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