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 (has_aperf) {
 677                 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 678                         old->aperf = new->aperf - old->aperf;
 679                         old->mperf = new->mperf - old->mperf;
 680                 } else {
 681
 682                         if (!aperf_mperf_unstable) {
 683                                 fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
 684                                 fprintf(stderr, "* Frequency results do not cover entire interval *\n");
 685                                 fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
 686
 687                                 aperf_mperf_unstable = 1;
 688                         }
 689                         /*
 690                          * mperf delta is likely a huge "positive" number
 691                          * can not use it for calculating c0 time
 692                          */
 693                         skip_c0 = 1;
 694                         skip_c1 = 1;
 695                 }
 696         }
 697
 698
 699         if (use_c1_residency_msr) {
 700                 /*
 701                  * Some models have a dedicated C1 residency MSR,
 702                  * which should be more accurate than the derivation below.
 703                  */
 704         } else {
 705                 /*
 706                  * As counter collection is not atomic,
 707                  * it is possible for mperf's non-halted cycles + idle states
 708                  * to exceed TSC's all cycles: show c1 = 0% in that case.
 709                  */
 710                 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 711                         old->c1 = 0;
 712                 else {
 713                         /* normal case, derive c1 */
 714                         old->c1 = old->tsc - old->mperf - core_delta->c3
 715                                 - core_delta->c6 - core_delta->c7;
 716                 }
 717         }
 718
 719         if (old->mperf == 0) {
 720                 if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
 721                 old->mperf = 1; /* divide by 0 protection */
 722         }
 723
 724         old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
 725         old->extra_delta32 &= 0xFFFFFFFF;
 726
 727         old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
 728
 729         /*
 730          * Extra MSR is just a snapshot, simply copy latest w/o subtracting
 731          */
 732         old->extra_msr32 = new->extra_msr32;
 733         old->extra_msr64 = new->extra_msr64;
 734
 735         if (do_smi)
 736                 old->smi_count = new->smi_count - old->smi_count;
 737 }
 738
 739 int delta_cpu(struct thread_data *t, struct core_data *c,
 740         struct pkg_data *p, struct thread_data *t2,
 741         struct core_data *c2, struct pkg_data *p2)
 742 {
 743         /* calculate core delta only for 1st thread in core */
 744         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 745                 delta_core(c, c2);
 746
 747         /* always calculate thread delta */
 748         delta_thread(t, t2, c2);        /* c2 is core delta */
 749
 750         /* calculate package delta only for 1st core in package */
 751         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 752                 delta_package(p, p2);
 753
 754         return 0;
 755 }
 756
 757 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 758 {
 759         t->tsc = 0;
 760         t->aperf = 0;
 761         t->mperf = 0;
 762         t->c1 = 0;
 763
 764         t->smi_count = 0;
 765         t->extra_delta32 = 0;
 766         t->extra_delta64 = 0;
 767
 768         /* tells format_counters to dump all fields from this set */
 769         t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
 770
 771         c->c3 = 0;
 772         c->c6 = 0;
 773         c->c7 = 0;
 774         c->core_temp_c = 0;
 775
 776         p->pc2 = 0;
 777         p->pc3 = 0;
 778         p->pc6 = 0;
 779         p->pc7 = 0;
 780         p->pc8 = 0;
 781         p->pc9 = 0;
 782         p->pc10 = 0;
 783
 784         p->energy_pkg = 0;
 785         p->energy_dram = 0;
 786         p->energy_cores = 0;
 787         p->energy_gfx = 0;
 788         p->rapl_pkg_perf_status = 0;
 789         p->rapl_dram_perf_status = 0;
 790         p->pkg_temp_c = 0;
 791 }
 792 int sum_counters(struct thread_data *t, struct core_data *c,
 793         struct pkg_data *p)
 794 {
 795         average.threads.tsc += t->tsc;
 796         average.threads.aperf += t->aperf;
 797         average.threads.mperf += t->mperf;
 798         average.threads.c1 += t->c1;
 799
 800         average.threads.extra_delta32 += t->extra_delta32;
 801         average.threads.extra_delta64 += t->extra_delta64;
 802
 803         /* sum per-core values only for 1st thread in core */
 804         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 805                 return 0;
 806
 807         average.cores.c3 += c->c3;
 808         average.cores.c6 += c->c6;
 809         average.cores.c7 += c->c7;
 810
 811         average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
 812
 813         /* sum per-pkg values only for 1st core in pkg */
 814         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 815                 return 0;
 816
 817         average.packages.pc2 += p->pc2;
 818         average.packages.pc3 += p->pc3;
 819         average.packages.pc6 += p->pc6;
 820         average.packages.pc7 += p->pc7;
 821         average.packages.pc8 += p->pc8;
 822         average.packages.pc9 += p->pc9;
 823         average.packages.pc10 += p->pc10;
 824
 825         average.packages.energy_pkg += p->energy_pkg;
 826         average.packages.energy_dram += p->energy_dram;
 827         average.packages.energy_cores += p->energy_cores;
 828         average.packages.energy_gfx += p->energy_gfx;
 829
 830         average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
 831
 832         average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
 833         average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
 834         return 0;
 835 }
 836 /*
 837  * sum the counters for all cpus in the system
 838  * compute the weighted average
 839  */
 840 void compute_average(struct thread_data *t, struct core_data *c,
 841         struct pkg_data *p)
 842 {
 843         clear_counters(&average.threads, &average.cores, &average.packages);
 844
 845         for_all_cpus(sum_counters, t, c, p);
 846
 847         average.threads.tsc /= topo.num_cpus;
 848         average.threads.aperf /= topo.num_cpus;
 849         average.threads.mperf /= topo.num_cpus;
 850         average.threads.c1 /= topo.num_cpus;
 851
 852         average.threads.extra_delta32 /= topo.num_cpus;
 853         average.threads.extra_delta32 &= 0xFFFFFFFF;
 854
 855         average.threads.extra_delta64 /= topo.num_cpus;
 856
 857         average.cores.c3 /= topo.num_cores;
 858         average.cores.c6 /= topo.num_cores;
 859         average.cores.c7 /= topo.num_cores;
 860
 861         average.packages.pc2 /= topo.num_packages;
 862         average.packages.pc3 /= topo.num_packages;
 863         average.packages.pc6 /= topo.num_packages;
 864         average.packages.pc7 /= topo.num_packages;
 865
 866         average.packages.pc8 /= topo.num_packages;
 867         average.packages.pc9 /= topo.num_packages;
 868         average.packages.pc10 /= topo.num_packages;
 869 }
 870
 871 static unsigned long long rdtsc(void)
 872 {
 873         unsigned int low, high;
 874
 875         asm volatile("rdtsc" : "=a" (low), "=d" (high));
 876
 877         return low | ((unsigned long long)high) << 32;
 878 }
 879
 880
 881 /*
 882  * get_counters(...)
 883  * migrate to cpu
 884  * acquire and record local counters for that cpu
 885  */
 886 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 887 {
 888         int cpu = t->cpu_id;
 889         unsigned long long msr;
 890
 891         if (cpu_migrate(cpu)) {
 892                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
 893                 return -1;
 894         }
 895
 896         t->tsc = rdtsc();       /* we are running on local CPU of interest */
 897
 898         if (has_aperf) {
 899                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
 900                         return -3;
 901                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
 902                         return -4;
 903         }
 904
 905         if (do_smi) {
 906                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
 907                         return -5;
 908                 t->smi_count = msr & 0xFFFFFFFF;
 909         }
 910         if (extra_delta_offset32) {
 911                 if (get_msr(cpu, extra_delta_offset32, &msr))
 912                         return -5;
 913                 t->extra_delta32 = msr & 0xFFFFFFFF;
 914         }
 915
 916         if (extra_delta_offset64)
 917                 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
 918                         return -5;
 919
 920         if (extra_msr_offset32) {
 921                 if (get_msr(cpu, extra_msr_offset32, &msr))
 922                         return -5;
 923                 t->extra_msr32 = msr & 0xFFFFFFFF;
 924         }
 925
 926         if (extra_msr_offset64)
 927                 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
 928                         return -5;
 929
 930         if (use_c1_residency_msr) {
 931                 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
 932                         return -6;
 933         }
 934
 935         /* collect core counters only for 1st thread in core */
 936         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 937                 return 0;
 938
 939         if (do_nhm_cstates && !do_slm_cstates) {
 940                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
 941                         return -6;
 942         }
 943
 944         if (do_nhm_cstates) {
 945                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
 946                         return -7;
 947         }
 948
 949         if (do_snb_cstates)
 950                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
 951                         return -8;
 952
 953         if (do_dts) {
 954                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
 955                         return -9;
 956                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 957         }
 958
 959
 960         /* collect package counters only for 1st core in package */
 961         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 962                 return 0;
 963
 964         if (do_nhm_cstates && !do_slm_cstates) {
 965                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
 966                         return -9;
 967                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
 968                         return -10;
 969         }
 970         if (do_snb_cstates) {
 971                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
 972                         return -11;
 973                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
 974                         return -12;
 975         }
 976         if (do_c8_c9_c10) {
 977                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
 978                         return -13;
 979                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
 980                         return -13;
 981                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
 982                         return -13;
 983         }
 984         if (do_rapl & RAPL_PKG) {
 985                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
 986                         return -13;
 987                 p->energy_pkg = msr & 0xFFFFFFFF;
 988         }
 989         if (do_rapl & RAPL_CORES) {
 990                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
 991                         return -14;
 992                 p->energy_cores = msr & 0xFFFFFFFF;
 993         }
 994         if (do_rapl & RAPL_DRAM) {
 995                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
 996                         return -15;
 997                 p->energy_dram = msr & 0xFFFFFFFF;
 998         }
 999         if (do_rapl & RAPL_GFX) {
1000                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1001                         return -16;
1002                 p->energy_gfx = msr & 0xFFFFFFFF;
1003         }
1004         if (do_rapl & RAPL_PKG_PERF_STATUS) {
1005                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1006                         return -16;
1007                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1008         }
1009         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1010                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1011                         return -16;
1012                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1013         }
1014         if (do_ptm) {
1015                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1016                         return -17;
1017                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1018         }
1019         return 0;
1020 }
1021
1022 void print_verbose_header(void)
1023 {
1024         unsigned long long msr;
1025         unsigned int ratio;
1026
1027         if (!do_nhm_platform_info)
1028                 return;
1029
1030         get_msr(0, MSR_NHM_PLATFORM_INFO, &msr);
1031
1032         fprintf(stderr, "cpu0: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", msr);
1033
1034         ratio = (msr >> 40) & 0xFF;
1035         fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
1036                 ratio, bclk, ratio * bclk);
1037
1038         ratio = (msr >> 8) & 0xFF;
1039         fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
1040                 ratio, bclk, ratio * bclk);
1041
1042         get_msr(0, MSR_IA32_POWER_CTL, &msr);
1043         fprintf(stderr, "cpu0: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1044                 msr, msr & 0x2 ? "EN" : "DIS");
1045
1046         if (!do_ivt_turbo_ratio_limit)
1047                 goto print_nhm_turbo_ratio_limits;
1048
1049         get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr);
1050
1051         fprintf(stderr, "cpu0: MSR_IVT_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1052
1053         ratio = (msr >> 56) & 0xFF;
1054         if (ratio)
1055                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1056                         ratio, bclk, ratio * bclk);
1057
1058         ratio = (msr >> 48) & 0xFF;
1059         if (ratio)
1060                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1061                         ratio, bclk, ratio * bclk);
1062
1063         ratio = (msr >> 40) & 0xFF;
1064         if (ratio)
1065                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1066                         ratio, bclk, ratio * bclk);
1067
1068         ratio = (msr >> 32) & 0xFF;
1069         if (ratio)
1070                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1071                         ratio, bclk, ratio * bclk);
1072
1073         ratio = (msr >> 24) & 0xFF;
1074         if (ratio)
1075                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1076                         ratio, bclk, ratio * bclk);
1077
1078         ratio = (msr >> 16) & 0xFF;
1079         if (ratio)
1080                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1081                         ratio, bclk, ratio * bclk);
1082
1083         ratio = (msr >> 8) & 0xFF;
1084         if (ratio)
1085                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1086                         ratio, bclk, ratio * bclk);
1087
1088         ratio = (msr >> 0) & 0xFF;
1089         if (ratio)
1090                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1091                         ratio, bclk, ratio * bclk);
1092
1093 print_nhm_turbo_ratio_limits:
1094         get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1095
1096 #define SNB_C1_AUTO_UNDEMOTE              (1UL << 27)
1097 #define SNB_C3_AUTO_UNDEMOTE              (1UL << 28)
1098
1099         fprintf(stderr, "cpu0: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", msr);
1100
1101         fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: ",
1102                 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1103                 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1104                 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1105                 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1106                 (msr & (1 << 15)) ? "" : "UN",
1107                 (unsigned int)msr & 7);
1108
1109
1110         switch(msr & 0x7) {
1111         case 0:
1112                 fprintf(stderr, do_slm_cstates ? "no pkg states" : "pc0");
1113                 break;
1114         case 1:
1115                 fprintf(stderr, do_slm_cstates ? "no pkg states" : do_snb_cstates ? "pc2" : "pc0");
1116                 break;
1117         case 2:
1118                 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc6-noret" : "pc3");
1119                 break;
1120         case 3:
1121                 fprintf(stderr, do_slm_cstates ? "invalid" : "pc6");
1122                 break;
1123         case 4:
1124                 fprintf(stderr, do_slm_cstates ? "pc4" : "pc7");
1125                 break;
1126         case 5:
1127                 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc7s" : "invalid");
1128                 break;
1129         case 6:
1130                 fprintf(stderr, do_slm_cstates ? "pc6" : "invalid");
1131                 break;
1132         case 7:
1133                 fprintf(stderr, do_slm_cstates ? "pc7" : "unlimited");
1134                 break;
1135         default:
1136                 fprintf(stderr, "invalid");
1137         }
1138         fprintf(stderr, ")\n");
1139
1140         if (!do_nhm_turbo_ratio_limit)
1141                 return;
1142
1143         get_msr(0, MSR_NHM_TURBO_RATIO_LIMIT, &msr);
1144
1145         fprintf(stderr, "cpu0: MSR_NHM_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1146
1147         ratio = (msr >> 56) & 0xFF;
1148         if (ratio)
1149                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1150                         ratio, bclk, ratio * bclk);
1151
1152         ratio = (msr >> 48) & 0xFF;
1153         if (ratio)
1154                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1155                         ratio, bclk, ratio * bclk);
1156
1157         ratio = (msr >> 40) & 0xFF;
1158         if (ratio)
1159                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1160                         ratio, bclk, ratio * bclk);
1161
1162         ratio = (msr >> 32) & 0xFF;
1163         if (ratio)
1164                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1165                         ratio, bclk, ratio * bclk);
1166
1167         ratio = (msr >> 24) & 0xFF;
1168         if (ratio)
1169                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1170                         ratio, bclk, ratio * bclk);
1171
1172         ratio = (msr >> 16) & 0xFF;
1173         if (ratio)
1174                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1175                         ratio, bclk, ratio * bclk);
1176
1177         ratio = (msr >> 8) & 0xFF;
1178         if (ratio)
1179                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1180                         ratio, bclk, ratio * bclk);
1181
1182         ratio = (msr >> 0) & 0xFF;
1183         if (ratio)
1184                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1185                         ratio, bclk, ratio * bclk);
1186
1187 }
1188
1189 void free_all_buffers(void)
1190 {
1191         CPU_FREE(cpu_present_set);
1192         cpu_present_set = NULL;
1193         cpu_present_set = 0;
1194
1195         CPU_FREE(cpu_affinity_set);
1196         cpu_affinity_set = NULL;
1197         cpu_affinity_setsize = 0;
1198
1199         free(thread_even);
1200         free(core_even);
1201         free(package_even);
1202
1203         thread_even = NULL;
1204         core_even = NULL;
1205         package_even = NULL;
1206
1207         free(thread_odd);
1208         free(core_odd);
1209         free(package_odd);
1210
1211         thread_odd = NULL;
1212         core_odd = NULL;
1213         package_odd = NULL;
1214
1215         free(output_buffer);
1216         output_buffer = NULL;
1217         outp = NULL;
1218 }
1219
1220 /*
1221  * Open a file, and exit on failure
1222  */
1223 FILE *fopen_or_die(const char *path, const char *mode)
1224 {
1225         FILE *filep = fopen(path, "r");
1226         if (!filep)
1227                 err(1, "%s: open failed", path);
1228         return filep;
1229 }
1230
1231 /*
1232  * Parse a file containing a single int.
1233  */
1234 int parse_int_file(const char *fmt, ...)
1235 {
1236         va_list args;
1237         char path[PATH_MAX];
1238         FILE *filep;
1239         int value;
1240
1241         va_start(args, fmt);
1242         vsnprintf(path, sizeof(path), fmt, args);
1243         va_end(args);
1244         filep = fopen_or_die(path, "r");
1245         if (fscanf(filep, "%d", &value) != 1)
1246                 err(1, "%s: failed to parse number from file", path);
1247         fclose(filep);
1248         return value;
1249 }
1250
1251 /*
1252  * cpu_is_first_sibling_in_core(cpu)
1253  * return 1 if given CPU is 1st HT sibling in the core
1254  */
1255 int cpu_is_first_sibling_in_core(int cpu)
1256 {
1257         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1258 }
1259
1260 /*
1261  * cpu_is_first_core_in_package(cpu)
1262  * return 1 if given CPU is 1st core in package
1263  */
1264 int cpu_is_first_core_in_package(int cpu)
1265 {
1266         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1267 }
1268
1269 int get_physical_package_id(int cpu)
1270 {
1271         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1272 }
1273
1274 int get_core_id(int cpu)
1275 {
1276         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1277 }
1278
1279 int get_num_ht_siblings(int cpu)
1280 {
1281         char path[80];
1282         FILE *filep;
1283         int sib1, sib2;
1284         int matches;
1285         char character;
1286
1287         sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1288         filep = fopen_or_die(path, "r");
1289         /*
1290          * file format:
1291          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1292          * otherwinse 1 sibling (self).
1293          */
1294         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1295
1296         fclose(filep);
1297
1298         if (matches == 3)
1299                 return 2;
1300         else
1301                 return 1;
1302 }
1303
1304 /*
1305  * run func(thread, core, package) in topology order
1306  * skip non-present cpus
1307  */
1308
1309 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1310         struct pkg_data *, struct thread_data *, struct core_data *,
1311         struct pkg_data *), struct thread_data *thread_base,
1312         struct core_data *core_base, struct pkg_data *pkg_base,
1313         struct thread_data *thread_base2, struct core_data *core_base2,
1314         struct pkg_data *pkg_base2)
1315 {
1316         int retval, pkg_no, core_no, thread_no;
1317
1318         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1319                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1320                         for (thread_no = 0; thread_no <
1321                                 topo.num_threads_per_core; ++thread_no) {
1322                                 struct thread_data *t, *t2;
1323                                 struct core_data *c, *c2;
1324                                 struct pkg_data *p, *p2;
1325
1326                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1327
1328                                 if (cpu_is_not_present(t->cpu_id))
1329                                         continue;
1330
1331                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1332
1333                                 c = GET_CORE(core_base, core_no, pkg_no);
1334                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
1335
1336                                 p = GET_PKG(pkg_base, pkg_no);
1337                                 p2 = GET_PKG(pkg_base2, pkg_no);
1338
1339                                 retval = func(t, c, p, t2, c2, p2);
1340                                 if (retval)
1341                                         return retval;
1342                         }
1343                 }
1344         }
1345         return 0;
1346 }
1347
1348 /*
1349  * run func(cpu) on every cpu in /proc/stat
1350  * return max_cpu number
1351  */
1352 int for_all_proc_cpus(int (func)(int))
1353 {
1354         FILE *fp;
1355         int cpu_num;
1356         int retval;
1357
1358         fp = fopen_or_die(proc_stat, "r");
1359
1360         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1361         if (retval != 0)
1362                 err(1, "%s: failed to parse format", proc_stat);
1363
1364         while (1) {
1365                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1366                 if (retval != 1)
1367                         break;
1368
1369                 retval = func(cpu_num);
1370                 if (retval) {
1371                         fclose(fp);
1372                         return(retval);
1373                 }
1374         }
1375         fclose(fp);
1376         return 0;
1377 }
1378
1379 void re_initialize(void)
1380 {
1381         free_all_buffers();
1382         setup_all_buffers();
1383         printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1384 }
1385
1386
1387 /*
1388  * count_cpus()
1389  * remember the last one seen, it will be the max
1390  */
1391 int count_cpus(int cpu)
1392 {
1393         if (topo.max_cpu_num < cpu)
1394                 topo.max_cpu_num = cpu;
1395
1396         topo.num_cpus += 1;
1397         return 0;
1398 }
1399 int mark_cpu_present(int cpu)
1400 {
1401         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1402         return 0;
1403 }
1404
1405 void turbostat_loop()
1406 {
1407         int retval;
1408         int restarted = 0;
1409
1410 restart:
1411         restarted++;
1412
1413         retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1414         if (retval < -1) {
1415                 exit(retval);
1416         } else if (retval == -1) {
1417                 if (restarted > 1) {
1418                         exit(retval);
1419                 }
1420                 re_initialize();
1421                 goto restart;
1422         }
1423         restarted = 0;
1424         gettimeofday(&tv_even, (struct timezone *)NULL);
1425
1426         while (1) {
1427                 if (for_all_proc_cpus(cpu_is_not_present)) {
1428                         re_initialize();
1429                         goto restart;
1430                 }
1431                 sleep(interval_sec);
1432                 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1433                 if (retval < -1) {
1434                         exit(retval);
1435                 } else if (retval == -1) {
1436                         re_initialize();
1437                         goto restart;
1438                 }
1439                 gettimeofday(&tv_odd, (struct timezone *)NULL);
1440                 timersub(&tv_odd, &tv_even, &tv_delta);
1441                 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1442                 compute_average(EVEN_COUNTERS);
1443                 format_all_counters(EVEN_COUNTERS);
1444                 flush_stdout();
1445                 sleep(interval_sec);
1446                 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1447                 if (retval < -1) {
1448                         exit(retval);
1449                 } else if (retval == -1) {
1450                         re_initialize();
1451                         goto restart;
1452                 }
1453                 gettimeofday(&tv_even, (struct timezone *)NULL);
1454                 timersub(&tv_even, &tv_odd, &tv_delta);
1455                 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1456                 compute_average(ODD_COUNTERS);
1457                 format_all_counters(ODD_COUNTERS);
1458                 flush_stdout();
1459         }
1460 }
1461
1462 void check_dev_msr()
1463 {
1464         struct stat sb;
1465
1466         if (stat("/dev/cpu/0/msr", &sb))
1467                 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
1468 }
1469
1470 void check_permissions()
1471 {
1472         struct __user_cap_header_struct cap_header_data;
1473         cap_user_header_t cap_header = &cap_header_data;
1474         struct __user_cap_data_struct cap_data_data;
1475         cap_user_data_t cap_data = &cap_data_data;
1476         extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
1477         int do_exit = 0;
1478
1479         /* check for CAP_SYS_RAWIO */
1480         cap_header->pid = getpid();
1481         cap_header->version = _LINUX_CAPABILITY_VERSION;
1482         if (capget(cap_header, cap_data) < 0)
1483                 err(-6, "capget(2) failed");
1484
1485         if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1486                 do_exit++;
1487                 warnx("capget(CAP_SYS_RAWIO) failed,"
1488                         " try \"# setcap cap_sys_rawio=ep %s\"", progname);
1489         }
1490
1491         /* test file permissions */
1492         if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1493                 do_exit++;
1494                 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1495         }
1496
1497         /* if all else fails, thell them to be root */
1498         if (do_exit)
1499                 if (getuid() != 0)
1500                         warnx("... or simply run as root");
1501
1502         if (do_exit)
1503                 exit(-6);
1504 }
1505
1506 /*
1507  * NHM adds support for additional MSRs:
1508  *
1509  * MSR_SMI_COUNT                   0x00000034
1510  *
1511  * MSR_NHM_PLATFORM_INFO           0x000000ce
1512  * MSR_NHM_SNB_PKG_CST_CFG_CTL     0x000000e2
1513  *
1514  * MSR_PKG_C3_RESIDENCY            0x000003f8
1515  * MSR_PKG_C6_RESIDENCY            0x000003f9
1516  * MSR_CORE_C3_RESIDENCY           0x000003fc
1517  * MSR_CORE_C6_RESIDENCY           0x000003fd
1518  *
1519  */
1520 int has_nhm_msrs(unsigned int family, unsigned int model)
1521 {
1522         if (!genuine_intel)
1523                 return 0;
1524
1525         if (family != 6)
1526                 return 0;
1527
1528         switch (model) {
1529         case 0x1A:      /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1530         case 0x1E:      /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1531         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
1532         case 0x25:      /* Westmere Client - Clarkdale, Arrandale */
1533         case 0x2C:      /* Westmere EP - Gulftown */
1534         case 0x2A:      /* SNB */
1535         case 0x2D:      /* SNB Xeon */
1536         case 0x3A:      /* IVB */
1537         case 0x3E:      /* IVB Xeon */
1538         case 0x3C:      /* HSW */
1539         case 0x3F:      /* HSX */
1540         case 0x45:      /* HSW */
1541         case 0x46:      /* HSW */
1542         case 0x37:      /* BYT */
1543         case 0x4D:      /* AVN */
1544         case 0x3D:      /* BDW */
1545         case 0x4F:      /* BDX */
1546         case 0x56:      /* BDX-DE */
1547         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1548         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1549                 return 1;
1550         default:
1551                 return 0;
1552         }
1553 }
1554 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
1555 {
1556         if (!has_nhm_msrs(family, model))
1557                 return 0;
1558
1559         switch (model) {
1560         /* Nehalem compatible, but do not include turbo-ratio limit support */
1561         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1562         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1563                 return 0;
1564         default:
1565                 return 1;
1566         }
1567 }
1568 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1569 {
1570         if (!genuine_intel)
1571                 return 0;
1572
1573         if (family != 6)
1574                 return 0;
1575
1576         switch (model) {
1577         case 0x3E:      /* IVB Xeon */
1578                 return 1;
1579         default:
1580                 return 0;
1581         }
1582 }
1583
1584 /*
1585  * print_epb()
1586  * Decode the ENERGY_PERF_BIAS MSR
1587  */
1588 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1589 {
1590         unsigned long long msr;
1591         char *epb_string;
1592         int cpu;
1593
1594         if (!has_epb)
1595                 return 0;
1596
1597         cpu = t->cpu_id;
1598
1599         /* EPB is per-package */
1600         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1601                 return 0;
1602
1603         if (cpu_migrate(cpu)) {
1604                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1605                 return -1;
1606         }
1607
1608         if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
1609                 return 0;
1610
1611         switch (msr & 0x7) {
1612         case ENERGY_PERF_BIAS_PERFORMANCE:
1613                 epb_string = "performance";
1614                 break;
1615         case ENERGY_PERF_BIAS_NORMAL:
1616                 epb_string = "balanced";
1617                 break;
1618         case ENERGY_PERF_BIAS_POWERSAVE:
1619                 epb_string = "powersave";
1620                 break;
1621         default:
1622                 epb_string = "custom";
1623                 break;
1624         }
1625         fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
1626
1627         return 0;
1628 }
1629
1630 /*
1631  * print_perf_limit()
1632  */
1633 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1634 {
1635         unsigned long long msr;
1636         int cpu;
1637
1638         cpu = t->cpu_id;
1639
1640         /* per-package */
1641         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1642                 return 0;
1643
1644         if (cpu_migrate(cpu)) {
1645                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1646                 return -1;
1647         }
1648
1649         if (do_core_perf_limit_reasons) {
1650                 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
1651                 fprintf(stderr, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1652                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
1653                         (msr & 1 << 0) ? "PROCHOT, " : "",
1654                         (msr & 1 << 1) ? "ThermStatus, " : "",
1655                         (msr & 1 << 2) ? "bit2, " : "",
1656                         (msr & 1 << 4) ? "Graphics, " : "",
1657                         (msr & 1 << 5) ? "Auto-HWP, " : "",
1658                         (msr & 1 << 6) ? "VR-Therm, " : "",
1659                         (msr & 1 << 8) ? "Amps, " : "",
1660                         (msr & 1 << 9) ? "CorePwr, " : "",
1661                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1662                         (msr & 1 << 11) ? "PkgPwrL2, " : "",
1663                         (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
1664                         (msr & 1 << 13) ? "Transitions, " : "",
1665                         (msr & 1 << 14) ? "bit14, " : "",
1666                         (msr & 1 << 15) ? "bit15, " : "");
1667                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
1668                         (msr & 1 << 16) ? "PROCHOT, " : "",
1669                         (msr & 1 << 17) ? "ThermStatus, " : "",
1670                         (msr & 1 << 18) ? "bit18, " : "",
1671                         (msr & 1 << 20) ? "Graphics, " : "",
1672                         (msr & 1 << 21) ? "Auto-HWP, " : "",
1673                         (msr & 1 << 22) ? "VR-Therm, " : "",
1674                         (msr & 1 << 24) ? "Amps, " : "",
1675                         (msr & 1 << 25) ? "CorePwr, " : "",
1676                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1677                         (msr & 1 << 27) ? "PkgPwrL2, " : "",
1678                         (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
1679                         (msr & 1 << 29) ? "Transitions, " : "",
1680                         (msr & 1 << 30) ? "bit30, " : "",
1681                         (msr & 1 << 31) ? "bit31, " : "");
1682
1683         }
1684         if (do_gfx_perf_limit_reasons) {
1685                 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
1686                 fprintf(stderr, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1687                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s)",
1688                         (msr & 1 << 0) ? "PROCHOT, " : "",
1689                         (msr & 1 << 1) ? "ThermStatus, " : "",
1690                         (msr & 1 << 4) ? "Graphics, " : "",
1691                         (msr & 1 << 6) ? "VR-Therm, " : "",
1692                         (msr & 1 << 8) ? "Amps, " : "",
1693                         (msr & 1 << 9) ? "GFXPwr, " : "",
1694                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1695                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
1696                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s)\n",
1697                         (msr & 1 << 16) ? "PROCHOT, " : "",
1698                         (msr & 1 << 17) ? "ThermStatus, " : "",
1699                         (msr & 1 << 20) ? "Graphics, " : "",
1700                         (msr & 1 << 22) ? "VR-Therm, " : "",
1701                         (msr & 1 << 24) ? "Amps, " : "",
1702                         (msr & 1 << 25) ? "GFXPwr, " : "",
1703                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1704                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
1705         }
1706         if (do_ring_perf_limit_reasons) {
1707                 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
1708                 fprintf(stderr, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1709                 fprintf(stderr, " (Active: %s%s%s%s%s%s)",
1710                         (msr & 1 << 0) ? "PROCHOT, " : "",
1711                         (msr & 1 << 1) ? "ThermStatus, " : "",
1712                         (msr & 1 << 6) ? "VR-Therm, " : "",
1713                         (msr & 1 << 8) ? "Amps, " : "",
1714                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
1715                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
1716                 fprintf(stderr, " (Logged: %s%s%s%s%s%s)\n",
1717                         (msr & 1 << 16) ? "PROCHOT, " : "",
1718                         (msr & 1 << 17) ? "ThermStatus, " : "",
1719                         (msr & 1 << 22) ? "VR-Therm, " : "",
1720                         (msr & 1 << 24) ? "Amps, " : "",
1721                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
1722                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
1723         }
1724         return 0;
1725 }
1726
1727 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
1728 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
1729
1730 double get_tdp(model)
1731 {
1732         unsigned long long msr;
1733
1734         if (do_rapl & RAPL_PKG_POWER_INFO)
1735                 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
1736                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
1737
1738         switch (model) {
1739         case 0x37:
1740         case 0x4D:
1741                 return 30.0;
1742         default:
1743                 return 135.0;
1744         }
1745 }
1746
1747
1748 /*
1749  * rapl_probe()
1750  *
1751  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
1752  */
1753 void rapl_probe(unsigned int family, unsigned int model)
1754 {
1755         unsigned long long msr;
1756         unsigned int time_unit;
1757         double tdp;
1758
1759         if (!genuine_intel)
1760                 return;
1761
1762         if (family != 6)
1763                 return;
1764
1765         switch (model) {
1766         case 0x2A:
1767         case 0x3A:
1768         case 0x3C:      /* HSW */
1769         case 0x45:      /* HSW */
1770         case 0x46:      /* HSW */
1771         case 0x3D:      /* BDW */
1772                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1773                 break;
1774         case 0x3F:      /* HSX */
1775         case 0x4F:      /* BDX */
1776         case 0x56:      /* BDX-DE */
1777                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1778                 break;
1779         case 0x2D:
1780         case 0x3E:
1781                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1782                 break;
1783         case 0x37:      /* BYT */
1784         case 0x4D:      /* AVN */
1785                 do_rapl = RAPL_PKG | RAPL_CORES ;
1786                 break;
1787         default:
1788                 return;
1789         }
1790
1791         /* units on package 0, verify later other packages match */
1792         if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1793                 return;
1794
1795         rapl_power_units = 1.0 / (1 << (msr & 0xF));
1796         if (model == 0x37)
1797                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1798         else
1799                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1800
1801         time_unit = msr >> 16 & 0xF;
1802         if (time_unit == 0)
1803                 time_unit = 0xA;
1804
1805         rapl_time_units = 1.0 / (1 << (time_unit));
1806
1807         tdp = get_tdp(model);
1808
1809         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1810         if (verbose)
1811                 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1812
1813         return;
1814 }
1815
1816 void perf_limit_reasons_probe(family, model)
1817 {
1818         if (!genuine_intel)
1819                 return;
1820
1821         if (family != 6)
1822                 return;
1823
1824         switch (model) {
1825         case 0x3C:      /* HSW */
1826         case 0x45:      /* HSW */
1827         case 0x46:      /* HSW */
1828                 do_gfx_perf_limit_reasons = 1;
1829         case 0x3F:      /* HSX */
1830                 do_core_perf_limit_reasons = 1;
1831                 do_ring_perf_limit_reasons = 1;
1832         default:
1833                 return;
1834         }
1835 }
1836
1837 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1838 {
1839         unsigned long long msr;
1840         unsigned int dts;
1841         int cpu;
1842
1843         if (!(do_dts || do_ptm))
1844                 return 0;
1845
1846         cpu = t->cpu_id;
1847
1848         /* DTS is per-core, no need to print for each thread */
1849         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1850                 return 0;
1851
1852         if (cpu_migrate(cpu)) {
1853                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1854                 return -1;
1855         }
1856
1857         if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
1858                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1859                         return 0;
1860
1861                 dts = (msr >> 16) & 0x7F;
1862                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
1863                         cpu, msr, tcc_activation_temp - dts);
1864
1865 #ifdef  THERM_DEBUG
1866                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
1867                         return 0;
1868
1869                 dts = (msr >> 16) & 0x7F;
1870                 dts2 = (msr >> 8) & 0x7F;
1871                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1872                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1873 #endif
1874         }
1875
1876
1877         if (do_dts) {
1878                 unsigned int resolution;
1879
1880                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1881                         return 0;
1882
1883                 dts = (msr >> 16) & 0x7F;
1884                 resolution = (msr >> 27) & 0xF;
1885                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
1886                         cpu, msr, tcc_activation_temp - dts, resolution);
1887
1888 #ifdef THERM_DEBUG
1889                 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
1890                         return 0;
1891
1892                 dts = (msr >> 16) & 0x7F;
1893                 dts2 = (msr >> 8) & 0x7F;
1894                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1895                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1896 #endif
1897         }
1898
1899         return 0;
1900 }
1901
1902 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
1903 {
1904         fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
1905                 cpu, label,
1906                 ((msr >> 15) & 1) ? "EN" : "DIS",
1907                 ((msr >> 0) & 0x7FFF) * rapl_power_units,
1908                 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
1909                 (((msr >> 16) & 1) ? "EN" : "DIS"));
1910
1911         return;
1912 }
1913
1914 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1915 {
1916         unsigned long long msr;
1917         int cpu;
1918
1919         if (!do_rapl)
1920                 return 0;
1921
1922         /* RAPL counters are per package, so print only for 1st thread/package */
1923         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1924                 return 0;
1925
1926         cpu = t->cpu_id;
1927         if (cpu_migrate(cpu)) {
1928                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1929                 return -1;
1930         }
1931
1932         if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
1933                 return -1;
1934
1935         if (verbose) {
1936                 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
1937                         "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
1938                         rapl_power_units, rapl_energy_units, rapl_time_units);
1939         }
1940         if (do_rapl & RAPL_PKG_POWER_INFO) {
1941
1942                 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
1943                         return -5;
1944
1945
1946                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1947                         cpu, msr,
1948                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1949                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1950                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1951                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1952
1953         }
1954         if (do_rapl & RAPL_PKG) {
1955
1956                 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
1957                         return -9;
1958
1959                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
1960                         cpu, msr, (msr >> 63) & 1 ? "": "UN");
1961
1962                 print_power_limit_msr(cpu, msr, "PKG Limit #1");
1963                 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
1964                         cpu,
1965                         ((msr >> 47) & 1) ? "EN" : "DIS",
1966                         ((msr >> 32) & 0x7FFF) * rapl_power_units,
1967                         (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
1968                         ((msr >> 48) & 1) ? "EN" : "DIS");
1969         }
1970
1971         if (do_rapl & RAPL_DRAM) {
1972                 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
1973                         return -6;
1974
1975
1976                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1977                         cpu, msr,
1978                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1979                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1980                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1981                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1982
1983
1984                 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
1985                         return -9;
1986                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
1987                                 cpu, msr, (msr >> 31) & 1 ? "": "UN");
1988
1989                 print_power_limit_msr(cpu, msr, "DRAM Limit");
1990         }
1991         if (do_rapl & RAPL_CORE_POLICY) {
1992                 if (verbose) {
1993                         if (get_msr(cpu, MSR_PP0_POLICY, &msr))
1994                                 return -7;
1995
1996                         fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
1997                 }
1998         }
1999         if (do_rapl & RAPL_CORES) {
2000                 if (verbose) {
2001
2002                         if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
2003                                 return -9;
2004                         fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2005                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2006                         print_power_limit_msr(cpu, msr, "Cores Limit");
2007                 }
2008         }
2009         if (do_rapl & RAPL_GFX) {
2010                 if (verbose) {
2011                         if (get_msr(cpu, MSR_PP1_POLICY, &msr))
2012                                 return -8;
2013
2014                         fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
2015
2016                         if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
2017                                 return -9;
2018                         fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2019                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2020                         print_power_limit_msr(cpu, msr, "GFX Limit");
2021                 }
2022         }
2023         return 0;
2024 }
2025
2026 /*
2027  * SNB adds support for additional MSRs:
2028  *
2029  * MSR_PKG_C7_RESIDENCY            0x000003fa
2030  * MSR_CORE_C7_RESIDENCY           0x000003fe
2031  * MSR_PKG_C2_RESIDENCY            0x0000060d
2032  */
2033
2034 int has_snb_msrs(unsigned int family, unsigned int model)
2035 {
2036         if (!genuine_intel)
2037                 return 0;
2038
2039         switch (model) {
2040         case 0x2A:
2041         case 0x2D:
2042         case 0x3A:      /* IVB */
2043         case 0x3E:      /* IVB Xeon */
2044         case 0x3C:      /* HSW */
2045         case 0x3F:      /* HSW */
2046         case 0x45:      /* HSW */
2047         case 0x46:      /* HSW */
2048         case 0x3D:      /* BDW */
2049         case 0x4F:      /* BDX */
2050         case 0x56:      /* BDX-DE */
2051                 return 1;
2052         }
2053         return 0;
2054 }
2055
2056 /*
2057  * HSW adds support for additional MSRs:
2058  *
2059  * MSR_PKG_C8_RESIDENCY            0x00000630
2060  * MSR_PKG_C9_RESIDENCY            0x00000631
2061  * MSR_PKG_C10_RESIDENCY           0x00000632
2062  */
2063 int has_hsw_msrs(unsigned int family, unsigned int model)
2064 {
2065         if (!genuine_intel)
2066                 return 0;
2067
2068         switch (model) {
2069         case 0x45:      /* HSW */
2070         case 0x3D:      /* BDW */
2071                 return 1;
2072         }
2073         return 0;
2074 }
2075
2076
2077 int is_slm(unsigned int family, unsigned int model)
2078 {
2079         if (!genuine_intel)
2080                 return 0;
2081         switch (model) {
2082         case 0x37:      /* BYT */
2083         case 0x4D:      /* AVN */
2084                 return 1;
2085         }
2086         return 0;
2087 }
2088
2089 #define SLM_BCLK_FREQS 5
2090 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
2091
2092 double slm_bclk(void)
2093 {
2094         unsigned long long msr = 3;
2095         unsigned int i;
2096         double freq;
2097
2098         if (get_msr(0, MSR_FSB_FREQ, &msr))
2099                 fprintf(stderr, "SLM BCLK: unknown\n");
2100
2101         i = msr & 0xf;
2102         if (i >= SLM_BCLK_FREQS) {
2103                 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
2104                 msr = 3;
2105         }
2106         freq = slm_freq_table[i];
2107
2108         fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
2109
2110         return freq;
2111 }
2112
2113 double discover_bclk(unsigned int family, unsigned int model)
2114 {
2115         if (has_snb_msrs(family, model))
2116                 return 100.00;
2117         else if (is_slm(family, model))
2118                 return slm_bclk();
2119         else
2120                 return 133.33;
2121 }
2122
2123 /*
2124  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
2125  * the Thermal Control Circuit (TCC) activates.
2126  * This is usually equal to tjMax.
2127  *
2128  * Older processors do not have this MSR, so there we guess,
2129  * but also allow cmdline over-ride with -T.
2130  *
2131  * Several MSR temperature values are in units of degrees-C
2132  * below this value, including the Digital Thermal Sensor (DTS),
2133  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
2134  */
2135 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2136 {
2137         unsigned long long msr;
2138         unsigned int target_c_local;
2139         int cpu;
2140
2141         /* tcc_activation_temp is used only for dts or ptm */
2142         if (!(do_dts || do_ptm))
2143                 return 0;
2144
2145         /* this is a per-package concept */
2146         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2147                 return 0;
2148
2149         cpu = t->cpu_id;
2150         if (cpu_migrate(cpu)) {
2151                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2152                 return -1;
2153         }
2154
2155         if (tcc_activation_temp_override != 0) {
2156                 tcc_activation_temp = tcc_activation_temp_override;
2157                 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
2158                         cpu, tcc_activation_temp);
2159                 return 0;
2160         }
2161
2162         /* Temperature Target MSR is Nehalem and newer only */
2163         if (!do_nhm_platform_info)
2164                 goto guess;
2165
2166         if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
2167                 goto guess;
2168
2169         target_c_local = (msr >> 16) & 0xFF;
2170
2171         if (verbose)
2172                 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2173                         cpu, msr, target_c_local);
2174
2175         if (!target_c_local)
2176                 goto guess;
2177
2178         tcc_activation_temp = target_c_local;
2179
2180         return 0;
2181
2182 guess:
2183         tcc_activation_temp = TJMAX_DEFAULT;
2184         fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2185                 cpu, tcc_activation_temp);
2186
2187         return 0;
2188 }
2189 void check_cpuid()
2190 {
2191         unsigned int eax, ebx, ecx, edx, max_level;
2192         unsigned int fms, family, model, stepping;
2193
2194         eax = ebx = ecx = edx = 0;
2195
2196         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2197
2198         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2199                 genuine_intel = 1;
2200
2201         if (verbose)
2202                 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2203                         (char *)&ebx, (char *)&edx, (char *)&ecx);
2204
2205         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2206         family = (fms >> 8) & 0xf;
2207         model = (fms >> 4) & 0xf;
2208         stepping = fms & 0xf;
2209         if (family == 6 || family == 0xf)
2210                 model += ((fms >> 16) & 0xf) << 4;
2211
2212         if (verbose)
2213                 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2214                         max_level, family, model, stepping, family, model, stepping);
2215
2216         if (!(edx & (1 << 5)))
2217                 errx(1, "CPUID: no MSR");
2218
2219         /*
2220          * check max extended function levels of CPUID.
2221          * This is needed to check for invariant TSC.
2222          * This check is valid for both Intel and AMD.
2223          */
2224         ebx = ecx = edx = 0;
2225         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2226
2227         if (max_level >= 0x80000007) {
2228
2229                 /*
2230                  * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2231                  * this check is valid for both Intel and AMD
2232                  */
2233                 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2234                 has_invariant_tsc = edx & (1 << 8);
2235         }
2236
2237         /*
2238          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2239          * this check is valid for both Intel and AMD
2240          */
2241
2242         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2243         has_aperf = ecx & (1 << 0);
2244         do_dts = eax & (1 << 0);
2245         do_ptm = eax & (1 << 6);
2246         has_epb = ecx & (1 << 3);
2247
2248         if (verbose)
2249                 fprintf(stderr, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sEPB\n",
2250                         has_aperf ? "" : "No ",
2251                         do_dts ? "" : "No ",
2252                         do_ptm ? "" : "No ",
2253                         has_epb ? "" : "No ");
2254
2255         do_nhm_platform_info = do_nhm_cstates = do_smi = has_nhm_msrs(family, model);
2256         do_snb_cstates = has_snb_msrs(family, model);
2257         do_c8_c9_c10 = has_hsw_msrs(family, model);
2258         do_slm_cstates = is_slm(family, model);
2259         bclk = discover_bclk(family, model);
2260
2261         do_nhm_turbo_ratio_limit = has_nhm_turbo_ratio_limit(family, model);
2262         do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model);
2263         rapl_probe(family, model);
2264         perf_limit_reasons_probe(family, model);
2265
2266         return;
2267 }
2268
2269
2270 void usage()
2271 {
2272         errx(1, "%s: [-v][-R][-T][-p|-P|-S][-c MSR#][-C MSR#][-m MSR#][-M MSR#][-i interval_sec | command ...]\n",
2273              progname);
2274 }
2275
2276
2277 /*
2278  * in /dev/cpu/ return success for names that are numbers
2279  * ie. filter out ".", "..", "microcode".
2280  */
2281 int dir_filter(const struct dirent *dirp)
2282 {
2283         if (isdigit(dirp->d_name[0]))
2284                 return 1;
2285         else
2286                 return 0;
2287 }
2288
2289 int open_dev_cpu_msr(int dummy1)
2290 {
2291         return 0;
2292 }
2293
2294 void topology_probe()
2295 {
2296         int i;
2297         int max_core_id = 0;
2298         int max_package_id = 0;
2299         int max_siblings = 0;
2300         struct cpu_topology {
2301                 int core_id;
2302                 int physical_package_id;
2303         } *cpus;
2304
2305         /* Initialize num_cpus, max_cpu_num */
2306         topo.num_cpus = 0;
2307         topo.max_cpu_num = 0;
2308         for_all_proc_cpus(count_cpus);
2309         if (!summary_only && topo.num_cpus > 1)
2310                 show_cpu = 1;
2311
2312         if (verbose > 1)
2313                 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2314
2315         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
2316         if (cpus == NULL)
2317                 err(1, "calloc cpus");
2318
2319         /*
2320          * Allocate and initialize cpu_present_set
2321          */
2322         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2323         if (cpu_present_set == NULL)
2324                 err(3, "CPU_ALLOC");
2325         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2326         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2327         for_all_proc_cpus(mark_cpu_present);
2328
2329         /*
2330          * Allocate and initialize cpu_affinity_set
2331          */
2332         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2333         if (cpu_affinity_set == NULL)
2334                 err(3, "CPU_ALLOC");
2335         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2336         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2337
2338
2339         /*
2340          * For online cpus
2341          * find max_core_id, max_package_id
2342          */
2343         for (i = 0; i <= topo.max_cpu_num; ++i) {
2344                 int siblings;
2345
2346                 if (cpu_is_not_present(i)) {
2347                         if (verbose > 1)
2348                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2349                         continue;
2350                 }
2351                 cpus[i].core_id = get_core_id(i);
2352                 if (cpus[i].core_id > max_core_id)
2353                         max_core_id = cpus[i].core_id;
2354
2355                 cpus[i].physical_package_id = get_physical_package_id(i);
2356                 if (cpus[i].physical_package_id > max_package_id)
2357                         max_package_id = cpus[i].physical_package_id;
2358
2359                 siblings = get_num_ht_siblings(i);
2360                 if (siblings > max_siblings)
2361                         max_siblings = siblings;
2362                 if (verbose > 1)
2363                         fprintf(stderr, "cpu %d pkg %d core %d\n",
2364                                 i, cpus[i].physical_package_id, cpus[i].core_id);
2365         }
2366         topo.num_cores_per_pkg = max_core_id + 1;
2367         if (verbose > 1)
2368                 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2369                         max_core_id, topo.num_cores_per_pkg);
2370         if (!summary_only && topo.num_cores_per_pkg > 1)
2371                 show_core = 1;
2372
2373         topo.num_packages = max_package_id + 1;
2374         if (verbose > 1)
2375                 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2376                         max_package_id, topo.num_packages);
2377         if (!summary_only && topo.num_packages > 1)
2378                 show_pkg = 1;
2379
2380         topo.num_threads_per_core = max_siblings;
2381         if (verbose > 1)
2382                 fprintf(stderr, "max_siblings %d\n", max_siblings);
2383
2384         free(cpus);
2385 }
2386
2387 void
2388 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2389 {
2390         int i;
2391
2392         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2393                 topo.num_packages, sizeof(struct thread_data));
2394         if (*t == NULL)
2395                 goto error;
2396
2397         for (i = 0; i < topo.num_threads_per_core *
2398                 topo.num_cores_per_pkg * topo.num_packages; i++)
2399                 (*t)[i].cpu_id = -1;
2400
2401         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2402                 sizeof(struct core_data));
2403         if (*c == NULL)
2404                 goto error;
2405
2406         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2407                 (*c)[i].core_id = -1;
2408
2409         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2410         if (*p == NULL)
2411                 goto error;
2412
2413         for (i = 0; i < topo.num_packages; i++)
2414                 (*p)[i].package_id = i;
2415
2416         return;
2417 error:
2418         err(1, "calloc counters");
2419 }
2420 /*
2421  * init_counter()
2422  *
2423  * set cpu_id, core_num, pkg_num
2424  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2425  *
2426  * increment topo.num_cores when 1st core in pkg seen
2427  */
2428 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2429         struct pkg_data *pkg_base, int thread_num, int core_num,
2430         int pkg_num, int cpu_id)
2431 {
2432         struct thread_data *t;
2433         struct core_data *c;
2434         struct pkg_data *p;
2435
2436         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2437         c = GET_CORE(core_base, core_num, pkg_num);
2438         p = GET_PKG(pkg_base, pkg_num);
2439
2440         t->cpu_id = cpu_id;
2441         if (thread_num == 0) {
2442                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2443                 if (cpu_is_first_core_in_package(cpu_id))
2444                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2445         }
2446
2447         c->core_id = core_num;
2448         p->package_id = pkg_num;
2449 }
2450
2451
2452 int initialize_counters(int cpu_id)
2453 {
2454         int my_thread_id, my_core_id, my_package_id;
2455
2456         my_package_id = get_physical_package_id(cpu_id);
2457         my_core_id = get_core_id(cpu_id);
2458
2459         if (cpu_is_first_sibling_in_core(cpu_id)) {
2460                 my_thread_id = 0;
2461                 topo.num_cores++;
2462         } else {
2463                 my_thread_id = 1;
2464         }
2465
2466         init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2467         init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2468         return 0;
2469 }
2470
2471 void allocate_output_buffer()
2472 {
2473         output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
2474         outp = output_buffer;
2475         if (outp == NULL)
2476                 err(-1, "calloc output buffer");
2477 }
2478
2479 void setup_all_buffers(void)
2480 {
2481         topology_probe();
2482         allocate_counters(&thread_even, &core_even, &package_even);
2483         allocate_counters(&thread_odd, &core_odd, &package_odd);
2484         allocate_output_buffer();
2485         for_all_proc_cpus(initialize_counters);
2486 }
2487
2488 void turbostat_init()
2489 {
2490         check_dev_msr();
2491         check_permissions();
2492         check_cpuid();
2493
2494         setup_all_buffers();
2495
2496         if (verbose)
2497                 print_verbose_header();
2498
2499         if (verbose)
2500                 for_all_cpus(print_epb, ODD_COUNTERS);
2501
2502         if (verbose)
2503                 for_all_cpus(print_perf_limit, ODD_COUNTERS);
2504
2505         if (verbose)
2506                 for_all_cpus(print_rapl, ODD_COUNTERS);
2507
2508         for_all_cpus(set_temperature_target, ODD_COUNTERS);
2509
2510         if (verbose)
2511                 for_all_cpus(print_thermal, ODD_COUNTERS);
2512 }
2513
2514 int fork_it(char **argv)
2515 {
2516         pid_t child_pid;
2517         int status;
2518
2519         status = for_all_cpus(get_counters, EVEN_COUNTERS);
2520         if (status)
2521                 exit(status);
2522         /* clear affinity side-effect of get_counters() */
2523         sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
2524         gettimeofday(&tv_even, (struct timezone *)NULL);
2525
2526         child_pid = fork();
2527         if (!child_pid) {
2528                 /* child */
2529                 execvp(argv[0], argv);
2530         } else {
2531
2532                 /* parent */
2533                 if (child_pid == -1)
2534                         err(1, "fork");
2535
2536                 signal(SIGINT, SIG_IGN);
2537                 signal(SIGQUIT, SIG_IGN);
2538                 if (waitpid(child_pid, &status, 0) == -1)
2539                         err(status, "waitpid");
2540         }
2541         /*
2542          * n.b. fork_it() does not check for errors from for_all_cpus()
2543          * because re-starting is problematic when forking
2544          */
2545         for_all_cpus(get_counters, ODD_COUNTERS);
2546         gettimeofday(&tv_odd, (struct timezone *)NULL);
2547         timersub(&tv_odd, &tv_even, &tv_delta);
2548         for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
2549         compute_average(EVEN_COUNTERS);
2550         format_all_counters(EVEN_COUNTERS);
2551         flush_stderr();
2552
2553         fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
2554
2555         return status;
2556 }
2557
2558 int get_and_dump_counters(void)
2559 {
2560         int status;
2561
2562         status = for_all_cpus(get_counters, ODD_COUNTERS);
2563         if (status)
2564                 return status;
2565
2566         status = for_all_cpus(dump_counters, ODD_COUNTERS);
2567         if (status)
2568                 return status;
2569
2570         flush_stdout();
2571
2572         return status;
2573 }
2574
2575 void cmdline(int argc, char **argv)
2576 {
2577         int opt;
2578
2579         progname = argv[0];
2580
2581         while ((opt = getopt(argc, argv, "+pPsSvi:c:C:m:M:RJT:")) != -1) {
2582                 switch (opt) {
2583                 case 'p':
2584                         show_core_only++;
2585                         break;
2586                 case 'P':
2587                         show_pkg_only++;
2588                         break;
2589                 case 's':
2590                         dump_only++;
2591                         break;
2592                 case 'S':
2593                         summary_only++;
2594                         break;
2595                 case 'v':
2596                         verbose++;
2597                         break;
2598                 case 'i':
2599                         interval_sec = atoi(optarg);
2600                         break;
2601                 case 'c':
2602                         sscanf(optarg, "%x", &extra_delta_offset32);
2603                         break;
2604                 case 'C':
2605                         sscanf(optarg, "%x", &extra_delta_offset64);
2606                         break;
2607                 case 'm':
2608                         sscanf(optarg, "%x", &extra_msr_offset32);
2609                         break;
2610                 case 'M':
2611                         sscanf(optarg, "%x", &extra_msr_offset64);
2612                         break;
2613                 case 'R':
2614                         rapl_verbose++;
2615                         break;
2616                 case 'T':
2617                         tcc_activation_temp_override = atoi(optarg);
2618                         break;
2619                 case 'J':
2620                         rapl_joules++;
2621                         break;
2622
2623                 default:
2624                         usage();
2625                 }
2626         }
2627 }
2628
2629 int main(int argc, char **argv)
2630 {
2631         cmdline(argc, argv);
2632
2633         if (verbose)
2634                 fprintf(stderr, "turbostat v3.9 23-Jan, 2015"
2635                         " - Len Brown <lenb@kernel.org>\n");
2636
2637         turbostat_init();
2638
2639         /* dump counters and exit */
2640         if (dump_only)
2641                 return get_and_dump_counters();
2642
2643         /*
2644          * if any params left, it must be a command to fork
2645          */
2646         if (argc - optind)
2647                 return fork_it(argv + optind);
2648         else
2649                 turbostat_loop();
2650
2651         return 0;
2652 }