         int num_cores_per_pkg;
 194         int num_threads_per_core;
 195 } topo;
 196
 197 struct timeval tv_even, tv_odd, tv_delta;
 198
 199 void setup_all_buffers(void);
 200
 201 int cpu_is_not_present(int cpu)
 202 {
 203         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 204 }
 205 /*
 206  * run func(thread, core, package) in topology order
 207  * skip non-present cpus
 208  */
 209
 210 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 211         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 212 {
 213         int retval, pkg_no, core_no, thread_no;
 214
 215         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 216                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 217                         for (thread_no = 0; thread_no <
 218                                 topo.num_threads_per_core; ++thread_no) {
 219                                 struct thread_data *t;
 220                                 struct core_data *c;
 221                                 struct pkg_data *p;
 222
 223                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 224
 225                                 if (cpu_is_not_present(t->cpu_id))
 226                                         continue;
 227
 228                                 c = GET_CORE(core_base, core_no, pkg_no);
 229                                 p = GET_PKG(pkg_base, pkg_no);
 230
 231                                 retval = func(t, c, p);
 232                                 if (retval)
 233                                         return retval;
 234                         }
 235                 }
 236         }
 237         return 0;
 238 }
 239
 240 int cpu_migrate(int cpu)
 241 {
 242         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 243         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 244         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 245                 return -1;
 246         else
 247                 return 0;
 248 }
 249
 250 int get_msr(int cpu, off_t offset, unsigned long long *msr)
 251 {
 252         ssize_t retval;
 253         char pathname[32];
 254         int fd;
 255
 256         sprintf(pathname, "/dev/cpu/%d/msr", cpu);
 257         fd = open(pathname, O_RDONLY);
 258         if (fd < 0)
 259                 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
 260
 261         retval = pread(fd, msr, sizeof *msr, offset);
 262         close(fd);
 263
 264         if (retval != sizeof *msr)
 265                 err(-1, "%s offset 0x%llx read failed", pathname, (unsigned long long)offset);
 266
 267         return 0;
 268 }
 269
 270 /*
 271  * Example Format w/ field column widths:
 272  *
 273  *  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
 274  * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
 275  */
 276
 277 void print_header(void)
 278 {
 279         if (show_pkg)
 280                 outp += sprintf(outp, " Package");
 281         if (show_core)
 282                 outp += sprintf(outp, "    Core");
 283         if (show_cpu)
 284                 outp += sprintf(outp, "     CPU");
 285         if (has_aperf)
 286                 outp += sprintf(outp, " Avg_MHz");
 287         if (has_aperf)
 288                 outp += sprintf(outp, "   %%Busy");
 289         if (has_aperf)
 290                 outp += sprintf(outp, " Bzy_MHz");
 291         outp += sprintf(outp, " TSC_MHz");
 292         if (do_smi)
 293                 outp += sprintf(outp, "     SMI");
 294         if (extra_delta_offset32)
 295                 outp += sprintf(outp, "  count 0x%03X", extra_delta_offset32);
 296         if (extra_delta_offset64)
 297                 outp += sprintf(outp, "  COUNT 0x%03X", extra_delta_offset64);
 298         if (extra_msr_offset32)
 299                 outp += sprintf(outp, "   MSR 0x%03X", extra_msr_offset32);
 300         if (extra_msr_offset64)
 301                 outp += sprintf(outp, "           MSR 0x%03X", extra_msr_offset64);
 302         if (do_nhm_cstates)
 303                 outp += sprintf(outp, "  CPU%%c1");
 304         if (do_nhm_cstates && !do_slm_cstates)
 305                 outp += sprintf(outp, "  CPU%%c3");
 306         if (do_nhm_cstates)
 307                 outp += sprintf(outp, "  CPU%%c6");
 308         if (do_snb_cstates)
 309                 outp += sprintf(outp, "  CPU%%c7");
 310
 311         if (do_dts)
 312                 outp += sprintf(outp, " CoreTmp");
 313         if (do_ptm)
 314                 outp += sprintf(outp, "  PkgTmp");
 315
 316         if (do_snb_cstates)
 317                 outp += sprintf(outp, " Pkg%%pc2");
 318         if (do_nhm_cstates && !do_slm_cstates)
 319                 outp += sprintf(outp, " Pkg%%pc3");
 320         if (do_nhm_cstates && !do_slm_cstates)
 321                 outp += sprintf(outp, " Pkg%%pc6");
 322         if (do_snb_cstates)
 323                 outp += sprintf(outp, " Pkg%%pc7");
 324         if (do_c8_c9_c10) {
 325                 outp += sprintf(outp, " Pkg%%pc8");
 326                 outp += sprintf(outp, " Pkg%%pc9");
 327                 outp += sprintf(outp, " Pk%%pc10");
 328         }
 329
 330         if (do_rapl && !rapl_joules) {
 331                 if (do_rapl & RAPL_PKG)
 332                         outp += sprintf(outp, " PkgWatt");
 333                 if (do_rapl & RAPL_CORES)
 334                         outp += sprintf(outp, " CorWatt");
 335                 if (do_rapl & RAPL_GFX)
 336                         outp += sprintf(outp, " GFXWatt");
 337                 if (do_rapl & RAPL_DRAM)
 338                         outp += sprintf(outp, " RAMWatt");
 339                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 340                         outp += sprintf(outp, "   PKG_%%");
 341                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 342                         outp += sprintf(outp, "   RAM_%%");
 343         } else if (do_rapl && rapl_joules) {
 344                 if (do_rapl & RAPL_PKG)
 345                         outp += sprintf(outp, "   Pkg_J");
 346                 if (do_rapl & RAPL_CORES)
 347                         outp += sprintf(outp, "   Cor_J");
 348                 if (do_rapl & RAPL_GFX)
 349                         outp += sprintf(outp, "   GFX_J");
 350                 if (do_rapl & RAPL_DRAM)
 351                         outp += sprintf(outp, "   RAM_W");
 352                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 353                         outp += sprintf(outp, "   PKG_%%");
 354                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 355                         outp += sprintf(outp, "   RAM_%%");
 356                 outp += sprintf(outp, "   time");
 357
 358         }
 359         outp += sprintf(outp, "\n");
 360 }
 361
 362 int dump_counters(struct thread_data *t, struct core_data *c,
 363         struct pkg_data *p)
 364 {
 365         outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
 366
 367         if (t) {
 368                 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
 369                         t->cpu_id, t->flags);
 370                 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
 371                 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
 372                 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
 373                 outp += sprintf(outp, "c1: %016llX\n", t->c1);
 374                 outp += sprintf(outp, "msr0x%x: %08llX\n",
 375                         extra_delta_offset32, t->extra_delta32);
 376                 outp += sprintf(outp, "msr0x%x: %016llX\n",
 377                         extra_delta_offset64, t->extra_delta64);
 378                 outp += sprintf(outp, "msr0x%x: %08llX\n",
 379                         extra_msr_offset32, t->extra_msr32);
 380                 outp += sprintf(outp, "msr0x%x: %016llX\n",
 381                         extra_msr_offset64, t->extra_msr64);
 382                 if (do_smi)
 383                         outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
 384         }
 385
 386         if (c) {
 387                 outp += sprintf(outp, "core: %d\n", c->core_id);
 388                 outp += sprintf(outp, "c3: %016llX\n", c->c3);
 389                 outp += sprintf(outp, "c6: %016llX\n", c->c6);
 390                 outp += sprintf(outp, "c7: %016llX\n", c->c7);
 391                 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
 392         }
 393
 394         if (p) {
 395                 outp += sprintf(outp, "package: %d\n", p->package_id);
 396                 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
 397                 outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
 398                 outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
 399                 outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
 400                 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
 401                 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
 402                 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
 403                 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
 404                 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
 405                 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
 406                 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
 407                 outp += sprintf(outp, "Throttle PKG: %0X\n",
 408                         p->rapl_pkg_perf_status);
 409                 outp += sprintf(outp, "Throttle RAM: %0X\n",
 410                         p->rapl_dram_perf_status);
 411                 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
 412         }
 413
 414         outp += sprintf(outp, "\n");
 415
 416         return 0;
 417 }
 418
 419 /*
 420  * column formatting convention & formats
 421  */
 422 int format_counters(struct thread_data *t, struct core_data *c,
 423         struct pkg_data *p)
 424 {
 425         double interval_float;
 426         char *fmt8;
 427
 428          /* if showing only 1st thread in core and this isn't one, bail out */
 429         if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 430                 return 0;
 431
 432          /* if showing only 1st thread in pkg and this isn't one, bail out */
 433         if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 434                 return 0;
 435
 436         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 437
 438         /* topo columns, print blanks on 1st (average) line */
 439         if (t == &average.threads) {
 440                 if (show_pkg)
 441                         outp += sprintf(outp, "       -");
 442                 if (show_core)
 443                         outp += sprintf(outp, "       -");
 444                 if (show_cpu)
 445                         outp += sprintf(outp, "       -");
 446         } else {
 447                 if (show_pkg) {
 448                         if (p)
 449                                 outp += sprintf(outp, "%8d", p->package_id);
 450                         else
 451                                 outp += sprintf(outp, "       -");
 452                 }
 453                 if (show_core) {
 454                         if (c)
 455                                 outp += sprintf(outp, "%8d", c->core_id);
 456                         else
 457                                 outp += sprintf(outp, "       -");
 458                 }
 459                 if (show_cpu)
 460                         outp += sprintf(outp, "%8d", t->cpu_id);
 461         }
 462
 463         /* Avg_MHz */
 464         if (has_aperf)
 465                 outp += sprintf(outp, "%8.0f",
 466                         1.0 / units * t->aperf / interval_float);
 467
 468         /* %Busy */
 469         if (has_aperf) {
 470                 if (!skip_c0)
 471                         outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc);
 472                 else
 473                         outp += sprintf(outp, "********");
 474         }
 475
 476         /* Bzy_MHz */
 477         if (has_aperf)
 478                 outp += sprintf(outp, "%8.0f",
 479                         1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
 480
 481         /* TSC_MHz */
 482         outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
 483
 484         /* SMI */
 485         if (do_smi)
 486                 outp += sprintf(outp, "%8d", t->smi_count);
 487
 488         /* delta */
 489         if (extra_delta_offset32)
 490                 outp += sprintf(outp, "  %11llu", t->extra_delta32);
 491
 492         /* DELTA */
 493         if (extra_delta_offset64)
 494                 outp += sprintf(outp, "  %11llu", t->extra_delta64);
 495         /* msr */
 496         if (extra_msr_offset32)
 497                 outp += sprintf(outp, "  0x%08llx", t->extra_msr32);
 498
 499         /* MSR */
 500         if (extra_msr_offset64)
 501                 outp += sprintf(outp, "  0x%016llx", t->extra_msr64);
 502
 503         if (do_nhm_cstates) {
 504                 if (!skip_c1)
 505                         outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
 506                 else
 507                         outp += sprintf(outp, "********");
 508         }
 509
 510         /* print per-core data only for 1st thread in core */
 511         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 512                 goto done;
 513
 514         if (do_nhm_cstates && !do_slm_cstates)
 515                 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
 516         if (do_nhm_cstates)
 517                 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
 518         if (do_snb_cstates)
 519                 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
 520
 521         if (do_dts)
 522                 outp += sprintf(outp, "%8d", c->core_temp_c);
 523
 524         /* print per-package data only for 1st core in package */
 525         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 526                 goto done;
 527
 528         if (do_ptm)
 529                 outp += sprintf(outp, "%8d", p->pkg_temp_c);
 530
 531         if (do_snb_cstates)
 532                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
 533         if (do_nhm_cstates && !do_slm_cstates)
 534                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
 535         if (do_nhm_cstates && !do_slm_cstates)
 536                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
 537         if (do_snb_cstates)
 538                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
 539         if (do_c8_c9_c10) {
 540                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
 541                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
 542                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
 543         }
 544
 545         /*
 546          * If measurement interval exceeds minimum RAPL Joule Counter range,
 547          * indicate that results are suspect by printing "**" in fraction place.
 548          */
 549         if (interval_float < rapl_joule_counter_range)
 550                 fmt8 = "%8.2f";
 551         else
 552                 fmt8 = " %6.0f**";
 553
 554         if (do_rapl && !rapl_joules) {
 555                 if (do_rapl & RAPL_PKG)
 556                         outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
 557                 if (do_rapl & RAPL_CORES)
 558                         outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
 559                 if (do_rapl & RAPL_GFX)
 560                         outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
 561                 if (do_rapl & RAPL_DRAM)
 562                         outp += sprintf(outp, fmt8, p->energy_dram * rapl_energy_units / interval_float);
 563                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 564                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
 565                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 566                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
 567         } else if (do_rapl && rapl_joules) {
 568                 if (do_rapl & RAPL_PKG)
 569                         outp += sprintf(outp, fmt8,
 570                                         p->energy_pkg * rapl_energy_units);
 571                 if (do_rapl & RAPL_CORES)
 572                         outp += sprintf(outp, fmt8,
 573                                         p->energy_cores * rapl_energy_units);
 574                 if (do_rapl & RAPL_GFX)
 575                         outp += sprintf(outp, fmt8,
 576                                         p->energy_gfx * rapl_energy_units);
 577                 if (do_rapl & RAPL_DRAM)
 578                         outp += sprintf(outp, fmt8,
 579                                         p->energy_dram * rapl_energy_units);
 580                 if (do_rapl & RAPL_PKG_PERF_STATUS)
 581                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
 582                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
 583                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
 584
 585                 outp += sprintf(outp, fmt8, interval_float);
 586         }
 587 done:
 588         outp += sprintf(outp, "\n");
 589
 590         return 0;
 591 }
 592
 593 void flush_stdout()
 594 {
 595         fputs(output_buffer, stdout);
 596         fflush(stdout);
 597         outp = output_buffer;
 598 }
 599 void flush_stderr()
 600 {
 601         fputs(output_buffer, stderr);
 602         outp = output_buffer;
 603 }
 604 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 605 {
 606         static int printed;
 607
 608         if (!printed || !summary_only)
 609                 print_header();
 610
 611         if (topo.num_cpus > 1)
 612                 format_counters(&average.threads, &average.cores,
 613                         &average.packages);
 614
 615         printed = 1;
 616
 617         if (summary_only)
 618                 return;
 619
 620         for_all_cpus(format_counters, t, c, p);
 621 }
 622
 623 #define DELTA_WRAP32(new, old)                  \
 624         if (new > old) {                        \
 625                 old = new - old;                \
 626         } else {                                \
 627                 old = 0x100000000 + new - old;  \
 628         }
 629
 630 void
 631 delta_package(struct pkg_data *new, struct pkg_data *old)
 632 {
 633         old->pc2 = new->pc2 - old->pc2;
 634         old->pc3 = new->pc3 - old->pc3;
 635         old->pc6 = new->pc6 - old->pc6;
 636         old->pc7 = new->pc7 - old->pc7;
 637         old->pc8 = new->pc8 - old->pc8;
 638         old->pc9 = new->pc9 - old->pc9;
 639         old->pc10 = new->pc10 - old->pc10;
 640         old->pkg_temp_c = new->pkg_temp_c;
 641
 642         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 643         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 644         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 645         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 646         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 647         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 648 }
 649
 650 void
 651 delta_core(struct core_data *new, struct core_data *old)
 652 {
 653         old->c3 = new->c3 - old->c3;
 654         old->c6 = new->c6 - old->c6;
 655         old->c7 = new->c7 - old->c7;
 656         old->core_temp_c = new->core_temp_c;
 657 }
 658
 659 /*
 660  * old = new - old
 661  */
 662 void
 663 delta_thread(struct thread_data *new, struct thread_data *old,
 664         struct core_data *core_delta)
 665 {
 666         old->tsc = new->tsc - old->tsc;
 667
 668         /* check for TSC < 1 Mcycles over interval */
 669         if (old->tsc < (1000 * 1000))
 670                 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
 671                      "You can disable all c-states by booting with \"idle=poll\"\n"
 672                      "or just the deep ones with \"processor.max_cstate=1\"");
 673
 674         old->c1 = new->c1 - old->c1;
 675
 676         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 677                 old->aperf = new->aperf - old->aperf;
 678                 old->mperf = new->mperf - old->mperf;
 679         } else {
 680
 681                 if (!aperf_mperf_unstable) {
 682                         fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
 683                         fprintf(stderr, "* Frequency results do not cover entire interval *\n");
 684                         fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
 685
 686                         aperf_mperf_unstable = 1;
 687                 }
 688                 /*
 689                  * mperf delta is likely a huge "positive" number
 690                  * can not use it for calculating c0 time
 691                  */
 692                 skip_c0 = 1;
 693                 skip_c1 = 1;
 694         }
 695
 696
 697         if (use_c1_residency_msr) {
 698                 /*
 699                  * Some models have a dedicated C1 residency MSR,
 700                  * which should be more accurate than the derivation below.
 701                  */
 702         } else {
 703                 /*
 704                  * As counter collection is not atomic,
 705                  * it is possible for mperf's non-halted cycles + idle states
 706                  * to exceed TSC's all cycles: show c1 = 0% in that case.
 707                  */
 708                 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 709                         old->c1 = 0;
 710                 else {
 711                         /* normal case, derive c1 */
 712                         old->c1 = old->tsc - old->mperf - core_delta->c3
 713                                 - core_delta->c6 - core_delta->c7;
 714                 }
 715         }
 716
 717         if (old->mperf == 0) {
 718                 if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
 719                 old->mperf = 1; /* divide by 0 protection */
 720         }
 721
 722         old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
 723         old->extra_delta32 &= 0xFFFFFFFF;
 724
 725         old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
 726
 727         /*
 728          * Extra MSR is just a snapshot, simply copy latest w/o subtracting
 729          */
 730         old->extra_msr32 = new->extra_msr32;
 731         old->extra_msr64 = new->extra_msr64;
 732
 733         if (do_smi)
 734                 old->smi_count = new->smi_count - old->smi_count;
 735 }
 736
 737 int delta_cpu(struct thread_data *t, struct core_data *c,
 738         struct pkg_data *p, struct thread_data *t2,
 739         struct core_data *c2, struct pkg_data *p2)
 740 {
 741         /* calculate core delta only for 1st thread in core */
 742         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 743                 delta_core(c, c2);
 744
 745         /* always calculate thread delta */
 746         delta_thread(t, t2, c2);        /* c2 is core delta */
 747
 748         /* calculate package delta only for 1st core in package */
 749         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 750                 delta_package(p, p2);
 751
 752         return 0;
 753 }
 754
 755 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 756 {
 757         t->tsc = 0;
 758         t->aperf = 0;
 759         t->mperf = 0;
 760         t->c1 = 0;
 761
 762         t->smi_count = 0;
 763         t->extra_delta32 = 0;
 764         t->extra_delta64 = 0;
 765
 766         /* tells format_counters to dump all fields from this set */
 767         t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
 768
 769         c->c3 = 0;
 770         c->c6 = 0;
 771         c->c7 = 0;
 772         c->core_temp_c = 0;
 773
 774         p->pc2 = 0;
 775         p->pc3 = 0;
 776         p->pc6 = 0;
 777         p->pc7 = 0;
 778         p->pc8 = 0;
 779         p->pc9 = 0;
 780         p->pc10 = 0;
 781
 782         p->energy_pkg = 0;
 783         p->energy_dram = 0;
 784         p->energy_cores = 0;
 785         p->energy_gfx = 0;
 786         p->rapl_pkg_perf_status = 0;
 787         p->rapl_dram_perf_status = 0;
 788         p->pkg_temp_c = 0;
 789 }
 790 int sum_counters(struct thread_data *t, struct core_data *c,
 791         struct pkg_data *p)
 792 {
 793         average.threads.tsc += t->tsc;
 794         average.threads.aperf += t->aperf;
 795         average.threads.mperf += t->mperf;
 796         average.threads.c1 += t->c1;
 797
 798         average.threads.extra_delta32 += t->extra_delta32;
 799         average.threads.extra_delta64 += t->extra_delta64;
 800
 801         /* sum per-core values only for 1st thread in core */
 802         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 803                 return 0;
 804
 805         average.cores.c3 += c->c3;
 806         average.cores.c6 += c->c6;
 807         average.cores.c7 += c->c7;
 808
 809         average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
 810
 811         /* sum per-pkg values only for 1st core in pkg */
 812         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 813                 return 0;
 814
 815         average.packages.pc2 += p->pc2;
 816         average.packages.pc3 += p->pc3;
 817         average.packages.pc6 += p->pc6;
 818         average.packages.pc7 += p->pc7;
 819         average.packages.pc8 += p->pc8;
 820         average.packages.pc9 += p->pc9;
 821         average.packages.pc10 += p->pc10;
 822
 823         average.packages.energy_pkg += p->energy_pkg;
 824         average.packages.energy_dram += p->energy_dram;
 825         average.packages.energy_cores += p->energy_cores;
 826         average.packages.energy_gfx += p->energy_gfx;
 827
 828         average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
 829
 830         average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
 831         average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
 832         return 0;
 833 }
 834 /*
 835  * sum the counters for all cpus in the system
 836  * compute the weighted average
 837  */
 838 void compute_average(struct thread_data *t, struct core_data *c,
 839         struct pkg_data *p)
 840 {
 841         clear_counters(&average.threads, &average.cores, &average.packages);
 842
 843         for_all_cpus(sum_counters, t, c, p);
 844
 845         average.threads.tsc /= topo.num_cpus;
 846         average.threads.aperf /= topo.num_cpus;
 847         average.threads.mperf /= topo.num_cpus;
 848         average.threads.c1 /= topo.num_cpus;
 849
 850         average.threads.extra_delta32 /= topo.num_cpus;
 851         average.threads.extra_delta32 &= 0xFFFFFFFF;
 852
 853         average.threads.extra_delta64 /= topo.num_cpus;
 854
 855         average.cores.c3 /= topo.num_cores;
 856         average.cores.c6 /= topo.num_cores;
 857         average.cores.c7 /= topo.num_cores;
 858
 859         average.packages.pc2 /= topo.num_packages;
 860         average.packages.pc3 /= topo.num_packages;
 861         average.packages.pc6 /= topo.num_packages;
 862         average.packages.pc7 /= topo.num_packages;
 863
 864         average.packages.pc8 /= topo.num_packages;
 865         average.packages.pc9 /= topo.num_packages;
 866         average.packages.pc10 /= topo.num_packages;
 867 }
 868
 869 static unsigned long long rdtsc(void)
 870 {
 871         unsigned int low, high;
 872
 873         asm volatile("rdtsc" : "=a" (low), "=d" (high));
 874
 875         return low | ((unsigned long long)high) << 32;
 876 }
 877
 878
 879 /*
 880  * get_counters(...)
 881  * migrate to cpu
 882  * acquire and record local counters for that cpu
 883  */
 884 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 885 {
 886         int cpu = t->cpu_id;
 887         unsigned long long msr;
 888
 889         if (cpu_migrate(cpu)) {
 890                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
 891                 return -1;
 892         }
 893
 894         t->tsc = rdtsc();       /* we are running on local CPU of interest */
 895
 896         if (has_aperf) {
 897                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
 898                         return -3;
 899                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
 900                         return -4;
 901         }
 902
 903         if (do_smi) {
 904                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
 905                         return -5;
 906                 t->smi_count = msr & 0xFFFFFFFF;
 907         }
 908         if (extra_delta_offset32) {
 909                 if (get_msr(cpu, extra_delta_offset32, &msr))
 910                         return -5;
 911                 t->extra_delta32 = msr & 0xFFFFFFFF;
 912         }
 913
 914         if (extra_delta_offset64)
 915                 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
 916                         return -5;
 917
 918         if (extra_msr_offset32) {
 919                 if (get_msr(cpu, extra_msr_offset32, &msr))
 920                         return -5;
 921                 t->extra_msr32 = msr & 0xFFFFFFFF;
 922         }
 923
 924         if (extra_msr_offset64)
 925                 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
 926                         return -5;
 927
 928         if (use_c1_residency_msr) {
 929                 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
 930                         return -6;
 931         }
 932
 933         /* collect core counters only for 1st thread in core */
 934         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 935                 return 0;
 936
 937         if (do_nhm_cstates && !do_slm_cstates) {
 938                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
 939                         return -6;
 940         }
 941
 942         if (do_nhm_cstates) {
 943                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
 944                         return -7;
 945         }
 946
 947         if (do_snb_cstates)
 948                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
 949                         return -8;
 950
 951         if (do_dts) {
 952                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
 953                         return -9;
 954                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 955         }
 956
 957
 958         /* collect package counters only for 1st core in package */
 959         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 960                 return 0;
 961
 962         if (do_nhm_cstates && !do_slm_cstates) {
 963                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
 964                         return -9;
 965                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
 966                         return -10;
 967         }
 968         if (do_snb_cstates) {
 969                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
 970                         return -11;
 971                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
 972                         return -12;
 973         }
 974         if (do_c8_c9_c10) {
 975                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
 976                         return -13;
 977                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
 978                         return -13;
 979                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
 980                         return -13;
 981         }
 982         if (do_rapl & RAPL_PKG) {
 983                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
 984                         return -13;
 985                 p->energy_pkg = msr & 0xFFFFFFFF;
 986         }
 987         if (do_rapl & RAPL_CORES) {
 988                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
 989                         return -14;
 990                 p->energy_cores = msr & 0xFFFFFFFF;
 991         }
 992         if (do_rapl & RAPL_DRAM) {
 993                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
 994                         return -15;
 995                 p->energy_dram = msr & 0xFFFFFFFF;
 996         }
 997         if (do_rapl & RAPL_GFX) {
 998                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
 999                         return -16;
1000                 p->energy_gfx = msr & 0xFFFFFFFF;
1001         }
1002         if (do_rapl & RAPL_PKG_PERF_STATUS) {
1003                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1004                         return -16;
1005                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1006         }
1007         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1008                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1009                         return -16;
1010                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1011         }
1012         if (do_ptm) {
1013                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1014                         return -17;
1015                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1016         }
1017         return 0;
1018 }
1019
1020 void print_verbose_header(void)
1021 {
1022         unsigned long long msr;
1023         unsigned int ratio;
1024
1025         if (!do_nhm_platform_info)
1026                 return;
1027
1028         get_msr(0, MSR_NHM_PLATFORM_INFO, &msr);
1029
1030         fprintf(stderr, "cpu0: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", msr);
1031
1032         ratio = (msr >> 40) & 0xFF;
1033         fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
1034                 ratio, bclk, ratio * bclk);
1035
1036         ratio = (msr >> 8) & 0xFF;
1037         fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
1038                 ratio, bclk, ratio * bclk);
1039
1040         get_msr(0, MSR_IA32_POWER_CTL, &msr);
1041         fprintf(stderr, "cpu0: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1042                 msr, msr & 0x2 ? "EN" : "DIS");
1043
1044         if (!do_ivt_turbo_ratio_limit)
1045                 goto print_nhm_turbo_ratio_limits;
1046
1047         get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr);
1048
1049         fprintf(stderr, "cpu0: MSR_IVT_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1050
1051         ratio = (msr >> 56) & 0xFF;
1052         if (ratio)
1053                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1054                         ratio, bclk, ratio * bclk);
1055
1056         ratio = (msr >> 48) & 0xFF;
1057         if (ratio)
1058                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1059                         ratio, bclk, ratio * bclk);
1060
1061         ratio = (msr >> 40) & 0xFF;
1062         if (ratio)
1063                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1064                         ratio, bclk, ratio * bclk);
1065
1066         ratio = (msr >> 32) & 0xFF;
1067         if (ratio)
1068                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1069                         ratio, bclk, ratio * bclk);
1070
1071         ratio = (msr >> 24) & 0xFF;
1072         if (ratio)
1073                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1074                         ratio, bclk, ratio * bclk);
1075
1076         ratio = (msr >> 16) & 0xFF;
1077         if (ratio)
1078                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1079                         ratio, bclk, ratio * bclk);
1080
1081         ratio = (msr >> 8) & 0xFF;
1082         if (ratio)
1083                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1084                         ratio, bclk, ratio * bclk);
1085
1086         ratio = (msr >> 0) & 0xFF;
1087         if (ratio)
1088                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1089                         ratio, bclk, ratio * bclk);
1090
1091 print_nhm_turbo_ratio_limits:
1092         get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1093
1094 #define SNB_C1_AUTO_UNDEMOTE              (1UL << 27)
1095 #define SNB_C3_AUTO_UNDEMOTE              (1UL << 28)
1096
1097         fprintf(stderr, "cpu0: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", msr);
1098
1099         fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: ",
1100                 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1101                 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1102                 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1103                 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1104                 (msr & (1 << 15)) ? "" : "UN",
1105                 (unsigned int)msr & 7);
1106
1107
1108         switch(msr & 0x7) {
1109         case 0:
1110                 fprintf(stderr, do_slm_cstates ? "no pkg states" : "pc0");
1111                 break;
1112         case 1:
1113                 fprintf(stderr, do_slm_cstates ? "no pkg states" : do_snb_cstates ? "pc2" : "pc0");
1114                 break;
1115         case 2:
1116                 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc6-noret" : "pc3");
1117                 break;
1118         case 3:
1119                 fprintf(stderr, do_slm_cstates ? "invalid" : "pc6");
1120                 break;
1121         case 4:
1122                 fprintf(stderr, do_slm_cstates ? "pc4" : "pc7");
1123                 break;
1124         case 5:
1125                 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc7s" : "invalid");
1126                 break;
1127         case 6:
1128                 fprintf(stderr, do_slm_cstates ? "pc6" : "invalid");
1129                 break;
1130         case 7:
1131                 fprintf(stderr, do_slm_cstates ? "pc7" : "unlimited");
1132                 break;
1133         default:
1134                 fprintf(stderr, "invalid");
1135         }
1136         fprintf(stderr, ")\n");
1137
1138         if (!do_nhm_turbo_ratio_limit)
1139                 return;
1140
1141         get_msr(0, MSR_NHM_TURBO_RATIO_LIMIT, &msr);
1142
1143         fprintf(stderr, "cpu0: MSR_NHM_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1144
1145         ratio = (msr >> 56) & 0xFF;
1146         if (ratio)
1147                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1148                         ratio, bclk, ratio * bclk);
1149
1150         ratio = (msr >> 48) & 0xFF;
1151         if (ratio)
1152                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1153                         ratio, bclk, ratio * bclk);
1154
1155         ratio = (msr >> 40) & 0xFF;
1156         if (ratio)
1157                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1158                         ratio, bclk, ratio * bclk);
1159
1160         ratio = (msr >> 32) & 0xFF;
1161         if (ratio)
1162                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1163                         ratio, bclk, ratio * bclk);
1164
1165         ratio = (msr >> 24) & 0xFF;
1166         if (ratio)
1167                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1168                         ratio, bclk, ratio * bclk);
1169
1170         ratio = (msr >> 16) & 0xFF;
1171         if (ratio)
1172                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1173                         ratio, bclk, ratio * bclk);
1174
1175         ratio = (msr >> 8) & 0xFF;
1176         if (ratio)
1177                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1178                         ratio, bclk, ratio * bclk);
1179
1180         ratio = (msr >> 0) & 0xFF;
1181         if (ratio)
1182                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1183                         ratio, bclk, ratio * bclk);
1184
1185 }
1186
1187 void free_all_buffers(void)
1188 {
1189         CPU_FREE(cpu_present_set);
1190         cpu_present_set = NULL;
1191         cpu_present_set = 0;
1192
1193         CPU_FREE(cpu_affinity_set);
1194         cpu_affinity_set = NULL;
1195         cpu_affinity_setsize = 0;
1196
1197         free(thread_even);
1198         free(core_even);
1199         free(package_even);
1200
1201         thread_even = NULL;
1202         core_even = NULL;
1203         package_even = NULL;
1204
1205         free(thread_odd);
1206         free(core_odd);
1207         free(package_odd);
1208
1209         thread_odd = NULL;
1210         core_odd = NULL;
1211         package_odd = NULL;
1212
1213         free(output_buffer);
1214         output_buffer = NULL;
1215         outp = NULL;
1216 }
1217
1218 /*
1219  * Open a file, and exit on failure
1220  */
1221 FILE *fopen_or_die(const char *path, const char *mode)
1222 {
1223         FILE *filep = fopen(path, "r");
1224         if (!filep)
1225                 err(1, "%s: open failed", path);
1226         return filep;
1227 }
1228
1229 /*
1230  * Parse a file containing a single int.
1231  */
1232 int parse_int_file(const char *fmt, ...)
1233 {
1234         va_list args;
1235         char path[PATH_MAX];
1236         FILE *filep;
1237         int value;
1238
1239         va_start(args, fmt);
1240         vsnprintf(path, sizeof(path), fmt, args);
1241         va_end(args);
1242         filep = fopen_or_die(path, "r");
1243         if (fscanf(filep, "%d", &value) != 1)
1244                 err(1, "%s: failed to parse number from file", path);
1245         fclose(filep);
1246         return value;
1247 }
1248
1249 /*
1250  * cpu_is_first_sibling_in_core(cpu)
1251  * return 1 if given CPU is 1st HT sibling in the core
1252  */
1253 int cpu_is_first_sibling_in_core(int cpu)
1254 {
1255         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1256 }
1257
1258 /*
1259  * cpu_is_first_core_in_package(cpu)
1260  * return 1 if given CPU is 1st core in package
1261  */
1262 int cpu_is_first_core_in_package(int cpu)
1263 {
1264         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1265 }
1266
1267 int get_physical_package_id(int cpu)
1268 {
1269         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1270 }
1271
1272 int get_core_id(int cpu)
1273 {
1274         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1275 }
1276
1277 int get_num_ht_siblings(int cpu)
1278 {
1279         char path[80];
1280         FILE *filep;
1281         int sib1, sib2;
1282         int matches;
1283         char character;
1284
1285         sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1286         filep = fopen_or_die(path, "r");
1287         /*
1288          * file format:
1289          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1290          * otherwinse 1 sibling (self).
1291          */
1292         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1293
1294         fclose(filep);
1295
1296         if (matches == 3)
1297                 return 2;
1298         else
1299                 return 1;
1300 }
1301
1302 /*
1303  * run func(thread, core, package) in topology order
1304  * skip non-present cpus
1305  */
1306
1307 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1308         struct pkg_data *, struct thread_data *, struct core_data *,
1309         struct pkg_data *), struct thread_data *thread_base,
1310         struct core_data *core_base, struct pkg_data *pkg_base,
1311         struct thread_data *thread_base2, struct core_data *core_base2,
1312         struct pkg_data *pkg_base2)
1313 {
1314         int retval, pkg_no, core_no, thread_no;
1315
1316         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1317                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1318                         for (thread_no = 0; thread_no <
1319                                 topo.num_threads_per_core; ++thread_no) {
1320                                 struct thread_data *t, *t2;
1321                                 struct core_data *c, *c2;
1322                                 struct pkg_data *p, *p2;
1323
1324                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1325
1326                                 if (cpu_is_not_present(t->cpu_id))
1327                                         continue;
1328
1329                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1330
1331                                 c = GET_CORE(core_base, core_no, pkg_no);
1332                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
1333
1334                                 p = GET_PKG(pkg_base, pkg_no);
1335                                 p2 = GET_PKG(pkg_base2, pkg_no);
1336
1337                                 retval = func(t, c, p, t2, c2, p2);
1338                                 if (retval)
1339                                         return retval;
1340                         }
1341                 }
1342         }
1343         return 0;
1344 }
1345
1346 /*
1347  * run func(cpu) on every cpu in /proc/stat
1348  * return max_cpu number
1349  */
1350 int for_all_proc_cpus(int (func)(int))
1351 {
1352         FILE *fp;
1353         int cpu_num;
1354         int retval;
1355
1356         fp = fopen_or_die(proc_stat, "r");
1357
1358         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1359         if (retval != 0)
1360                 err(1, "%s: failed to parse format", proc_stat);
1361
1362         while (1) {
1363                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1364                 if (retval != 1)
1365                         break;
1366
1367                 retval = func(cpu_num);
1368                 if (retval) {
1369                         fclose(fp);
1370                         return(retval);
1371                 }
1372         }
1373         fclose(fp);
1374         return 0;
1375 }
1376
1377 void re_initialize(void)
1378 {
1379         free_all_buffers();
1380         setup_all_buffers();
1381         printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1382 }
1383
1384
1385 /*
1386  * count_cpus()
1387  * remember the last one seen, it will be the max
1388  */
1389 int count_cpus(int cpu)
1390 {
1391         if (topo.max_cpu_num < cpu)
1392                 topo.max_cpu_num = cpu;
1393
1394         topo.num_cpus += 1;
1395         return 0;
1396 }
1397 int mark_cpu_present(int cpu)
1398 {
1399         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1400         return 0;
1401 }
1402
1403 void turbostat_loop()
1404 {
1405         int retval;
1406         int restarted = 0;
1407
1408 restart:
1409         restarted++;
1410
1411         retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1412         if (retval < -1) {
1413                 exit(retval);
1414         } else if (retval == -1) {
1415                 if (restarted > 1) {
1416                         exit(retval);
1417                 }
1418                 re_initialize();
1419                 goto restart;
1420         }
1421         restarted = 0;
1422         gettimeofday(&tv_even, (struct timezone *)NULL);
1423
1424         while (1) {
1425                 if (for_all_proc_cpus(cpu_is_not_present)) {
1426                         re_initialize();
1427                         goto restart;
1428                 }
1429                 sleep(interval_sec);
1430                 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1431                 if (retval < -1) {
1432                         exit(retval);
1433                 } else if (retval == -1) {
1434                         re_initialize();
1435                         goto restart;
1436                 }
1437                 gettimeofday(&tv_odd, (struct timezone *)NULL);
1438                 timersub(&tv_odd, &tv_even, &tv_delta);
1439                 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1440                 compute_average(EVEN_COUNTERS);
1441                 format_all_counters(EVEN_COUNTERS);
1442                 flush_stdout();
1443                 sleep(interval_sec);
1444                 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1445                 if (retval < -1) {
1446                         exit(retval);
1447                 } else if (retval == -1) {
1448                         re_initialize();
1449                         goto restart;
1450                 }
1451                 gettimeofday(&tv_even, (struct timezone *)NULL);
1452                 timersub(&tv_even, &tv_odd, &tv_delta);
1453                 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1454                 compute_average(ODD_COUNTERS);
1455                 format_all_counters(ODD_COUNTERS);
1456                 flush_stdout();
1457         }
1458 }
1459
1460 void check_dev_msr()
1461 {
1462         struct stat sb;
1463
1464         if (stat("/dev/cpu/0/msr", &sb))
1465                 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
1466 }
1467
1468 void check_permissions()
1469 {
1470         struct __user_cap_header_struct cap_header_data;
1471         cap_user_header_t cap_header = &cap_header_data;
1472         struct __user_cap_data_struct cap_data_data;
1473         cap_user_data_t cap_data = &cap_data_data;
1474         extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
1475         int do_exit = 0;
1476
1477         /* check for CAP_SYS_RAWIO */
1478         cap_header->pid = getpid();
1479         cap_header->version = _LINUX_CAPABILITY_VERSION;
1480         if (capget(cap_header, cap_data) < 0)
1481                 err(-6, "capget(2) failed");
1482
1483         if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1484                 do_exit++;
1485                 warnx("capget(CAP_SYS_RAWIO) failed,"
1486                         " try \"# setcap cap_sys_rawio=ep %s\"", progname);
1487         }
1488
1489         /* test file permissions */
1490         if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1491                 do_exit++;
1492                 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1493         }
1494
1495         /* if all else fails, thell them to be root */
1496         if (do_exit)
1497                 if (getuid() != 0)
1498                         warnx("... or simply run as root");
1499
1500         if (do_exit)
1501                 exit(-6);
1502 }
1503
1504 /*
1505  * NHM adds support for additional MSRs:
1506  *
1507  * MSR_SMI_COUNT                   0x00000034
1508  *
1509  * MSR_NHM_PLATFORM_INFO           0x000000ce
1510  * MSR_NHM_SNB_PKG_CST_CFG_CTL     0x000000e2
1511  *
1512  * MSR_PKG_C3_RESIDENCY            0x000003f8
1513  * MSR_PKG_C6_RESIDENCY            0x000003f9
1514  * MSR_CORE_C3_RESIDENCY           0x000003fc
1515  * MSR_CORE_C6_RESIDENCY           0x000003fd
1516  *
1517  */
1518 int has_nhm_msrs(unsigned int family, unsigned int model)
1519 {
1520         if (!genuine_intel)
1521                 return 0;
1522
1523         if (family != 6)
1524                 return 0;
1525
1526         switch (model) {
1527         case 0x1A:      /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1528         case 0x1E:      /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1529         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
1530         case 0x25:      /* Westmere Client - Clarkdale, Arrandale */
1531         case 0x2C:      /* Westmere EP - Gulftown */
1532         case 0x2A:      /* SNB */
1533         case 0x2D:      /* SNB Xeon */
1534         case 0x3A:      /* IVB */
1535         case 0x3E:      /* IVB Xeon */
1536         case 0x3C:      /* HSW */
1537         case 0x3F:      /* HSX */
1538         case 0x45:      /* HSW */
1539         case 0x46:      /* HSW */
1540         case 0x37:      /* BYT */
1541         case 0x4D:      /* AVN */
1542         case 0x3D:      /* BDW */
1543         case 0x4F:      /* BDX */
1544         case 0x56:      /* BDX-DE */
1545         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1546         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1547                 return 1;
1548         default:
1549                 return 0;
1550         }
1551 }
1552 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
1553 {
1554         if (!has_nhm_msrs(family, model))
1555                 return 0;
1556
1557         switch (model) {
1558         /* Nehalem compatible, but do not include turbo-ratio limit support */
1559         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1560         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1561                 return 0;
1562         default:
1563                 return 1;
1564         }
1565 }
1566 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1567 {
1568         if (!genuine_intel)
1569                 return 0;
1570
1571         if (family != 6)
1572                 return 0;
1573
1574         switch (model) {
1575         case 0x3E:      /* IVB Xeon */
1576                 return 1;
1577         default:
1578                 return 0;
1579         }
1580 }
1581
1582 /*
1583  * print_epb()
1584  * Decode the ENERGY_PERF_BIAS MSR
1585  */
1586 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1587 {
1588         unsigned long long msr;
1589         char *epb_string;
1590         int cpu;
1591
1592         if (!has_epb)
1593                 return 0;
1594
1595         cpu = t->cpu_id;
1596
1597         /* EPB is per-package */
1598         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1599                 return 0;
1600
1601         if (cpu_migrate(cpu)) {
1602                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1603                 return -1;
1604         }
1605
1606         if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
1607                 return 0;
1608
1609         switch (msr & 0x7) {
1610         case ENERGY_PERF_BIAS_PERFORMANCE:
1611                 epb_string = "performance";
1612                 break;
1613         case ENERGY_PERF_BIAS_NORMAL:
1614                 epb_string = "balanced";
1615                 break;
1616         case ENERGY_PERF_BIAS_POWERSAVE:
1617                 epb_string = "powersave";
1618                 break;
1619         default:
1620                 epb_string = "custom";
1621                 break;
1622         }
1623         fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
1624
1625         return 0;
1626 }
1627
1628 /*
1629  * print_perf_limit()
1630  */
1631 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1632 {
1633         unsigned long long msr;
1634         int cpu;
1635
1636         cpu = t->cpu_id;
1637
1638         /* per-package */
1639         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1640                 return 0;
1641
1642         if (cpu_migrate(cpu)) {
1643                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1644                 return -1;
1645         }
1646
1647         if (do_core_perf_limit_reasons) {
1648                 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
1649                 fprintf(stderr, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1650                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
1651                         (msr & 1 << 0) ? "PROCHOT, " : "",
1652                         (msr & 1 << 1) ? "ThermStatus, " : "",
1653                         (msr & 1 << 2) ? "bit2, " : "",
1654                         (msr & 1 << 4) ? "Graphics, " : "",
1655                         (msr & 1 << 5) ? "Auto-HWP, " : "",
1656                         (msr & 1 << 6) ? "VR-Therm, " : "",
1657                         (msr & 1 << 8) ? "Amps, " : "",
1658                         (msr & 1 << 9) ? "CorePwr, " : "",
1659                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1660                         (msr & 1 << 11) ? "PkgPwrL2, " : "",
1661                         (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
1662                         (msr & 1 << 13) ? "Transitions, " : "",
1663                         (msr & 1 << 14) ? "bit14, " : "",
1664                         (msr & 1 << 15) ? "bit15, " : "");
1665                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
1666                         (msr & 1 << 16) ? "PROCHOT, " : "",
1667                         (msr & 1 << 17) ? "ThermStatus, " : "",
1668                         (msr & 1 << 18) ? "bit18, " : "",
1669                         (msr & 1 << 20) ? "Graphics, " : "",
1670                         (msr & 1 << 21) ? "Auto-HWP, " : "",
1671                         (msr & 1 << 22) ? "VR-Therm, " : "",
1672                         (msr & 1 << 24) ? "Amps, " : "",
1673                         (msr & 1 << 25) ? "CorePwr, " : "",
1674                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1675                         (msr & 1 << 27) ? "PkgPwrL2, " : "",
1676                         (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
1677                         (msr & 1 << 29) ? "Transitions, " : "",
1678                         (msr & 1 << 30) ? "bit30, " : "",
1679                         (msr & 1 << 31) ? "bit31, " : "");
1680
1681         }
1682         if (do_gfx_perf_limit_reasons) {
1683                 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
1684                 fprintf(stderr, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1685                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s)",
1686                         (msr & 1 << 0) ? "PROCHOT, " : "",
1687                         (msr & 1 << 1) ? "ThermStatus, " : "",
1688                         (msr & 1 << 4) ? "Graphics, " : "",
1689                         (msr & 1 << 6) ? "VR-Therm, " : "",
1690                         (msr & 1 << 8) ? "Amps, " : "",
1691                         (msr & 1 << 9) ? "GFXPwr, " : "",
1692                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1693                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
1694                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s)\n",
1695                         (msr & 1 << 16) ? "PROCHOT, " : "",
1696                         (msr & 1 << 17) ? "ThermStatus, " : "",
1697                         (msr & 1 << 20) ? "Graphics, " : "",
1698                         (msr & 1 << 22) ? "VR-Therm, " : "",
1699                         (msr & 1 << 24) ? "Amps, " : "",
1700                         (msr & 1 << 25) ? "GFXPwr, " : "",
1701                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1702                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
1703         }
1704         if (do_ring_perf_limit_reasons) {
1705                 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
1706                 fprintf(stderr, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1707                 fprintf(stderr, " (Active: %s%s%s%s%s%s)",
1708                         (msr & 1 << 0) ? "PROCHOT, " : "",
1709                         (msr & 1 << 1) ? "ThermStatus, " : "",
1710                         (msr & 1 << 6) ? "VR-Therm, " : "",
1711                         (msr & 1 << 8) ? "Amps, " : "",
1712                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1713                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
1714                 fprintf(stderr, " (Logged: %s%s%s%s%s%s)\n",
1715                         (msr & 1 << 16) ? "PROCHOT, " : "",
1716                         (msr & 1 << 17) ? "ThermStatus, " : "",
1717                         (msr & 1 << 22) ? "VR-Therm, " : "",
1718                         (msr & 1 << 24) ? "Amps, " : "",
1719                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1720                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
1721         }
1722         return 0;
1723 }
1724
1725 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
1726 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
1727
1728 double get_tdp(model)
1729 {
1730         unsigned long long msr;
1731
1732         if (do_rapl & RAPL_PKG_POWER_INFO)
1733                 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
1734                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
1735
1736         switch (model) {
1737         case 0x37:
1738         case 0x4D:
1739                 return 30.0;
1740         default:
1741                 return 135.0;
1742         }
1743 }
1744
1745
1746 /*
1747  * rapl_probe()
1748  *
1749  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
1750  */
1751 void rapl_probe(unsigned int family, unsigned int model)
1752 {
1753         unsigned long long msr;
1754         unsigned int time_unit;
1755         double tdp;
1756
1757         if (!genuine_intel)
1758                 return;
1759
1760         if (family != 6)
1761                 return;
1762
1763         switch (model) {
1764         case 0x2A:
1765         case 0x3A:
1766         case 0x3C:      /* HSW */
1767         case 0x45:      /* HSW */
1768         case 0x46:      /* HSW */
1769         case 0x3D:      /* BDW */
1770                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1771                 break;
1772         case 0x3F:      /* HSX */
1773         case 0x4F:      /* BDX */
1774         case 0x56:      /* BDX-DE */
1775                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1776                 break;
1777         case 0x2D:
1778         case 0x3E:
1779                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1780                 break;
1781         case 0x37:      /* BYT */
1782         case 0x4D:      /* AVN */
1783                 do_rapl = RAPL_PKG | RAPL_CORES ;
1784                 break;
1785         default:
1786                 return;
1787         }
1788
1789         /* units on package 0, verify later other packages match */
1790         if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1791                 return;
1792
1793         rapl_power_units = 1.0 / (1 << (msr & 0xF));
1794         if (model == 0x37)
1795                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1796         else
1797                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1798
1799         time_unit = msr >> 16 & 0xF;
1800         if (time_unit == 0)
1801                 time_unit = 0xA;
1802
1803         rapl_time_units = 1.0 / (1 << (time_unit));
1804
1805         tdp = get_tdp(model);
1806
1807         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1808         if (verbose)
1809                 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1810
1811         return;
1812 }
1813
1814 void perf_limit_reasons_probe(family, model)
1815 {
1816         if (!genuine_intel)
1817                 return;
1818
1819         if (family != 6)
1820                 return;
1821
1822         switch (model) {
1823         case 0x3C:      /* HSW */
1824         case 0x45:      /* HSW */
1825         case 0x46:      /* HSW */
1826                 do_gfx_perf_limit_reasons = 1;
1827         case 0x3F:      /* HSX */
1828                 do_core_perf_limit_reasons = 1;
1829                 do_ring_perf_limit_reasons = 1;
1830         default:
1831                 return;
1832         }
1833 }
1834
1835 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1836 {
1837         unsigned long long msr;
1838         unsigned int dts;
1839         int cpu;
1840
1841         if (!(do_dts || do_ptm))
1842                 return 0;
1843
1844         cpu = t->cpu_id;
1845
1846         /* DTS is per-core, no need to print for each thread */
1847         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1848                 return 0;
1849
1850         if (cpu_migrate(cpu)) {
1851                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1852                 return -1;
1853         }
1854
1855         if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
1856                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1857                         return 0;
1858
1859                 dts = (msr >> 16) & 0x7F;
1860                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
1861                         cpu, msr, tcc_activation_temp - dts);
1862
1863 #ifdef  THERM_DEBUG
1864                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
1865                         return 0;
1866
1867                 dts = (msr >> 16) & 0x7F;
1868                 dts2 = (msr >> 8) & 0x7F;
1869                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1870                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1871 #endif
1872         }
1873
1874
1875         if (do_dts) {
1876                 unsigned int resolution;
1877
1878                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1879                         return 0;
1880
1881                 dts = (msr >> 16) & 0x7F;
1882                 resolution = (msr >> 27) & 0xF;
1883                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
1884                         cpu, msr, tcc_activation_temp - dts, resolution);
1885
1886 #ifdef THERM_DEBUG
1887                 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
1888                         return 0;
1889
1890                 dts = (msr >> 16) & 0x7F;
1891                 dts2 = (msr >> 8) & 0x7F;
1892                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1893                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1894 #endif
1895         }
1896
1897         return 0;
1898 }
1899
1900 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
1901 {
1902         fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
1903                 cpu, label,
1904                 ((msr >> 15) & 1) ? "EN" : "DIS",
1905                 ((msr >> 0) & 0x7FFF) * rapl_power_units,
1906                 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
1907                 (((msr >> 16) & 1) ? "EN" : "DIS"));
1908
1909         return;
1910 }
1911
1912 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1913 {
1914         unsigned long long msr;
1915         int cpu;
1916
1917         if (!do_rapl)
1918                 return 0;
1919
1920         /* RAPL counters are per package, so print only for 1st thread/package */
1921         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1922                 return 0;
1923
1924         cpu = t->cpu_id;
1925         if (cpu_migrate(cpu)) {
1926                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1927                 return -1;
1928         }
1929
1930         if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
1931                 return -1;
1932
1933         if (verbose) {
1934                 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
1935                         "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
1936                         rapl_power_units, rapl_energy_units, rapl_time_units);
1937         }
1938         if (do_rapl & RAPL_PKG_POWER_INFO) {
1939
1940                 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
1941                         return -5;
1942
1943
1944                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1945                         cpu, msr,
1946                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1947                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1948                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1949                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1950
1951         }
1952         if (do_rapl & RAPL_PKG) {
1953
1954                 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
1955                         return -9;
1956
1957                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
1958                         cpu, msr, (msr >> 63) & 1 ? "": "UN");
1959
1960                 print_power_limit_msr(cpu, msr, "PKG Limit #1");
1961                 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
1962                         cpu,
1963                         ((msr >> 47) & 1) ? "EN" : "DIS",
1964                         ((msr >> 32) & 0x7FFF) * rapl_power_units,
1965                         (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
1966                         ((msr >> 48) & 1) ? "EN" : "DIS");
1967         }
1968
1969         if (do_rapl & RAPL_DRAM) {
1970                 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
1971                         return -6;
1972
1973
1974                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1975                         cpu, msr,
1976                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1977                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1978                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1979                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1980
1981
1982                 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
1983                         return -9;
1984                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
1985                                 cpu, msr, (msr >> 31) & 1 ? "": "UN");
1986
1987                 print_power_limit_msr(cpu, msr, "DRAM Limit");
1988         }
1989         if (do_rapl & RAPL_CORE_POLICY) {
1990                 if (verbose) {
1991                         if (get_msr(cpu, MSR_PP0_POLICY, &msr))
1992                                 return -7;
1993
1994                         fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
1995                 }
1996         }
1997         if (do_rapl & RAPL_CORES) {
1998                 if (verbose) {
1999
2000                         if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
2001                                 return -9;
2002                         fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2003                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2004                         print_power_limit_msr(cpu, msr, "Cores Limit");
2005                 }
2006         }
2007         if (do_rapl & RAPL_GFX) {
2008                 if (verbose) {
2009                         if (get_msr(cpu, MSR_PP1_POLICY, &msr))
2010                                 return -8;
2011
2012                         fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
2013
2014                         if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
2015                                 return -9;
2016                         fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2017                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2018                         print_power_limit_msr(cpu, msr, "GFX Limit");
2019                 }
2020         }
2021         return 0;
2022 }
2023
2024 /*
2025  * SNB adds support for additional MSRs:
2026  *
2027  * MSR_PKG_C7_RESIDENCY            0x000003fa
2028  * MSR_CORE_C7_RESIDENCY           0x000003fe
2029  * MSR_PKG_C2_RESIDENCY            0x0000060d
2030  */
2031
2032 int has_snb_msrs(unsigned int family, unsigned int model)
2033 {
2034         if (!genuine_intel)
2035                 return 0;
2036
2037         switch (model) {
2038         case 0x2A:
2039         case 0x2D:
2040         case 0x3A:      /* IVB */
2041         case 0x3E:      /* IVB Xeon */
2042         case 0x3C:      /* HSW */
2043         case 0x3F:      /* HSW */
2044         case 0x45:      /* HSW */
2045         case 0x46:      /* HSW */
2046         case 0x3D:      /* BDW */
2047         case 0x4F:      /* BDX */
2048         case 0x56:      /* BDX-DE */
2049                 return 1;
2050         }
2051         return 0;
2052 }
2053
2054 /*
2055  * HSW adds support for additional MSRs:
2056  *
2057  * MSR_PKG_C8_RESIDENCY            0x00000630
2058  * MSR_PKG_C9_RESIDENCY            0x00000631
2059  * MSR_PKG_C10_RESIDENCY           0x00000632
2060  */
2061 int has_hsw_msrs(unsigned int family, unsigned int model)
2062 {
2063         if (!genuine_intel)
2064                 return 0;
2065
2066         switch (model) {
2067         case 0x45:      /* HSW */
2068         case 0x3D:      /* BDW */
2069                 return 1;
2070         }
2071         return 0;
2072 }
2073
2074
2075 int is_slm(unsigned int family, unsigned int model)
2076 {
2077         if (!genuine_intel)
2078                 return 0;
2079         switch (model) {
2080         case 0x37:      /* BYT */
2081         case 0x4D:      /* AVN */
2082                 return 1;
2083         }
2084         return 0;
2085 }
2086
2087 #define SLM_BCLK_FREQS 5
2088 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
2089
2090 double slm_bclk(void)
2091 {
2092         unsigned long long msr = 3;
2093         unsigned int i;
2094         double freq;
2095
2096         if (get_msr(0, MSR_FSB_FREQ, &msr))
2097                 fprintf(stderr, "SLM BCLK: unknown\n");
2098
2099         i = msr & 0xf;
2100         if (i >= SLM_BCLK_FREQS) {
2101                 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
2102                 msr = 3;
2103         }
2104         freq = slm_freq_table[i];
2105
2106         fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
2107
2108         return freq;
2109 }
2110
2111 double discover_bclk(unsigned int family, unsigned int model)
2112 {
2113         if (has_snb_msrs(family, model))
2114                 return 100.00;
2115         else if (is_slm(family, model))
2116                 return slm_bclk();
2117         else
2118                 return 133.33;
2119 }
2120
2121 /*
2122  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
2123  * the Thermal Control Circuit (TCC) activates.
2124  * This is usually equal to tjMax.
2125  *
2126  * Older processors do not have this MSR, so there we guess,
2127  * but also allow cmdline over-ride with -T.
2128  *
2129  * Several MSR temperature values are in units of degrees-C
2130  * below this value, including the Digital Thermal Sensor (DTS),
2131  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
2132  */
2133 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2134 {
2135         unsigned long long msr;
2136         unsigned int target_c_local;
2137         int cpu;
2138
2139         /* tcc_activation_temp is used only for dts or ptm */
2140         if (!(do_dts || do_ptm))
2141                 return 0;
2142
2143         /* this is a per-package concept */
2144         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2145                 return 0;
2146
2147         cpu = t->cpu_id;
2148         if (cpu_migrate(cpu)) {
2149                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2150                 return -1;
2151         }
2152
2153         if (tcc_activation_temp_override != 0) {
2154                 tcc_activation_temp = tcc_activation_temp_override;
2155                 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
2156                         cpu, tcc_activation_temp);
2157                 return 0;
2158         }
2159
2160         /* Temperature Target MSR is Nehalem and newer only */
2161         if (!do_nhm_platform_info)
2162                 goto guess;
2163
2164         if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
2165                 goto guess;
2166
2167         target_c_local = (msr >> 16) & 0xFF;
2168
2169         if (verbose)
2170                 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2171                         cpu, msr, target_c_local);
2172
2173         if (!target_c_local)
2174                 goto guess;
2175
2176         tcc_activation_temp = target_c_local;
2177
2178         return 0;
2179
2180 guess:
2181         tcc_activation_temp = TJMAX_DEFAULT;
2182         fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2183                 cpu, tcc_activation_temp);
2184
2185         return 0;
2186 }
2187 void check_cpuid()
2188 {
2189         unsigned int eax, ebx, ecx, edx, max_level;
2190         unsigned int fms, family, model, stepping;
2191
2192         eax = ebx = ecx = edx = 0;
2193
2194         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2195
2196         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2197                 genuine_intel = 1;
2198
2199         if (verbose)
2200                 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2201                         (char *)&ebx, (char *)&edx, (char *)&ecx);
2202
2203         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2204         family = (fms >> 8) & 0xf;
2205         model = (fms >> 4) & 0xf;
2206         stepping = fms & 0xf;
2207         if (family == 6 || family == 0xf)
2208                 model += ((fms >> 16) & 0xf) << 4;
2209
2210         if (verbose)
2211                 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2212                         max_level, family, model, stepping, family, model, stepping);
2213
2214         if (!(edx & (1 << 5)))
2215                 errx(1, "CPUID: no MSR");
2216
2217         /*
2218          * check max extended function levels of CPUID.
2219          * This is needed to check for invariant TSC.
2220          * This check is valid for both Intel and AMD.
2221          */
2222         ebx = ecx = edx = 0;
2223         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2224
2225         if (max_level >= 0x80000007) {
2226
2227                 /*
2228                  * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2229                  * this check is valid for both Intel and AMD
2230                  */
2231                 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2232                 has_invariant_tsc = edx & (1 << 8);
2233         }
2234
2235         /*
2236          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2237          * this check is valid for both Intel and AMD
2238          */
2239
2240         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2241         has_aperf = ecx & (1 << 0);
2242         do_dts = eax & (1 << 0);
2243         do_ptm = eax & (1 << 6);
2244         has_epb = ecx & (1 << 3);
2245
2246         if (verbose)
2247                 fprintf(stderr, "CPUID(6): %s%s%s%s\n",
2248                         has_aperf ? "APERF" : "No APERF!",
2249                         do_dts ? ", DTS" : "",
2250                         do_ptm ? ", PTM": "",
2251                         has_epb ? ", EPB": "");
2252
2253         if (!has_aperf)
2254                 errx(-1, "No APERF");
2255
2256         do_nhm_platform_info = do_nhm_cstates = do_smi = has_nhm_msrs(family, model);
2257         do_snb_cstates = has_snb_msrs(family, model);
2258         do_c8_c9_c10 = has_hsw_msrs(family, model);
2259         do_slm_cstates = is_slm(family, model);
2260         bclk = discover_bclk(family, model);
2261
2262         do_nhm_turbo_ratio_limit = has_nhm_turbo_ratio_limit(family, model);
2263         do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model);
2264         rapl_probe(family, model);
2265         perf_limit_reasons_probe(family, model);
2266
2267         return;
2268 }
2269
2270
2271 void usage()
2272 {
2273         errx(1, "%s: [-v][-R][-T][-p|-P|-S][-c MSR#][-C MSR#][-m MSR#][-M MSR#][-i interval_sec | command ...]\n",
2274              progname);
2275 }
2276
2277
2278 /*
2279  * in /dev/cpu/ return success for names that are numbers
2280  * ie. filter out ".", "..", "microcode".
2281  */
2282 int dir_filter(const struct dirent *dirp)
2283 {
2284         if (isdigit(dirp->d_name[0]))
2285                 return 1;
2286         else
2287                 return 0;
2288 }
2289
2290 int open_dev_cpu_msr(int dummy1)
2291 {
2292         return 0;
2293 }
2294
2295 void topology_probe()
2296 {
2297         int i;
2298         int max_core_id = 0;
2299         int max_package_id = 0;
2300         int max_siblings = 0;
2301         struct cpu_topology {
2302                 int core_id;
2303                 int physical_package_id;
2304         } *cpus;
2305
2306         /* Initialize num_cpus, max_cpu_num */
2307         topo.num_cpus = 0;
2308         topo.max_cpu_num = 0;
2309         for_all_proc_cpus(count_cpus);
2310         if (!summary_only && topo.num_cpus > 1)
2311                 show_cpu = 1;
2312
2313         if (verbose > 1)
2314                 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2315
2316         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
2317         if (cpus == NULL)
2318                 err(1, "calloc cpus");
2319
2320         /*
2321          * Allocate and initialize cpu_present_set
2322          */
2323         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2324         if (cpu_present_set == NULL)
2325                 err(3, "CPU_ALLOC");
2326         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2327         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2328         for_all_proc_cpus(mark_cpu_present);
2329
2330         /*
2331          * Allocate and initialize cpu_affinity_set
2332          */
2333         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2334         if (cpu_affinity_set == NULL)
2335                 err(3, "CPU_ALLOC");
2336         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2337         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2338
2339
2340         /*
2341          * For online cpus
2342          * find max_core_id, max_package_id
2343          */
2344         for (i = 0; i <= topo.max_cpu_num; ++i) {
2345                 int siblings;
2346
2347                 if (cpu_is_not_present(i)) {
2348                         if (verbose > 1)
2349                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2350                         continue;
2351                 }
2352                 cpus[i].core_id = get_core_id(i);
2353                 if (cpus[i].core_id > max_core_id)
2354                         max_core_id = cpus[i].core_id;
2355
2356                 cpus[i].physical_package_id = get_physical_package_id(i);
2357                 if (cpus[i].physical_package_id > max_package_id)
2358                         max_package_id = cpus[i].physical_package_id;
2359
2360                 siblings = get_num_ht_siblings(i);
2361                 if (siblings > max_siblings)
2362                         max_siblings = siblings;
2363                 if (verbose > 1)
2364                         fprintf(stderr, "cpu %d pkg %d core %d\n",
2365                                 i, cpus[i].physical_package_id, cpus[i].core_id);
2366         }
2367         topo.num_cores_per_pkg = max_core_id + 1;
2368         if (verbose > 1)
2369                 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2370                         max_core_id, topo.num_cores_per_pkg);
2371         if (!summary_only && topo.num_cores_per_pkg > 1)
2372                 show_core = 1;
2373
2374         topo.num_packages = max_package_id + 1;
2375         if (verbose > 1)
2376                 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2377                         max_package_id, topo.num_packages);
2378         if (!summary_only && topo.num_packages > 1)
2379                 show_pkg = 1;
2380
2381         topo.num_threads_per_core = max_siblings;
2382         if (verbose > 1)
2383                 fprintf(stderr, "max_siblings %d\n", max_siblings);
2384
2385         free(cpus);
2386 }
2387
2388 void
2389 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2390 {
2391         int i;
2392
2393         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2394                 topo.num_packages, sizeof(struct thread_data));
2395         if (*t == NULL)
2396                 goto error;
2397
2398         for (i = 0; i < topo.num_threads_per_core *
2399                 topo.num_cores_per_pkg * topo.num_packages; i++)
2400                 (*t)[i].cpu_id = -1;
2401
2402         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2403                 sizeof(struct core_data));
2404         if (*c == NULL)
2405                 goto error;
2406
2407         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2408                 (*c)[i].core_id = -1;
2409
2410         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2411         if (*p == NULL)
2412                 goto error;
2413
2414         for (i = 0; i < topo.num_packages; i++)
2415                 (*p)[i].package_id = i;
2416
2417         return;
2418 error:
2419         err(1, "calloc counters");
2420 }
2421 /*
2422  * init_counter()
2423  *
2424  * set cpu_id, core_num, pkg_num
2425  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2426  *
2427  * increment topo.num_cores when 1st core in pkg seen
2428  */
2429 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2430         struct pkg_data *pkg_base, int thread_num, int core_num,
2431         int pkg_num, int cpu_id)
2432 {
2433         struct thread_data *t;
2434         struct core_data *c;
2435         struct pkg_data *p;
2436
2437         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2438         c = GET_CORE(core_base, core_num, pkg_num);
2439         p = GET_PKG(pkg_base, pkg_num);
2440
2441         t->cpu_id = cpu_id;
2442         if (thread_num == 0) {
2443                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2444                 if (cpu_is_first_core_in_package(cpu_id))
2445                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2446         }
2447
2448         c->core_id = core_num;
2449         p->package_id = pkg_num;
2450 }
2451
2452
2453 int initialize_counters(int cpu_id)
2454 {
2455         int my_thread_id, my_core_id, my_package_id;
2456
2457         my_package_id = get_physical_package_id(cpu_id);
2458         my_core_id = get_core_id(cpu_id);
2459
2460         if (cpu_is_first_sibling_in_core(cpu_id)) {
2461                 my_thread_id = 0;
2462                 topo.num_cores++;
2463         } else {
2464                 my_thread_id = 1;
2465         }
2466
2467         init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2468         init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2469         return 0;
2470 }
2471
2472 void allocate_output_buffer()
2473 {
2474         output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
2475         outp = output_buffer;
2476         if (outp == NULL)
2477                 err(-1, "calloc output buffer");
2478 }
2479
2480 void setup_all_buffers(void)
2481 {
2482         topology_probe();
2483         allocate_counters(&thread_even, &core_even, &package_even);
2484         allocate_counters(&thread_odd, &core_odd, &package_odd);
2485         allocate_output_buffer();
2486         for_all_proc_cpus(initialize_counters);
2487 }
2488
2489 void turbostat_init()
2490 {
2491         check_dev_msr();
2492         check_permissions();
2493         check_cpuid();
2494
2495         setup_all_buffers();
2496
2497         if (verbose)
2498                 print_verbose_header();
2499
2500         if (verbose)
2501                 for_all_cpus(print_epb, ODD_COUNTERS);
2502
2503         if (verbose)
2504                 for_all_cpus(print_perf_limit, ODD_COUNTERS);
2505
2506         if (verbose)
2507                 for_all_cpus(print_rapl, ODD_COUNTERS);
2508
2509         for_all_cpus(set_temperature_target, ODD_COUNTERS);
2510
2511         if (verbose)
2512                 for_all_cpus(print_thermal, ODD_COUNTERS);
2513 }
2514
2515 int fork_it(char **argv)
2516 {
2517         pid_t child_pid;
2518         int status;
2519
2520         status = for_all_cpus(get_counters, EVEN_COUNTERS);
2521         if (status)
2522                 exit(status);
2523         /* clear affinity side-effect of get_counters() */
2524         sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
2525         gettimeofday(&tv_even, (struct timezone *)NULL);
2526
2527         child_pid = fork();
2528         if (!child_pid) {
2529                 /* child */
2530                 execvp(argv[0], argv);
2531         } else {
2532
2533                 /* parent */
2534                 if (child_pid == -1)
2535                         err(1, "fork");
2536
2537                 signal(SIGINT, SIG_IGN);
2538                 signal(SIGQUIT, SIG_IGN);
2539                 if (waitpid(child_pid, &status, 0) == -1)
2540                         err(status, "waitpid");
2541         }
2542         /*
2543          * n.b. fork_it() does not check for errors from for_all_cpus()
2544          * because re-starting is problematic when forking
2545          */
2546         for_all_cpus(get_counters, ODD_COUNTERS);
2547         gettimeofday(&tv_odd, (struct timezone *)NULL);
2548         timersub(&tv_odd, &tv_even, &tv_delta);
2549         for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
2550         compute_average(EVEN_COUNTERS);
2551         format_all_counters(EVEN_COUNTERS);
2552         flush_stderr();
2553
2554         fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
2555
2556         return status;
2557 }
2558
2559 int get_and_dump_counters(void)
2560 {
2561         int status;
2562
2563         status = for_all_cpus(get_counters, ODD_COUNTERS);
2564         if (status)
2565                 return status;
2566
2567         status = for_all_cpus(dump_counters, ODD_COUNTERS);
2568         if (status)
2569                 return status;
2570
2571         flush_stdout();
2572
2573         return status;
2574 }
2575
2576 void cmdline(int argc, char **argv)
2577 {
2578         int opt;
2579
2580         progname = argv[0];
2581
2582         while ((opt = getopt(argc, argv, "+pPsSvi:c:C:m:M:RJT:")) != -1) {
2583                 switch (opt) {
2584                 case 'p':
2585                         show_core_only++;
2586                         break;
2587                 case 'P':
2588                         show_pkg_only++;
2589                         break;
2590                 case 's':
2591                         dump_only++;
2592                         break;
2593                 case 'S':
2594                         summary_only++;
2595                         break;
2596                 case 'v':
2597                         verbose++;
2598                         break;
2599                 case 'i':
2600                         interval_sec = atoi(optarg);
2601                         break;
2602                 case 'c':
2603                         sscanf(optarg, "%x", &extra_delta_offset32);
2604                         break;
2605                 case 'C':
2606                         sscanf(optarg, "%x", &extra_delta_offset64);
2607                         break;
2608                 case 'm':
2609                         sscanf(optarg, "%x", &extra_msr_offset32);
2610                         break;
2611                 case 'M':
2612                         sscanf(optarg, "%x", &extra_msr_offset64);
2613                         break;
2614                 case 'R':
2615                         rapl_verbose++;
2616                         break;
2617                 case 'T':
2618                         tcc_activation_temp_override = atoi(optarg);
2619                         break;
2620                 case 'J':
2621                         rapl_joules++;
2622                         break;
2623
2624                 default:
2625                         usage();
2626                 }
2627         }
2628 }
2629
2630 int main(int argc, char **argv)
2631 {
2632         cmdline(argc, argv);
2633
2634         if (verbose)
2635                 fprintf(stderr, "turbostat v3.8 14-Aug 2014"
2636                         " - Len Brown <lenb@kernel.org>\n");
2637
2638         turbostat_init();
2639
2640         /* dump counters and exit */
2641         if (dump_only)
2642                 return get_and_dump_counters();
2643
2644         /*
2645          * if any params left, it must be a command to fork
2646          */
2647         if (argc - optind)
2648                 return fork_it(argv + optind);
2649         else
2650                 turbostat_loop();
2651
2652         return 0;
2653 }