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