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 _Bool do_dts;
  76 static _Bool do_ptm;
  77 static unsigned int tcc_activation_temp;
  78
  79 static unsigned int do_rapl;
  80 static double rapl_energy_units;
  81
  82 #define RAPL_PKG                (1 << 0)
  83                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  84                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  85 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  86                                         /* 0x613 MSR_PKG_PERF_STATUS */
  87 #define RAPL_PKG_POWER_INFO     (1 << 2)
  88                                         /* 0x614 MSR_PKG_POWER_INFO */
  89
  90 #define RAPL_DRAM               (1 << 3)
  91                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
  92                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
  93                                         /* 0x61c MSR_DRAM_POWER_INFO */
  94 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
  95                                         /* 0x61b MSR_DRAM_PERF_STATUS */
  96
  97 #define RAPL_CORES              (1 << 5)
  98                                         /* 0x638 MSR_PP0_POWER_LIMIT */
  99                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 100 #define RAPL_CORE_POLICY        (1 << 6)
 101                                         /* 0x63a MSR_PP0_POLICY */
 102
 103
 104 #define RAPL_GFX                (1 << 7)
 105                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 106                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 107                                         /* 0x642 MSR_PP1_POLICY */
 108 #define TJMAX_DEFAULT   100
 109
 110 int backwards_count;
 111
 112 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
 113 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
 114
 115 struct thread_data {
 116         unsigned long long tsc;
 117         unsigned long long aperf;
 118         unsigned long long mperf;
 119         unsigned long long c1;
 120         unsigned int smi_count;
 121         unsigned int cpu_id;
 122         unsigned int flags;
 123 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 124 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 125 } *thread_even, *thread_odd;
 126
 127 struct core_data {
 128         unsigned long long c3;
 129         unsigned long long c6;
 130         unsigned long long c7;
 131         unsigned int core_temp_c;
 132         unsigned int core_id;
 133 } *core_even, *core_odd;
 134
 135 struct pkg_data {
 136         unsigned long long pc2;
 137         unsigned long long pc3;
 138         unsigned long long pc6;
 139         unsigned long long pc7;
 140         unsigned long long pc8;
 141         unsigned long long pc9;
 142         unsigned long long pc10;
 143         unsigned int package_id;
 144         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 145         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 146         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 147         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 148         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 149         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 150         unsigned int tcc_activation_temp;
 151         unsigned int pkg_temp_c;
 152 } *package_even, *package_odd;
 153
 154 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 155 #define EVEN_COUNTERS thread_even, core_even, package_even
 156 static _Bool is_even = 1;
 157
 158 static _Bool allocated = 0;
 159 static _Bool initialized = 0;
 160
 161 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 162         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 163                 topo.num_threads_per_core + \
 164                 (core_no) * topo.num_threads_per_core + (thread_no))
 165 #define GET_CORE(core_base, core_no, pkg_no) \
 166         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 167 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 168
 169 struct topo_params {
 170         int num_packages;
 171         int num_cpus;
 172         int num_cores;
 173         int max_cpu_num;
 174         int num_cores_per_pkg;
 175         int num_threads_per_core;
 176 } topo;
 177
 178 struct timeval tv_even, tv_odd, tv_delta;
 179
 180 enum return_values {
 181         OK = 0,
 182         ERR_CPU_MIGRATE,
 183         ERR_CPU_SAVE_SCHED_AFFINITY,
 184         ERR_MSR_IA32_APERF,
 185         ERR_MSR_IA32_MPERF,
 186         ERR_MSR_SMI_COUNT,
 187         ERR_MSR_CORE_C3_RESIDENCY,
 188         ERR_MSR_CORE_C6_RESIDENCY,
 189         ERR_MSR_CORE_C7_RESIDENCY,
 190         ERR_MSR_IA32_THERM_STATUS,
 191         ERR_MSR_PKG_C3_RESIDENCY,
 192         ERR_MSR_PKG_C6_RESIDENCY,
 193         ERR_MSR_PKG_C2_RESIDENCY,
 194         ERR_MSR_PKG_C7_RESIDENCY,
 195         ERR_MSR_PKG_C8_RESIDENCY,
 196         ERR_MSR_PKG_C9_RESIDENCY,
 197         ERR_MSR_PKG_C10_RESIDENCY,
 198         ERR_MSR_PKG_ENERGY_STATUS,
 199         ERR_MSR_PKG_POWER_INFO,
 200         ERR_MSR_PP0_ENERGY_STATUS,
 201         ERR_MSR_DRAM_ENERGY_STATUS,
 202         ERR_MSR_PP1_ENERGY_STATUS,
 203         ERR_MSR_PKG_PERF_STATUS,
 204         ERR_MSR_DRAM_PERF_STATUS,
 205         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 206         ERR_MSR_IA32_TSC,
 207         ERR_CPU_NOT_PRESENT,
 208         ERR_NO_MSR,
 209         ERR_CANT_OPEN_MSR,
 210         ERR_CANT_OPEN_FILE,
 211         ERR_CANT_READ_NUMBER,
 212         ERR_CANT_READ_PROC_STAT,
 213         ERR_NO_INVARIANT_TSC,
 214         ERR_NO_APERF,
 215         ERR_CALLOC,
 216         ERR_CPU_ALLOC,
 217         ERR_NOT_ROOT,
 218         UNSUPPORTED_CPU,
 219 };
 220
 221 static int setup_all_buffers(void);
 222
 223 static int
 224 cpu_is_not_present(int cpu)
 225 {
 226         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 227 }
 228 /*
 229  * run func(thread, core, package) in topology order
 230  * skip non-present cpus
 231  */
 232
 233 static int __attribute__((warn_unused_result))
 234 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 235         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 236 {
 237         int retval, pkg_no, core_no, thread_no;
 238
 239         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 240                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 241                         for (thread_no = 0; thread_no <
 242                                 topo.num_threads_per_core; ++thread_no) {
 243                                 struct thread_data *t;
 244                                 struct core_data *c;
 245                                 struct pkg_data *p;
 246
 247                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 248
 249                                 if (cpu_is_not_present(t->cpu_id))
 250                                         continue;
 251
 252                                 c = GET_CORE(core_base, core_no, pkg_no);
 253                                 p = GET_PKG(pkg_base, pkg_no);
 254
 255                                 retval = func(t, c, p);
 256                                 if (retval)
 257                                         return retval;
 258                         }
 259                 }
 260         }
 261         return 0;
 262 }
 263
 264 static int __attribute__((warn_unused_result))
 265 open_msr(int cpu, _Bool multiple_read)
 266 {
 267         char pathname[32];
 268         int fd;
 269
 270         /*
 271          * If we need to do multiple read, let's migrate to the CPU
 272          * Otherwise, we would lose time calling functions on another CPU
 273          *
 274          * If we are not yet initialized (cpu_affinity_setsize = 0),
 275          * we need to skip this optimisation.
 276          */
 277         if (multiple_read && cpu_affinity_setsize) {
 278                 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 279                 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 280                 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
 281                         ERROR("Could not migrate to CPU %d", cpu);
 282                         return -ERR_CPU_MIGRATE;
 283                 }
 284         }
 285
 286         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 287         fd = open(pathname, O_RDONLY);
 288         if (fd < 0)
 289                 return -ERR_CANT_OPEN_MSR;
 290         return fd;
 291 }
 292
 293 static int __attribute__((warn_unused_result))
 294 read_msr(int fd, off_t offset, unsigned long long *msr)
 295 {
 296         ssize_t retval;
 297
 298         retval = pread(fd, msr, sizeof *msr, offset);
 299
 300         if (retval != sizeof *msr) {
 301                 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
 302                 return -1;
 303         }
 304         return 0;
 305 }
 306
 307 static int __attribute__((warn_unused_result))
 308 get_msr(int cpu, off_t offset, unsigned long long *msr)
 309 {
 310         ssize_t retval;
 311         int fd;
 312
 313         fd = open_msr(cpu, 0);
 314         if (fd < 0)
 315                 return fd;
 316         retval = read_msr(fd, offset, msr);
 317         close(fd);
 318         return retval;
 319 }
 320
 321 #define DELTA_WRAP32(new, old)                  \
 322         if (new > old) {                        \
 323                 old = new - old;                \
 324         } else {                                \
 325                 old = 0x100000000 + new - old;  \
 326         }
 327
 328 static void
 329 delta_package(struct pkg_data *new, struct pkg_data *old)
 330 {
 331         old->pc2 = new->pc2 - old->pc2;
 332         old->pc3 = new->pc3 - old->pc3;
 333         old->pc6 = new->pc6 - old->pc6;
 334         old->pc7 = new->pc7 - old->pc7;
 335         old->pc8 = new->pc8 - old->pc8;
 336         old->pc9 = new->pc9 - old->pc9;
 337         old->pc10 = new->pc10 - old->pc10;
 338         old->pkg_temp_c = new->pkg_temp_c;
 339
 340         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 341         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 342         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 343         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 344         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 345         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 346 }
 347
 348 static void
 349 delta_core(struct core_data *new, struct core_data *old)
 350 {
 351         old->c3 = new->c3 - old->c3;
 352         old->c6 = new->c6 - old->c6;
 353         old->c7 = new->c7 - old->c7;
 354         old->core_temp_c = new->core_temp_c;
 355 }
 356
 357 /*
 358  * old = new - old
 359  */
 360 static int __attribute__((warn_unused_result))
 361 delta_thread(struct thread_data *new, struct thread_data *old,
 362         struct core_data *core_delta)
 363 {
 364         old->tsc = new->tsc - old->tsc;
 365
 366         /* check for TSC < 1 Mcycles over interval */
 367         if (old->tsc < (1000 * 1000)) {
 368                 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
 369                 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
 370                 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
 371                 return -1;
 372         }
 373
 374         old->c1 = new->c1 - old->c1;
 375
 376         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 377                 old->aperf = new->aperf - old->aperf;
 378                 old->mperf = new->mperf - old->mperf;
 379         } else {
 380                 if (!aperf_mperf_unstable) {
 381                         WARNING(" APERF or MPERF went backwards * ");
 382                         WARNING("* Frequency results do not cover entire interval *");
 383                         WARNING("* fix this by running Linux-2.6.30 or later *");
 384
 385                         aperf_mperf_unstable = 1;
 386                 }
 387         }
 388
 389
 390         /*
 391          * As counter collection is not atomic,
 392          * it is possible for mperf's non-halted cycles + idle states
 393          * to exceed TSC's all cycles: show c1 = 0% in that case.
 394          */
 395         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 396                 old->c1 = 0;
 397         else {
 398                 /* normal case, derive c1 */
 399                 old->c1 = old->tsc - old->mperf - core_delta->c3
 400                         - core_delta->c6 - core_delta->c7;
 401         }
 402
 403         if (old->mperf == 0) {
 404                 WARNING("cpu%d MPERF 0!", old->cpu_id);
 405                 old->mperf = 1; /* divide by 0 protection */
 406         }
 407
 408         old->smi_count = new->smi_count - old->smi_count;
 409
 410         return 0;
 411 }
 412
 413 static int __attribute__((warn_unused_result))
 414 delta_cpu(struct thread_data *t, struct core_data *c,
 415         struct pkg_data *p, struct thread_data *t2,
 416         struct core_data *c2, struct pkg_data *p2)
 417 {
 418         int ret;
 419
 420         /* calculate core delta only for 1st thread in core */
 421         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 422                 delta_core(c, c2);
 423
 424         /* always calculate thread delta */
 425         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 426         if (ret != 0)
 427                 return ret;
 428
 429         /* calculate package delta only for 1st core in package */
 430         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 431                 delta_package(p, p2);
 432
 433         return 0;
 434 }
 435
 436 /*
 437  * get_counters(...)
 438  * migrate to cpu
 439  * acquire and record local counters for that cpu
 440  */
 441 static int __attribute__((warn_unused_result))
 442 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 443 {
 444         int cpu = t->cpu_id;
 445         unsigned long long msr;
 446         int msr_fd;
 447         int retval = 0;
 448
 449         msr_fd = open_msr(cpu, 1);
 450         if (msr_fd < 0)
 451                 return msr_fd;
 452
 453 #define READ_MSR(msr, dst)                      \
 454 do {                                            \
 455         if (read_msr(msr_fd, msr, dst)) {       \
 456                 retval = -ERR_##msr;            \
 457                 goto out;                       \
 458         }                                       \
 459 } while (0)
 460
 461         READ_MSR(MSR_IA32_TSC, &t->tsc);
 462
 463         READ_MSR(MSR_IA32_APERF, &t->aperf);
 464         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 465
 466         READ_MSR(MSR_SMI_COUNT, &msr);
 467         t->smi_count = msr & 0xFFFFFFFF;
 468
 469         /* collect core counters only for 1st thread in core */
 470         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 471                 retval = 0;
 472                 goto out;
 473         }
 474
 475         if (do_core_cstate & (1 << 3))
 476                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 477         if (do_core_cstate & (1 << 6))
 478                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 479         if (do_core_cstate & (1 << 7))
 480                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 481
 482         if (do_dts) {
 483                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 484                 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 485         }
 486
 487         /* collect package counters only for 1st core in package */
 488         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 489                 retval = 0;
 490                 goto out;
 491         }
 492
 493         if (do_pkg_cstate & (1 << 2))
 494                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 495         if (do_pkg_cstate & (1 << 3))
 496                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 497         if (do_pkg_cstate & (1 << 6))
 498                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 499         if (do_pkg_cstate & (1 << 7))
 500                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 501         if (do_pkg_cstate & (1 << 8))
 502                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 503         if (do_pkg_cstate & (1 << 9))
 504                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 505         if (do_pkg_cstate & (1 << 10))
 506                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 507
 508         if (do_rapl & RAPL_PKG) {
 509                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 510                 p->energy_pkg = msr & 0xFFFFFFFF;
 511         }
 512         if (do_rapl & RAPL_CORES) {
 513                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 514                 p->energy_cores = msr & 0xFFFFFFFF;
 515         }
 516         if (do_rapl & RAPL_DRAM) {
 517                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 518                 p->energy_dram = msr & 0xFFFFFFFF;
 519         }
 520         if (do_rapl & RAPL_GFX) {
 521                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 522                 p->energy_gfx = msr & 0xFFFFFFFF;
 523         }
 524         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 525                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 526                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 527         }
 528         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 529                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 530                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 531         }
 532         if (do_ptm) {
 533                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 534                 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 535         }
 536
 537 out:
 538         close(msr_fd);
 539         return retval;
 540 }
 541
 542 static void
 543 free_all_buffers(void)
 544 {
 545         allocated = 0;
 546         initialized = 0;
 547
 548         CPU_FREE(cpu_present_set);
 549         cpu_present_set = NULL;
 550         cpu_present_set = 0;
 551
 552         CPU_FREE(cpu_affinity_set);
 553         cpu_affinity_set = NULL;
 554         cpu_affinity_setsize = 0;
 555
 556         CPU_FREE(cpu_saved_affinity_set);
 557         cpu_saved_affinity_set = NULL;
 558         cpu_saved_affinity_setsize = 0;
 559
 560         free(thread_even);
 561         free(core_even);
 562         free(package_even);
 563
 564         thread_even = NULL;
 565         core_even = NULL;
 566         package_even = NULL;
 567
 568         free(thread_odd);
 569         free(core_odd);
 570         free(package_odd);
 571
 572         thread_odd = NULL;
 573         core_odd = NULL;
 574         package_odd = NULL;
 575 }
 576
 577 /*
 578  * Parse a file containing a single int.
 579  */
 580 static int __attribute__ ((format(printf,1,2)))
 581 parse_int_file(const char *fmt, ...)
 582 {
 583         va_list args;
 584         char path[PATH_MAX];
 585         FILE *filep;
 586         int value;
 587
 588         va_start(args, fmt);
 589         vsnprintf(path, sizeof(path), fmt, args);
 590         va_end(args);
 591         filep = fopen(path, "r");
 592         if (!filep) {
 593                 ERROR("%s: open failed", path);
 594                 return -ERR_CANT_OPEN_FILE;
 595         }
 596         if (fscanf(filep, "%d", &value) != 1) {
 597                 ERROR("%s: failed to parse number from file", path);
 598                 return -ERR_CANT_READ_NUMBER;
 599         }
 600         fclose(filep);
 601         return value;
 602 }
 603
 604 /*
 605  * cpu_is_first_sibling_in_core(cpu)
 606  * return 1 if given CPU is 1st HT sibling in the core
 607  */
 608 static int
 609 cpu_is_first_sibling_in_core(int cpu)
 610 {
 611         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 612 }
 613
 614 /*
 615  * cpu_is_first_core_in_package(cpu)
 616  * return 1 if given CPU is 1st core in package
 617  */
 618 static int
 619 cpu_is_first_core_in_package(int cpu)
 620 {
 621         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 622 }
 623
 624 static int
 625 get_physical_package_id(int cpu)
 626 {
 627         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 628 }
 629
 630 static int
 631 get_core_id(int cpu)
 632 {
 633         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 634 }
 635
 636 static int
 637 get_num_ht_siblings(int cpu)
 638 {
 639         char path[80];
 640         FILE *filep;
 641         int sib1, sib2;
 642         int matches;
 643         char character;
 644
 645         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 646         filep = fopen(path, "r");
 647         if (!filep) {
 648                 ERROR("%s: open failed", path);
 649                 return -ERR_CANT_OPEN_FILE;
 650         }
 651         /*
 652          * file format:
 653          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 654          * otherwinse 1 sibling (self).
 655          */
 656         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 657
 658         fclose(filep);
 659
 660         if (matches == 3)
 661                 return 2;
 662         else
 663                 return 1;
 664 }
 665
 666 /*
 667  * run func(thread, core, package) in topology order
 668  * skip non-present cpus
 669  */
 670
 671
 672 static int __attribute__((warn_unused_result))
 673 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 674         struct pkg_data *, struct thread_data *, struct core_data *,
 675         struct pkg_data *), struct thread_data *thread_base,
 676         struct core_data *core_base, struct pkg_data *pkg_base,
 677         struct thread_data *thread_base2, struct core_data *core_base2,
 678         struct pkg_data *pkg_base2)
 679 {
 680         int retval, pkg_no, core_no, thread_no;
 681
 682         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 683                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 684                         for (thread_no = 0; thread_no <
 685                                 topo.num_threads_per_core; ++thread_no) {
 686                                 struct thread_data *t, *t2;
 687                                 struct core_data *c, *c2;
 688                                 struct pkg_data *p, *p2;
 689
 690                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 691
 692                                 if (cpu_is_not_present(t->cpu_id))
 693                                         continue;
 694
 695                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 696
 697                                 c = GET_CORE(core_base, core_no, pkg_no);
 698                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 699
 700                                 p = GET_PKG(pkg_base, pkg_no);
 701                                 p2 = GET_PKG(pkg_base2, pkg_no);
 702
 703                                 retval = func(t, c, p, t2, c2, p2);
 704                                 if (retval)
 705                                         return retval;
 706                         }
 707                 }
 708         }
 709         return 0;
 710 }
 711
 712 /*
 713  * run func(cpu) on every cpu in /proc/stat
 714  * return max_cpu number
 715  */
 716 static int __attribute__((warn_unused_result))
 717 for_all_proc_cpus(int (func)(int))
 718 {
 719         FILE *fp;
 720         int cpu_num;
 721         int retval;
 722
 723         fp = fopen(proc_stat, "r");
 724         if (!fp) {
 725                 ERROR("%s: open failed", proc_stat);
 726                 return -ERR_CANT_OPEN_FILE;
 727         }
 728
 729         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 730         if (retval != 0) {
 731                 ERROR("%s: failed to parse format", proc_stat);
 732                 return -ERR_CANT_READ_PROC_STAT;
 733         }
 734
 735         while (1) {
 736                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 737                 if (retval != 1)
 738                         break;
 739
 740                 retval = func(cpu_num);
 741                 if (retval) {
 742                         fclose(fp);
 743                         return(retval);
 744                 }
 745         }
 746         fclose(fp);
 747         return 0;
 748 }
 749
 750 /*
 751  * count_cpus()
 752  * remember the last one seen, it will be the max
 753  */
 754 static int
 755 count_cpus(int cpu)
 756 {
 757         if (topo.max_cpu_num < cpu)
 758                 topo.max_cpu_num = cpu;
 759
 760         topo.num_cpus += 1;
 761         return 0;
 762 }
 763 static int
 764 mark_cpu_present(int cpu)
 765 {
 766         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 767         return 0;
 768 }
 769
 770
 771 static void
 772 turbostat_submit (const char *plugin_instance,
 773         const char *type, const char *type_instance,
 774         gauge_t value)
 775 {
 776         value_list_t vl = VALUE_LIST_INIT;
 777         value_t v;
 778
 779         v.gauge = value;
 780         vl.values = &v;
 781         vl.values_len = 1;
 782         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 783         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 784         if (plugin_instance != NULL)
 785                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 786         sstrncpy (vl.type, type, sizeof (vl.type));
 787         if (type_instance != NULL)
 788                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 789
 790         plugin_dispatch_values (&vl);
 791 }
 792
 793 /*
 794  * column formatting convention & formats
 795  * package: "pk" 2 columns %2d
 796  * core: "cor" 3 columns %3d
 797  * CPU: "CPU" 3 columns %3d
 798  * Pkg_W: %6.2
 799  * Cor_W: %6.2
 800  * GFX_W: %5.2
 801  * RAM_W: %5.2
 802  * GHz: "GHz" 3 columns %3.2
 803  * TSC: "TSC" 3 columns %3.2
 804  * SMI: "SMI" 4 columns %4d
 805  * percentage " %pc3" %6.2
 806  * Perf Status percentage: %5.2
 807  * "CTMP" 4 columns %4d
 808  */
 809 #define NAME_LEN 12
 810 static int
 811 submit_counters(struct thread_data *t, struct core_data *c,
 812         struct pkg_data *p)
 813 {
 814         char name[NAME_LEN];
 815         double interval_float;
 816
 817         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 818
 819         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 820
 821         if (!aperf_mperf_unstable)
 822                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 823         if (!aperf_mperf_unstable)
 824                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 825
 826         /* GHz */
 827         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 828                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 829
 830         /* SMI */
 831         turbostat_submit(NULL, "current", name, t->smi_count);
 832
 833         /* print per-core data only for 1st thread in core */
 834         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 835                 goto done;
 836
 837         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 838
 839         if (do_core_cstate & (1 << 3))
 840                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 841         if (do_core_cstate & (1 << 6))
 842                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 843         if (do_core_cstate & (1 << 7))
 844                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 845
 846         if (do_dts)
 847                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 848
 849         /* print per-package data only for 1st core in package */
 850         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 851                 goto done;
 852
 853         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 854
 855         if (do_ptm)
 856                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 857
 858         if (do_pkg_cstate & (1 << 2))
 859                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 860         if (do_pkg_cstate & (1 << 3))
 861                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 862         if (do_pkg_cstate & (1 << 6))
 863                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 864         if (do_pkg_cstate & (1 << 7))
 865                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 866         if (do_pkg_cstate & (1 << 8))
 867                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 868         if (do_pkg_cstate & (1 << 9))
 869                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 870         if (do_pkg_cstate & (1 << 10))
 871                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 872
 873         if (do_rapl) {
 874                 if (do_rapl & RAPL_PKG)
 875                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 876                 if (do_rapl & RAPL_CORES)
 877                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 878                 if (do_rapl & RAPL_GFX)
 879                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 880                 if (do_rapl & RAPL_DRAM)
 881                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 882         }
 883 done:
 884         return 0;
 885 }
 886
 887 static int
 888 turbostat_read(user_data_t * not_used)
 889 {
 890         int ret;
 891
 892         if (!allocated) {
 893                 if ((ret = setup_all_buffers()) < 0)
 894                         return ret;
 895         }
 896
 897         if (for_all_proc_cpus(cpu_is_not_present)) {
 898                 free_all_buffers();
 899                 if ((ret = setup_all_buffers()) < 0)
 900                         return ret;
 901                 if (for_all_proc_cpus(cpu_is_not_present))
 902                         return -ERR_CPU_NOT_PRESENT;
 903         }
 904
 905         /* Saving the scheduling affinity, as it will be modified by get_counters */
 906         if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0)
 907                 return -ERR_CPU_SAVE_SCHED_AFFINITY;
 908
 909         if (!initialized) {
 910                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 911                         goto out;
 912                 gettimeofday(&tv_even, (struct timezone *)NULL);
 913                 is_even = 1;
 914                 initialized = 1;
 915                 ret = 0;
 916                 goto out;
 917         }
 918
 919         if (is_even) {
 920                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 921                         goto out;
 922                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 923                 is_even = 0;
 924                 timersub(&tv_odd, &tv_even, &tv_delta);
 925                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 926                         goto out;
 927                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 928                         goto out;
 929         } else {
 930                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 931                         goto out;
 932                 gettimeofday(&tv_even, (struct timezone *)NULL);
 933                 is_even = 1;
 934                 timersub(&tv_even, &tv_odd, &tv_delta);
 935                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 936                         goto out;
 937                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 938                         goto out;
 939         }
 940         ret = 0;
 941 out:
 942         /*
 943          * Let's restore the affinity
 944          * This might fail if the number of CPU changed, but we can't do anything in that case..
 945          */
 946         (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
 947         return ret;
 948 }
 949
 950 static int __attribute__((warn_unused_result))
 951 check_dev_msr()
 952 {
 953         struct stat sb;
 954
 955         if (stat("/dev/cpu/0/msr", &sb)) {
 956                 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
 957                 return -ERR_NO_MSR;
 958         }
 959         return 0;
 960 }
 961
 962 static int __attribute__((warn_unused_result))
 963 check_super_user()
 964 {
 965         if (getuid() != 0) {
 966                 ERROR("must be root");
 967                 return -ERR_NOT_ROOT;
 968         }
 969         return 0;
 970 }
 971
 972 /*
 973  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
 974  * the Thermal Control Circuit (TCC) activates.
 975  * This is usually equal to tjMax.
 976  *
 977  * Older processors do not have this MSR, so there we guess,
 978  * but also allow cmdline over-ride with -T.
 979  *
 980  * Several MSR temperature values are in units of degrees-C
 981  * below this value, including the Digital Thermal Sensor (DTS),
 982  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
 983  */
 984 static int __attribute__((warn_unused_result))
 985 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 986 {
 987         unsigned long long msr;
 988         unsigned int target_c_local;
 989
 990         /* tcc_activation_temp is used only for dts or ptm */
 991         if (!(do_dts || do_ptm))
 992                 return 0;
 993
 994         /* this is a per-package concept */
 995         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 996                 return 0;
 997
 998         if (tcc_activation_temp != 0) {
 999                 p->tcc_activation_temp = tcc_activation_temp;
1000                 return 0;
1001         }
1002
1003         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
1004                 goto guess;
1005
1006         target_c_local = (msr >> 16) & 0x7F;
1007
1008         if (target_c_local < 85 || target_c_local > 127)
1009                 goto guess;
1010
1011         p->tcc_activation_temp = target_c_local;
1012
1013         return 0;
1014
1015 guess:
1016         p->tcc_activation_temp = TJMAX_DEFAULT;
1017         WARNING("cpu%d: Guessing tjMax %d C, Please use TCCActivationTemp to specify",
1018                 t->cpu_id, p->tcc_activation_temp);
1019
1020         return 0;
1021 }
1022
1023 /*
1024  * Identify the functionality of the CPU
1025  */
1026 static int __attribute__((warn_unused_result))
1027 probe_cpu()
1028 {
1029         unsigned int eax, ebx, ecx, edx, max_level;
1030         unsigned int fms, family, model;
1031
1032         /* CPUID(0):
1033          * - EAX: Maximum Input Value for Basic CPUID Information
1034          * - EBX: "Genu" (0x756e6547)
1035          * - EDX: "ineI" (0x49656e69)
1036          * - ECX: "ntel" (0x6c65746e)
1037          */
1038         max_level = ebx = ecx = edx = 0;
1039         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1040         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1041                 ERROR("Unsupported CPU");
1042                 return -UNSUPPORTED_CPU;
1043         }
1044
1045         /* CPUID(1):
1046          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1047          *  + 4-7:   Model ID
1048          *  + 8-11:  Family ID
1049          *  + 12-13: Processor type
1050          *  + 16-19: Extended Model ID
1051          *  + 20-27: Extended Family ID
1052          * - EDX: Feature Information:
1053          *  + 5: Support for MSR read/write operations
1054          */
1055         fms = ebx = ecx = edx = 0;
1056         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1057         family = (fms >> 8) & 0xf;
1058         model = (fms >> 4) & 0xf;
1059         if (family == 0xf)
1060                 family += (fms >> 20) & 0xf;
1061         if (family == 6 || family == 0xf)
1062                 model += ((fms >> 16) & 0xf) << 4;
1063         if (!(edx & (1 << 5))) {
1064                 ERROR("CPUID: no MSR");
1065                 return -ERR_NO_MSR;
1066         }
1067
1068         /*
1069          * CPUID(0x80000000):
1070          * - EAX: Maximum Input Value for Extended Function CPUID Information
1071          *
1072          * This allows us to verify if the CPUID(0x80000007) can be called
1073          *
1074          * This check is valid for both Intel and AMD.
1075          */
1076         max_level = ebx = ecx = edx = 0;
1077         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1078         if (max_level < 0x80000007) {
1079                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1080                 return -ERR_NO_INVARIANT_TSC;
1081         }
1082
1083         /*
1084          * CPUID(0x80000007):
1085          * - EDX:
1086          *  + 8: Invariant TSC available if set
1087          *
1088          * This check is valid for both Intel and AMD
1089          */
1090         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1091         if (!(edx & (1 << 8))) {
1092                 ERROR("No invariant TSC");
1093                 return -ERR_NO_INVARIANT_TSC;
1094         }
1095
1096         /*
1097          * CPUID(6):
1098          * - EAX:
1099          *  + 0: Digital temperature sensor is supported if set
1100          *  + 6: Package thermal management is supported if set
1101          * - ECX:
1102          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1103          *  + 3: The processor supports performance-energy bias preference if set.
1104          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1105          *
1106          * This check is valid for both Intel and AMD
1107          */
1108         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1109         do_dts = eax & (1 << 0);
1110         do_ptm = eax & (1 << 6);
1111         if (!(ecx & (1 << 0))) {
1112                 ERROR("No APERF");
1113                 return -ERR_NO_APERF;
1114         }
1115
1116         /*
1117          * Enable or disable C states depending on the model and family
1118          */
1119         if (family == 6) {
1120                 switch (model) {
1121                 /* Atom (partial) */
1122                 case 0x27:
1123                         do_core_cstate = 0;
1124                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1125                         break;
1126                 /* Silvermont */
1127                 case 0x37: /* BYT */
1128                 case 0x4A:
1129                 case 0x4D: /* AVN */
1130                 case 0x5A:
1131                 case 0x5D:
1132                         do_core_cstate = (1 << 1) | (1 << 6);
1133                         do_pkg_cstate = (1 << 6);
1134                         break;
1135                 /* Nehalem */
1136                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1137                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1138                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1139                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1140                         do_core_cstate = (1 << 3) | (1 << 6);
1141                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1142                         break;
1143                 /* Westmere */
1144                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1145                 case 0x2C: /* Westmere EP - Gulftown */
1146                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1147                         do_core_cstate = (1 << 3) | (1 << 6);
1148                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1149                         break;
1150                 /* Sandy Bridge */
1151                 case 0x2A: /* SNB */
1152                 case 0x2D: /* SNB Xeon */
1153                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1154                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1155                         break;
1156                 /* Ivy Bridge */
1157                 case 0x3A: /* IVB */
1158                 case 0x3E: /* IVB Xeon */
1159                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1160                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1161                         break;
1162                 /* Haswell Bridge */
1163                 case 0x3C: /* HSW */
1164                 case 0x3F: /* HSW */
1165                 case 0x46: /* HSW */
1166                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1167                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1168                         break;
1169                 case 0x45: /* HSW */
1170                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1171                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1172                         break;
1173                 /* Broadwel */
1174                 case 0x4F: /* BDW */
1175                 case 0x56: /* BDX-DE */
1176                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1177                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1178                         break;
1179                 case 0x3D: /* BDW */
1180                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1181                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1182                         break;
1183                 default:
1184                         ERROR("Unsupported CPU");
1185                 }
1186                 switch (model) {
1187                 case 0x2A:
1188                 case 0x3A:
1189                 case 0x3C:
1190                 case 0x45:
1191                 case 0x46:
1192                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1193                         break;
1194                 case 0x3F:
1195                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1196                         break;
1197                 case 0x2D:
1198                 case 0x3E:
1199                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1200                         break;
1201                 case 0x37:
1202                 case 0x4D:
1203                         do_rapl = RAPL_PKG | RAPL_CORES;
1204                         break;
1205                 default:
1206                         do_rapl = 0;
1207                 }
1208         } else {
1209                 ERROR("Unsupported CPU");
1210                 return -UNSUPPORTED_CPU;
1211         }
1212
1213         if (do_rapl) {
1214                 unsigned long msr;
1215                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1216                         return 0;
1217
1218                 if (model == 0x37)
1219                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1220                 else
1221                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1222         }
1223
1224         return 0;
1225 }
1226
1227
1228
1229 static int __attribute__((warn_unused_result))
1230 topology_probe()
1231 {
1232         int i;
1233         int ret;
1234         int max_core_id = 0;
1235         int max_package_id = 0;
1236         int max_siblings = 0;
1237         struct cpu_topology {
1238                 int core_id;
1239                 int physical_package_id;
1240         } *cpus;
1241
1242         /* Initialize num_cpus, max_cpu_num */
1243         topo.num_cpus = 0;
1244         topo.max_cpu_num = 0;
1245         ret = for_all_proc_cpus(count_cpus);
1246         if (ret < 0)
1247                 return ret;
1248
1249         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1250
1251         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1252         if (cpus == NULL) {
1253                 ERROR("calloc cpus");
1254                 return -ERR_CALLOC;
1255         }
1256
1257         /*
1258          * Allocate and initialize cpu_present_set
1259          */
1260         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1261         if (cpu_present_set == NULL) {
1262                 free(cpus);
1263                 ERROR("CPU_ALLOC");
1264                 return -ERR_CPU_ALLOC;
1265         }
1266         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1267         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1268         ret = for_all_proc_cpus(mark_cpu_present);
1269         if (ret < 0) {
1270                 free(cpus);
1271                 return ret;
1272         }
1273
1274         /*
1275          * Allocate and initialize cpu_affinity_set
1276          */
1277         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1278         if (cpu_affinity_set == NULL) {
1279                 free(cpus);
1280                 ERROR("CPU_ALLOC");
1281                 return -ERR_CPU_ALLOC;
1282         }
1283         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1284         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1285
1286
1287         /*
1288          * Allocate and initialize cpu_saved_affinity_set
1289          */
1290         cpu_saved_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1291         if (cpu_saved_affinity_set == NULL) {
1292                 free(cpus);
1293                 ERROR("CPU_ALLOC");
1294                 return -ERR_CPU_ALLOC;
1295         }
1296         cpu_saved_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1297         CPU_ZERO_S(cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1298
1299
1300         /*
1301          * For online cpus
1302          * find max_core_id, max_package_id
1303          */
1304         for (i = 0; i <= topo.max_cpu_num; ++i) {
1305                 int siblings;
1306
1307                 if (cpu_is_not_present(i)) {
1308                         WARNING("cpu%d NOT PRESENT", i);
1309                         continue;
1310                 }
1311                 cpus[i].core_id = get_core_id(i);
1312                 if (cpus[i].core_id < 0)
1313                         return cpus[i].core_id;
1314                 if (cpus[i].core_id > max_core_id)
1315                         max_core_id = cpus[i].core_id;
1316
1317                 cpus[i].physical_package_id = get_physical_package_id(i);
1318                 if (cpus[i].physical_package_id < 0)
1319                         return cpus[i].physical_package_id;
1320                 if (cpus[i].physical_package_id > max_package_id)
1321                         max_package_id = cpus[i].physical_package_id;
1322
1323                 siblings = get_num_ht_siblings(i);
1324                 if (siblings < 0)
1325                         return siblings;
1326                 if (siblings > max_siblings)
1327                         max_siblings = siblings;
1328                 DEBUG("cpu %d pkg %d core %d\n",
1329                         i, cpus[i].physical_package_id, cpus[i].core_id);
1330         }
1331         topo.num_cores_per_pkg = max_core_id + 1;
1332         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1333                 max_core_id, topo.num_cores_per_pkg);
1334
1335         topo.num_packages = max_package_id + 1;
1336         DEBUG("max_package_id %d, sizing for %d packages\n",
1337                 max_package_id, topo.num_packages);
1338
1339         topo.num_threads_per_core = max_siblings;
1340         DEBUG("max_siblings %d\n", max_siblings);
1341
1342         free(cpus);
1343         return 0;
1344 }
1345
1346 static int
1347 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1348 {
1349         int i;
1350
1351         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1352                 topo.num_packages, sizeof(struct thread_data));
1353         if (*t == NULL)
1354                 goto error;
1355
1356         for (i = 0; i < topo.num_threads_per_core *
1357                 topo.num_cores_per_pkg * topo.num_packages; i++)
1358                 (*t)[i].cpu_id = -1;
1359
1360         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1361                 sizeof(struct core_data));
1362         if (*c == NULL)
1363                 goto error;
1364
1365         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1366                 (*c)[i].core_id = -1;
1367
1368         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1369         if (*p == NULL)
1370                 goto error;
1371
1372         for (i = 0; i < topo.num_packages; i++)
1373                 (*p)[i].package_id = i;
1374
1375         return 0;
1376 error:
1377         ERROR("calloc counters");
1378         return -ERR_CALLOC;
1379 }
1380 /*
1381  * init_counter()
1382  *
1383  * set cpu_id, core_num, pkg_num
1384  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1385  *
1386  * increment topo.num_cores when 1st core in pkg seen
1387  */
1388 static int
1389 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1390         struct pkg_data *pkg_base, int thread_num, int core_num,
1391         int pkg_num, int cpu_id)
1392 {
1393         int ret;
1394         struct thread_data *t;
1395         struct core_data *c;
1396         struct pkg_data *p;
1397
1398         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1399         c = GET_CORE(core_base, core_num, pkg_num);
1400         p = GET_PKG(pkg_base, pkg_num);
1401
1402         t->cpu_id = cpu_id;
1403         if (thread_num == 0) {
1404                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1405                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1406                         return ret;
1407                 } else if (ret != 0) {
1408                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1409                 }
1410         }
1411
1412         c->core_id = core_num;
1413         p->package_id = pkg_num;
1414
1415         return 0;
1416 }
1417
1418
1419 static int
1420 initialize_counters(int cpu_id)
1421 {
1422         int my_thread_id, my_core_id, my_package_id;
1423         int ret;
1424
1425         my_package_id = get_physical_package_id(cpu_id);
1426         if (my_package_id < 0)
1427                 return my_package_id;
1428         my_core_id = get_core_id(cpu_id);
1429         if (my_core_id < 0)
1430                 return my_core_id;
1431
1432         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1433                 return ret;
1434         } else if (ret != 0) {
1435                 my_thread_id = 0;
1436                 topo.num_cores++;
1437         } else {
1438                 my_thread_id = 1;
1439         }
1440
1441         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1442         if (ret < 0)
1443                 return ret;
1444         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1445         if (ret < 0)
1446                 return ret;
1447         return 0;
1448 }
1449
1450 #define DO_OR_GOTO_ERR(something) \
1451 do {                         \
1452         ret = (something);     \
1453         if (ret < 0)         \
1454                 goto err;    \
1455 } while (0)
1456
1457 static int setup_all_buffers(void)
1458 {
1459         int ret;
1460
1461         DO_OR_GOTO_ERR(topology_probe());
1462         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1463         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1464         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1465
1466         allocated = 1;
1467         return 0;
1468 err:
1469         free_all_buffers();
1470         return ret;
1471 }
1472
1473 static int
1474 turbostat_init(void)
1475 {
1476         int ret;
1477
1478         DO_OR_GOTO_ERR(check_super_user());
1479         DO_OR_GOTO_ERR(probe_cpu());
1480         DO_OR_GOTO_ERR(check_dev_msr());
1481         DO_OR_GOTO_ERR(setup_all_buffers());
1482         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1483         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1484
1485         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1486
1487         return 0;
1488 err:
1489         free_all_buffers();
1490         return ret;
1491 }
1492
1493 static const char *config_keys[] =
1494 {
1495         "TCCActivationTemp",
1496 };
1497 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1498
1499 static int
1500 turbostat_config(const char *key, const char *value)
1501 {
1502         long unsigned int tmp_val;
1503         char *end;
1504
1505         if (strcasecmp("TCCActivationTemp", key) == 0) {
1506                 tmp_val = strtoul(value, &end, 0);
1507                 if (*end != '\0' || tmp_val > UINT_MAX)
1508                         return -1;
1509                 tcc_activation_temp = (unsigned int) tmp_val;
1510         } else {
1511                 return -1;
1512         }
1513         return 0;
1514 }
1515
1516 void module_register(void);
1517 void module_register(void)
1518 {
1519         plugin_register_init(PLUGIN_NAME, turbostat_init);
1520         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1521 }