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
 274         ssnprintf(pathname, 32, "/dev/cpu/%d/msr", cpu);
 275         return open(pathname, O_RDONLY);
 276 }
 277
 278 static int __attribute__((warn_unused_result))
 279 read_msr(int fd, off_t offset, unsigned long long *msr)
 280 {
 281         ssize_t retval;
 282
 283         retval = pread(fd, msr, sizeof *msr, offset);
 284
 285         if (retval != sizeof *msr) {
 286                 ERROR ("MSR offset 0x%llx read failed", (unsigned long long)offset);
 287                 return -1;
 288         }
 289         return 0;
 290 }
 291
 292 static int __attribute__((warn_unused_result))
 293 get_msr(int cpu, off_t offset, unsigned long long *msr)
 294 {
 295         ssize_t retval;
 296         int fd;
 297
 298         fd = open_msr(cpu);
 299         if (fd < 0)
 300                 return -1;
 301         retval = read_msr(fd, offset, msr);
 302         close(fd);
 303         return retval;
 304 }
 305
 306 #define DELTA_WRAP32(new, old)                  \
 307         if (new > old) {                        \
 308                 old = new - old;                \
 309         } else {                                \
 310                 old = 0x100000000 + new - old;  \
 311         }
 312
 313 static void
 314 delta_package(struct pkg_data *new, struct pkg_data *old)
 315 {
 316         old->pc2 = new->pc2 - old->pc2;
 317         old->pc3 = new->pc3 - old->pc3;
 318         old->pc6 = new->pc6 - old->pc6;
 319         old->pc7 = new->pc7 - old->pc7;
 320         old->pc8 = new->pc8 - old->pc8;
 321         old->pc9 = new->pc9 - old->pc9;
 322         old->pc10 = new->pc10 - old->pc10;
 323         old->pkg_temp_c = new->pkg_temp_c;
 324
 325         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 326         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 327         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 328         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 329         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 330         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 331 }
 332
 333 static void
 334 delta_core(struct core_data *new, struct core_data *old)
 335 {
 336         old->c3 = new->c3 - old->c3;
 337         old->c6 = new->c6 - old->c6;
 338         old->c7 = new->c7 - old->c7;
 339         old->core_temp_c = new->core_temp_c;
 340 }
 341
 342 /*
 343  * old = new - old
 344  */
 345 static int __attribute__((warn_unused_result))
 346 delta_thread(struct thread_data *new, struct thread_data *old,
 347         struct core_data *core_delta)
 348 {
 349         old->tsc = new->tsc - old->tsc;
 350
 351         /* check for TSC < 1 Mcycles over interval */
 352         if (old->tsc < (1000 * 1000)) {
 353                 WARNING("Insanely slow TSC rate, TSC stops in idle?\n"
 354                         "You can disable all c-states by booting with \"idle=poll\"\n"
 355                         "or just the deep ones with \"processor.max_cstate=1\"");
 356                 return -1;
 357         }
 358
 359         old->c1 = new->c1 - old->c1;
 360
 361         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 362                 old->aperf = new->aperf - old->aperf;
 363                 old->mperf = new->mperf - old->mperf;
 364         } else {
 365
 366                 if (!aperf_mperf_unstable) {
 367                         WARNING("%s: APERF or MPERF went backwards *\n", progname);
 368                         WARNING("* Frequency results do not cover entire interval *\n");
 369                         WARNING("* fix this by running Linux-2.6.30 or later *\n");
 370
 371                         aperf_mperf_unstable = 1;
 372                 }
 373                 /*
 374                  * mperf delta is likely a huge "positive" number
 375                  * can not use it for calculating c0 time
 376                  */
 377                 skip_c0 = 1;
 378                 skip_c1 = 1;
 379         }
 380
 381
 382         /*
 383          * As counter collection is not atomic,
 384          * it is possible for mperf's non-halted cycles + idle states
 385          * to exceed TSC's all cycles: show c1 = 0% in that case.
 386          */
 387         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 388                 old->c1 = 0;
 389         else {
 390                 /* normal case, derive c1 */
 391                 old->c1 = old->tsc - old->mperf - core_delta->c3
 392                         - core_delta->c6 - core_delta->c7;
 393         }
 394
 395         if (old->mperf == 0) {
 396                 WARNING("cpu%d MPERF 0!\n", old->cpu_id);
 397                 old->mperf = 1; /* divide by 0 protection */
 398         }
 399
 400         old->smi_count = new->smi_count - old->smi_count;
 401
 402         return 0;
 403 }
 404
 405 static int __attribute__((warn_unused_result))
 406 delta_cpu(struct thread_data *t, struct core_data *c,
 407         struct pkg_data *p, struct thread_data *t2,
 408         struct core_data *c2, struct pkg_data *p2)
 409 {
 410         int ret;
 411
 412         /* calculate core delta only for 1st thread in core */
 413         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 414                 delta_core(c, c2);
 415
 416         /* always calculate thread delta */
 417         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 418         if (ret != 0)
 419                 return ret;
 420
 421         /* calculate package delta only for 1st core in package */
 422         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 423                 delta_package(p, p2);
 424
 425         return 0;
 426 }
 427
 428 /*
 429  * get_counters(...)
 430  * migrate to cpu
 431  * acquire and record local counters for that cpu
 432  */
 433 static int __attribute__((warn_unused_result))
 434 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 435 {
 436         int cpu = t->cpu_id;
 437         unsigned long long msr;
 438         int msr_fd;
 439         int retval = 0;
 440
 441         if (cpu_migrate(cpu)) {
 442                 WARNING("Could not migrate to CPU %d\n", cpu);
 443                 return -ERR_CPU_MIGRATE;
 444         }
 445
 446         msr_fd = open_msr(cpu);
 447         if (msr_fd < 0)
 448                 return -ERR_CANT_OPEN_MSR;
 449
 450 #define READ_MSR(msr, dst)                      \
 451 do {                                            \
 452         if (read_msr(msr_fd, msr, dst)) {       \
 453                 retval = -ERR_##msr;            \
 454                 goto out;                       \
 455         }                                       \
 456 } while (0)
 457
 458         READ_MSR(MSR_IA32_TSC, &t->tsc);
 459
 460         READ_MSR(MSR_IA32_APERF, &t->aperf);
 461         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 462
 463         READ_MSR(MSR_SMI_COUNT, &msr);
 464         t->smi_count = msr & 0xFFFFFFFF;
 465
 466         /* collect core counters only for 1st thread in core */
 467         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 468                 retval = 0;
 469                 goto out;
 470         }
 471
 472         if (do_core_cstate & (1 << 3))
 473                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 474         if (do_core_cstate & (1 << 6))
 475                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 476         if (do_core_cstate & (1 << 7))
 477                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 478
 479         if (do_dts) {
 480                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 481                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 482         }
 483
 484         /* collect package counters only for 1st core in package */
 485         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 486                 retval = 0;
 487                 goto out;
 488         }
 489
 490         if (do_pkg_cstate & (1 << 2))
 491                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 492         if (do_pkg_cstate & (1 << 3))
 493                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 494         if (do_pkg_cstate & (1 << 6))
 495                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 496         if (do_pkg_cstate & (1 << 7))
 497                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 498         if (do_pkg_cstate & (1 << 8))
 499                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 500         if (do_pkg_cstate & (1 << 9))
 501                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 502         if (do_pkg_cstate & (1 << 10))
 503                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 504
 505         if (do_rapl & RAPL_PKG) {
 506                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 507                 p->energy_pkg = msr & 0xFFFFFFFF;
 508         }
 509         if (do_rapl & RAPL_CORES) {
 510                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 511                 p->energy_cores = msr & 0xFFFFFFFF;
 512         }
 513         if (do_rapl & RAPL_DRAM) {
 514                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 515                 p->energy_dram = msr & 0xFFFFFFFF;
 516         }
 517         if (do_rapl & RAPL_GFX) {
 518                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 519                 p->energy_gfx = msr & 0xFFFFFFFF;
 520         }
 521         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 522                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 523                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 524         }
 525         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 526                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 527                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 528         }
 529         if (do_ptm) {
 530                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 531                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 532         }
 533
 534 out:
 535         close(msr_fd);
 536         return retval;
 537 }
 538
 539 static void
 540 free_all_buffers(void)
 541 {
 542         allocated = 0;
 543         initialized = 0;
 544
 545         CPU_FREE(cpu_present_set);
 546         cpu_present_set = NULL;
 547         cpu_present_set = 0;
 548
 549         CPU_FREE(cpu_affinity_set);
 550         cpu_affinity_set = NULL;
 551         cpu_affinity_setsize = 0;
 552
 553         free(thread_even);
 554         free(core_even);
 555         free(package_even);
 556
 557         thread_even = NULL;
 558         core_even = NULL;
 559         package_even = NULL;
 560
 561         free(thread_odd);
 562         free(core_odd);
 563         free(package_odd);
 564
 565         thread_odd = NULL;
 566         core_odd = NULL;
 567         package_odd = NULL;
 568 }
 569
 570 /*
 571  * Parse a file containing a single int.
 572  */
 573 static int __attribute__ ((format(printf,1,2)))
 574 parse_int_file(const char *fmt, ...)
 575 {
 576         va_list args;
 577         char path[PATH_MAX];
 578         FILE *filep;
 579         int value;
 580
 581         va_start(args, fmt);
 582         vsnprintf(path, sizeof(path), fmt, args);
 583         va_end(args);
 584         filep = fopen(path, "r");
 585         if (!filep) {
 586                 ERROR("%s: open failed", path);
 587                 return -ERR_CANT_OPEN_FILE;
 588         }
 589         if (fscanf(filep, "%d", &value) != 1) {
 590                 ERROR("%s: failed to parse number from file", path);
 591                 return -ERR_CANT_READ_NUMBER;
 592         }
 593         fclose(filep);
 594         return value;
 595 }
 596
 597 /*
 598  * cpu_is_first_sibling_in_core(cpu)
 599  * return 1 if given CPU is 1st HT sibling in the core
 600  */
 601 static int
 602 cpu_is_first_sibling_in_core(int cpu)
 603 {
 604         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 605 }
 606
 607 /*
 608  * cpu_is_first_core_in_package(cpu)
 609  * return 1 if given CPU is 1st core in package
 610  */
 611 static int
 612 cpu_is_first_core_in_package(int cpu)
 613 {
 614         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 615 }
 616
 617 static int
 618 get_physical_package_id(int cpu)
 619 {
 620         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 621 }
 622
 623 static int
 624 get_core_id(int cpu)
 625 {
 626         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 627 }
 628
 629 static int
 630 get_num_ht_siblings(int cpu)
 631 {
 632         char path[80];
 633         FILE *filep;
 634         int sib1, sib2;
 635         int matches;
 636         char character;
 637
 638         ssnprintf(path, 80, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 639         filep = fopen(path, "r");
 640         if (!filep) {
 641                 ERROR("%s: open failed", path);
 642                 return -ERR_CANT_OPEN_FILE;
 643         }
 644         /*
 645          * file format:
 646          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 647          * otherwinse 1 sibling (self).
 648          */
 649         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 650
 651         fclose(filep);
 652
 653         if (matches == 3)
 654                 return 2;
 655         else
 656                 return 1;
 657 }
 658
 659 /*
 660  * run func(thread, core, package) in topology order
 661  * skip non-present cpus
 662  */
 663
 664
 665 static int __attribute__((warn_unused_result))
 666 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 667         struct pkg_data *, struct thread_data *, struct core_data *,
 668         struct pkg_data *), struct thread_data *thread_base,
 669         struct core_data *core_base, struct pkg_data *pkg_base,
 670         struct thread_data *thread_base2, struct core_data *core_base2,
 671         struct pkg_data *pkg_base2)
 672 {
 673         int retval, pkg_no, core_no, thread_no;
 674
 675         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 676                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 677                         for (thread_no = 0; thread_no <
 678                                 topo.num_threads_per_core; ++thread_no) {
 679                                 struct thread_data *t, *t2;
 680                                 struct core_data *c, *c2;
 681                                 struct pkg_data *p, *p2;
 682
 683                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 684
 685                                 if (cpu_is_not_present(t->cpu_id))
 686                                         continue;
 687
 688                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 689
 690                                 c = GET_CORE(core_base, core_no, pkg_no);
 691                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 692
 693                                 p = GET_PKG(pkg_base, pkg_no);
 694                                 p2 = GET_PKG(pkg_base2, pkg_no);
 695
 696                                 retval = func(t, c, p, t2, c2, p2);
 697                                 if (retval)
 698                                         return retval;
 699                         }
 700                 }
 701         }
 702         return 0;
 703 }
 704
 705 /*
 706  * run func(cpu) on every cpu in /proc/stat
 707  * return max_cpu number
 708  */
 709 static int __attribute__((warn_unused_result))
 710 for_all_proc_cpus(int (func)(int))
 711 {
 712         FILE *fp;
 713         int cpu_num;
 714         int retval;
 715
 716         fp = fopen(proc_stat, "r");
 717         if (!fp) {
 718                 ERROR("%s: open failed", proc_stat);
 719                 return -ERR_CANT_OPEN_FILE;
 720         }
 721
 722         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 723         if (retval != 0) {
 724                 ERROR("%s: failed to parse format", proc_stat);
 725                 return -ERR_CANT_READ_PROC_STAT;
 726         }
 727
 728         while (1) {
 729                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 730                 if (retval != 1)
 731                         break;
 732
 733                 retval = func(cpu_num);
 734                 if (retval) {
 735                         fclose(fp);
 736                         return(retval);
 737                 }
 738         }
 739         fclose(fp);
 740         return 0;
 741 }
 742
 743 /*
 744  * count_cpus()
 745  * remember the last one seen, it will be the max
 746  */
 747 static int
 748 count_cpus(int cpu)
 749 {
 750         if (topo.max_cpu_num < cpu)
 751                 topo.max_cpu_num = cpu;
 752
 753         topo.num_cpus += 1;
 754         return 0;
 755 }
 756 static int
 757 mark_cpu_present(int cpu)
 758 {
 759         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 760         return 0;
 761 }
 762
 763
 764 static void
 765 turbostat_submit (const char *plugin_instance,
 766         const char *type, const char *type_instance,
 767         gauge_t value)
 768 {
 769         value_list_t vl = VALUE_LIST_INIT;
 770         value_t v;
 771
 772         v.gauge = value;
 773         vl.values = &v;
 774         vl.values_len = 1;
 775         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 776         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 777         if (plugin_instance != NULL)
 778                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 779         sstrncpy (vl.type, type, sizeof (vl.type));
 780         if (type_instance != NULL)
 781                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 782
 783         plugin_dispatch_values (&vl);
 784 }
 785
 786 /*
 787  * column formatting convention & formats
 788  * package: "pk" 2 columns %2d
 789  * core: "cor" 3 columns %3d
 790  * CPU: "CPU" 3 columns %3d
 791  * Pkg_W: %6.2
 792  * Cor_W: %6.2
 793  * GFX_W: %5.2
 794  * RAM_W: %5.2
 795  * GHz: "GHz" 3 columns %3.2
 796  * TSC: "TSC" 3 columns %3.2
 797  * SMI: "SMI" 4 columns %4d
 798  * percentage " %pc3" %6.2
 799  * Perf Status percentage: %5.2
 800  * "CTMP" 4 columns %4d
 801  */
 802 #define NAME_LEN 12
 803 static int
 804 submit_counters(struct thread_data *t, struct core_data *c,
 805         struct pkg_data *p)
 806 {
 807         char name[NAME_LEN];
 808         double interval_float;
 809
 810         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 811
 812         snprintf(name, NAME_LEN, "cpu%02d", t->cpu_id);
 813
 814         if (!skip_c0)
 815                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 816         if (!skip_c1)
 817                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 818
 819         /* GHz */
 820         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 821                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 822
 823         /* SMI */
 824         turbostat_submit(NULL, "current", name, t->smi_count);
 825
 826         /* print per-core data only for 1st thread in core */
 827         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 828                 goto done;
 829
 830         snprintf(name, NAME_LEN, "core%02d", c->core_id);
 831
 832         if (do_core_cstate & (1 << 3))
 833                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 834         if (do_core_cstate & (1 << 6))
 835                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 836         if (do_core_cstate & (1 << 7))
 837                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 838
 839         if (do_dts)
 840                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 841
 842         /* print per-package data only for 1st core in package */
 843         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 844                 goto done;
 845
 846         snprintf(name, NAME_LEN, "pkg%02d", p->package_id);
 847
 848         if (do_ptm)
 849                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 850
 851         if (do_pkg_cstate & (1 << 2))
 852                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 853         if (do_pkg_cstate & (1 << 3))
 854                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 855         if (do_pkg_cstate & (1 << 6))
 856                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 857         if (do_pkg_cstate & (1 << 7))
 858                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 859         if (do_pkg_cstate & (1 << 8))
 860                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 861         if (do_pkg_cstate & (1 << 9))
 862                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 863         if (do_pkg_cstate & (1 << 10))
 864                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 865
 866         if (do_rapl) {
 867                 if (do_rapl & RAPL_PKG)
 868                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 869                 if (do_rapl & RAPL_CORES)
 870                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 871                 if (do_rapl & RAPL_GFX)
 872                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 873                 if (do_rapl & RAPL_DRAM)
 874                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 875         }
 876 done:
 877         return 0;
 878 }
 879
 880 static int
 881 turbostat_read(user_data_t * not_used)
 882 {
 883         int ret;
 884
 885         if (!allocated) {
 886                 if ((ret = setup_all_buffers()) < 0)
 887                         return ret;
 888         }
 889
 890         if (for_all_proc_cpus(cpu_is_not_present)) {
 891                 free_all_buffers();
 892                 if ((ret = setup_all_buffers()) < 0)
 893                         return ret;
 894                 if (for_all_proc_cpus(cpu_is_not_present))
 895                         return -ERR_CPU_NOT_PRESENT;
 896         }
 897
 898         if (!initialized) {
 899                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 900                         return ret;
 901                 gettimeofday(&tv_even, (struct timezone *)NULL);
 902                 is_even = 1;
 903                 initialized = 1;
 904                 return 0;
 905         }
 906
 907         if (is_even) {
 908                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 909                         return ret;
 910                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 911                 is_even = 0;
 912                 timersub(&tv_odd, &tv_even, &tv_delta);
 913                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 914                         return ret;
 915                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 916                         return ret;
 917         } else {
 918                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 919                         return ret;
 920                 gettimeofday(&tv_even, (struct timezone *)NULL);
 921                 is_even = 1;
 922                 timersub(&tv_even, &tv_odd, &tv_delta);
 923                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 924                         return ret;
 925                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 926                         return ret;
 927         }
 928         return 0;
 929 }
 930
 931 static int __attribute__((warn_unused_result))
 932 check_dev_msr()
 933 {
 934         struct stat sb;
 935
 936         if (stat("/dev/cpu/0/msr", &sb)) {
 937                 ERROR("no /dev/cpu/0/msr\n"
 938                         "Try \"# modprobe msr\"");
 939                 return -ERR_NO_MSR;
 940         }
 941         return 0;
 942 }
 943
 944 static int __attribute__((warn_unused_result))
 945 check_super_user()
 946 {
 947         if (getuid() != 0) {
 948                 ERROR("must be root");
 949                 return -ERR_NOT_ROOT;
 950         }
 951         return 0;
 952 }
 953
 954 /*
 955  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
 956  * the Thermal Control Circuit (TCC) activates.
 957  * This is usually equal to tjMax.
 958  *
 959  * Older processors do not have this MSR, so there we guess,
 960  * but also allow cmdline over-ride with -T.
 961  *
 962  * Several MSR temperature values are in units of degrees-C
 963  * below this value, including the Digital Thermal Sensor (DTS),
 964  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
 965  */
 966 static int __attribute__((warn_unused_result))
 967 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 968 {
 969         unsigned long long msr;
 970         unsigned int target_c_local;
 971         int cpu;
 972
 973         /* tcc_activation_temp is used only for dts or ptm */
 974         if (!(do_dts || do_ptm))
 975                 return 0;
 976
 977         /* this is a per-package concept */
 978         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 979                 return 0;
 980
 981         cpu = t->cpu_id;
 982         if (cpu_migrate(cpu)) {
 983                 ERROR("Could not migrate to CPU %d\n", cpu);
 984                 return -ERR_CPU_MIGRATE;
 985         }
 986
 987         if (tcc_activation_temp_override != 0) {
 988                 tcc_activation_temp = tcc_activation_temp_override;
 989                 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
 990                         cpu, tcc_activation_temp);
 991                 return 0;
 992         }
 993
 994         if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
 995                 goto guess;
 996
 997         target_c_local = (msr >> 16) & 0x7F;
 998
 999         if (target_c_local < 85 || target_c_local > 127)
1000                 goto guess;
1001
1002         tcc_activation_temp = target_c_local;
1003
1004         return 0;
1005
1006 guess:
1007         tcc_activation_temp = TJMAX_DEFAULT;
1008         WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1009                 cpu, tcc_activation_temp);
1010
1011         return 0;
1012 }
1013
1014 /*
1015  * Identify the functionality of the CPU
1016  */
1017 static int __attribute__((warn_unused_result))
1018 probe_cpu()
1019 {
1020         unsigned int eax, ebx, ecx, edx, max_level;
1021         unsigned int fms, family, model;
1022
1023         /* CPUID(0):
1024          * - EAX: Maximum Input Value for Basic CPUID Information
1025          * - EBX: "Genu" (0x756e6547)
1026          * - EDX: "ineI" (0x49656e69)
1027          * - ECX: "ntel" (0x6c65746e)
1028          */
1029         max_level = ebx = ecx = edx = 0;
1030         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1031         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1032                 ERROR("Unsupported CPU");
1033                 return -UNSUPPORTED_CPU;
1034         }
1035
1036         /* CPUID(1):
1037          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1038          *  + 4-7:   Model ID
1039          *  + 8-11:  Family ID
1040          *  + 12-13: Processor type
1041          *  + 16-19: Extended Model ID
1042          *  + 20-27: Extended Family ID
1043          * - EDX: Feature Information:
1044          *  + 5: Support for MSR read/write operations
1045          */
1046         fms = ebx = ecx = edx = 0;
1047         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1048         family = (fms >> 8) & 0xf;
1049         model = (fms >> 4) & 0xf;
1050         if (family == 0xf)
1051                 family += (fms >> 20) & 0xf;
1052         if (family == 6 || family == 0xf)
1053                 model += ((fms >> 16) & 0xf) << 4;
1054         if (!(edx & (1 << 5))) {
1055                 ERROR("CPUID: no MSR");
1056                 return -ERR_NO_MSR;
1057         }
1058
1059         /*
1060          * CPUID(0x80000000):
1061          * - EAX: Maximum Input Value for Extended Function CPUID Information
1062          *
1063          * This allows us to verify if the CPUID(0x80000007) can be called
1064          *
1065          * This check is valid for both Intel and AMD.
1066          */
1067         max_level = ebx = ecx = edx = 0;
1068         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1069         if (max_level < 0x80000007) {
1070                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1071                 return -ERR_NO_INVARIANT_TSC;
1072         }
1073
1074         /*
1075          * CPUID(0x80000007):
1076          * - EDX:
1077          *  + 8: Invariant TSC available if set
1078          *
1079          * This check is valid for both Intel and AMD
1080          */
1081         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1082         if (!(edx & (1 << 8))) {
1083                 ERROR("No invariant TSC");
1084                 return -ERR_NO_INVARIANT_TSC;
1085         }
1086
1087         /*
1088          * CPUID(6):
1089          * - EAX:
1090          *  + 0: Digital temperature sensor is supported if set
1091          *  + 6: Package thermal management is supported if set
1092          * - ECX:
1093          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1094          *  + 3: The processor supports performance-energy bias preference if set.
1095          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1096          *
1097          * This check is valid for both Intel and AMD
1098          */
1099         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1100         do_dts = eax & (1 << 0);
1101         do_ptm = eax & (1 << 6);
1102         if (!(ecx & (1 << 0))) {
1103                 ERROR("No APERF");
1104                 return -ERR_NO_APERF;
1105         }
1106
1107         /*
1108          * Enable or disable C states depending on the model and family
1109          */
1110         if (family == 6) {
1111                 switch (model) {
1112                 /* Atom (partial) */
1113                 case 0x27:
1114                         do_core_cstate = 0;
1115                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1116                         break;
1117                 /* Silvermont */
1118                 case 0x37: /* BYT */
1119                 case 0x4A:
1120                 case 0x4D: /* AVN */
1121                 case 0x5A:
1122                 case 0x5D:
1123                         do_core_cstate = (1 << 1) | (1 << 6);
1124                         do_pkg_cstate = (1 << 6);
1125                         break;
1126                 /* Nehalem */
1127                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1128                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1129                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1130                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1131                         do_core_cstate = (1 << 3) | (1 << 6);
1132                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1133                         break;
1134                 /* Westmere */
1135                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1136                 case 0x2C: /* Westmere EP - Gulftown */
1137                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1138                         do_core_cstate = (1 << 3) | (1 << 6);
1139                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1140                         break;
1141                 /* Sandy Bridge */
1142                 case 0x2A: /* SNB */
1143                 case 0x2D: /* SNB Xeon */
1144                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1145                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1146                         break;
1147                 /* Ivy Bridge */
1148                 case 0x3A: /* IVB */
1149                 case 0x3E: /* IVB Xeon */
1150                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1151                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1152                         break;
1153                 /* Haswell Bridge */
1154                 case 0x3C: /* HSW */
1155                 case 0x3F: /* HSW */
1156                 case 0x46: /* HSW */
1157                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1158                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1159                         break;
1160                 case 0x45: /* HSW */
1161                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1162                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1163                         break;
1164                 /* Broadwel */
1165                 case 0x4F: /* BDW */
1166                 case 0x56: /* BDX-DE */
1167                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1168                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1169                         break;
1170                 case 0x3D: /* BDW */
1171                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1172                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1173                         break;
1174                 default:
1175                         ERROR("Unsupported CPU");
1176                 }
1177                 switch (model) {
1178                 case 0x2A:
1179                 case 0x3A:
1180                 case 0x3C:
1181                 case 0x45:
1182                 case 0x46:
1183                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1184                         break;
1185                 case 0x3F:
1186                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1187                         break;
1188                 case 0x2D:
1189                 case 0x3E:
1190                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1191                         break;
1192                 case 0x37:
1193                 case 0x4D:
1194                         do_rapl = RAPL_PKG | RAPL_CORES;
1195                         break;
1196                 default:
1197                         do_rapl = 0;
1198                 }
1199         } else {
1200                 ERROR("Unsupported CPU");
1201                 return -UNSUPPORTED_CPU;
1202         }
1203
1204         if (do_rapl) {
1205                 unsigned long msr;
1206                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1207                         return 0;
1208
1209                 if (model == 0x37)
1210                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1211                 else
1212                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1213         }
1214
1215         return 0;
1216 }
1217
1218
1219
1220 static int __attribute__((warn_unused_result))
1221 topology_probe()
1222 {
1223         int i;
1224         int ret;
1225         int max_core_id = 0;
1226         int max_package_id = 0;
1227         int max_siblings = 0;
1228         struct cpu_topology {
1229                 int core_id;
1230                 int physical_package_id;
1231         } *cpus;
1232
1233         /* Initialize num_cpus, max_cpu_num */
1234         topo.num_cpus = 0;
1235         topo.max_cpu_num = 0;
1236         ret = for_all_proc_cpus(count_cpus);
1237         if (ret < 0)
1238                 return ret;
1239
1240         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1241
1242         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1243         if (cpus == NULL) {
1244                 ERROR("calloc cpus");
1245                 return -ERR_CALLOC;
1246         }
1247
1248         /*
1249          * Allocate and initialize cpu_present_set
1250          */
1251         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1252         if (cpu_present_set == NULL) {
1253                 free(cpus);
1254                 ERROR("CPU_ALLOC");
1255                 return -ERR_CPU_ALLOC;
1256         }
1257         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1258         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1259         ret = for_all_proc_cpus(mark_cpu_present);
1260         if (ret < 0) {
1261                 free(cpus);
1262                 return ret;
1263         }
1264
1265         /*
1266          * Allocate and initialize cpu_affinity_set
1267          */
1268         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1269         if (cpu_affinity_set == NULL) {
1270                 free(cpus);
1271                 ERROR("CPU_ALLOC");
1272                 return -ERR_CPU_ALLOC;
1273         }
1274         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1275         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1276
1277
1278         /*
1279          * For online cpus
1280          * find max_core_id, max_package_id
1281          */
1282         for (i = 0; i <= topo.max_cpu_num; ++i) {
1283                 int siblings;
1284
1285                 if (cpu_is_not_present(i)) {
1286                         //if (verbose > 1)
1287                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1288                         continue;
1289                 }
1290                 cpus[i].core_id = get_core_id(i);
1291                 if (cpus[i].core_id < 0)
1292                         return cpus[i].core_id;
1293                 if (cpus[i].core_id > max_core_id)
1294                         max_core_id = cpus[i].core_id;
1295
1296                 cpus[i].physical_package_id = get_physical_package_id(i);
1297                 if (cpus[i].physical_package_id < 0)
1298                         return cpus[i].physical_package_id;
1299                 if (cpus[i].physical_package_id > max_package_id)
1300                         max_package_id = cpus[i].physical_package_id;
1301
1302                 siblings = get_num_ht_siblings(i);
1303                 if (siblings < 0)
1304                         return siblings;
1305                 if (siblings > max_siblings)
1306                         max_siblings = siblings;
1307                 DEBUG("cpu %d pkg %d core %d\n",
1308                         i, cpus[i].physical_package_id, cpus[i].core_id);
1309         }
1310         topo.num_cores_per_pkg = max_core_id + 1;
1311         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1312                 max_core_id, topo.num_cores_per_pkg);
1313
1314         topo.num_packages = max_package_id + 1;
1315         DEBUG("max_package_id %d, sizing for %d packages\n",
1316                 max_package_id, topo.num_packages);
1317
1318         topo.num_threads_per_core = max_siblings;
1319         DEBUG("max_siblings %d\n", max_siblings);
1320
1321         free(cpus);
1322         return 0;
1323 }
1324
1325 static int
1326 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1327 {
1328         int i;
1329
1330         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1331                 topo.num_packages, sizeof(struct thread_data));
1332         if (*t == NULL)
1333                 goto error;
1334
1335         for (i = 0; i < topo.num_threads_per_core *
1336                 topo.num_cores_per_pkg * topo.num_packages; i++)
1337                 (*t)[i].cpu_id = -1;
1338
1339         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1340                 sizeof(struct core_data));
1341         if (*c == NULL)
1342                 goto error;
1343
1344         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1345                 (*c)[i].core_id = -1;
1346
1347         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1348         if (*p == NULL)
1349                 goto error;
1350
1351         for (i = 0; i < topo.num_packages; i++)
1352                 (*p)[i].package_id = i;
1353
1354         return 0;
1355 error:
1356         ERROR("calloc counters");
1357         return -ERR_CALLOC;
1358 }
1359 /*
1360  * init_counter()
1361  *
1362  * set cpu_id, core_num, pkg_num
1363  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1364  *
1365  * increment topo.num_cores when 1st core in pkg seen
1366  */
1367 static int
1368 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1369         struct pkg_data *pkg_base, int thread_num, int core_num,
1370         int pkg_num, int cpu_id)
1371 {
1372         int ret;
1373         struct thread_data *t;
1374         struct core_data *c;
1375         struct pkg_data *p;
1376
1377         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1378         c = GET_CORE(core_base, core_num, pkg_num);
1379         p = GET_PKG(pkg_base, pkg_num);
1380
1381         t->cpu_id = cpu_id;
1382         if (thread_num == 0) {
1383                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1384                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1385                         return ret;
1386                 } else if (ret != 0) {
1387                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1388                 }
1389         }
1390
1391         c->core_id = core_num;
1392         p->package_id = pkg_num;
1393
1394         return 0;
1395 }
1396
1397
1398 static int
1399 initialize_counters(int cpu_id)
1400 {
1401         int my_thread_id, my_core_id, my_package_id;
1402         int ret;
1403
1404         my_package_id = get_physical_package_id(cpu_id);
1405         if (my_package_id < 0)
1406                 return my_package_id;
1407         my_core_id = get_core_id(cpu_id);
1408         if (my_core_id < 0)
1409                 return my_core_id;
1410
1411         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1412                 return ret;
1413         } else if (ret != 0) {
1414                 my_thread_id = 0;
1415                 topo.num_cores++;
1416         } else {
1417                 my_thread_id = 1;
1418         }
1419
1420         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1421         if (ret < 0)
1422                 return ret;
1423         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1424         if (ret < 0)
1425                 return ret;
1426         return 0;
1427 }
1428
1429 #define DO_OR_GOTO_ERR(something) \
1430 do {                         \
1431         ret = (something);     \
1432         if (ret < 0)         \
1433                 goto err;    \
1434 } while (0)
1435
1436 static int setup_all_buffers(void)
1437 {
1438         int ret;
1439
1440         DO_OR_GOTO_ERR(topology_probe());
1441         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1442         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1443         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1444
1445         allocated = 1;
1446         return 0;
1447 err:
1448         free_all_buffers();
1449         return ret;
1450 }
1451
1452 static int
1453 turbostat_init(void)
1454 {
1455         int ret;
1456
1457         DO_OR_GOTO_ERR(check_super_user());
1458         DO_OR_GOTO_ERR(probe_cpu());
1459         DO_OR_GOTO_ERR(check_dev_msr());
1460         DO_OR_GOTO_ERR(setup_all_buffers());
1461         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1462
1463         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1464
1465         return 0;
1466 err:
1467         free_all_buffers();
1468         return ret;
1469 }
1470
1471 void module_register(void);
1472 void module_register(void)
1473 {
1474         plugin_register_init(PLUGIN_NAME, turbostat_init);
1475 }