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_core_cstate;
  67 static unsigned int do_pkg_cstate;
  68 static unsigned int do_rapl;
  69 static unsigned int do_dts;
  70 static unsigned int do_ptm;
  71 static unsigned int tcc_activation_temp;
  72 static unsigned int tcc_activation_temp_override;
  73 static double rapl_energy_units;
  74
  75 #define RAPL_PKG                (1 << 0)
  76                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  77                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  78 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  79                                         /* 0x613 MSR_PKG_PERF_STATUS */
  80 #define RAPL_PKG_POWER_INFO     (1 << 2)
  81                                         /* 0x614 MSR_PKG_POWER_INFO */
  82
  83 #define RAPL_DRAM               (1 << 3)
  84                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
  85                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
  86                                         /* 0x61c MSR_DRAM_POWER_INFO */
  87 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
  88                                         /* 0x61b MSR_DRAM_PERF_STATUS */
  89
  90 #define RAPL_CORES              (1 << 5)
  91                                         /* 0x638 MSR_PP0_POWER_LIMIT */
  92                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
  93 #define RAPL_CORE_POLICY        (1 << 6)
  94                                         /* 0x63a MSR_PP0_POLICY */
  95
  96
  97 #define RAPL_GFX                (1 << 7)
  98                                         /* 0x640 MSR_PP1_POWER_LIMIT */
  99                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 100                                         /* 0x642 MSR_PP1_POLICY */
 101 #define TJMAX_DEFAULT   100
 102
 103 int aperf_mperf_unstable;
 104 int backwards_count;
 105 char *progname;
 106
 107 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 108 size_t cpu_present_setsize, cpu_affinity_setsize;
 109
 110 struct thread_data {
 111         unsigned long long tsc;
 112         unsigned long long aperf;
 113         unsigned long long mperf;
 114         unsigned long long c1;
 115         unsigned int smi_count;
 116         unsigned int cpu_id;
 117         unsigned int flags;
 118 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 119 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 120 } *thread_even, *thread_odd;
 121
 122 struct core_data {
 123         unsigned long long c3;
 124         unsigned long long c6;
 125         unsigned long long c7;
 126         unsigned int core_temp_c;
 127         unsigned int core_id;
 128 } *core_even, *core_odd;
 129
 130 struct pkg_data {
 131         unsigned long long pc2;
 132         unsigned long long pc3;
 133         unsigned long long pc6;
 134         unsigned long long pc7;
 135         unsigned long long pc8;
 136         unsigned long long pc9;
 137         unsigned long long pc10;
 138         unsigned int package_id;
 139         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 140         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 141         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 142         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 143         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 144         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 145         unsigned int pkg_temp_c;
 146
 147 } *package_even, *package_odd;
 148
 149 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 150 #define EVEN_COUNTERS thread_even, core_even, package_even
 151 static _Bool is_even = 1;
 152
 153 static _Bool allocated = 0;
 154 static _Bool initialized = 0;
 155
 156 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 157         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 158                 topo.num_threads_per_core + \
 159                 (core_no) * topo.num_threads_per_core + (thread_no))
 160 #define GET_CORE(core_base, core_no, pkg_no) \
 161         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 162 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 163
 164 struct topo_params {
 165         int num_packages;
 166         int num_cpus;
 167         int num_cores;
 168         int max_cpu_num;
 169         int num_cores_per_pkg;
 170         int num_threads_per_core;
 171 } topo;
 172
 173 struct timeval tv_even, tv_odd, tv_delta;
 174
 175 enum return_values {
 176         OK = 0,
 177         ERR_CPU_MIGRATE,
 178         ERR_MSR_IA32_APERF,
 179         ERR_MSR_IA32_MPERF,
 180         ERR_MSR_SMI_COUNT,
 181         ERR_MSR_CORE_C3_RESIDENCY,
 182         ERR_MSR_CORE_C6_RESIDENCY,
 183         ERR_MSR_CORE_C7_RESIDENCY,
 184         ERR_MSR_IA32_THERM_STATUS,
 185         ERR_MSR_PKG_C3_RESIDENCY,
 186         ERR_MSR_PKG_C6_RESIDENCY,
 187         ERR_MSR_PKG_C2_RESIDENCY,
 188         ERR_MSR_PKG_C7_RESIDENCY,
 189         ERR_MSR_PKG_C8_RESIDENCY,
 190         ERR_MSR_PKG_C9_RESIDENCY,
 191         ERR_MSR_PKG_C10_RESIDENCY,
 192         ERR_MSR_PKG_ENERGY_STATUS,
 193         ERR_MSR_PKG_POWER_INFO,
 194         ERR_MSR_PP0_ENERGY_STATUS,
 195         ERR_MSR_DRAM_ENERGY_STATUS,
 196         ERR_MSR_PP1_ENERGY_STATUS,
 197         ERR_MSR_PKG_PERF_STATUS,
 198         ERR_MSR_DRAM_PERF_STATUS,
 199         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 200         ERR_MSR_IA32_TSC,
 201         ERR_CPU_NOT_PRESENT,
 202         ERR_NO_MSR,
 203         ERR_CANT_OPEN_MSR,
 204         ERR_CANT_OPEN_FILE,
 205         ERR_CANT_READ_NUMBER,
 206         ERR_CANT_READ_PROC_STAT,
 207         ERR_NO_INVARIANT_TSC,
 208         ERR_NO_APERF,
 209         ERR_CALLOC,
 210         ERR_CPU_ALLOC,
 211         ERR_NOT_ROOT,
 212         UNSUPPORTED_CPU,
 213 };
 214
 215 static int setup_all_buffers(void);
 216
 217 static int
 218 cpu_is_not_present(int cpu)
 219 {
 220         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 221 }
 222 /*
 223  * run func(thread, core, package) in topology order
 224  * skip non-present cpus
 225  */
 226
 227 static int __attribute__((warn_unused_result))
 228 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 229         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 230 {
 231         int retval, pkg_no, core_no, thread_no;
 232
 233         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 234                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 235                         for (thread_no = 0; thread_no <
 236                                 topo.num_threads_per_core; ++thread_no) {
 237                                 struct thread_data *t;
 238                                 struct core_data *c;
 239                                 struct pkg_data *p;
 240
 241                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 242
 243                                 if (cpu_is_not_present(t->cpu_id))
 244                                         continue;
 245
 246                                 c = GET_CORE(core_base, core_no, pkg_no);
 247                                 p = GET_PKG(pkg_base, pkg_no);
 248
 249                                 retval = func(t, c, p);
 250                                 if (retval)
 251                                         return retval;
 252                         }
 253                 }
 254         }
 255         return 0;
 256 }
 257
 258 static int __attribute__((warn_unused_result))
 259 cpu_migrate(int cpu)
 260 {
 261         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 262         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 263         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 264                 return -ERR_CPU_MIGRATE;
 265         else
 266                 return 0;
 267 }
 268
 269 static int __attribute__((warn_unused_result))
 270 open_msr(int cpu)
 271 {
 272         char pathname[32];
 273         int fd;
 274
 275         /* FIXME: Do we really need this, why? */
 276         if (cpu_migrate(cpu)) {
 277                 ERROR("Could not migrate to CPU %d\n", cpu);
 278                 return -ERR_CPU_MIGRATE;
 279         }
 280
 281         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 282         fd = open(pathname, O_RDONLY);
 283         if (fd < 0)
 284                 return -ERR_CANT_OPEN_MSR;
 285         return fd;
 286 }
 287
 288 static int __attribute__((warn_unused_result))
 289 read_msr(int fd, off_t offset, unsigned long long *msr)
 290 {
 291         ssize_t retval;
 292
 293         retval = pread(fd, msr, sizeof *msr, offset);
 294
 295         if (retval != sizeof *msr) {
 296                 ERROR ("MSR offset 0x%llx read failed", (unsigned long long)offset);
 297                 return -1;
 298         }
 299         return 0;
 300 }
 301
 302 static int __attribute__((warn_unused_result))
 303 get_msr(int cpu, off_t offset, unsigned long long *msr)
 304 {
 305         ssize_t retval;
 306         int fd;
 307
 308         fd = open_msr(cpu);
 309         if (fd < 0)
 310                 return fd;
 311         retval = read_msr(fd, offset, msr);
 312         close(fd);
 313         return retval;
 314 }
 315
 316 #define DELTA_WRAP32(new, old)                  \
 317         if (new > old) {                        \
 318                 old = new - old;                \
 319         } else {                                \
 320                 old = 0x100000000 + new - old;  \
 321         }
 322
 323 static void
 324 delta_package(struct pkg_data *new, struct pkg_data *old)
 325 {
 326         old->pc2 = new->pc2 - old->pc2;
 327         old->pc3 = new->pc3 - old->pc3;
 328         old->pc6 = new->pc6 - old->pc6;
 329         old->pc7 = new->pc7 - old->pc7;
 330         old->pc8 = new->pc8 - old->pc8;
 331         old->pc9 = new->pc9 - old->pc9;
 332         old->pc10 = new->pc10 - old->pc10;
 333         old->pkg_temp_c = new->pkg_temp_c;
 334
 335         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 336         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 337         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 338         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 339         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 340         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 341 }
 342
 343 static void
 344 delta_core(struct core_data *new, struct core_data *old)
 345 {
 346         old->c3 = new->c3 - old->c3;
 347         old->c6 = new->c6 - old->c6;
 348         old->c7 = new->c7 - old->c7;
 349         old->core_temp_c = new->core_temp_c;
 350 }
 351
 352 /*
 353  * old = new - old
 354  */
 355 static int __attribute__((warn_unused_result))
 356 delta_thread(struct thread_data *new, struct thread_data *old,
 357         struct core_data *core_delta)
 358 {
 359         old->tsc = new->tsc - old->tsc;
 360
 361         /* check for TSC < 1 Mcycles over interval */
 362         if (old->tsc < (1000 * 1000)) {
 363                 WARNING("Insanely slow TSC rate, TSC stops in idle?\n"
 364                         "You can disable all c-states by booting with \"idle=poll\"\n"
 365                         "or just the deep ones with \"processor.max_cstate=1\"");
 366                 return -1;
 367         }
 368
 369         old->c1 = new->c1 - old->c1;
 370
 371         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 372                 old->aperf = new->aperf - old->aperf;
 373                 old->mperf = new->mperf - old->mperf;
 374         } else {
 375
 376                 if (!aperf_mperf_unstable) {
 377                         WARNING("%s: APERF or MPERF went backwards *\n", progname);
 378                         WARNING("* Frequency results do not cover entire interval *\n");
 379                         WARNING("* fix this by running Linux-2.6.30 or later *\n");
 380
 381                         aperf_mperf_unstable = 1;
 382                 }
 383                 /*
 384                  * mperf delta is likely a huge "positive" number
 385                  * can not use it for calculating c0 time
 386                  */
 387                 skip_c0 = 1;
 388                 skip_c1 = 1;
 389         }
 390
 391
 392         /*
 393          * As counter collection is not atomic,
 394          * it is possible for mperf's non-halted cycles + idle states
 395          * to exceed TSC's all cycles: show c1 = 0% in that case.
 396          */
 397         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 398                 old->c1 = 0;
 399         else {
 400                 /* normal case, derive c1 */
 401                 old->c1 = old->tsc - old->mperf - core_delta->c3
 402                         - core_delta->c6 - core_delta->c7;
 403         }
 404
 405         if (old->mperf == 0) {
 406                 WARNING("cpu%d MPERF 0!\n", old->cpu_id);
 407                 old->mperf = 1; /* divide by 0 protection */
 408         }
 409
 410         old->smi_count = new->smi_count - old->smi_count;
 411
 412         return 0;
 413 }
 414
 415 static int __attribute__((warn_unused_result))
 416 delta_cpu(struct thread_data *t, struct core_data *c,
 417         struct pkg_data *p, struct thread_data *t2,
 418         struct core_data *c2, struct pkg_data *p2)
 419 {
 420         int ret;
 421
 422         /* calculate core delta only for 1st thread in core */
 423         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 424                 delta_core(c, c2);
 425
 426         /* always calculate thread delta */
 427         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 428         if (ret != 0)
 429                 return ret;
 430
 431         /* calculate package delta only for 1st core in package */
 432         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 433                 delta_package(p, p2);
 434
 435         return 0;
 436 }
 437
 438 /*
 439  * get_counters(...)
 440  * migrate to cpu
 441  * acquire and record local counters for that cpu
 442  */
 443 static int __attribute__((warn_unused_result))
 444 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 445 {
 446         int cpu = t->cpu_id;
 447         unsigned long long msr;
 448         int msr_fd;
 449         int retval = 0;
 450
 451         msr_fd = open_msr(cpu);
 452         if (msr_fd < 0)
 453                 return msr_fd;
 454
 455 #define READ_MSR(msr, dst)                      \
 456 do {                                            \
 457         if (read_msr(msr_fd, msr, dst)) {       \
 458                 retval = -ERR_##msr;            \
 459                 goto out;                       \
 460         }                                       \
 461 } while (0)
 462
 463         READ_MSR(MSR_IA32_TSC, &t->tsc);
 464
 465         READ_MSR(MSR_IA32_APERF, &t->aperf);
 466         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 467
 468         READ_MSR(MSR_SMI_COUNT, &msr);
 469         t->smi_count = msr & 0xFFFFFFFF;
 470
 471         /* collect core counters only for 1st thread in core */
 472         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 473                 retval = 0;
 474                 goto out;
 475         }
 476
 477         if (do_core_cstate & (1 << 3))
 478                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 479         if (do_core_cstate & (1 << 6))
 480                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 481         if (do_core_cstate & (1 << 7))
 482                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 483
 484         if (do_dts) {
 485                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 486                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 487         }
 488
 489         /* collect package counters only for 1st core in package */
 490         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 491                 retval = 0;
 492                 goto out;
 493         }
 494
 495         if (do_pkg_cstate & (1 << 2))
 496                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 497         if (do_pkg_cstate & (1 << 3))
 498                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 499         if (do_pkg_cstate & (1 << 6))
 500                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 501         if (do_pkg_cstate & (1 << 7))
 502                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 503         if (do_pkg_cstate & (1 << 8))
 504                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 505         if (do_pkg_cstate & (1 << 9))
 506                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 507         if (do_pkg_cstate & (1 << 10))
 508                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 509
 510         if (do_rapl & RAPL_PKG) {
 511                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 512                 p->energy_pkg = msr & 0xFFFFFFFF;
 513         }
 514         if (do_rapl & RAPL_CORES) {
 515                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 516                 p->energy_cores = msr & 0xFFFFFFFF;
 517         }
 518         if (do_rapl & RAPL_DRAM) {
 519                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 520                 p->energy_dram = msr & 0xFFFFFFFF;
 521         }
 522         if (do_rapl & RAPL_GFX) {
 523                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 524                 p->energy_gfx = msr & 0xFFFFFFFF;
 525         }
 526         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 527                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 528                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 529         }
 530         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 531                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 532                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 533         }
 534         if (do_ptm) {
 535                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 536                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 537         }
 538
 539 out:
 540         close(msr_fd);
 541         return retval;
 542 }
 543
 544 static void
 545 free_all_buffers(void)
 546 {
 547         allocated = 0;
 548         initialized = 0;
 549
 550         CPU_FREE(cpu_present_set);
 551         cpu_present_set = NULL;
 552         cpu_present_set = 0;
 553
 554         CPU_FREE(cpu_affinity_set);
 555         cpu_affinity_set = NULL;
 556         cpu_affinity_setsize = 0;
 557
 558         free(thread_even);
 559         free(core_even);
 560         free(package_even);
 561
 562         thread_even = NULL;
 563         core_even = NULL;
 564         package_even = NULL;
 565
 566         free(thread_odd);
 567         free(core_odd);
 568         free(package_odd);
 569
 570         thread_odd = NULL;
 571         core_odd = NULL;
 572         package_odd = NULL;
 573 }
 574
 575 /*
 576  * Parse a file containing a single int.
 577  */
 578 static int __attribute__ ((format(printf,1,2)))
 579 parse_int_file(const char *fmt, ...)
 580 {
 581         va_list args;
 582         char path[PATH_MAX];
 583         FILE *filep;
 584         int value;
 585
 586         va_start(args, fmt);
 587         vsnprintf(path, sizeof(path), fmt, args);
 588         va_end(args);
 589         filep = fopen(path, "r");
 590         if (!filep) {
 591                 ERROR("%s: open failed", path);
 592                 return -ERR_CANT_OPEN_FILE;
 593         }
 594         if (fscanf(filep, "%d", &value) != 1) {
 595                 ERROR("%s: failed to parse number from file", path);
 596                 return -ERR_CANT_READ_NUMBER;
 597         }
 598         fclose(filep);
 599         return value;
 600 }
 601
 602 /*
 603  * cpu_is_first_sibling_in_core(cpu)
 604  * return 1 if given CPU is 1st HT sibling in the core
 605  */
 606 static int
 607 cpu_is_first_sibling_in_core(int cpu)
 608 {
 609         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 610 }
 611
 612 /*
 613  * cpu_is_first_core_in_package(cpu)
 614  * return 1 if given CPU is 1st core in package
 615  */
 616 static int
 617 cpu_is_first_core_in_package(int cpu)
 618 {
 619         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 620 }
 621
 622 static int
 623 get_physical_package_id(int cpu)
 624 {
 625         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 626 }
 627
 628 static int
 629 get_core_id(int cpu)
 630 {
 631         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 632 }
 633
 634 static int
 635 get_num_ht_siblings(int cpu)
 636 {
 637         char path[80];
 638         FILE *filep;
 639         int sib1, sib2;
 640         int matches;
 641         char character;
 642
 643         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 644         filep = fopen(path, "r");
 645         if (!filep) {
 646                 ERROR("%s: open failed", path);
 647                 return -ERR_CANT_OPEN_FILE;
 648         }
 649         /*
 650          * file format:
 651          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 652          * otherwinse 1 sibling (self).
 653          */
 654         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 655
 656         fclose(filep);
 657
 658         if (matches == 3)
 659                 return 2;
 660         else
 661                 return 1;
 662 }
 663
 664 /*
 665  * run func(thread, core, package) in topology order
 666  * skip non-present cpus
 667  */
 668
 669
 670 static int __attribute__((warn_unused_result))
 671 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 672         struct pkg_data *, struct thread_data *, struct core_data *,
 673         struct pkg_data *), struct thread_data *thread_base,
 674         struct core_data *core_base, struct pkg_data *pkg_base,
 675         struct thread_data *thread_base2, struct core_data *core_base2,
 676         struct pkg_data *pkg_base2)
 677 {
 678         int retval, pkg_no, core_no, thread_no;
 679
 680         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 681                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 682                         for (thread_no = 0; thread_no <
 683                                 topo.num_threads_per_core; ++thread_no) {
 684                                 struct thread_data *t, *t2;
 685                                 struct core_data *c, *c2;
 686                                 struct pkg_data *p, *p2;
 687
 688                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 689
 690                                 if (cpu_is_not_present(t->cpu_id))
 691                                         continue;
 692
 693                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 694
 695                                 c = GET_CORE(core_base, core_no, pkg_no);
 696                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 697
 698                                 p = GET_PKG(pkg_base, pkg_no);
 699                                 p2 = GET_PKG(pkg_base2, pkg_no);
 700
 701                                 retval = func(t, c, p, t2, c2, p2);
 702                                 if (retval)
 703                                         return retval;
 704                         }
 705                 }
 706         }
 707         return 0;
 708 }
 709
 710 /*
 711  * run func(cpu) on every cpu in /proc/stat
 712  * return max_cpu number
 713  */
 714 static int __attribute__((warn_unused_result))
 715 for_all_proc_cpus(int (func)(int))
 716 {
 717         FILE *fp;
 718         int cpu_num;
 719         int retval;
 720
 721         fp = fopen(proc_stat, "r");
 722         if (!fp) {
 723                 ERROR("%s: open failed", proc_stat);
 724                 return -ERR_CANT_OPEN_FILE;
 725         }
 726
 727         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 728         if (retval != 0) {
 729                 ERROR("%s: failed to parse format", proc_stat);
 730                 return -ERR_CANT_READ_PROC_STAT;
 731         }
 732
 733         while (1) {
 734                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 735                 if (retval != 1)
 736                         break;
 737
 738                 retval = func(cpu_num);
 739                 if (retval) {
 740                         fclose(fp);
 741                         return(retval);
 742                 }
 743         }
 744         fclose(fp);
 745         return 0;
 746 }
 747
 748 /*
 749  * count_cpus()
 750  * remember the last one seen, it will be the max
 751  */
 752 static int
 753 count_cpus(int cpu)
 754 {
 755         if (topo.max_cpu_num < cpu)
 756                 topo.max_cpu_num = cpu;
 757
 758         topo.num_cpus += 1;
 759         return 0;
 760 }
 761 static int
 762 mark_cpu_present(int cpu)
 763 {
 764         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 765         return 0;
 766 }
 767
 768
 769 static void
 770 turbostat_submit (const char *plugin_instance,
 771         const char *type, const char *type_instance,
 772         gauge_t value)
 773 {
 774         value_list_t vl = VALUE_LIST_INIT;
 775         value_t v;
 776
 777         v.gauge = value;
 778         vl.values = &v;
 779         vl.values_len = 1;
 780         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 781         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 782         if (plugin_instance != NULL)
 783                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 784         sstrncpy (vl.type, type, sizeof (vl.type));
 785         if (type_instance != NULL)
 786                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 787
 788         plugin_dispatch_values (&vl);
 789 }
 790
 791 /*
 792  * column formatting convention & formats
 793  * package: "pk" 2 columns %2d
 794  * core: "cor" 3 columns %3d
 795  * CPU: "CPU" 3 columns %3d
 796  * Pkg_W: %6.2
 797  * Cor_W: %6.2
 798  * GFX_W: %5.2
 799  * RAM_W: %5.2
 800  * GHz: "GHz" 3 columns %3.2
 801  * TSC: "TSC" 3 columns %3.2
 802  * SMI: "SMI" 4 columns %4d
 803  * percentage " %pc3" %6.2
 804  * Perf Status percentage: %5.2
 805  * "CTMP" 4 columns %4d
 806  */
 807 #define NAME_LEN 12
 808 static int
 809 submit_counters(struct thread_data *t, struct core_data *c,
 810         struct pkg_data *p)
 811 {
 812         char name[NAME_LEN];
 813         double interval_float;
 814
 815         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 816
 817         snprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 818
 819         if (!skip_c0)
 820                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 821         if (!skip_c1)
 822                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 823
 824         /* GHz */
 825         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 826                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 827
 828         /* SMI */
 829         turbostat_submit(NULL, "current", name, t->smi_count);
 830
 831         /* print per-core data only for 1st thread in core */
 832         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 833                 goto done;
 834
 835         snprintf(name, sizeof(name), "core%02d", c->core_id);
 836
 837         if (do_core_cstate & (1 << 3))
 838                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 839         if (do_core_cstate & (1 << 6))
 840                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 841         if (do_core_cstate & (1 << 7))
 842                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 843
 844         if (do_dts)
 845                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 846
 847         /* print per-package data only for 1st core in package */
 848         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 849                 goto done;
 850
 851         snprintf(name, sizeof(name), "pkg%02d", p->package_id);
 852
 853         if (do_ptm)
 854                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 855
 856         if (do_pkg_cstate & (1 << 2))
 857                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 858         if (do_pkg_cstate & (1 << 3))
 859                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 860         if (do_pkg_cstate & (1 << 6))
 861                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 862         if (do_pkg_cstate & (1 << 7))
 863                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 864         if (do_pkg_cstate & (1 << 8))
 865                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 866         if (do_pkg_cstate & (1 << 9))
 867                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 868         if (do_pkg_cstate & (1 << 10))
 869                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 870
 871         if (do_rapl) {
 872                 if (do_rapl & RAPL_PKG)
 873                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 874                 if (do_rapl & RAPL_CORES)
 875                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 876                 if (do_rapl & RAPL_GFX)
 877                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 878                 if (do_rapl & RAPL_DRAM)
 879                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 880         }
 881 done:
 882         return 0;
 883 }
 884
 885 static int
 886 turbostat_read(user_data_t * not_used)
 887 {
 888         int ret;
 889
 890         if (!allocated) {
 891                 if ((ret = setup_all_buffers()) < 0)
 892                         return ret;
 893         }
 894
 895         if (for_all_proc_cpus(cpu_is_not_present)) {
 896                 free_all_buffers();
 897                 if ((ret = setup_all_buffers()) < 0)
 898                         return ret;
 899                 if (for_all_proc_cpus(cpu_is_not_present))
 900                         return -ERR_CPU_NOT_PRESENT;
 901         }
 902
 903         if (!initialized) {
 904                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 905                         return ret;
 906                 gettimeofday(&tv_even, (struct timezone *)NULL);
 907                 is_even = 1;
 908                 initialized = 1;
 909                 return 0;
 910         }
 911
 912         if (is_even) {
 913                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 914                         return ret;
 915                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 916                 is_even = 0;
 917                 timersub(&tv_odd, &tv_even, &tv_delta);
 918                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 919                         return ret;
 920                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 921                         return ret;
 922         } else {
 923                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 924                         return ret;
 925                 gettimeofday(&tv_even, (struct timezone *)NULL);
 926                 is_even = 1;
 927                 timersub(&tv_even, &tv_odd, &tv_delta);
 928                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 929                         return ret;
 930                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 931                         return ret;
 932         }
 933         return 0;
 934 }
 935
 936 static int __attribute__((warn_unused_result))
 937 check_dev_msr()
 938 {
 939         struct stat sb;
 940
 941         if (stat("/dev/cpu/0/msr", &sb)) {
 942                 ERROR("no /dev/cpu/0/msr\n"
 943                         "Try \"# modprobe msr\"");
 944                 return -ERR_NO_MSR;
 945         }
 946         return 0;
 947 }
 948
 949 static int __attribute__((warn_unused_result))
 950 check_super_user()
 951 {
 952         if (getuid() != 0) {
 953                 ERROR("must be root");
 954                 return -ERR_NOT_ROOT;
 955         }
 956         return 0;
 957 }
 958
 959 /*
 960  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
 961  * the Thermal Control Circuit (TCC) activates.
 962  * This is usually equal to tjMax.
 963  *
 964  * Older processors do not have this MSR, so there we guess,
 965  * but also allow cmdline over-ride with -T.
 966  *
 967  * Several MSR temperature values are in units of degrees-C
 968  * below this value, including the Digital Thermal Sensor (DTS),
 969  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
 970  */
 971 static int __attribute__((warn_unused_result))
 972 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 973 {
 974         unsigned long long msr;
 975         unsigned int target_c_local;
 976
 977         /* tcc_activation_temp is used only for dts or ptm */
 978         if (!(do_dts || do_ptm))
 979                 return 0;
 980
 981         /* this is a per-package concept */
 982         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 983                 return 0;
 984
 985         if (tcc_activation_temp_override != 0) {
 986                 tcc_activation_temp = tcc_activation_temp_override;
 987                 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
 988                         cpu, tcc_activation_temp);
 989                 return 0;
 990         }
 991
 992         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
 993                 goto guess;
 994
 995         target_c_local = (msr >> 16) & 0x7F;
 996
 997         if (target_c_local < 85 || target_c_local > 127)
 998                 goto guess;
 999
1000         tcc_activation_temp = target_c_local;
1001
1002         return 0;
1003
1004 guess:
1005         tcc_activation_temp = TJMAX_DEFAULT;
1006         WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1007                 t->cpu_id, tcc_activation_temp);
1008
1009         return 0;
1010 }
1011
1012 /*
1013  * Identify the functionality of the CPU
1014  */
1015 static int __attribute__((warn_unused_result))
1016 probe_cpu()
1017 {
1018         unsigned int eax, ebx, ecx, edx, max_level;
1019         unsigned int fms, family, model;
1020
1021         /* CPUID(0):
1022          * - EAX: Maximum Input Value for Basic CPUID Information
1023          * - EBX: "Genu" (0x756e6547)
1024          * - EDX: "ineI" (0x49656e69)
1025          * - ECX: "ntel" (0x6c65746e)
1026          */
1027         max_level = ebx = ecx = edx = 0;
1028         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1029         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1030                 ERROR("Unsupported CPU");
1031                 return -UNSUPPORTED_CPU;
1032         }
1033
1034         /* CPUID(1):
1035          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1036          *  + 4-7:   Model ID
1037          *  + 8-11:  Family ID
1038          *  + 12-13: Processor type
1039          *  + 16-19: Extended Model ID
1040          *  + 20-27: Extended Family ID
1041          * - EDX: Feature Information:
1042          *  + 5: Support for MSR read/write operations
1043          */
1044         fms = ebx = ecx = edx = 0;
1045         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1046         family = (fms >> 8) & 0xf;
1047         model = (fms >> 4) & 0xf;
1048         if (family == 0xf)
1049                 family += (fms >> 20) & 0xf;
1050         if (family == 6 || family == 0xf)
1051                 model += ((fms >> 16) & 0xf) << 4;
1052         if (!(edx & (1 << 5))) {
1053                 ERROR("CPUID: no MSR");
1054                 return -ERR_NO_MSR;
1055         }
1056
1057         /*
1058          * CPUID(0x80000000):
1059          * - EAX: Maximum Input Value for Extended Function CPUID Information
1060          *
1061          * This allows us to verify if the CPUID(0x80000007) can be called
1062          *
1063          * This check is valid for both Intel and AMD.
1064          */
1065         max_level = ebx = ecx = edx = 0;
1066         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1067         if (max_level < 0x80000007) {
1068                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1069                 return -ERR_NO_INVARIANT_TSC;
1070         }
1071
1072         /*
1073          * CPUID(0x80000007):
1074          * - EDX:
1075          *  + 8: Invariant TSC available if set
1076          *
1077          * This check is valid for both Intel and AMD
1078          */
1079         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1080         if (!(edx & (1 << 8))) {
1081                 ERROR("No invariant TSC");
1082                 return -ERR_NO_INVARIANT_TSC;
1083         }
1084
1085         /*
1086          * CPUID(6):
1087          * - EAX:
1088          *  + 0: Digital temperature sensor is supported if set
1089          *  + 6: Package thermal management is supported if set
1090          * - ECX:
1091          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1092          *  + 3: The processor supports performance-energy bias preference if set.
1093          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1094          *
1095          * This check is valid for both Intel and AMD
1096          */
1097         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1098         do_dts = eax & (1 << 0);
1099         do_ptm = eax & (1 << 6);
1100         if (!(ecx & (1 << 0))) {
1101                 ERROR("No APERF");
1102                 return -ERR_NO_APERF;
1103         }
1104
1105         /*
1106          * Enable or disable C states depending on the model and family
1107          */
1108         if (family == 6) {
1109                 switch (model) {
1110                 /* Atom (partial) */
1111                 case 0x27:
1112                         do_core_cstate = 0;
1113                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1114                         break;
1115                 /* Silvermont */
1116                 case 0x37: /* BYT */
1117                 case 0x4A:
1118                 case 0x4D: /* AVN */
1119                 case 0x5A:
1120                 case 0x5D:
1121                         do_core_cstate = (1 << 1) | (1 << 6);
1122                         do_pkg_cstate = (1 << 6);
1123                         break;
1124                 /* Nehalem */
1125                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1126                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1127                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1128                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1129                         do_core_cstate = (1 << 3) | (1 << 6);
1130                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1131                         break;
1132                 /* Westmere */
1133                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1134                 case 0x2C: /* Westmere EP - Gulftown */
1135                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1136                         do_core_cstate = (1 << 3) | (1 << 6);
1137                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1138                         break;
1139                 /* Sandy Bridge */
1140                 case 0x2A: /* SNB */
1141                 case 0x2D: /* SNB Xeon */
1142                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1143                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1144                         break;
1145                 /* Ivy Bridge */
1146                 case 0x3A: /* IVB */
1147                 case 0x3E: /* IVB Xeon */
1148                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1149                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1150                         break;
1151                 /* Haswell Bridge */
1152                 case 0x3C: /* HSW */
1153                 case 0x3F: /* HSW */
1154                 case 0x46: /* HSW */
1155                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1156                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1157                         break;
1158                 case 0x45: /* HSW */
1159                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1160                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1161                         break;
1162                 /* Broadwel */
1163                 case 0x4F: /* BDW */
1164                 case 0x56: /* BDX-DE */
1165                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1166                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1167                         break;
1168                 case 0x3D: /* BDW */
1169                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1170                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1171                         break;
1172                 default:
1173                         ERROR("Unsupported CPU");
1174                 }
1175                 switch (model) {
1176                 case 0x2A:
1177                 case 0x3A:
1178                 case 0x3C:
1179                 case 0x45:
1180                 case 0x46:
1181                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1182                         break;
1183                 case 0x3F:
1184                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1185                         break;
1186                 case 0x2D:
1187                 case 0x3E:
1188                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1189                         break;
1190                 case 0x37:
1191                 case 0x4D:
1192                         do_rapl = RAPL_PKG | RAPL_CORES;
1193                         break;
1194                 default:
1195                         do_rapl = 0;
1196                 }
1197         } else {
1198                 ERROR("Unsupported CPU");
1199                 return -UNSUPPORTED_CPU;
1200         }
1201
1202         if (do_rapl) {
1203                 unsigned long msr;
1204                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1205                         return 0;
1206
1207                 if (model == 0x37)
1208                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1209                 else
1210                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1211         }
1212
1213         return 0;
1214 }
1215
1216
1217
1218 static int __attribute__((warn_unused_result))
1219 topology_probe()
1220 {
1221         int i;
1222         int ret;
1223         int max_core_id = 0;
1224         int max_package_id = 0;
1225         int max_siblings = 0;
1226         struct cpu_topology {
1227                 int core_id;
1228                 int physical_package_id;
1229         } *cpus;
1230
1231         /* Initialize num_cpus, max_cpu_num */
1232         topo.num_cpus = 0;
1233         topo.max_cpu_num = 0;
1234         ret = for_all_proc_cpus(count_cpus);
1235         if (ret < 0)
1236                 return ret;
1237
1238         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1239
1240         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1241         if (cpus == NULL) {
1242                 ERROR("calloc cpus");
1243                 return -ERR_CALLOC;
1244         }
1245
1246         /*
1247          * Allocate and initialize cpu_present_set
1248          */
1249         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1250         if (cpu_present_set == NULL) {
1251                 free(cpus);
1252                 ERROR("CPU_ALLOC");
1253                 return -ERR_CPU_ALLOC;
1254         }
1255         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1256         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1257         ret = for_all_proc_cpus(mark_cpu_present);
1258         if (ret < 0) {
1259                 free(cpus);
1260                 return ret;
1261         }
1262
1263         /*
1264          * Allocate and initialize cpu_affinity_set
1265          */
1266         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1267         if (cpu_affinity_set == NULL) {
1268                 free(cpus);
1269                 ERROR("CPU_ALLOC");
1270                 return -ERR_CPU_ALLOC;
1271         }
1272         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1273         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1274
1275
1276         /*
1277          * For online cpus
1278          * find max_core_id, max_package_id
1279          */
1280         for (i = 0; i <= topo.max_cpu_num; ++i) {
1281                 int siblings;
1282
1283                 if (cpu_is_not_present(i)) {
1284                         //if (verbose > 1)
1285                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1286                         continue;
1287                 }
1288                 cpus[i].core_id = get_core_id(i);
1289                 if (cpus[i].core_id < 0)
1290                         return cpus[i].core_id;
1291                 if (cpus[i].core_id > max_core_id)
1292                         max_core_id = cpus[i].core_id;
1293
1294                 cpus[i].physical_package_id = get_physical_package_id(i);
1295                 if (cpus[i].physical_package_id < 0)
1296                         return cpus[i].physical_package_id;
1297                 if (cpus[i].physical_package_id > max_package_id)
1298                         max_package_id = cpus[i].physical_package_id;
1299
1300                 siblings = get_num_ht_siblings(i);
1301                 if (siblings < 0)
1302                         return siblings;
1303                 if (siblings > max_siblings)
1304                         max_siblings = siblings;
1305                 DEBUG("cpu %d pkg %d core %d\n",
1306                         i, cpus[i].physical_package_id, cpus[i].core_id);
1307         }
1308         topo.num_cores_per_pkg = max_core_id + 1;
1309         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1310                 max_core_id, topo.num_cores_per_pkg);
1311
1312         topo.num_packages = max_package_id + 1;
1313         DEBUG("max_package_id %d, sizing for %d packages\n",
1314                 max_package_id, topo.num_packages);
1315
1316         topo.num_threads_per_core = max_siblings;
1317         DEBUG("max_siblings %d\n", max_siblings);
1318
1319         free(cpus);
1320         return 0;
1321 }
1322
1323 static int
1324 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1325 {
1326         int i;
1327
1328         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1329                 topo.num_packages, sizeof(struct thread_data));
1330         if (*t == NULL)
1331                 goto error;
1332
1333         for (i = 0; i < topo.num_threads_per_core *
1334                 topo.num_cores_per_pkg * topo.num_packages; i++)
1335                 (*t)[i].cpu_id = -1;
1336
1337         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1338                 sizeof(struct core_data));
1339         if (*c == NULL)
1340                 goto error;
1341
1342         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1343                 (*c)[i].core_id = -1;
1344
1345         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1346         if (*p == NULL)
1347                 goto error;
1348
1349         for (i = 0; i < topo.num_packages; i++)
1350                 (*p)[i].package_id = i;
1351
1352         return 0;
1353 error:
1354         ERROR("calloc counters");
1355         return -ERR_CALLOC;
1356 }
1357 /*
1358  * init_counter()
1359  *
1360  * set cpu_id, core_num, pkg_num
1361  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1362  *
1363  * increment topo.num_cores when 1st core in pkg seen
1364  */
1365 static int
1366 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1367         struct pkg_data *pkg_base, int thread_num, int core_num,
1368         int pkg_num, int cpu_id)
1369 {
1370         int ret;
1371         struct thread_data *t;
1372         struct core_data *c;
1373         struct pkg_data *p;
1374
1375         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1376         c = GET_CORE(core_base, core_num, pkg_num);
1377         p = GET_PKG(pkg_base, pkg_num);
1378
1379         t->cpu_id = cpu_id;
1380         if (thread_num == 0) {
1381                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1382                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1383                         return ret;
1384                 } else if (ret != 0) {
1385                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1386                 }
1387         }
1388
1389         c->core_id = core_num;
1390         p->package_id = pkg_num;
1391
1392         return 0;
1393 }
1394
1395
1396 static int
1397 initialize_counters(int cpu_id)
1398 {
1399         int my_thread_id, my_core_id, my_package_id;
1400         int ret;
1401
1402         my_package_id = get_physical_package_id(cpu_id);
1403         if (my_package_id < 0)
1404                 return my_package_id;
1405         my_core_id = get_core_id(cpu_id);
1406         if (my_core_id < 0)
1407                 return my_core_id;
1408
1409         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1410                 return ret;
1411         } else if (ret != 0) {
1412                 my_thread_id = 0;
1413                 topo.num_cores++;
1414         } else {
1415                 my_thread_id = 1;
1416         }
1417
1418         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1419         if (ret < 0)
1420                 return ret;
1421         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1422         if (ret < 0)
1423                 return ret;
1424         return 0;
1425 }
1426
1427 #define DO_OR_GOTO_ERR(something) \
1428 do {                         \
1429         ret = (something);     \
1430         if (ret < 0)         \
1431                 goto err;    \
1432 } while (0)
1433
1434 static int setup_all_buffers(void)
1435 {
1436         int ret;
1437
1438         DO_OR_GOTO_ERR(topology_probe());
1439         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1440         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1441         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1442
1443         allocated = 1;
1444         return 0;
1445 err:
1446         free_all_buffers();
1447         return ret;
1448 }
1449
1450 static int
1451 turbostat_init(void)
1452 {
1453         int ret;
1454
1455         DO_OR_GOTO_ERR(check_super_user());
1456         DO_OR_GOTO_ERR(probe_cpu());
1457         DO_OR_GOTO_ERR(check_dev_msr());
1458         DO_OR_GOTO_ERR(setup_all_buffers());
1459         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1460
1461         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1462
1463         return 0;
1464 err:
1465         free_all_buffers();
1466         return ret;
1467 }
1468
1469 void module_register(void);
1470 void module_register(void)
1471 {
1472         plugin_register_init(PLUGIN_NAME, turbostat_init);
1473 }