src/turbostat.c

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