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