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
 259 /*****************************
 260  *  MSR Manipulation helpers *
 261  *****************************/
 262
 263 /*
 264  * Open a MSR device for reading
 265  * Can change the scheduling affinity of the current process if multiple_read is 1
 266  */
 267 static int __attribute__((warn_unused_result))
 268 open_msr(int cpu, _Bool multiple_read)
 269 {
 270         char pathname[32];
 271         int fd;
 272
 273         /*
 274          * If we need to do multiple read, let's migrate to the CPU
 275          * Otherwise, we would lose time calling functions on another CPU
 276          *
 277          * If we are not yet initialized (cpu_affinity_setsize = 0),
 278          * we need to skip this optimisation.
 279          */
 280         if (multiple_read && cpu_affinity_setsize) {
 281                 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 282                 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 283                 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
 284                         ERROR("Could not migrate to CPU %d", cpu);
 285                         return -ERR_CPU_MIGRATE;
 286                 }
 287         }
 288
 289         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 290         fd = open(pathname, O_RDONLY);
 291         if (fd < 0)
 292                 return -ERR_CANT_OPEN_MSR;
 293         return fd;
 294 }
 295
 296 /*
 297  * Read a single MSR from an open file descriptor
 298  */
 299 static int __attribute__((warn_unused_result))
 300 read_msr(int fd, off_t offset, unsigned long long *msr)
 301 {
 302         ssize_t retval;
 303
 304         retval = pread(fd, msr, sizeof *msr, offset);
 305
 306         if (retval != sizeof *msr) {
 307                 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
 308                 return -1;
 309         }
 310         return 0;
 311 }
 312
 313 /*
 314  * Open a MSR device for reading, read the value asked for and close it.
 315  * This call will not affect the scheduling affinity of this thread.
 316  */
 317 static int __attribute__((warn_unused_result))
 318 get_msr(int cpu, off_t offset, unsigned long long *msr)
 319 {
 320         ssize_t retval;
 321         int fd;
 322
 323         fd = open_msr(cpu, 0);
 324         if (fd < 0)
 325                 return fd;
 326         retval = read_msr(fd, offset, msr);
 327         close(fd);
 328         return retval;
 329 }
 330
 331
 332 /********************************
 333  * Raw data acquisition (1 CPU) *
 334  ********************************/
 335
 336 /*
 337  * Read every data avalaible for a single CPU
 338  *
 339  * Core data is shared for all threads in one core: extracted only for the first thread
 340  * Package data is shared for all core in one package: extracted only for the first thread of the first core
 341  *
 342  * Side effect: migrates to the targeted CPU
 343  */
 344 static int __attribute__((warn_unused_result))
 345 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 346 {
 347         int cpu = t->cpu_id;
 348         unsigned long long msr;
 349         int msr_fd;
 350         int retval = 0;
 351
 352         msr_fd = open_msr(cpu, 1);
 353         if (msr_fd < 0)
 354                 return msr_fd;
 355
 356 #define READ_MSR(msr, dst)                      \
 357 do {                                            \
 358         if (read_msr(msr_fd, msr, dst)) {       \
 359                 retval = -ERR_##msr;            \
 360                 goto out;                       \
 361         }                                       \
 362 } while (0)
 363
 364         READ_MSR(MSR_IA32_TSC, &t->tsc);
 365
 366         READ_MSR(MSR_IA32_APERF, &t->aperf);
 367         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 368
 369         READ_MSR(MSR_SMI_COUNT, &msr);
 370         t->smi_count = msr & 0xFFFFFFFF;
 371
 372         /* collect core counters only for 1st thread in core */
 373         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 374                 retval = 0;
 375                 goto out;
 376         }
 377
 378         if (do_core_cstate & (1 << 3))
 379                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 380         if (do_core_cstate & (1 << 6))
 381                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 382         if (do_core_cstate & (1 << 7))
 383                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 384
 385         if (do_dts) {
 386                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 387                 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 388         }
 389
 390         /* collect package counters only for 1st core in package */
 391         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 392                 retval = 0;
 393                 goto out;
 394         }
 395
 396         if (do_pkg_cstate & (1 << 2))
 397                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 398         if (do_pkg_cstate & (1 << 3))
 399                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 400         if (do_pkg_cstate & (1 << 6))
 401                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 402         if (do_pkg_cstate & (1 << 7))
 403                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 404         if (do_pkg_cstate & (1 << 8))
 405                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 406         if (do_pkg_cstate & (1 << 9))
 407                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 408         if (do_pkg_cstate & (1 << 10))
 409                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 410
 411         if (do_rapl & RAPL_PKG) {
 412                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 413                 p->energy_pkg = msr & 0xFFFFFFFF;
 414         }
 415         if (do_rapl & RAPL_CORES) {
 416                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 417                 p->energy_cores = msr & 0xFFFFFFFF;
 418         }
 419         if (do_rapl & RAPL_DRAM) {
 420                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 421                 p->energy_dram = msr & 0xFFFFFFFF;
 422         }
 423         if (do_rapl & RAPL_GFX) {
 424                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 425                 p->energy_gfx = msr & 0xFFFFFFFF;
 426         }
 427         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 428                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 429                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 430         }
 431         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 432                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 433                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 434         }
 435         if (do_ptm) {
 436                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 437                 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 438         }
 439
 440 out:
 441         close(msr_fd);
 442         return retval;
 443 }
 444
 445
 446 /**********************************
 447  * Evaluating the changes (1 CPU) *
 448  **********************************/
 449
 450 /*
 451  * Do delta = new - old on 32bits cyclique intergers
 452  */
 453 #define DELTA_WRAP32(delta, new, old)                   \
 454         if (new > old) {                                \
 455                 delta = new - old;                      \
 456         } else {                                        \
 457                 delta = 0x100000000 + new - old;        \
 458         }
 459
 460 /*
 461  * Extract the evolution old->new in delta at a package level
 462  * (some are not new-delta, e.g. temperature)
 463  */
 464 static inline void
 465 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
 466 {
 467         delta->pc2 = new->pc2 - old->pc2;
 468         delta->pc3 = new->pc3 - old->pc3;
 469         delta->pc6 = new->pc6 - old->pc6;
 470         delta->pc7 = new->pc7 - old->pc7;
 471         delta->pc8 = new->pc8 - old->pc8;
 472         delta->pc9 = new->pc9 - old->pc9;
 473         delta->pc10 = new->pc10 - old->pc10;
 474         delta->pkg_temp_c = new->pkg_temp_c;
 475
 476         DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
 477         DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
 478         DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
 479         DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
 480         DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 481         DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 482 }
 483
 484 /*
 485  * Extract the evolution old->new in delta at a core level
 486  * (some are not new-delta, e.g. temperature)
 487  */
 488 static inline void
 489 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
 490 {
 491         delta->c3 = new->c3 - old->c3;
 492         delta->c6 = new->c6 - old->c6;
 493         delta->c7 = new->c7 - old->c7;
 494         delta->core_temp_c = new->core_temp_c;
 495 }
 496
 497 /*
 498  * Extract the evolution old->new in delta at a package level
 499  * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
 500  */
 501 static inline int __attribute__((warn_unused_result))
 502 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
 503         const struct core_data *core_delta)
 504 {
 505         delta->tsc = new->tsc - old->tsc;
 506
 507         /* check for TSC < 1 Mcycles over interval */
 508         if (delta->tsc < (1000 * 1000)) {
 509                 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
 510                 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
 511                 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
 512                 return -1;
 513         }
 514
 515         delta->c1 = new->c1 - old->c1;
 516
 517         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 518                 delta->aperf = new->aperf - old->aperf;
 519                 delta->mperf = new->mperf - old->mperf;
 520         } else {
 521                 if (!aperf_mperf_unstable) {
 522                         WARNING(" APERF or MPERF went backwards * ");
 523                         WARNING("* Frequency results do not cover entire interval *");
 524                         WARNING("* fix this by running Linux-2.6.30 or later *");
 525
 526                         aperf_mperf_unstable = 1;
 527                 }
 528         }
 529
 530         /*
 531          * As counter collection is not atomic,
 532          * it is possible for mperf's non-halted cycles + idle states
 533          * to exceed TSC's all cycles: show c1 = 0% in that case.
 534          */
 535         if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
 536                 delta->c1 = 0;
 537         else {
 538                 /* normal case, derive c1 */
 539                 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
 540                         - core_delta->c6 - core_delta->c7;
 541         }
 542
 543         if (delta->mperf == 0) {
 544                 WARNING("cpu%d MPERF 0!", old->cpu_id);
 545                 delta->mperf = 1;       /* divide by 0 protection */
 546         }
 547
 548         delta->smi_count = new->smi_count - old->smi_count;
 549
 550         return 0;
 551 }
 552
 553 /**********************************
 554  * Submitting the results (1 CPU) *
 555  **********************************/
 556
 557 /*
 558  * Submit one gauge value
 559  */
 560 static void
 561 turbostat_submit (const char *plugin_instance,
 562         const char *type, const char *type_instance,
 563         gauge_t value)
 564 {
 565         value_list_t vl = VALUE_LIST_INIT;
 566         value_t v;
 567
 568         v.gauge = value;
 569         vl.values = &v;
 570         vl.values_len = 1;
 571         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 572         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 573         if (plugin_instance != NULL)
 574                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 575         sstrncpy (vl.type, type, sizeof (vl.type));
 576         if (type_instance != NULL)
 577                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 578
 579         plugin_dispatch_values (&vl);
 580 }
 581
 582 /*
 583  * Submit every data for a single CPU
 584  *
 585  * Core data is shared for all threads in one core: submitted only for the first thread
 586  * Package data is shared for all core in one package: submitted only for the first thread of the first core
 587  */
 588 static int
 589 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 590 {
 591         char name[12];
 592         double interval_float;
 593
 594         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 595
 596         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 597
 598         if (!aperf_mperf_unstable)
 599                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 600         if (!aperf_mperf_unstable)
 601                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 602
 603         /* GHz */
 604         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 605                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 606
 607         /* SMI */
 608         turbostat_submit(NULL, "current", name, t->smi_count);
 609
 610         /* submit per-core data only for 1st thread in core */
 611         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 612                 goto done;
 613
 614         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 615
 616         if (do_core_cstate & (1 << 3))
 617                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 618         if (do_core_cstate & (1 << 6))
 619                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 620         if (do_core_cstate & (1 << 7))
 621                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 622
 623         if (do_dts)
 624                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 625
 626         /* submit per-package data only for 1st core in package */
 627         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 628                 goto done;
 629
 630         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 631
 632         if (do_ptm)
 633                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 634
 635         if (do_pkg_cstate & (1 << 2))
 636                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 637         if (do_pkg_cstate & (1 << 3))
 638                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 639         if (do_pkg_cstate & (1 << 6))
 640                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 641         if (do_pkg_cstate & (1 << 7))
 642                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 643         if (do_pkg_cstate & (1 << 8))
 644                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 645         if (do_pkg_cstate & (1 << 9))
 646                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 647         if (do_pkg_cstate & (1 << 10))
 648                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 649
 650         if (do_rapl) {
 651                 if (do_rapl & RAPL_PKG)
 652                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 653                 if (do_rapl & RAPL_CORES)
 654                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 655                 if (do_rapl & RAPL_GFX)
 656                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 657                 if (do_rapl & RAPL_DRAM)
 658                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 659         }
 660 done:
 661         return 0;
 662 }
 663
 664 /**********************************
 665  * Looping function over all CPUs *
 666  **********************************/
 667
 668 /*
 669  * Check if a given cpu id is in our compiled list of existing CPUs
 670  *
 671  * CHECKME: Why do we need this?
 672  */
 673 static int
 674 cpu_is_not_present(int cpu)
 675 {
 676         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 677 }
 678
 679 /*
 680  * Loop on all CPUs in topological order
 681  *
 682  * Skip 'non-present' cpus (CHECKME: Why do we need this?)
 683  * Return the error code at the first error or 0
 684  */
 685 static int __attribute__((warn_unused_result))
 686 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 687         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 688 {
 689         int retval, pkg_no, core_no, thread_no;
 690
 691         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 692                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 693                         for (thread_no = 0; thread_no <
 694                                 topo.num_threads_per_core; ++thread_no) {
 695                                 struct thread_data *t;
 696                                 struct core_data *c;
 697                                 struct pkg_data *p;
 698
 699                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 700
 701                                 if (cpu_is_not_present(t->cpu_id))
 702                                         continue;
 703
 704                                 c = GET_CORE(core_base, core_no, pkg_no);
 705                                 p = GET_PKG(pkg_base, pkg_no);
 706
 707                                 retval = func(t, c, p);
 708                                 if (retval)
 709                                         return retval;
 710                         }
 711                 }
 712         }
 713         return 0;
 714 }
 715
 716 /*
 717  * Dedicated loop: Extract every data evolution for all CPU
 718  *
 719  * Core data is shared for all threads in one core: extracted only for the first thread
 720  * Package data is shared for all core in one package: extracted only for the first thread of the first core
 721  */
 722 static int __attribute__((warn_unused_result))
 723 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,
 724                    const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
 725 {
 726         int retval, pkg_no, core_no, thread_no;
 727
 728         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 729                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 730                         for (thread_no = 0; thread_no <
 731                                 topo.num_threads_per_core; ++thread_no) {
 732                                 struct thread_data *t_delta;
 733                                 const struct thread_data *t_old, *t_new;
 734                                 struct core_data *c_delta;
 735
 736                                 /* Get correct pointers for threads */
 737                                 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
 738                                 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
 739                                 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
 740
 741                                 /* Skip threads that disappeared */
 742                                 if (cpu_is_not_present(t_delta->cpu_id))
 743                                         continue;
 744
 745                                 /* c_delta is always required for delta_thread */
 746                                 c_delta = GET_CORE(core_delta, core_no, pkg_no);
 747
 748                                 /* calculate core delta only for 1st thread in core */
 749                                 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
 750                                         const struct core_data *c_old, *c_new;
 751
 752                                         c_new = GET_CORE(core_new_base, core_no, pkg_no);
 753                                         c_old = GET_CORE(core_old_base, core_no, pkg_no);
 754
 755                                         delta_core(c_delta, c_new, c_old);
 756                                 }
 757
 758                                 /* Always calculate thread delta */
 759                                 retval = delta_thread(t_delta, t_new, t_old, c_delta);
 760                                 if (retval)
 761                                         return retval;
 762
 763                                 /* calculate package delta only for 1st core in package */
 764                                 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
 765                                         struct pkg_data *p_delta;
 766                                         const struct pkg_data *p_old, *p_new;
 767
 768                                         p_delta = GET_PKG(package_delta, pkg_no);
 769                                         p_new = GET_PKG(pkg_new_base, pkg_no);
 770                                         p_old = GET_PKG(pkg_old_base, pkg_no);
 771
 772                                         delta_package(p_delta, p_new, p_old);
 773                                 }
 774                         }
 775                 }
 776         }
 777         return 0;
 778 }
 779
 780
 781 static void
 782 free_all_buffers(void)
 783 {
 784         allocated = 0;
 785         initialized = 0;
 786
 787         CPU_FREE(cpu_present_set);
 788         cpu_present_set = NULL;
 789         cpu_present_set = 0;
 790
 791         CPU_FREE(cpu_affinity_set);
 792         cpu_affinity_set = NULL;
 793         cpu_affinity_setsize = 0;
 794
 795         CPU_FREE(cpu_saved_affinity_set);
 796         cpu_saved_affinity_set = NULL;
 797         cpu_saved_affinity_setsize = 0;
 798
 799         free(thread_even);
 800         free(core_even);
 801         free(package_even);
 802
 803         thread_even = NULL;
 804         core_even = NULL;
 805         package_even = NULL;
 806
 807         free(thread_odd);
 808         free(core_odd);
 809         free(package_odd);
 810
 811         thread_odd = NULL;
 812         core_odd = NULL;
 813         package_odd = NULL;
 814
 815         free(thread_delta);
 816         free(core_delta);
 817         free(package_delta);
 818
 819         thread_delta = NULL;
 820         core_delta = NULL;
 821         package_delta = NULL;
 822 }
 823
 824 /*
 825  * Parse a file containing a single int.
 826  */
 827 static int __attribute__ ((format(printf,1,2)))
 828 parse_int_file(const char *fmt, ...)
 829 {
 830         va_list args;
 831         char path[PATH_MAX];
 832         FILE *filep;
 833         int value;
 834
 835         va_start(args, fmt);
 836         vsnprintf(path, sizeof(path), fmt, args);
 837         va_end(args);
 838         filep = fopen(path, "r");
 839         if (!filep) {
 840                 ERROR("%s: open failed", path);
 841                 return -ERR_CANT_OPEN_FILE;
 842         }
 843         if (fscanf(filep, "%d", &value) != 1) {
 844                 ERROR("%s: failed to parse number from file", path);
 845                 return -ERR_CANT_READ_NUMBER;
 846         }
 847         fclose(filep);
 848         return value;
 849 }
 850
 851 /*
 852  * cpu_is_first_sibling_in_core(cpu)
 853  * return 1 if given CPU is 1st HT sibling in the core
 854  */
 855 static int
 856 cpu_is_first_sibling_in_core(int cpu)
 857 {
 858         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 859 }
 860
 861 /*
 862  * cpu_is_first_core_in_package(cpu)
 863  * return 1 if given CPU is 1st core in package
 864  */
 865 static int
 866 cpu_is_first_core_in_package(int cpu)
 867 {
 868         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 869 }
 870
 871 static int
 872 get_physical_package_id(int cpu)
 873 {
 874         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 875 }
 876
 877 static int
 878 get_core_id(int cpu)
 879 {
 880         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 881 }
 882
 883 static int
 884 get_num_ht_siblings(int cpu)
 885 {
 886         char path[80];
 887         FILE *filep;
 888         int sib1, sib2;
 889         int matches;
 890         char character;
 891
 892         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 893         filep = fopen(path, "r");
 894         if (!filep) {
 895                 ERROR("%s: open failed", path);
 896                 return -ERR_CANT_OPEN_FILE;
 897         }
 898         /*
 899          * file format:
 900          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 901          * otherwinse 1 sibling (self).
 902          */
 903         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 904
 905         fclose(filep);
 906
 907         if (matches == 3)
 908                 return 2;
 909         else
 910                 return 1;
 911 }
 912
 913 /*
 914  * run func(cpu) on every cpu in /proc/stat
 915  * return max_cpu number
 916  */
 917 static int __attribute__((warn_unused_result))
 918 for_all_proc_cpus(int (func)(int))
 919 {
 920         FILE *fp;
 921         int cpu_num;
 922         int retval;
 923
 924         fp = fopen("/proc/stat", "r");
 925         if (!fp) {
 926                 ERROR("Failed to open /proc/stat");
 927                 return -ERR_CANT_OPEN_FILE;
 928         }
 929
 930         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 931         if (retval != 0) {
 932                 ERROR("Failed to parse /proc/stat");
 933                 fclose(fp);
 934                 return -ERR_CANT_READ_PROC_STAT;
 935         }
 936
 937         while (1) {
 938                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 939                 if (retval != 1)
 940                         break;
 941
 942                 retval = func(cpu_num);
 943                 if (retval) {
 944                         fclose(fp);
 945                         return(retval);
 946                 }
 947         }
 948         fclose(fp);
 949         return 0;
 950 }
 951
 952 /*
 953  * count_cpus()
 954  * remember the last one seen, it will be the max
 955  */
 956 static int
 957 count_cpus(int cpu)
 958 {
 959         if (topo.max_cpu_num < cpu)
 960                 topo.max_cpu_num = cpu;
 961
 962         topo.num_cpus += 1;
 963         return 0;
 964 }
 965 static int
 966 mark_cpu_present(int cpu)
 967 {
 968         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 969         return 0;
 970 }
 971
 972
 973 static int setup_all_buffers(void);
 974
 975 static int
 976 turbostat_read(user_data_t * not_used)
 977 {
 978         int ret;
 979
 980         if (!allocated) {
 981                 if ((ret = setup_all_buffers()) < 0)
 982                         return ret;
 983         }
 984
 985         if (for_all_proc_cpus(cpu_is_not_present)) {
 986                 free_all_buffers();
 987                 if ((ret = setup_all_buffers()) < 0)
 988                         return ret;
 989                 if (for_all_proc_cpus(cpu_is_not_present))
 990                         return -ERR_CPU_NOT_PRESENT;
 991         }
 992
 993         /* Saving the scheduling affinity, as it will be modified by get_counters */
 994         if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0)
 995                 return -ERR_CPU_SAVE_SCHED_AFFINITY;
 996
 997         if (!initialized) {
 998                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 999                         goto out;
1000                 gettimeofday(&tv_even, (struct timezone *)NULL);
1001                 is_even = 1;
1002                 initialized = 1;
1003                 ret = 0;
1004                 goto out;
1005         }
1006
1007         if (is_even) {
1008                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1009                         goto out;
1010                 gettimeofday(&tv_odd, (struct timezone *)NULL);
1011                 is_even = 0;
1012                 timersub(&tv_odd, &tv_even, &tv_delta);
1013                 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1014                         goto out;
1015                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1016                         goto out;
1017         } else {
1018                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1019                         goto out;
1020                 gettimeofday(&tv_even, (struct timezone *)NULL);
1021                 is_even = 1;
1022                 timersub(&tv_even, &tv_odd, &tv_delta);
1023                 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1024                         goto out;
1025                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1026                         goto out;
1027         }
1028         ret = 0;
1029 out:
1030         /*
1031          * Let's restore the affinity
1032          * This might fail if the number of CPU changed, but we can't do anything in that case..
1033          */
1034         (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1035         return ret;
1036 }
1037
1038 static int __attribute__((warn_unused_result))
1039 check_dev_msr()
1040 {
1041         struct stat sb;
1042
1043         if (stat("/dev/cpu/0/msr", &sb)) {
1044                 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
1045                 return -ERR_NO_MSR;
1046         }
1047         return 0;
1048 }
1049
1050 static int __attribute__((warn_unused_result))
1051 check_super_user()
1052 {
1053         if (getuid() != 0) {
1054                 ERROR("must be root");
1055                 return -ERR_NOT_ROOT;
1056         }
1057         return 0;
1058 }
1059
1060 /*
1061  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1062  * the Thermal Control Circuit (TCC) activates.
1063  * This is usually equal to tjMax.
1064  *
1065  * Older processors do not have this MSR, so there we guess,
1066  * but also allow conficuration over-ride with "TCCActivationTemp".
1067  *
1068  * Several MSR temperature values are in units of degrees-C
1069  * below this value, including the Digital Thermal Sensor (DTS),
1070  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1071  */
1072 static int __attribute__((warn_unused_result))
1073 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1074 {
1075         unsigned long long msr;
1076         unsigned int target_c_local;
1077
1078         /* tcc_activation_temp is used only for dts or ptm */
1079         if (!(do_dts || do_ptm))
1080                 return 0;
1081
1082         /* this is a per-package concept */
1083         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1084                 return 0;
1085
1086         if (tcc_activation_temp != 0) {
1087                 p->tcc_activation_temp = tcc_activation_temp;
1088                 return 0;
1089         }
1090
1091         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
1092                 goto guess;
1093
1094         target_c_local = (msr >> 16) & 0xFF;
1095
1096         if (!target_c_local)
1097                 goto guess;
1098
1099         p->tcc_activation_temp = target_c_local;
1100
1101         return 0;
1102
1103 guess:
1104         p->tcc_activation_temp = TJMAX_DEFAULT;
1105         WARNING("cpu%d: Guessing tjMax %d C, Please use TCCActivationTemp to specify",
1106                 t->cpu_id, p->tcc_activation_temp);
1107
1108         return 0;
1109 }
1110
1111 /*
1112  * Identify the functionality of the CPU
1113  */
1114 static int __attribute__((warn_unused_result))
1115 probe_cpu()
1116 {
1117         unsigned int eax, ebx, ecx, edx, max_level;
1118         unsigned int fms, family, model;
1119
1120         /* CPUID(0):
1121          * - EAX: Maximum Input Value for Basic CPUID Information
1122          * - EBX: "Genu" (0x756e6547)
1123          * - EDX: "ineI" (0x49656e69)
1124          * - ECX: "ntel" (0x6c65746e)
1125          */
1126         max_level = ebx = ecx = edx = 0;
1127         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1128         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1129                 ERROR("Unsupported CPU");
1130                 return -UNSUPPORTED_CPU;
1131         }
1132
1133         /* CPUID(1):
1134          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1135          *  + 4-7:   Model ID
1136          *  + 8-11:  Family ID
1137          *  + 12-13: Processor type
1138          *  + 16-19: Extended Model ID
1139          *  + 20-27: Extended Family ID
1140          * - EDX: Feature Information:
1141          *  + 5: Support for MSR read/write operations
1142          */
1143         fms = ebx = ecx = edx = 0;
1144         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1145         family = (fms >> 8) & 0xf;
1146         model = (fms >> 4) & 0xf;
1147         if (family == 0xf)
1148                 family += (fms >> 20) & 0xf;
1149         if (family == 6 || family == 0xf)
1150                 model += ((fms >> 16) & 0xf) << 4;
1151         if (!(edx & (1 << 5))) {
1152                 ERROR("CPUID: no MSR");
1153                 return -ERR_NO_MSR;
1154         }
1155
1156         /*
1157          * CPUID(0x80000000):
1158          * - EAX: Maximum Input Value for Extended Function CPUID Information
1159          *
1160          * This allows us to verify if the CPUID(0x80000007) can be called
1161          *
1162          * This check is valid for both Intel and AMD.
1163          */
1164         max_level = ebx = ecx = edx = 0;
1165         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1166         if (max_level < 0x80000007) {
1167                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1168                 return -ERR_NO_INVARIANT_TSC;
1169         }
1170
1171         /*
1172          * CPUID(0x80000007):
1173          * - EDX:
1174          *  + 8: Invariant TSC available if set
1175          *
1176          * This check is valid for both Intel and AMD
1177          */
1178         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1179         if (!(edx & (1 << 8))) {
1180                 ERROR("No invariant TSC");
1181                 return -ERR_NO_INVARIANT_TSC;
1182         }
1183
1184         /*
1185          * CPUID(6):
1186          * - EAX:
1187          *  + 0: Digital temperature sensor is supported if set
1188          *  + 6: Package thermal management is supported if set
1189          * - ECX:
1190          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1191          *  + 3: The processor supports performance-energy bias preference if set.
1192          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1193          *
1194          * This check is valid for both Intel and AMD
1195          */
1196         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1197         do_dts = eax & (1 << 0);
1198         do_ptm = eax & (1 << 6);
1199         if (!(ecx & (1 << 0))) {
1200                 ERROR("No APERF");
1201                 return -ERR_NO_APERF;
1202         }
1203
1204         /*
1205          * Enable or disable C states depending on the model and family
1206          */
1207         if (family == 6) {
1208                 switch (model) {
1209                 /* Atom (partial) */
1210                 case 0x27:
1211                         do_core_cstate = 0;
1212                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1213                         break;
1214                 /* Silvermont */
1215                 case 0x37: /* BYT */
1216                 case 0x4A:
1217                 case 0x4D: /* AVN */
1218                 case 0x5A:
1219                 case 0x5D:
1220                         do_core_cstate = (1 << 1) | (1 << 6);
1221                         do_pkg_cstate = (1 << 6);
1222                         break;
1223                 /* Nehalem */
1224                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1225                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1226                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1227                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1228                         do_core_cstate = (1 << 3) | (1 << 6);
1229                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1230                         break;
1231                 /* Westmere */
1232                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1233                 case 0x2C: /* Westmere EP - Gulftown */
1234                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1235                         do_core_cstate = (1 << 3) | (1 << 6);
1236                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1237                         break;
1238                 /* Sandy Bridge */
1239                 case 0x2A: /* SNB */
1240                 case 0x2D: /* SNB Xeon */
1241                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1242                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1243                         break;
1244                 /* Ivy Bridge */
1245                 case 0x3A: /* IVB */
1246                 case 0x3E: /* IVB Xeon */
1247                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1248                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1249                         break;
1250                 /* Haswell Bridge */
1251                 case 0x3C: /* HSW */
1252                 case 0x3F: /* HSW */
1253                 case 0x46: /* HSW */
1254                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1255                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1256                         break;
1257                 case 0x45: /* HSW */
1258                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1259                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1260                         break;
1261                 /* Broadwel */
1262                 case 0x4F: /* BDW */
1263                 case 0x56: /* BDX-DE */
1264                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1265                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1266                         break;
1267                 case 0x3D: /* BDW */
1268                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1269                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1270                         break;
1271                 default:
1272                         ERROR("Unsupported CPU");
1273                 }
1274                 switch (model) {
1275                 case 0x2A:
1276                 case 0x3A:
1277                 case 0x3C:
1278                 case 0x45:
1279                 case 0x46:
1280                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1281                         break;
1282                 case 0x3F:
1283                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1284                         break;
1285                 case 0x2D:
1286                 case 0x3E:
1287                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1288                         break;
1289                 case 0x37:
1290                 case 0x4D:
1291                         do_rapl = RAPL_PKG | RAPL_CORES;
1292                         break;
1293                 default:
1294                         do_rapl = 0;
1295                 }
1296         } else {
1297                 ERROR("Unsupported CPU");
1298                 return -UNSUPPORTED_CPU;
1299         }
1300
1301         if (do_rapl) {
1302                 unsigned long msr;
1303                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1304                         return 0;
1305
1306                 if (model == 0x37)
1307                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1308                 else
1309                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1310         }
1311
1312         return 0;
1313 }
1314
1315
1316
1317 static int __attribute__((warn_unused_result))
1318 topology_probe()
1319 {
1320         int i;
1321         int ret;
1322         int max_core_id = 0;
1323         int max_package_id = 0;
1324         int max_siblings = 0;
1325         struct cpu_topology {
1326                 int core_id;
1327                 int physical_package_id;
1328         } *cpus;
1329
1330         /* Initialize num_cpus, max_cpu_num */
1331         topo.num_cpus = 0;
1332         topo.max_cpu_num = 0;
1333         ret = for_all_proc_cpus(count_cpus);
1334         if (ret < 0)
1335                 return ret;
1336
1337         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1338
1339         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1340         if (cpus == NULL) {
1341                 ERROR("calloc cpus");
1342                 return -ERR_CALLOC;
1343         }
1344
1345         /*
1346          * Allocate and initialize cpu_present_set
1347          */
1348         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1349         if (cpu_present_set == NULL) {
1350                 free(cpus);
1351                 ERROR("CPU_ALLOC");
1352                 return -ERR_CPU_ALLOC;
1353         }
1354         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1355         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1356         ret = for_all_proc_cpus(mark_cpu_present);
1357         if (ret < 0) {
1358                 free(cpus);
1359                 return ret;
1360         }
1361
1362         /*
1363          * Allocate and initialize cpu_affinity_set
1364          */
1365         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1366         if (cpu_affinity_set == NULL) {
1367                 free(cpus);
1368                 ERROR("CPU_ALLOC");
1369                 return -ERR_CPU_ALLOC;
1370         }
1371         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1372         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1373
1374
1375         /*
1376          * Allocate and initialize cpu_saved_affinity_set
1377          */
1378         cpu_saved_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1379         if (cpu_saved_affinity_set == NULL) {
1380                 free(cpus);
1381                 ERROR("CPU_ALLOC");
1382                 return -ERR_CPU_ALLOC;
1383         }
1384         cpu_saved_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1385         CPU_ZERO_S(cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1386
1387
1388         /*
1389          * For online cpus
1390          * find max_core_id, max_package_id
1391          */
1392         for (i = 0; i <= topo.max_cpu_num; ++i) {
1393                 int siblings;
1394
1395                 if (cpu_is_not_present(i)) {
1396                         WARNING("cpu%d NOT PRESENT", i);
1397                         continue;
1398                 }
1399                 cpus[i].core_id = get_core_id(i);
1400                 if (cpus[i].core_id < 0)
1401                         return cpus[i].core_id;
1402                 if (cpus[i].core_id > max_core_id)
1403                         max_core_id = cpus[i].core_id;
1404
1405                 cpus[i].physical_package_id = get_physical_package_id(i);
1406                 if (cpus[i].physical_package_id < 0)
1407                         return cpus[i].physical_package_id;
1408                 if (cpus[i].physical_package_id > max_package_id)
1409                         max_package_id = cpus[i].physical_package_id;
1410
1411                 siblings = get_num_ht_siblings(i);
1412                 if (siblings < 0)
1413                         return siblings;
1414                 if (siblings > max_siblings)
1415                         max_siblings = siblings;
1416                 DEBUG("cpu %d pkg %d core %d\n",
1417                         i, cpus[i].physical_package_id, cpus[i].core_id);
1418         }
1419         topo.num_cores_per_pkg = max_core_id + 1;
1420         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1421                 max_core_id, topo.num_cores_per_pkg);
1422
1423         topo.num_packages = max_package_id + 1;
1424         DEBUG("max_package_id %d, sizing for %d packages\n",
1425                 max_package_id, topo.num_packages);
1426
1427         topo.num_threads_per_core = max_siblings;
1428         DEBUG("max_siblings %d\n", max_siblings);
1429
1430         free(cpus);
1431         return 0;
1432 }
1433
1434 static int
1435 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1436 {
1437         int i;
1438
1439         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1440                 topo.num_packages, sizeof(struct thread_data));
1441         if (*t == NULL)
1442                 goto error;
1443
1444         for (i = 0; i < topo.num_threads_per_core *
1445                 topo.num_cores_per_pkg * topo.num_packages; i++)
1446                 (*t)[i].cpu_id = -1;
1447
1448         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1449                 sizeof(struct core_data));
1450         if (*c == NULL)
1451                 goto error;
1452
1453         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1454                 (*c)[i].core_id = -1;
1455
1456         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1457         if (*p == NULL)
1458                 goto error;
1459
1460         for (i = 0; i < topo.num_packages; i++)
1461                 (*p)[i].package_id = i;
1462
1463         return 0;
1464 error:
1465         ERROR("calloc counters");
1466         return -ERR_CALLOC;
1467 }
1468 /*
1469  * init_counter()
1470  *
1471  * set cpu_id, core_num, pkg_num
1472  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1473  *
1474  * increment topo.num_cores when 1st core in pkg seen
1475  */
1476 static int
1477 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1478         struct pkg_data *pkg_base, int thread_num, int core_num,
1479         int pkg_num, int cpu_id)
1480 {
1481         int ret;
1482         struct thread_data *t;
1483         struct core_data *c;
1484         struct pkg_data *p;
1485
1486         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1487         c = GET_CORE(core_base, core_num, pkg_num);
1488         p = GET_PKG(pkg_base, pkg_num);
1489
1490         t->cpu_id = cpu_id;
1491         if (thread_num == 0) {
1492                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1493                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1494                         return ret;
1495                 } else if (ret != 0) {
1496                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1497                 }
1498         }
1499
1500         c->core_id = core_num;
1501         p->package_id = pkg_num;
1502
1503         return 0;
1504 }
1505
1506
1507 static int
1508 initialize_counters(int cpu_id)
1509 {
1510         int my_thread_id, my_core_id, my_package_id;
1511         int ret;
1512
1513         my_package_id = get_physical_package_id(cpu_id);
1514         if (my_package_id < 0)
1515                 return my_package_id;
1516         my_core_id = get_core_id(cpu_id);
1517         if (my_core_id < 0)
1518                 return my_core_id;
1519
1520         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1521                 return ret;
1522         } else if (ret != 0) {
1523                 my_thread_id = 0;
1524                 topo.num_cores++;
1525         } else {
1526                 my_thread_id = 1;
1527         }
1528
1529         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1530         if (ret < 0)
1531                 return ret;
1532         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1533         if (ret < 0)
1534                 return ret;
1535         ret = init_counter(DELTA_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1536         if (ret < 0)
1537                 return ret;
1538         return 0;
1539 }
1540
1541 #define DO_OR_GOTO_ERR(something) \
1542 do {                         \
1543         ret = (something);     \
1544         if (ret < 0)         \
1545                 goto err;    \
1546 } while (0)
1547
1548 static int setup_all_buffers(void)
1549 {
1550         int ret;
1551
1552         DO_OR_GOTO_ERR(topology_probe());
1553         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1554         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1555         DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1556         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1557         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1558         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1559
1560         allocated = 1;
1561         return 0;
1562 err:
1563         free_all_buffers();
1564         return ret;
1565 }
1566
1567 static int
1568 turbostat_init(void)
1569 {
1570         int ret;
1571
1572         DO_OR_GOTO_ERR(check_super_user());
1573         DO_OR_GOTO_ERR(probe_cpu());
1574         DO_OR_GOTO_ERR(check_dev_msr());
1575         DO_OR_GOTO_ERR(setup_all_buffers());
1576
1577         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1578
1579         return 0;
1580 err:
1581         free_all_buffers();
1582         return ret;
1583 }
1584
1585 static const char *config_keys[] =
1586 {
1587         "TCCActivationTemp",
1588 };
1589 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1590
1591 static int
1592 turbostat_config(const char *key, const char *value)
1593 {
1594         long unsigned int tmp_val;
1595         char *end;
1596
1597         if (strcasecmp("TCCActivationTemp", key) == 0) {
1598                 tmp_val = strtoul(value, &end, 0);
1599                 if (*end != '\0' || tmp_val > UINT_MAX)
1600                         return -1;
1601                 tcc_activation_temp = (unsigned int) tmp_val;
1602         } else {
1603                 return -1;
1604         }
1605         return 0;
1606 }
1607
1608 void module_register(void);
1609 void module_register(void)
1610 {
1611         plugin_register_init(PLUGIN_NAME, turbostat_init);
1612         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1613 }