src/turbostat.c

   1 /*
   2  * turbostat -- Log CPU frequency and C-state residency
   3  * on modern Intel turbo-capable processors for collectd.
   4  *
   5  * Based on the 'turbostat' tool of the Linux kernel, found at
   6  * linux/tools/power/x86/turbostat/turbostat.c:
   7  * ----
   8  * Copyright (c) 2013 Intel Corporation.
   9  * Len Brown <len.brown@intel.com>
  10  *
  11  * This program is free software; you can redistribute it and/or modify it
  12  * under the terms and conditions of the GNU General Public License,
  13  * version 2, as published by the Free Software Foundation.
  14  *
  15  * This program is distributed in the hope it will be useful, but WITHOUT
  16  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  17  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  18  * more details.
  19  *
  20  * You should have received a copy of the GNU General Public License along with
  21  * this program; if not, write to the Free Software Foundation, Inc.,
  22  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  23  * ----
  24  * Ported to collectd by Vincent Brillault <git@lerya.net>
  25  */
  26
  27 /*
  28  * _GNU_SOURCE is required because of the following functions:
  29  * - CPU_ISSET_S
  30  * - CPU_ZERO_S
  31  * - CPU_SET_S
  32  * - CPU_FREE
  33  * - CPU_ALLOC
  34  * - CPU_ALLOC_SIZE
  35  */
  36 #define _GNU_SOURCE
  37
  38 #include <asm/msr-index.h>
  39 #include <stdarg.h>
  40 #include <stdio.h>
  41 #include <err.h>
  42 #include <unistd.h>
  43 #include <sys/types.h>
  44 #include <sys/wait.h>
  45 #include <sys/stat.h>
  46 #include <sys/resource.h>
  47 #include <fcntl.h>
  48 #include <signal.h>
  49 #include <sys/time.h>
  50 #include <stdlib.h>
  51 #include <dirent.h>
  52 #include <string.h>
  53 #include <ctype.h>
  54 #include <sched.h>
  55 #include <cpuid.h>
  56
  57 #include "collectd.h"
  58 #include "common.h"
  59 #include "plugin.h"
  60
  61 #define PLUGIN_NAME "turbostat"
  62
  63 /*
  64  * This tool uses the Model-Specific Registers (MSRs) present on Intel processors.
  65  * The general description each of these registers, depending on the architecture,
  66  * can be found in the Intel® 64 and IA-32 Architectures Software Developer Manual,
  67  * Volume 3 Chapter 35.
  68  */
  69
  70 /*
  71  * If set, aperf_mperf_unstable disables a/mperf based stats.
  72  * This includes: C0 & C1 states, frequency
  73  *
  74  * This value is automatically set if mperf or aperf go backward
  75  */
  76 static _Bool aperf_mperf_unstable;
  77
  78 /*
  79  * Bitmask of the list of core C states supported by the processor.
  80  * Currently supported C-states (by this plugin): 3, 6, 7
  81  */
  82 static unsigned int do_core_cstate;
  83
  84 /*
  85  * Bitmask of the list of pacages C states supported by the processor.
  86  * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
  87  */
  88 static unsigned int do_pkg_cstate;
  89
  90 /*
  91  * Boolean indicating if the processor supports 'Digital temperature sensor'
  92  * This feature enables the monitoring of the temperature of each core
  93  *
  94  * This feature has two limitations:
  95  *  - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
  96  *  - Temperatures above the tcc_activation_temp are not recorded
  97  */
  98 static _Bool do_dts;
  99
 100 /*
 101  * Boolean indicating if the processor supports 'Package thermal management'
 102  * This feature allows the monitoring of the temperature of each package
 103  *
 104  * This feature has two limitations:
 105  *  - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
 106  *  - Temperatures above the tcc_activation_temp are not recorded
 107  */
 108 static _Bool do_ptm;
 109
 110 /*
 111  * Thermal Control Circuit Activation Temperature as configured by the user.
 112  * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
 113  * and should only be used if the automated detection fails.
 114  */
 115 static unsigned int tcc_activation_temp;
 116
 117 static unsigned int do_rapl;
 118 static double rapl_energy_units;
 119
 120 #define RAPL_PKG                (1 << 0)
 121                                         /* 0x610 MSR_PKG_POWER_LIMIT */
 122                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
 123 #define RAPL_PKG_PERF_STATUS    (1 << 1)
 124                                         /* 0x613 MSR_PKG_PERF_STATUS */
 125 #define RAPL_PKG_POWER_INFO     (1 << 2)
 126                                         /* 0x614 MSR_PKG_POWER_INFO */
 127
 128 #define RAPL_DRAM               (1 << 3)
 129                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
 130                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
 131                                         /* 0x61c MSR_DRAM_POWER_INFO */
 132 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
 133                                         /* 0x61b MSR_DRAM_PERF_STATUS */
 134
 135 #define RAPL_CORES              (1 << 5)
 136                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 137                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 138 #define RAPL_CORE_POLICY        (1 << 6)
 139                                         /* 0x63a MSR_PP0_POLICY */
 140
 141
 142 #define RAPL_GFX                (1 << 7)
 143                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 144                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 145                                         /* 0x642 MSR_PP1_POLICY */
 146 #define TJMAX_DEFAULT   100
 147
 148 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
 149 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
 150
 151 struct thread_data {
 152         unsigned long long tsc;
 153         unsigned long long aperf;
 154         unsigned long long mperf;
 155         unsigned long long c1;
 156         unsigned int smi_count;
 157         unsigned int cpu_id;
 158         unsigned int flags;
 159 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 160 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 161 } *thread_delta, *thread_even, *thread_odd;
 162
 163 struct core_data {
 164         unsigned long long c3;
 165         unsigned long long c6;
 166         unsigned long long c7;
 167         unsigned int core_temp_c;
 168         unsigned int core_id;
 169 } *core_delta, *core_even, *core_odd;
 170
 171 struct pkg_data {
 172         unsigned long long pc2;
 173         unsigned long long pc3;
 174         unsigned long long pc6;
 175         unsigned long long pc7;
 176         unsigned long long pc8;
 177         unsigned long long pc9;
 178         unsigned long long pc10;
 179         unsigned int package_id;
 180         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 181         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 182         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 183         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 184         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 185         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 186         unsigned int tcc_activation_temp;
 187         unsigned int pkg_temp_c;
 188 } *package_delta, *package_even, *package_odd;
 189
 190 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
 191 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 192 #define EVEN_COUNTERS thread_even, core_even, package_even
 193 static _Bool is_even = 1;
 194
 195 static _Bool allocated = 0;
 196 static _Bool initialized = 0;
 197
 198 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 199         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 200                 topo.num_threads_per_core + \
 201                 (core_no) * topo.num_threads_per_core + (thread_no))
 202 #define GET_CORE(core_base, core_no, pkg_no) \
 203         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 204 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 205
 206 struct topo_params {
 207         int num_packages;
 208         int num_cpus;
 209         int num_cores;
 210         int max_cpu_num;
 211         int num_cores_per_pkg;
 212         int num_threads_per_core;
 213 } topo;
 214
 215 struct timeval tv_even, tv_odd, tv_delta;
 216
 217 enum return_values {
 218         OK = 0,
 219         ERR_CPU_MIGRATE,
 220         ERR_CPU_SAVE_SCHED_AFFINITY,
 221         ERR_MSR_IA32_APERF,
 222         ERR_MSR_IA32_MPERF,
 223         ERR_MSR_SMI_COUNT,
 224         ERR_MSR_CORE_C3_RESIDENCY,
 225         ERR_MSR_CORE_C6_RESIDENCY,
 226         ERR_MSR_CORE_C7_RESIDENCY,
 227         ERR_MSR_IA32_THERM_STATUS,
 228         ERR_MSR_PKG_C3_RESIDENCY,
 229         ERR_MSR_PKG_C6_RESIDENCY,
 230         ERR_MSR_PKG_C2_RESIDENCY,
 231         ERR_MSR_PKG_C7_RESIDENCY,
 232         ERR_MSR_PKG_C8_RESIDENCY,
 233         ERR_MSR_PKG_C9_RESIDENCY,
 234         ERR_MSR_PKG_C10_RESIDENCY,
 235         ERR_MSR_PKG_ENERGY_STATUS,
 236         ERR_MSR_PKG_POWER_INFO,
 237         ERR_MSR_PP0_ENERGY_STATUS,
 238         ERR_MSR_DRAM_ENERGY_STATUS,
 239         ERR_MSR_PP1_ENERGY_STATUS,
 240         ERR_MSR_PKG_PERF_STATUS,
 241         ERR_MSR_DRAM_PERF_STATUS,
 242         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 243         ERR_MSR_IA32_TSC,
 244         ERR_CPU_NOT_PRESENT,
 245         ERR_NO_MSR,
 246         ERR_CANT_OPEN_MSR,
 247         ERR_CANT_OPEN_FILE,
 248         ERR_CANT_READ_NUMBER,
 249         ERR_CANT_READ_PROC_STAT,
 250         ERR_NO_INVARIANT_TSC,
 251         ERR_NO_APERF,
 252         ERR_CALLOC,
 253         ERR_CPU_ALLOC,
 254         ERR_NOT_ROOT,
 255         UNSUPPORTED_CPU,
 256 };
 257
 258 static int setup_all_buffers(void);
 259
 260 static int
 261 cpu_is_not_present(int cpu)
 262 {
 263         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 264 }
 265 /*
 266  * run func(thread, core, package) in topology order
 267  * skip non-present cpus
 268  */
 269
 270 static int __attribute__((warn_unused_result))
 271 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 272         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 273 {
 274         int retval, pkg_no, core_no, thread_no;
 275
 276         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 277                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 278                         for (thread_no = 0; thread_no <
 279                                 topo.num_threads_per_core; ++thread_no) {
 280                                 struct thread_data *t;
 281                                 struct core_data *c;
 282                                 struct pkg_data *p;
 283
 284                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 285
 286                                 if (cpu_is_not_present(t->cpu_id))
 287                                         continue;
 288
 289                                 c = GET_CORE(core_base, core_no, pkg_no);
 290                                 p = GET_PKG(pkg_base, pkg_no);
 291
 292                                 retval = func(t, c, p);
 293                                 if (retval)
 294                                         return retval;
 295                         }
 296                 }
 297         }
 298         return 0;
 299 }
 300
 301 /*
 302  * Open a MSR device for reading
 303  * Can change the scheduling affinity of the current process if multiple_read is 1
 304  */
 305 static int __attribute__((warn_unused_result))
 306 open_msr(int cpu, _Bool multiple_read)
 307 {
 308         char pathname[32];
 309         int fd;
 310
 311         /*
 312          * If we need to do multiple read, let's migrate to the CPU
 313          * Otherwise, we would lose time calling functions on another CPU
 314          *
 315          * If we are not yet initialized (cpu_affinity_setsize = 0),
 316          * we need to skip this optimisation.
 317          */
 318         if (multiple_read && cpu_affinity_setsize) {
 319                 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 320                 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 321                 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
 322                         ERROR("Could not migrate to CPU %d", cpu);
 323                         return -ERR_CPU_MIGRATE;
 324                 }
 325         }
 326
 327         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 328         fd = open(pathname, O_RDONLY);
 329         if (fd < 0)
 330                 return -ERR_CANT_OPEN_MSR;
 331         return fd;
 332 }
 333
 334 /*
 335  * Read a single MSR from an open file descriptor
 336  */
 337 static int __attribute__((warn_unused_result))
 338 read_msr(int fd, off_t offset, unsigned long long *msr)
 339 {
 340         ssize_t retval;
 341
 342         retval = pread(fd, msr, sizeof *msr, offset);
 343
 344         if (retval != sizeof *msr) {
 345                 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
 346                 return -1;
 347         }
 348         return 0;
 349 }
 350
 351 /*
 352  * Open a MSR device for reading, read the value asked for and close it.
 353  * This call will not affect the scheduling affinity of this thread.
 354  */
 355 static int __attribute__((warn_unused_result))
 356 get_msr(int cpu, off_t offset, unsigned long long *msr)
 357 {
 358         ssize_t retval;
 359         int fd;
 360
 361         fd = open_msr(cpu, 0);
 362         if (fd < 0)
 363                 return fd;
 364         retval = read_msr(fd, offset, msr);
 365         close(fd);
 366         return retval;
 367 }
 368
 369 #define DELTA_WRAP32(delta, new, old)                   \
 370         if (new > old) {                                \
 371                 delta = new - old;                      \
 372         } else {                                        \
 373                 delta = 0x100000000 + new - old;        \
 374         }
 375
 376 static void
 377 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
 378 {
 379         delta->pc2 = new->pc2 - old->pc2;
 380         delta->pc3 = new->pc3 - old->pc3;
 381         delta->pc6 = new->pc6 - old->pc6;
 382         delta->pc7 = new->pc7 - old->pc7;
 383         delta->pc8 = new->pc8 - old->pc8;
 384         delta->pc9 = new->pc9 - old->pc9;
 385         delta->pc10 = new->pc10 - old->pc10;
 386         delta->pkg_temp_c = new->pkg_temp_c;
 387
 388         DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
 389         DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
 390         DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
 391         DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
 392         DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 393         DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 394 }
 395
 396 static void
 397 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
 398 {
 399         delta->c3 = new->c3 - old->c3;
 400         delta->c6 = new->c6 - old->c6;
 401         delta->c7 = new->c7 - old->c7;
 402         delta->core_temp_c = new->core_temp_c;
 403 }
 404
 405 static int __attribute__((warn_unused_result))
 406 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
 407         const struct core_data *core_delta)
 408 {
 409         delta->tsc = new->tsc - old->tsc;
 410
 411         /* check for TSC < 1 Mcycles over interval */
 412         if (delta->tsc < (1000 * 1000)) {
 413                 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
 414                 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
 415                 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
 416                 return -1;
 417         }
 418
 419         delta->c1 = new->c1 - old->c1;
 420
 421         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 422                 delta->aperf = new->aperf - old->aperf;
 423                 delta->mperf = new->mperf - old->mperf;
 424         } else {
 425                 if (!aperf_mperf_unstable) {
 426                         WARNING(" APERF or MPERF went backwards * ");
 427                         WARNING("* Frequency results do not cover entire interval *");
 428                         WARNING("* fix this by running Linux-2.6.30 or later *");
 429
 430                         aperf_mperf_unstable = 1;
 431                 }
 432         }
 433
 434         /*
 435          * As counter collection is not atomic,
 436          * it is possible for mperf's non-halted cycles + idle states
 437          * to exceed TSC's all cycles: show c1 = 0% in that case.
 438          */
 439         if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
 440                 delta->c1 = 0;
 441         else {
 442                 /* normal case, derive c1 */
 443                 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
 444                         - core_delta->c6 - core_delta->c7;
 445         }
 446
 447         if (delta->mperf == 0) {
 448                 WARNING("cpu%d MPERF 0!", old->cpu_id);
 449                 delta->mperf = 1;       /* divide by 0 protection */
 450         }
 451
 452         delta->smi_count = new->smi_count - old->smi_count;
 453
 454         return 0;
 455 }
 456
 457 static int __attribute__((warn_unused_result))
 458 delta_cpu(struct thread_data *t_delta, struct core_data *c_delta, struct pkg_data *p_delta,
 459           const struct thread_data *t_new, const struct core_data *c_new, const struct pkg_data *p_new,
 460           const struct thread_data *t_old, const struct core_data *c_old, const struct pkg_data *p_old)
 461 {
 462         int ret;
 463
 464         /* calculate core delta only for 1st thread in core */
 465         if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 466                 delta_core(c_delta, c_new, c_old);
 467
 468         /* always calculate thread delta */
 469         ret = delta_thread(t_delta, t_new, t_old, c_delta);
 470         if (ret != 0)
 471                 return ret;
 472
 473         /* calculate package delta only for 1st core in package */
 474         if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 475                 delta_package(p_delta, p_new, p_old);
 476
 477         return 0;
 478 }
 479
 480
 481 /*
 482  * get_counters(...)
 483  * migrate to cpu
 484  * acquire and record local counters for that cpu
 485  */
 486 static int __attribute__((warn_unused_result))
 487 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 488 {
 489         int cpu = t->cpu_id;
 490         unsigned long long msr;
 491         int msr_fd;
 492         int retval = 0;
 493
 494         msr_fd = open_msr(cpu, 1);
 495         if (msr_fd < 0)
 496                 return msr_fd;
 497
 498 #define READ_MSR(msr, dst)                      \
 499 do {                                            \
 500         if (read_msr(msr_fd, msr, dst)) {       \
 501                 retval = -ERR_##msr;            \
 502                 goto out;                       \
 503         }                                       \
 504 } while (0)
 505
 506         READ_MSR(MSR_IA32_TSC, &t->tsc);
 507
 508         READ_MSR(MSR_IA32_APERF, &t->aperf);
 509         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 510
 511         READ_MSR(MSR_SMI_COUNT, &msr);
 512         t->smi_count = msr & 0xFFFFFFFF;
 513
 514         /* collect core counters only for 1st thread in core */
 515         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 516                 retval = 0;
 517                 goto out;
 518         }
 519
 520         if (do_core_cstate & (1 << 3))
 521                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 522         if (do_core_cstate & (1 << 6))
 523                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 524         if (do_core_cstate & (1 << 7))
 525                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 526
 527         if (do_dts) {
 528                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 529                 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 530         }
 531
 532         /* collect package counters only for 1st core in package */
 533         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 534                 retval = 0;
 535                 goto out;
 536         }
 537
 538         if (do_pkg_cstate & (1 << 2))
 539                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 540         if (do_pkg_cstate & (1 << 3))
 541                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 542         if (do_pkg_cstate & (1 << 6))
 543                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 544         if (do_pkg_cstate & (1 << 7))
 545                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 546         if (do_pkg_cstate & (1 << 8))
 547                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 548         if (do_pkg_cstate & (1 << 9))
 549                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 550         if (do_pkg_cstate & (1 << 10))
 551                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 552
 553         if (do_rapl & RAPL_PKG) {
 554                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 555                 p->energy_pkg = msr & 0xFFFFFFFF;
 556         }
 557         if (do_rapl & RAPL_CORES) {
 558                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 559                 p->energy_cores = msr & 0xFFFFFFFF;
 560         }
 561         if (do_rapl & RAPL_DRAM) {
 562                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 563                 p->energy_dram = msr & 0xFFFFFFFF;
 564         }
 565         if (do_rapl & RAPL_GFX) {
 566                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 567                 p->energy_gfx = msr & 0xFFFFFFFF;
 568         }
 569         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 570                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 571                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 572         }
 573         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 574                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 575                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 576         }
 577         if (do_ptm) {
 578                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 579                 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 580         }
 581
 582 out:
 583         close(msr_fd);
 584         return retval;
 585 }
 586
 587 static void
 588 free_all_buffers(void)
 589 {
 590         allocated = 0;
 591         initialized = 0;
 592
 593         CPU_FREE(cpu_present_set);
 594         cpu_present_set = NULL;
 595         cpu_present_set = 0;
 596
 597         CPU_FREE(cpu_affinity_set);
 598         cpu_affinity_set = NULL;
 599         cpu_affinity_setsize = 0;
 600
 601         CPU_FREE(cpu_saved_affinity_set);
 602         cpu_saved_affinity_set = NULL;
 603         cpu_saved_affinity_setsize = 0;
 604
 605         free(thread_even);
 606         free(core_even);
 607         free(package_even);
 608
 609         thread_even = NULL;
 610         core_even = NULL;
 611         package_even = NULL;
 612
 613         free(thread_odd);
 614         free(core_odd);
 615         free(package_odd);
 616
 617         thread_odd = NULL;
 618         core_odd = NULL;
 619         package_odd = NULL;
 620
 621         free(thread_delta);
 622         free(core_delta);
 623         free(package_delta);
 624
 625         thread_delta = NULL;
 626         core_delta = NULL;
 627         package_delta = NULL;
 628 }
 629
 630 /*
 631  * Parse a file containing a single int.
 632  */
 633 static int __attribute__ ((format(printf,1,2)))
 634 parse_int_file(const char *fmt, ...)
 635 {
 636         va_list args;
 637         char path[PATH_MAX];
 638         FILE *filep;
 639         int value;
 640
 641         va_start(args, fmt);
 642         vsnprintf(path, sizeof(path), fmt, args);
 643         va_end(args);
 644         filep = fopen(path, "r");
 645         if (!filep) {
 646                 ERROR("%s: open failed", path);
 647                 return -ERR_CANT_OPEN_FILE;
 648         }
 649         if (fscanf(filep, "%d", &value) != 1) {
 650                 ERROR("%s: failed to parse number from file", path);
 651                 return -ERR_CANT_READ_NUMBER;
 652         }
 653         fclose(filep);
 654         return value;
 655 }
 656
 657 /*
 658  * cpu_is_first_sibling_in_core(cpu)
 659  * return 1 if given CPU is 1st HT sibling in the core
 660  */
 661 static int
 662 cpu_is_first_sibling_in_core(int cpu)
 663 {
 664         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 665 }
 666
 667 /*
 668  * cpu_is_first_core_in_package(cpu)
 669  * return 1 if given CPU is 1st core in package
 670  */
 671 static int
 672 cpu_is_first_core_in_package(int cpu)
 673 {
 674         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 675 }
 676
 677 static int
 678 get_physical_package_id(int cpu)
 679 {
 680         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 681 }
 682
 683 static int
 684 get_core_id(int cpu)
 685 {
 686         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 687 }
 688
 689 static int
 690 get_num_ht_siblings(int cpu)
 691 {
 692         char path[80];
 693         FILE *filep;
 694         int sib1, sib2;
 695         int matches;
 696         char character;
 697
 698         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 699         filep = fopen(path, "r");
 700         if (!filep) {
 701                 ERROR("%s: open failed", path);
 702                 return -ERR_CANT_OPEN_FILE;
 703         }
 704         /*
 705          * file format:
 706          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 707          * otherwinse 1 sibling (self).
 708          */
 709         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 710
 711         fclose(filep);
 712
 713         if (matches == 3)
 714                 return 2;
 715         else
 716                 return 1;
 717 }
 718
 719 static int __attribute__((warn_unused_result))
 720 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,
 721                    const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
 722 {
 723         int retval, pkg_no, core_no, thread_no;
 724
 725         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 726                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 727                         for (thread_no = 0; thread_no <
 728                                 topo.num_threads_per_core; ++thread_no) {
 729                                 struct thread_data *t_delta;
 730                                 const struct thread_data *t_old, *t_new;
 731                                 struct core_data *c_delta;
 732                                 const struct core_data *c_old, *c_new;
 733                                 struct pkg_data *p_delta;
 734                                 const struct pkg_data *p_old, *p_new;
 735
 736                                 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
 737                                 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
 738                                 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
 739                                 if (cpu_is_not_present(t_delta->cpu_id))
 740                                         continue;
 741
 742                                 c_delta = GET_CORE(core_delta, core_no, pkg_no);
 743                                 c_new = GET_CORE(core_new_base, core_no, pkg_no);
 744                                 c_old = GET_CORE(core_old_base, core_no, pkg_no);
 745
 746                                 p_delta = GET_PKG(package_delta, pkg_no);
 747                                 p_new = GET_PKG(pkg_new_base, pkg_no);
 748                                 p_old = GET_PKG(pkg_old_base, pkg_no);
 749
 750                                 retval = delta_cpu(t_delta, c_delta, p_delta,
 751                                                    t_new, c_new, p_new,
 752                                                    t_old, c_old, p_old);
 753                                 if (retval)
 754                                         return retval;
 755                         }
 756                 }
 757         }
 758         return 0;
 759 }
 760
 761 /*
 762  * run func(cpu) on every cpu in /proc/stat
 763  * return max_cpu number
 764  */
 765 static int __attribute__((warn_unused_result))
 766 for_all_proc_cpus(int (func)(int))
 767 {
 768         FILE *fp;
 769         int cpu_num;
 770         int retval;
 771
 772         fp = fopen("/proc/stat", "r");
 773         if (!fp) {
 774                 ERROR("Failed to open /proc/stat");
 775                 return -ERR_CANT_OPEN_FILE;
 776         }
 777
 778         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 779         if (retval != 0) {
 780                 ERROR("Failed to parse /proc/stat");
 781                 fclose(fp);
 782                 return -ERR_CANT_READ_PROC_STAT;
 783         }
 784
 785         while (1) {
 786                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 787                 if (retval != 1)
 788                         break;
 789
 790                 retval = func(cpu_num);
 791                 if (retval) {
 792                         fclose(fp);
 793                         return(retval);
 794                 }
 795         }
 796         fclose(fp);
 797         return 0;
 798 }
 799
 800 /*
 801  * count_cpus()
 802  * remember the last one seen, it will be the max
 803  */
 804 static int
 805 count_cpus(int cpu)
 806 {
 807         if (topo.max_cpu_num < cpu)
 808                 topo.max_cpu_num = cpu;
 809
 810         topo.num_cpus += 1;
 811         return 0;
 812 }
 813 static int
 814 mark_cpu_present(int cpu)
 815 {
 816         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 817         return 0;
 818 }
 819
 820
 821 static void
 822 turbostat_submit (const char *plugin_instance,
 823         const char *type, const char *type_instance,
 824         gauge_t value)
 825 {
 826         value_list_t vl = VALUE_LIST_INIT;
 827         value_t v;
 828
 829         v.gauge = value;
 830         vl.values = &v;
 831         vl.values_len = 1;
 832         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 833         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 834         if (plugin_instance != NULL)
 835                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 836         sstrncpy (vl.type, type, sizeof (vl.type));
 837         if (type_instance != NULL)
 838                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 839
 840         plugin_dispatch_values (&vl);
 841 }
 842
 843 /*
 844  * column formatting convention & formats
 845  * package: "pk" 2 columns %2d
 846  * core: "cor" 3 columns %3d
 847  * CPU: "CPU" 3 columns %3d
 848  * Pkg_W: %6.2
 849  * Cor_W: %6.2
 850  * GFX_W: %5.2
 851  * RAM_W: %5.2
 852  * GHz: "GHz" 3 columns %3.2
 853  * TSC: "TSC" 3 columns %3.2
 854  * SMI: "SMI" 4 columns %4d
 855  * percentage " %pc3" %6.2
 856  * Perf Status percentage: %5.2
 857  * "CTMP" 4 columns %4d
 858  */
 859 static int
 860 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 861 {
 862         char name[12];
 863         double interval_float;
 864
 865         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 866
 867         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 868
 869         if (!aperf_mperf_unstable)
 870                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 871         if (!aperf_mperf_unstable)
 872                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 873
 874         /* GHz */
 875         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 876                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 877
 878         /* SMI */
 879         turbostat_submit(NULL, "current", name, t->smi_count);
 880
 881         /* print per-core data only for 1st thread in core */
 882         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 883                 goto done;
 884
 885         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 886
 887         if (do_core_cstate & (1 << 3))
 888                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 889         if (do_core_cstate & (1 << 6))
 890                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 891         if (do_core_cstate & (1 << 7))
 892                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 893
 894         if (do_dts)
 895                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 896
 897         /* print per-package data only for 1st core in package */
 898         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 899                 goto done;
 900
 901         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 902
 903         if (do_ptm)
 904                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 905
 906         if (do_pkg_cstate & (1 << 2))
 907                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 908         if (do_pkg_cstate & (1 << 3))
 909                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 910         if (do_pkg_cstate & (1 << 6))
 911                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 912         if (do_pkg_cstate & (1 << 7))
 913                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 914         if (do_pkg_cstate & (1 << 8))
 915                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 916         if (do_pkg_cstate & (1 << 9))
 917                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 918         if (do_pkg_cstate & (1 << 10))
 919                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 920
 921         if (do_rapl) {
 922                 if (do_rapl & RAPL_PKG)
 923                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 924                 if (do_rapl & RAPL_CORES)
 925                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 926                 if (do_rapl & RAPL_GFX)
 927                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 928                 if (do_rapl & RAPL_DRAM)
 929                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 930         }
 931 done:
 932         return 0;
 933 }
 934
 935 static int
 936 turbostat_read(user_data_t * not_used)
 937 {
 938         int ret;
 939
 940         if (!allocated) {
 941                 if ((ret = setup_all_buffers()) < 0)
 942                         return ret;
 943         }
 944
 945         if (for_all_proc_cpus(cpu_is_not_present)) {
 946                 free_all_buffers();
 947                 if ((ret = setup_all_buffers()) < 0)
 948                         return ret;
 949                 if (for_all_proc_cpus(cpu_is_not_present))
 950                         return -ERR_CPU_NOT_PRESENT;
 951         }
 952
 953         /* Saving the scheduling affinity, as it will be modified by get_counters */
 954         if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0)
 955                 return -ERR_CPU_SAVE_SCHED_AFFINITY;
 956
 957         if (!initialized) {
 958                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 959                         goto out;
 960                 gettimeofday(&tv_even, (struct timezone *)NULL);
 961                 is_even = 1;
 962                 initialized = 1;
 963                 ret = 0;
 964                 goto out;
 965         }
 966
 967         if (is_even) {
 968                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 969                         goto out;
 970                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 971                 is_even = 0;
 972                 timersub(&tv_odd, &tv_even, &tv_delta);
 973                 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 974                         goto out;
 975                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
 976                         goto out;
 977         } else {
 978                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 979                         goto out;
 980                 gettimeofday(&tv_even, (struct timezone *)NULL);
 981                 is_even = 1;
 982                 timersub(&tv_even, &tv_odd, &tv_delta);
 983                 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 984                         goto out;
 985                 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
 986                         goto out;
 987         }
 988         ret = 0;
 989 out:
 990         /*
 991          * Let's restore the affinity
 992          * This might fail if the number of CPU changed, but we can't do anything in that case..
 993          */
 994         (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
 995         return ret;
 996 }
 997
 998 static int __attribute__((warn_unused_result))
 999 check_dev_msr()
1000 {
1001         struct stat sb;
1002
1003         if (stat("/dev/cpu/0/msr", &sb)) {
1004                 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
1005                 return -ERR_NO_MSR;
1006         }
1007         return 0;
1008 }
1009
1010 static int __attribute__((warn_unused_result))
1011 check_super_user()
1012 {
1013         if (getuid() != 0) {
1014                 ERROR("must be root");
1015                 return -ERR_NOT_ROOT;
1016         }
1017         return 0;
1018 }
1019
1020 /*
1021  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1022  * the Thermal Control Circuit (TCC) activates.
1023  * This is usually equal to tjMax.
1024  *
1025  * Older processors do not have this MSR, so there we guess,
1026  * but also allow conficuration over-ride with "TCCActivationTemp".
1027  *
1028  * Several MSR temperature values are in units of degrees-C
1029  * below this value, including the Digital Thermal Sensor (DTS),
1030  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1031  */
1032 static int __attribute__((warn_unused_result))
1033 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1034 {
1035         unsigned long long msr;
1036         unsigned int target_c_local;
1037
1038         /* tcc_activation_temp is used only for dts or ptm */
1039         if (!(do_dts || do_ptm))
1040                 return 0;
1041
1042         /* this is a per-package concept */
1043         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1044                 return 0;
1045
1046         if (tcc_activation_temp != 0) {
1047                 p->tcc_activation_temp = tcc_activation_temp;
1048                 return 0;
1049         }
1050
1051         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
1052                 goto guess;
1053
1054         target_c_local = (msr >> 16) & 0x7F;
1055
1056         if (target_c_local < 85 || target_c_local > 127)
1057                 goto guess;
1058
1059         p->tcc_activation_temp = target_c_local;
1060
1061         return 0;
1062
1063 guess:
1064         p->tcc_activation_temp = TJMAX_DEFAULT;
1065         WARNING("cpu%d: Guessing tjMax %d C, Please use TCCActivationTemp to specify",
1066                 t->cpu_id, p->tcc_activation_temp);
1067
1068         return 0;
1069 }
1070
1071 /*
1072  * Identify the functionality of the CPU
1073  */
1074 static int __attribute__((warn_unused_result))
1075 probe_cpu()
1076 {
1077         unsigned int eax, ebx, ecx, edx, max_level;
1078         unsigned int fms, family, model;
1079
1080         /* CPUID(0):
1081          * - EAX: Maximum Input Value for Basic CPUID Information
1082          * - EBX: "Genu" (0x756e6547)
1083          * - EDX: "ineI" (0x49656e69)
1084          * - ECX: "ntel" (0x6c65746e)
1085          */
1086         max_level = ebx = ecx = edx = 0;
1087         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1088         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1089                 ERROR("Unsupported CPU");
1090                 return -UNSUPPORTED_CPU;
1091         }
1092
1093         /* CPUID(1):
1094          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1095          *  + 4-7:   Model ID
1096          *  + 8-11:  Family ID
1097          *  + 12-13: Processor type
1098          *  + 16-19: Extended Model ID
1099          *  + 20-27: Extended Family ID
1100          * - EDX: Feature Information:
1101          *  + 5: Support for MSR read/write operations
1102          */
1103         fms = ebx = ecx = edx = 0;
1104         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1105         family = (fms >> 8) & 0xf;
1106         model = (fms >> 4) & 0xf;
1107         if (family == 0xf)
1108                 family += (fms >> 20) & 0xf;
1109         if (family == 6 || family == 0xf)
1110                 model += ((fms >> 16) & 0xf) << 4;
1111         if (!(edx & (1 << 5))) {
1112                 ERROR("CPUID: no MSR");
1113                 return -ERR_NO_MSR;
1114         }
1115
1116         /*
1117          * CPUID(0x80000000):
1118          * - EAX: Maximum Input Value for Extended Function CPUID Information
1119          *
1120          * This allows us to verify if the CPUID(0x80000007) can be called
1121          *
1122          * This check is valid for both Intel and AMD.
1123          */
1124         max_level = ebx = ecx = edx = 0;
1125         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1126         if (max_level < 0x80000007) {
1127                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1128                 return -ERR_NO_INVARIANT_TSC;
1129         }
1130
1131         /*
1132          * CPUID(0x80000007):
1133          * - EDX:
1134          *  + 8: Invariant TSC available if set
1135          *
1136          * This check is valid for both Intel and AMD
1137          */
1138         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1139         if (!(edx & (1 << 8))) {
1140                 ERROR("No invariant TSC");
1141                 return -ERR_NO_INVARIANT_TSC;
1142         }
1143
1144         /*
1145          * CPUID(6):
1146          * - EAX:
1147          *  + 0: Digital temperature sensor is supported if set
1148          *  + 6: Package thermal management is supported if set
1149          * - ECX:
1150          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1151          *  + 3: The processor supports performance-energy bias preference if set.
1152          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1153          *
1154          * This check is valid for both Intel and AMD
1155          */
1156         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1157         do_dts = eax & (1 << 0);
1158         do_ptm = eax & (1 << 6);
1159         if (!(ecx & (1 << 0))) {
1160                 ERROR("No APERF");
1161                 return -ERR_NO_APERF;
1162         }
1163
1164         /*
1165          * Enable or disable C states depending on the model and family
1166          */
1167         if (family == 6) {
1168                 switch (model) {
1169                 /* Atom (partial) */
1170                 case 0x27:
1171                         do_core_cstate = 0;
1172                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1173                         break;
1174                 /* Silvermont */
1175                 case 0x37: /* BYT */
1176                 case 0x4A:
1177                 case 0x4D: /* AVN */
1178                 case 0x5A:
1179                 case 0x5D:
1180                         do_core_cstate = (1 << 1) | (1 << 6);
1181                         do_pkg_cstate = (1 << 6);
1182                         break;
1183                 /* Nehalem */
1184                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1185                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1186                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1187                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1188                         do_core_cstate = (1 << 3) | (1 << 6);
1189                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1190                         break;
1191                 /* Westmere */
1192                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1193                 case 0x2C: /* Westmere EP - Gulftown */
1194                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1195                         do_core_cstate = (1 << 3) | (1 << 6);
1196                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1197                         break;
1198                 /* Sandy Bridge */
1199                 case 0x2A: /* SNB */
1200                 case 0x2D: /* SNB Xeon */
1201                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1202                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1203                         break;
1204                 /* Ivy Bridge */
1205                 case 0x3A: /* IVB */
1206                 case 0x3E: /* IVB Xeon */
1207                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1208                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1209                         break;
1210                 /* Haswell Bridge */
1211                 case 0x3C: /* HSW */
1212                 case 0x3F: /* HSW */
1213                 case 0x46: /* HSW */
1214                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1215                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1216                         break;
1217                 case 0x45: /* HSW */
1218                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1219                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1220                         break;
1221                 /* Broadwel */
1222                 case 0x4F: /* BDW */
1223                 case 0x56: /* BDX-DE */
1224                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1225                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1226                         break;
1227                 case 0x3D: /* BDW */
1228                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1229                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1230                         break;
1231                 default:
1232                         ERROR("Unsupported CPU");
1233                 }
1234                 switch (model) {
1235                 case 0x2A:
1236                 case 0x3A:
1237                 case 0x3C:
1238                 case 0x45:
1239                 case 0x46:
1240                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1241                         break;
1242                 case 0x3F:
1243                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1244                         break;
1245                 case 0x2D:
1246                 case 0x3E:
1247                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1248                         break;
1249                 case 0x37:
1250                 case 0x4D:
1251                         do_rapl = RAPL_PKG | RAPL_CORES;
1252                         break;
1253                 default:
1254                         do_rapl = 0;
1255                 }
1256         } else {
1257                 ERROR("Unsupported CPU");
1258                 return -UNSUPPORTED_CPU;
1259         }
1260
1261         if (do_rapl) {
1262                 unsigned long msr;
1263                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1264                         return 0;
1265
1266                 if (model == 0x37)
1267                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1268                 else
1269                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1270         }
1271
1272         return 0;
1273 }
1274
1275
1276
1277 static int __attribute__((warn_unused_result))
1278 topology_probe()
1279 {
1280         int i;
1281         int ret;
1282         int max_core_id = 0;
1283         int max_package_id = 0;
1284         int max_siblings = 0;
1285         struct cpu_topology {
1286                 int core_id;
1287                 int physical_package_id;
1288         } *cpus;
1289
1290         /* Initialize num_cpus, max_cpu_num */
1291         topo.num_cpus = 0;
1292         topo.max_cpu_num = 0;
1293         ret = for_all_proc_cpus(count_cpus);
1294         if (ret < 0)
1295                 return ret;
1296
1297         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1298
1299         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1300         if (cpus == NULL) {
1301                 ERROR("calloc cpus");
1302                 return -ERR_CALLOC;
1303         }
1304
1305         /*
1306          * Allocate and initialize cpu_present_set
1307          */
1308         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1309         if (cpu_present_set == NULL) {
1310                 free(cpus);
1311                 ERROR("CPU_ALLOC");
1312                 return -ERR_CPU_ALLOC;
1313         }
1314         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1315         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1316         ret = for_all_proc_cpus(mark_cpu_present);
1317         if (ret < 0) {
1318                 free(cpus);
1319                 return ret;
1320         }
1321
1322         /*
1323          * Allocate and initialize cpu_affinity_set
1324          */
1325         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1326         if (cpu_affinity_set == NULL) {
1327                 free(cpus);
1328                 ERROR("CPU_ALLOC");
1329                 return -ERR_CPU_ALLOC;
1330         }
1331         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1332         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1333
1334
1335         /*
1336          * Allocate and initialize cpu_saved_affinity_set
1337          */
1338         cpu_saved_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1339         if (cpu_saved_affinity_set == NULL) {
1340                 free(cpus);
1341                 ERROR("CPU_ALLOC");
1342                 return -ERR_CPU_ALLOC;
1343         }
1344         cpu_saved_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1345         CPU_ZERO_S(cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1346
1347
1348         /*
1349          * For online cpus
1350          * find max_core_id, max_package_id
1351          */
1352         for (i = 0; i <= topo.max_cpu_num; ++i) {
1353                 int siblings;
1354
1355                 if (cpu_is_not_present(i)) {
1356                         WARNING("cpu%d NOT PRESENT", i);
1357                         continue;
1358                 }
1359                 cpus[i].core_id = get_core_id(i);
1360                 if (cpus[i].core_id < 0)
1361                         return cpus[i].core_id;
1362                 if (cpus[i].core_id > max_core_id)
1363                         max_core_id = cpus[i].core_id;
1364
1365                 cpus[i].physical_package_id = get_physical_package_id(i);
1366                 if (cpus[i].physical_package_id < 0)
1367                         return cpus[i].physical_package_id;
1368                 if (cpus[i].physical_package_id > max_package_id)
1369                         max_package_id = cpus[i].physical_package_id;
1370
1371                 siblings = get_num_ht_siblings(i);
1372                 if (siblings < 0)
1373                         return siblings;
1374                 if (siblings > max_siblings)
1375                         max_siblings = siblings;
1376                 DEBUG("cpu %d pkg %d core %d\n",
1377                         i, cpus[i].physical_package_id, cpus[i].core_id);
1378         }
1379         topo.num_cores_per_pkg = max_core_id + 1;
1380         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1381                 max_core_id, topo.num_cores_per_pkg);
1382
1383         topo.num_packages = max_package_id + 1;
1384         DEBUG("max_package_id %d, sizing for %d packages\n",
1385                 max_package_id, topo.num_packages);
1386
1387         topo.num_threads_per_core = max_siblings;
1388         DEBUG("max_siblings %d\n", max_siblings);
1389
1390         free(cpus);
1391         return 0;
1392 }
1393
1394 static int
1395 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1396 {
1397         int i;
1398
1399         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1400                 topo.num_packages, sizeof(struct thread_data));
1401         if (*t == NULL)
1402                 goto error;
1403
1404         for (i = 0; i < topo.num_threads_per_core *
1405                 topo.num_cores_per_pkg * topo.num_packages; i++)
1406                 (*t)[i].cpu_id = -1;
1407
1408         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1409                 sizeof(struct core_data));
1410         if (*c == NULL)
1411                 goto error;
1412
1413         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1414                 (*c)[i].core_id = -1;
1415
1416         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1417         if (*p == NULL)
1418                 goto error;
1419
1420         for (i = 0; i < topo.num_packages; i++)
1421                 (*p)[i].package_id = i;
1422
1423         return 0;
1424 error:
1425         ERROR("calloc counters");
1426         return -ERR_CALLOC;
1427 }
1428 /*
1429  * init_counter()
1430  *
1431  * set cpu_id, core_num, pkg_num
1432  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1433  *
1434  * increment topo.num_cores when 1st core in pkg seen
1435  */
1436 static int
1437 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1438         struct pkg_data *pkg_base, int thread_num, int core_num,
1439         int pkg_num, int cpu_id)
1440 {
1441         int ret;
1442         struct thread_data *t;
1443         struct core_data *c;
1444         struct pkg_data *p;
1445
1446         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1447         c = GET_CORE(core_base, core_num, pkg_num);
1448         p = GET_PKG(pkg_base, pkg_num);
1449
1450         t->cpu_id = cpu_id;
1451         if (thread_num == 0) {
1452                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1453                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1454                         return ret;
1455                 } else if (ret != 0) {
1456                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1457                 }
1458         }
1459
1460         c->core_id = core_num;
1461         p->package_id = pkg_num;
1462
1463         return 0;
1464 }
1465
1466
1467 static int
1468 initialize_counters(int cpu_id)
1469 {
1470         int my_thread_id, my_core_id, my_package_id;
1471         int ret;
1472
1473         my_package_id = get_physical_package_id(cpu_id);
1474         if (my_package_id < 0)
1475                 return my_package_id;
1476         my_core_id = get_core_id(cpu_id);
1477         if (my_core_id < 0)
1478                 return my_core_id;
1479
1480         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1481                 return ret;
1482         } else if (ret != 0) {
1483                 my_thread_id = 0;
1484                 topo.num_cores++;
1485         } else {
1486                 my_thread_id = 1;
1487         }
1488
1489         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1490         if (ret < 0)
1491                 return ret;
1492         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1493         if (ret < 0)
1494                 return ret;
1495         ret = init_counter(DELTA_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1496         if (ret < 0)
1497                 return ret;
1498         return 0;
1499 }
1500
1501 #define DO_OR_GOTO_ERR(something) \
1502 do {                         \
1503         ret = (something);     \
1504         if (ret < 0)         \
1505                 goto err;    \
1506 } while (0)
1507
1508 static int setup_all_buffers(void)
1509 {
1510         int ret;
1511
1512         DO_OR_GOTO_ERR(topology_probe());
1513         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1514         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1515         DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1516         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1517         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1518         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1519
1520         allocated = 1;
1521         return 0;
1522 err:
1523         free_all_buffers();
1524         return ret;
1525 }
1526
1527 static int
1528 turbostat_init(void)
1529 {
1530         int ret;
1531
1532         DO_OR_GOTO_ERR(check_super_user());
1533         DO_OR_GOTO_ERR(probe_cpu());
1534         DO_OR_GOTO_ERR(check_dev_msr());
1535         DO_OR_GOTO_ERR(setup_all_buffers());
1536
1537         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1538
1539         return 0;
1540 err:
1541         free_all_buffers();
1542         return ret;
1543 }
1544
1545 static const char *config_keys[] =
1546 {
1547         "TCCActivationTemp",
1548 };
1549 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1550
1551 static int
1552 turbostat_config(const char *key, const char *value)
1553 {
1554         long unsigned int tmp_val;
1555         char *end;
1556
1557         if (strcasecmp("TCCActivationTemp", key) == 0) {
1558                 tmp_val = strtoul(value, &end, 0);
1559                 if (*end != '\0' || tmp_val > UINT_MAX)
1560                         return -1;
1561                 tcc_activation_temp = (unsigned int) tmp_val;
1562         } else {
1563                 return -1;
1564         }
1565         return 0;
1566 }
1567
1568 void module_register(void);
1569 void module_register(void)
1570 {
1571         plugin_register_init(PLUGIN_NAME, turbostat_init);
1572         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1573 }