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