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 #define _GNU_SOURCE
  28 #include <asm/msr-index.h>
  29 #include <stdarg.h>
  30 #include <stdio.h>
  31 #include <err.h>
  32 #include <unistd.h>
  33 #include <sys/types.h>
  34 #include <sys/wait.h>
  35 #include <sys/stat.h>
  36 #include <sys/resource.h>
  37 #include <fcntl.h>
  38 #include <signal.h>
  39 #include <sys/time.h>
  40 #include <stdlib.h>
  41 #include <dirent.h>
  42 #include <string.h>
  43 #include <ctype.h>
  44 #include <sched.h>
  45 #include <cpuid.h>
  46
  47 #include "collectd.h"
  48 #include "common.h"
  49 #include "plugin.h"
  50
  51 #define PLUGIN_NAME "turbostat"
  52
  53 static const char *proc_stat = "/proc/stat";
  54 static unsigned int interval_sec = 5;   /* set with -i interval_sec */
  55 static unsigned int skip_c0;
  56 static unsigned int skip_c1;
  57 static unsigned int do_nhm_cstates;
  58 static unsigned int do_snb_cstates;
  59 static unsigned int do_c8_c9_c10;
  60 static unsigned int do_slm_cstates;
  61 static unsigned int has_aperf;
  62 static unsigned int has_epb;
  63 static unsigned int units = 1000000000; /* Ghz etc */
  64 static unsigned int genuine_intel;
  65 static unsigned int has_invariant_tsc;
  66 static unsigned int do_nehalem_platform_info;
  67 static int do_smi;
  68 static unsigned int show_pkg;
  69 static unsigned int show_core;
  70 static unsigned int show_cpu;
  71 static unsigned int do_rapl;
  72 static unsigned int do_dts;
  73 static unsigned int do_ptm;
  74 static unsigned int tcc_activation_temp;
  75 static unsigned int tcc_activation_temp_override;
  76 static double rapl_power_units, rapl_energy_units, rapl_time_units;
  77 static double rapl_joule_counter_range;
  78
  79 #define RAPL_PKG                (1 << 0)
  80                                         /* 0x610 MSR_PKG_POWER_LIMIT */
  81                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
  82 #define RAPL_PKG_PERF_STATUS    (1 << 1)
  83                                         /* 0x613 MSR_PKG_PERF_STATUS */
  84 #define RAPL_PKG_POWER_INFO     (1 << 2)
  85                                         /* 0x614 MSR_PKG_POWER_INFO */
  86
  87 #define RAPL_DRAM               (1 << 3)
  88                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
  89                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
  90                                         /* 0x61c MSR_DRAM_POWER_INFO */
  91 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
  92                                         /* 0x61b MSR_DRAM_PERF_STATUS */
  93
  94 #define RAPL_CORES              (1 << 5)
  95                                         /* 0x638 MSR_PP0_POWER_LIMIT */
  96                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
  97 #define RAPL_CORE_POLICY        (1 << 6)
  98                                         /* 0x63a MSR_PP0_POLICY */
  99
 100
 101 #define RAPL_GFX                (1 << 7)
 102                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 103                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 104                                         /* 0x642 MSR_PP1_POLICY */
 105 #define TJMAX_DEFAULT   100
 106
 107 int aperf_mperf_unstable;
 108 int backwards_count;
 109 char *progname;
 110
 111 cpu_set_t *cpu_present_set, *cpu_affinity_set;
 112 size_t cpu_present_setsize, cpu_affinity_setsize;
 113
 114 struct thread_data {
 115         unsigned long long tsc;
 116         unsigned long long aperf;
 117         unsigned long long mperf;
 118         unsigned long long c1;
 119         unsigned int smi_count;
 120         unsigned int cpu_id;
 121         unsigned int flags;
 122 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 123 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 124 } *thread_even, *thread_odd;
 125
 126 struct core_data {
 127         unsigned long long c3;
 128         unsigned long long c6;
 129         unsigned long long c7;
 130         unsigned int core_temp_c;
 131         unsigned int core_id;
 132 } *core_even, *core_odd;
 133
 134 struct pkg_data {
 135         unsigned long long pc2;
 136         unsigned long long pc3;
 137         unsigned long long pc6;
 138         unsigned long long pc7;
 139         unsigned long long pc8;
 140         unsigned long long pc9;
 141         unsigned long long pc10;
 142         unsigned int package_id;
 143         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 144         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 145         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 146         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 147         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 148         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 149         unsigned int pkg_temp_c;
 150
 151 } *package_even, *package_odd;
 152
 153 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 154 #define EVEN_COUNTERS thread_even, core_even, package_even
 155 static _Bool is_even = 1;
 156
 157 static _Bool allocated = 0;
 158 static _Bool initialized = 0;
 159
 160 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
 161         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
 162                 topo.num_threads_per_core + \
 163                 (core_no) * topo.num_threads_per_core + (thread_no))
 164 #define GET_CORE(core_base, core_no, pkg_no) \
 165         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
 166 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 167
 168 struct topo_params {
 169         int num_packages;
 170         int num_cpus;
 171         int num_cores;
 172         int max_cpu_num;
 173         int num_cores_per_pkg;
 174         int num_threads_per_core;
 175 } topo;
 176
 177 struct timeval tv_even, tv_odd, tv_delta;
 178
 179 enum return_values {
 180         OK = 0,
 181         ERR_CPU_MIGRATE,
 182         ERR_MSR_IA32_APERF,
 183         ERR_MSR_IA32_MPERF,
 184         ERR_MSR_SMI_COUNT,
 185         ERR_MSR_CORE_C3_RESIDENCY,
 186         ERR_MSR_CORE_C6_RESIDENCY,
 187         ERR_MSR_CORE_C7_RESIDENCY,
 188         ERR_MSR_IA32_THERM_STATUS,
 189         ERR_MSR_PKG_C3_RESIDENCY,
 190         ERR_MSR_PKG_C6_RESIDENCY,
 191         ERR_MSR_PKG_C2_RESIDENCY,
 192         ERR_MSR_PKG_C7_RESIDENCY,
 193         ERR_MSR_PKG_C8_RESIDENCY,
 194         ERR_MSR_PKG_C9_RESIDENCY,
 195         ERR_MSR_PKG_C10_RESIDENCY,
 196         ERR_MSR_PKG_ENERGY_STATUS,
 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_CPU_NOT_PRESENT,
 204         ERR_NO_MSR,
 205         ERR_CANT_OPEN_FILE,
 206         ERR_CANT_READ_NUMBER,
 207         ERR_CANT_READ_PROC_STAT,
 208         ERR_NO_INVARIANT_TSC,
 209         ERR_NO_APERF,
 210         ERR_CALLOC,
 211         ERR_CPU_ALLOC,
 212         ERR_NOT_ROOT,
 213 };
 214
 215 #define __must_check __attribute__((warn_unused_result))
 216
 217 static int setup_all_buffers(void);
 218
 219 static int
 220 cpu_is_not_present(int cpu)
 221 {
 222         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 223 }
 224 /*
 225  * run func(thread, core, package) in topology order
 226  * skip non-present cpus
 227  */
 228
 229 static int __must_check
 230 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 231         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 232 {
 233         int retval, pkg_no, core_no, thread_no;
 234
 235         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 236                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 237                         for (thread_no = 0; thread_no <
 238                                 topo.num_threads_per_core; ++thread_no) {
 239                                 struct thread_data *t;
 240                                 struct core_data *c;
 241                                 struct pkg_data *p;
 242
 243                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 244
 245                                 if (cpu_is_not_present(t->cpu_id))
 246                                         continue;
 247
 248                                 c = GET_CORE(core_base, core_no, pkg_no);
 249                                 p = GET_PKG(pkg_base, pkg_no);
 250
 251                                 retval = func(t, c, p);
 252                                 if (retval)
 253                                         return retval;
 254                         }
 255                 }
 256         }
 257         return 0;
 258 }
 259
 260 static int __must_check
 261 cpu_migrate(int cpu)
 262 {
 263         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 264         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 265         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 266                 return -ERR_CPU_MIGRATE;
 267         else
 268                 return 0;
 269 }
 270
 271 static int __must_check
 272 get_msr(int cpu, off_t offset, unsigned long long *msr)
 273 {
 274         ssize_t retval;
 275         char pathname[32];
 276         int fd;
 277
 278         sprintf(pathname, "/dev/cpu/%d/msr", cpu);
 279         fd = open(pathname, O_RDONLY);
 280         if (fd < 0)
 281                 return -1;
 282
 283         retval = pread(fd, msr, sizeof *msr, offset);
 284         close(fd);
 285
 286         if (retval != sizeof *msr) {
 287                 ERROR ("%s offset 0x%llx read failed\n", pathname, (unsigned long long)offset);
 288                 return -1;
 289         }
 290
 291         return 0;
 292 }
 293
 294 #define DELTA_WRAP32(new, old)                  \
 295         if (new > old) {                        \
 296                 old = new - old;                \
 297         } else {                                \
 298                 old = 0x100000000 + new - old;  \
 299         }
 300
 301 static void
 302 delta_package(struct pkg_data *new, struct pkg_data *old)
 303 {
 304         old->pc2 = new->pc2 - old->pc2;
 305         old->pc3 = new->pc3 - old->pc3;
 306         old->pc6 = new->pc6 - old->pc6;
 307         old->pc7 = new->pc7 - old->pc7;
 308         old->pc8 = new->pc8 - old->pc8;
 309         old->pc9 = new->pc9 - old->pc9;
 310         old->pc10 = new->pc10 - old->pc10;
 311         old->pkg_temp_c = new->pkg_temp_c;
 312
 313         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
 314         DELTA_WRAP32(new->energy_cores, old->energy_cores);
 315         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
 316         DELTA_WRAP32(new->energy_dram, old->energy_dram);
 317         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
 318         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
 319 }
 320
 321 static void
 322 delta_core(struct core_data *new, struct core_data *old)
 323 {
 324         old->c3 = new->c3 - old->c3;
 325         old->c6 = new->c6 - old->c6;
 326         old->c7 = new->c7 - old->c7;
 327         old->core_temp_c = new->core_temp_c;
 328 }
 329
 330 /*
 331  * old = new - old
 332  */
 333 static int __must_check
 334 delta_thread(struct thread_data *new, struct thread_data *old,
 335         struct core_data *core_delta)
 336 {
 337         old->tsc = new->tsc - old->tsc;
 338
 339         /* check for TSC < 1 Mcycles over interval */
 340         if (old->tsc < (1000 * 1000)) {
 341                 WARNING("Insanely slow TSC rate, TSC stops in idle?\n"
 342                         "You can disable all c-states by booting with \"idle=poll\"\n"
 343                         "or just the deep ones with \"processor.max_cstate=1\"");
 344                 return -1;
 345         }
 346
 347         old->c1 = new->c1 - old->c1;
 348
 349         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
 350                 old->aperf = new->aperf - old->aperf;
 351                 old->mperf = new->mperf - old->mperf;
 352         } else {
 353
 354                 if (!aperf_mperf_unstable) {
 355                         WARNING("%s: APERF or MPERF went backwards *\n", progname);
 356                         WARNING("* Frequency results do not cover entire interval *\n");
 357                         WARNING("* fix this by running Linux-2.6.30 or later *\n");
 358
 359                         aperf_mperf_unstable = 1;
 360                 }
 361                 /*
 362                  * mperf delta is likely a huge "positive" number
 363                  * can not use it for calculating c0 time
 364                  */
 365                 skip_c0 = 1;
 366                 skip_c1 = 1;
 367         }
 368
 369
 370         /*
 371          * As counter collection is not atomic,
 372          * it is possible for mperf's non-halted cycles + idle states
 373          * to exceed TSC's all cycles: show c1 = 0% in that case.
 374          */
 375         if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
 376                 old->c1 = 0;
 377         else {
 378                 /* normal case, derive c1 */
 379                 old->c1 = old->tsc - old->mperf - core_delta->c3
 380                         - core_delta->c6 - core_delta->c7;
 381         }
 382
 383         if (old->mperf == 0) {
 384                 WARNING("cpu%d MPERF 0!\n", old->cpu_id);
 385                 old->mperf = 1; /* divide by 0 protection */
 386         }
 387
 388         if (do_smi)
 389                 old->smi_count = new->smi_count - old->smi_count;
 390
 391         return 0;
 392 }
 393
 394 static int __must_check
 395 delta_cpu(struct thread_data *t, struct core_data *c,
 396         struct pkg_data *p, struct thread_data *t2,
 397         struct core_data *c2, struct pkg_data *p2)
 398 {
 399         int ret;
 400
 401         /* calculate core delta only for 1st thread in core */
 402         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
 403                 delta_core(c, c2);
 404
 405         /* always calculate thread delta */
 406         ret = delta_thread(t, t2, c2);  /* c2 is core delta */
 407         if (ret != 0)
 408                 return ret;
 409
 410         /* calculate package delta only for 1st core in package */
 411         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
 412                 delta_package(p, p2);
 413
 414         return 0;
 415 }
 416
 417 static unsigned long long
 418 rdtsc(void)
 419 {
 420         unsigned int low, high;
 421
 422         asm volatile("rdtsc" : "=a" (low), "=d" (high));
 423
 424         return low | ((unsigned long long)high) << 32;
 425 }
 426
 427
 428 /*
 429  * get_counters(...)
 430  * migrate to cpu
 431  * acquire and record local counters for that cpu
 432  */
 433 static int __must_check
 434 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
 435 {
 436         int cpu = t->cpu_id;
 437         unsigned long long msr;
 438
 439         if (cpu_migrate(cpu)) {
 440                 WARNING("Could not migrate to CPU %d\n", cpu);
 441                 return -ERR_CPU_MIGRATE;
 442         }
 443
 444         t->tsc = rdtsc();       /* we are running on local CPU of interest */
 445
 446         if (has_aperf) {
 447                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
 448                         return -ERR_MSR_IA32_APERF;
 449                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
 450                         return -ERR_MSR_IA32_MPERF;
 451         }
 452
 453         if (do_smi) {
 454                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
 455                         return -ERR_MSR_SMI_COUNT;
 456                 t->smi_count = msr & 0xFFFFFFFF;
 457         }
 458
 459         /* collect core counters only for 1st thread in core */
 460         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 461                 return 0;
 462
 463         if (do_nhm_cstates && !do_slm_cstates) {
 464                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
 465                         return -ERR_MSR_CORE_C3_RESIDENCY;
 466         }
 467
 468         if (do_nhm_cstates) {
 469                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
 470                         return -ERR_MSR_CORE_C6_RESIDENCY;
 471         }
 472
 473         if (do_snb_cstates)
 474                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
 475                         return -ERR_MSR_CORE_C7_RESIDENCY;
 476
 477         if (do_dts) {
 478                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
 479                         return -ERR_MSR_IA32_THERM_STATUS;
 480                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 481         }
 482
 483
 484         /* collect package counters only for 1st core in package */
 485         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 486                 return 0;
 487
 488         if (do_nhm_cstates && !do_slm_cstates) {
 489                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
 490                         return -ERR_MSR_PKG_C3_RESIDENCY;
 491                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
 492                         return -ERR_MSR_PKG_C6_RESIDENCY;
 493         }
 494         if (do_snb_cstates) {
 495                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
 496                         return -ERR_MSR_PKG_C2_RESIDENCY;
 497                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
 498                         return -ERR_MSR_PKG_C7_RESIDENCY;
 499         }
 500         if (do_c8_c9_c10) {
 501                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
 502                         return -ERR_MSR_PKG_C8_RESIDENCY;
 503                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
 504                         return -ERR_MSR_PKG_C9_RESIDENCY;
 505                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
 506                         return -ERR_MSR_PKG_C10_RESIDENCY;
 507         }
 508         if (do_rapl & RAPL_PKG) {
 509                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
 510                         return -ERR_MSR_PKG_ENERGY_STATUS;
 511                 p->energy_pkg = msr & 0xFFFFFFFF;
 512         }
 513         if (do_rapl & RAPL_CORES) {
 514                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
 515                         return MSR_PP0_ENERGY_STATUS;
 516                 p->energy_cores = msr & 0xFFFFFFFF;
 517         }
 518         if (do_rapl & RAPL_DRAM) {
 519                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
 520                         return -ERR_MSR_DRAM_ENERGY_STATUS;
 521                 p->energy_dram = msr & 0xFFFFFFFF;
 522         }
 523         if (do_rapl & RAPL_GFX) {
 524                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
 525                         return -ERR_MSR_PP1_ENERGY_STATUS;
 526                 p->energy_gfx = msr & 0xFFFFFFFF;
 527         }
 528         if (do_rapl & RAPL_PKG_PERF_STATUS) {
 529                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
 530                         return -ERR_MSR_PKG_PERF_STATUS;
 531                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
 532         }
 533         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
 534                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
 535                         return -ERR_MSR_DRAM_PERF_STATUS;
 536                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
 537         }
 538         if (do_ptm) {
 539                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
 540                         return -ERR_MSR_IA32_PACKAGE_THERM_STATUS;
 541                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
 542         }
 543         return 0;
 544 }
 545
 546 static void
 547 free_all_buffers(void)
 548 {
 549         allocated = 0;
 550         initialized = 0;
 551
 552         CPU_FREE(cpu_present_set);
 553         cpu_present_set = NULL;
 554         cpu_present_set = 0;
 555
 556         CPU_FREE(cpu_affinity_set);
 557         cpu_affinity_set = NULL;
 558         cpu_affinity_setsize = 0;
 559
 560         free(thread_even);
 561         free(core_even);
 562         free(package_even);
 563
 564         thread_even = NULL;
 565         core_even = NULL;
 566         package_even = NULL;
 567
 568         free(thread_odd);
 569         free(core_odd);
 570         free(package_odd);
 571
 572         thread_odd = NULL;
 573         core_odd = NULL;
 574         package_odd = NULL;
 575 }
 576
 577 /*
 578  * Parse a file containing a single int.
 579  */
 580 static int
 581 parse_int_file(const char *fmt, ...)
 582 {
 583         va_list args;
 584         char path[PATH_MAX];
 585         FILE *filep;
 586         int value;
 587
 588         va_start(args, fmt);
 589         vsnprintf(path, sizeof(path), fmt, args);
 590         va_end(args);
 591         filep = fopen(path, "r");
 592         if (!filep) {
 593                 ERROR("%s: open failed", path);
 594                 return -ERR_CANT_OPEN_FILE;
 595         }
 596         if (fscanf(filep, "%d", &value) != 1) {
 597                 ERROR("%s: failed to parse number from file", path);
 598                 return -ERR_CANT_READ_NUMBER;
 599         }
 600         fclose(filep);
 601         return value;
 602 }
 603
 604 /*
 605  * cpu_is_first_sibling_in_core(cpu)
 606  * return 1 if given CPU is 1st HT sibling in the core
 607  */
 608 static int
 609 cpu_is_first_sibling_in_core(int cpu)
 610 {
 611         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 612 }
 613
 614 /*
 615  * cpu_is_first_core_in_package(cpu)
 616  * return 1 if given CPU is 1st core in package
 617  */
 618 static int
 619 cpu_is_first_core_in_package(int cpu)
 620 {
 621         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
 622 }
 623
 624 static int
 625 get_physical_package_id(int cpu)
 626 {
 627         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
 628 }
 629
 630 static int
 631 get_core_id(int cpu)
 632 {
 633         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
 634 }
 635
 636 static int
 637 get_num_ht_siblings(int cpu)
 638 {
 639         char path[80];
 640         FILE *filep;
 641         int sib1, sib2;
 642         int matches;
 643         char character;
 644
 645         sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
 646         filep = fopen(path, "r");
 647         if (!filep) {
 648                 ERROR("%s: open failed", path);
 649                 return -ERR_CANT_OPEN_FILE;
 650         }
 651         /*
 652          * file format:
 653          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
 654          * otherwinse 1 sibling (self).
 655          */
 656         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
 657
 658         fclose(filep);
 659
 660         if (matches == 3)
 661                 return 2;
 662         else
 663                 return 1;
 664 }
 665
 666 /*
 667  * run func(thread, core, package) in topology order
 668  * skip non-present cpus
 669  */
 670
 671
 672 static int __must_check
 673 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
 674         struct pkg_data *, struct thread_data *, struct core_data *,
 675         struct pkg_data *), struct thread_data *thread_base,
 676         struct core_data *core_base, struct pkg_data *pkg_base,
 677         struct thread_data *thread_base2, struct core_data *core_base2,
 678         struct pkg_data *pkg_base2)
 679 {
 680         int retval, pkg_no, core_no, thread_no;
 681
 682         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 683                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
 684                         for (thread_no = 0; thread_no <
 685                                 topo.num_threads_per_core; ++thread_no) {
 686                                 struct thread_data *t, *t2;
 687                                 struct core_data *c, *c2;
 688                                 struct pkg_data *p, *p2;
 689
 690                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
 691
 692                                 if (cpu_is_not_present(t->cpu_id))
 693                                         continue;
 694
 695                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
 696
 697                                 c = GET_CORE(core_base, core_no, pkg_no);
 698                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
 699
 700                                 p = GET_PKG(pkg_base, pkg_no);
 701                                 p2 = GET_PKG(pkg_base2, pkg_no);
 702
 703                                 retval = func(t, c, p, t2, c2, p2);
 704                                 if (retval)
 705                                         return retval;
 706                         }
 707                 }
 708         }
 709         return 0;
 710 }
 711
 712 /*
 713  * run func(cpu) on every cpu in /proc/stat
 714  * return max_cpu number
 715  */
 716 static int __must_check
 717 for_all_proc_cpus(int (func)(int))
 718 {
 719         FILE *fp;
 720         int cpu_num;
 721         int retval;
 722
 723         fp = fopen(proc_stat, "r");
 724         if (!fp) {
 725                 ERROR("%s: open failed", proc_stat);
 726                 return -ERR_CANT_OPEN_FILE;
 727         }
 728
 729         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
 730         if (retval != 0) {
 731                 ERROR("%s: failed to parse format", proc_stat);
 732                 return -ERR_CANT_READ_PROC_STAT;
 733         }
 734
 735         while (1) {
 736                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
 737                 if (retval != 1)
 738                         break;
 739
 740                 retval = func(cpu_num);
 741                 if (retval) {
 742                         fclose(fp);
 743                         return(retval);
 744                 }
 745         }
 746         fclose(fp);
 747         return 0;
 748 }
 749
 750 /*
 751  * count_cpus()
 752  * remember the last one seen, it will be the max
 753  */
 754 static int
 755 count_cpus(int cpu)
 756 {
 757         if (topo.max_cpu_num < cpu)
 758                 topo.max_cpu_num = cpu;
 759
 760         topo.num_cpus += 1;
 761         return 0;
 762 }
 763 static int
 764 mark_cpu_present(int cpu)
 765 {
 766         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
 767         return 0;
 768 }
 769
 770
 771 static void
 772 turbostat_submit (const char *plugin_instance,
 773         const char *type, const char *type_instance,
 774         gauge_t value)
 775 {
 776         value_list_t vl = VALUE_LIST_INIT;
 777         value_t v;
 778
 779         v.gauge = value;
 780         vl.values = &v;
 781         vl.values_len = 1;
 782         sstrncpy (vl.host, hostname_g, sizeof (vl.host));
 783         sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
 784         if (plugin_instance != NULL)
 785                 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
 786         sstrncpy (vl.type, type, sizeof (vl.type));
 787         if (type_instance != NULL)
 788                 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
 789
 790         plugin_dispatch_values (&vl);
 791 }
 792
 793 /*
 794  * column formatting convention & formats
 795  * package: "pk" 2 columns %2d
 796  * core: "cor" 3 columns %3d
 797  * CPU: "CPU" 3 columns %3d
 798  * Pkg_W: %6.2
 799  * Cor_W: %6.2
 800  * GFX_W: %5.2
 801  * RAM_W: %5.2
 802  * GHz: "GHz" 3 columns %3.2
 803  * TSC: "TSC" 3 columns %3.2
 804  * SMI: "SMI" 4 columns %4d
 805  * percentage " %pc3" %6.2
 806  * Perf Status percentage: %5.2
 807  * "CTMP" 4 columns %4d
 808  */
 809 #define NAME_LEN 12
 810 static int
 811 submit_counters(struct thread_data *t, struct core_data *c,
 812         struct pkg_data *p)
 813 {
 814         char name[NAME_LEN];
 815         double interval_float;
 816
 817         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 818
 819         snprintf(name, NAME_LEN, "cpu%02d", t->cpu_id);
 820
 821         if (do_nhm_cstates) {
 822                 if (!skip_c0)
 823                         turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
 824                 if (!skip_c1)
 825                         turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
 826         }
 827
 828         /* GHz */
 829         if (has_aperf && ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc))))
 830                 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
 831
 832         /* SMI */
 833         if (do_smi)
 834                 turbostat_submit(NULL, "current", name, t->smi_count);
 835
 836         /* print per-core data only for 1st thread in core */
 837         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 838                 goto done;
 839
 840         snprintf(name, NAME_LEN, "core%02d", c->core_id);
 841
 842         if (do_nhm_cstates && !do_slm_cstates)
 843                 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
 844         if (do_nhm_cstates)
 845                 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
 846         if (do_snb_cstates)
 847                 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
 848
 849         if (do_dts)
 850                 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
 851
 852         /* print per-package data only for 1st core in package */
 853         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 854                 goto done;
 855
 856         snprintf(name, NAME_LEN, "pkg%02d", p->package_id);
 857
 858         if (do_ptm)
 859                 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
 860
 861         if (do_snb_cstates)
 862                 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
 863         if (do_nhm_cstates && !do_slm_cstates)
 864                 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
 865         if (do_nhm_cstates && !do_slm_cstates)
 866                 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
 867         if (do_snb_cstates)
 868                 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
 869         if (do_c8_c9_c10) {
 870                 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
 871                 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
 872                 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
 873         }
 874
 875         if (do_rapl) {
 876                 if (do_rapl & RAPL_PKG)
 877                         turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
 878                 if (do_rapl & RAPL_CORES)
 879                         turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
 880                 if (do_rapl & RAPL_GFX)
 881                         turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
 882                 if (do_rapl & RAPL_DRAM)
 883                         turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
 884         }
 885 done:
 886         return 0;
 887 }
 888
 889 static int
 890 turbostat_read(user_data_t * not_used)
 891 {
 892         int ret;
 893
 894         if (!allocated) {
 895                 if ((ret = setup_all_buffers()) < 0)
 896                         return ret;
 897         }
 898
 899         if (for_all_proc_cpus(cpu_is_not_present)) {
 900                 free_all_buffers();
 901                 if ((ret = setup_all_buffers()) < 0)
 902                         return ret;
 903                 if (for_all_proc_cpus(cpu_is_not_present))
 904                         return -ERR_CPU_NOT_PRESENT;
 905         }
 906
 907         if (!initialized) {
 908                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 909                         return ret;
 910                 gettimeofday(&tv_even, (struct timezone *)NULL);
 911                 is_even = 1;
 912                 initialized = 1;
 913                 return 0;
 914         }
 915
 916         if (is_even) {
 917                 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
 918                         return ret;
 919                 gettimeofday(&tv_odd, (struct timezone *)NULL);
 920                 is_even = 0;
 921                 timersub(&tv_odd, &tv_even, &tv_delta);
 922                 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
 923                         return ret;
 924                 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
 925                         return ret;
 926         } else {
 927                 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
 928                         return ret;
 929                 gettimeofday(&tv_even, (struct timezone *)NULL);
 930                 is_even = 1;
 931                 timersub(&tv_even, &tv_odd, &tv_delta);
 932                 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
 933                         return ret;
 934                 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
 935                         return ret;
 936         }
 937         return 0;
 938 }
 939
 940 static int __must_check
 941 check_dev_msr()
 942 {
 943         struct stat sb;
 944
 945         if (stat("/dev/cpu/0/msr", &sb)) {
 946                 ERROR("no /dev/cpu/0/msr\n"
 947                         "Try \"# modprobe msr\"");
 948                 return -ERR_NO_MSR;
 949         }
 950         return 0;
 951 }
 952
 953 static int __must_check
 954 check_super_user()
 955 {
 956         if (getuid() != 0) {
 957                 ERROR("must be root");
 958                 return -ERR_NOT_ROOT;
 959         }
 960         return 0;
 961 }
 962
 963
 964 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
 965 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
 966
 967 static double
 968 get_tdp(unsigned int model)
 969 {
 970         unsigned long long msr;
 971
 972         if (do_rapl & RAPL_PKG_POWER_INFO)
 973                 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
 974                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
 975
 976         switch (model) {
 977         case 0x37:
 978         case 0x4D:
 979                 return 30.0;
 980         default:
 981                 return 135.0;
 982         }
 983 }
 984
 985
 986 /*
 987  * rapl_probe()
 988  *
 989  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
 990  */
 991 static void
 992 rapl_probe(unsigned int family, unsigned int model)
 993 {
 994         unsigned long long msr;
 995         unsigned int time_unit;
 996         double tdp;
 997
 998         if (!genuine_intel)
 999                 return;
1000
1001         if (family != 6)
1002                 return;
1003
1004         switch (model) {
1005         case 0x2A:
1006         case 0x3A:
1007         case 0x3C:      /* HSW */
1008         case 0x45:      /* HSW */
1009         case 0x46:      /* HSW */
1010                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1011                 break;
1012         case 0x3F:      /* HSX */
1013                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1014                 break;
1015         case 0x2D:
1016         case 0x3E:
1017                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1018                 break;
1019         case 0x37:      /* BYT */
1020         case 0x4D:      /* AVN */
1021                 do_rapl = RAPL_PKG | RAPL_CORES ;
1022                 break;
1023         default:
1024                 return;
1025         }
1026
1027         /* units on package 0, verify later other packages match */
1028         if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1029                 return;
1030
1031         rapl_power_units = 1.0 / (1 << (msr & 0xF));
1032         if (model == 0x37)
1033                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1034         else
1035                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1036
1037         time_unit = msr >> 16 & 0xF;
1038         if (time_unit == 0)
1039                 time_unit = 0xA;
1040
1041         rapl_time_units = 1.0 / (1 << (time_unit));
1042
1043         tdp = get_tdp(model);
1044
1045         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1046 //      if (verbose)
1047 //              fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1048
1049         return;
1050 }
1051
1052 static int
1053 is_snb(unsigned int family, unsigned int model)
1054 {
1055         if (!genuine_intel)
1056                 return 0;
1057
1058         switch (model) {
1059         case 0x2A:
1060         case 0x2D:
1061         case 0x3A:      /* IVB */
1062         case 0x3E:      /* IVB Xeon */
1063         case 0x3C:      /* HSW */
1064         case 0x3F:      /* HSW */
1065         case 0x45:      /* HSW */
1066         case 0x46:      /* HSW */
1067                 return 1;
1068         }
1069         return 0;
1070 }
1071
1072 static int
1073 has_c8_c9_c10(unsigned int family, unsigned int model)
1074 {
1075         if (!genuine_intel)
1076                 return 0;
1077
1078         switch (model) {
1079         case 0x45:
1080                 return 1;
1081         }
1082         return 0;
1083 }
1084
1085
1086 static int
1087 is_slm(unsigned int family, unsigned int model)
1088 {
1089         if (!genuine_intel)
1090                 return 0;
1091         switch (model) {
1092         case 0x37:      /* BYT */
1093         case 0x4D:      /* AVN */
1094                 return 1;
1095         }
1096         return 0;
1097 }
1098
1099 /*
1100  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1101  * the Thermal Control Circuit (TCC) activates.
1102  * This is usually equal to tjMax.
1103  *
1104  * Older processors do not have this MSR, so there we guess,
1105  * but also allow cmdline over-ride with -T.
1106  *
1107  * Several MSR temperature values are in units of degrees-C
1108  * below this value, including the Digital Thermal Sensor (DTS),
1109  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1110  */
1111 static int __must_check
1112 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1113 {
1114         unsigned long long msr;
1115         unsigned int target_c_local;
1116         int cpu;
1117
1118         /* tcc_activation_temp is used only for dts or ptm */
1119         if (!(do_dts || do_ptm))
1120                 return 0;
1121
1122         /* this is a per-package concept */
1123         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1124                 return 0;
1125
1126         cpu = t->cpu_id;
1127         if (cpu_migrate(cpu)) {
1128                 ERROR("Could not migrate to CPU %d\n", cpu);
1129                 return -ERR_CPU_MIGRATE;
1130         }
1131
1132         if (tcc_activation_temp_override != 0) {
1133                 tcc_activation_temp = tcc_activation_temp_override;
1134                 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
1135                         cpu, tcc_activation_temp);
1136                 return 0;
1137         }
1138
1139         /* Temperature Target MSR is Nehalem and newer only */
1140         if (!do_nehalem_platform_info)
1141                 goto guess;
1142
1143         if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
1144                 goto guess;
1145
1146         target_c_local = (msr >> 16) & 0x7F;
1147
1148         if (target_c_local < 85 || target_c_local > 127)
1149                 goto guess;
1150
1151         tcc_activation_temp = target_c_local;
1152
1153         return 0;
1154
1155 guess:
1156         tcc_activation_temp = TJMAX_DEFAULT;
1157         WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1158                 cpu, tcc_activation_temp);
1159
1160         return 0;
1161 }
1162
1163 static int __must_check
1164 check_cpuid()
1165 {
1166         unsigned int eax, ebx, ecx, edx, max_level;
1167         unsigned int fms, family, model;
1168
1169         eax = ebx = ecx = edx = 0;
1170
1171         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1172
1173         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1174                 genuine_intel = 1;
1175
1176         fms = 0;
1177         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1178         family = (fms >> 8) & 0xf;
1179         model = (fms >> 4) & 0xf;
1180         if (family == 6 || family == 0xf)
1181                 model += ((fms >> 16) & 0xf) << 4;
1182
1183         if (!(edx & (1 << 5))) {
1184                 ERROR("CPUID: no MSR");
1185                 return -ERR_NO_MSR;
1186         }
1187
1188         /*
1189          * check max extended function levels of CPUID.
1190          * This is needed to check for invariant TSC.
1191          * This check is valid for both Intel and AMD.
1192          */
1193         ebx = ecx = edx = 0;
1194         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1195
1196         if (max_level < 0x80000007) {
1197                 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1198                 return -ERR_NO_INVARIANT_TSC;
1199         }
1200
1201         /*
1202          * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1203          * this check is valid for both Intel and AMD
1204          */
1205         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1206         has_invariant_tsc = edx & (1 << 8);
1207
1208         if (!has_invariant_tsc) {
1209                 ERROR("No invariant TSC");
1210                 return -ERR_NO_INVARIANT_TSC;
1211         }
1212
1213         /*
1214          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1215          * this check is valid for both Intel and AMD
1216          */
1217
1218         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1219         has_aperf = ecx & (1 << 0);
1220         do_dts = eax & (1 << 0);
1221         do_ptm = eax & (1 << 6);
1222         has_epb = ecx & (1 << 3);
1223
1224         if (!has_aperf) {
1225                 ERROR("No APERF");
1226                 return -ERR_NO_APERF;
1227         }
1228
1229         do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1230         do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */
1231         do_smi = do_nhm_cstates;
1232         do_snb_cstates = is_snb(family, model);
1233         do_c8_c9_c10 = has_c8_c9_c10(family, model);
1234         do_slm_cstates = is_slm(family, model);
1235
1236         rapl_probe(family, model);
1237
1238         return 0;
1239 }
1240
1241
1242
1243 static int __must_check
1244 topology_probe()
1245 {
1246         int i;
1247         int ret;
1248         int max_core_id = 0;
1249         int max_package_id = 0;
1250         int max_siblings = 0;
1251         struct cpu_topology {
1252                 int core_id;
1253                 int physical_package_id;
1254         } *cpus;
1255
1256         /* Initialize num_cpus, max_cpu_num */
1257         topo.num_cpus = 0;
1258         topo.max_cpu_num = 0;
1259         ret = for_all_proc_cpus(count_cpus);
1260         if (ret < 0)
1261                 return ret;
1262         if (topo.num_cpus > 1)
1263                 show_cpu = 1;
1264
1265         DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1266
1267         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1268         if (cpus == NULL) {
1269                 ERROR("calloc cpus");
1270                 return -ERR_CALLOC;
1271         }
1272
1273         /*
1274          * Allocate and initialize cpu_present_set
1275          */
1276         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1277         if (cpu_present_set == NULL) {
1278                 free(cpus);
1279                 ERROR("CPU_ALLOC");
1280                 return -ERR_CPU_ALLOC;
1281         }
1282         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1283         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1284         ret = for_all_proc_cpus(mark_cpu_present);
1285         if (ret < 0) {
1286                 free(cpus);
1287                 return ret;
1288         }
1289
1290         /*
1291          * Allocate and initialize cpu_affinity_set
1292          */
1293         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1294         if (cpu_affinity_set == NULL) {
1295                 free(cpus);
1296                 ERROR("CPU_ALLOC");
1297                 return -ERR_CPU_ALLOC;
1298         }
1299         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1300         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1301
1302
1303         /*
1304          * For online cpus
1305          * find max_core_id, max_package_id
1306          */
1307         for (i = 0; i <= topo.max_cpu_num; ++i) {
1308                 int siblings;
1309
1310                 if (cpu_is_not_present(i)) {
1311                         //if (verbose > 1)
1312                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1313                         continue;
1314                 }
1315                 cpus[i].core_id = get_core_id(i);
1316                 if (cpus[i].core_id < 0)
1317                         return cpus[i].core_id;
1318                 if (cpus[i].core_id > max_core_id)
1319                         max_core_id = cpus[i].core_id;
1320
1321                 cpus[i].physical_package_id = get_physical_package_id(i);
1322                 if (cpus[i].physical_package_id < 0)
1323                         return cpus[i].physical_package_id;
1324                 if (cpus[i].physical_package_id > max_package_id)
1325                         max_package_id = cpus[i].physical_package_id;
1326
1327                 siblings = get_num_ht_siblings(i);
1328                 if (siblings < 0)
1329                         return siblings;
1330                 if (siblings > max_siblings)
1331                         max_siblings = siblings;
1332                 DEBUG("cpu %d pkg %d core %d\n",
1333                         i, cpus[i].physical_package_id, cpus[i].core_id);
1334         }
1335         topo.num_cores_per_pkg = max_core_id + 1;
1336         DEBUG("max_core_id %d, sizing for %d cores per package\n",
1337                 max_core_id, topo.num_cores_per_pkg);
1338         if (topo.num_cores_per_pkg > 1)
1339                 show_core = 1;
1340
1341         topo.num_packages = max_package_id + 1;
1342         DEBUG("max_package_id %d, sizing for %d packages\n",
1343                 max_package_id, topo.num_packages);
1344         if (topo.num_packages > 1)
1345                 show_pkg = 1;
1346
1347         topo.num_threads_per_core = max_siblings;
1348         DEBUG("max_siblings %d\n", max_siblings);
1349
1350         free(cpus);
1351         return 0;
1352 }
1353
1354 static int
1355 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1356 {
1357         int i;
1358
1359         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1360                 topo.num_packages, sizeof(struct thread_data));
1361         if (*t == NULL)
1362                 goto error;
1363
1364         for (i = 0; i < topo.num_threads_per_core *
1365                 topo.num_cores_per_pkg * topo.num_packages; i++)
1366                 (*t)[i].cpu_id = -1;
1367
1368         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1369                 sizeof(struct core_data));
1370         if (*c == NULL)
1371                 goto error;
1372
1373         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1374                 (*c)[i].core_id = -1;
1375
1376         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1377         if (*p == NULL)
1378                 goto error;
1379
1380         for (i = 0; i < topo.num_packages; i++)
1381                 (*p)[i].package_id = i;
1382
1383         return 0;
1384 error:
1385         ERROR("calloc counters");
1386         return -ERR_CALLOC;
1387 }
1388 /*
1389  * init_counter()
1390  *
1391  * set cpu_id, core_num, pkg_num
1392  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1393  *
1394  * increment topo.num_cores when 1st core in pkg seen
1395  */
1396 static int
1397 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1398         struct pkg_data *pkg_base, int thread_num, int core_num,
1399         int pkg_num, int cpu_id)
1400 {
1401         int ret;
1402         struct thread_data *t;
1403         struct core_data *c;
1404         struct pkg_data *p;
1405
1406         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1407         c = GET_CORE(core_base, core_num, pkg_num);
1408         p = GET_PKG(pkg_base, pkg_num);
1409
1410         t->cpu_id = cpu_id;
1411         if (thread_num == 0) {
1412                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1413                 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1414                         return ret;
1415                 } else if (ret != 0) {
1416                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1417                 }
1418         }
1419
1420         c->core_id = core_num;
1421         p->package_id = pkg_num;
1422
1423         return 0;
1424 }
1425
1426
1427 static int
1428 initialize_counters(int cpu_id)
1429 {
1430         int my_thread_id, my_core_id, my_package_id;
1431         int ret;
1432
1433         my_package_id = get_physical_package_id(cpu_id);
1434         if (my_package_id < 0)
1435                 return my_package_id;
1436         my_core_id = get_core_id(cpu_id);
1437         if (my_core_id < 0)
1438                 return my_core_id;
1439
1440         if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1441                 return ret;
1442         } else if (ret != 0) {
1443                 my_thread_id = 0;
1444                 topo.num_cores++;
1445         } else {
1446                 my_thread_id = 1;
1447         }
1448
1449         ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1450         if (ret < 0)
1451                 return ret;
1452         ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1453         if (ret < 0)
1454                 return ret;
1455         return 0;
1456 }
1457
1458 #define DO_OR_GOTO_ERR(something) \
1459 do {                         \
1460         ret = (something);     \
1461         if (ret < 0)         \
1462                 goto err;    \
1463 } while (0)
1464
1465 static int setup_all_buffers(void)
1466 {
1467         int ret;
1468
1469         DO_OR_GOTO_ERR(topology_probe());
1470         DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1471         DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1472         DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1473
1474         allocated = 1;
1475         return 0;
1476 err:
1477         free_all_buffers();
1478         return ret;
1479 }
1480
1481 static int
1482 turbostat_init(void)
1483 {
1484         int ret;
1485         struct timespec ts;
1486
1487         DO_OR_GOTO_ERR(check_cpuid());
1488         DO_OR_GOTO_ERR(check_dev_msr());
1489         DO_OR_GOTO_ERR(check_super_user());
1490         DO_OR_GOTO_ERR(setup_all_buffers());
1491         DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1492
1493         ts.tv_sec = interval_sec;
1494         ts.tv_nsec = 0;
1495
1496         plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, &ts, NULL);
1497
1498         return 0;
1499 err:
1500         free_all_buffers();
1501         return ret;
1502 }
1503
1504 static const char *config_keys[] =
1505 {
1506         "Interval",
1507 };
1508 static int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1509
1510 static int
1511 turbostat_config(const char *key, const char *value)
1512 {
1513         if (strcasecmp("Interval", key) == 0)
1514                 interval_sec = atoi(value);
1515         else
1516                 return -1;
1517         return 0;
1518 }
1519
1520 void module_register(void);
1521 void module_register(void)
1522 {
1523         plugin_register_init(PLUGIN_NAME, turbostat_init);
1524         plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1525 }