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
  65 /*
  66  * If set, aperf_mperf_unstable disables a/mperf based stats.
  67  * This includes: C0 & C1 states, frequency
  68  *
  69  * This value is automatically set if mperf or aperf decreases
  70  */
  71 static _Bool aperf_mperf_unstable;
  72
  73 static unsigned int do_core_cstate;
  74 static unsigned int do_pkg_cstate;
  75 static unsigned int do_rapl;
  76 static unsigned int do_dts;
  77 static unsigned int do_ptm;
  78 static unsigned int tcc_activation_temp;
  79 static double rapl_energy_units;
  80
  81 #define RAPL_PKG                (1 << 0)
  82                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  83                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  84 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  85                                         /* 0x613 MSR_PKG_PERF_STATUS */
  86 #define RAPL_PKG_POWER_INFO     (1 << 2)
  87                                         /* 0x614 MSR_PKG_POWER_INFO */
  88
  89 #define RAPL_DRAM               (1 << 3)
  90                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
  91                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
  92                                         /* 0x61c MSR_DRAM_POWER_INFO */
  93 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
  94                                         /* 0x61b MSR_DRAM_PERF_STATUS */
  95
  96 #define RAPL_CORES              (1 << 5)
  97                                         /* 0x638 MSR_PP0_POWER_LIMIT */
  98                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
  99 #define RAPL_CORE_POLICY        (1 << 6)
 100                                         /* 0x63a MSR_PP0_POLICY */
 101
 102
 103 #define RAPL_GFX                (1 << 7)
 104                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 105                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 106                                         /* 0x642 MSR_PP1_POLICY */
 107 #define TJMAX_DEFAULT   100
 108
 109 int backwards_count;
 110
 111 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
 112 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
 113
 114 struct thread_data {
 115         unsigned long long tsc;
 116         unsigned long long aperf;
 117         unsigned long long mperf;
 118         unsigned long long c1;
 119         unsigned int smi_count;
 120         unsigned int cpu_id;
 121         unsigned int flags;
 122 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 123 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 124 } *thread_even, *thread_odd;
 125
 126 struct core_data {
 127         unsigned long long c3;
 128         unsigned long long c6;
 129         unsigned long long c7;
 130         unsigned int core_temp_c;
 131         unsigned int core_id;
 132 } *core_even, *core_odd;
 133
 134 struct pkg_data {
 135         unsigned long long pc2;
 136         unsigned long long pc3;
 137         unsigned long long pc6;
 138         unsigned long long pc7;
 139         unsigned long long pc8;
 140         unsigned long long pc9;
 141         unsigned long long pc10;
 142         unsigned int package_id;
 143         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 144         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 145         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 146         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 147         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 148         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 149         unsigned int tcc_activation_temp;
 150         unsigned int pkg_temp_c;
 151 } *package_even, *package_odd;
 152
 153 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 154 #define EVEN_COUNTERS thread_even, core_even, package_even
 155 static _Bool is_even = 1;
 156
 157 static _Bool allocated = 0;
 158 static _Bool initialized = 0;
 159
 160 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 161         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 162                 topo.num_threads_per_core + \
 163                 (core_no) * topo.num_threads_per_core + (thread_no))
 164 #define GET_CORE(core_base, core_no, pkg_no) \
 165         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 166 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 167
 168 struct topo_params {
 169         int num_packages;
 170         int num_cpus;
 171         int num_cores;
 172         int max_cpu_num;
 173         int num_cores_per_pkg;
 174         int num_threads_per_core;
 175 } topo;
 176
 177 struct timeval tv_even, tv_odd, tv_delta;
 178
 179 enum return_values {
 180         OK = 0,
 181         ERR_CPU_MIGRATE,
 182         ERR_CPU_SAVE_SCHED_AFFINITY,
 183         ERR_MSR_IA32_APERF,
 184         ERR_MSR_IA32_MPERF,
 185         ERR_MSR_SMI_COUNT,
 186         ERR_MSR_CORE_C3_RESIDENCY,
 187         ERR_MSR_CORE_C6_RESIDENCY,
 188         ERR_MSR_CORE_C7_RESIDENCY,
 189         ERR_MSR_IA32_THERM_STATUS,
 190         ERR_MSR_PKG_C3_RESIDENCY,
 191         ERR_MSR_PKG_C6_RESIDENCY,
 192         ERR_MSR_PKG_C2_RESIDENCY,
 193         ERR_MSR_PKG_C7_RESIDENCY,
 194         ERR_MSR_PKG_C8_RESIDENCY,
 195         ERR_MSR_PKG_C9_RESIDENCY,
 196         ERR_MSR_PKG_C10_RESIDENCY,
 197         ERR_MSR_PKG_ENERGY_STATUS,
 198         ERR_MSR_PKG_POWER_INFO,
 199         ERR_MSR_PP0_ENERGY_STATUS,
 200         ERR_MSR_DRAM_ENERGY_STATUS,
 201         ERR_MSR_PP1_ENERGY_STATUS,
 202         ERR_MSR_PKG_PERF_STATUS,
 203         ERR_MSR_DRAM_PERF_STATUS,
 204         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 205         ERR_MSR_IA32_TSC,
 206         ERR_CPU_NOT_PRESENT,
 207         ERR_NO_MSR,
 208         ERR_CANT_OPEN_MSR,
 209         ERR_CANT_OPEN_FILE,
 210         ERR_CANT_READ_NUMBER,
 211         ERR_CANT_READ_PROC_STAT,
 212         ERR_NO_INVARIANT_TSC,
 213         ERR_NO_APERF,
 214         ERR_CALLOC,
 215         ERR_CPU_ALLOC,
 216         ERR_NOT_ROOT,
 217         UNSUPPORTED_CPU,
 218 };
 219
 220 static int setup_all_buffers(void);
 221
 222 static int
 223 cpu_is_not_present(int cpu)
 224 {
 225         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 226 }
 227 /*
 228  * run func(thread, core, package) in topology order
 229  * skip non-present cpus
 230  */
 231
 232 static int __attribute__((warn_unused_result))
 233 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 234         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 235 {
 236         int retval, pkg_no, core_no, thread_no;
 237
 238         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 239                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 240                         for (thread_no = 0; thread_no <
 241                                 topo.num_threads_per_core; ++thread_no) {
 242                                 struct thread_data *t;
 243                                 struct core_data *c;
 244                                 struct pkg_data *p;
 245
 246                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 247
 248                                 if (cpu_is_not_present(t->cpu_id))
 249                                         continue;
 250
 251                                 c = GET_CORE(core_base, core_no, pkg_no);
 252                                 p = GET_PKG(pkg_base, pkg_no);
 253
 254                                 retval = func(t, c, p);
 255                                 if (retval)
 256                                         return retval;
 257                         }
 258                 }
 259         }
 260         return 0;
 261 }
 262
 263 static int __attribute__((warn_unused_result))
 264 open_msr(int cpu, _Bool multiple_read)
 265 {
 266         char pathname[32];
 267         int fd;
 268
 269         /*
 270          * If we need to do multiple read, let's migrate to the CPU
 271          * Otherwise, we would lose time calling functions on another CPU
 272          *
 273          * If we are not yet initialized (cpu_affinity_setsize = 0),
 274          * we need to skip this optimisation.
 275          */
 276         if (multiple_read && cpu_affinity_setsize) {
 277                 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 278                 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 279                 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
 280                         ERROR("Could not migrate to CPU %d", cpu);
 281                         return -ERR_CPU_MIGRATE;
 282                 }
 283         }
 284
 285         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 286         fd = open(pathname, O_RDONLY);
 287         if (fd < 0)
 288                 return -ERR_CANT_OPEN_MSR;
 289         return fd;
 290 }
 291
 292 static int __attribute__((warn_unused_result))
 293 read_msr(int fd, off_t offset, unsigned long long *msr)
 294 {
 295         ssize_t retval;
 296
 297         retval = pread(fd, msr, sizeof *msr, offset);
 298
 299         if (retval != sizeof *msr) {
 300                 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
 301                 return -1;
 302         }
 303         return 0;
 304 }
 305
 306 static int __attribute__((warn_unused_result))
 307 get_msr(int cpu, off_t offset, unsigned long long *msr)
 308 {
 309         ssize_t retval;
 310         int fd;
 311
 312         fd = open_msr(cpu, 0);
 313         if (fd < 0)
 314                 return fd;
 315         retval = read_msr(fd, offset, msr);
 316         close(fd);
 317         return retval;
 318 }
 319
 320 #define DELTA_WRAP32(new, old)                  \
 321         if (new > old) {                        \
 322                 old = new - old;                \
 323         } else {                                \
 324                 old = 0x100000000 + new - old;  \
 325         }
 326
 327 static void
 328 delta_package(struct pkg_data *new, struct pkg_data *old)
 329 {
 330         old->pc2 = new->pc2 - old->pc2;
 331         old->pc3 = new->pc3 - old->pc3;
 332         old->pc6 = new->pc6 - old->pc6;
 333         old->pc7 = new->pc7 - old->pc7;
 334         old->pc8 = new->pc8 - old->pc8;
 335         old->pc9 = new->pc9 - old->pc9;
 336         old->pc10 = new->pc10 - old->pc10;
 337         old->pkg_temp_c = new->pkg_temp_c;
 338
 339         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 340         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 341         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 342         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 343         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 344         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 345 }
 346
 347 static void
 348 delta_core(struct core_data *new, struct core_data *old)
 349 {
 350         old->c3 = new->c3 - old->c3;
 351         old->c6 = new->c6 - old->c6;
 352         old->c7 = new->c7 - old->c7;
 353         old->core_temp_c = new->core_temp_c;
 354 }
 355
 356 /*
 357  * old = new - old
 358  */
 359 static int __attribute__((warn_unused_result))
 360 delta_thread(struct thread_data *new, struct thread_data *old,
 361         struct core_data *core_delta)
 362 {
 363         old->tsc = new->tsc - old->tsc;
 364
 365         /* check for TSC < 1 Mcycles over interval */
 366         if (old->tsc < (1000 * 1000)) {
 367                 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
 368                 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
 369                 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
 370                 return -1;
 371         }
 372
 373         old->c1 = new->c1 - old->c1;
 374
 375         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 376                 old->aperf = new->aperf - old->aperf;
 377                 old->mperf = new->mperf - old->mperf;
 378         } else {
 379                 if (!aperf_mperf_unstable) {
 380                         WARNING(" APERF or MPERF went backwards * ");
 381                         WARNING("* Frequency results do not cover entire interval *");
 382                         WARNING("* fix this by running Linux-2.6.30 or later *");
 383
 384                         aperf_mperf_unstable = 1;
 385                 }
 386         }
 387
 388
 389         /*
 390          * As counter collection is not atomic,
 391          * it is possible for mperf's non-halted cycles + idle states
 392          * to exceed TSC's all cycles: show c1 = 0% in that case.
 393          */
 394         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 395                 old->c1 = 0;
 396         else {
 397                 /* normal case, derive c1 */
 398                 old->c1 = old->tsc - old->mperf - core_delta->c3
 399                         - core_delta->c6 - core_delta->c7;
 400         }
 401
 402         if (old->mperf == 0) {
 403                 WARNING("cpu%d MPERF 0!", old->cpu_id);
 404                 old->mperf = 1; /* divide by 0 protection */
 405         }
 406
 407         old->smi_count = new->smi_count - old->smi_count;
 408
 409         return 0;
 410 }
 411
 412 static int __attribute__((warn_unused_result))
 413 delta_cpu(struct thread_data *t, struct core_data *c,
 414         struct pkg_data *p, struct thread_data *t2,
 415         struct core_data *c2, struct pkg_data *p2)
 416 {
 417         int ret;
 418
 419         /* calculate core delta only for 1st thread in core */
 420         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 421                 delta_core(c, c2);
 422
 423         /* always calculate thread delta */
 424         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 425         if (ret != 0)
 426                 return ret;
 427
 428         /* calculate package delta only for 1st core in package */
 429         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 430                 delta_package(p, p2);
 431
 432         return 0;
 433 }
 434
 435 /*
 436  * get_counters(...)
 437  * migrate to cpu
 438  * acquire and record local counters for that cpu
 439  */
 440 static int __attribute__((warn_unused_result))
 441 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 442 {
 443         int cpu = t->cpu_id;
 444         unsigned long long msr;
 445         int msr_fd;
 446         int retval = 0;
 447
 448         msr_fd = open_msr(cpu, 1);
 449         if (msr_fd < 0)
 450                 return msr_fd;
 451
 452 #define READ_MSR(msr, dst)                      \
 453 do {                                            \
 454         if (read_msr(msr_fd, msr, dst)) {       \
 455                 retval = -ERR_##msr;            \
 456                 goto out;                       \
 457         }                                       \
 458 } while (0)
 459
 460         READ_MSR(MSR_IA32_TSC, &t->tsc);
 461
 462         READ_MSR(MSR_IA32_APERF, &t->aperf);
 463         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 464
 465         READ_MSR(MSR_SMI_COUNT, &msr);
 466         t->smi_count = msr & 0xFFFFFFFF;
 467
 468         /* collect core counters only for 1st thread in core */
 469         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 470                 retval = 0;
 471                 goto out;
 472         }
 473
 474         if (do_core_cstate & (1 << 3))
 475                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 476         if (do_core_cstate & (1 << 6))
 477                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 478         if (do_core_cstate & (1 << 7))
 479                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 480
 481         if (do_dts) {
 482                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 483                 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 484         }
 485
 486         /* collect package counters only for 1st core in package */
 487         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 488                 retval = 0;
 489                 goto out;
 490         }
 491
 492         if (do_pkg_cstate & (1 << 2))
 493                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 494         if (do_pkg_cstate & (1 << 3))
 495                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 496         if (do_pkg_cstate & (1 << 6))
 497                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 498         if (do_pkg_cstate & (1 << 7))
 499                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 500         if (do_pkg_cstate & (1 << 8))
 501                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 502         if (do_pkg_cstate & (1 << 9))
 503                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 504         if (do_pkg_cstate & (1 << 10))
 505                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 506
 507         if (do_rapl & RAPL_PKG) {
 508                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 509                 p->energy_pkg = msr & 0xFFFFFFFF;
 510         }
 511         if (do_rapl & RAPL_CORES) {
 512                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 513                 p->energy_cores = msr & 0xFFFFFFFF;
 514         }
 515         if (do_rapl & RAPL_DRAM) {
 516                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 517                 p->energy_dram = msr & 0xFFFFFFFF;
 518         }
 519         if (do_rapl & RAPL_GFX) {
 520                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 521                 p->energy_gfx = msr & 0xFFFFFFFF;
 522         }
 523         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 524                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 525                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 526         }
 527         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 528                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 529                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 530         }
 531         if (do_ptm) {
 532                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 533                 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 534         }
 535
 536 out:
 537         close(msr_fd);
 538         return retval;
 539 }
 540
 541 static void
 542 free_all_buffers(void)
 543 {
 544         allocated = 0;
 545         initialized = 0;
 546
 547         CPU_FREE(cpu_present_set);
 548         cpu_present_set = NULL;
 549         cpu_present_set = 0;
 550
 551         CPU_FREE(cpu_affinity_set);
 552         cpu_affinity_set = NULL;
 553         cpu_affinity_setsize = 0;
 554
 555         CPU_FREE(cpu_saved_affinity_set);
 556         cpu_saved_affinity_set = NULL;
 557         cpu_saved_affinity_setsize = 0;
 558
 559         free(thread_even);
 560         free(core_even);
 561         free(package_even);
 562
 563         thread_even = NULL;
 564         core_even = NULL;
 565         package_even = NULL;
 566
 567         free(thread_odd);
 568         free(core_odd);
 569         free(package_odd);
 570
 571         thread_odd = NULL;
 572         core_odd = NULL;
 573         package_odd = NULL;
 574 }
 575
 576 /*
 577  * Parse a file containing a single int.
 578  */
 579 static int __attribute__ ((format(printf,1,2)))
 580 parse_int_file(const char *fmt, ...)
 581 {
 582         va_list args;
 583         char path[PATH_MAX];
 584         FILE *filep;
 585         int value;
 586
 587         va_start(args, fmt);
 588         vsnprintf(path, sizeof(path), fmt, args);
 589         va_end(args);
 590         filep = fopen(path, "r");
 591         if (!filep) {
 592                 ERROR("%s: open failed", path);
 593                 return -ERR_CANT_OPEN_FILE;
 594         }
 595         if (fscanf(filep, "%d", &value) != 1) {
 596                 ERROR("%s: failed to parse number from file", path);
 597                 return -ERR_CANT_READ_NUMBER;
 598         }
 599         fclose(filep);
 600         return value;
 601 }
 602
 603 /*
 604  * cpu_is_first_sibling_in_core(cpu)
 605  * return 1 if given CPU is 1st HT sibling in the core
 606  */
 607 static int
 608 cpu_is_first_sibling_in_core(int cpu)
 609 {
 610         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 611 }
 612
 613 /*
 614  * cpu_is_first_core_in_package(cpu)
 615  * return 1 if given CPU is 1st core in package
 616  */
 617 static int
 618 cpu_is_first_core_in_package(int cpu)
 619 {
 620         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 621 }
 622
 623 static int
 624 get_physical_package_id(int cpu)
 625 {
 626         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 627 }
 628
 629 static int
 630 get_core_id(int cpu)
 631 {
 632         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 633 }
 634
 635 static int
 636 get_num_ht_siblings(int cpu)
 637 {
 638         char path[80];
 639         FILE *filep;
 640         int sib1, sib2;
 641         int matches;
 642         char character;
 643
 644         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 645         filep = fopen(path, "r");
 646         if (!filep) {
 647                 ERROR("%s: open failed", path);
 648                 return -ERR_CANT_OPEN_FILE;
 649         }
 650         /*
 651          * file format:
 652          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 653          * otherwinse 1 sibling (self).
 654          */
 655         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 656
 657         fclose(filep);
 658
 659         if (matches == 3)
 660                 return 2;
 661         else
 662                 return 1;
 663 }
 664
 665 /*
 666  * run func(thread, core, package) in topology order
 667  * skip non-present cpus
 668  */
 669
 670
 671 static int __attribute__((warn_unused_result))
 672 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 673         struct pkg_data *, struct thread_data *, struct core_data *,
 674         struct pkg_data *), struct thread_data *thread_base,
 675         struct core_data *core_base, struct pkg_data *pkg_base,
 676         struct thread_data *thread_base2, struct core_data *core_base2,
 677         struct pkg_data *pkg_base2)
 678 {
 679         int retval, pkg_no, core_no, thread_no;
 680
 681         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 682                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 683                         for (thread_no = 0; thread_no <
 684                                 topo.num_threads_per_core; ++thread_no) {
 685                                 struct thread_data *t, *t2;
 686                                 struct core_data *c, *c2;
 687                                 struct pkg_data *p, *p2;
 688
 689                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 690
 691                                 if (cpu_is_not_present(t->cpu_id))
 692                                         continue;
 693
 694                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 695
 696                                 c = GET_CORE(core_base, core_no, pkg_no);
 697                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 698
 699                                 p = GET_PKG(pkg_base, pkg_no);
 700                                 p2 = GET_PKG(pkg_base2, pkg_no);
 701
 702                                 retval = func(t, c, p, t2, c2, p2);
 703                                 if (retval)
 704                                         return retval;
 705                         }
 706                 }
 707         }
 708         return 0;
 709 }
 710
 711 /*
 712  * run func(cpu) on every cpu in /proc/stat
 713  * return max_cpu number
 714  */
 715 static int __attribute__((warn_unused_result))
 716 for_all_proc_cpus(int (func)(int))
 717 {
 718         FILE *fp;
 719         int cpu_num;
 720         int retval;
 721
 722         fp = fopen(proc_stat, "r");
 723         if (!fp) {
 724                 ERROR("%s: open failed", proc_stat);
 725                 return -ERR_CANT_OPEN_FILE;
 726         }
 727
 728         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 729         if (retval != 0) {
 730                 ERROR("%s: failed to parse format", proc_stat);
 731                 return -ERR_CANT_READ_PROC_STAT;
 732         }
 733
 734         while (1) {
 735                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 736                 if (retval != 1)
 737                         break;
 738
 739                 retval = func(cpu_num);
 740                 if (retval) {
 741                         fclose(fp);
 742                         return(retval);
 743                 }
 744         }
 745         fclose(fp);
 746         return 0;
 747 }
 748
 749 /*
 750  * count_cpus()
 751  * remember the last one seen, it will be the max
 752  */
 753 static int
 754 count_cpus(int cpu)
 755 {
 756         if (topo.max_cpu_num < cpu)
 757                 topo.max_cpu_num = cpu;
 758
 759         topo.num_cpus += 1;
 760         return 0;
 761 }
 762 static int
 763 mark_cpu_present(int cpu)
 764 {
 765         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 766         return 0;
 767 }
 768
 769
 770 static void
 771 turbostat_submit (const char *plugin_instance,
 772         const char *type, const char *type_instance,
 773         gauge_t value)
 774 {
 775         value_list_t vl = VALUE_LIST_INIT;
 776         value_t v;
 777
 778         v.gauge = value;
 779         vl.values = &v;
 780         vl.values_len = 1;
 781         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 782         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 783         if (plugin_instance != NULL)
 784                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 785         sstrncpy (vl.type, type, sizeof (vl.type));
 786         if (type_instance != NULL)
 787                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 788
 789         plugin_dispatch_values (&vl);
 790 }
 791
 792 /*
 793  * column formatting convention & formats
 794  * package: "pk" 2 columns %2d
 795  * core: "cor" 3 columns %3d
 796  * CPU: "CPU" 3 columns %3d
 797  * Pkg_W: %6.2
 798  * Cor_W: %6.2
 799  * GFX_W: %5.2
 800  * RAM_W: %5.2
 801  * GHz: "GHz" 3 columns %3.2
 802  * TSC: "TSC" 3 columns %3.2
 803  * SMI: "SMI" 4 columns %4d
 804  * percentage " %pc3" %6.2
 805  * Perf Status percentage: %5.2
 806  * "CTMP" 4 columns %4d
 807  */
 808 #define NAME_LEN 12
 809 static int
 810 submit_counters(struct thread_data *t, struct core_data *c,
 811         struct pkg_data *p)
 812 {
 813         char name[NAME_LEN];
 814         double interval_float;
 815
 816         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 817
 818         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 819
 820         if (!aperf_mperf_unstable)
 821                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 822         if (!aperf_mperf_unstable)
 823                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 824
 825         /* GHz */
 826         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 827                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 828
 829         /* SMI */
 830         turbostat_submit(NULL, "current", name, t->smi_count);
 831
 832         /* print per-core data only for 1st thread in core */
 833         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 834                 goto done;
 835
 836         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 837
 838         if (do_core_cstate & (1 << 3))
 839                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 840         if (do_core_cstate & (1 << 6))
 841                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 842         if (do_core_cstate & (1 << 7))
 843                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 844
 845         if (do_dts)
 846                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 847
 848         /* print per-package data only for 1st core in package */
 849         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 850                 goto done;
 851
 852         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 853
 854         if (do_ptm)
 855                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 856
 857         if (do_pkg_cstate & (1 << 2))
 858                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 859         if (do_pkg_cstate & (1 << 3))
 860                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 861         if (do_pkg_cstate & (1 << 6))
 862                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 863         if (do_pkg_cstate & (1 << 7))
 864                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 865         if (do_pkg_cstate & (1 << 8))
 866                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 867         if (do_pkg_cstate & (1 << 9))
 868                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 869         if (do_pkg_cstate & (1 << 10))
 870                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 871
 872         if (do_rapl) {
 873                 if (do_rapl & RAPL_PKG)
 874                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 875                 if (do_rapl & RAPL_CORES)
 876                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 877                 if (do_rapl & RAPL_GFX)
 878                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 879                 if (do_rapl & RAPL_DRAM)
 880                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 881         }
 882 done:
 883         return 0;
 884 }
 885
 886 static int
 887 turbostat_read(user_data_t * not_used)
 888 {
 889         int ret;
 890
 891         if (!allocated) {
 892                 if ((ret = setup_all_buffers()) < 0)
 893                         return ret;
 894         }
 895
 896         if (for_all_proc_cpus(cpu_is_not_present)) {
 897                 free_all_buffers();
 898                 if ((ret = setup_all_buffers()) < 0)
 899                         return ret;
 900                 if (for_all_proc_cpus(cpu_is_not_present))
 901                         return -ERR_CPU_NOT_PRESENT;
 902         }
 903
 904         /* Saving the scheduling affinity, as it will be modified by get_counters */
 905         if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0)
 906                 return -ERR_CPU_SAVE_SCHED_AFFINITY;
 907
 908         if (!initialized) {
 909                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 910                         goto out;
 911                 gettimeofday(&tv_even, (struct timezone *)NULL);
 912                 is_even = 1;
 913                 initialized = 1;
 914                 ret = 0;
 915                 goto out;
 916         }
 917
 918         if (is_even) {
 919                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 920                         goto out;
 921                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 922                 is_even = 0;
 923                 timersub(&tv_odd, &tv_even, &tv_delta);
 924                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 925                         goto out;
 926                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 927                         goto out;
 928         } else {
 929                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 930                         goto out;
 931                 gettimeofday(&tv_even, (struct timezone *)NULL);
 932                 is_even = 1;
 933                 timersub(&tv_even, &tv_odd, &tv_delta);
 934                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 935                         goto out;
 936                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 937                         goto out;
 938         }
 939         ret = 0;
 940 out:
 941         /*
 942          * Let's restore the affinity
 943          * This might fail if the number of CPU changed, but we can't do anything in that case..
 944          */
 945         (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
 946         return ret;
 947 }
 948
 949 static int __attribute__((warn_unused_result))
 950 check_dev_msr()
 951 {
 952         struct stat sb;
 953
 954         if (stat("/dev/cpu/0/msr", &sb)) {
 955                 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
 956                 return -ERR_NO_MSR;
 957         }
 958         return 0;
 959 }
 960
 961 static int __attribute__((warn_unused_result))
 962 check_super_user()
 963 {
 964         if (getuid() != 0) {
 965                 ERROR("must be root");
 966                 return -ERR_NOT_ROOT;
 967         }
 968         return 0;
 969 }
 970
 971 /*
 972  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
 973  * the Thermal Control Circuit (TCC) activates.
 974  * This is usually equal to tjMax.
 975  *
 976  * Older processors do not have this MSR, so there we guess,
 977  * but also allow cmdline over-ride with -T.
 978  *
 979  * Several MSR temperature values are in units of degrees-C
 980  * below this value, including the Digital Thermal Sensor (DTS),
 981  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
 982  */
 983 static int __attribute__((warn_unused_result))
 984 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 985 {
 986         unsigned long long msr;
 987         unsigned int target_c_local;
 988
 989         /* tcc_activation_temp is used only for dts or ptm */
 990         if (!(do_dts || do_ptm))
 991                 return 0;
 992
 993         /* this is a per-package concept */
 994         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 995                 return 0;
 996
 997         if (tcc_activation_temp != 0) {
 998                 p->tcc_activation_temp = tcc_activation_temp;
 999                 return 0;
1000         }
1001
1002         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
1003                 goto guess;
1004
1005         target_c_local = (msr >> 16) & 0x7F;
1006
1007         if (target_c_local < 85 || target_c_local > 127)
1008                 goto guess;
1009
1010         p->tcc_activation_temp = target_c_local;
1011
1012         return 0;
1013
1014 guess:
1015         p->tcc_activation_temp = TJMAX_DEFAULT;
1016         WARNING("cpu%d: Guessing tjMax %d C, Please use TCCActivationTemp to specify",
1017                 t->cpu_id, p->tcc_activation_temp);
1018
1019         return 0;
1020 }
1021
1022 /*
1023  * Identify the functionality of the CPU
1024  */
1025 static int __attribute__((warn_unused_result))
1026 probe_cpu()
1027 {
1028         unsigned int eax, ebx, ecx, edx, max_level;
1029         unsigned int fms, family, model;
1030
1031         /* CPUID(0):
1032          * - EAX: Maximum Input Value for Basic CPUID Information
1033          * - EBX: "Genu" (0x756e6547)
1034          * - EDX: "ineI" (0x49656e69)
1035          * - ECX: "ntel" (0x6c65746e)
1036          */
1037         max_level = ebx = ecx = edx = 0;
1038         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1039         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1040                 ERROR("Unsupported CPU");
1041                 return -UNSUPPORTED_CPU;
1042         }
1043
1044         /* CPUID(1):
1045          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1046          *  + 4-7:   Model ID
1047          *  + 8-11:  Family ID
1048          *  + 12-13: Processor type
1049          *  + 16-19: Extended Model ID
1050          *  + 20-27: Extended Family ID
1051          * - EDX: Feature Information:
1052          *  + 5: Support for MSR read/write operations
1053          */
1054         fms = ebx = ecx = edx = 0;
1055         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1056         family = (fms >> 8) & 0xf;
1057         model = (fms >> 4) & 0xf;
1058         if (family == 0xf)
1059                 family += (fms >> 20) & 0xf;
1060         if (family == 6 || family == 0xf)
1061                 model += ((fms >> 16) & 0xf) << 4;
1062         if (!(edx & (1 << 5))) {
1063                 ERROR("CPUID: no MSR");
1064                 return -ERR_NO_MSR;
1065         }
1066
1067         /*
1068          * CPUID(0x80000000):
1069          * - EAX: Maximum Input Value for Extended Function CPUID Information
1070          *
1071          * This allows us to verify if the CPUID(0x80000007) can be called
1072          *
1073          * This check is valid for both Intel and AMD.
1074          */
1075         max_level = ebx = ecx = edx = 0;
1076         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1077         if (max_level < 0x80000007) {
1078                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1079                 return -ERR_NO_INVARIANT_TSC;
1080         }
1081
1082         /*
1083          * CPUID(0x80000007):
1084          * - EDX:
1085          *  + 8: Invariant TSC available if set
1086          *
1087          * This check is valid for both Intel and AMD
1088          */
1089         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1090         if (!(edx & (1 << 8))) {
1091                 ERROR("No invariant TSC");
1092                 return -ERR_NO_INVARIANT_TSC;
1093         }
1094
1095         /*
1096          * CPUID(6):
1097          * - EAX:
1098          *  + 0: Digital temperature sensor is supported if set
1099          *  + 6: Package thermal management is supported if set
1100          * - ECX:
1101          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1102          *  + 3: The processor supports performance-energy bias preference if set.
1103          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1104          *
1105          * This check is valid for both Intel and AMD
1106          */
1107         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1108         do_dts = eax & (1 << 0);
1109         do_ptm = eax & (1 << 6);
1110         if (!(ecx & (1 << 0))) {
1111                 ERROR("No APERF");
1112                 return -ERR_NO_APERF;
1113         }
1114
1115         /*
1116          * Enable or disable C states depending on the model and family
1117          */
1118         if (family == 6) {
1119                 switch (model) {
1120                 /* Atom (partial) */
1121                 case 0x27:
1122                         do_core_cstate = 0;
1123                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1124                         break;
1125                 /* Silvermont */
1126                 case 0x37: /* BYT */
1127                 case 0x4A:
1128                 case 0x4D: /* AVN */
1129                 case 0x5A:
1130                 case 0x5D:
1131                         do_core_cstate = (1 << 1) | (1 << 6);
1132                         do_pkg_cstate = (1 << 6);
1133                         break;
1134                 /* Nehalem */
1135                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1136                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1137                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1138                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1139                         do_core_cstate = (1 << 3) | (1 << 6);
1140                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1141                         break;
1142                 /* Westmere */
1143                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1144                 case 0x2C: /* Westmere EP - Gulftown */
1145                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1146                         do_core_cstate = (1 << 3) | (1 << 6);
1147                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1148                         break;
1149                 /* Sandy Bridge */
1150                 case 0x2A: /* SNB */
1151                 case 0x2D: /* SNB Xeon */
1152                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1153                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1154                         break;
1155                 /* Ivy Bridge */
1156                 case 0x3A: /* IVB */
1157                 case 0x3E: /* IVB Xeon */
1158                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1159                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1160                         break;
1161                 /* Haswell Bridge */
1162                 case 0x3C: /* HSW */
1163                 case 0x3F: /* HSW */
1164                 case 0x46: /* HSW */
1165                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1166                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1167                         break;
1168                 case 0x45: /* HSW */
1169                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1170                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1171                         break;
1172                 /* Broadwel */
1173                 case 0x4F: /* BDW */
1174                 case 0x56: /* BDX-DE */
1175                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1176                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1177                         break;
1178                 case 0x3D: /* BDW */
1179                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1180                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1181                         break;
1182                 default:
1183                         ERROR("Unsupported CPU");
1184                 }
1185                 switch (model) {
1186                 case 0x2A:
1187                 case 0x3A:
1188                 case 0x3C:
1189                 case 0x45:
1190                 case 0x46:
1191                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1192                         break;
1193                 case 0x3F:
1194                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1195                         break;
1196                 case 0x2D:
1197                 case 0x3E:
1198                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1199                         break;
1200                 case 0x37:
1201                 case 0x4D:
1202                         do_rapl = RAPL_PKG | RAPL_CORES;
1203                         break;
1204                 default:
1205                         do_rapl = 0;
1206                 }
1207         } else {
1208                 ERROR("Unsupported CPU");
1209                 return -UNSUPPORTED_CPU;
1210         }
1211
1212         if (do_rapl) {
1213                 unsigned long msr;
1214                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1215                         return 0;
1216
1217                 if (model == 0x37)
1218                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1219                 else
1220                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1221         }
1222
1223         return 0;
1224 }
1225
1226
1227
1228 static int __attribute__((warn_unused_result))
1229 topology_probe()
1230 {
1231         int i;
1232         int ret;
1233         int max_core_id = 0;
1234         int max_package_id = 0;
1235         int max_siblings = 0;
1236         struct cpu_topology {
1237                 int core_id;
1238                 int physical_package_id;
1239         } *cpus;
1240
1241         /* Initialize num_cpus, max_cpu_num */
1242         topo.num_cpus = 0;
1243         topo.max_cpu_num = 0;
1244         ret = for_all_proc_cpus(count_cpus);
1245         if (ret < 0)
1246                 return ret;
1247
1248         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1249
1250         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1251         if (cpus == NULL) {
1252                 ERROR("calloc cpus");
1253                 return -ERR_CALLOC;
1254         }
1255
1256         /*
1257          * Allocate and initialize cpu_present_set
1258          */
1259         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1260         if (cpu_present_set == NULL) {
1261                 free(cpus);
1262                 ERROR("CPU_ALLOC");
1263                 return -ERR_CPU_ALLOC;
1264         }
1265         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1266         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1267         ret = for_all_proc_cpus(mark_cpu_present);
1268         if (ret < 0) {
1269                 free(cpus);
1270                 return ret;
1271         }
1272
1273         /*
1274          * Allocate and initialize cpu_affinity_set
1275          */
1276         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1277         if (cpu_affinity_set == NULL) {
1278                 free(cpus);
1279                 ERROR("CPU_ALLOC");
1280                 return -ERR_CPU_ALLOC;
1281         }
1282         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1283         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1284
1285
1286         /*
1287          * Allocate and initialize cpu_saved_affinity_set
1288          */
1289         cpu_saved_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1290         if (cpu_saved_affinity_set == NULL) {
1291                 free(cpus);
1292                 ERROR("CPU_ALLOC");
1293                 return -ERR_CPU_ALLOC;
1294         }
1295         cpu_saved_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1296         CPU_ZERO_S(cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1297
1298
1299         /*
1300          * For online cpus
1301          * find max_core_id, max_package_id
1302          */
1303         for (i = 0; i <= topo.max_cpu_num; ++i) {
1304                 int siblings;
1305
1306                 if (cpu_is_not_present(i)) {
1307                         WARNING("cpu%d NOT PRESENT", i);
1308                         continue;
1309                 }
1310                 cpus[i].core_id = get_core_id(i);
1311                 if (cpus[i].core_id < 0)
1312                         return cpus[i].core_id;
1313                 if (cpus[i].core_id > max_core_id)
1314                         max_core_id = cpus[i].core_id;
1315
1316                 cpus[i].physical_package_id = get_physical_package_id(i);
1317                 if (cpus[i].physical_package_id < 0)
1318                         return cpus[i].physical_package_id;
1319                 if (cpus[i].physical_package_id > max_package_id)
1320                         max_package_id = cpus[i].physical_package_id;
1321
1322                 siblings = get_num_ht_siblings(i);
1323                 if (siblings < 0)
1324                         return siblings;
1325                 if (siblings > max_siblings)
1326                         max_siblings = siblings;
1327                 DEBUG("cpu %d pkg %d core %d\n",
1328                         i, cpus[i].physical_package_id, cpus[i].core_id);
1329         }
1330         topo.num_cores_per_pkg = max_core_id + 1;
1331         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1332                 max_core_id, topo.num_cores_per_pkg);
1333
1334         topo.num_packages = max_package_id + 1;
1335         DEBUG("max_package_id %d, sizing for %d packages\n",
1336                 max_package_id, topo.num_packages);
1337
1338         topo.num_threads_per_core = max_siblings;
1339         DEBUG("max_siblings %d\n", max_siblings);
1340
1341         free(cpus);
1342         return 0;
1343 }
1344
1345 static int
1346 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1347 {
1348         int i;
1349
1350         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1351                 topo.num_packages, sizeof(struct thread_data));
1352         if (*t == NULL)
1353                 goto error;
1354
1355         for (i = 0; i < topo.num_threads_per_core *
1356                 topo.num_cores_per_pkg * topo.num_packages; i++)
1357                 (*t)[i].cpu_id = -1;
1358
1359         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1360                 sizeof(struct core_data));
1361         if (*c == NULL)
1362                 goto error;
1363
1364         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1365                 (*c)[i].core_id = -1;
1366
1367         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1368         if (*p == NULL)
1369                 goto error;
1370
1371         for (i = 0; i < topo.num_packages; i++)
1372                 (*p)[i].package_id = i;
1373
1374         return 0;
1375 error:
1376         ERROR("calloc counters");
1377         return -ERR_CALLOC;
1378 }
1379 /*
1380  * init_counter()
1381  *
1382  * set cpu_id, core_num, pkg_num
1383  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1384  *
1385  * increment topo.num_cores when 1st core in pkg seen
1386  */
1387 static int
1388 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1389         struct pkg_data *pkg_base, int thread_num, int core_num,
1390         int pkg_num, int cpu_id)
1391 {
1392         int ret;
1393         struct thread_data *t;
1394         struct core_data *c;
1395         struct pkg_data *p;
1396
1397         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1398         c = GET_CORE(core_base, core_num, pkg_num);
1399         p = GET_PKG(pkg_base, pkg_num);
1400
1401         t->cpu_id = cpu_id;
1402         if (thread_num == 0) {
1403                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1404                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1405                         return ret;
1406                 } else if (ret != 0) {
1407                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1408                 }
1409         }
1410
1411         c->core_id = core_num;
1412         p->package_id = pkg_num;
1413
1414         return 0;
1415 }
1416
1417
1418 static int
1419 initialize_counters(int cpu_id)
1420 {
1421         int my_thread_id, my_core_id, my_package_id;
1422         int ret;
1423
1424         my_package_id = get_physical_package_id(cpu_id);
1425         if (my_package_id < 0)
1426                 return my_package_id;
1427         my_core_id = get_core_id(cpu_id);
1428         if (my_core_id < 0)
1429                 return my_core_id;
1430
1431         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1432                 return ret;
1433         } else if (ret != 0) {
1434                 my_thread_id = 0;
1435                 topo.num_cores++;
1436         } else {
1437                 my_thread_id = 1;
1438         }
1439
1440         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1441         if (ret < 0)
1442                 return ret;
1443         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1444         if (ret < 0)
1445                 return ret;
1446         return 0;
1447 }
1448
1449 #define DO_OR_GOTO_ERR(something) \
1450 do {                         \
1451         ret = (something);     \
1452         if (ret < 0)         \
1453                 goto err;    \
1454 } while (0)
1455
1456 static int setup_all_buffers(void)
1457 {
1458         int ret;
1459
1460         DO_OR_GOTO_ERR(topology_probe());
1461         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1462         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1463         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1464
1465         allocated = 1;
1466         return 0;
1467 err:
1468         free_all_buffers();
1469         return ret;
1470 }
1471
1472 static int
1473 turbostat_init(void)
1474 {
1475         int ret;
1476
1477         DO_OR_GOTO_ERR(check_super_user());
1478         DO_OR_GOTO_ERR(probe_cpu());
1479         DO_OR_GOTO_ERR(check_dev_msr());
1480         DO_OR_GOTO_ERR(setup_all_buffers());
1481         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1482         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1483
1484         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1485
1486         return 0;
1487 err:
1488         free_all_buffers();
1489         return ret;
1490 }
1491
1492 static const char *config_keys[] =
1493 {
1494         "TCCActivationTemp",
1495 };
1496 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1497
1498 static int
1499 turbostat_config(const char *key, const char *value)
1500 {
1501         long unsigned int tmp_val;
1502         char *end;
1503
1504         if (strcasecmp("TCCActivationTemp", key) == 0) {
1505                 tmp_val = strtoul(value, &end, 0);
1506                 if (*end != '\0' || tmp_val > UINT_MAX)
1507                         return -1;
1508                 tcc_activation_temp = (unsigned int) tmp_val;
1509         } else {
1510                 return -1;
1511         }
1512         return 0;
1513 }
1514
1515 void module_register(void);
1516 void module_register(void)
1517 {
1518         plugin_register_init(PLUGIN_NAME, turbostat_init);
1519         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1520 }