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 double rapl_energy_units;
  73
  74 #define RAPL_PKG                (1 << 0)
  75                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  76                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  77 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  78                                         /* 0x613 MSR_PKG_PERF_STATUS */
  79 #define RAPL_PKG_POWER_INFO     (1 << 2)
  80                                         /* 0x614 MSR_PKG_POWER_INFO */
  81
  82 #define RAPL_DRAM               (1 << 3)
  83                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
  84                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
  85                                         /* 0x61c MSR_DRAM_POWER_INFO */
  86 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
  87                                         /* 0x61b MSR_DRAM_PERF_STATUS */
  88
  89 #define RAPL_CORES              (1 << 5)
  90                                         /* 0x638 MSR_PP0_POWER_LIMIT */
  91                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
  92 #define RAPL_CORE_POLICY        (1 << 6)
  93                                         /* 0x63a MSR_PP0_POLICY */
  94
  95
  96 #define RAPL_GFX                (1 << 7)
  97                                         /* 0x640 MSR_PP1_POWER_LIMIT */
  98                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
  99                                         /* 0x642 MSR_PP1_POLICY */
 100 #define TJMAX_DEFAULT   100
 101
 102 int aperf_mperf_unstable;
 103 int backwards_count;
 104
 105 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 106 size_t cpu_present_setsize, cpu_affinity_setsize;
 107
 108 struct thread_data {
 109         unsigned long long tsc;
 110         unsigned long long aperf;
 111         unsigned long long mperf;
 112         unsigned long long c1;
 113         unsigned int smi_count;
 114         unsigned int cpu_id;
 115         unsigned int flags;
 116 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 117 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 118 } *thread_even, *thread_odd;
 119
 120 struct core_data {
 121         unsigned long long c3;
 122         unsigned long long c6;
 123         unsigned long long c7;
 124         unsigned int core_temp_c;
 125         unsigned int core_id;
 126 } *core_even, *core_odd;
 127
 128 struct pkg_data {
 129         unsigned long long pc2;
 130         unsigned long long pc3;
 131         unsigned long long pc6;
 132         unsigned long long pc7;
 133         unsigned long long pc8;
 134         unsigned long long pc9;
 135         unsigned long long pc10;
 136         unsigned int package_id;
 137         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 138         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 139         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 140         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 141         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 142         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 143         unsigned int tcc_activation_temp;
 144         unsigned int pkg_temp_c;
 145 } *package_even, *package_odd;
 146
 147 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 148 #define EVEN_COUNTERS thread_even, core_even, package_even
 149 static _Bool is_even = 1;
 150
 151 static _Bool allocated = 0;
 152 static _Bool initialized = 0;
 153
 154 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 155         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 156                 topo.num_threads_per_core + \
 157                 (core_no) * topo.num_threads_per_core + (thread_no))
 158 #define GET_CORE(core_base, core_no, pkg_no) \
 159         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 160 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 161
 162 struct topo_params {
 163         int num_packages;
 164         int num_cpus;
 165         int num_cores;
 166         int max_cpu_num;
 167         int num_cores_per_pkg;
 168         int num_threads_per_core;
 169 } topo;
 170
 171 struct timeval tv_even, tv_odd, tv_delta;
 172
 173 enum return_values {
 174         OK = 0,
 175         ERR_CPU_MIGRATE,
 176         ERR_MSR_IA32_APERF,
 177         ERR_MSR_IA32_MPERF,
 178         ERR_MSR_SMI_COUNT,
 179         ERR_MSR_CORE_C3_RESIDENCY,
 180         ERR_MSR_CORE_C6_RESIDENCY,
 181         ERR_MSR_CORE_C7_RESIDENCY,
 182         ERR_MSR_IA32_THERM_STATUS,
 183         ERR_MSR_PKG_C3_RESIDENCY,
 184         ERR_MSR_PKG_C6_RESIDENCY,
 185         ERR_MSR_PKG_C2_RESIDENCY,
 186         ERR_MSR_PKG_C7_RESIDENCY,
 187         ERR_MSR_PKG_C8_RESIDENCY,
 188         ERR_MSR_PKG_C9_RESIDENCY,
 189         ERR_MSR_PKG_C10_RESIDENCY,
 190         ERR_MSR_PKG_ENERGY_STATUS,
 191         ERR_MSR_PKG_POWER_INFO,
 192         ERR_MSR_PP0_ENERGY_STATUS,
 193         ERR_MSR_DRAM_ENERGY_STATUS,
 194         ERR_MSR_PP1_ENERGY_STATUS,
 195         ERR_MSR_PKG_PERF_STATUS,
 196         ERR_MSR_DRAM_PERF_STATUS,
 197         ERR_MSR_IA32_PACKAGE_THERM_STATUS,
 198         ERR_MSR_IA32_TSC,
 199         ERR_CPU_NOT_PRESENT,
 200         ERR_NO_MSR,
 201         ERR_CANT_OPEN_MSR,
 202         ERR_CANT_OPEN_FILE,
 203         ERR_CANT_READ_NUMBER,
 204         ERR_CANT_READ_PROC_STAT,
 205         ERR_NO_INVARIANT_TSC,
 206         ERR_NO_APERF,
 207         ERR_CALLOC,
 208         ERR_CPU_ALLOC,
 209         ERR_NOT_ROOT,
 210         UNSUPPORTED_CPU,
 211 };
 212
 213 static int setup_all_buffers(void);
 214
 215 static int
 216 cpu_is_not_present(int cpu)
 217 {
 218         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 219 }
 220 /*
 221  * run func(thread, core, package) in topology order
 222  * skip non-present cpus
 223  */
 224
 225 static int __attribute__((warn_unused_result))
 226 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 227         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 228 {
 229         int retval, pkg_no, core_no, thread_no;
 230
 231         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 232                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 233                         for (thread_no = 0; thread_no <
 234                                 topo.num_threads_per_core; ++thread_no) {
 235                                 struct thread_data *t;
 236                                 struct core_data *c;
 237                                 struct pkg_data *p;
 238
 239                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 240
 241                                 if (cpu_is_not_present(t->cpu_id))
 242                                         continue;
 243
 244                                 c = GET_CORE(core_base, core_no, pkg_no);
 245                                 p = GET_PKG(pkg_base, pkg_no);
 246
 247                                 retval = func(t, c, p);
 248                                 if (retval)
 249                                         return retval;
 250                         }
 251                 }
 252         }
 253         return 0;
 254 }
 255
 256 static int __attribute__((warn_unused_result))
 257 cpu_migrate(int cpu)
 258 {
 259         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 260         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 261         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 262                 return -ERR_CPU_MIGRATE;
 263         else
 264                 return 0;
 265 }
 266
 267 static int __attribute__((warn_unused_result))
 268 open_msr(int cpu)
 269 {
 270         char pathname[32];
 271         int fd;
 272
 273         /* FIXME: Do we really need this, why? */
 274         if (cpu_migrate(cpu)) {
 275                 ERROR("Could not migrate to CPU %d", cpu);
 276                 return -ERR_CPU_MIGRATE;
 277         }
 278
 279         ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
 280         fd = open(pathname, O_RDONLY);
 281         if (fd < 0)
 282                 return -ERR_CANT_OPEN_MSR;
 283         return fd;
 284 }
 285
 286 static int __attribute__((warn_unused_result))
 287 read_msr(int fd, off_t offset, unsigned long long *msr)
 288 {
 289         ssize_t retval;
 290
 291         retval = pread(fd, msr, sizeof *msr, offset);
 292
 293         if (retval != sizeof *msr) {
 294                 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
 295                 return -1;
 296         }
 297         return 0;
 298 }
 299
 300 static int __attribute__((warn_unused_result))
 301 get_msr(int cpu, off_t offset, unsigned long long *msr)
 302 {
 303         ssize_t retval;
 304         int fd;
 305
 306         fd = open_msr(cpu);
 307         if (fd < 0)
 308                 return fd;
 309         retval = read_msr(fd, offset, msr);
 310         close(fd);
 311         return retval;
 312 }
 313
 314 #define DELTA_WRAP32(new, old)                  \
 315         if (new > old) {                        \
 316                 old = new - old;                \
 317         } else {                                \
 318                 old = 0x100000000 + new - old;  \
 319         }
 320
 321 static void
 322 delta_package(struct pkg_data *new, struct pkg_data *old)
 323 {
 324         old->pc2 = new->pc2 - old->pc2;
 325         old->pc3 = new->pc3 - old->pc3;
 326         old->pc6 = new->pc6 - old->pc6;
 327         old->pc7 = new->pc7 - old->pc7;
 328         old->pc8 = new->pc8 - old->pc8;
 329         old->pc9 = new->pc9 - old->pc9;
 330         old->pc10 = new->pc10 - old->pc10;
 331         old->pkg_temp_c = new->pkg_temp_c;
 332
 333         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 334         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 335         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 336         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 337         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 338         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 339 }
 340
 341 static void
 342 delta_core(struct core_data *new, struct core_data *old)
 343 {
 344         old->c3 = new->c3 - old->c3;
 345         old->c6 = new->c6 - old->c6;
 346         old->c7 = new->c7 - old->c7;
 347         old->core_temp_c = new->core_temp_c;
 348 }
 349
 350 /*
 351  * old = new - old
 352  */
 353 static int __attribute__((warn_unused_result))
 354 delta_thread(struct thread_data *new, struct thread_data *old,
 355         struct core_data *core_delta)
 356 {
 357         old->tsc = new->tsc - old->tsc;
 358
 359         /* check for TSC < 1 Mcycles over interval */
 360         if (old->tsc < (1000 * 1000)) {
 361                 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
 362                 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
 363                 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
 364                 return -1;
 365         }
 366
 367         old->c1 = new->c1 - old->c1;
 368
 369         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 370                 old->aperf = new->aperf - old->aperf;
 371                 old->mperf = new->mperf - old->mperf;
 372         } else {
 373
 374                 if (!aperf_mperf_unstable) {
 375                         WARNING(" APERF or MPERF went backwards * ");
 376                         WARNING("* Frequency results do not cover entire interval *");
 377                         WARNING("* fix this by running Linux-2.6.30 or later *");
 378
 379                         aperf_mperf_unstable = 1;
 380                 }
 381                 /*
 382                  * mperf delta is likely a huge "positive" number
 383                  * can not use it for calculating c0 time
 384                  */
 385                 skip_c0 = 1;
 386                 skip_c1 = 1;
 387         }
 388
 389
 390         /*
 391          * As counter collection is not atomic,
 392          * it is possible for mperf's non-halted cycles + idle states
 393          * to exceed TSC's all cycles: show c1 = 0% in that case.
 394          */
 395         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 396                 old->c1 = 0;
 397         else {
 398                 /* normal case, derive c1 */
 399                 old->c1 = old->tsc - old->mperf - core_delta->c3
 400                         - core_delta->c6 - core_delta->c7;
 401         }
 402
 403         if (old->mperf == 0) {
 404                 WARNING("cpu%d MPERF 0!", old->cpu_id);
 405                 old->mperf = 1; /* divide by 0 protection */
 406         }
 407
 408         old->smi_count = new->smi_count - old->smi_count;
 409
 410         return 0;
 411 }
 412
 413 static int __attribute__((warn_unused_result))
 414 delta_cpu(struct thread_data *t, struct core_data *c,
 415         struct pkg_data *p, struct thread_data *t2,
 416         struct core_data *c2, struct pkg_data *p2)
 417 {
 418         int ret;
 419
 420         /* calculate core delta only for 1st thread in core */
 421         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 422                 delta_core(c, c2);
 423
 424         /* always calculate thread delta */
 425         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 426         if (ret != 0)
 427                 return ret;
 428
 429         /* calculate package delta only for 1st core in package */
 430         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 431                 delta_package(p, p2);
 432
 433         return 0;
 434 }
 435
 436 /*
 437  * get_counters(...)
 438  * migrate to cpu
 439  * acquire and record local counters for that cpu
 440  */
 441 static int __attribute__((warn_unused_result))
 442 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 443 {
 444         int cpu = t->cpu_id;
 445         unsigned long long msr;
 446         int msr_fd;
 447         int retval = 0;
 448
 449         msr_fd = open_msr(cpu);
 450         if (msr_fd < 0)
 451                 return msr_fd;
 452
 453 #define READ_MSR(msr, dst)                      \
 454 do {                                            \
 455         if (read_msr(msr_fd, msr, dst)) {       \
 456                 retval = -ERR_##msr;            \
 457                 goto out;                       \
 458         }                                       \
 459 } while (0)
 460
 461         READ_MSR(MSR_IA32_TSC, &t->tsc);
 462
 463         READ_MSR(MSR_IA32_APERF, &t->aperf);
 464         READ_MSR(MSR_IA32_MPERF, &t->mperf);
 465
 466         READ_MSR(MSR_SMI_COUNT, &msr);
 467         t->smi_count = msr & 0xFFFFFFFF;
 468
 469         /* collect core counters only for 1st thread in core */
 470         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
 471                 retval = 0;
 472                 goto out;
 473         }
 474
 475         if (do_core_cstate & (1 << 3))
 476                 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
 477         if (do_core_cstate & (1 << 6))
 478                 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
 479         if (do_core_cstate & (1 << 7))
 480                 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
 481
 482         if (do_dts) {
 483                 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
 484                 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 485         }
 486
 487         /* collect package counters only for 1st core in package */
 488         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
 489                 retval = 0;
 490                 goto out;
 491         }
 492
 493         if (do_pkg_cstate & (1 << 2))
 494                 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
 495         if (do_pkg_cstate & (1 << 3))
 496                 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
 497         if (do_pkg_cstate & (1 << 6))
 498                 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
 499         if (do_pkg_cstate & (1 << 7))
 500                 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
 501         if (do_pkg_cstate & (1 << 8))
 502                 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
 503         if (do_pkg_cstate & (1 << 9))
 504                 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
 505         if (do_pkg_cstate & (1 << 10))
 506                 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
 507
 508         if (do_rapl & RAPL_PKG) {
 509                 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
 510                 p->energy_pkg = msr & 0xFFFFFFFF;
 511         }
 512         if (do_rapl & RAPL_CORES) {
 513                 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
 514                 p->energy_cores = msr & 0xFFFFFFFF;
 515         }
 516         if (do_rapl & RAPL_DRAM) {
 517                 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
 518                 p->energy_dram = msr & 0xFFFFFFFF;
 519         }
 520         if (do_rapl & RAPL_GFX) {
 521                 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
 522                 p->energy_gfx = msr & 0xFFFFFFFF;
 523         }
 524         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 525                 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
 526                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 527         }
 528         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 529                 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
 530                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 531         }
 532         if (do_ptm) {
 533                 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
 534                 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
 535         }
 536
 537 out:
 538         close(msr_fd);
 539         return retval;
 540 }
 541
 542 static void
 543 free_all_buffers(void)
 544 {
 545         allocated = 0;
 546         initialized = 0;
 547
 548         CPU_FREE(cpu_present_set);
 549         cpu_present_set = NULL;
 550         cpu_present_set = 0;
 551
 552         CPU_FREE(cpu_affinity_set);
 553         cpu_affinity_set = NULL;
 554         cpu_affinity_setsize = 0;
 555
 556         free(thread_even);
 557         free(core_even);
 558         free(package_even);
 559
 560         thread_even = NULL;
 561         core_even = NULL;
 562         package_even = NULL;
 563
 564         free(thread_odd);
 565         free(core_odd);
 566         free(package_odd);
 567
 568         thread_odd = NULL;
 569         core_odd = NULL;
 570         package_odd = NULL;
 571 }
 572
 573 /*
 574  * Parse a file containing a single int.
 575  */
 576 static int __attribute__ ((format(printf,1,2)))
 577 parse_int_file(const char *fmt, ...)
 578 {
 579         va_list args;
 580         char path[PATH_MAX];
 581         FILE *filep;
 582         int value;
 583
 584         va_start(args, fmt);
 585         vsnprintf(path, sizeof(path), fmt, args);
 586         va_end(args);
 587         filep = fopen(path, "r");
 588         if (!filep) {
 589                 ERROR("%s: open failed", path);
 590                 return -ERR_CANT_OPEN_FILE;
 591         }
 592         if (fscanf(filep, "%d", &value) != 1) {
 593                 ERROR("%s: failed to parse number from file", path);
 594                 return -ERR_CANT_READ_NUMBER;
 595         }
 596         fclose(filep);
 597         return value;
 598 }
 599
 600 /*
 601  * cpu_is_first_sibling_in_core(cpu)
 602  * return 1 if given CPU is 1st HT sibling in the core
 603  */
 604 static int
 605 cpu_is_first_sibling_in_core(int cpu)
 606 {
 607         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 608 }
 609
 610 /*
 611  * cpu_is_first_core_in_package(cpu)
 612  * return 1 if given CPU is 1st core in package
 613  */
 614 static int
 615 cpu_is_first_core_in_package(int cpu)
 616 {
 617         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 618 }
 619
 620 static int
 621 get_physical_package_id(int cpu)
 622 {
 623         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 624 }
 625
 626 static int
 627 get_core_id(int cpu)
 628 {
 629         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 630 }
 631
 632 static int
 633 get_num_ht_siblings(int cpu)
 634 {
 635         char path[80];
 636         FILE *filep;
 637         int sib1, sib2;
 638         int matches;
 639         char character;
 640
 641         ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 642         filep = fopen(path, "r");
 643         if (!filep) {
 644                 ERROR("%s: open failed", path);
 645                 return -ERR_CANT_OPEN_FILE;
 646         }
 647         /*
 648          * file format:
 649          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 650          * otherwinse 1 sibling (self).
 651          */
 652         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 653
 654         fclose(filep);
 655
 656         if (matches == 3)
 657                 return 2;
 658         else
 659                 return 1;
 660 }
 661
 662 /*
 663  * run func(thread, core, package) in topology order
 664  * skip non-present cpus
 665  */
 666
 667
 668 static int __attribute__((warn_unused_result))
 669 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 670         struct pkg_data *, struct thread_data *, struct core_data *,
 671         struct pkg_data *), struct thread_data *thread_base,
 672         struct core_data *core_base, struct pkg_data *pkg_base,
 673         struct thread_data *thread_base2, struct core_data *core_base2,
 674         struct pkg_data *pkg_base2)
 675 {
 676         int retval, pkg_no, core_no, thread_no;
 677
 678         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 679                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 680                         for (thread_no = 0; thread_no <
 681                                 topo.num_threads_per_core; ++thread_no) {
 682                                 struct thread_data *t, *t2;
 683                                 struct core_data *c, *c2;
 684                                 struct pkg_data *p, *p2;
 685
 686                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 687
 688                                 if (cpu_is_not_present(t->cpu_id))
 689                                         continue;
 690
 691                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 692
 693                                 c = GET_CORE(core_base, core_no, pkg_no);
 694                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 695
 696                                 p = GET_PKG(pkg_base, pkg_no);
 697                                 p2 = GET_PKG(pkg_base2, pkg_no);
 698
 699                                 retval = func(t, c, p, t2, c2, p2);
 700                                 if (retval)
 701                                         return retval;
 702                         }
 703                 }
 704         }
 705         return 0;
 706 }
 707
 708 /*
 709  * run func(cpu) on every cpu in /proc/stat
 710  * return max_cpu number
 711  */
 712 static int __attribute__((warn_unused_result))
 713 for_all_proc_cpus(int (func)(int))
 714 {
 715         FILE *fp;
 716         int cpu_num;
 717         int retval;
 718
 719         fp = fopen(proc_stat, "r");
 720         if (!fp) {
 721                 ERROR("%s: open failed", proc_stat);
 722                 return -ERR_CANT_OPEN_FILE;
 723         }
 724
 725         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 726         if (retval != 0) {
 727                 ERROR("%s: failed to parse format", proc_stat);
 728                 return -ERR_CANT_READ_PROC_STAT;
 729         }
 730
 731         while (1) {
 732                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 733                 if (retval != 1)
 734                         break;
 735
 736                 retval = func(cpu_num);
 737                 if (retval) {
 738                         fclose(fp);
 739                         return(retval);
 740                 }
 741         }
 742         fclose(fp);
 743         return 0;
 744 }
 745
 746 /*
 747  * count_cpus()
 748  * remember the last one seen, it will be the max
 749  */
 750 static int
 751 count_cpus(int cpu)
 752 {
 753         if (topo.max_cpu_num < cpu)
 754                 topo.max_cpu_num = cpu;
 755
 756         topo.num_cpus += 1;
 757         return 0;
 758 }
 759 static int
 760 mark_cpu_present(int cpu)
 761 {
 762         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 763         return 0;
 764 }
 765
 766
 767 static void
 768 turbostat_submit (const char *plugin_instance,
 769         const char *type, const char *type_instance,
 770         gauge_t value)
 771 {
 772         value_list_t vl = VALUE_LIST_INIT;
 773         value_t v;
 774
 775         v.gauge = value;
 776         vl.values = &v;
 777         vl.values_len = 1;
 778         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 779         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 780         if (plugin_instance != NULL)
 781                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 782         sstrncpy (vl.type, type, sizeof (vl.type));
 783         if (type_instance != NULL)
 784                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 785
 786         plugin_dispatch_values (&vl);
 787 }
 788
 789 /*
 790  * column formatting convention & formats
 791  * package: "pk" 2 columns %2d
 792  * core: "cor" 3 columns %3d
 793  * CPU: "CPU" 3 columns %3d
 794  * Pkg_W: %6.2
 795  * Cor_W: %6.2
 796  * GFX_W: %5.2
 797  * RAM_W: %5.2
 798  * GHz: "GHz" 3 columns %3.2
 799  * TSC: "TSC" 3 columns %3.2
 800  * SMI: "SMI" 4 columns %4d
 801  * percentage " %pc3" %6.2
 802  * Perf Status percentage: %5.2
 803  * "CTMP" 4 columns %4d
 804  */
 805 #define NAME_LEN 12
 806 static int
 807 submit_counters(struct thread_data *t, struct core_data *c,
 808         struct pkg_data *p)
 809 {
 810         char name[NAME_LEN];
 811         double interval_float;
 812
 813         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 814
 815         ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
 816
 817         if (!skip_c0)
 818                 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 819         if (!skip_c1)
 820                 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 821
 822         /* GHz */
 823         if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
 824                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
 825
 826         /* SMI */
 827         turbostat_submit(NULL, "current", name, t->smi_count);
 828
 829         /* print per-core data only for 1st thread in core */
 830         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 831                 goto done;
 832
 833         ssnprintf(name, sizeof(name), "core%02d", c->core_id);
 834
 835         if (do_core_cstate & (1 << 3))
 836                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 837         if (do_core_cstate & (1 << 6))
 838                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 839         if (do_core_cstate & (1 << 7))
 840                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 841
 842         if (do_dts)
 843                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 844
 845         /* print per-package data only for 1st core in package */
 846         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 847                 goto done;
 848
 849         ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
 850
 851         if (do_ptm)
 852                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 853
 854         if (do_pkg_cstate & (1 << 2))
 855                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 856         if (do_pkg_cstate & (1 << 3))
 857                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 858         if (do_pkg_cstate & (1 << 6))
 859                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 860         if (do_pkg_cstate & (1 << 7))
 861                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 862         if (do_pkg_cstate & (1 << 8))
 863                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 864         if (do_pkg_cstate & (1 << 9))
 865                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 866         if (do_pkg_cstate & (1 << 10))
 867                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 868
 869         if (do_rapl) {
 870                 if (do_rapl & RAPL_PKG)
 871                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 872                 if (do_rapl & RAPL_CORES)
 873                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 874                 if (do_rapl & RAPL_GFX)
 875                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 876                 if (do_rapl & RAPL_DRAM)
 877                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 878         }
 879 done:
 880         return 0;
 881 }
 882
 883 static int
 884 turbostat_read(user_data_t * not_used)
 885 {
 886         int ret;
 887
 888         if (!allocated) {
 889                 if ((ret = setup_all_buffers()) < 0)
 890                         return ret;
 891         }
 892
 893         if (for_all_proc_cpus(cpu_is_not_present)) {
 894                 free_all_buffers();
 895                 if ((ret = setup_all_buffers()) < 0)
 896                         return ret;
 897                 if (for_all_proc_cpus(cpu_is_not_present))
 898                         return -ERR_CPU_NOT_PRESENT;
 899         }
 900
 901         if (!initialized) {
 902                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 903                         return ret;
 904                 gettimeofday(&tv_even, (struct timezone *)NULL);
 905                 is_even = 1;
 906                 initialized = 1;
 907                 return 0;
 908         }
 909
 910         if (is_even) {
 911                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 912                         return ret;
 913                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 914                 is_even = 0;
 915                 timersub(&tv_odd, &tv_even, &tv_delta);
 916                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 917                         return ret;
 918                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 919                         return ret;
 920         } else {
 921                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 922                         return ret;
 923                 gettimeofday(&tv_even, (struct timezone *)NULL);
 924                 is_even = 1;
 925                 timersub(&tv_even, &tv_odd, &tv_delta);
 926                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 927                         return ret;
 928                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 929                         return ret;
 930         }
 931         return 0;
 932 }
 933
 934 static int __attribute__((warn_unused_result))
 935 check_dev_msr()
 936 {
 937         struct stat sb;
 938
 939         if (stat("/dev/cpu/0/msr", &sb)) {
 940                 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
 941                 return -ERR_NO_MSR;
 942         }
 943         return 0;
 944 }
 945
 946 static int __attribute__((warn_unused_result))
 947 check_super_user()
 948 {
 949         if (getuid() != 0) {
 950                 ERROR("must be root");
 951                 return -ERR_NOT_ROOT;
 952         }
 953         return 0;
 954 }
 955
 956 /*
 957  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
 958  * the Thermal Control Circuit (TCC) activates.
 959  * This is usually equal to tjMax.
 960  *
 961  * Older processors do not have this MSR, so there we guess,
 962  * but also allow cmdline over-ride with -T.
 963  *
 964  * Several MSR temperature values are in units of degrees-C
 965  * below this value, including the Digital Thermal Sensor (DTS),
 966  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
 967  */
 968 static int __attribute__((warn_unused_result))
 969 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 970 {
 971         unsigned long long msr;
 972         unsigned int target_c_local;
 973
 974         /* tcc_activation_temp is used only for dts or ptm */
 975         if (!(do_dts || do_ptm))
 976                 return 0;
 977
 978         /* this is a per-package concept */
 979         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 980                 return 0;
 981
 982         if (tcc_activation_temp != 0) {
 983                 p->tcc_activation_temp = tcc_activation_temp;
 984                 return 0;
 985         }
 986
 987         if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
 988                 goto guess;
 989
 990         target_c_local = (msr >> 16) & 0x7F;
 991
 992         if (target_c_local < 85 || target_c_local > 127)
 993                 goto guess;
 994
 995         p->tcc_activation_temp = target_c_local;
 996
 997         return 0;
 998
 999 guess:
1000         p->tcc_activation_temp = TJMAX_DEFAULT;
1001         WARNING("cpu%d: Guessing tjMax %d C, Please use TCCActivationTemp to specify",
1002                 t->cpu_id, p->tcc_activation_temp);
1003
1004         return 0;
1005 }
1006
1007 /*
1008  * Identify the functionality of the CPU
1009  */
1010 static int __attribute__((warn_unused_result))
1011 probe_cpu()
1012 {
1013         unsigned int eax, ebx, ecx, edx, max_level;
1014         unsigned int fms, family, model;
1015
1016         /* CPUID(0):
1017          * - EAX: Maximum Input Value for Basic CPUID Information
1018          * - EBX: "Genu" (0x756e6547)
1019          * - EDX: "ineI" (0x49656e69)
1020          * - ECX: "ntel" (0x6c65746e)
1021          */
1022         max_level = ebx = ecx = edx = 0;
1023         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1024         if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1025                 ERROR("Unsupported CPU");
1026                 return -UNSUPPORTED_CPU;
1027         }
1028
1029         /* CPUID(1):
1030          * - EAX: Version Information: Type, Family, Model, and Stepping ID
1031          *  + 4-7:   Model ID
1032          *  + 8-11:  Family ID
1033          *  + 12-13: Processor type
1034          *  + 16-19: Extended Model ID
1035          *  + 20-27: Extended Family ID
1036          * - EDX: Feature Information:
1037          *  + 5: Support for MSR read/write operations
1038          */
1039         fms = ebx = ecx = edx = 0;
1040         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1041         family = (fms >> 8) & 0xf;
1042         model = (fms >> 4) & 0xf;
1043         if (family == 0xf)
1044                 family += (fms >> 20) & 0xf;
1045         if (family == 6 || family == 0xf)
1046                 model += ((fms >> 16) & 0xf) << 4;
1047         if (!(edx & (1 << 5))) {
1048                 ERROR("CPUID: no MSR");
1049                 return -ERR_NO_MSR;
1050         }
1051
1052         /*
1053          * CPUID(0x80000000):
1054          * - EAX: Maximum Input Value for Extended Function CPUID Information
1055          *
1056          * This allows us to verify if the CPUID(0x80000007) can be called
1057          *
1058          * This check is valid for both Intel and AMD.
1059          */
1060         max_level = ebx = ecx = edx = 0;
1061         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1062         if (max_level < 0x80000007) {
1063                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1064                 return -ERR_NO_INVARIANT_TSC;
1065         }
1066
1067         /*
1068          * CPUID(0x80000007):
1069          * - EDX:
1070          *  + 8: Invariant TSC available if set
1071          *
1072          * This check is valid for both Intel and AMD
1073          */
1074         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1075         if (!(edx & (1 << 8))) {
1076                 ERROR("No invariant TSC");
1077                 return -ERR_NO_INVARIANT_TSC;
1078         }
1079
1080         /*
1081          * CPUID(6):
1082          * - EAX:
1083          *  + 0: Digital temperature sensor is supported if set
1084          *  + 6: Package thermal management is supported if set
1085          * - ECX:
1086          *  + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1087          *  + 3: The processor supports performance-energy bias preference if set.
1088          *       It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1089          *
1090          * This check is valid for both Intel and AMD
1091          */
1092         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1093         do_dts = eax & (1 << 0);
1094         do_ptm = eax & (1 << 6);
1095         if (!(ecx & (1 << 0))) {
1096                 ERROR("No APERF");
1097                 return -ERR_NO_APERF;
1098         }
1099
1100         /*
1101          * Enable or disable C states depending on the model and family
1102          */
1103         if (family == 6) {
1104                 switch (model) {
1105                 /* Atom (partial) */
1106                 case 0x27:
1107                         do_core_cstate = 0;
1108                         do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1109                         break;
1110                 /* Silvermont */
1111                 case 0x37: /* BYT */
1112                 case 0x4A:
1113                 case 0x4D: /* AVN */
1114                 case 0x5A:
1115                 case 0x5D:
1116                         do_core_cstate = (1 << 1) | (1 << 6);
1117                         do_pkg_cstate = (1 << 6);
1118                         break;
1119                 /* Nehalem */
1120                 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1121                 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1122                 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1123                 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1124                         do_core_cstate = (1 << 3) | (1 << 6);
1125                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1126                         break;
1127                 /* Westmere */
1128                 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1129                 case 0x2C: /* Westmere EP - Gulftown */
1130                 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1131                         do_core_cstate = (1 << 3) | (1 << 6);
1132                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1133                         break;
1134                 /* Sandy Bridge */
1135                 case 0x2A: /* SNB */
1136                 case 0x2D: /* SNB Xeon */
1137                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1138                         do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1139                         break;
1140                 /* Ivy Bridge */
1141                 case 0x3A: /* IVB */
1142                 case 0x3E: /* IVB Xeon */
1143                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1144                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1145                         break;
1146                 /* Haswell Bridge */
1147                 case 0x3C: /* HSW */
1148                 case 0x3F: /* HSW */
1149                 case 0x46: /* HSW */
1150                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1151                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1152                         break;
1153                 case 0x45: /* HSW */
1154                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1155                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1156                         break;
1157                 /* Broadwel */
1158                 case 0x4F: /* BDW */
1159                 case 0x56: /* BDX-DE */
1160                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1161                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1162                         break;
1163                 case 0x3D: /* BDW */
1164                         do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1165                         do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1166                         break;
1167                 default:
1168                         ERROR("Unsupported CPU");
1169                 }
1170                 switch (model) {
1171                 case 0x2A:
1172                 case 0x3A:
1173                 case 0x3C:
1174                 case 0x45:
1175                 case 0x46:
1176                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1177                         break;
1178                 case 0x3F:
1179                         do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1180                         break;
1181                 case 0x2D:
1182                 case 0x3E:
1183                         do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1184                         break;
1185                 case 0x37:
1186                 case 0x4D:
1187                         do_rapl = RAPL_PKG | RAPL_CORES;
1188                         break;
1189                 default:
1190                         do_rapl = 0;
1191                 }
1192         } else {
1193                 ERROR("Unsupported CPU");
1194                 return -UNSUPPORTED_CPU;
1195         }
1196
1197         if (do_rapl) {
1198                 unsigned long msr;
1199                 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1200                         return 0;
1201
1202                 if (model == 0x37)
1203                         rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1204                 else
1205                         rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1206         }
1207
1208         return 0;
1209 }
1210
1211
1212
1213 static int __attribute__((warn_unused_result))
1214 topology_probe()
1215 {
1216         int i;
1217         int ret;
1218         int max_core_id = 0;
1219         int max_package_id = 0;
1220         int max_siblings = 0;
1221         struct cpu_topology {
1222                 int core_id;
1223                 int physical_package_id;
1224         } *cpus;
1225
1226         /* Initialize num_cpus, max_cpu_num */
1227         topo.num_cpus = 0;
1228         topo.max_cpu_num = 0;
1229         ret = for_all_proc_cpus(count_cpus);
1230         if (ret < 0)
1231                 return ret;
1232
1233         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1234
1235         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1236         if (cpus == NULL) {
1237                 ERROR("calloc cpus");
1238                 return -ERR_CALLOC;
1239         }
1240
1241         /*
1242          * Allocate and initialize cpu_present_set
1243          */
1244         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1245         if (cpu_present_set == NULL) {
1246                 free(cpus);
1247                 ERROR("CPU_ALLOC");
1248                 return -ERR_CPU_ALLOC;
1249         }
1250         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1251         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1252         ret = for_all_proc_cpus(mark_cpu_present);
1253         if (ret < 0) {
1254                 free(cpus);
1255                 return ret;
1256         }
1257
1258         /*
1259          * Allocate and initialize cpu_affinity_set
1260          */
1261         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1262         if (cpu_affinity_set == NULL) {
1263                 free(cpus);
1264                 ERROR("CPU_ALLOC");
1265                 return -ERR_CPU_ALLOC;
1266         }
1267         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1268         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1269
1270
1271         /*
1272          * For online cpus
1273          * find max_core_id, max_package_id
1274          */
1275         for (i = 0; i <= topo.max_cpu_num; ++i) {
1276                 int siblings;
1277
1278                 if (cpu_is_not_present(i)) {
1279                         WARNING("cpu%d NOT PRESENT", i);
1280                         continue;
1281                 }
1282                 cpus[i].core_id = get_core_id(i);
1283                 if (cpus[i].core_id < 0)
1284                         return cpus[i].core_id;
1285                 if (cpus[i].core_id > max_core_id)
1286                         max_core_id = cpus[i].core_id;
1287
1288                 cpus[i].physical_package_id = get_physical_package_id(i);
1289                 if (cpus[i].physical_package_id < 0)
1290                         return cpus[i].physical_package_id;
1291                 if (cpus[i].physical_package_id > max_package_id)
1292                         max_package_id = cpus[i].physical_package_id;
1293
1294                 siblings = get_num_ht_siblings(i);
1295                 if (siblings < 0)
1296                         return siblings;
1297                 if (siblings > max_siblings)
1298                         max_siblings = siblings;
1299                 DEBUG("cpu %d pkg %d core %d\n",
1300                         i, cpus[i].physical_package_id, cpus[i].core_id);
1301         }
1302         topo.num_cores_per_pkg = max_core_id + 1;
1303         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1304                 max_core_id, topo.num_cores_per_pkg);
1305
1306         topo.num_packages = max_package_id + 1;
1307         DEBUG("max_package_id %d, sizing for %d packages\n",
1308                 max_package_id, topo.num_packages);
1309
1310         topo.num_threads_per_core = max_siblings;
1311         DEBUG("max_siblings %d\n", max_siblings);
1312
1313         free(cpus);
1314         return 0;
1315 }
1316
1317 static int
1318 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1319 {
1320         int i;
1321
1322         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1323                 topo.num_packages, sizeof(struct thread_data));
1324         if (*t == NULL)
1325                 goto error;
1326
1327         for (i = 0; i < topo.num_threads_per_core *
1328                 topo.num_cores_per_pkg * topo.num_packages; i++)
1329                 (*t)[i].cpu_id = -1;
1330
1331         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1332                 sizeof(struct core_data));
1333         if (*c == NULL)
1334                 goto error;
1335
1336         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1337                 (*c)[i].core_id = -1;
1338
1339         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1340         if (*p == NULL)
1341                 goto error;
1342
1343         for (i = 0; i < topo.num_packages; i++)
1344                 (*p)[i].package_id = i;
1345
1346         return 0;
1347 error:
1348         ERROR("calloc counters");
1349         return -ERR_CALLOC;
1350 }
1351 /*
1352  * init_counter()
1353  *
1354  * set cpu_id, core_num, pkg_num
1355  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1356  *
1357  * increment topo.num_cores when 1st core in pkg seen
1358  */
1359 static int
1360 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1361         struct pkg_data *pkg_base, int thread_num, int core_num,
1362         int pkg_num, int cpu_id)
1363 {
1364         int ret;
1365         struct thread_data *t;
1366         struct core_data *c;
1367         struct pkg_data *p;
1368
1369         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1370         c = GET_CORE(core_base, core_num, pkg_num);
1371         p = GET_PKG(pkg_base, pkg_num);
1372
1373         t->cpu_id = cpu_id;
1374         if (thread_num == 0) {
1375                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1376                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1377                         return ret;
1378                 } else if (ret != 0) {
1379                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1380                 }
1381         }
1382
1383         c->core_id = core_num;
1384         p->package_id = pkg_num;
1385
1386         return 0;
1387 }
1388
1389
1390 static int
1391 initialize_counters(int cpu_id)
1392 {
1393         int my_thread_id, my_core_id, my_package_id;
1394         int ret;
1395
1396         my_package_id = get_physical_package_id(cpu_id);
1397         if (my_package_id < 0)
1398                 return my_package_id;
1399         my_core_id = get_core_id(cpu_id);
1400         if (my_core_id < 0)
1401                 return my_core_id;
1402
1403         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1404                 return ret;
1405         } else if (ret != 0) {
1406                 my_thread_id = 0;
1407                 topo.num_cores++;
1408         } else {
1409                 my_thread_id = 1;
1410         }
1411
1412         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1413         if (ret < 0)
1414                 return ret;
1415         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1416         if (ret < 0)
1417                 return ret;
1418         return 0;
1419 }
1420
1421 #define DO_OR_GOTO_ERR(something) \
1422 do {                         \
1423         ret = (something);     \
1424         if (ret < 0)         \
1425                 goto err;    \
1426 } while (0)
1427
1428 static int setup_all_buffers(void)
1429 {
1430         int ret;
1431
1432         DO_OR_GOTO_ERR(topology_probe());
1433         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1434         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1435         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1436
1437         allocated = 1;
1438         return 0;
1439 err:
1440         free_all_buffers();
1441         return ret;
1442 }
1443
1444 static int
1445 turbostat_init(void)
1446 {
1447         int ret;
1448
1449         DO_OR_GOTO_ERR(check_super_user());
1450         DO_OR_GOTO_ERR(probe_cpu());
1451         DO_OR_GOTO_ERR(check_dev_msr());
1452         DO_OR_GOTO_ERR(setup_all_buffers());
1453         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1454         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1455
1456         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1457
1458         return 0;
1459 err:
1460         free_all_buffers();
1461         return ret;
1462 }
1463
1464 static const char *config_keys[] =
1465 {
1466         "TCCActivationTemp",
1467 };
1468 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1469
1470 static int
1471 turbostat_config(const char *key, const char *value)
1472 {
1473         long unsigned int tmp_val;
1474         char *end;
1475
1476         if (strcasecmp("TCCActivationTemp", key) == 0) {
1477                 tmp_val = strtoul(value, &end, 0);
1478                 if (*end != '\0' || tmp_val > UINT_MAX)
1479                         return -1;
1480                 tcc_activation_temp = (unsigned int) tmp_val;
1481         } else {
1482                 return -1;
1483         }
1484         return 0;
1485 }
1486
1487 void module_register(void);
1488 void module_register(void)
1489 {
1490         plugin_register_init(PLUGIN_NAME, turbostat_init);
1491         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1492 }