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 "collectd.h"
  39 #include "common.h"
  40 #include "plugin.h"
  41 #include "utils_time.h"
  42
  43 #include <asm/msr-index.h>
  44 #include <stdarg.h>
  45 #include <stdio.h>
  46 #include <err.h>
  47 #include <unistd.h>
  48 #include <sys/types.h>
  49 #include <sys/wait.h>
  50 #include <sys/stat.h>
  51 #include <sys/resource.h>
  52 #include <fcntl.h>
  53 #include <signal.h>
  54 #include <stdlib.h>
  55 #include <dirent.h>
  56 #include <string.h>
  57 #include <ctype.h>
  58 #include <sched.h>
  59 #include <cpuid.h>
  60 #include <sys/capability.h>
  61
  62 #define PLUGIN_NAME "turbostat"
  63
  64 /*
  65  * This tool uses the Model-Specific Registers (MSRs) present on Intel processors.
  66  * The general description each of these registers, depending on the architecture,
  67  * can be found in the Intel® 64 and IA-32 Architectures Software Developer Manual,
  68  * Volume 3 Chapter 35.
  69  */
  70
  71 /*
  72  * If set, aperf_mperf_unstable disables a/mperf based stats.
  73  * This includes: C0 & C1 states, frequency
  74  *
  75  * This value is automatically set if mperf or aperf go backward
  76  */
  77 static _Bool aperf_mperf_unstable;
  78
  79 /*
  80  * Bitmask of the list of core C states supported by the processor.
  81  * Currently supported C-states (by this plugin): 3, 6, 7
  82  */
  83 static unsigned int do_core_cstate;
  84 static unsigned int config_core_cstate;
  85 static _Bool apply_config_core_cstate;
  86
  87 /*
  88  * Bitmask of the list of pacages C states supported by the processor.
  89  * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
  90  */
  91 static unsigned int do_pkg_cstate;
  92 static unsigned int config_pkg_cstate;
  93 static _Bool apply_config_pkg_cstate;
  94
  95 /*
  96  * Boolean indicating if the processor supports 'I/O System-Management Interrupt counter'
  97  */
  98 static _Bool do_smi;
  99 static _Bool config_smi;
 100 static _Bool apply_config_smi;
 101
 102 /*
 103  * Boolean indicating if the processor supports 'Digital temperature sensor'
 104  * This feature enables the monitoring of the temperature of each core
 105  *
 106  * This feature has two limitations:
 107  *  - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
 108  *  - Temperatures above the tcc_activation_temp are not recorded
 109  */
 110 static _Bool do_dts;
 111 static _Bool config_dts;
 112 static _Bool apply_config_dts;
 113
 114 /*
 115  * Boolean indicating if the processor supports 'Package thermal management'
 116  * This feature allows the monitoring of the temperature of each package
 117  *
 118  * This feature has two limitations:
 119  *  - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
 120  *  - Temperatures above the tcc_activation_temp are not recorded
 121  */
 122 static _Bool do_ptm;
 123 static _Bool config_ptm;
 124 static _Bool apply_config_ptm;
 125
 126 /*
 127  * Thermal Control Circuit Activation Temperature as configured by the user.
 128  * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
 129  * and should only be used if the automated detection fails.
 130  */
 131 static unsigned int tcc_activation_temp;
 132
 133 static unsigned int do_rapl;
 134 static unsigned int config_rapl;
 135 static _Bool apply_config_rapl;
 136 static double rapl_energy_units;
 137
 138 #define RAPL_PKG                (1 << 0)
 139                                         /* 0x610 MSR_PKG_POWER_LIMIT */
 140                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
 141 #define RAPL_DRAM               (1 << 1)
 142                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
 143                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
 144                                         /* 0x61c MSR_DRAM_POWER_INFO */
 145 #define RAPL_CORES              (1 << 2)
 146                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 147                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 148
 149 #define RAPL_GFX                (1 << 3)
 150                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 151                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 152                                         /* 0x642 MSR_PP1_POLICY */
 153 #define TJMAX_DEFAULT   100
 154
 155 static cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
 156 static size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
 157
 158 static struct thread_data {
 159         unsigned long long tsc;
 160         unsigned long long aperf;
 161         unsigned long long mperf;
 162         unsigned long long c1;
 163         unsigned int smi_count;
 164         unsigned int cpu_id;
 165         unsigned int flags;
 166 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 167 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 168 } *thread_delta, *thread_even, *thread_odd;
 169
 170 static struct core_data {
 171         unsigned long long c3;
 172         unsigned long long c6;
 173         unsigned long long c7;
 174         unsigned int core_temp_c;
 175         unsigned int core_id;
 176 } *core_delta, *core_even, *core_odd;
 177
 178 static struct pkg_data {
 179         unsigned long long pc2;
 180         unsigned long long pc3;
 181         unsigned long long pc6;
 182         unsigned long long pc7;
 183         unsigned long long pc8;
 184         unsigned long long pc9;
 185         unsigned long long pc10;
 186         unsigned int package_id;
 187         uint32_t energy_pkg;    /* MSR_PKG_ENERGY_STATUS */
 188         uint32_t energy_dram;   /* MSR_DRAM_ENERGY_STATUS */
 189         uint32_t energy_cores;  /* MSR_PP0_ENERGY_STATUS */
 190         uint32_t energy_gfx;    /* MSR_PP1_ENERGY_STATUS */
 191         unsigned int tcc_activation_temp;
 192         unsigned int pkg_temp_c;
 193 } *package_delta, *package_even, *package_odd;
 194
 195 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
 196 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 197 #define EVEN_COUNTERS thread_even, core_even, package_even
 198 static _Bool is_even = 1;
 199
 200 static _Bool allocated = 0;
 201 static _Bool initialized = 0;
 202
 203 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 204         (thread_base + \
 205                 (pkg_no) * topology.num_cores * topology.num_threads + \
 206                 (core_no) * topology.num_threads + \
 207                 (thread_no))
 208 #define GET_CORE(core_base, core_no, pkg_no) \
 209         (core_base + \
 210                 (pkg_no) * topology.num_cores + \
 211                 (core_no))
 212 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 213
 214 struct cpu_topology {
 215         unsigned int package_id;
 216         unsigned int core_id;
 217         _Bool first_core_in_package;
 218         _Bool first_thread_in_core;
 219 };
 220
 221 static struct topology {
 222         unsigned int max_cpu_id;
 223         unsigned int num_packages;
 224         unsigned int num_cores;
 225         unsigned int num_threads;
 226         struct cpu_topology *cpus;
 227 } topology;
 228
 229 static cdtime_t time_even, time_odd, time_delta;
 230
 231 static const char *config_keys[] =
 232 {
 233         "CoreCstates",
 234         "PackageCstates",
 235         "SystemManagementInterrupt",
 236         "DigitalTemperatureSensor",
 237         "PackageThermalManagement",
 238         "TCCActivationTemp",
 239         "RunningAveragePowerLimit",
 240 };
 241 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
 242
 243 /*****************************
 244  *  MSR Manipulation helpers *
 245  *****************************/
 246
 247 /*
 248  * Open a MSR device for reading
 249  * Can change the scheduling affinity of the current process if multiple_read is 1
 250  */
 251 static int __attribute__((warn_unused_result))
 252 open_msr(unsigned int cpu, _Bool multiple_read)
 253 {
 254         char pathname[32];
 255         int fd;
 256
 257         /*
 258          * If we need to do multiple read, let's migrate to the CPU
 259          * Otherwise, we would lose time calling functions on another CPU
 260          *
 261          * If we are not yet initialized (cpu_affinity_setsize = 0),
 262          * we need to skip this optimisation.
 263          */
 264         if (multiple_read && cpu_affinity_setsize) {
 265                 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 266                 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 267                 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
 268                         ERROR("Turbostat plugin: Could not migrate to CPU %d", cpu);
 269                         return -1;
 270                 }
 271         }
 272
 273         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 274         fd = open(pathname, O_RDONLY);
 275         if (fd < 0) {
 276                 ERROR("Turbostat plugin: failed to open %s", pathname);
 277                 return -1;
 278         }
 279         return fd;
 280 }
 281
 282 /*
 283  * Read a single MSR from an open file descriptor
 284  */
 285 static int __attribute__((warn_unused_result))
 286 read_msr(int fd, off_t offset, unsigned long long *msr)
 287 {
 288         ssize_t retval;
 289
 290         retval = pread(fd, msr, sizeof *msr, offset);
 291
 292         if (retval != sizeof *msr) {
 293                 ERROR("Turbostat plugin: MSR offset 0x%llx read failed",
 294                       (unsigned long long)offset);
 295                 return -1;
 296         }
 297         return 0;
 298 }
 299
 300 /*
 301  * Open a MSR device for reading, read the value asked for and close it.
 302  * This call will not affect the scheduling affinity of this thread.
 303  */
 304 static ssize_t __attribute__((warn_unused_result))
 305 get_msr(unsigned int cpu, off_t offset, unsigned long long *msr)
 306 {
 307         ssize_t retval;
 308         int fd;
 309
 310         fd = open_msr(cpu, 0);
 311         if (fd < 0)
 312                 return fd;
 313         retval = read_msr(fd, offset, msr);
 314         close(fd);
 315         return retval;
 316 }
 317
 318
 319 /********************************
 320  * Raw data acquisition (1 CPU) *
 321  ********************************/
 322
 323 /*
 324  * Read every data avalaible for a single CPU
 325  *
 326  * Core data is shared for all threads in one core: extracted only for the first thread
 327  * Package data is shared for all core in one package: extracted only for the first thread of the first core
 328  *
 329  * Side effect: migrates to the targeted CPU
 330  */
 331 static int __attribute__((warn_unused_result))
 332 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 333 {
 334         unsigned int cpu = t->cpu_id;
 335         unsigned long long msr;
 336         int msr_fd;
 337         int retval = 0;
 338
 339         msr_fd = open_msr(cpu, 1);
 340         if (msr_fd < 0)
 341                 return msr_fd;
 342
 343 #define READ_MSR(msr, dst)                                              \
 344 do {                                                                    \
 345         if (read_msr(msr_fd, msr, dst)) {                               \
 346                 ERROR("Turbostat plugin: Unable to read " #msr);        \
 347                 retval = -1;                                            \
 348                 goto out;                                               \
 349         }                                                               \
 350 } while (0)
 351
 352         READ_MSR(MSR_IA32_TSC, &t->tsc);
 353
 354         READ_MSR(MSR_IA32_APERF, &t->aperf);
 355         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 356
 357         if (do_smi) {
 358                 READ_MSR(MSR_SMI_COUNT, &msr);
 359                 t->smi_count = msr & 0xFFFFFFFF;
 360         }
 361
 362         /* collect core counters only for 1st thread in core */
 363         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 364                 retval = 0;
 365                 goto out;
 366         }
 367
 368         if (do_core_cstate & (1 << 3))
 369                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 370         if (do_core_cstate & (1 << 6))
 371                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 372         if (do_core_cstate & (1 << 7))
 373                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 374
 375         if (do_dts) {
 376                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 377                 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 378         }
 379
 380         /* collect package counters only for 1st core in package */
 381         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 382                 retval = 0;
 383                 goto out;
 384         }
 385
 386         if (do_pkg_cstate & (1 << 2))
 387                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 388         if (do_pkg_cstate & (1 << 3))
 389                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 390         if (do_pkg_cstate & (1 << 6))
 391                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 392         if (do_pkg_cstate & (1 << 7))
 393                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 394         if (do_pkg_cstate & (1 << 8))
 395                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 396         if (do_pkg_cstate & (1 << 9))
 397                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 398         if (do_pkg_cstate & (1 << 10))
 399                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 400
 401         if (do_rapl & RAPL_PKG) {
 402                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 403                 p->energy_pkg = msr & 0xFFFFFFFF;
 404         }
 405         if (do_rapl & RAPL_CORES) {
 406                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 407                 p->energy_cores = msr & 0xFFFFFFFF;
 408         }
 409         if (do_rapl & RAPL_DRAM) {
 410                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 411                 p->energy_dram = msr & 0xFFFFFFFF;
 412         }
 413         if (do_rapl & RAPL_GFX) {
 414                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 415                 p->energy_gfx = msr & 0xFFFFFFFF;
 416         }
 417         if (do_ptm) {
 418                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 419                 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 420         }
 421
 422 out:
 423         close(msr_fd);
 424         return retval;
 425 }
 426
 427
 428 /**********************************
 429  * Evaluating the changes (1 CPU) *
 430  **********************************/
 431
 432 /*
 433  * Extract the evolution old->new in delta at a package level
 434  * (some are not new-delta, e.g. temperature)
 435  */
 436 static inline void
 437 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
 438 {
 439         delta->pc2 = new->pc2 - old->pc2;
 440         delta->pc3 = new->pc3 - old->pc3;
 441         delta->pc6 = new->pc6 - old->pc6;
 442         delta->pc7 = new->pc7 - old->pc7;
 443         delta->pc8 = new->pc8 - old->pc8;
 444         delta->pc9 = new->pc9 - old->pc9;
 445         delta->pc10 = new->pc10 - old->pc10;
 446         delta->pkg_temp_c = new->pkg_temp_c;
 447
 448         delta->energy_pkg = new->energy_pkg - old->energy_pkg;
 449         delta->energy_cores = new->energy_cores - old->energy_cores;
 450         delta->energy_gfx = new->energy_gfx - old->energy_gfx;
 451         delta->energy_dram = new->energy_dram - old->energy_dram;
 452 }
 453
 454 /*
 455  * Extract the evolution old->new in delta at a core level
 456  * (some are not new-delta, e.g. temperature)
 457  */
 458 static inline void
 459 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
 460 {
 461         delta->c3 = new->c3 - old->c3;
 462         delta->c6 = new->c6 - old->c6;
 463         delta->c7 = new->c7 - old->c7;
 464         delta->core_temp_c = new->core_temp_c;
 465 }
 466
 467 /*
 468  * Extract the evolution old->new in delta at a package level
 469  * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
 470  */
 471 static inline int __attribute__((warn_unused_result))
 472 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
 473         const struct core_data *core_delta)
 474 {
 475         delta->tsc = new->tsc - old->tsc;
 476
 477         /* check for TSC < 1 Mcycles over interval */
 478         if (delta->tsc < (1000 * 1000)) {
 479                 WARNING("Turbostat plugin: Insanely slow TSC rate, TSC stops "
 480                         "in idle? You can disable all c-states by booting with"
 481                         " 'idle=poll' or just the deep ones with"
 482                         " 'processor.max_cstate=1'");
 483                 return -1;
 484         }
 485
 486         delta->c1 = new->c1 - old->c1;
 487
 488         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 489                 delta->aperf = new->aperf - old->aperf;
 490                 delta->mperf = new->mperf - old->mperf;
 491         } else {
 492                 if (!aperf_mperf_unstable) {
 493                         WARNING("Turbostat plugin: APERF or MPERF went "
 494                                 "backwards. Frequency results do not cover "
 495                                 "the entire interval. Fix this by running "
 496                                 "Linux-2.6.30 or later.");
 497
 498                         aperf_mperf_unstable = 1;
 499                 }
 500         }
 501
 502         /*
 503          * As counter collection is not atomic,
 504          * it is possible for mperf's non-halted cycles + idle states
 505          * to exceed TSC's all cycles: show c1 = 0% in that case.
 506          */
 507         if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
 508                 delta->c1 = 0;
 509         else {
 510                 /* normal case, derive c1 */
 511                 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
 512                         - core_delta->c6 - core_delta->c7;
 513         }
 514
 515         if (delta->mperf == 0) {
 516                 WARNING("Turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
 517                 delta->mperf = 1;       /* divide by 0 protection */
 518         }
 519
 520         if (do_smi)
 521                 delta->smi_count = new->smi_count - old->smi_count;
 522
 523         return 0;
 524 }
 525
 526 /**********************************
 527  * Submitting the results (1 CPU) *
 528  **********************************/
 529
 530 /*
 531  * Submit one gauge value
 532  */
 533 static void
 534 turbostat_submit (const char *plugin_instance,
 535         const char *type, const char *type_instance,
 536         gauge_t value)
 537 {
 538         value_list_t vl = VALUE_LIST_INIT;
 539         value_t v;
 540
 541         v.gauge = value;
 542         vl.values = &v;
 543         vl.values_len = 1;
 544         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 545         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 546         if (plugin_instance != NULL)
 547                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 548         sstrncpy (vl.type, type, sizeof (vl.type));
 549         if (type_instance != NULL)
 550                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 551
 552         plugin_dispatch_values (&vl);
 553 }
 554
 555 /*
 556  * Submit every data for a single CPU
 557  *
 558  * Core data is shared for all threads in one core: submitted only for the first thread
 559  * Package data is shared for all core in one package: submitted only for the first thread of the first core
 560  */
 561 static int
 562 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 563 {
 564         char name[DATA_MAX_NAME_LEN];
 565         double interval_float;
 566
 567         interval_float = CDTIME_T_TO_DOUBLE(time_delta);
 568
 569         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 570
 571         if (!aperf_mperf_unstable)
 572                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 573         if (!aperf_mperf_unstable)
 574                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 575
 576         turbostat_submit("Average", "frequency", name, 1.0 / 1000000 * t->aperf / interval_float);
 577
 578         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 579                 turbostat_submit("Buzy", "frequency", name, 1.0 * t->tsc / 1000000 * t->aperf / t->mperf / interval_float);
 580
 581         /* Sanity check (should stay stable) */
 582         turbostat_submit("TSC", "gauge", name, 1.0 * t->tsc / 1000000 / interval_float);
 583
 584         /* SMI */
 585         if (do_smi)
 586                 turbostat_submit(NULL, "current", name, t->smi_count);
 587
 588         /* submit per-core data only for 1st thread in core */
 589         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 590                 goto done;
 591
 592         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 593
 594         if (do_core_cstate & (1 << 3))
 595                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 596         if (do_core_cstate & (1 << 6))
 597                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 598         if (do_core_cstate & (1 << 7))
 599                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 600
 601         if (do_dts)
 602                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 603
 604         /* submit per-package data only for 1st core in package */
 605         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 606                 goto done;
 607
 608         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 609
 610         if (do_ptm)
 611                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 612
 613         if (do_pkg_cstate & (1 << 2))
 614                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 615         if (do_pkg_cstate & (1 << 3))
 616                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 617         if (do_pkg_cstate & (1 << 6))
 618                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 619         if (do_pkg_cstate & (1 << 7))
 620                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 621         if (do_pkg_cstate & (1 << 8))
 622                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 623         if (do_pkg_cstate & (1 << 9))
 624                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 625         if (do_pkg_cstate & (1 << 10))
 626                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 627
 628         if (do_rapl) {
 629                 if (do_rapl & RAPL_PKG)
 630                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 631                 if (do_rapl & RAPL_CORES)
 632                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 633                 if (do_rapl & RAPL_GFX)
 634                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 635                 if (do_rapl & RAPL_DRAM)
 636                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 637         }
 638 done:
 639         return 0;
 640 }
 641
 642
 643 /**********************************
 644  * Looping function over all CPUs *
 645  **********************************/
 646
 647 /*
 648  * Check if a given cpu id is in our compiled list of existing CPUs
 649  */
 650 static int
 651 cpu_is_not_present(unsigned int cpu)
 652 {
 653         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 654 }
 655
 656 /*
 657  * Loop on all CPUs in topological order
 658  *
 659  * Skip non-present cpus
 660  * Return the error code at the first error or 0
 661  */
 662 static int __attribute__((warn_unused_result))
 663 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 664         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 665 {
 666         int retval;
 667         unsigned int pkg_no, core_no, thread_no;
 668
 669         for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
 670                 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
 671                         for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
 672                                 struct thread_data *t;
 673                                 struct core_data *c;
 674                                 struct pkg_data *p;
 675
 676                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 677
 678                                 if (cpu_is_not_present(t->cpu_id))
 679                                         continue;
 680
 681                                 c = GET_CORE(core_base, core_no, pkg_no);
 682                                 p = GET_PKG(pkg_base, pkg_no);
 683
 684                                 retval = func(t, c, p);
 685                                 if (retval)
 686                                         return retval;
 687                         }
 688                 }
 689         }
 690         return 0;
 691 }
 692
 693 /*
 694  * Dedicated loop: Extract every data evolution for all CPU
 695  *
 696  * Skip non-present cpus
 697  * Return the error code at the first error or 0
 698  *
 699  * Core data is shared for all threads in one core: extracted only for the first thread
 700  * Package data is shared for all core in one package: extracted only for the first thread of the first core
 701  */
 702 static int __attribute__((warn_unused_result))
 703 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,
 704                    const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
 705 {
 706         int retval;
 707         unsigned int pkg_no, core_no, thread_no;
 708
 709         for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
 710                 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
 711                         for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
 712                                 struct thread_data *t_delta;
 713                                 const struct thread_data *t_old, *t_new;
 714                                 struct core_data *c_delta;
 715
 716                                 /* Get correct pointers for threads */
 717                                 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
 718                                 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
 719                                 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
 720
 721                                 /* Skip threads that disappeared */
 722                                 if (cpu_is_not_present(t_delta->cpu_id))
 723                                         continue;
 724
 725                                 /* c_delta is always required for delta_thread */
 726                                 c_delta = GET_CORE(core_delta, core_no, pkg_no);
 727
 728                                 /* calculate core delta only for 1st thread in core */
 729                                 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
 730                                         const struct core_data *c_old, *c_new;
 731
 732                                         c_new = GET_CORE(core_new_base, core_no, pkg_no);
 733                                         c_old = GET_CORE(core_old_base, core_no, pkg_no);
 734
 735                                         delta_core(c_delta, c_new, c_old);
 736                                 }
 737
 738                                 /* Always calculate thread delta */
 739                                 retval = delta_thread(t_delta, t_new, t_old, c_delta);
 740                                 if (retval)
 741                                         return retval;
 742
 743                                 /* calculate package delta only for 1st core in package */
 744                                 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
 745                                         struct pkg_data *p_delta;
 746                                         const struct pkg_data *p_old, *p_new;
 747
 748                                         p_delta = GET_PKG(package_delta, pkg_no);
 749                                         p_new = GET_PKG(pkg_new_base, pkg_no);
 750                                         p_old = GET_PKG(pkg_old_base, pkg_no);
 751
 752                                         delta_package(p_delta, p_new, p_old);
 753                                 }
 754                         }
 755                 }
 756         }
 757         return 0;
 758 }
 759
 760
 761 /***************
 762  * CPU Probing *
 763  ***************/
 764
 765 /*
 766  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
 767  * the Thermal Control Circuit (TCC) activates.
 768  * This is usually equal to tjMax.
 769  *
 770  * Older processors do not have this MSR, so there we guess,
 771  * but also allow conficuration over-ride with "TCCActivationTemp".
 772  *
 773  * Several MSR temperature values are in units of degrees-C
 774  * below this value, including the Digital Thermal Sensor (DTS),
 775  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
 776  */
 777 static int __attribute__((warn_unused_result))
 778 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 779 {
 780         unsigned long long msr;
 781         unsigned int target_c_local;
 782
 783         /* tcc_activation_temp is used only for dts or ptm */
 784         if (!(do_dts || do_ptm))
 785                 return 0;
 786
 787         /* this is a per-package concept */
 788         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 789                 return 0;
 790
 791         if (tcc_activation_temp != 0) {
 792                 p->tcc_activation_temp = tcc_activation_temp;
 793                 return 0;
 794         }
 795
 796         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
 797                 goto guess;
 798
 799         target_c_local = (msr >> 16) & 0xFF;
 800
 801         if (!target_c_local)
 802                 goto guess;
 803
 804         p->tcc_activation_temp = target_c_local;
 805
 806         return 0;
 807
 808 guess:
 809         p->tcc_activation_temp = TJMAX_DEFAULT;
 810         WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
 811                 " Please use TCCActivationTemp to specify it.",
 812                 t->cpu_id, p->tcc_activation_temp);
 813
 814         return 0;
 815 }
 816
 817 /*
 818  * Identify the functionality of the CPU
 819  */
 820 static int __attribute__((warn_unused_result))
 821 probe_cpu()
 822 {
 823         unsigned int eax, ebx, ecx, edx, max_level;
 824         unsigned int fms, family, model;
 825
 826         /* CPUID(0):
 827          * - EAX: Maximum Input Value for Basic CPUID Information
 828          * - EBX: "Genu" (0x756e6547)
 829          * - EDX: "ineI" (0x49656e69)
 830          * - ECX: "ntel" (0x6c65746e)
 831          */
 832         max_level = ebx = ecx = edx = 0;
 833         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
 834         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
 835                 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
 836                 return -1;
 837         }
 838
 839         /* CPUID(1):
 840          * - EAX: Version Information: Type, Family, Model, and Stepping ID
 841          *  + 4-7:   Model ID
 842          *  + 8-11:  Family ID
 843          *  + 12-13: Processor type
 844          *  + 16-19: Extended Model ID
 845          *  + 20-27: Extended Family ID
 846          * - EDX: Feature Information:
 847          *  + 5: Support for MSR read/write operations
 848          */
 849         fms = ebx = ecx = edx = 0;
 850         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
 851         family = (fms >> 8) & 0xf;
 852         model = (fms >> 4) & 0xf;
 853         if (family == 0xf)
 854                 family += (fms >> 20) & 0xf;
 855         if (family == 6 || family == 0xf)
 856                 model += ((fms >> 16) & 0xf) << 4;
 857         if (!(edx & (1 << 5))) {
 858                 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
 859                 return -1;
 860         }
 861
 862         /*
 863          * CPUID(6):
 864          * - EAX:
 865          *  + 0: Digital temperature sensor is supported if set
 866          *  + 6: Package thermal management is supported if set
 867          * - ECX:
 868          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
 869          *  + 3: The processor supports performance-energy bias preference if set.
 870          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
 871          *
 872          * This check is valid for both Intel and AMD
 873          */
 874         eax = ebx = ecx = edx = 0;
 875         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
 876         do_dts = eax & (1 << 0);
 877         do_ptm = eax & (1 << 6);
 878         if (!(ecx & (1 << 0))) {
 879                 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
 880                 return -1;
 881         }
 882
 883         /*
 884          * Enable or disable C states depending on the model and family
 885          */
 886         if (family == 6) {
 887                 switch (model) {
 888                 /* Atom (partial) */
 889                 case 0x27:
 890                         do_smi = 0;
 891                         do_core_cstate = 0;
 892                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
 893                         break;
 894                 /* Silvermont */
 895                 case 0x37: /* BYT */
 896                 case 0x4D: /* AVN */
 897                         do_smi = 1;
 898                         do_core_cstate = (1 << 1) | (1 << 6);
 899                         do_pkg_cstate = (1 << 6);
 900                         break;
 901                 /* Nehalem */
 902                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
 903                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
 904                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
 905                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
 906                         do_smi = 1;
 907                         do_core_cstate = (1 << 3) | (1 << 6);
 908                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 909                         break;
 910                 /* Westmere */
 911                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
 912                 case 0x2C: /* Westmere EP - Gulftown */
 913                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
 914                         do_smi = 1;
 915                         do_core_cstate = (1 << 3) | (1 << 6);
 916                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 917                         break;
 918                 /* Sandy Bridge */
 919                 case 0x2A: /* SNB */
 920                 case 0x2D: /* SNB Xeon */
 921                         do_smi = 1;
 922                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 923                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
 924                         break;
 925                 /* Ivy Bridge */
 926                 case 0x3A: /* IVB */
 927                 case 0x3E: /* IVB Xeon */
 928                         do_smi = 1;
 929                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 930                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
 931                         break;
 932                 /* Haswell Bridge */
 933                 case 0x3C: /* HSW */
 934                 case 0x3F: /* HSW */
 935                 case 0x46: /* HSW */
 936                         do_smi = 1;
 937                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 938                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
 939                         break;
 940                 case 0x45: /* HSW */
 941                         do_smi = 1;
 942                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 943                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
 944                         break;
 945                 /* Broadwel */
 946                 case 0x4F: /* BDW */
 947                 case 0x56: /* BDX-DE */
 948                         do_smi = 1;
 949                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 950                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
 951                         break;
 952                 case 0x3D: /* BDW */
 953                         do_smi = 1;
 954                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
 955                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
 956                         break;
 957                 default:
 958                         do_smi = 0;
 959                         do_core_cstate = 0;
 960                         do_pkg_cstate = 0;
 961                         break;
 962                 }
 963                 switch (model) {
 964                 case 0x2A: /* SNB */
 965                 case 0x3A: /* IVB */
 966                 case 0x3C: /* HSW */
 967                 case 0x45: /* HSW */
 968                 case 0x46: /* HSW */
 969                 case 0x3D: /* BDW */
 970                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_GFX;
 971                         break;
 972                 case 0x3F: /* HSX */
 973                 case 0x4F: /* BDX */
 974                 case 0x56: /* BDX-DE */
 975                         do_rapl = RAPL_PKG | RAPL_DRAM ;
 976                         break;
 977                 case 0x2D: /* SNB Xeon */
 978                 case 0x3E: /* IVB Xeon */
 979                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_DRAM;
 980                         break;
 981                 case 0x37: /* BYT */
 982                 case 0x4D: /* AVN */
 983                         do_rapl = RAPL_PKG | RAPL_CORES;
 984                         break;
 985                 default:
 986                         do_rapl = 0;
 987                 }
 988         } else {
 989                 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
 990                       "model: %#x)", family, model);
 991                 return -1;
 992         }
 993
 994         /* Override detected values with configuration */
 995         if (apply_config_core_cstate)
 996                 do_core_cstate = config_core_cstate;
 997         if (apply_config_pkg_cstate)
 998                 do_pkg_cstate = config_pkg_cstate;
 999         if (apply_config_smi)
1000                 do_smi = config_smi;
1001         if (apply_config_dts)
1002                 do_dts = config_dts;
1003         if (apply_config_ptm)
1004                 do_ptm = config_ptm;
1005         if (apply_config_rapl)
1006                 do_rapl = config_rapl;
1007
1008         if (do_rapl) {
1009                 unsigned long long msr;
1010                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1011                         return 0;
1012
1013                 if (model == 0x37)
1014                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1015                 else
1016                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1017         }
1018
1019         return 0;
1020 }
1021
1022
1023 /********************
1024  * Topology Probing *
1025  ********************/
1026
1027 /*
1028  * Read a single int from a file.
1029  */
1030 static int __attribute__ ((format(printf,1,2)))
1031 parse_int_file(const char *fmt, ...)
1032 {
1033         va_list args;
1034         char path[PATH_MAX];
1035         FILE *filep;
1036         int len, value;
1037
1038         va_start(args, fmt);
1039         len = vsnprintf(path, sizeof(path), fmt, args);
1040         va_end(args);
1041         if (len < 0 || len >= PATH_MAX) {
1042                 ERROR("Turbostat plugin: path truncated: '%s'", path);
1043                 return -1;
1044         }
1045
1046         filep = fopen(path, "r");
1047         if (!filep) {
1048                 ERROR("Turbostat plugin: Failed to open '%s'", path);
1049                 return -1;
1050         }
1051         if (fscanf(filep, "%d", &value) != 1) {
1052                 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1053                 return -1;
1054         }
1055         fclose(filep);
1056         return value;
1057 }
1058
1059 static int
1060 get_threads_on_core(unsigned int cpu)
1061 {
1062         char path[80];
1063         FILE *filep;
1064         int sib1, sib2;
1065         int matches;
1066         char character;
1067
1068         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1069         filep = fopen(path, "r");
1070         if (!filep) {
1071                 ERROR("Turbostat plugin: Failed to open '%s'", path);
1072                 return -1;
1073         }
1074         /*
1075          * file format:
1076          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1077          * otherwinse 1 sibling (self).
1078          */
1079         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1080
1081         fclose(filep);
1082
1083         if (matches == 3)
1084                 return 2;
1085         else
1086                 return 1;
1087 }
1088
1089 /*
1090  * run func(cpu) on every cpu in /proc/stat
1091  * return max_cpu number
1092  */
1093 static int __attribute__((warn_unused_result))
1094 for_all_proc_cpus(int (func)(unsigned int))
1095 {
1096         FILE *fp;
1097         unsigned int cpu_num;
1098         int retval;
1099
1100         fp = fopen("/proc/stat", "r");
1101         if (!fp) {
1102                 ERROR("Turbostat plugin: Failed to open /proc/stat");
1103                 return -1;
1104         }
1105
1106         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1107         if (retval != 0) {
1108                 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1109                 fclose(fp);
1110                 return -1;
1111         }
1112
1113         while (1) {
1114                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1115                 if (retval != 1)
1116                         break;
1117
1118                 retval = func(cpu_num);
1119                 if (retval) {
1120                         fclose(fp);
1121                         return(retval);
1122                 }
1123         }
1124         fclose(fp);
1125         return 0;
1126 }
1127
1128 /*
1129  * Update the stored topology.max_cpu_id
1130  */
1131 static int
1132 update_max_cpu_id(unsigned int cpu)
1133 {
1134         if (topology.max_cpu_id < cpu)
1135                 topology.max_cpu_id = cpu;
1136         return 0;
1137 }
1138
1139 static int
1140 mark_cpu_present(unsigned int cpu)
1141 {
1142         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1143         return 0;
1144 }
1145
1146 static int __attribute__((warn_unused_result))
1147 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1148         *set = CPU_ALLOC(topology.max_cpu_id  + 1);
1149         if (*set == NULL) {
1150                 ERROR("Turbostat plugin: Unable to allocate CPU state");
1151                 return -1;
1152         }
1153         *size = CPU_ALLOC_SIZE(topology.max_cpu_id  + 1);
1154         CPU_ZERO_S(*size, *set);
1155         return 0;
1156 }
1157
1158 /*
1159  * Build a local representation of the cpu distribution
1160  */
1161 static int __attribute__((warn_unused_result))
1162 topology_probe()
1163 {
1164         unsigned int i;
1165         int ret;
1166         unsigned int max_package_id, max_core_id, max_thread_id;
1167         max_package_id = max_core_id = max_thread_id = 0;
1168
1169         /* Clean topology */
1170         free(topology.cpus);
1171         memset(&topology, 0, sizeof(topology));
1172
1173         ret = for_all_proc_cpus(update_max_cpu_id);
1174         if (ret != 0)
1175                 goto err;
1176
1177         topology.cpus = calloc(1, (topology.max_cpu_id  + 1) * sizeof(struct cpu_topology));
1178         if (topology.cpus == NULL) {
1179                 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1180                 return -1;
1181         }
1182
1183         ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1184         if (ret != 0)
1185                 goto err;
1186         ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1187         if (ret != 0)
1188                 goto err;
1189         ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1190         if (ret != 0)
1191                 goto err;
1192
1193         ret = for_all_proc_cpus(mark_cpu_present);
1194         if (ret != 0)
1195                 goto err;
1196
1197         /*
1198          * For online cpus
1199          * find max_core_id, max_package_id
1200          */
1201         for (i = 0; i <= topology.max_cpu_id; ++i) {
1202                 unsigned int num_threads;
1203                 struct cpu_topology *cpu = &topology.cpus[i];
1204
1205                 if (cpu_is_not_present(i)) {
1206                         WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1207                         continue;
1208                 }
1209
1210                 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1211                 if (ret < 0)
1212                         goto err;
1213                 else
1214                         cpu->package_id = (unsigned int) ret;
1215                 if (cpu->package_id > max_package_id)
1216                         max_package_id = cpu->package_id;
1217
1218                 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1219                 if (ret < 0)
1220                         goto err;
1221                 else
1222                         cpu->core_id = (unsigned int) ret;
1223                 if (cpu->core_id > max_core_id)
1224                         max_core_id = cpu->core_id;
1225                 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1226                 if (ret < 0)
1227                         goto err;
1228                 else if ((unsigned int) ret == i)
1229                         cpu->first_core_in_package = 1;
1230
1231                 ret = get_threads_on_core(i);
1232                 if (ret < 0)
1233                         goto err;
1234                 else
1235                         num_threads = (unsigned int) ret;
1236                 if (num_threads > max_thread_id)
1237                         max_thread_id = num_threads;
1238                 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1239                 if (ret < 0)
1240                         goto err;
1241                 else if ((unsigned int) ret == i)
1242                         cpu->first_thread_in_core = 1;
1243
1244                 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1245                         i, cpu->package_id, cpu->core_id);
1246         }
1247         /* Num is max + 1 (need to count 0) */
1248         topology.num_packages = max_package_id + 1;
1249         topology.num_cores = max_core_id + 1;
1250         topology.num_threads = max_thread_id + 1;
1251
1252         return 0;
1253 err:
1254         free(topology.cpus);
1255         return ret;
1256 }
1257
1258
1259 /************************
1260  * Main alloc/init/free *
1261  ************************/
1262
1263 static int
1264 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1265 {
1266         unsigned int i;
1267         unsigned int total_threads, total_cores;
1268
1269         total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1270         *threads = calloc(total_threads, sizeof(struct thread_data));
1271         if (*threads == NULL)
1272                 goto err;
1273
1274         total_cores = topology.num_cores * topology.num_packages;
1275         *cores = calloc(total_cores, sizeof(struct core_data));
1276         if (*cores == NULL)
1277                 goto err_clean_threads;
1278
1279         *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1280         if (*packages == NULL)
1281                 goto err_clean_cores;
1282
1283         for (i = 0; i < topology.num_packages; i++)
1284                 (*packages)[i].package_id = i;
1285
1286         return 0;
1287
1288 err_clean_cores:
1289         free(*cores);
1290 err_clean_threads:
1291         free(*threads);
1292 err:
1293         ERROR("Turbostat plugin: Failled to allocate memory for counters");
1294         return -1;
1295 }
1296
1297 static void
1298 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1299         struct pkg_data *pkg_base, unsigned int cpu_id)
1300 {
1301         struct thread_data *t;
1302         struct core_data *c;
1303         struct pkg_data *p;
1304         struct cpu_topology *cpu = &topology.cpus[cpu_id];
1305
1306         t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1307         c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1308         p = GET_PKG(pkg_base, cpu->package_id);
1309
1310         t->cpu_id = cpu_id;
1311         if (cpu->first_thread_in_core)
1312                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1313         if (cpu->first_core_in_package)
1314                 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1315
1316         c->core_id = cpu->core_id;
1317         p->package_id = cpu->package_id;
1318 }
1319
1320 static void
1321 initialize_counters(void)
1322 {
1323         unsigned int cpu_id;
1324
1325         for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1326                 if (cpu_is_not_present(cpu_id))
1327                         continue;
1328                 init_counter(EVEN_COUNTERS, cpu_id);
1329                 init_counter(ODD_COUNTERS, cpu_id);
1330                 init_counter(DELTA_COUNTERS, cpu_id);
1331         }
1332 }
1333
1334
1335
1336 static void
1337 free_all_buffers(void)
1338 {
1339         allocated = 0;
1340         initialized = 0;
1341
1342         CPU_FREE(cpu_present_set);
1343         cpu_present_set = NULL;
1344         cpu_present_set = 0;
1345
1346         CPU_FREE(cpu_affinity_set);
1347         cpu_affinity_set = NULL;
1348         cpu_affinity_setsize = 0;
1349
1350         CPU_FREE(cpu_saved_affinity_set);
1351         cpu_saved_affinity_set = NULL;
1352         cpu_saved_affinity_setsize = 0;
1353
1354         free(thread_even);
1355         free(core_even);
1356         free(package_even);
1357
1358         thread_even = NULL;
1359         core_even = NULL;
1360         package_even = NULL;
1361
1362         free(thread_odd);
1363         free(core_odd);
1364         free(package_odd);
1365
1366         thread_odd = NULL;
1367         core_odd = NULL;
1368         package_odd = NULL;
1369
1370         free(thread_delta);
1371         free(core_delta);
1372         free(package_delta);
1373
1374         thread_delta = NULL;
1375         core_delta = NULL;
1376         package_delta = NULL;
1377 }
1378
1379
1380 /**********************
1381  * Collectd functions *
1382  **********************/
1383
1384 #define DO_OR_GOTO_ERR(something) \
1385 do {                              \
1386         ret = (something);        \
1387         if (ret < 0)              \
1388                 goto err;         \
1389 } while (0)
1390
1391 static int setup_all_buffers(void)
1392 {
1393         int ret;
1394
1395         DO_OR_GOTO_ERR(topology_probe());
1396         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1397         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1398         DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1399         initialize_counters();
1400         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1401         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1402
1403         allocated = 1;
1404         return 0;
1405 err:
1406         free_all_buffers();
1407         return ret;
1408 }
1409
1410 static int
1411 turbostat_read(void)
1412 {
1413         int ret;
1414
1415         if (!allocated) {
1416                 if ((ret = setup_all_buffers()) < 0)
1417                         return ret;
1418         }
1419
1420         if (for_all_proc_cpus(cpu_is_not_present)) {
1421                 free_all_buffers();
1422                 if ((ret = setup_all_buffers()) < 0)
1423                         return ret;
1424                 if (for_all_proc_cpus(cpu_is_not_present)) {
1425                         ERROR("Turbostat plugin: CPU appeared just after "
1426                               "initialization");
1427                         return -1;
1428                 }
1429         }
1430
1431         /* Saving the scheduling affinity, as it will be modified by get_counters */
1432         if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1433                 ERROR("Turbostat plugin: Unable to save the CPU affinity");
1434                 return -1;
1435         }
1436
1437         if (!initialized) {
1438                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1439                         goto out;
1440                 time_even = cdtime();
1441                 is_even = 1;
1442                 initialized = 1;
1443                 ret = 0;
1444                 goto out;
1445         }
1446
1447         if (is_even) {
1448                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1449                         goto out;
1450                 time_odd = cdtime();
1451                 is_even = 0;
1452                 time_delta = time_odd - time_even;
1453                 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1454                         goto out;
1455                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1456                         goto out;
1457         } else {
1458                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1459                         goto out;
1460                 time_even = cdtime();
1461                 is_even = 1;
1462                 time_delta = time_even - time_odd;
1463                 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1464                         goto out;
1465                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1466                         goto out;
1467         }
1468         ret = 0;
1469 out:
1470         /*
1471          * Let's restore the affinity
1472          * This might fail if the number of CPU changed, but we can't do anything in that case..
1473          */
1474         (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1475         return ret;
1476 }
1477
1478 static int
1479 check_permissions(void)
1480 {
1481         struct __user_cap_header_struct cap_header_data;
1482         cap_user_header_t cap_header = &cap_header_data;
1483         struct __user_cap_data_struct cap_data_data;
1484         cap_user_data_t cap_data = &cap_data_data;
1485         int ret = 0;
1486
1487         if (getuid() == 0) {
1488                 /* We have everything we need */
1489                 return 0;
1490         }
1491
1492         /* check for CAP_SYS_RAWIO */
1493         cap_header->pid = getpid();
1494         cap_header->version = _LINUX_CAPABILITY_VERSION;
1495         if (capget(cap_header, cap_data) < 0) {
1496                 ERROR("Turbostat plugin: capget failed");
1497                 return -1;
1498         }
1499
1500         if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1501                 WARNING("Turbostat plugin: Collectd doesn't have the "
1502                         "CAP_SYS_RAWIO capability. If you don't want to run "
1503                         "collectd as root, try running \"setcap "
1504                         "cap_sys_rawio=ep\" on collectd binary");
1505                 ret = -1;
1506         }
1507
1508         if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1509                 WARNING("Turbostat plugin: Collectd cannot open"
1510                         "/dev/cpu/0/msr. If you don't want to run collectd as "
1511                         "root, you need to change the ownership (chown) and "
1512                         "permissions on /dev/cpu/*/msr to allow such access");
1513                 ret = -1;
1514         }
1515
1516         if (ret != 0)
1517                 ERROR("Turbostat plugin: Initialization failed: this plugin "
1518                       "requires collectd to either to run as root or give "
1519                       "collectd a special capability (CAP_SYS_RAWIO) and read "
1520                       "access to /dev/cpu/*/msr (see previous warnings)");
1521         return ret;
1522 }
1523
1524 static int
1525 turbostat_init(void)
1526 {
1527         struct stat sb;
1528         int ret;
1529
1530         if (stat("/dev/cpu/0/msr", &sb)) {
1531                 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1532                       " does not exist while the CPU supports MSR. You may be "
1533                       "missing the corresponding kernel module, please try '# "
1534                       "modprobe msr'");
1535                 return -1;
1536         }
1537
1538         DO_OR_GOTO_ERR(check_permissions());
1539
1540         DO_OR_GOTO_ERR(probe_cpu());
1541
1542         DO_OR_GOTO_ERR(setup_all_buffers());
1543
1544         plugin_register_read(PLUGIN_NAME, turbostat_read);
1545
1546         return 0;
1547 err:
1548         free_all_buffers();
1549         return ret;
1550 }
1551
1552 static int
1553 turbostat_config(const char *key, const char *value)
1554 {
1555         long unsigned int tmp_val;
1556         char *end;
1557
1558         if (strcasecmp("CoreCstates", key) == 0) {
1559                 tmp_val = strtoul(value, &end, 0);
1560                 if (*end != '\0' || tmp_val > UINT_MAX) {
1561                         ERROR("Turbostat plugin: Invalid CoreCstates '%s'",
1562                               value);
1563                         return -1;
1564                 }
1565                 config_core_cstate = (unsigned int) tmp_val;
1566                 apply_config_core_cstate = 1;
1567         } else if (strcasecmp("PackageCstates", key) == 0) {
1568                 tmp_val = strtoul(value, &end, 0);
1569                 if (*end != '\0' || tmp_val > UINT_MAX) {
1570                         ERROR("Turbostat plugin: Invalid PackageCstates '%s'",
1571                               value);
1572                         return -1;
1573                 }
1574                 config_pkg_cstate = (unsigned int) tmp_val;
1575                 apply_config_pkg_cstate = 1;
1576         } else if (strcasecmp("SystemManagementInterrupt", key) == 0) {
1577                 config_smi = IS_TRUE(value);
1578                 apply_config_smi = 1;
1579         } else if (strcasecmp("DigitalTemperatureSensor", key) == 0) {
1580                 config_dts = IS_TRUE(value);
1581                 apply_config_dts = 1;
1582         } else if (strcasecmp("PackageThermalManagement", key) == 0) {
1583                 config_ptm = IS_TRUE(value);
1584                 apply_config_ptm = 1;
1585         } else if (strcasecmp("RunningAveragePowerLimit", key) == 0) {
1586                 tmp_val = strtoul(value, &end, 0);
1587                 if (*end != '\0' || tmp_val > UINT_MAX) {
1588                         ERROR("Turbostat plugin: Invalid RunningAveragePowerLimit '%s'",
1589                               value);
1590                         return -1;
1591                 }
1592                 config_rapl = (unsigned int) tmp_val;
1593                 apply_config_rapl = 1;
1594         } else if (strcasecmp("TCCActivationTemp", key) == 0) {
1595                 tmp_val = strtoul(value, &end, 0);
1596                 if (*end != '\0' || tmp_val > UINT_MAX) {
1597                         ERROR("Turbostat plugin: Invalid TCCActivationTemp '%s'",
1598                               value);
1599                         return -1;
1600                 }
1601                 tcc_activation_temp = (unsigned int) tmp_val;
1602         } else {
1603                 ERROR("Turbostat plugin: Invalid configuration option '%s'",
1604                       key);
1605                 return -1;
1606         }
1607         return 0;
1608 }
1609
1610 void module_register(void)
1611 {
1612         plugin_register_init(PLUGIN_NAME, turbostat_init);
1613         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1614 }