src/turbostat.c

   1 /*
   2  * turbostat -- Log CPU frequency and C-state residency
   3  * on modern Intel turbo-capable processors for collectd.
   4  *
   5  * Based on the 'turbostat' tool of the Linux kernel, found at
   6  * linux/tools/power/x86/turbostat/turbostat.c:
   7  * ----
   8  * Copyright (c) 2013 Intel Corporation.
   9  * Len Brown <len.brown@intel.com>
  10  *
  11  * This program is free software; you can redistribute it and/or modify it
  12  * under the terms and conditions of the GNU General Public License,
  13  * version 2, as published by the Free Software Foundation.
  14  *
  15  * This program is distributed in the hope it will be useful, but WITHOUT
  16  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  17  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  18  * more details.
  19  *
  20  * You should have received a copy of the GNU General Public License along with
  21  * this program; if not, write to the Free Software Foundation, Inc.,
  22  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  23  * ----
  24  * Ported to collectd by Vincent Brillault <git@lerya.net>
  25  */
  26
  27 /*
  28  * _GNU_SOURCE is required because of the following functions:
  29  * - CPU_ISSET_S
  30  * - CPU_ZERO_S
  31  * - CPU_SET_S
  32  * - CPU_FREE
  33  * - CPU_ALLOC
  34  * - CPU_ALLOC_SIZE
  35  */
  36 #define _GNU_SOURCE
  37
  38 #include <asm/msr-index.h>
  39 #include <stdarg.h>
  40 #include <stdio.h>
  41 #include <err.h>
  42 #include <unistd.h>
  43 #include <sys/types.h>
  44 #include <sys/wait.h>
  45 #include <sys/stat.h>
  46 #include <sys/resource.h>
  47 #include <fcntl.h>
  48 #include <signal.h>
  49 #include <sys/time.h>
  50 #include <stdlib.h>
  51 #include <dirent.h>
  52 #include <string.h>
  53 #include <ctype.h>
  54 #include <sched.h>
  55 #include <cpuid.h>
  56
  57 #include "collectd.h"
  58 #include "common.h"
  59 #include "plugin.h"
  60
  61 #define PLUGIN_NAME "turbostat"
  62
  63 static const char *proc_stat = "/proc/stat";
  64 static unsigned int skip_c0;
  65 static unsigned int skip_c1;
  66 static unsigned int do_nhm_cstates;
  67 static unsigned int do_snb_cstates;
  68 static unsigned int do_c8_c9_c10;
  69 static unsigned int do_slm_cstates;
  70 static unsigned int has_aperf;
  71 static unsigned int has_epb;
  72 static unsigned int genuine_intel;
  73 static unsigned int has_invariant_tsc;
  74 static unsigned int do_nehalem_platform_info;
  75 static int do_smi;
  76 static unsigned int do_rapl;
  77 static unsigned int do_dts;
  78 static unsigned int do_ptm;
  79 static unsigned int tcc_activation_temp;
  80 static unsigned int tcc_activation_temp_override;
  81 static double rapl_power_units, rapl_energy_units, rapl_time_units;
  82 static double rapl_joule_counter_range;
  83
  84 #define RAPL_PKG                (1 << 0)
  85                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  86                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  87 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  88                                         /* 0x613 MSR_PKG_PERF_STATUS */
  89 #define RAPL_PKG_POWER_INFO     (1 << 2)
  90                                         /* 0x614 MSR_PKG_POWER_INFO */
  91
  92 #define RAPL_DRAM               (1 << 3)
  93                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
  94                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
  95                                         /* 0x61c MSR_DRAM_POWER_INFO */
  96 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
  97                                         /* 0x61b MSR_DRAM_PERF_STATUS */
  98
  99 #define RAPL_CORES              (1 << 5)
 100                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 101                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 102 #define RAPL_CORE_POLICY        (1 << 6)
 103                                         /* 0x63a MSR_PP0_POLICY */
 104
 105
 106 #define RAPL_GFX                (1 << 7)
 107                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 108                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 109                                         /* 0x642 MSR_PP1_POLICY */
 110 #define TJMAX_DEFAULT   100
 111
 112 int aperf_mperf_unstable;
 113 int backwards_count;
 114 char *progname;
 115
 116 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 117 size_t cpu_present_setsize, cpu_affinity_setsize;
 118
 119 struct thread_data {
 120         unsigned long long tsc;
 121         unsigned long long aperf;
 122         unsigned long long mperf;
 123         unsigned long long c1;
 124         unsigned int smi_count;
 125         unsigned int cpu_id;
 126         unsigned int flags;
 127 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 128 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 129 } *thread_even, *thread_odd;
 130
 131 struct core_data {
 132         unsigned long long c3;
 133         unsigned long long c6;
 134         unsigned long long c7;
 135         unsigned int core_temp_c;
 136         unsigned int core_id;
 137 } *core_even, *core_odd;
 138
 139 struct pkg_data {
 140         unsigned long long pc2;
 141         unsigned long long pc3;
 142         unsigned long long pc6;
 143         unsigned long long pc7;
 144         unsigned long long pc8;
 145         unsigned long long pc9;
 146         unsigned long long pc10;
 147         unsigned int package_id;
 148         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 149         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 150         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 151         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 152         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 153         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 154         unsigned int pkg_temp_c;
 155
 156 } *package_even, *package_odd;
 157
 158 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 159 #define EVEN_COUNTERS thread_even, core_even, package_even
 160 static _Bool is_even = 1;
 161
 162 static _Bool allocated = 0;
 163 static _Bool initialized = 0;
 164
 165 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 166         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 167                 topo.num_threads_per_core + \
 168                 (core_no) * topo.num_threads_per_core + (thread_no))
 169 #define GET_CORE(core_base, core_no, pkg_no) \
 170         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 171 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 172
 173 struct topo_params {
 174         int num_packages;
 175         int num_cpus;
 176         int num_cores;
 177         int max_cpu_num;
 178         int num_cores_per_pkg;
 179         int num_threads_per_core;
 180 } topo;
 181
 182 struct timeval tv_even, tv_odd, tv_delta;
 183
 184 enum return_values {
 185         OK = 0,
 186         ERR_CPU_MIGRATE,
 187         ERR_MSR_IA32_APERF,
 188         ERR_MSR_IA32_MPERF,
 189         ERR_MSR_SMI_COUNT,
 190         ERR_MSR_CORE_C3_RESIDENCY,
 191         ERR_MSR_CORE_C6_RESIDENCY,
 192         ERR_MSR_CORE_C7_RESIDENCY,
 193         ERR_MSR_IA32_THERM_STATUS,
 194         ERR_MSR_PKG_C3_RESIDENCY,
 195         ERR_MSR_PKG_C6_RESIDENCY,
 196         ERR_MSR_PKG_C2_RESIDENCY,
 197         ERR_MSR_PKG_C7_RESIDENCY,
 198         ERR_MSR_PKG_C8_RESIDENCY,
 199         ERR_MSR_PKG_C9_RESIDENCY,
 200         ERR_MSR_PKG_C10_RESIDENCY,
 201         ERR_MSR_PKG_ENERGY_STATUS,
 202         ERR_MSR_PP0_ENERGY_STATUS,
 203         ERR_MSR_DRAM_ENERGY_STATUS,
 204         ERR_MSR_PP1_ENERGY_STATUS,
 205         ERR_MSR_PKG_PERF_STATUS,
 206         ERR_MSR_DRAM_PERF_STATUS,
 207         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 208         ERR_CPU_NOT_PRESENT,
 209         ERR_NO_MSR,
 210         ERR_CANT_OPEN_FILE,
 211         ERR_CANT_READ_NUMBER,
 212         ERR_CANT_READ_PROC_STAT,
 213         ERR_NO_INVARIANT_TSC,
 214         ERR_NO_APERF,
 215         ERR_CALLOC,
 216         ERR_CPU_ALLOC,
 217         ERR_NOT_ROOT,
 218 };
 219
 220 static int setup_all_buffers(void);
 221
 222 static int
 223 cpu_is_not_present(int cpu)
 224 {
 225         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 226 }
 227 /*
 228  * run func(thread, core, package) in topology order
 229  * skip non-present cpus
 230  */
 231
 232 static int __attribute__((warn_unused_result))
 233 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 234         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 235 {
 236         int retval, pkg_no, core_no, thread_no;
 237
 238         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 239                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 240                         for (thread_no = 0; thread_no <
 241                                 topo.num_threads_per_core; ++thread_no) {
 242                                 struct thread_data *t;
 243                                 struct core_data *c;
 244                                 struct pkg_data *p;
 245
 246                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 247
 248                                 if (cpu_is_not_present(t->cpu_id))
 249                                         continue;
 250
 251                                 c = GET_CORE(core_base, core_no, pkg_no);
 252                                 p = GET_PKG(pkg_base, pkg_no);
 253
 254                                 retval = func(t, c, p);
 255                                 if (retval)
 256                                         return retval;
 257                         }
 258                 }
 259         }
 260         return 0;
 261 }
 262
 263 static int __attribute__((warn_unused_result))
 264 cpu_migrate(int cpu)
 265 {
 266         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 267         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 268         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 269                 return -ERR_CPU_MIGRATE;
 270         else
 271                 return 0;
 272 }
 273
 274 static int __attribute__((warn_unused_result))
 275 get_msr(int cpu, off_t offset, unsigned long long *msr)
 276 {
 277         ssize_t retval;
 278         char pathname[32];
 279         int fd;
 280
 281         ssnprintf(pathname, 32, "/dev/cpu/%d/msr", cpu);
 282         fd = open(pathname, O_RDONLY);
 283         if (fd < 0)
 284                 return -1;
 285
 286         retval = pread(fd, msr, sizeof *msr, offset);
 287         close(fd);
 288
 289         if (retval != sizeof *msr) {
 290                 ERROR ("%s offset 0x%llx read failed\n", pathname, (unsigned long long)offset);
 291                 return -1;
 292         }
 293
 294         return 0;
 295 }
 296
 297 #define DELTA_WRAP32(new, old)                  \
 298         if (new > old) {                        \
 299                 old = new - old;                \
 300         } else {                                \
 301                 old = 0x100000000 + new - old;  \
 302         }
 303
 304 static void
 305 delta_package(struct pkg_data *new, struct pkg_data *old)
 306 {
 307         old->pc2 = new->pc2 - old->pc2;
 308         old->pc3 = new->pc3 - old->pc3;
 309         old->pc6 = new->pc6 - old->pc6;
 310         old->pc7 = new->pc7 - old->pc7;
 311         old->pc8 = new->pc8 - old->pc8;
 312         old->pc9 = new->pc9 - old->pc9;
 313         old->pc10 = new->pc10 - old->pc10;
 314         old->pkg_temp_c = new->pkg_temp_c;
 315
 316         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 317         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 318         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 319         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 320         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 321         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 322 }
 323
 324 static void
 325 delta_core(struct core_data *new, struct core_data *old)
 326 {
 327         old->c3 = new->c3 - old->c3;
 328         old->c6 = new->c6 - old->c6;
 329         old->c7 = new->c7 - old->c7;
 330         old->core_temp_c = new->core_temp_c;
 331 }
 332
 333 /*
 334  * old = new - old
 335  */
 336 static int __attribute__((warn_unused_result))
 337 delta_thread(struct thread_data *new, struct thread_data *old,
 338         struct core_data *core_delta)
 339 {
 340         old->tsc = new->tsc - old->tsc;
 341
 342         /* check for TSC < 1 Mcycles over interval */
 343         if (old->tsc < (1000 * 1000)) {
 344                 WARNING("Insanely slow TSC rate, TSC stops in idle?\n"
 345                         "You can disable all c-states by booting with \"idle=poll\"\n"
 346                         "or just the deep ones with \"processor.max_cstate=1\"");
 347                 return -1;
 348         }
 349
 350         old->c1 = new->c1 - old->c1;
 351
 352         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 353                 old->aperf = new->aperf - old->aperf;
 354                 old->mperf = new->mperf - old->mperf;
 355         } else {
 356
 357                 if (!aperf_mperf_unstable) {
 358                         WARNING("%s: APERF or MPERF went backwards *\n", progname);
 359                         WARNING("* Frequency results do not cover entire interval *\n");
 360                         WARNING("* fix this by running Linux-2.6.30 or later *\n");
 361
 362                         aperf_mperf_unstable = 1;
 363                 }
 364                 /*
 365                  * mperf delta is likely a huge "positive" number
 366                  * can not use it for calculating c0 time
 367                  */
 368                 skip_c0 = 1;
 369                 skip_c1 = 1;
 370         }
 371
 372
 373         /*
 374          * As counter collection is not atomic,
 375          * it is possible for mperf's non-halted cycles + idle states
 376          * to exceed TSC's all cycles: show c1 = 0% in that case.
 377          */
 378         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 379                 old->c1 = 0;
 380         else {
 381                 /* normal case, derive c1 */
 382                 old->c1 = old->tsc - old->mperf - core_delta->c3
 383                         - core_delta->c6 - core_delta->c7;
 384         }
 385
 386         if (old->mperf == 0) {
 387                 WARNING("cpu%d MPERF 0!\n", old->cpu_id);
 388                 old->mperf = 1; /* divide by 0 protection */
 389         }
 390
 391         if (do_smi)
 392                 old->smi_count = new->smi_count - old->smi_count;
 393
 394         return 0;
 395 }
 396
 397 static int __attribute__((warn_unused_result))
 398 delta_cpu(struct thread_data *t, struct core_data *c,
 399         struct pkg_data *p, struct thread_data *t2,
 400         struct core_data *c2, struct pkg_data *p2)
 401 {
 402         int ret;
 403
 404         /* calculate core delta only for 1st thread in core */
 405         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 406                 delta_core(c, c2);
 407
 408         /* always calculate thread delta */
 409         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 410         if (ret != 0)
 411                 return ret;
 412
 413         /* calculate package delta only for 1st core in package */
 414         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 415                 delta_package(p, p2);
 416
 417         return 0;
 418 }
 419
 420 static unsigned long long
 421 rdtsc(void)
 422 {
 423         unsigned int low, high;
 424
 425         asm volatile("rdtsc" : "=a" (low), "=d" (high));
 426
 427         return low | ((unsigned long long)high) << 32;
 428 }
 429
 430
 431 /*
 432  * get_counters(...)
 433  * migrate to cpu
 434  * acquire and record local counters for that cpu
 435  */
 436 static int __attribute__((warn_unused_result))
 437 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 438 {
 439         int cpu = t->cpu_id;
 440         unsigned long long msr;
 441
 442         if (cpu_migrate(cpu)) {
 443                 WARNING("Could not migrate to CPU %d\n", cpu);
 444                 return -ERR_CPU_MIGRATE;
 445         }
 446
 447         t->tsc = rdtsc();       /* we are running on local CPU of interest */
 448
 449         if (has_aperf) {
 450                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
 451                         return -ERR_MSR_IA32_APERF;
 452                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
 453                         return -ERR_MSR_IA32_MPERF;
 454         }
 455
 456         if (do_smi) {
 457                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
 458                         return -ERR_MSR_SMI_COUNT;
 459                 t->smi_count = msr & 0xFFFFFFFF;
 460         }
 461
 462         /* collect core counters only for 1st thread in core */
 463         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 464                 return 0;
 465
 466         if (do_nhm_cstates && !do_slm_cstates) {
 467                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
 468                         return -ERR_MSR_CORE_C3_RESIDENCY;
 469         }
 470
 471         if (do_nhm_cstates) {
 472                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
 473                         return -ERR_MSR_CORE_C6_RESIDENCY;
 474         }
 475
 476         if (do_snb_cstates)
 477                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
 478                         return -ERR_MSR_CORE_C7_RESIDENCY;
 479
 480         if (do_dts) {
 481                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
 482                         return -ERR_MSR_IA32_THERM_STATUS;
 483                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 484         }
 485
 486
 487         /* collect package counters only for 1st core in package */
 488         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 489                 return 0;
 490
 491         if (do_nhm_cstates && !do_slm_cstates) {
 492                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
 493                         return -ERR_MSR_PKG_C3_RESIDENCY;
 494                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
 495                         return -ERR_MSR_PKG_C6_RESIDENCY;
 496         }
 497         if (do_snb_cstates) {
 498                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
 499                         return -ERR_MSR_PKG_C2_RESIDENCY;
 500                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
 501                         return -ERR_MSR_PKG_C7_RESIDENCY;
 502         }
 503         if (do_c8_c9_c10) {
 504                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
 505                         return -ERR_MSR_PKG_C8_RESIDENCY;
 506                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
 507                         return -ERR_MSR_PKG_C9_RESIDENCY;
 508                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
 509                         return -ERR_MSR_PKG_C10_RESIDENCY;
 510         }
 511         if (do_rapl & RAPL_PKG) {
 512                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
 513                         return -ERR_MSR_PKG_ENERGY_STATUS;
 514                 p->energy_pkg = msr & 0xFFFFFFFF;
 515         }
 516         if (do_rapl & RAPL_CORES) {
 517                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
 518                         return MSR_PP0_ENERGY_STATUS;
 519                 p->energy_cores = msr & 0xFFFFFFFF;
 520         }
 521         if (do_rapl & RAPL_DRAM) {
 522                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
 523                         return -ERR_MSR_DRAM_ENERGY_STATUS;
 524                 p->energy_dram = msr & 0xFFFFFFFF;
 525         }
 526         if (do_rapl & RAPL_GFX) {
 527                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
 528                         return -ERR_MSR_PP1_ENERGY_STATUS;
 529                 p->energy_gfx = msr & 0xFFFFFFFF;
 530         }
 531         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 532                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
 533                         return -ERR_MSR_PKG_PERF_STATUS;
 534                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 535         }
 536         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 537                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
 538                         return -ERR_MSR_DRAM_PERF_STATUS;
 539                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 540         }
 541         if (do_ptm) {
 542                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
 543                         return -ERR_MSR_IA32_PACKAGE_THERM_STATUS;
 544                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 545         }
 546         return 0;
 547 }
 548
 549 static void
 550 free_all_buffers(void)
 551 {
 552         allocated = 0;
 553         initialized = 0;
 554
 555         CPU_FREE(cpu_present_set);
 556         cpu_present_set = NULL;
 557         cpu_present_set = 0;
 558
 559         CPU_FREE(cpu_affinity_set);
 560         cpu_affinity_set = NULL;
 561         cpu_affinity_setsize = 0;
 562
 563         free(thread_even);
 564         free(core_even);
 565         free(package_even);
 566
 567         thread_even = NULL;
 568         core_even = NULL;
 569         package_even = NULL;
 570
 571         free(thread_odd);
 572         free(core_odd);
 573         free(package_odd);
 574
 575         thread_odd = NULL;
 576         core_odd = NULL;
 577         package_odd = NULL;
 578 }
 579
 580 /*
 581  * Parse a file containing a single int.
 582  */
 583 static int __attribute__ ((format(printf,1,2)))
 584 parse_int_file(const char *fmt, ...)
 585 {
 586         va_list args;
 587         char path[PATH_MAX];
 588         FILE *filep;
 589         int value;
 590
 591         va_start(args, fmt);
 592         vsnprintf(path, sizeof(path), fmt, args);
 593         va_end(args);
 594         filep = fopen(path, "r");
 595         if (!filep) {
 596                 ERROR("%s: open failed", path);
 597                 return -ERR_CANT_OPEN_FILE;
 598         }
 599         if (fscanf(filep, "%d", &value) != 1) {
 600                 ERROR("%s: failed to parse number from file", path);
 601                 return -ERR_CANT_READ_NUMBER;
 602         }
 603         fclose(filep);
 604         return value;
 605 }
 606
 607 /*
 608  * cpu_is_first_sibling_in_core(cpu)
 609  * return 1 if given CPU is 1st HT sibling in the core
 610  */
 611 static int
 612 cpu_is_first_sibling_in_core(int cpu)
 613 {
 614         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 615 }
 616
 617 /*
 618  * cpu_is_first_core_in_package(cpu)
 619  * return 1 if given CPU is 1st core in package
 620  */
 621 static int
 622 cpu_is_first_core_in_package(int cpu)
 623 {
 624         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 625 }
 626
 627 static int
 628 get_physical_package_id(int cpu)
 629 {
 630         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 631 }
 632
 633 static int
 634 get_core_id(int cpu)
 635 {
 636         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 637 }
 638
 639 static int
 640 get_num_ht_siblings(int cpu)
 641 {
 642         char path[80];
 643         FILE *filep;
 644         int sib1, sib2;
 645         int matches;
 646         char character;
 647
 648         ssnprintf(path, 80, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 649         filep = fopen(path, "r");
 650         if (!filep) {
 651                 ERROR("%s: open failed", path);
 652                 return -ERR_CANT_OPEN_FILE;
 653         }
 654         /*
 655          * file format:
 656          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 657          * otherwinse 1 sibling (self).
 658          */
 659         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 660
 661         fclose(filep);
 662
 663         if (matches == 3)
 664                 return 2;
 665         else
 666                 return 1;
 667 }
 668
 669 /*
 670  * run func(thread, core, package) in topology order
 671  * skip non-present cpus
 672  */
 673
 674
 675 static int __attribute__((warn_unused_result))
 676 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 677         struct pkg_data *, struct thread_data *, struct core_data *,
 678         struct pkg_data *), struct thread_data *thread_base,
 679         struct core_data *core_base, struct pkg_data *pkg_base,
 680         struct thread_data *thread_base2, struct core_data *core_base2,
 681         struct pkg_data *pkg_base2)
 682 {
 683         int retval, pkg_no, core_no, thread_no;
 684
 685         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 686                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 687                         for (thread_no = 0; thread_no <
 688                                 topo.num_threads_per_core; ++thread_no) {
 689                                 struct thread_data *t, *t2;
 690                                 struct core_data *c, *c2;
 691                                 struct pkg_data *p, *p2;
 692
 693                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 694
 695                                 if (cpu_is_not_present(t->cpu_id))
 696                                         continue;
 697
 698                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 699
 700                                 c = GET_CORE(core_base, core_no, pkg_no);
 701                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 702
 703                                 p = GET_PKG(pkg_base, pkg_no);
 704                                 p2 = GET_PKG(pkg_base2, pkg_no);
 705
 706                                 retval = func(t, c, p, t2, c2, p2);
 707                                 if (retval)
 708                                         return retval;
 709                         }
 710                 }
 711         }
 712         return 0;
 713 }
 714
 715 /*
 716  * run func(cpu) on every cpu in /proc/stat
 717  * return max_cpu number
 718  */
 719 static int __attribute__((warn_unused_result))
 720 for_all_proc_cpus(int (func)(int))
 721 {
 722         FILE *fp;
 723         int cpu_num;
 724         int retval;
 725
 726         fp = fopen(proc_stat, "r");
 727         if (!fp) {
 728                 ERROR("%s: open failed", proc_stat);
 729                 return -ERR_CANT_OPEN_FILE;
 730         }
 731
 732         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 733         if (retval != 0) {
 734                 ERROR("%s: failed to parse format", proc_stat);
 735                 return -ERR_CANT_READ_PROC_STAT;
 736         }
 737
 738         while (1) {
 739                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 740                 if (retval != 1)
 741                         break;
 742
 743                 retval = func(cpu_num);
 744                 if (retval) {
 745                         fclose(fp);
 746                         return(retval);
 747                 }
 748         }
 749         fclose(fp);
 750         return 0;
 751 }
 752
 753 /*
 754  * count_cpus()
 755  * remember the last one seen, it will be the max
 756  */
 757 static int
 758 count_cpus(int cpu)
 759 {
 760         if (topo.max_cpu_num < cpu)
 761                 topo.max_cpu_num = cpu;
 762
 763         topo.num_cpus += 1;
 764         return 0;
 765 }
 766 static int
 767 mark_cpu_present(int cpu)
 768 {
 769         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 770         return 0;
 771 }
 772
 773
 774 static void
 775 turbostat_submit (const char *plugin_instance,
 776         const char *type, const char *type_instance,
 777         gauge_t value)
 778 {
 779         value_list_t vl = VALUE_LIST_INIT;
 780         value_t v;
 781
 782         v.gauge = value;
 783         vl.values = &v;
 784         vl.values_len = 1;
 785         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 786         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 787         if (plugin_instance != NULL)
 788                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 789         sstrncpy (vl.type, type, sizeof (vl.type));
 790         if (type_instance != NULL)
 791                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 792
 793         plugin_dispatch_values (&vl);
 794 }
 795
 796 /*
 797  * column formatting convention & formats
 798  * package: "pk" 2 columns %2d
 799  * core: "cor" 3 columns %3d
 800  * CPU: "CPU" 3 columns %3d
 801  * Pkg_W: %6.2
 802  * Cor_W: %6.2
 803  * GFX_W: %5.2
 804  * RAM_W: %5.2
 805  * GHz: "GHz" 3 columns %3.2
 806  * TSC: "TSC" 3 columns %3.2
 807  * SMI: "SMI" 4 columns %4d
 808  * percentage " %pc3" %6.2
 809  * Perf Status percentage: %5.2
 810  * "CTMP" 4 columns %4d
 811  */
 812 #define NAME_LEN 12
 813 static int
 814 submit_counters(struct thread_data *t, struct core_data *c,
 815         struct pkg_data *p)
 816 {
 817         char name[NAME_LEN];
 818         double interval_float;
 819
 820         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 821
 822         snprintf(name, NAME_LEN, "cpu%02d", t->cpu_id);
 823
 824         if (do_nhm_cstates) {
 825                 if (!skip_c0)
 826                         turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 827                 if (!skip_c1)
 828                         turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 829         }
 830
 831         /* GHz */
 832         if (has_aperf && ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc))))
 833                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 834
 835         /* SMI */
 836         if (do_smi)
 837                 turbostat_submit(NULL, "current", name, t->smi_count);
 838
 839         /* print per-core data only for 1st thread in core */
 840         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 841                 goto done;
 842
 843         snprintf(name, NAME_LEN, "core%02d", c->core_id);
 844
 845         if (do_nhm_cstates && !do_slm_cstates)
 846                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 847         if (do_nhm_cstates)
 848                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 849         if (do_snb_cstates)
 850                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 851
 852         if (do_dts)
 853                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 854
 855         /* print per-package data only for 1st core in package */
 856         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 857                 goto done;
 858
 859         snprintf(name, NAME_LEN, "pkg%02d", p->package_id);
 860
 861         if (do_ptm)
 862                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 863
 864         if (do_snb_cstates)
 865                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 866         if (do_nhm_cstates && !do_slm_cstates)
 867                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 868         if (do_nhm_cstates && !do_slm_cstates)
 869                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 870         if (do_snb_cstates)
 871                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 872         if (do_c8_c9_c10) {
 873                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 874                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 875                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 876         }
 877
 878         if (do_rapl) {
 879                 if (do_rapl & RAPL_PKG)
 880                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 881                 if (do_rapl & RAPL_CORES)
 882                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 883                 if (do_rapl & RAPL_GFX)
 884                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 885                 if (do_rapl & RAPL_DRAM)
 886                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 887         }
 888 done:
 889         return 0;
 890 }
 891
 892 static int
 893 turbostat_read(user_data_t * not_used)
 894 {
 895         int ret;
 896
 897         if (!allocated) {
 898                 if ((ret = setup_all_buffers()) < 0)
 899                         return ret;
 900         }
 901
 902         if (for_all_proc_cpus(cpu_is_not_present)) {
 903                 free_all_buffers();
 904                 if ((ret = setup_all_buffers()) < 0)
 905                         return ret;
 906                 if (for_all_proc_cpus(cpu_is_not_present))
 907                         return -ERR_CPU_NOT_PRESENT;
 908         }
 909
 910         if (!initialized) {
 911                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 912                         return ret;
 913                 gettimeofday(&tv_even, (struct timezone *)NULL);
 914                 is_even = 1;
 915                 initialized = 1;
 916                 return 0;
 917         }
 918
 919         if (is_even) {
 920                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 921                         return ret;
 922                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 923                 is_even = 0;
 924                 timersub(&tv_odd, &tv_even, &tv_delta);
 925                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 926                         return ret;
 927                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 928                         return ret;
 929         } else {
 930                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 931                         return ret;
 932                 gettimeofday(&tv_even, (struct timezone *)NULL);
 933                 is_even = 1;
 934                 timersub(&tv_even, &tv_odd, &tv_delta);
 935                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 936                         return ret;
 937                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 938                         return ret;
 939         }
 940         return 0;
 941 }
 942
 943 static int __attribute__((warn_unused_result))
 944 check_dev_msr()
 945 {
 946         struct stat sb;
 947
 948         if (stat("/dev/cpu/0/msr", &sb)) {
 949                 ERROR("no /dev/cpu/0/msr\n"
 950                         "Try \"# modprobe msr\"");
 951                 return -ERR_NO_MSR;
 952         }
 953         return 0;
 954 }
 955
 956 static int __attribute__((warn_unused_result))
 957 check_super_user()
 958 {
 959         if (getuid() != 0) {
 960                 ERROR("must be root");
 961                 return -ERR_NOT_ROOT;
 962         }
 963         return 0;
 964 }
 965
 966
 967 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
 968 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
 969
 970 static double
 971 get_tdp(unsigned int model)
 972 {
 973         unsigned long long msr;
 974
 975         if (do_rapl & RAPL_PKG_POWER_INFO)
 976                 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
 977                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
 978
 979         switch (model) {
 980         case 0x37:
 981         case 0x4D:
 982                 return 30.0;
 983         default:
 984                 return 135.0;
 985         }
 986 }
 987
 988
 989 /*
 990  * rapl_probe()
 991  *
 992  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
 993  */
 994 static void
 995 rapl_probe(unsigned int family, unsigned int model)
 996 {
 997         unsigned long long msr;
 998         unsigned int time_unit;
 999         double tdp;
1000
1001         if (!genuine_intel)
1002                 return;
1003
1004         if (family != 6)
1005                 return;
1006
1007         switch (model) {
1008         case 0x2A:
1009         case 0x3A:
1010         case 0x3C:      /* HSW */
1011         case 0x45:      /* HSW */
1012         case 0x46:      /* HSW */
1013                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1014                 break;
1015         case 0x3F:      /* HSX */
1016                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1017                 break;
1018         case 0x2D:
1019         case 0x3E:
1020                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1021                 break;
1022         case 0x37:      /* BYT */
1023         case 0x4D:      /* AVN */
1024                 do_rapl = RAPL_PKG | RAPL_CORES ;
1025                 break;
1026         default:
1027                 return;
1028         }
1029
1030         /* units on package 0, verify later other packages match */
1031         if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1032                 return;
1033
1034         rapl_power_units = 1.0 / (1 << (msr & 0xF));
1035         if (model == 0x37)
1036                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1037         else
1038                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1039
1040         time_unit = msr >> 16 & 0xF;
1041         if (time_unit == 0)
1042                 time_unit = 0xA;
1043
1044         rapl_time_units = 1.0 / (1 << (time_unit));
1045
1046         tdp = get_tdp(model);
1047
1048         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1049 //      if (verbose)
1050 //              fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1051
1052         return;
1053 }
1054
1055 static int
1056 is_snb(unsigned int family, unsigned int model)
1057 {
1058         if (!genuine_intel)
1059                 return 0;
1060
1061         switch (model) {
1062         case 0x2A:
1063         case 0x2D:
1064         case 0x3A:      /* IVB */
1065         case 0x3E:      /* IVB Xeon */
1066         case 0x3C:      /* HSW */
1067         case 0x3F:      /* HSW */
1068         case 0x45:      /* HSW */
1069         case 0x46:      /* HSW */
1070                 return 1;
1071         }
1072         return 0;
1073 }
1074
1075 static int
1076 has_c8_c9_c10(unsigned int family, unsigned int model)
1077 {
1078         if (!genuine_intel)
1079                 return 0;
1080
1081         switch (model) {
1082         case 0x45:
1083                 return 1;
1084         }
1085         return 0;
1086 }
1087
1088
1089 static int
1090 is_slm(unsigned int family, unsigned int model)
1091 {
1092         if (!genuine_intel)
1093                 return 0;
1094         switch (model) {
1095         case 0x37:      /* BYT */
1096         case 0x4D:      /* AVN */
1097                 return 1;
1098         }
1099         return 0;
1100 }
1101
1102 /*
1103  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1104  * the Thermal Control Circuit (TCC) activates.
1105  * This is usually equal to tjMax.
1106  *
1107  * Older processors do not have this MSR, so there we guess,
1108  * but also allow cmdline over-ride with -T.
1109  *
1110  * Several MSR temperature values are in units of degrees-C
1111  * below this value, including the Digital Thermal Sensor (DTS),
1112  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1113  */
1114 static int __attribute__((warn_unused_result))
1115 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1116 {
1117         unsigned long long msr;
1118         unsigned int target_c_local;
1119         int cpu;
1120
1121         /* tcc_activation_temp is used only for dts or ptm */
1122         if (!(do_dts || do_ptm))
1123                 return 0;
1124
1125         /* this is a per-package concept */
1126         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1127                 return 0;
1128
1129         cpu = t->cpu_id;
1130         if (cpu_migrate(cpu)) {
1131                 ERROR("Could not migrate to CPU %d\n", cpu);
1132                 return -ERR_CPU_MIGRATE;
1133         }
1134
1135         if (tcc_activation_temp_override != 0) {
1136                 tcc_activation_temp = tcc_activation_temp_override;
1137                 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
1138                         cpu, tcc_activation_temp);
1139                 return 0;
1140         }
1141
1142         /* Temperature Target MSR is Nehalem and newer only */
1143         if (!do_nehalem_platform_info)
1144                 goto guess;
1145
1146         if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
1147                 goto guess;
1148
1149         target_c_local = (msr >> 16) & 0x7F;
1150
1151         if (target_c_local < 85 || target_c_local > 127)
1152                 goto guess;
1153
1154         tcc_activation_temp = target_c_local;
1155
1156         return 0;
1157
1158 guess:
1159         tcc_activation_temp = TJMAX_DEFAULT;
1160         WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1161                 cpu, tcc_activation_temp);
1162
1163         return 0;
1164 }
1165
1166 static int __attribute__((warn_unused_result))
1167 check_cpuid()
1168 {
1169         unsigned int eax, ebx, ecx, edx, max_level;
1170         unsigned int fms, family, model;
1171
1172         eax = ebx = ecx = edx = 0;
1173
1174         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1175
1176         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1177                 genuine_intel = 1;
1178
1179         fms = 0;
1180         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1181         family = (fms >> 8) & 0xf;
1182         model = (fms >> 4) & 0xf;
1183         if (family == 6 || family == 0xf)
1184                 model += ((fms >> 16) & 0xf) << 4;
1185
1186         if (!(edx & (1 << 5))) {
1187                 ERROR("CPUID: no MSR");
1188                 return -ERR_NO_MSR;
1189         }
1190
1191         /*
1192          * check max extended function levels of CPUID.
1193          * This is needed to check for invariant TSC.
1194          * This check is valid for both Intel and AMD.
1195          */
1196         ebx = ecx = edx = 0;
1197         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1198
1199         if (max_level < 0x80000007) {
1200                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1201                 return -ERR_NO_INVARIANT_TSC;
1202         }
1203
1204         /*
1205          * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1206          * this check is valid for both Intel and AMD
1207          */
1208         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1209         has_invariant_tsc = edx & (1 << 8);
1210
1211         if (!has_invariant_tsc) {
1212                 ERROR("No invariant TSC");
1213                 return -ERR_NO_INVARIANT_TSC;
1214         }
1215
1216         /*
1217          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1218          * this check is valid for both Intel and AMD
1219          */
1220
1221         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1222         has_aperf = ecx & (1 << 0);
1223         do_dts = eax & (1 << 0);
1224         do_ptm = eax & (1 << 6);
1225         has_epb = ecx & (1 << 3);
1226
1227         if (!has_aperf) {
1228                 ERROR("No APERF");
1229                 return -ERR_NO_APERF;
1230         }
1231
1232         do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1233         do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */
1234         do_smi = do_nhm_cstates;
1235         do_snb_cstates = is_snb(family, model);
1236         do_c8_c9_c10 = has_c8_c9_c10(family, model);
1237         do_slm_cstates = is_slm(family, model);
1238
1239         rapl_probe(family, model);
1240
1241         return 0;
1242 }
1243
1244
1245
1246 static int __attribute__((warn_unused_result))
1247 topology_probe()
1248 {
1249         int i;
1250         int ret;
1251         int max_core_id = 0;
1252         int max_package_id = 0;
1253         int max_siblings = 0;
1254         struct cpu_topology {
1255                 int core_id;
1256                 int physical_package_id;
1257         } *cpus;
1258
1259         /* Initialize num_cpus, max_cpu_num */
1260         topo.num_cpus = 0;
1261         topo.max_cpu_num = 0;
1262         ret = for_all_proc_cpus(count_cpus);
1263         if (ret < 0)
1264                 return ret;
1265
1266         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1267
1268         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1269         if (cpus == NULL) {
1270                 ERROR("calloc cpus");
1271                 return -ERR_CALLOC;
1272         }
1273
1274         /*
1275          * Allocate and initialize cpu_present_set
1276          */
1277         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1278         if (cpu_present_set == NULL) {
1279                 free(cpus);
1280                 ERROR("CPU_ALLOC");
1281                 return -ERR_CPU_ALLOC;
1282         }
1283         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1284         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1285         ret = for_all_proc_cpus(mark_cpu_present);
1286         if (ret < 0) {
1287                 free(cpus);
1288                 return ret;
1289         }
1290
1291         /*
1292          * Allocate and initialize cpu_affinity_set
1293          */
1294         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1295         if (cpu_affinity_set == NULL) {
1296                 free(cpus);
1297                 ERROR("CPU_ALLOC");
1298                 return -ERR_CPU_ALLOC;
1299         }
1300         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1301         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1302
1303
1304         /*
1305          * For online cpus
1306          * find max_core_id, max_package_id
1307          */
1308         for (i = 0; i <= topo.max_cpu_num; ++i) {
1309                 int siblings;
1310
1311                 if (cpu_is_not_present(i)) {
1312                         //if (verbose > 1)
1313                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1314                         continue;
1315                 }
1316                 cpus[i].core_id = get_core_id(i);
1317                 if (cpus[i].core_id < 0)
1318                         return cpus[i].core_id;
1319                 if (cpus[i].core_id > max_core_id)
1320                         max_core_id = cpus[i].core_id;
1321
1322                 cpus[i].physical_package_id = get_physical_package_id(i);
1323                 if (cpus[i].physical_package_id < 0)
1324                         return cpus[i].physical_package_id;
1325                 if (cpus[i].physical_package_id > max_package_id)
1326                         max_package_id = cpus[i].physical_package_id;
1327
1328                 siblings = get_num_ht_siblings(i);
1329                 if (siblings < 0)
1330                         return siblings;
1331                 if (siblings > max_siblings)
1332                         max_siblings = siblings;
1333                 DEBUG("cpu %d pkg %d core %d\n",
1334                         i, cpus[i].physical_package_id, cpus[i].core_id);
1335         }
1336         topo.num_cores_per_pkg = max_core_id + 1;
1337         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1338                 max_core_id, topo.num_cores_per_pkg);
1339
1340         topo.num_packages = max_package_id + 1;
1341         DEBUG("max_package_id %d, sizing for %d packages\n",
1342                 max_package_id, topo.num_packages);
1343
1344         topo.num_threads_per_core = max_siblings;
1345         DEBUG("max_siblings %d\n", max_siblings);
1346
1347         free(cpus);
1348         return 0;
1349 }
1350
1351 static int
1352 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1353 {
1354         int i;
1355
1356         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1357                 topo.num_packages, sizeof(struct thread_data));
1358         if (*t == NULL)
1359                 goto error;
1360
1361         for (i = 0; i < topo.num_threads_per_core *
1362                 topo.num_cores_per_pkg * topo.num_packages; i++)
1363                 (*t)[i].cpu_id = -1;
1364
1365         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1366                 sizeof(struct core_data));
1367         if (*c == NULL)
1368                 goto error;
1369
1370         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1371                 (*c)[i].core_id = -1;
1372
1373         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1374         if (*p == NULL)
1375                 goto error;
1376
1377         for (i = 0; i < topo.num_packages; i++)
1378                 (*p)[i].package_id = i;
1379
1380         return 0;
1381 error:
1382         ERROR("calloc counters");
1383         return -ERR_CALLOC;
1384 }
1385 /*
1386  * init_counter()
1387  *
1388  * set cpu_id, core_num, pkg_num
1389  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1390  *
1391  * increment topo.num_cores when 1st core in pkg seen
1392  */
1393 static int
1394 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1395         struct pkg_data *pkg_base, int thread_num, int core_num,
1396         int pkg_num, int cpu_id)
1397 {
1398         int ret;
1399         struct thread_data *t;
1400         struct core_data *c;
1401         struct pkg_data *p;
1402
1403         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1404         c = GET_CORE(core_base, core_num, pkg_num);
1405         p = GET_PKG(pkg_base, pkg_num);
1406
1407         t->cpu_id = cpu_id;
1408         if (thread_num == 0) {
1409                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1410                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1411                         return ret;
1412                 } else if (ret != 0) {
1413                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1414                 }
1415         }
1416
1417         c->core_id = core_num;
1418         p->package_id = pkg_num;
1419
1420         return 0;
1421 }
1422
1423
1424 static int
1425 initialize_counters(int cpu_id)
1426 {
1427         int my_thread_id, my_core_id, my_package_id;
1428         int ret;
1429
1430         my_package_id = get_physical_package_id(cpu_id);
1431         if (my_package_id < 0)
1432                 return my_package_id;
1433         my_core_id = get_core_id(cpu_id);
1434         if (my_core_id < 0)
1435                 return my_core_id;
1436
1437         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1438                 return ret;
1439         } else if (ret != 0) {
1440                 my_thread_id = 0;
1441                 topo.num_cores++;
1442         } else {
1443                 my_thread_id = 1;
1444         }
1445
1446         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1447         if (ret < 0)
1448                 return ret;
1449         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1450         if (ret < 0)
1451                 return ret;
1452         return 0;
1453 }
1454
1455 #define DO_OR_GOTO_ERR(something) \
1456 do {                         \
1457         ret = (something);     \
1458         if (ret < 0)         \
1459                 goto err;    \
1460 } while (0)
1461
1462 static int setup_all_buffers(void)
1463 {
1464         int ret;
1465
1466         DO_OR_GOTO_ERR(topology_probe());
1467         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1468         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1469         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1470
1471         allocated = 1;
1472         return 0;
1473 err:
1474         free_all_buffers();
1475         return ret;
1476 }
1477
1478 static int
1479 turbostat_init(void)
1480 {
1481         int ret;
1482
1483         DO_OR_GOTO_ERR(check_cpuid());
1484         DO_OR_GOTO_ERR(check_dev_msr());
1485         DO_OR_GOTO_ERR(check_super_user());
1486         DO_OR_GOTO_ERR(setup_all_buffers());
1487         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1488
1489         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1490
1491         return 0;
1492 err:
1493         free_all_buffers();
1494         return ret;
1495 }
1496
1497 void module_register(void);
1498 void module_register(void)
1499 {
1500         plugin_register_init(PLUGIN_NAME, turbostat_init);
1501 }