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 static const char *proc_stat = "/proc/stat";
  64 static unsigned int skip_c0;
  65 static unsigned int skip_c1;
  66 static unsigned int do_core_cstate;
  67 static unsigned int do_pkg_cstate;
  68 static unsigned int do_rapl;
  69 static unsigned int do_dts;
  70 static unsigned int do_ptm;
  71 static unsigned int tcc_activation_temp;
  72 static unsigned int tcc_activation_temp_override;
  73 static double rapl_energy_units;
  74
  75 #define RAPL_PKG                (1 << 0)
  76                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  77                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  78 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  79                                         /* 0x613 MSR_PKG_PERF_STATUS */
  80 #define RAPL_PKG_POWER_INFO     (1 << 2)
  81                                         /* 0x614 MSR_PKG_POWER_INFO */
  82
  83 #define RAPL_DRAM               (1 << 3)
  84                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
  85                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
  86                                         /* 0x61c MSR_DRAM_POWER_INFO */
  87 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
  88                                         /* 0x61b MSR_DRAM_PERF_STATUS */
  89
  90 #define RAPL_CORES              (1 << 5)
  91                                         /* 0x638 MSR_PP0_POWER_LIMIT */
  92                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
  93 #define RAPL_CORE_POLICY        (1 << 6)
  94                                         /* 0x63a MSR_PP0_POLICY */
  95
  96
  97 #define RAPL_GFX                (1 << 7)
  98                                         /* 0x640 MSR_PP1_POWER_LIMIT */
  99                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 100                                         /* 0x642 MSR_PP1_POLICY */
 101 #define TJMAX_DEFAULT   100
 102
 103 int aperf_mperf_unstable;
 104 int backwards_count;
 105
 106 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 107 size_t cpu_present_setsize, cpu_affinity_setsize;
 108
 109 struct thread_data {
 110         unsigned long long tsc;
 111         unsigned long long aperf;
 112         unsigned long long mperf;
 113         unsigned long long c1;
 114         unsigned int smi_count;
 115         unsigned int cpu_id;
 116         unsigned int flags;
 117 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 118 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 119 } *thread_even, *thread_odd;
 120
 121 struct core_data {
 122         unsigned long long c3;
 123         unsigned long long c6;
 124         unsigned long long c7;
 125         unsigned int core_temp_c;
 126         unsigned int core_id;
 127 } *core_even, *core_odd;
 128
 129 struct pkg_data {
 130         unsigned long long pc2;
 131         unsigned long long pc3;
 132         unsigned long long pc6;
 133         unsigned long long pc7;
 134         unsigned long long pc8;
 135         unsigned long long pc9;
 136         unsigned long long pc10;
 137         unsigned int package_id;
 138         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 139         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 140         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 141         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 142         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 143         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 144         unsigned int pkg_temp_c;
 145
 146 } *package_even, *package_odd;
 147
 148 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 149 #define EVEN_COUNTERS thread_even, core_even, package_even
 150 static _Bool is_even = 1;
 151
 152 static _Bool allocated = 0;
 153 static _Bool initialized = 0;
 154
 155 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 156         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 157                 topo.num_threads_per_core + \
 158                 (core_no) * topo.num_threads_per_core + (thread_no))
 159 #define GET_CORE(core_base, core_no, pkg_no) \
 160         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 161 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 162
 163 struct topo_params {
 164         int num_packages;
 165         int num_cpus;
 166         int num_cores;
 167         int max_cpu_num;
 168         int num_cores_per_pkg;
 169         int num_threads_per_core;
 170 } topo;
 171
 172 struct timeval tv_even, tv_odd, tv_delta;
 173
 174 enum return_values {
 175         OK = 0,
 176         ERR_CPU_MIGRATE,
 177         ERR_MSR_IA32_APERF,
 178         ERR_MSR_IA32_MPERF,
 179         ERR_MSR_SMI_COUNT,
 180         ERR_MSR_CORE_C3_RESIDENCY,
 181         ERR_MSR_CORE_C6_RESIDENCY,
 182         ERR_MSR_CORE_C7_RESIDENCY,
 183         ERR_MSR_IA32_THERM_STATUS,
 184         ERR_MSR_PKG_C3_RESIDENCY,
 185         ERR_MSR_PKG_C6_RESIDENCY,
 186         ERR_MSR_PKG_C2_RESIDENCY,
 187         ERR_MSR_PKG_C7_RESIDENCY,
 188         ERR_MSR_PKG_C8_RESIDENCY,
 189         ERR_MSR_PKG_C9_RESIDENCY,
 190         ERR_MSR_PKG_C10_RESIDENCY,
 191         ERR_MSR_PKG_ENERGY_STATUS,
 192         ERR_MSR_PKG_POWER_INFO,
 193         ERR_MSR_PP0_ENERGY_STATUS,
 194         ERR_MSR_DRAM_ENERGY_STATUS,
 195         ERR_MSR_PP1_ENERGY_STATUS,
 196         ERR_MSR_PKG_PERF_STATUS,
 197         ERR_MSR_DRAM_PERF_STATUS,
 198         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 199         ERR_MSR_IA32_TSC,
 200         ERR_CPU_NOT_PRESENT,
 201         ERR_NO_MSR,
 202         ERR_CANT_OPEN_MSR,
 203         ERR_CANT_OPEN_FILE,
 204         ERR_CANT_READ_NUMBER,
 205         ERR_CANT_READ_PROC_STAT,
 206         ERR_NO_INVARIANT_TSC,
 207         ERR_NO_APERF,
 208         ERR_CALLOC,
 209         ERR_CPU_ALLOC,
 210         ERR_NOT_ROOT,
 211         UNSUPPORTED_CPU,
 212 };
 213
 214 static int setup_all_buffers(void);
 215
 216 static int
 217 cpu_is_not_present(int cpu)
 218 {
 219         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 220 }
 221 /*
 222  * run func(thread, core, package) in topology order
 223  * skip non-present cpus
 224  */
 225
 226 static int __attribute__((warn_unused_result))
 227 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 228         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 229 {
 230         int retval, pkg_no, core_no, thread_no;
 231
 232         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 233                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 234                         for (thread_no = 0; thread_no <
 235                                 topo.num_threads_per_core; ++thread_no) {
 236                                 struct thread_data *t;
 237                                 struct core_data *c;
 238                                 struct pkg_data *p;
 239
 240                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 241
 242                                 if (cpu_is_not_present(t->cpu_id))
 243                                         continue;
 244
 245                                 c = GET_CORE(core_base, core_no, pkg_no);
 246                                 p = GET_PKG(pkg_base, pkg_no);
 247
 248                                 retval = func(t, c, p);
 249                                 if (retval)
 250                                         return retval;
 251                         }
 252                 }
 253         }
 254         return 0;
 255 }
 256
 257 static int __attribute__((warn_unused_result))
 258 cpu_migrate(int cpu)
 259 {
 260         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 261         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 262         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 263                 return -ERR_CPU_MIGRATE;
 264         else
 265                 return 0;
 266 }
 267
 268 static int __attribute__((warn_unused_result))
 269 open_msr(int cpu)
 270 {
 271         char pathname[32];
 272         int fd;
 273
 274         /* FIXME: Do we really need this, why? */
 275         if (cpu_migrate(cpu)) {
 276                 ERROR("Could not migrate to CPU %d", cpu);
 277                 return -ERR_CPU_MIGRATE;
 278         }
 279
 280         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 281         fd = open(pathname, O_RDONLY);
 282         if (fd < 0)
 283                 return -ERR_CANT_OPEN_MSR;
 284         return fd;
 285 }
 286
 287 static int __attribute__((warn_unused_result))
 288 read_msr(int fd, off_t offset, unsigned long long *msr)
 289 {
 290         ssize_t retval;
 291
 292         retval = pread(fd, msr, sizeof *msr, offset);
 293
 294         if (retval != sizeof *msr) {
 295                 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
 296                 return -1;
 297         }
 298         return 0;
 299 }
 300
 301 static int __attribute__((warn_unused_result))
 302 get_msr(int cpu, off_t offset, unsigned long long *msr)
 303 {
 304         ssize_t retval;
 305         int fd;
 306
 307         fd = open_msr(cpu);
 308         if (fd < 0)
 309                 return fd;
 310         retval = read_msr(fd, offset, msr);
 311         close(fd);
 312         return retval;
 313 }
 314
 315 #define DELTA_WRAP32(new, old)                  \
 316         if (new > old) {                        \
 317                 old = new - old;                \
 318         } else {                                \
 319                 old = 0x100000000 + new - old;  \
 320         }
 321
 322 static void
 323 delta_package(struct pkg_data *new, struct pkg_data *old)
 324 {
 325         old->pc2 = new->pc2 - old->pc2;
 326         old->pc3 = new->pc3 - old->pc3;
 327         old->pc6 = new->pc6 - old->pc6;
 328         old->pc7 = new->pc7 - old->pc7;
 329         old->pc8 = new->pc8 - old->pc8;
 330         old->pc9 = new->pc9 - old->pc9;
 331         old->pc10 = new->pc10 - old->pc10;
 332         old->pkg_temp_c = new->pkg_temp_c;
 333
 334         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 335         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 336         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 337         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 338         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 339         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 340 }
 341
 342 static void
 343 delta_core(struct core_data *new, struct core_data *old)
 344 {
 345         old->c3 = new->c3 - old->c3;
 346         old->c6 = new->c6 - old->c6;
 347         old->c7 = new->c7 - old->c7;
 348         old->core_temp_c = new->core_temp_c;
 349 }
 350
 351 /*
 352  * old = new - old
 353  */
 354 static int __attribute__((warn_unused_result))
 355 delta_thread(struct thread_data *new, struct thread_data *old,
 356         struct core_data *core_delta)
 357 {
 358         old->tsc = new->tsc - old->tsc;
 359
 360         /* check for TSC < 1 Mcycles over interval */
 361         if (old->tsc < (1000 * 1000)) {
 362                 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
 363                 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
 364                 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
 365                 return -1;
 366         }
 367
 368         old->c1 = new->c1 - old->c1;
 369
 370         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 371                 old->aperf = new->aperf - old->aperf;
 372                 old->mperf = new->mperf - old->mperf;
 373         } else {
 374
 375                 if (!aperf_mperf_unstable) {
 376                         WARNING(" APERF or MPERF went backwards * ");
 377                         WARNING("* Frequency results do not cover entire interval *");
 378                         WARNING("* fix this by running Linux-2.6.30 or later *");
 379
 380                         aperf_mperf_unstable = 1;
 381                 }
 382                 /*
 383                  * mperf delta is likely a huge "positive" number
 384                  * can not use it for calculating c0 time
 385                  */
 386                 skip_c0 = 1;
 387                 skip_c1 = 1;
 388         }
 389
 390
 391         /*
 392          * As counter collection is not atomic,
 393          * it is possible for mperf's non-halted cycles + idle states
 394          * to exceed TSC's all cycles: show c1 = 0% in that case.
 395          */
 396         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 397                 old->c1 = 0;
 398         else {
 399                 /* normal case, derive c1 */
 400                 old->c1 = old->tsc - old->mperf - core_delta->c3
 401                         - core_delta->c6 - core_delta->c7;
 402         }
 403
 404         if (old->mperf == 0) {
 405                 WARNING("cpu%d MPERF 0!", old->cpu_id);
 406                 old->mperf = 1; /* divide by 0 protection */
 407         }
 408
 409         old->smi_count = new->smi_count - old->smi_count;
 410
 411         return 0;
 412 }
 413
 414 static int __attribute__((warn_unused_result))
 415 delta_cpu(struct thread_data *t, struct core_data *c,
 416         struct pkg_data *p, struct thread_data *t2,
 417         struct core_data *c2, struct pkg_data *p2)
 418 {
 419         int ret;
 420
 421         /* calculate core delta only for 1st thread in core */
 422         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 423                 delta_core(c, c2);
 424
 425         /* always calculate thread delta */
 426         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 427         if (ret != 0)
 428                 return ret;
 429
 430         /* calculate package delta only for 1st core in package */
 431         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 432                 delta_package(p, p2);
 433
 434         return 0;
 435 }
 436
 437 /*
 438  * get_counters(...)
 439  * migrate to cpu
 440  * acquire and record local counters for that cpu
 441  */
 442 static int __attribute__((warn_unused_result))
 443 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 444 {
 445         int cpu = t->cpu_id;
 446         unsigned long long msr;
 447         int msr_fd;
 448         int retval = 0;
 449
 450         msr_fd = open_msr(cpu);
 451         if (msr_fd < 0)
 452                 return msr_fd;
 453
 454 #define READ_MSR(msr, dst)                      \
 455 do {                                            \
 456         if (read_msr(msr_fd, msr, dst)) {       \
 457                 retval = -ERR_##msr;            \
 458                 goto out;                       \
 459         }                                       \
 460 } while (0)
 461
 462         READ_MSR(MSR_IA32_TSC, &t->tsc);
 463
 464         READ_MSR(MSR_IA32_APERF, &t->aperf);
 465         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 466
 467         READ_MSR(MSR_SMI_COUNT, &msr);
 468         t->smi_count = msr & 0xFFFFFFFF;
 469
 470         /* collect core counters only for 1st thread in core */
 471         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 472                 retval = 0;
 473                 goto out;
 474         }
 475
 476         if (do_core_cstate & (1 << 3))
 477                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 478         if (do_core_cstate & (1 << 6))
 479                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 480         if (do_core_cstate & (1 << 7))
 481                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 482
 483         if (do_dts) {
 484                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 485                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 486         }
 487
 488         /* collect package counters only for 1st core in package */
 489         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 490                 retval = 0;
 491                 goto out;
 492         }
 493
 494         if (do_pkg_cstate & (1 << 2))
 495                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 496         if (do_pkg_cstate & (1 << 3))
 497                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 498         if (do_pkg_cstate & (1 << 6))
 499                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 500         if (do_pkg_cstate & (1 << 7))
 501                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 502         if (do_pkg_cstate & (1 << 8))
 503                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 504         if (do_pkg_cstate & (1 << 9))
 505                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 506         if (do_pkg_cstate & (1 << 10))
 507                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 508
 509         if (do_rapl & RAPL_PKG) {
 510                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 511                 p->energy_pkg = msr & 0xFFFFFFFF;
 512         }
 513         if (do_rapl & RAPL_CORES) {
 514                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 515                 p->energy_cores = msr & 0xFFFFFFFF;
 516         }
 517         if (do_rapl & RAPL_DRAM) {
 518                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 519                 p->energy_dram = msr & 0xFFFFFFFF;
 520         }
 521         if (do_rapl & RAPL_GFX) {
 522                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 523                 p->energy_gfx = msr & 0xFFFFFFFF;
 524         }
 525         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 526                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 527                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 528         }
 529         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 530                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 531                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 532         }
 533         if (do_ptm) {
 534                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 535                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 536         }
 537
 538 out:
 539         close(msr_fd);
 540         return retval;
 541 }
 542
 543 static void
 544 free_all_buffers(void)
 545 {
 546         allocated = 0;
 547         initialized = 0;
 548
 549         CPU_FREE(cpu_present_set);
 550         cpu_present_set = NULL;
 551         cpu_present_set = 0;
 552
 553         CPU_FREE(cpu_affinity_set);
 554         cpu_affinity_set = NULL;
 555         cpu_affinity_setsize = 0;
 556
 557         free(thread_even);
 558         free(core_even);
 559         free(package_even);
 560
 561         thread_even = NULL;
 562         core_even = NULL;
 563         package_even = NULL;
 564
 565         free(thread_odd);
 566         free(core_odd);
 567         free(package_odd);
 568
 569         thread_odd = NULL;
 570         core_odd = NULL;
 571         package_odd = NULL;
 572 }
 573
 574 /*
 575  * Parse a file containing a single int.
 576  */
 577 static int __attribute__ ((format(printf,1,2)))
 578 parse_int_file(const char *fmt, ...)
 579 {
 580         va_list args;
 581         char path[PATH_MAX];
 582         FILE *filep;
 583         int value;
 584
 585         va_start(args, fmt);
 586         vsnprintf(path, sizeof(path), fmt, args);
 587         va_end(args);
 588         filep = fopen(path, "r");
 589         if (!filep) {
 590                 ERROR("%s: open failed", path);
 591                 return -ERR_CANT_OPEN_FILE;
 592         }
 593         if (fscanf(filep, "%d", &value) != 1) {
 594                 ERROR("%s: failed to parse number from file", path);
 595                 return -ERR_CANT_READ_NUMBER;
 596         }
 597         fclose(filep);
 598         return value;
 599 }
 600
 601 /*
 602  * cpu_is_first_sibling_in_core(cpu)
 603  * return 1 if given CPU is 1st HT sibling in the core
 604  */
 605 static int
 606 cpu_is_first_sibling_in_core(int cpu)
 607 {
 608         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 609 }
 610
 611 /*
 612  * cpu_is_first_core_in_package(cpu)
 613  * return 1 if given CPU is 1st core in package
 614  */
 615 static int
 616 cpu_is_first_core_in_package(int cpu)
 617 {
 618         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 619 }
 620
 621 static int
 622 get_physical_package_id(int cpu)
 623 {
 624         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 625 }
 626
 627 static int
 628 get_core_id(int cpu)
 629 {
 630         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 631 }
 632
 633 static int
 634 get_num_ht_siblings(int cpu)
 635 {
 636         char path[80];
 637         FILE *filep;
 638         int sib1, sib2;
 639         int matches;
 640         char character;
 641
 642         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 643         filep = fopen(path, "r");
 644         if (!filep) {
 645                 ERROR("%s: open failed", path);
 646                 return -ERR_CANT_OPEN_FILE;
 647         }
 648         /*
 649          * file format:
 650          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 651          * otherwinse 1 sibling (self).
 652          */
 653         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 654
 655         fclose(filep);
 656
 657         if (matches == 3)
 658                 return 2;
 659         else
 660                 return 1;
 661 }
 662
 663 /*
 664  * run func(thread, core, package) in topology order
 665  * skip non-present cpus
 666  */
 667
 668
 669 static int __attribute__((warn_unused_result))
 670 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 671         struct pkg_data *, struct thread_data *, struct core_data *,
 672         struct pkg_data *), struct thread_data *thread_base,
 673         struct core_data *core_base, struct pkg_data *pkg_base,
 674         struct thread_data *thread_base2, struct core_data *core_base2,
 675         struct pkg_data *pkg_base2)
 676 {
 677         int retval, pkg_no, core_no, thread_no;
 678
 679         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 680                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 681                         for (thread_no = 0; thread_no <
 682                                 topo.num_threads_per_core; ++thread_no) {
 683                                 struct thread_data *t, *t2;
 684                                 struct core_data *c, *c2;
 685                                 struct pkg_data *p, *p2;
 686
 687                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 688
 689                                 if (cpu_is_not_present(t->cpu_id))
 690                                         continue;
 691
 692                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 693
 694                                 c = GET_CORE(core_base, core_no, pkg_no);
 695                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 696
 697                                 p = GET_PKG(pkg_base, pkg_no);
 698                                 p2 = GET_PKG(pkg_base2, pkg_no);
 699
 700                                 retval = func(t, c, p, t2, c2, p2);
 701                                 if (retval)
 702                                         return retval;
 703                         }
 704                 }
 705         }
 706         return 0;
 707 }
 708
 709 /*
 710  * run func(cpu) on every cpu in /proc/stat
 711  * return max_cpu number
 712  */
 713 static int __attribute__((warn_unused_result))
 714 for_all_proc_cpus(int (func)(int))
 715 {
 716         FILE *fp;
 717         int cpu_num;
 718         int retval;
 719
 720         fp = fopen(proc_stat, "r");
 721         if (!fp) {
 722                 ERROR("%s: open failed", proc_stat);
 723                 return -ERR_CANT_OPEN_FILE;
 724         }
 725
 726         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 727         if (retval != 0) {
 728                 ERROR("%s: failed to parse format", proc_stat);
 729                 return -ERR_CANT_READ_PROC_STAT;
 730         }
 731
 732         while (1) {
 733                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 734                 if (retval != 1)
 735                         break;
 736
 737                 retval = func(cpu_num);
 738                 if (retval) {
 739                         fclose(fp);
 740                         return(retval);
 741                 }
 742         }
 743         fclose(fp);
 744         return 0;
 745 }
 746
 747 /*
 748  * count_cpus()
 749  * remember the last one seen, it will be the max
 750  */
 751 static int
 752 count_cpus(int cpu)
 753 {
 754         if (topo.max_cpu_num < cpu)
 755                 topo.max_cpu_num = cpu;
 756
 757         topo.num_cpus += 1;
 758         return 0;
 759 }
 760 static int
 761 mark_cpu_present(int cpu)
 762 {
 763         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 764         return 0;
 765 }
 766
 767
 768 static void
 769 turbostat_submit (const char *plugin_instance,
 770         const char *type, const char *type_instance,
 771         gauge_t value)
 772 {
 773         value_list_t vl = VALUE_LIST_INIT;
 774         value_t v;
 775
 776         v.gauge = value;
 777         vl.values = &v;
 778         vl.values_len = 1;
 779         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 780         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 781         if (plugin_instance != NULL)
 782                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 783         sstrncpy (vl.type, type, sizeof (vl.type));
 784         if (type_instance != NULL)
 785                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 786
 787         plugin_dispatch_values (&vl);
 788 }
 789
 790 /*
 791  * column formatting convention & formats
 792  * package: "pk" 2 columns %2d
 793  * core: "cor" 3 columns %3d
 794  * CPU: "CPU" 3 columns %3d
 795  * Pkg_W: %6.2
 796  * Cor_W: %6.2
 797  * GFX_W: %5.2
 798  * RAM_W: %5.2
 799  * GHz: "GHz" 3 columns %3.2
 800  * TSC: "TSC" 3 columns %3.2
 801  * SMI: "SMI" 4 columns %4d
 802  * percentage " %pc3" %6.2
 803  * Perf Status percentage: %5.2
 804  * "CTMP" 4 columns %4d
 805  */
 806 #define NAME_LEN 12
 807 static int
 808 submit_counters(struct thread_data *t, struct core_data *c,
 809         struct pkg_data *p)
 810 {
 811         char name[NAME_LEN];
 812         double interval_float;
 813
 814         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 815
 816         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 817
 818         if (!skip_c0)
 819                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 820         if (!skip_c1)
 821                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 822
 823         /* GHz */
 824         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 825                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 826
 827         /* SMI */
 828         turbostat_submit(NULL, "current", name, t->smi_count);
 829
 830         /* print per-core data only for 1st thread in core */
 831         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 832                 goto done;
 833
 834         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 835
 836         if (do_core_cstate & (1 << 3))
 837                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 838         if (do_core_cstate & (1 << 6))
 839                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 840         if (do_core_cstate & (1 << 7))
 841                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 842
 843         if (do_dts)
 844                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 845
 846         /* print per-package data only for 1st core in package */
 847         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 848                 goto done;
 849
 850         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 851
 852         if (do_ptm)
 853                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 854
 855         if (do_pkg_cstate & (1 << 2))
 856                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 857         if (do_pkg_cstate & (1 << 3))
 858                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 859         if (do_pkg_cstate & (1 << 6))
 860                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 861         if (do_pkg_cstate & (1 << 7))
 862                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 863         if (do_pkg_cstate & (1 << 8))
 864                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 865         if (do_pkg_cstate & (1 << 9))
 866                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 867         if (do_pkg_cstate & (1 << 10))
 868                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 869
 870         if (do_rapl) {
 871                 if (do_rapl & RAPL_PKG)
 872                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 873                 if (do_rapl & RAPL_CORES)
 874                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 875                 if (do_rapl & RAPL_GFX)
 876                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 877                 if (do_rapl & RAPL_DRAM)
 878                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 879         }
 880 done:
 881         return 0;
 882 }
 883
 884 static int
 885 turbostat_read(user_data_t * not_used)
 886 {
 887         int ret;
 888
 889         if (!allocated) {
 890                 if ((ret = setup_all_buffers()) < 0)
 891                         return ret;
 892         }
 893
 894         if (for_all_proc_cpus(cpu_is_not_present)) {
 895                 free_all_buffers();
 896                 if ((ret = setup_all_buffers()) < 0)
 897                         return ret;
 898                 if (for_all_proc_cpus(cpu_is_not_present))
 899                         return -ERR_CPU_NOT_PRESENT;
 900         }
 901
 902         if (!initialized) {
 903                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 904                         return ret;
 905                 gettimeofday(&tv_even, (struct timezone *)NULL);
 906                 is_even = 1;
 907                 initialized = 1;
 908                 return 0;
 909         }
 910
 911         if (is_even) {
 912                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 913                         return ret;
 914                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 915                 is_even = 0;
 916                 timersub(&tv_odd, &tv_even, &tv_delta);
 917                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 918                         return ret;
 919                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 920                         return ret;
 921         } else {
 922                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 923                         return ret;
 924                 gettimeofday(&tv_even, (struct timezone *)NULL);
 925                 is_even = 1;
 926                 timersub(&tv_even, &tv_odd, &tv_delta);
 927                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 928                         return ret;
 929                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 930                         return ret;
 931         }
 932         return 0;
 933 }
 934
 935 static int __attribute__((warn_unused_result))
 936 check_dev_msr()
 937 {
 938         struct stat sb;
 939
 940         if (stat("/dev/cpu/0/msr", &sb)) {
 941                 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
 942                 return -ERR_NO_MSR;
 943         }
 944         return 0;
 945 }
 946
 947 static int __attribute__((warn_unused_result))
 948 check_super_user()
 949 {
 950         if (getuid() != 0) {
 951                 ERROR("must be root");
 952                 return -ERR_NOT_ROOT;
 953         }
 954         return 0;
 955 }
 956
 957 /*
 958  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
 959  * the Thermal Control Circuit (TCC) activates.
 960  * This is usually equal to tjMax.
 961  *
 962  * Older processors do not have this MSR, so there we guess,
 963  * but also allow cmdline over-ride with -T.
 964  *
 965  * Several MSR temperature values are in units of degrees-C
 966  * below this value, including the Digital Thermal Sensor (DTS),
 967  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
 968  */
 969 static int __attribute__((warn_unused_result))
 970 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 971 {
 972         unsigned long long msr;
 973         unsigned int target_c_local;
 974
 975         /* tcc_activation_temp is used only for dts or ptm */
 976         if (!(do_dts || do_ptm))
 977                 return 0;
 978
 979         /* this is a per-package concept */
 980         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 981                 return 0;
 982
 983         if (tcc_activation_temp_override != 0) {
 984                 tcc_activation_temp = tcc_activation_temp_override;
 985                 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
 986                         cpu, tcc_activation_temp);
 987                 return 0;
 988         }
 989
 990         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
 991                 goto guess;
 992
 993         target_c_local = (msr >> 16) & 0x7F;
 994
 995         if (target_c_local < 85 || target_c_local > 127)
 996                 goto guess;
 997
 998         tcc_activation_temp = target_c_local;
 999
1000         return 0;
1001
1002 guess:
1003         tcc_activation_temp = TJMAX_DEFAULT;
1004         WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify",
1005                 t->cpu_id, tcc_activation_temp);
1006
1007         return 0;
1008 }
1009
1010 /*
1011  * Identify the functionality of the CPU
1012  */
1013 static int __attribute__((warn_unused_result))
1014 probe_cpu()
1015 {
1016         unsigned int eax, ebx, ecx, edx, max_level;
1017         unsigned int fms, family, model;
1018
1019         /* CPUID(0):
1020          * - EAX: Maximum Input Value for Basic CPUID Information
1021          * - EBX: "Genu" (0x756e6547)
1022          * - EDX: "ineI" (0x49656e69)
1023          * - ECX: "ntel" (0x6c65746e)
1024          */
1025         max_level = ebx = ecx = edx = 0;
1026         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1027         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1028                 ERROR("Unsupported CPU");
1029                 return -UNSUPPORTED_CPU;
1030         }
1031
1032         /* CPUID(1):
1033          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1034          *  + 4-7:   Model ID
1035          *  + 8-11:  Family ID
1036          *  + 12-13: Processor type
1037          *  + 16-19: Extended Model ID
1038          *  + 20-27: Extended Family ID
1039          * - EDX: Feature Information:
1040          *  + 5: Support for MSR read/write operations
1041          */
1042         fms = ebx = ecx = edx = 0;
1043         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1044         family = (fms >> 8) & 0xf;
1045         model = (fms >> 4) & 0xf;
1046         if (family == 0xf)
1047                 family += (fms >> 20) & 0xf;
1048         if (family == 6 || family == 0xf)
1049                 model += ((fms >> 16) & 0xf) << 4;
1050         if (!(edx & (1 << 5))) {
1051                 ERROR("CPUID: no MSR");
1052                 return -ERR_NO_MSR;
1053         }
1054
1055         /*
1056          * CPUID(0x80000000):
1057          * - EAX: Maximum Input Value for Extended Function CPUID Information
1058          *
1059          * This allows us to verify if the CPUID(0x80000007) can be called
1060          *
1061          * This check is valid for both Intel and AMD.
1062          */
1063         max_level = ebx = ecx = edx = 0;
1064         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1065         if (max_level < 0x80000007) {
1066                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1067                 return -ERR_NO_INVARIANT_TSC;
1068         }
1069
1070         /*
1071          * CPUID(0x80000007):
1072          * - EDX:
1073          *  + 8: Invariant TSC available if set
1074          *
1075          * This check is valid for both Intel and AMD
1076          */
1077         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1078         if (!(edx & (1 << 8))) {
1079                 ERROR("No invariant TSC");
1080                 return -ERR_NO_INVARIANT_TSC;
1081         }
1082
1083         /*
1084          * CPUID(6):
1085          * - EAX:
1086          *  + 0: Digital temperature sensor is supported if set
1087          *  + 6: Package thermal management is supported if set
1088          * - ECX:
1089          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1090          *  + 3: The processor supports performance-energy bias preference if set.
1091          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1092          *
1093          * This check is valid for both Intel and AMD
1094          */
1095         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1096         do_dts = eax & (1 << 0);
1097         do_ptm = eax & (1 << 6);
1098         if (!(ecx & (1 << 0))) {
1099                 ERROR("No APERF");
1100                 return -ERR_NO_APERF;
1101         }
1102
1103         /*
1104          * Enable or disable C states depending on the model and family
1105          */
1106         if (family == 6) {
1107                 switch (model) {
1108                 /* Atom (partial) */
1109                 case 0x27:
1110                         do_core_cstate = 0;
1111                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1112                         break;
1113                 /* Silvermont */
1114                 case 0x37: /* BYT */
1115                 case 0x4A:
1116                 case 0x4D: /* AVN */
1117                 case 0x5A:
1118                 case 0x5D:
1119                         do_core_cstate = (1 << 1) | (1 << 6);
1120                         do_pkg_cstate = (1 << 6);
1121                         break;
1122                 /* Nehalem */
1123                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1124                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1125                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1126                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1127                         do_core_cstate = (1 << 3) | (1 << 6);
1128                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1129                         break;
1130                 /* Westmere */
1131                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1132                 case 0x2C: /* Westmere EP - Gulftown */
1133                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1134                         do_core_cstate = (1 << 3) | (1 << 6);
1135                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1136                         break;
1137                 /* Sandy Bridge */
1138                 case 0x2A: /* SNB */
1139                 case 0x2D: /* SNB Xeon */
1140                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1141                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1142                         break;
1143                 /* Ivy Bridge */
1144                 case 0x3A: /* IVB */
1145                 case 0x3E: /* IVB Xeon */
1146                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1147                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1148                         break;
1149                 /* Haswell Bridge */
1150                 case 0x3C: /* HSW */
1151                 case 0x3F: /* HSW */
1152                 case 0x46: /* HSW */
1153                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1154                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1155                         break;
1156                 case 0x45: /* HSW */
1157                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1158                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1159                         break;
1160                 /* Broadwel */
1161                 case 0x4F: /* BDW */
1162                 case 0x56: /* BDX-DE */
1163                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1164                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1165                         break;
1166                 case 0x3D: /* BDW */
1167                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1168                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1169                         break;
1170                 default:
1171                         ERROR("Unsupported CPU");
1172                 }
1173                 switch (model) {
1174                 case 0x2A:
1175                 case 0x3A:
1176                 case 0x3C:
1177                 case 0x45:
1178                 case 0x46:
1179                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1180                         break;
1181                 case 0x3F:
1182                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1183                         break;
1184                 case 0x2D:
1185                 case 0x3E:
1186                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1187                         break;
1188                 case 0x37:
1189                 case 0x4D:
1190                         do_rapl = RAPL_PKG | RAPL_CORES;
1191                         break;
1192                 default:
1193                         do_rapl = 0;
1194                 }
1195         } else {
1196                 ERROR("Unsupported CPU");
1197                 return -UNSUPPORTED_CPU;
1198         }
1199
1200         if (do_rapl) {
1201                 unsigned long msr;
1202                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1203                         return 0;
1204
1205                 if (model == 0x37)
1206                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1207                 else
1208                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1209         }
1210
1211         return 0;
1212 }
1213
1214
1215
1216 static int __attribute__((warn_unused_result))
1217 topology_probe()
1218 {
1219         int i;
1220         int ret;
1221         int max_core_id = 0;
1222         int max_package_id = 0;
1223         int max_siblings = 0;
1224         struct cpu_topology {
1225                 int core_id;
1226                 int physical_package_id;
1227         } *cpus;
1228
1229         /* Initialize num_cpus, max_cpu_num */
1230         topo.num_cpus = 0;
1231         topo.max_cpu_num = 0;
1232         ret = for_all_proc_cpus(count_cpus);
1233         if (ret < 0)
1234                 return ret;
1235
1236         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1237
1238         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1239         if (cpus == NULL) {
1240                 ERROR("calloc cpus");
1241                 return -ERR_CALLOC;
1242         }
1243
1244         /*
1245          * Allocate and initialize cpu_present_set
1246          */
1247         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1248         if (cpu_present_set == NULL) {
1249                 free(cpus);
1250                 ERROR("CPU_ALLOC");
1251                 return -ERR_CPU_ALLOC;
1252         }
1253         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1254         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1255         ret = for_all_proc_cpus(mark_cpu_present);
1256         if (ret < 0) {
1257                 free(cpus);
1258                 return ret;
1259         }
1260
1261         /*
1262          * Allocate and initialize cpu_affinity_set
1263          */
1264         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1265         if (cpu_affinity_set == NULL) {
1266                 free(cpus);
1267                 ERROR("CPU_ALLOC");
1268                 return -ERR_CPU_ALLOC;
1269         }
1270         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1271         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1272
1273
1274         /*
1275          * For online cpus
1276          * find max_core_id, max_package_id
1277          */
1278         for (i = 0; i <= topo.max_cpu_num; ++i) {
1279                 int siblings;
1280
1281                 if (cpu_is_not_present(i)) {
1282                         WARNING("cpu%d NOT PRESENT", i);
1283                         continue;
1284                 }
1285                 cpus[i].core_id = get_core_id(i);
1286                 if (cpus[i].core_id < 0)
1287                         return cpus[i].core_id;
1288                 if (cpus[i].core_id > max_core_id)
1289                         max_core_id = cpus[i].core_id;
1290
1291                 cpus[i].physical_package_id = get_physical_package_id(i);
1292                 if (cpus[i].physical_package_id < 0)
1293                         return cpus[i].physical_package_id;
1294                 if (cpus[i].physical_package_id > max_package_id)
1295                         max_package_id = cpus[i].physical_package_id;
1296
1297                 siblings = get_num_ht_siblings(i);
1298                 if (siblings < 0)
1299                         return siblings;
1300                 if (siblings > max_siblings)
1301                         max_siblings = siblings;
1302                 DEBUG("cpu %d pkg %d core %d\n",
1303                         i, cpus[i].physical_package_id, cpus[i].core_id);
1304         }
1305         topo.num_cores_per_pkg = max_core_id + 1;
1306         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1307                 max_core_id, topo.num_cores_per_pkg);
1308
1309         topo.num_packages = max_package_id + 1;
1310         DEBUG("max_package_id %d, sizing for %d packages\n",
1311                 max_package_id, topo.num_packages);
1312
1313         topo.num_threads_per_core = max_siblings;
1314         DEBUG("max_siblings %d\n", max_siblings);
1315
1316         free(cpus);
1317         return 0;
1318 }
1319
1320 static int
1321 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1322 {
1323         int i;
1324
1325         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1326                 topo.num_packages, sizeof(struct thread_data));
1327         if (*t == NULL)
1328                 goto error;
1329
1330         for (i = 0; i < topo.num_threads_per_core *
1331                 topo.num_cores_per_pkg * topo.num_packages; i++)
1332                 (*t)[i].cpu_id = -1;
1333
1334         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1335                 sizeof(struct core_data));
1336         if (*c == NULL)
1337                 goto error;
1338
1339         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1340                 (*c)[i].core_id = -1;
1341
1342         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1343         if (*p == NULL)
1344                 goto error;
1345
1346         for (i = 0; i < topo.num_packages; i++)
1347                 (*p)[i].package_id = i;
1348
1349         return 0;
1350 error:
1351         ERROR("calloc counters");
1352         return -ERR_CALLOC;
1353 }
1354 /*
1355  * init_counter()
1356  *
1357  * set cpu_id, core_num, pkg_num
1358  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1359  *
1360  * increment topo.num_cores when 1st core in pkg seen
1361  */
1362 static int
1363 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1364         struct pkg_data *pkg_base, int thread_num, int core_num,
1365         int pkg_num, int cpu_id)
1366 {
1367         int ret;
1368         struct thread_data *t;
1369         struct core_data *c;
1370         struct pkg_data *p;
1371
1372         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1373         c = GET_CORE(core_base, core_num, pkg_num);
1374         p = GET_PKG(pkg_base, pkg_num);
1375
1376         t->cpu_id = cpu_id;
1377         if (thread_num == 0) {
1378                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1379                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1380                         return ret;
1381                 } else if (ret != 0) {
1382                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1383                 }
1384         }
1385
1386         c->core_id = core_num;
1387         p->package_id = pkg_num;
1388
1389         return 0;
1390 }
1391
1392
1393 static int
1394 initialize_counters(int cpu_id)
1395 {
1396         int my_thread_id, my_core_id, my_package_id;
1397         int ret;
1398
1399         my_package_id = get_physical_package_id(cpu_id);
1400         if (my_package_id < 0)
1401                 return my_package_id;
1402         my_core_id = get_core_id(cpu_id);
1403         if (my_core_id < 0)
1404                 return my_core_id;
1405
1406         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1407                 return ret;
1408         } else if (ret != 0) {
1409                 my_thread_id = 0;
1410                 topo.num_cores++;
1411         } else {
1412                 my_thread_id = 1;
1413         }
1414
1415         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1416         if (ret < 0)
1417                 return ret;
1418         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1419         if (ret < 0)
1420                 return ret;
1421         return 0;
1422 }
1423
1424 #define DO_OR_GOTO_ERR(something) \
1425 do {                         \
1426         ret = (something);     \
1427         if (ret < 0)         \
1428                 goto err;    \
1429 } while (0)
1430
1431 static int setup_all_buffers(void)
1432 {
1433         int ret;
1434
1435         DO_OR_GOTO_ERR(topology_probe());
1436         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1437         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1438         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1439
1440         allocated = 1;
1441         return 0;
1442 err:
1443         free_all_buffers();
1444         return ret;
1445 }
1446
1447 static int
1448 turbostat_init(void)
1449 {
1450         int ret;
1451
1452         DO_OR_GOTO_ERR(check_super_user());
1453         DO_OR_GOTO_ERR(probe_cpu());
1454         DO_OR_GOTO_ERR(check_dev_msr());
1455         DO_OR_GOTO_ERR(setup_all_buffers());
1456         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1457
1458         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1459
1460         return 0;
1461 err:
1462         free_all_buffers();
1463         return ret;
1464 }
1465
1466 void module_register(void);
1467 void module_register(void)
1468 {
1469         plugin_register_init(PLUGIN_NAME, turbostat_init);
1470 }