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 /*
  64  * This tool uses the Model-Specific Registers (MSRs) present on Intel processors.
  65  * The general description each of these registers, depending on the architecture,
  66  * can be found in the Intel® 64 and IA-32 Architectures Software Developer Manual,
  67  * Volume 3 Chapter 35.
  68  */
  69
  70 /*
  71  * If set, aperf_mperf_unstable disables a/mperf based stats.
  72  * This includes: C0 & C1 states, frequency
  73  *
  74  * This value is automatically set if mperf or aperf go backward
  75  */
  76 static _Bool aperf_mperf_unstable;
  77
  78 /*
  79  * Bitmask of the list of core C states supported by the processor.
  80  * Currently supported C-states (by this plugin): 3, 6, 7
  81  */
  82 static unsigned int do_core_cstate;
  83
  84 /*
  85  * Bitmask of the list of pacages C states supported by the processor.
  86  * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
  87  */
  88 static unsigned int do_pkg_cstate;
  89
  90 /*
  91  * Boolean indicating if the processor supports 'Digital temperature sensor'
  92  * This feature enables the monitoring of the temperature of each core
  93  *
  94  * This feature has two limitations:
  95  *  - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
  96  *  - Temperatures above the tcc_activation_temp are not recorded
  97  */
  98 static _Bool do_dts;
  99
 100 /*
 101  * Boolean indicating if the processor supports 'Package thermal management'
 102  * This feature allows the monitoring of the temperature of each package
 103  *
 104  * This feature has two limitations:
 105  *  - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
 106  *  - Temperatures above the tcc_activation_temp are not recorded
 107  */
 108 static _Bool do_ptm;
 109
 110 /*
 111  * Thermal Control Circuit Activation Temperature as configured by the user.
 112  * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
 113  * and should only be used if the automated detection fails.
 114  */
 115 static unsigned int tcc_activation_temp;
 116
 117 static unsigned int do_rapl;
 118 static double rapl_energy_units;
 119
 120 #define RAPL_PKG                (1 << 0)
 121                                         /* 0x610 MSR_PKG_POWER_LIMIT */
 122                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
 123 #define RAPL_PKG_PERF_STATUS    (1 << 1)
 124                                         /* 0x613 MSR_PKG_PERF_STATUS */
 125 #define RAPL_PKG_POWER_INFO     (1 << 2)
 126                                         /* 0x614 MSR_PKG_POWER_INFO */
 127
 128 #define RAPL_DRAM               (1 << 3)
 129                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
 130                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
 131                                         /* 0x61c MSR_DRAM_POWER_INFO */
 132 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
 133                                         /* 0x61b MSR_DRAM_PERF_STATUS */
 134
 135 #define RAPL_CORES              (1 << 5)
 136                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 137                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 138 #define RAPL_CORE_POLICY        (1 << 6)
 139                                         /* 0x63a MSR_PP0_POLICY */
 140
 141
 142 #define RAPL_GFX                (1 << 7)
 143                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 144                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 145                                         /* 0x642 MSR_PP1_POLICY */
 146 #define TJMAX_DEFAULT   100
 147
 148 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
 149 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
 150
 151 struct thread_data {
 152         unsigned long long tsc;
 153         unsigned long long aperf;
 154         unsigned long long mperf;
 155         unsigned long long c1;
 156         unsigned int smi_count;
 157         unsigned int cpu_id;
 158         unsigned int flags;
 159 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 160 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 161 } *thread_delta, *thread_even, *thread_odd;
 162
 163 struct core_data {
 164         unsigned long long c3;
 165         unsigned long long c6;
 166         unsigned long long c7;
 167         unsigned int core_temp_c;
 168         unsigned int core_id;
 169 } *core_delta, *core_even, *core_odd;
 170
 171 struct pkg_data {
 172         unsigned long long pc2;
 173         unsigned long long pc3;
 174         unsigned long long pc6;
 175         unsigned long long pc7;
 176         unsigned long long pc8;
 177         unsigned long long pc9;
 178         unsigned long long pc10;
 179         unsigned int package_id;
 180         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 181         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 182         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 183         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 184         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 185         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 186         unsigned int tcc_activation_temp;
 187         unsigned int pkg_temp_c;
 188 } *package_delta, *package_even, *package_odd;
 189
 190 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
 191 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 192 #define EVEN_COUNTERS thread_even, core_even, package_even
 193 static _Bool is_even = 1;
 194
 195 static _Bool allocated = 0;
 196 static _Bool initialized = 0;
 197
 198 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 199         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 200                 topo.num_threads_per_core + \
 201                 (core_no) * topo.num_threads_per_core + (thread_no))
 202 #define GET_CORE(core_base, core_no, pkg_no) \
 203         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 204 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 205
 206 struct topo_params {
 207         int num_packages;
 208         int num_cpus;
 209         int num_cores;
 210         int max_cpu_num;
 211         int num_cores_per_pkg;
 212         int num_threads_per_core;
 213 } topo;
 214
 215 struct timeval tv_even, tv_odd, tv_delta;
 216
 217 enum return_values {
 218         OK = 0,
 219         ERR_CPU_MIGRATE,
 220         ERR_CPU_SAVE_SCHED_AFFINITY,
 221         ERR_MSR_IA32_APERF,
 222         ERR_MSR_IA32_MPERF,
 223         ERR_MSR_SMI_COUNT,
 224         ERR_MSR_CORE_C3_RESIDENCY,
 225         ERR_MSR_CORE_C6_RESIDENCY,
 226         ERR_MSR_CORE_C7_RESIDENCY,
 227         ERR_MSR_IA32_THERM_STATUS,
 228         ERR_MSR_PKG_C3_RESIDENCY,
 229         ERR_MSR_PKG_C6_RESIDENCY,
 230         ERR_MSR_PKG_C2_RESIDENCY,
 231         ERR_MSR_PKG_C7_RESIDENCY,
 232         ERR_MSR_PKG_C8_RESIDENCY,
 233         ERR_MSR_PKG_C9_RESIDENCY,
 234         ERR_MSR_PKG_C10_RESIDENCY,
 235         ERR_MSR_PKG_ENERGY_STATUS,
 236         ERR_MSR_PKG_POWER_INFO,
 237         ERR_MSR_PP0_ENERGY_STATUS,
 238         ERR_MSR_DRAM_ENERGY_STATUS,
 239         ERR_MSR_PP1_ENERGY_STATUS,
 240         ERR_MSR_PKG_PERF_STATUS,
 241         ERR_MSR_DRAM_PERF_STATUS,
 242         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 243         ERR_MSR_IA32_TSC,
 244         ERR_CPU_NOT_PRESENT,
 245         ERR_NO_MSR,
 246         ERR_CANT_OPEN_MSR,
 247         ERR_CANT_OPEN_FILE,
 248         ERR_CANT_READ_NUMBER,
 249         ERR_CANT_READ_PROC_STAT,
 250         ERR_NO_INVARIANT_TSC,
 251         ERR_NO_APERF,
 252         ERR_CALLOC,
 253         ERR_CPU_ALLOC,
 254         ERR_NOT_ROOT,
 255         UNSUPPORTED_CPU,
 256 };
 257
 258 static int setup_all_buffers(void);
 259
 260 static int
 261 cpu_is_not_present(int cpu)
 262 {
 263         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 264 }
 265 /*
 266  * run func(thread, core, package) in topology order
 267  * skip non-present cpus
 268  */
 269
 270 static int __attribute__((warn_unused_result))
 271 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 272         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 273 {
 274         int retval, pkg_no, core_no, thread_no;
 275
 276         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 277                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 278                         for (thread_no = 0; thread_no <
 279                                 topo.num_threads_per_core; ++thread_no) {
 280                                 struct thread_data *t;
 281                                 struct core_data *c;
 282                                 struct pkg_data *p;
 283
 284                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 285
 286                                 if (cpu_is_not_present(t->cpu_id))
 287                                         continue;
 288
 289                                 c = GET_CORE(core_base, core_no, pkg_no);
 290                                 p = GET_PKG(pkg_base, pkg_no);
 291
 292                                 retval = func(t, c, p);
 293                                 if (retval)
 294                                         return retval;
 295                         }
 296                 }
 297         }
 298         return 0;
 299 }
 300
 301 /*
 302  * Open a MSR device for reading
 303  * Can change the scheduling affinity of the current process if multiple_read is 1
 304  */
 305 static int __attribute__((warn_unused_result))
 306 open_msr(int cpu, _Bool multiple_read)
 307 {
 308         char pathname[32];
 309         int fd;
 310
 311         /*
 312          * If we need to do multiple read, let's migrate to the CPU
 313          * Otherwise, we would lose time calling functions on another CPU
 314          *
 315          * If we are not yet initialized (cpu_affinity_setsize = 0),
 316          * we need to skip this optimisation.
 317          */
 318         if (multiple_read && cpu_affinity_setsize) {
 319                 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 320                 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 321                 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
 322                         ERROR("Could not migrate to CPU %d", cpu);
 323                         return -ERR_CPU_MIGRATE;
 324                 }
 325         }
 326
 327         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 328         fd = open(pathname, O_RDONLY);
 329         if (fd < 0)
 330                 return -ERR_CANT_OPEN_MSR;
 331         return fd;
 332 }
 333
 334 /*
 335  * Read a single MSR from an open file descriptor
 336  */
 337 static int __attribute__((warn_unused_result))
 338 read_msr(int fd, off_t offset, unsigned long long *msr)
 339 {
 340         ssize_t retval;
 341
 342         retval = pread(fd, msr, sizeof *msr, offset);
 343
 344         if (retval != sizeof *msr) {
 345                 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
 346                 return -1;
 347         }
 348         return 0;
 349 }
 350
 351 /*
 352  * Open a MSR device for reading, read the value asked for and close it.
 353  * This call will not affect the scheduling affinity of this thread.
 354  */
 355 static int __attribute__((warn_unused_result))
 356 get_msr(int cpu, off_t offset, unsigned long long *msr)
 357 {
 358         ssize_t retval;
 359         int fd;
 360
 361         fd = open_msr(cpu, 0);
 362         if (fd < 0)
 363                 return fd;
 364         retval = read_msr(fd, offset, msr);
 365         close(fd);
 366         return retval;
 367 }
 368
 369 #define DELTA_WRAP32(delta, new, old)                   \
 370         if (new > old) {                                \
 371                 delta = new - old;                      \
 372         } else {                                        \
 373                 delta = 0x100000000 + new - old;        \
 374         }
 375
 376 /*
 377  * Extract the evolution old->new in delta at a package level
 378  * (some are not new-delta, e.g. temperature)
 379  */
 380 static inline void
 381 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
 382 {
 383         delta->pc2 = new->pc2 - old->pc2;
 384         delta->pc3 = new->pc3 - old->pc3;
 385         delta->pc6 = new->pc6 - old->pc6;
 386         delta->pc7 = new->pc7 - old->pc7;
 387         delta->pc8 = new->pc8 - old->pc8;
 388         delta->pc9 = new->pc9 - old->pc9;
 389         delta->pc10 = new->pc10 - old->pc10;
 390         delta->pkg_temp_c = new->pkg_temp_c;
 391
 392         DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
 393         DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
 394         DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
 395         DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
 396         DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 397         DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 398 }
 399
 400 /*
 401  * Extract the evolution old->new in delta at a core level
 402  * (some are not new-delta, e.g. temperature)
 403  */
 404 static inline void
 405 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
 406 {
 407         delta->c3 = new->c3 - old->c3;
 408         delta->c6 = new->c6 - old->c6;
 409         delta->c7 = new->c7 - old->c7;
 410         delta->core_temp_c = new->core_temp_c;
 411 }
 412
 413 /*
 414  * Extract the evolution old->new in delta at a package level
 415  * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
 416  */
 417 static inline int __attribute__((warn_unused_result))
 418 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
 419         const struct core_data *core_delta)
 420 {
 421         delta->tsc = new->tsc - old->tsc;
 422
 423         /* check for TSC < 1 Mcycles over interval */
 424         if (delta->tsc < (1000 * 1000)) {
 425                 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
 426                 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
 427                 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
 428                 return -1;
 429         }
 430
 431         delta->c1 = new->c1 - old->c1;
 432
 433         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 434                 delta->aperf = new->aperf - old->aperf;
 435                 delta->mperf = new->mperf - old->mperf;
 436         } else {
 437                 if (!aperf_mperf_unstable) {
 438                         WARNING(" APERF or MPERF went backwards * ");
 439                         WARNING("* Frequency results do not cover entire interval *");
 440                         WARNING("* fix this by running Linux-2.6.30 or later *");
 441
 442                         aperf_mperf_unstable = 1;
 443                 }
 444         }
 445
 446         /*
 447          * As counter collection is not atomic,
 448          * it is possible for mperf's non-halted cycles + idle states
 449          * to exceed TSC's all cycles: show c1 = 0% in that case.
 450          */
 451         if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
 452                 delta->c1 = 0;
 453         else {
 454                 /* normal case, derive c1 */
 455                 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
 456                         - core_delta->c6 - core_delta->c7;
 457         }
 458
 459         if (delta->mperf == 0) {
 460                 WARNING("cpu%d MPERF 0!", old->cpu_id);
 461                 delta->mperf = 1;       /* divide by 0 protection */
 462         }
 463
 464         delta->smi_count = new->smi_count - old->smi_count;
 465
 466         return 0;
 467 }
 468
 469 /*
 470  * get_counters(...)
 471  * migrate to cpu
 472  * acquire and record local counters for that cpu
 473  */
 474 static int __attribute__((warn_unused_result))
 475 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 476 {
 477         int cpu = t->cpu_id;
 478         unsigned long long msr;
 479         int msr_fd;
 480         int retval = 0;
 481
 482         msr_fd = open_msr(cpu, 1);
 483         if (msr_fd < 0)
 484                 return msr_fd;
 485
 486 #define READ_MSR(msr, dst)                      \
 487 do {                                            \
 488         if (read_msr(msr_fd, msr, dst)) {       \
 489                 retval = -ERR_##msr;            \
 490                 goto out;                       \
 491         }                                       \
 492 } while (0)
 493
 494         READ_MSR(MSR_IA32_TSC, &t->tsc);
 495
 496         READ_MSR(MSR_IA32_APERF, &t->aperf);
 497         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 498
 499         READ_MSR(MSR_SMI_COUNT, &msr);
 500         t->smi_count = msr & 0xFFFFFFFF;
 501
 502         /* collect core counters only for 1st thread in core */
 503         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 504                 retval = 0;
 505                 goto out;
 506         }
 507
 508         if (do_core_cstate & (1 << 3))
 509                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 510         if (do_core_cstate & (1 << 6))
 511                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 512         if (do_core_cstate & (1 << 7))
 513                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 514
 515         if (do_dts) {
 516                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 517                 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 518         }
 519
 520         /* collect package counters only for 1st core in package */
 521         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 522                 retval = 0;
 523                 goto out;
 524         }
 525
 526         if (do_pkg_cstate & (1 << 2))
 527                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 528         if (do_pkg_cstate & (1 << 3))
 529                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 530         if (do_pkg_cstate & (1 << 6))
 531                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 532         if (do_pkg_cstate & (1 << 7))
 533                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 534         if (do_pkg_cstate & (1 << 8))
 535                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 536         if (do_pkg_cstate & (1 << 9))
 537                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 538         if (do_pkg_cstate & (1 << 10))
 539                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 540
 541         if (do_rapl & RAPL_PKG) {
 542                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 543                 p->energy_pkg = msr & 0xFFFFFFFF;
 544         }
 545         if (do_rapl & RAPL_CORES) {
 546                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 547                 p->energy_cores = msr & 0xFFFFFFFF;
 548         }
 549         if (do_rapl & RAPL_DRAM) {
 550                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 551                 p->energy_dram = msr & 0xFFFFFFFF;
 552         }
 553         if (do_rapl & RAPL_GFX) {
 554                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 555                 p->energy_gfx = msr & 0xFFFFFFFF;
 556         }
 557         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 558                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 559                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 560         }
 561         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 562                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 563                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 564         }
 565         if (do_ptm) {
 566                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 567                 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 568         }
 569
 570 out:
 571         close(msr_fd);
 572         return retval;
 573 }
 574
 575 static void
 576 free_all_buffers(void)
 577 {
 578         allocated = 0;
 579         initialized = 0;
 580
 581         CPU_FREE(cpu_present_set);
 582         cpu_present_set = NULL;
 583         cpu_present_set = 0;
 584
 585         CPU_FREE(cpu_affinity_set);
 586         cpu_affinity_set = NULL;
 587         cpu_affinity_setsize = 0;
 588
 589         CPU_FREE(cpu_saved_affinity_set);
 590         cpu_saved_affinity_set = NULL;
 591         cpu_saved_affinity_setsize = 0;
 592
 593         free(thread_even);
 594         free(core_even);
 595         free(package_even);
 596
 597         thread_even = NULL;
 598         core_even = NULL;
 599         package_even = NULL;
 600
 601         free(thread_odd);
 602         free(core_odd);
 603         free(package_odd);
 604
 605         thread_odd = NULL;
 606         core_odd = NULL;
 607         package_odd = NULL;
 608
 609         free(thread_delta);
 610         free(core_delta);
 611         free(package_delta);
 612
 613         thread_delta = NULL;
 614         core_delta = NULL;
 615         package_delta = NULL;
 616 }
 617
 618 /*
 619  * Parse a file containing a single int.
 620  */
 621 static int __attribute__ ((format(printf,1,2)))
 622 parse_int_file(const char *fmt, ...)
 623 {
 624         va_list args;
 625         char path[PATH_MAX];
 626         FILE *filep;
 627         int value;
 628
 629         va_start(args, fmt);
 630         vsnprintf(path, sizeof(path), fmt, args);
 631         va_end(args);
 632         filep = fopen(path, "r");
 633         if (!filep) {
 634                 ERROR("%s: open failed", path);
 635                 return -ERR_CANT_OPEN_FILE;
 636         }
 637         if (fscanf(filep, "%d", &value) != 1) {
 638                 ERROR("%s: failed to parse number from file", path);
 639                 return -ERR_CANT_READ_NUMBER;
 640         }
 641         fclose(filep);
 642         return value;
 643 }
 644
 645 /*
 646  * cpu_is_first_sibling_in_core(cpu)
 647  * return 1 if given CPU is 1st HT sibling in the core
 648  */
 649 static int
 650 cpu_is_first_sibling_in_core(int cpu)
 651 {
 652         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 653 }
 654
 655 /*
 656  * cpu_is_first_core_in_package(cpu)
 657  * return 1 if given CPU is 1st core in package
 658  */
 659 static int
 660 cpu_is_first_core_in_package(int cpu)
 661 {
 662         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 663 }
 664
 665 static int
 666 get_physical_package_id(int cpu)
 667 {
 668         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 669 }
 670
 671 static int
 672 get_core_id(int cpu)
 673 {
 674         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 675 }
 676
 677 static int
 678 get_num_ht_siblings(int cpu)
 679 {
 680         char path[80];
 681         FILE *filep;
 682         int sib1, sib2;
 683         int matches;
 684         char character;
 685
 686         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 687         filep = fopen(path, "r");
 688         if (!filep) {
 689                 ERROR("%s: open failed", path);
 690                 return -ERR_CANT_OPEN_FILE;
 691         }
 692         /*
 693          * file format:
 694          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 695          * otherwinse 1 sibling (self).
 696          */
 697         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 698
 699         fclose(filep);
 700
 701         if (matches == 3)
 702                 return 2;
 703         else
 704                 return 1;
 705 }
 706
 707 /*
 708  * Extract every data evolution for all CPU
 709  *
 710  * Core data is shared for all threads in one core: extracted only for the first thread
 711  * Package data is shared for all core in one package: extracted only for the first thread of the first core
 712  */
 713 static int __attribute__((warn_unused_result))
 714 for_all_cpus_delta(const struct thread_data *thread_new_base, const struct core_data *core_new_base, const struct pkg_data *pkg_new_base,
 715                    const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
 716 {
 717         int retval, pkg_no, core_no, thread_no;
 718
 719         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 720                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 721                         for (thread_no = 0; thread_no <
 722                                 topo.num_threads_per_core; ++thread_no) {
 723                                 struct thread_data *t_delta;
 724                                 const struct thread_data *t_old, *t_new;
 725                                 struct core_data *c_delta;
 726
 727                                 /* Get correct pointers for threads */
 728                                 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
 729                                 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
 730                                 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
 731
 732                                 /* Skip threads that disappeared */
 733                                 if (cpu_is_not_present(t_delta->cpu_id))
 734                                         continue;
 735
 736                                 /* c_delta is always required for delta_thread */
 737                                 c_delta = GET_CORE(core_delta, core_no, pkg_no);
 738
 739                                 /* calculate core delta only for 1st thread in core */
 740                                 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
 741                                         const struct core_data *c_old, *c_new;
 742
 743                                         c_new = GET_CORE(core_new_base, core_no, pkg_no);
 744                                         c_old = GET_CORE(core_old_base, core_no, pkg_no);
 745
 746                                         delta_core(c_delta, c_new, c_old);
 747                                 }
 748
 749                                 /* Always calculate thread delta */
 750                                 retval = delta_thread(t_delta, t_new, t_old, c_delta);
 751                                 if (retval)
 752                                         return retval;
 753
 754                                 /* calculate package delta only for 1st core in package */
 755                                 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
 756                                         struct pkg_data *p_delta;
 757                                         const struct pkg_data *p_old, *p_new;
 758
 759                                         p_delta = GET_PKG(package_delta, pkg_no);
 760                                         p_new = GET_PKG(pkg_new_base, pkg_no);
 761                                         p_old = GET_PKG(pkg_old_base, pkg_no);
 762
 763                                         delta_package(p_delta, p_new, p_old);
 764                                 }
 765                         }
 766                 }
 767         }
 768         return 0;
 769 }
 770
 771 /*
 772  * run func(cpu) on every cpu in /proc/stat
 773  * return max_cpu number
 774  */
 775 static int __attribute__((warn_unused_result))
 776 for_all_proc_cpus(int (func)(int))
 777 {
 778         FILE *fp;
 779         int cpu_num;
 780         int retval;
 781
 782         fp = fopen("/proc/stat", "r");
 783         if (!fp) {
 784                 ERROR("Failed to open /proc/stat");
 785                 return -ERR_CANT_OPEN_FILE;
 786         }
 787
 788         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 789         if (retval != 0) {
 790                 ERROR("Failed to parse /proc/stat");
 791                 fclose(fp);
 792                 return -ERR_CANT_READ_PROC_STAT;
 793         }
 794
 795         while (1) {
 796                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 797                 if (retval != 1)
 798                         break;
 799
 800                 retval = func(cpu_num);
 801                 if (retval) {
 802                         fclose(fp);
 803                         return(retval);
 804                 }
 805         }
 806         fclose(fp);
 807         return 0;
 808 }
 809
 810 /*
 811  * count_cpus()
 812  * remember the last one seen, it will be the max
 813  */
 814 static int
 815 count_cpus(int cpu)
 816 {
 817         if (topo.max_cpu_num < cpu)
 818                 topo.max_cpu_num = cpu;
 819
 820         topo.num_cpus += 1;
 821         return 0;
 822 }
 823 static int
 824 mark_cpu_present(int cpu)
 825 {
 826         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 827         return 0;
 828 }
 829
 830
 831 static void
 832 turbostat_submit (const char *plugin_instance,
 833         const char *type, const char *type_instance,
 834         gauge_t value)
 835 {
 836         value_list_t vl = VALUE_LIST_INIT;
 837         value_t v;
 838
 839         v.gauge = value;
 840         vl.values = &v;
 841         vl.values_len = 1;
 842         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 843         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 844         if (plugin_instance != NULL)
 845                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 846         sstrncpy (vl.type, type, sizeof (vl.type));
 847         if (type_instance != NULL)
 848                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 849
 850         plugin_dispatch_values (&vl);
 851 }
 852
 853 /*
 854  * column formatting convention & formats
 855  * package: "pk" 2 columns %2d
 856  * core: "cor" 3 columns %3d
 857  * CPU: "CPU" 3 columns %3d
 858  * Pkg_W: %6.2
 859  * Cor_W: %6.2
 860  * GFX_W: %5.2
 861  * RAM_W: %5.2
 862  * GHz: "GHz" 3 columns %3.2
 863  * TSC: "TSC" 3 columns %3.2
 864  * SMI: "SMI" 4 columns %4d
 865  * percentage " %pc3" %6.2
 866  * Perf Status percentage: %5.2
 867  * "CTMP" 4 columns %4d
 868  */
 869 static int
 870 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 871 {
 872         char name[12];
 873         double interval_float;
 874
 875         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 876
 877         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 878
 879         if (!aperf_mperf_unstable)
 880                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 881         if (!aperf_mperf_unstable)
 882                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 883
 884         /* GHz */
 885         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 886                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 887
 888         /* SMI */
 889         turbostat_submit(NULL, "current", name, t->smi_count);
 890
 891         /* print per-core data only for 1st thread in core */
 892         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 893                 goto done;
 894
 895         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 896
 897         if (do_core_cstate & (1 << 3))
 898                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 899         if (do_core_cstate & (1 << 6))
 900                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 901         if (do_core_cstate & (1 << 7))
 902                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 903
 904         if (do_dts)
 905                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 906
 907         /* print per-package data only for 1st core in package */
 908         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 909                 goto done;
 910
 911         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 912
 913         if (do_ptm)
 914                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 915
 916         if (do_pkg_cstate & (1 << 2))
 917                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 918         if (do_pkg_cstate & (1 << 3))
 919                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 920         if (do_pkg_cstate & (1 << 6))
 921                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 922         if (do_pkg_cstate & (1 << 7))
 923                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 924         if (do_pkg_cstate & (1 << 8))
 925                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 926         if (do_pkg_cstate & (1 << 9))
 927                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 928         if (do_pkg_cstate & (1 << 10))
 929                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 930
 931         if (do_rapl) {
 932                 if (do_rapl & RAPL_PKG)
 933                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 934                 if (do_rapl & RAPL_CORES)
 935                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 936                 if (do_rapl & RAPL_GFX)
 937                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 938                 if (do_rapl & RAPL_DRAM)
 939                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 940         }
 941 done:
 942         return 0;
 943 }
 944
 945 static int
 946 turbostat_read(user_data_t * not_used)
 947 {
 948         int ret;
 949
 950         if (!allocated) {
 951                 if ((ret = setup_all_buffers()) < 0)
 952                         return ret;
 953         }
 954
 955         if (for_all_proc_cpus(cpu_is_not_present)) {
 956                 free_all_buffers();
 957                 if ((ret = setup_all_buffers()) < 0)
 958                         return ret;
 959                 if (for_all_proc_cpus(cpu_is_not_present))
 960                         return -ERR_CPU_NOT_PRESENT;
 961         }
 962
 963         /* Saving the scheduling affinity, as it will be modified by get_counters */
 964         if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0)
 965                 return -ERR_CPU_SAVE_SCHED_AFFINITY;
 966
 967         if (!initialized) {
 968                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 969                         goto out;
 970                 gettimeofday(&tv_even, (struct timezone *)NULL);
 971                 is_even = 1;
 972                 initialized = 1;
 973                 ret = 0;
 974                 goto out;
 975         }
 976
 977         if (is_even) {
 978                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 979                         goto out;
 980                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 981                 is_even = 0;
 982                 timersub(&tv_odd, &tv_even, &tv_delta);
 983                 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 984                         goto out;
 985                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
 986                         goto out;
 987         } else {
 988                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 989                         goto out;
 990                 gettimeofday(&tv_even, (struct timezone *)NULL);
 991                 is_even = 1;
 992                 timersub(&tv_even, &tv_odd, &tv_delta);
 993                 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 994                         goto out;
 995                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
 996                         goto out;
 997         }
 998         ret = 0;
 999 out:
1000         /*
1001          * Let's restore the affinity
1002          * This might fail if the number of CPU changed, but we can't do anything in that case..
1003          */
1004         (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1005         return ret;
1006 }
1007
1008 static int __attribute__((warn_unused_result))
1009 check_dev_msr()
1010 {
1011         struct stat sb;
1012
1013         if (stat("/dev/cpu/0/msr", &sb)) {
1014                 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
1015                 return -ERR_NO_MSR;
1016         }
1017         return 0;
1018 }
1019
1020 static int __attribute__((warn_unused_result))
1021 check_super_user()
1022 {
1023         if (getuid() != 0) {
1024                 ERROR("must be root");
1025                 return -ERR_NOT_ROOT;
1026         }
1027         return 0;
1028 }
1029
1030 /*
1031  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1032  * the Thermal Control Circuit (TCC) activates.
1033  * This is usually equal to tjMax.
1034  *
1035  * Older processors do not have this MSR, so there we guess,
1036  * but also allow conficuration over-ride with "TCCActivationTemp".
1037  *
1038  * Several MSR temperature values are in units of degrees-C
1039  * below this value, including the Digital Thermal Sensor (DTS),
1040  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1041  */
1042 static int __attribute__((warn_unused_result))
1043 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1044 {
1045         unsigned long long msr;
1046         unsigned int target_c_local;
1047
1048         /* tcc_activation_temp is used only for dts or ptm */
1049         if (!(do_dts || do_ptm))
1050                 return 0;
1051
1052         /* this is a per-package concept */
1053         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1054                 return 0;
1055
1056         if (tcc_activation_temp != 0) {
1057                 p->tcc_activation_temp = tcc_activation_temp;
1058                 return 0;
1059         }
1060
1061         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
1062                 goto guess;
1063
1064         target_c_local = (msr >> 16) & 0xFF;
1065
1066         if (!target_c_local)
1067                 goto guess;
1068
1069         p->tcc_activation_temp = target_c_local;
1070
1071         return 0;
1072
1073 guess:
1074         p->tcc_activation_temp = TJMAX_DEFAULT;
1075         WARNING("cpu%d: Guessing tjMax %d C, Please use TCCActivationTemp to specify",
1076                 t->cpu_id, p->tcc_activation_temp);
1077
1078         return 0;
1079 }
1080
1081 /*
1082  * Identify the functionality of the CPU
1083  */
1084 static int __attribute__((warn_unused_result))
1085 probe_cpu()
1086 {
1087         unsigned int eax, ebx, ecx, edx, max_level;
1088         unsigned int fms, family, model;
1089
1090         /* CPUID(0):
1091          * - EAX: Maximum Input Value for Basic CPUID Information
1092          * - EBX: "Genu" (0x756e6547)
1093          * - EDX: "ineI" (0x49656e69)
1094          * - ECX: "ntel" (0x6c65746e)
1095          */
1096         max_level = ebx = ecx = edx = 0;
1097         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1098         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1099                 ERROR("Unsupported CPU");
1100                 return -UNSUPPORTED_CPU;
1101         }
1102
1103         /* CPUID(1):
1104          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1105          *  + 4-7:   Model ID
1106          *  + 8-11:  Family ID
1107          *  + 12-13: Processor type
1108          *  + 16-19: Extended Model ID
1109          *  + 20-27: Extended Family ID
1110          * - EDX: Feature Information:
1111          *  + 5: Support for MSR read/write operations
1112          */
1113         fms = ebx = ecx = edx = 0;
1114         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1115         family = (fms >> 8) & 0xf;
1116         model = (fms >> 4) & 0xf;
1117         if (family == 0xf)
1118                 family += (fms >> 20) & 0xf;
1119         if (family == 6 || family == 0xf)
1120                 model += ((fms >> 16) & 0xf) << 4;
1121         if (!(edx & (1 << 5))) {
1122                 ERROR("CPUID: no MSR");
1123                 return -ERR_NO_MSR;
1124         }
1125
1126         /*
1127          * CPUID(0x80000000):
1128          * - EAX: Maximum Input Value for Extended Function CPUID Information
1129          *
1130          * This allows us to verify if the CPUID(0x80000007) can be called
1131          *
1132          * This check is valid for both Intel and AMD.
1133          */
1134         max_level = ebx = ecx = edx = 0;
1135         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1136         if (max_level < 0x80000007) {
1137                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1138                 return -ERR_NO_INVARIANT_TSC;
1139         }
1140
1141         /*
1142          * CPUID(0x80000007):
1143          * - EDX:
1144          *  + 8: Invariant TSC available if set
1145          *
1146          * This check is valid for both Intel and AMD
1147          */
1148         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1149         if (!(edx & (1 << 8))) {
1150                 ERROR("No invariant TSC");
1151                 return -ERR_NO_INVARIANT_TSC;
1152         }
1153
1154         /*
1155          * CPUID(6):
1156          * - EAX:
1157          *  + 0: Digital temperature sensor is supported if set
1158          *  + 6: Package thermal management is supported if set
1159          * - ECX:
1160          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1161          *  + 3: The processor supports performance-energy bias preference if set.
1162          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1163          *
1164          * This check is valid for both Intel and AMD
1165          */
1166         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1167         do_dts = eax & (1 << 0);
1168         do_ptm = eax & (1 << 6);
1169         if (!(ecx & (1 << 0))) {
1170                 ERROR("No APERF");
1171                 return -ERR_NO_APERF;
1172         }
1173
1174         /*
1175          * Enable or disable C states depending on the model and family
1176          */
1177         if (family == 6) {
1178                 switch (model) {
1179                 /* Atom (partial) */
1180                 case 0x27:
1181                         do_core_cstate = 0;
1182                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1183                         break;
1184                 /* Silvermont */
1185                 case 0x37: /* BYT */
1186                 case 0x4A:
1187                 case 0x4D: /* AVN */
1188                 case 0x5A:
1189                 case 0x5D:
1190                         do_core_cstate = (1 << 1) | (1 << 6);
1191                         do_pkg_cstate = (1 << 6);
1192                         break;
1193                 /* Nehalem */
1194                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1195                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1196                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1197                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1198                         do_core_cstate = (1 << 3) | (1 << 6);
1199                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1200                         break;
1201                 /* Westmere */
1202                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1203                 case 0x2C: /* Westmere EP - Gulftown */
1204                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1205                         do_core_cstate = (1 << 3) | (1 << 6);
1206                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1207                         break;
1208                 /* Sandy Bridge */
1209                 case 0x2A: /* SNB */
1210                 case 0x2D: /* SNB Xeon */
1211                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1212                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1213                         break;
1214                 /* Ivy Bridge */
1215                 case 0x3A: /* IVB */
1216                 case 0x3E: /* IVB Xeon */
1217                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1218                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1219                         break;
1220                 /* Haswell Bridge */
1221                 case 0x3C: /* HSW */
1222                 case 0x3F: /* HSW */
1223                 case 0x46: /* HSW */
1224                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1225                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1226                         break;
1227                 case 0x45: /* HSW */
1228                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1229                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1230                         break;
1231                 /* Broadwel */
1232                 case 0x4F: /* BDW */
1233                 case 0x56: /* BDX-DE */
1234                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1235                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1236                         break;
1237                 case 0x3D: /* BDW */
1238                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1239                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1240                         break;
1241                 default:
1242                         ERROR("Unsupported CPU");
1243                 }
1244                 switch (model) {
1245                 case 0x2A:
1246                 case 0x3A:
1247                 case 0x3C:
1248                 case 0x45:
1249                 case 0x46:
1250                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1251                         break;
1252                 case 0x3F:
1253                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1254                         break;
1255                 case 0x2D:
1256                 case 0x3E:
1257                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1258                         break;
1259                 case 0x37:
1260                 case 0x4D:
1261                         do_rapl = RAPL_PKG | RAPL_CORES;
1262                         break;
1263                 default:
1264                         do_rapl = 0;
1265                 }
1266         } else {
1267                 ERROR("Unsupported CPU");
1268                 return -UNSUPPORTED_CPU;
1269         }
1270
1271         if (do_rapl) {
1272                 unsigned long msr;
1273                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1274                         return 0;
1275
1276                 if (model == 0x37)
1277                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1278                 else
1279                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1280         }
1281
1282         return 0;
1283 }
1284
1285
1286
1287 static int __attribute__((warn_unused_result))
1288 topology_probe()
1289 {
1290         int i;
1291         int ret;
1292         int max_core_id = 0;
1293         int max_package_id = 0;
1294         int max_siblings = 0;
1295         struct cpu_topology {
1296                 int core_id;
1297                 int physical_package_id;
1298         } *cpus;
1299
1300         /* Initialize num_cpus, max_cpu_num */
1301         topo.num_cpus = 0;
1302         topo.max_cpu_num = 0;
1303         ret = for_all_proc_cpus(count_cpus);
1304         if (ret < 0)
1305                 return ret;
1306
1307         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1308
1309         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1310         if (cpus == NULL) {
1311                 ERROR("calloc cpus");
1312                 return -ERR_CALLOC;
1313         }
1314
1315         /*
1316          * Allocate and initialize cpu_present_set
1317          */
1318         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1319         if (cpu_present_set == NULL) {
1320                 free(cpus);
1321                 ERROR("CPU_ALLOC");
1322                 return -ERR_CPU_ALLOC;
1323         }
1324         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1325         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1326         ret = for_all_proc_cpus(mark_cpu_present);
1327         if (ret < 0) {
1328                 free(cpus);
1329                 return ret;
1330         }
1331
1332         /*
1333          * Allocate and initialize cpu_affinity_set
1334          */
1335         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1336         if (cpu_affinity_set == NULL) {
1337                 free(cpus);
1338                 ERROR("CPU_ALLOC");
1339                 return -ERR_CPU_ALLOC;
1340         }
1341         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1342         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1343
1344
1345         /*
1346          * Allocate and initialize cpu_saved_affinity_set
1347          */
1348         cpu_saved_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1349         if (cpu_saved_affinity_set == NULL) {
1350                 free(cpus);
1351                 ERROR("CPU_ALLOC");
1352                 return -ERR_CPU_ALLOC;
1353         }
1354         cpu_saved_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1355         CPU_ZERO_S(cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1356
1357
1358         /*
1359          * For online cpus
1360          * find max_core_id, max_package_id
1361          */
1362         for (i = 0; i <= topo.max_cpu_num; ++i) {
1363                 int siblings;
1364
1365                 if (cpu_is_not_present(i)) {
1366                         WARNING("cpu%d NOT PRESENT", i);
1367                         continue;
1368                 }
1369                 cpus[i].core_id = get_core_id(i);
1370                 if (cpus[i].core_id < 0)
1371                         return cpus[i].core_id;
1372                 if (cpus[i].core_id > max_core_id)
1373                         max_core_id = cpus[i].core_id;
1374
1375                 cpus[i].physical_package_id = get_physical_package_id(i);
1376                 if (cpus[i].physical_package_id < 0)
1377                         return cpus[i].physical_package_id;
1378                 if (cpus[i].physical_package_id > max_package_id)
1379                         max_package_id = cpus[i].physical_package_id;
1380
1381                 siblings = get_num_ht_siblings(i);
1382                 if (siblings < 0)
1383                         return siblings;
1384                 if (siblings > max_siblings)
1385                         max_siblings = siblings;
1386                 DEBUG("cpu %d pkg %d core %d\n",
1387                         i, cpus[i].physical_package_id, cpus[i].core_id);
1388         }
1389         topo.num_cores_per_pkg = max_core_id + 1;
1390         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1391                 max_core_id, topo.num_cores_per_pkg);
1392
1393         topo.num_packages = max_package_id + 1;
1394         DEBUG("max_package_id %d, sizing for %d packages\n",
1395                 max_package_id, topo.num_packages);
1396
1397         topo.num_threads_per_core = max_siblings;
1398         DEBUG("max_siblings %d\n", max_siblings);
1399
1400         free(cpus);
1401         return 0;
1402 }
1403
1404 static int
1405 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1406 {
1407         int i;
1408
1409         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1410                 topo.num_packages, sizeof(struct thread_data));
1411         if (*t == NULL)
1412                 goto error;
1413
1414         for (i = 0; i < topo.num_threads_per_core *
1415                 topo.num_cores_per_pkg * topo.num_packages; i++)
1416                 (*t)[i].cpu_id = -1;
1417
1418         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1419                 sizeof(struct core_data));
1420         if (*c == NULL)
1421                 goto error;
1422
1423         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1424                 (*c)[i].core_id = -1;
1425
1426         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1427         if (*p == NULL)
1428                 goto error;
1429
1430         for (i = 0; i < topo.num_packages; i++)
1431                 (*p)[i].package_id = i;
1432
1433         return 0;
1434 error:
1435         ERROR("calloc counters");
1436         return -ERR_CALLOC;
1437 }
1438 /*
1439  * init_counter()
1440  *
1441  * set cpu_id, core_num, pkg_num
1442  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1443  *
1444  * increment topo.num_cores when 1st core in pkg seen
1445  */
1446 static int
1447 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1448         struct pkg_data *pkg_base, int thread_num, int core_num,
1449         int pkg_num, int cpu_id)
1450 {
1451         int ret;
1452         struct thread_data *t;
1453         struct core_data *c;
1454         struct pkg_data *p;
1455
1456         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1457         c = GET_CORE(core_base, core_num, pkg_num);
1458         p = GET_PKG(pkg_base, pkg_num);
1459
1460         t->cpu_id = cpu_id;
1461         if (thread_num == 0) {
1462                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1463                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1464                         return ret;
1465                 } else if (ret != 0) {
1466                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1467                 }
1468         }
1469
1470         c->core_id = core_num;
1471         p->package_id = pkg_num;
1472
1473         return 0;
1474 }
1475
1476
1477 static int
1478 initialize_counters(int cpu_id)
1479 {
1480         int my_thread_id, my_core_id, my_package_id;
1481         int ret;
1482
1483         my_package_id = get_physical_package_id(cpu_id);
1484         if (my_package_id < 0)
1485                 return my_package_id;
1486         my_core_id = get_core_id(cpu_id);
1487         if (my_core_id < 0)
1488                 return my_core_id;
1489
1490         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1491                 return ret;
1492         } else if (ret != 0) {
1493                 my_thread_id = 0;
1494                 topo.num_cores++;
1495         } else {
1496                 my_thread_id = 1;
1497         }
1498
1499         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1500         if (ret < 0)
1501                 return ret;
1502         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1503         if (ret < 0)
1504                 return ret;
1505         ret = init_counter(DELTA_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1506         if (ret < 0)
1507                 return ret;
1508         return 0;
1509 }
1510
1511 #define DO_OR_GOTO_ERR(something) \
1512 do {                         \
1513         ret = (something);     \
1514         if (ret < 0)         \
1515                 goto err;    \
1516 } while (0)
1517
1518 static int setup_all_buffers(void)
1519 {
1520         int ret;
1521
1522         DO_OR_GOTO_ERR(topology_probe());
1523         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1524         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1525         DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1526         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1527         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1528         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1529
1530         allocated = 1;
1531         return 0;
1532 err:
1533         free_all_buffers();
1534         return ret;
1535 }
1536
1537 static int
1538 turbostat_init(void)
1539 {
1540         int ret;
1541
1542         DO_OR_GOTO_ERR(check_super_user());
1543         DO_OR_GOTO_ERR(probe_cpu());
1544         DO_OR_GOTO_ERR(check_dev_msr());
1545         DO_OR_GOTO_ERR(setup_all_buffers());
1546
1547         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1548
1549         return 0;
1550 err:
1551         free_all_buffers();
1552         return ret;
1553 }
1554
1555 static const char *config_keys[] =
1556 {
1557         "TCCActivationTemp",
1558 };
1559 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1560
1561 static int
1562 turbostat_config(const char *key, const char *value)
1563 {
1564         long unsigned int tmp_val;
1565         char *end;
1566
1567         if (strcasecmp("TCCActivationTemp", key) == 0) {
1568                 tmp_val = strtoul(value, &end, 0);
1569                 if (*end != '\0' || tmp_val > UINT_MAX)
1570                         return -1;
1571                 tcc_activation_temp = (unsigned int) tmp_val;
1572         } else {
1573                 return -1;
1574         }
1575         return 0;
1576 }
1577
1578 void module_register(void);
1579 void module_register(void)
1580 {
1581         plugin_register_init(PLUGIN_NAME, turbostat_init);
1582         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1583 }