tools/power/x86/turbostat/turbostat.c

   1 /*
   2  * turbostat -- show CPU frequency and C-state residency
   3  * on modern Intel turbo-capable processors.
   4  *
   5  * Copyright (c) 2013 Intel Corporation.
   6  * Len Brown <len.brown@intel.com>
   7  *
   8  * This program is free software; you can redistribute it and/or modify it
   9  * under the terms and conditions of the GNU General Public License,
  10  * version 2, as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope it will be useful, but WITHOUT
  13  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  14  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  15  * more details.
  16  *
  17  * You should have received a copy of the GNU General Public License along with
  18  * this program; if not, write to the Free Software Foundation, Inc.,
  19  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  20  */
  21
  22 #define _GNU_SOURCE
  23 #include MSRHEADER
  24 #include INTEL_FAMILY_HEADER
  25 #include <stdarg.h>
  26 #include <stdio.h>
  27 #include <err.h>
  28 #include <unistd.h>
  29 #include <sys/types.h>
  30 #include <sys/wait.h>
  31 #include <sys/stat.h>
  32 #include <sys/select.h>
  33 #include <sys/resource.h>
  34 #include <fcntl.h>
  35 #include <signal.h>
  36 #include <sys/time.h>
  37 #include <stdlib.h>
  38 #include <getopt.h>
  39 #include <dirent.h>
  40 #include <string.h>
  41 #include <ctype.h>
  42 #include <sched.h>
  43 #include <time.h>
  44 #include <cpuid.h>
  45 #include <linux/capability.h>
  46 #include <errno.h>
  47
  48 char *proc_stat = "/proc/stat";
  49 FILE *outf;
  50 int *fd_percpu;
  51 struct timeval interval_tv = {5, 0};
  52 struct timespec interval_ts = {5, 0};
  53 struct timespec one_msec = {0, 1000000};
  54 unsigned int num_iterations;
  55 unsigned int debug;
  56 unsigned int quiet;
  57 unsigned int shown;
  58 unsigned int sums_need_wide_columns;
  59 unsigned int rapl_joules;
  60 unsigned int summary_only;
  61 unsigned int list_header_only;
  62 unsigned int dump_only;
  63 unsigned int do_snb_cstates;
  64 unsigned int do_knl_cstates;
  65 unsigned int do_slm_cstates;
  66 unsigned int do_cnl_cstates;
  67 unsigned int use_c1_residency_msr;
  68 unsigned int has_aperf;
  69 unsigned int has_epb;
  70 unsigned int do_irtl_snb;
  71 unsigned int do_irtl_hsw;
  72 unsigned int units = 1000000;   /* MHz etc */
  73 unsigned int genuine_intel;
  74 unsigned int has_invariant_tsc;
  75 unsigned int do_nhm_platform_info;
  76 unsigned int no_MSR_MISC_PWR_MGMT;
  77 unsigned int aperf_mperf_multiplier = 1;
  78 double bclk;
  79 double base_hz;
  80 unsigned int has_base_hz;
  81 double tsc_tweak = 1.0;
  82 unsigned int show_pkg_only;
  83 unsigned int show_core_only;
  84 char *output_buffer, *outp;
  85 unsigned int do_rapl;
  86 unsigned int do_dts;
  87 unsigned int do_ptm;
  88 unsigned long long  gfx_cur_rc6_ms;
  89 unsigned long long cpuidle_cur_cpu_lpi_us;
  90 unsigned long long cpuidle_cur_sys_lpi_us;
  91 unsigned int gfx_cur_mhz;
  92 unsigned int tcc_activation_temp;
  93 unsigned int tcc_activation_temp_override;
  94 double rapl_power_units, rapl_time_units;
  95 double rapl_dram_energy_units, rapl_energy_units;
  96 double rapl_joule_counter_range;
  97 unsigned int do_core_perf_limit_reasons;
  98 unsigned int has_automatic_cstate_conversion;
  99 unsigned int do_gfx_perf_limit_reasons;
 100 unsigned int do_ring_perf_limit_reasons;
 101 unsigned int crystal_hz;
 102 unsigned long long tsc_hz;
 103 int base_cpu;
 104 double discover_bclk(unsigned int family, unsigned int model);
 105 unsigned int has_hwp;   /* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
 106                         /* IA32_HWP_REQUEST, IA32_HWP_STATUS */
 107 unsigned int has_hwp_notify;            /* IA32_HWP_INTERRUPT */
 108 unsigned int has_hwp_activity_window;   /* IA32_HWP_REQUEST[bits 41:32] */
 109 unsigned int has_hwp_epp;               /* IA32_HWP_REQUEST[bits 31:24] */
 110 unsigned int has_hwp_pkg;               /* IA32_HWP_REQUEST_PKG */
 111 unsigned int has_misc_feature_control;
 112 unsigned int first_counter_read = 1;
 113
 114 #define RAPL_PKG                (1 << 0)
 115                                         /* 0x610 MSR_PKG_POWER_LIMIT */
 116                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
 117 #define RAPL_PKG_PERF_STATUS    (1 << 1)
 118                                         /* 0x613 MSR_PKG_PERF_STATUS */
 119 #define RAPL_PKG_POWER_INFO     (1 << 2)
 120                                         /* 0x614 MSR_PKG_POWER_INFO */
 121
 122 #define RAPL_DRAM               (1 << 3)
 123                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
 124                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
 125 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
 126                                         /* 0x61b MSR_DRAM_PERF_STATUS */
 127 #define RAPL_DRAM_POWER_INFO    (1 << 5)
 128                                         /* 0x61c MSR_DRAM_POWER_INFO */
 129
 130 #define RAPL_CORES_POWER_LIMIT  (1 << 6)
 131                                         /* 0x638 MSR_PP0_POWER_LIMIT */
 132 #define RAPL_CORE_POLICY        (1 << 7)
 133                                         /* 0x63a MSR_PP0_POLICY */
 134
 135 #define RAPL_GFX                (1 << 8)
 136                                         /* 0x640 MSR_PP1_POWER_LIMIT */
 137                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
 138                                         /* 0x642 MSR_PP1_POLICY */
 139
 140 #define RAPL_CORES_ENERGY_STATUS        (1 << 9)
 141                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
 142 #define RAPL_CORES (RAPL_CORES_ENERGY_STATUS | RAPL_CORES_POWER_LIMIT)
 143 #define TJMAX_DEFAULT   100
 144
 145 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 146
 147 /*
 148  * buffer size used by sscanf() for added column names
 149  * Usually truncated to 7 characters, but also handles 18 columns for raw 64-bit counters
 150  */
 151 #define NAME_BYTES 20
 152 #define PATH_BYTES 128
 153
 154 int backwards_count;
 155 char *progname;
 156
 157 #define CPU_SUBSET_MAXCPUS      1024    /* need to use before probe... */
 158 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_subset;
 159 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_subset_size;
 160 #define MAX_ADDED_COUNTERS 8
 161 #define MAX_ADDED_THREAD_COUNTERS 24
 162 #define BITMASK_SIZE 32
 163
 164 struct thread_data {
 165         struct timeval tv_begin;
 166         struct timeval tv_end;
 167         unsigned long long tsc;
 168         unsigned long long aperf;
 169         unsigned long long mperf;
 170         unsigned long long c1;
 171         unsigned long long  irq_count;
 172         unsigned int smi_count;
 173         unsigned int cpu_id;
 174         unsigned int apic_id;
 175         unsigned int x2apic_id;
 176         unsigned int flags;
 177 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
 178 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
 179         unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
 180 } *thread_even, *thread_odd;
 181
 182 struct core_data {
 183         unsigned long long c3;
 184         unsigned long long c6;
 185         unsigned long long c7;
 186         unsigned long long mc6_us;      /* duplicate as per-core for now, even though per module */
 187         unsigned int core_temp_c;
 188         unsigned int core_id;
 189         unsigned long long counter[MAX_ADDED_COUNTERS];
 190 } *core_even, *core_odd;
 191
 192 struct pkg_data {
 193         unsigned long long pc2;
 194         unsigned long long pc3;
 195         unsigned long long pc6;
 196         unsigned long long pc7;
 197         unsigned long long pc8;
 198         unsigned long long pc9;
 199         unsigned long long pc10;
 200         unsigned long long cpu_lpi;
 201         unsigned long long sys_lpi;
 202         unsigned long long pkg_wtd_core_c0;
 203         unsigned long long pkg_any_core_c0;
 204         unsigned long long pkg_any_gfxe_c0;
 205         unsigned long long pkg_both_core_gfxe_c0;
 206         long long gfx_rc6_ms;
 207         unsigned int gfx_mhz;
 208         unsigned int package_id;
 209         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
 210         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
 211         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
 212         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
 213         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
 214         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
 215         unsigned int pkg_temp_c;
 216         unsigned long long counter[MAX_ADDED_COUNTERS];
 217 } *package_even, *package_odd;
 218
 219 #define ODD_COUNTERS thread_odd, core_odd, package_odd
 220 #define EVEN_COUNTERS thread_even, core_even, package_even
 221
 222 #define GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no)          \
 223         ((thread_base) +                                                      \
 224          ((pkg_no) *                                                          \
 225           topo.nodes_per_pkg * topo.cores_per_node * topo.threads_per_core) + \
 226          ((node_no) * topo.cores_per_node * topo.threads_per_core) +          \
 227          ((core_no) * topo.threads_per_core) +                                \
 228          (thread_no))
 229
 230 #define GET_CORE(core_base, core_no, node_no, pkg_no)                   \
 231         ((core_base) +                                                  \
 232          ((pkg_no) *  topo.nodes_per_pkg * topo.cores_per_node) +       \
 233          ((node_no) * topo.cores_per_node) +                            \
 234          (core_no))
 235
 236
 237 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
 238
 239 enum counter_scope {SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE};
 240 enum counter_type {COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC};
 241 enum counter_format {FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT};
 242
 243 struct msr_counter {
 244         unsigned int msr_num;
 245         char name[NAME_BYTES];
 246         char path[PATH_BYTES];
 247         unsigned int width;
 248         enum counter_type type;
 249         enum counter_format format;
 250         struct msr_counter *next;
 251         unsigned int flags;
 252 #define FLAGS_HIDE      (1 << 0)
 253 #define FLAGS_SHOW      (1 << 1)
 254 #define SYSFS_PERCPU    (1 << 1)
 255 };
 256
 257 struct sys_counters {
 258         unsigned int added_thread_counters;
 259         unsigned int added_core_counters;
 260         unsigned int added_package_counters;
 261         struct msr_counter *tp;
 262         struct msr_counter *cp;
 263         struct msr_counter *pp;
 264 } sys;
 265
 266 struct system_summary {
 267         struct thread_data threads;
 268         struct core_data cores;
 269         struct pkg_data packages;
 270 } average;
 271
 272 struct cpu_topology {
 273         int physical_package_id;
 274         int logical_cpu_id;
 275         int physical_node_id;
 276         int logical_node_id;    /* 0-based count within the package */
 277         int physical_core_id;
 278         int thread_id;
 279         cpu_set_t *put_ids; /* Processing Unit/Thread IDs */
 280 } *cpus;
 281
 282 struct topo_params {
 283         int num_packages;
 284         int num_cpus;
 285         int num_cores;
 286         int max_cpu_num;
 287         int max_node_num;
 288         int nodes_per_pkg;
 289         int cores_per_node;
 290         int threads_per_core;
 291 } topo;
 292
 293 struct timeval tv_even, tv_odd, tv_delta;
 294
 295 int *irq_column_2_cpu;  /* /proc/interrupts column numbers */
 296 int *irqs_per_cpu;              /* indexed by cpu_num */
 297
 298 void setup_all_buffers(void);
 299
 300 int cpu_is_not_present(int cpu)
 301 {
 302         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
 303 }
 304 /*
 305  * run func(thread, core, package) in topology order
 306  * skip non-present cpus
 307  */
 308
 309 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
 310         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
 311 {
 312         int retval, pkg_no, core_no, thread_no, node_no;
 313
 314         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
 315                 for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
 316                         for (node_no = 0; node_no < topo.nodes_per_pkg;
 317                              node_no++) {
 318                                 for (thread_no = 0; thread_no <
 319                                         topo.threads_per_core; ++thread_no) {
 320                                         struct thread_data *t;
 321                                         struct core_data *c;
 322                                         struct pkg_data *p;
 323
 324                                         t = GET_THREAD(thread_base, thread_no,
 325                                                        core_no, node_no,
 326                                                        pkg_no);
 327
 328                                         if (cpu_is_not_present(t->cpu_id))
 329                                                 continue;
 330
 331                                         c = GET_CORE(core_base, core_no,
 332                                                      node_no, pkg_no);
 333                                         p = GET_PKG(pkg_base, pkg_no);
 334
 335                                         retval = func(t, c, p);
 336                                         if (retval)
 337                                                 return retval;
 338                                 }
 339                         }
 340                 }
 341         }
 342         return 0;
 343 }
 344
 345 int cpu_migrate(int cpu)
 346 {
 347         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
 348         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
 349         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
 350                 return -1;
 351         else
 352                 return 0;
 353 }
 354 int get_msr_fd(int cpu)
 355 {
 356         char pathname[32];
 357         int fd;
 358
 359         fd = fd_percpu[cpu];
 360
 361         if (fd)
 362                 return fd;
 363
 364         sprintf(pathname, "/dev/cpu/%d/msr", cpu);
 365         fd = open(pathname, O_RDONLY);
 366         if (fd < 0)
 367                 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
 368
 369         fd_percpu[cpu] = fd;
 370
 371         return fd;
 372 }
 373
 374 int get_msr(int cpu, off_t offset, unsigned long long *msr)
 375 {
 376         ssize_t retval;
 377
 378         retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
 379
 380         if (retval != sizeof *msr)
 381                 err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
 382
 383         return 0;
 384 }
 385
 386 /*
 387  * This list matches the column headers, except
 388  * 1. built-in only, the sysfs counters are not here -- we learn of those at run-time
 389  * 2. Core and CPU are moved to the end, we can't have strings that contain them
 390  *    matching on them for --show and --hide.
 391  */
 392 struct msr_counter bic[] = {
 393         { 0x0, "usec" },
 394         { 0x0, "Time_Of_Day_Seconds" },
 395         { 0x0, "Package" },
 396         { 0x0, "Node" },
 397         { 0x0, "Avg_MHz" },
 398         { 0x0, "Busy%" },
 399         { 0x0, "Bzy_MHz" },
 400         { 0x0, "TSC_MHz" },
 401         { 0x0, "IRQ" },
 402         { 0x0, "SMI", "", 32, 0, FORMAT_DELTA, NULL},
 403         { 0x0, "sysfs" },
 404         { 0x0, "CPU%c1" },
 405         { 0x0, "CPU%c3" },
 406         { 0x0, "CPU%c6" },
 407         { 0x0, "CPU%c7" },
 408         { 0x0, "ThreadC" },
 409         { 0x0, "CoreTmp" },
 410         { 0x0, "CoreCnt" },
 411         { 0x0, "PkgTmp" },
 412         { 0x0, "GFX%rc6" },
 413         { 0x0, "GFXMHz" },
 414         { 0x0, "Pkg%pc2" },
 415         { 0x0, "Pkg%pc3" },
 416         { 0x0, "Pkg%pc6" },
 417         { 0x0, "Pkg%pc7" },
 418         { 0x0, "Pkg%pc8" },
 419         { 0x0, "Pkg%pc9" },
 420         { 0x0, "Pk%pc10" },
 421         { 0x0, "CPU%LPI" },
 422         { 0x0, "SYS%LPI" },
 423         { 0x0, "PkgWatt" },
 424         { 0x0, "CorWatt" },
 425         { 0x0, "GFXWatt" },
 426         { 0x0, "PkgCnt" },
 427         { 0x0, "RAMWatt" },
 428         { 0x0, "PKG_%" },
 429         { 0x0, "RAM_%" },
 430         { 0x0, "Pkg_J" },
 431         { 0x0, "Cor_J" },
 432         { 0x0, "GFX_J" },
 433         { 0x0, "RAM_J" },
 434         { 0x0, "Mod%c6" },
 435         { 0x0, "Totl%C0" },
 436         { 0x0, "Any%C0" },
 437         { 0x0, "GFX%C0" },
 438         { 0x0, "CPUGFX%" },
 439         { 0x0, "Core" },
 440         { 0x0, "CPU" },
 441         { 0x0, "APIC" },
 442         { 0x0, "X2APIC" },
 443 };
 444
 445 #define MAX_BIC (sizeof(bic) / sizeof(struct msr_counter))
 446 #define BIC_USEC        (1ULL << 0)
 447 #define BIC_TOD         (1ULL << 1)
 448 #define BIC_Package     (1ULL << 2)
 449 #define BIC_Node        (1ULL << 3)
 450 #define BIC_Avg_MHz     (1ULL << 4)
 451 #define BIC_Busy        (1ULL << 5)
 452 #define BIC_Bzy_MHz     (1ULL << 6)
 453 #define BIC_TSC_MHz     (1ULL << 7)
 454 #define BIC_IRQ         (1ULL << 8)
 455 #define BIC_SMI         (1ULL << 9)
 456 #define BIC_sysfs       (1ULL << 10)
 457 #define BIC_CPU_c1      (1ULL << 11)
 458 #define BIC_CPU_c3      (1ULL << 12)
 459 #define BIC_CPU_c6      (1ULL << 13)
 460 #define BIC_CPU_c7      (1ULL << 14)
 461 #define BIC_ThreadC     (1ULL << 15)
 462 #define BIC_CoreTmp     (1ULL << 16)
 463 #define BIC_CoreCnt     (1ULL << 17)
 464 #define BIC_PkgTmp      (1ULL << 18)
 465 #define BIC_GFX_rc6     (1ULL << 19)
 466 #define BIC_GFXMHz      (1ULL << 20)
 467 #define BIC_Pkgpc2      (1ULL << 21)
 468 #define BIC_Pkgpc3      (1ULL << 22)
 469 #define BIC_Pkgpc6      (1ULL << 23)
 470 #define BIC_Pkgpc7      (1ULL << 24)
 471 #define BIC_Pkgpc8      (1ULL << 25)
 472 #define BIC_Pkgpc9      (1ULL << 26)
 473 #define BIC_Pkgpc10     (1ULL << 27)
 474 #define BIC_CPU_LPI     (1ULL << 28)
 475 #define BIC_SYS_LPI     (1ULL << 29)
 476 #define BIC_PkgWatt     (1ULL << 30)
 477 #define BIC_CorWatt     (1ULL << 31)
 478 #define BIC_GFXWatt     (1ULL << 32)
 479 #define BIC_PkgCnt      (1ULL << 33)
 480 #define BIC_RAMWatt     (1ULL << 34)
 481 #define BIC_PKG__       (1ULL << 35)
 482 #define BIC_RAM__       (1ULL << 36)
 483 #define BIC_Pkg_J       (1ULL << 37)
 484 #define BIC_Cor_J       (1ULL << 38)
 485 #define BIC_GFX_J       (1ULL << 39)
 486 #define BIC_RAM_J       (1ULL << 40)
 487 #define BIC_Mod_c6      (1ULL << 41)
 488 #define BIC_Totl_c0     (1ULL << 42)
 489 #define BIC_Any_c0      (1ULL << 43)
 490 #define BIC_GFX_c0      (1ULL << 44)
 491 #define BIC_CPUGFX      (1ULL << 45)
 492 #define BIC_Core        (1ULL << 46)
 493 #define BIC_CPU         (1ULL << 47)
 494 #define BIC_APIC        (1ULL << 48)
 495 #define BIC_X2APIC      (1ULL << 49)
 496
 497 #define BIC_DISABLED_BY_DEFAULT (BIC_USEC | BIC_TOD | BIC_APIC | BIC_X2APIC)
 498
 499 unsigned long long bic_enabled = (0xFFFFFFFFFFFFFFFFULL & ~BIC_DISABLED_BY_DEFAULT);
 500 unsigned long long bic_present = BIC_USEC | BIC_TOD | BIC_sysfs | BIC_APIC | BIC_X2APIC;
 501
 502 #define DO_BIC(COUNTER_NAME) (bic_enabled & bic_present & COUNTER_NAME)
 503 #define ENABLE_BIC(COUNTER_NAME) (bic_enabled |= COUNTER_NAME)
 504 #define BIC_PRESENT(COUNTER_BIT) (bic_present |= COUNTER_BIT)
 505 #define BIC_NOT_PRESENT(COUNTER_BIT) (bic_present &= ~COUNTER_BIT)
 506
 507
 508 #define MAX_DEFERRED 16
 509 char *deferred_skip_names[MAX_DEFERRED];
 510 int deferred_skip_index;
 511
 512 /*
 513  * HIDE_LIST - hide this list of counters, show the rest [default]
 514  * SHOW_LIST - show this list of counters, hide the rest
 515  */
 516 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
 517
 518 void help(void)
 519 {
 520         fprintf(outf,
 521         "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
 522         "\n"
 523         "Turbostat forks the specified COMMAND and prints statistics\n"
 524         "when COMMAND completes.\n"
 525         "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
 526         "to print statistics, until interrupted.\n"
 527         "  -a, --add    add a counter\n"
 528         "                 eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
 529         "  -c, --cpu    cpu-set limit output to summary plus cpu-set:\n"
 530         "                 {core | package | j,k,l..m,n-p }\n"
 531         "  -d, --debug  displays usec, Time_Of_Day_Seconds and more debugging\n"
 532         "  -D, --Dump   displays the raw counter values\n"
 533         "  -e, --enable [all | column]\n"
 534         "               shows all or the specified disabled column\n"
 535         "  -H, --hide [column|column,column,...]\n"
 536         "               hide the specified column(s)\n"
 537         "  -i, --interval sec.subsec\n"
 538         "               Override default 5-second measurement interval\n"
 539         "  -J, --Joules displays energy in Joules instead of Watts\n"
 540         "  -l, --list   list column headers only\n"
 541         "  -n, --num_iterations num\n"
 542         "               number of the measurement iterations\n"
 543         "  -o, --out file\n"
 544         "               create or truncate \"file\" for all output\n"
 545         "  -q, --quiet  skip decoding system configuration header\n"
 546         "  -s, --show [column|column,column,...]\n"
 547         "               show only the specified column(s)\n"
 548         "  -S, --Summary\n"
 549         "               limits output to 1-line system summary per interval\n"
 550         "  -T, --TCC temperature\n"
 551         "               sets the Thermal Control Circuit temperature in\n"
 552         "                 degrees Celsius\n"
 553         "  -h, --help   print this help message\n"
 554         "  -v, --version        print version information\n"
 555         "\n"
 556         "For more help, run \"man turbostat\"\n");
 557 }
 558
 559 /*
 560  * bic_lookup
 561  * for all the strings in comma separate name_list,
 562  * set the approprate bit in return value.
 563  */
 564 unsigned long long bic_lookup(char *name_list, enum show_hide_mode mode)
 565 {
 566         int i;
 567         unsigned long long retval = 0;
 568
 569         while (name_list) {
 570                 char *comma;
 571
 572                 comma = strchr(name_list, ',');
 573
 574                 if (comma)
 575                         *comma = '\0';
 576
 577                 if (!strcmp(name_list, "all"))
 578                         return ~0;
 579
 580                 for (i = 0; i < MAX_BIC; ++i) {
 581                         if (!strcmp(name_list, bic[i].name)) {
 582                                 retval |= (1ULL << i);
 583                                 break;
 584                         }
 585                 }
 586                 if (i == MAX_BIC) {
 587                         if (mode == SHOW_LIST) {
 588                                 fprintf(stderr, "Invalid counter name: %s\n", name_list);
 589                                 exit(-1);
 590                         }
 591                         deferred_skip_names[deferred_skip_index++] = name_list;
 592                         if (debug)
 593                                 fprintf(stderr, "deferred \"%s\"\n", name_list);
 594                         if (deferred_skip_index >= MAX_DEFERRED) {
 595                                 fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n",
 596                                         MAX_DEFERRED, name_list);
 597                                 help();
 598                                 exit(1);
 599                         }
 600                 }
 601
 602                 name_list = comma;
 603                 if (name_list)
 604                         name_list++;
 605
 606         }
 607         return retval;
 608 }
 609
 610
 611 void print_header(char *delim)
 612 {
 613         struct msr_counter *mp;
 614         int printed = 0;
 615
 616         if (DO_BIC(BIC_USEC))
 617                 outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
 618         if (DO_BIC(BIC_TOD))
 619                 outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
 620         if (DO_BIC(BIC_Package))
 621                 outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
 622         if (DO_BIC(BIC_Node))
 623                 outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
 624         if (DO_BIC(BIC_Core))
 625                 outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
 626         if (DO_BIC(BIC_CPU))
 627                 outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
 628         if (DO_BIC(BIC_APIC))
 629                 outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
 630         if (DO_BIC(BIC_X2APIC))
 631                 outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
 632         if (DO_BIC(BIC_Avg_MHz))
 633                 outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
 634         if (DO_BIC(BIC_Busy))
 635                 outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
 636         if (DO_BIC(BIC_Bzy_MHz))
 637                 outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
 638         if (DO_BIC(BIC_TSC_MHz))
 639                 outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
 640
 641         if (DO_BIC(BIC_IRQ)) {
 642                 if (sums_need_wide_columns)
 643                         outp += sprintf(outp, "%s     IRQ", (printed++ ? delim : ""));
 644                 else
 645                         outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
 646         }
 647
 648         if (DO_BIC(BIC_SMI))
 649                 outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
 650
 651         for (mp = sys.tp; mp; mp = mp->next) {
 652
 653                 if (mp->format == FORMAT_RAW) {
 654                         if (mp->width == 64)
 655                                 outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), mp->name);
 656                         else
 657                                 outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), mp->name);
 658                 } else {
 659                         if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
 660                                 outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), mp->name);
 661                         else
 662                                 outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), mp->name);
 663                 }
 664         }
 665
 666         if (DO_BIC(BIC_CPU_c1))
 667                 outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
 668         if (DO_BIC(BIC_CPU_c3) && !do_slm_cstates && !do_knl_cstates && !do_cnl_cstates)
 669                 outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
 670         if (DO_BIC(BIC_CPU_c6))
 671                 outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
 672         if (DO_BIC(BIC_CPU_c7))
 673                 outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
 674
 675         if (DO_BIC(BIC_Mod_c6))
 676                 outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
 677
 678         if (DO_BIC(BIC_CoreTmp))
 679                 outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
 680
 681         for (mp = sys.cp; mp; mp = mp->next) {
 682                 if (mp->format == FORMAT_RAW) {
 683                         if (mp->width == 64)
 684                                 outp += sprintf(outp, "%s%18.18s", delim, mp->name);
 685                         else
 686                                 outp += sprintf(outp, "%s%10.10s", delim, mp->name);
 687                 } else {
 688                         if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
 689                                 outp += sprintf(outp, "%s%8s", delim, mp->name);
 690                         else
 691                                 outp += sprintf(outp, "%s%s", delim, mp->name);
 692                 }
 693         }
 694
 695         if (DO_BIC(BIC_PkgTmp))
 696                 outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
 697
 698         if (DO_BIC(BIC_GFX_rc6))
 699                 outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
 700
 701         if (DO_BIC(BIC_GFXMHz))
 702                 outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
 703
 704         if (DO_BIC(BIC_Totl_c0))
 705                 outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
 706         if (DO_BIC(BIC_Any_c0))
 707                 outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
 708         if (DO_BIC(BIC_GFX_c0))
 709                 outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
 710         if (DO_BIC(BIC_CPUGFX))
 711                 outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
 712
 713         if (DO_BIC(BIC_Pkgpc2))
 714                 outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
 715         if (DO_BIC(BIC_Pkgpc3))
 716                 outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
 717         if (DO_BIC(BIC_Pkgpc6))
 718                 outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
 719         if (DO_BIC(BIC_Pkgpc7))
 720                 outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
 721         if (DO_BIC(BIC_Pkgpc8))
 722                 outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
 723         if (DO_BIC(BIC_Pkgpc9))
 724                 outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
 725         if (DO_BIC(BIC_Pkgpc10))
 726                 outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
 727         if (DO_BIC(BIC_CPU_LPI))
 728                 outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
 729         if (DO_BIC(BIC_SYS_LPI))
 730                 outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
 731
 732         if (do_rapl && !rapl_joules) {
 733                 if (DO_BIC(BIC_PkgWatt))
 734                         outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
 735                 if (DO_BIC(BIC_CorWatt))
 736                         outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
 737                 if (DO_BIC(BIC_GFXWatt))
 738                         outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
 739                 if (DO_BIC(BIC_RAMWatt))
 740                         outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
 741                 if (DO_BIC(BIC_PKG__))
 742                         outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
 743                 if (DO_BIC(BIC_RAM__))
 744                         outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
 745         } else if (do_rapl && rapl_joules) {
 746                 if (DO_BIC(BIC_Pkg_J))
 747                         outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
 748                 if (DO_BIC(BIC_Cor_J))
 749                         outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
 750                 if (DO_BIC(BIC_GFX_J))
 751                         outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
 752                 if (DO_BIC(BIC_RAM_J))
 753                         outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
 754                 if (DO_BIC(BIC_PKG__))
 755                         outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
 756                 if (DO_BIC(BIC_RAM__))
 757                         outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
 758         }
 759         for (mp = sys.pp; mp; mp = mp->next) {
 760                 if (mp->format == FORMAT_RAW) {
 761                         if (mp->width == 64)
 762                                 outp += sprintf(outp, "%s%18.18s", delim, mp->name);
 763                         else
 764                                 outp += sprintf(outp, "%s%10.10s", delim, mp->name);
 765                 } else {
 766                         if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
 767                                 outp += sprintf(outp, "%s%8s", delim, mp->name);
 768                         else
 769                                 outp += sprintf(outp, "%s%s", delim, mp->name);
 770                 }
 771         }
 772
 773         outp += sprintf(outp, "\n");
 774 }
 775
 776 int dump_counters(struct thread_data *t, struct core_data *c,
 777         struct pkg_data *p)
 778 {
 779         int i;
 780         struct msr_counter *mp;
 781
 782         outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
 783
 784         if (t) {
 785                 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
 786                         t->cpu_id, t->flags);
 787                 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
 788                 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
 789                 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
 790                 outp += sprintf(outp, "c1: %016llX\n", t->c1);
 791
 792                 if (DO_BIC(BIC_IRQ))
 793                         outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
 794                 if (DO_BIC(BIC_SMI))
 795                         outp += sprintf(outp, "SMI: %d\n", t->smi_count);
 796
 797                 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
 798                         outp += sprintf(outp, "tADDED [%d] msr0x%x: %08llX\n",
 799                                 i, mp->msr_num, t->counter[i]);
 800                 }
 801         }
 802
 803         if (c) {
 804                 outp += sprintf(outp, "core: %d\n", c->core_id);
 805                 outp += sprintf(outp, "c3: %016llX\n", c->c3);
 806                 outp += sprintf(outp, "c6: %016llX\n", c->c6);
 807                 outp += sprintf(outp, "c7: %016llX\n", c->c7);
 808                 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
 809
 810                 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
 811                         outp += sprintf(outp, "cADDED [%d] msr0x%x: %08llX\n",
 812                                 i, mp->msr_num, c->counter[i]);
 813                 }
 814                 outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
 815         }
 816
 817         if (p) {
 818                 outp += sprintf(outp, "package: %d\n", p->package_id);
 819
 820                 outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
 821                 outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
 822                 outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
 823                 outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
 824
 825                 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
 826                 if (DO_BIC(BIC_Pkgpc3))
 827                         outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
 828                 if (DO_BIC(BIC_Pkgpc6))
 829                         outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
 830                 if (DO_BIC(BIC_Pkgpc7))
 831                         outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
 832                 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
 833                 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
 834                 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
 835                 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
 836                 outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
 837                 outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
 838                 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
 839                 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
 840                 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
 841                 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
 842                 outp += sprintf(outp, "Throttle PKG: %0X\n",
 843                         p->rapl_pkg_perf_status);
 844                 outp += sprintf(outp, "Throttle RAM: %0X\n",
 845                         p->rapl_dram_perf_status);
 846                 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
 847
 848                 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
 849                         outp += sprintf(outp, "pADDED [%d] msr0x%x: %08llX\n",
 850                                 i, mp->msr_num, p->counter[i]);
 851                 }
 852         }
 853
 854         outp += sprintf(outp, "\n");
 855
 856         return 0;
 857 }
 858
 859 /*
 860  * column formatting convention & formats
 861  */
 862 int format_counters(struct thread_data *t, struct core_data *c,
 863         struct pkg_data *p)
 864 {
 865         double interval_float, tsc;
 866         char *fmt8;
 867         int i;
 868         struct msr_counter *mp;
 869         char *delim = "\t";
 870         int printed = 0;
 871
 872          /* if showing only 1st thread in core and this isn't one, bail out */
 873         if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
 874                 return 0;
 875
 876          /* if showing only 1st thread in pkg and this isn't one, bail out */
 877         if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
 878                 return 0;
 879
 880         /*if not summary line and --cpu is used */
 881         if ((t != &average.threads) &&
 882                 (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
 883                 return 0;
 884
 885         if (DO_BIC(BIC_USEC)) {
 886                 /* on each row, print how many usec each timestamp took to gather */
 887                 struct timeval tv;
 888
 889                 timersub(&t->tv_end, &t->tv_begin, &tv);
 890                 outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
 891         }
 892
 893         /* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
 894         if (DO_BIC(BIC_TOD))
 895                 outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
 896
 897         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
 898
 899         tsc = t->tsc * tsc_tweak;
 900
 901         /* topo columns, print blanks on 1st (average) line */
 902         if (t == &average.threads) {
 903                 if (DO_BIC(BIC_Package))
 904                         outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 905                 if (DO_BIC(BIC_Node))
 906                         outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 907                 if (DO_BIC(BIC_Core))
 908                         outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 909                 if (DO_BIC(BIC_CPU))
 910                         outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 911                 if (DO_BIC(BIC_APIC))
 912                         outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 913                 if (DO_BIC(BIC_X2APIC))
 914                         outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 915         } else {
 916                 if (DO_BIC(BIC_Package)) {
 917                         if (p)
 918                                 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->package_id);
 919                         else
 920                                 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 921                 }
 922                 if (DO_BIC(BIC_Node)) {
 923                         if (t)
 924                                 outp += sprintf(outp, "%s%d",
 925                                                 (printed++ ? delim : ""),
 926                                               cpus[t->cpu_id].physical_node_id);
 927                         else
 928                                 outp += sprintf(outp, "%s-",
 929                                                 (printed++ ? delim : ""));
 930                 }
 931                 if (DO_BIC(BIC_Core)) {
 932                         if (c)
 933                                 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_id);
 934                         else
 935                                 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
 936                 }
 937                 if (DO_BIC(BIC_CPU))
 938                         outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
 939                 if (DO_BIC(BIC_APIC))
 940                         outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->apic_id);
 941                 if (DO_BIC(BIC_X2APIC))
 942                         outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->x2apic_id);
 943         }
 944
 945         if (DO_BIC(BIC_Avg_MHz))
 946                 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
 947                         1.0 / units * t->aperf / interval_float);
 948
 949         if (DO_BIC(BIC_Busy))
 950                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->mperf/tsc);
 951
 952         if (DO_BIC(BIC_Bzy_MHz)) {
 953                 if (has_base_hz)
 954                         outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
 955                 else
 956                         outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
 957                                 tsc / units * t->aperf / t->mperf / interval_float);
 958         }
 959
 960         if (DO_BIC(BIC_TSC_MHz))
 961                 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc/units/interval_float);
 962
 963         /* IRQ */
 964         if (DO_BIC(BIC_IRQ)) {
 965                 if (sums_need_wide_columns)
 966                         outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
 967                 else
 968                         outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
 969         }
 970
 971         /* SMI */
 972         if (DO_BIC(BIC_SMI))
 973                 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
 974
 975         /* Added counters */
 976         for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
 977                 if (mp->format == FORMAT_RAW) {
 978                         if (mp->width == 32)
 979                                 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) t->counter[i]);
 980                         else
 981                                 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->counter[i]);
 982                 } else if (mp->format == FORMAT_DELTA) {
 983                         if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
 984                                 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->counter[i]);
 985                         else
 986                                 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->counter[i]);
 987                 } else if (mp->format == FORMAT_PERCENT) {
 988                         if (mp->type == COUNTER_USEC)
 989                                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), t->counter[i]/interval_float/10000);
 990                         else
 991                                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->counter[i]/tsc);
 992                 }
 993         }
 994
 995         /* C1 */
 996         if (DO_BIC(BIC_CPU_c1))
 997                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->c1/tsc);
 998
 999
1000         /* print per-core data only for 1st thread in core */
1001         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1002                 goto done;
1003
1004         if (DO_BIC(BIC_CPU_c3) && !do_slm_cstates && !do_knl_cstates && !do_cnl_cstates)
1005                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c3/tsc);
1006         if (DO_BIC(BIC_CPU_c6))
1007                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c6/tsc);
1008         if (DO_BIC(BIC_CPU_c7))
1009                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c7/tsc);
1010
1011         /* Mod%c6 */
1012         if (DO_BIC(BIC_Mod_c6))
1013                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->mc6_us / tsc);
1014
1015         if (DO_BIC(BIC_CoreTmp))
1016                 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
1017
1018         for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1019                 if (mp->format == FORMAT_RAW) {
1020                         if (mp->width == 32)
1021                                 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) c->counter[i]);
1022                         else
1023                                 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->counter[i]);
1024                 } else if (mp->format == FORMAT_DELTA) {
1025                         if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
1026                                 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->counter[i]);
1027                         else
1028                                 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->counter[i]);
1029                 } else if (mp->format == FORMAT_PERCENT) {
1030                         outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->counter[i]/tsc);
1031                 }
1032         }
1033
1034         /* print per-package data only for 1st core in package */
1035         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1036                 goto done;
1037
1038         /* PkgTmp */
1039         if (DO_BIC(BIC_PkgTmp))
1040                 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
1041
1042         /* GFXrc6 */
1043         if (DO_BIC(BIC_GFX_rc6)) {
1044                 if (p->gfx_rc6_ms == -1) {      /* detect GFX counter reset */
1045                         outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
1046                 } else {
1047                         outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
1048                                 p->gfx_rc6_ms / 10.0 / interval_float);
1049                 }
1050         }
1051
1052         /* GFXMHz */
1053         if (DO_BIC(BIC_GFXMHz))
1054                 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
1055
1056         /* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
1057         if (DO_BIC(BIC_Totl_c0))
1058                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_wtd_core_c0/tsc);
1059         if (DO_BIC(BIC_Any_c0))
1060                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_core_c0/tsc);
1061         if (DO_BIC(BIC_GFX_c0))
1062                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_gfxe_c0/tsc);
1063         if (DO_BIC(BIC_CPUGFX))
1064                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_both_core_gfxe_c0/tsc);
1065
1066         if (DO_BIC(BIC_Pkgpc2))
1067                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc2/tsc);
1068         if (DO_BIC(BIC_Pkgpc3))
1069                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc3/tsc);
1070         if (DO_BIC(BIC_Pkgpc6))
1071                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc6/tsc);
1072         if (DO_BIC(BIC_Pkgpc7))
1073                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc7/tsc);
1074         if (DO_BIC(BIC_Pkgpc8))
1075                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc8/tsc);
1076         if (DO_BIC(BIC_Pkgpc9))
1077                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc9/tsc);
1078         if (DO_BIC(BIC_Pkgpc10))
1079                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc10/tsc);
1080
1081         if (DO_BIC(BIC_CPU_LPI))
1082                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->cpu_lpi / 1000000.0 / interval_float);
1083         if (DO_BIC(BIC_SYS_LPI))
1084                 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->sys_lpi / 1000000.0 / interval_float);
1085
1086         /*
1087          * If measurement interval exceeds minimum RAPL Joule Counter range,
1088          * indicate that results are suspect by printing "**" in fraction place.
1089          */
1090         if (interval_float < rapl_joule_counter_range)
1091                 fmt8 = "%s%.2f";
1092         else
1093                 fmt8 = "%6.0f**";
1094
1095         if (DO_BIC(BIC_PkgWatt))
1096                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units / interval_float);
1097         if (DO_BIC(BIC_CorWatt))
1098                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units / interval_float);
1099         if (DO_BIC(BIC_GFXWatt))
1100                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units / interval_float);
1101         if (DO_BIC(BIC_RAMWatt))
1102                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_dram * rapl_dram_energy_units / interval_float);
1103         if (DO_BIC(BIC_Pkg_J))
1104                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units);
1105         if (DO_BIC(BIC_Cor_J))
1106                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units);
1107         if (DO_BIC(BIC_GFX_J))
1108                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units);
1109         if (DO_BIC(BIC_RAM_J))
1110                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_dram * rapl_dram_energy_units);
1111         if (DO_BIC(BIC_PKG__))
1112                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
1113         if (DO_BIC(BIC_RAM__))
1114                 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
1115
1116         for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1117                 if (mp->format == FORMAT_RAW) {
1118                         if (mp->width == 32)
1119                                 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) p->counter[i]);
1120                         else
1121                                 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->counter[i]);
1122                 } else if (mp->format == FORMAT_DELTA) {
1123                         if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
1124                                 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->counter[i]);
1125                         else
1126                                 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->counter[i]);
1127                 } else if (mp->format == FORMAT_PERCENT) {
1128                         outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->counter[i]/tsc);
1129                 }
1130         }
1131
1132 done:
1133         if (*(outp - 1) != '\n')
1134                 outp += sprintf(outp, "\n");
1135
1136         return 0;
1137 }
1138
1139 void flush_output_stdout(void)
1140 {
1141         FILE *filep;
1142
1143         if (outf == stderr)
1144                 filep = stdout;
1145         else
1146                 filep = outf;
1147
1148         fputs(output_buffer, filep);
1149         fflush(filep);
1150
1151         outp = output_buffer;
1152 }
1153 void flush_output_stderr(void)
1154 {
1155         fputs(output_buffer, outf);
1156         fflush(outf);
1157         outp = output_buffer;
1158 }
1159 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1160 {
1161         static int printed;
1162
1163         if (!printed || !summary_only)
1164                 print_header("\t");
1165
1166         format_counters(&average.threads, &average.cores, &average.packages);
1167
1168         printed = 1;
1169
1170         if (summary_only)
1171                 return;
1172
1173         for_all_cpus(format_counters, t, c, p);
1174 }
1175
1176 #define DELTA_WRAP32(new, old)                  \
1177         if (new > old) {                        \
1178                 old = new - old;                \
1179         } else {                                \
1180                 old = 0x100000000 + new - old;  \
1181         }
1182
1183 int
1184 delta_package(struct pkg_data *new, struct pkg_data *old)
1185 {
1186         int i;
1187         struct msr_counter *mp;
1188
1189
1190         if (DO_BIC(BIC_Totl_c0))
1191                 old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
1192         if (DO_BIC(BIC_Any_c0))
1193                 old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
1194         if (DO_BIC(BIC_GFX_c0))
1195                 old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
1196         if (DO_BIC(BIC_CPUGFX))
1197                 old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
1198
1199         old->pc2 = new->pc2 - old->pc2;
1200         if (DO_BIC(BIC_Pkgpc3))
1201                 old->pc3 = new->pc3 - old->pc3;
1202         if (DO_BIC(BIC_Pkgpc6))
1203                 old->pc6 = new->pc6 - old->pc6;
1204         if (DO_BIC(BIC_Pkgpc7))
1205                 old->pc7 = new->pc7 - old->pc7;
1206         old->pc8 = new->pc8 - old->pc8;
1207         old->pc9 = new->pc9 - old->pc9;
1208         old->pc10 = new->pc10 - old->pc10;
1209         old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
1210         old->sys_lpi = new->sys_lpi - old->sys_lpi;
1211         old->pkg_temp_c = new->pkg_temp_c;
1212
1213         /* flag an error when rc6 counter resets/wraps */
1214         if (old->gfx_rc6_ms >  new->gfx_rc6_ms)
1215                 old->gfx_rc6_ms = -1;
1216         else
1217                 old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
1218
1219         old->gfx_mhz = new->gfx_mhz;
1220
1221         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
1222         DELTA_WRAP32(new->energy_cores, old->energy_cores);
1223         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
1224         DELTA_WRAP32(new->energy_dram, old->energy_dram);
1225         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
1226         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
1227
1228         for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1229                 if (mp->format == FORMAT_RAW)
1230                         old->counter[i] = new->counter[i];
1231                 else
1232                         old->counter[i] = new->counter[i] - old->counter[i];
1233         }
1234
1235         return 0;
1236 }
1237
1238 void
1239 delta_core(struct core_data *new, struct core_data *old)
1240 {
1241         int i;
1242         struct msr_counter *mp;
1243
1244         old->c3 = new->c3 - old->c3;
1245         old->c6 = new->c6 - old->c6;
1246         old->c7 = new->c7 - old->c7;
1247         old->core_temp_c = new->core_temp_c;
1248         old->mc6_us = new->mc6_us - old->mc6_us;
1249
1250         for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1251                 if (mp->format == FORMAT_RAW)
1252                         old->counter[i] = new->counter[i];
1253                 else
1254                         old->counter[i] = new->counter[i] - old->counter[i];
1255         }
1256 }
1257
1258 /*
1259  * old = new - old
1260  */
1261 int
1262 delta_thread(struct thread_data *new, struct thread_data *old,
1263         struct core_data *core_delta)
1264 {
1265         int i;
1266         struct msr_counter *mp;
1267
1268         /* we run cpuid just the 1st time, copy the results */
1269         if (DO_BIC(BIC_APIC))
1270                 new->apic_id = old->apic_id;
1271         if (DO_BIC(BIC_X2APIC))
1272                 new->x2apic_id = old->x2apic_id;
1273
1274         /*
1275          * the timestamps from start of measurement interval are in "old"
1276          * the timestamp from end of measurement interval are in "new"
1277          * over-write old w/ new so we can print end of interval values
1278          */
1279
1280         old->tv_begin = new->tv_begin;
1281         old->tv_end = new->tv_end;
1282
1283         old->tsc = new->tsc - old->tsc;
1284
1285         /* check for TSC < 1 Mcycles over interval */
1286         if (old->tsc < (1000 * 1000))
1287                 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
1288                      "You can disable all c-states by booting with \"idle=poll\"\n"
1289                      "or just the deep ones with \"processor.max_cstate=1\"");
1290
1291         old->c1 = new->c1 - old->c1;
1292
1293         if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz)) {
1294                 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
1295                         old->aperf = new->aperf - old->aperf;
1296                         old->mperf = new->mperf - old->mperf;
1297                 } else {
1298                         return -1;
1299                 }
1300         }
1301
1302
1303         if (use_c1_residency_msr) {
1304                 /*
1305                  * Some models have a dedicated C1 residency MSR,
1306                  * which should be more accurate than the derivation below.
1307                  */
1308         } else {
1309                 /*
1310                  * As counter collection is not atomic,
1311                  * it is possible for mperf's non-halted cycles + idle states
1312                  * to exceed TSC's all cycles: show c1 = 0% in that case.
1313                  */
1314                 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
1315                         old->c1 = 0;
1316                 else {
1317                         /* normal case, derive c1 */
1318                         old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3
1319                                 - core_delta->c6 - core_delta->c7;
1320                 }
1321         }
1322
1323         if (old->mperf == 0) {
1324                 if (debug > 1)
1325                         fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
1326                 old->mperf = 1; /* divide by 0 protection */
1327         }
1328
1329         if (DO_BIC(BIC_IRQ))
1330                 old->irq_count = new->irq_count - old->irq_count;
1331
1332         if (DO_BIC(BIC_SMI))
1333                 old->smi_count = new->smi_count - old->smi_count;
1334
1335         for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1336                 if (mp->format == FORMAT_RAW)
1337                         old->counter[i] = new->counter[i];
1338                 else
1339                         old->counter[i] = new->counter[i] - old->counter[i];
1340         }
1341         return 0;
1342 }
1343
1344 int delta_cpu(struct thread_data *t, struct core_data *c,
1345         struct pkg_data *p, struct thread_data *t2,
1346         struct core_data *c2, struct pkg_data *p2)
1347 {
1348         int retval = 0;
1349
1350         /* calculate core delta only for 1st thread in core */
1351         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
1352                 delta_core(c, c2);
1353
1354         /* always calculate thread delta */
1355         retval = delta_thread(t, t2, c2);       /* c2 is core delta */
1356         if (retval)
1357                 return retval;
1358
1359         /* calculate package delta only for 1st core in package */
1360         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
1361                 retval = delta_package(p, p2);
1362
1363         return retval;
1364 }
1365
1366 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1367 {
1368         int i;
1369         struct msr_counter  *mp;
1370
1371         t->tv_begin.tv_sec = 0;
1372         t->tv_begin.tv_usec = 0;
1373         t->tv_end.tv_sec = 0;
1374         t->tv_end.tv_usec = 0;
1375
1376         t->tsc = 0;
1377         t->aperf = 0;
1378         t->mperf = 0;
1379         t->c1 = 0;
1380
1381         t->irq_count = 0;
1382         t->smi_count = 0;
1383
1384         /* tells format_counters to dump all fields from this set */
1385         t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
1386
1387         c->c3 = 0;
1388         c->c6 = 0;
1389         c->c7 = 0;
1390         c->mc6_us = 0;
1391         c->core_temp_c = 0;
1392
1393         p->pkg_wtd_core_c0 = 0;
1394         p->pkg_any_core_c0 = 0;
1395         p->pkg_any_gfxe_c0 = 0;
1396         p->pkg_both_core_gfxe_c0 = 0;
1397
1398         p->pc2 = 0;
1399         if (DO_BIC(BIC_Pkgpc3))
1400                 p->pc3 = 0;
1401         if (DO_BIC(BIC_Pkgpc6))
1402                 p->pc6 = 0;
1403         if (DO_BIC(BIC_Pkgpc7))
1404                 p->pc7 = 0;
1405         p->pc8 = 0;
1406         p->pc9 = 0;
1407         p->pc10 = 0;
1408         p->cpu_lpi = 0;
1409         p->sys_lpi = 0;
1410
1411         p->energy_pkg = 0;
1412         p->energy_dram = 0;
1413         p->energy_cores = 0;
1414         p->energy_gfx = 0;
1415         p->rapl_pkg_perf_status = 0;
1416         p->rapl_dram_perf_status = 0;
1417         p->pkg_temp_c = 0;
1418
1419         p->gfx_rc6_ms = 0;
1420         p->gfx_mhz = 0;
1421         for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
1422                 t->counter[i] = 0;
1423
1424         for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
1425                 c->counter[i] = 0;
1426
1427         for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
1428                 p->counter[i] = 0;
1429 }
1430 int sum_counters(struct thread_data *t, struct core_data *c,
1431         struct pkg_data *p)
1432 {
1433         int i;
1434         struct msr_counter *mp;
1435
1436         /* copy un-changing apic_id's */
1437         if (DO_BIC(BIC_APIC))
1438                 average.threads.apic_id = t->apic_id;
1439         if (DO_BIC(BIC_X2APIC))
1440                 average.threads.x2apic_id = t->x2apic_id;
1441
1442         /* remember first tv_begin */
1443         if (average.threads.tv_begin.tv_sec == 0)
1444                 average.threads.tv_begin = t->tv_begin;
1445
1446         /* remember last tv_end */
1447         average.threads.tv_end = t->tv_end;
1448
1449         average.threads.tsc += t->tsc;
1450         average.threads.aperf += t->aperf;
1451         average.threads.mperf += t->mperf;
1452         average.threads.c1 += t->c1;
1453
1454         average.threads.irq_count += t->irq_count;
1455         average.threads.smi_count += t->smi_count;
1456
1457         for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1458                 if (mp->format == FORMAT_RAW)
1459                         continue;
1460                 average.threads.counter[i] += t->counter[i];
1461         }
1462
1463         /* sum per-core values only for 1st thread in core */
1464         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1465                 return 0;
1466
1467         average.cores.c3 += c->c3;
1468         average.cores.c6 += c->c6;
1469         average.cores.c7 += c->c7;
1470         average.cores.mc6_us += c->mc6_us;
1471
1472         average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
1473
1474         for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1475                 if (mp->format == FORMAT_RAW)
1476                         continue;
1477                 average.cores.counter[i] += c->counter[i];
1478         }
1479
1480         /* sum per-pkg values only for 1st core in pkg */
1481         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1482                 return 0;
1483
1484         if (DO_BIC(BIC_Totl_c0))
1485                 average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
1486         if (DO_BIC(BIC_Any_c0))
1487                 average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
1488         if (DO_BIC(BIC_GFX_c0))
1489                 average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
1490         if (DO_BIC(BIC_CPUGFX))
1491                 average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
1492
1493         average.packages.pc2 += p->pc2;
1494         if (DO_BIC(BIC_Pkgpc3))
1495                 average.packages.pc3 += p->pc3;
1496         if (DO_BIC(BIC_Pkgpc6))
1497                 average.packages.pc6 += p->pc6;
1498         if (DO_BIC(BIC_Pkgpc7))
1499                 average.packages.pc7 += p->pc7;
1500         average.packages.pc8 += p->pc8;
1501         average.packages.pc9 += p->pc9;
1502         average.packages.pc10 += p->pc10;
1503
1504         average.packages.cpu_lpi = p->cpu_lpi;
1505         average.packages.sys_lpi = p->sys_lpi;
1506
1507         average.packages.energy_pkg += p->energy_pkg;
1508         average.packages.energy_dram += p->energy_dram;
1509         average.packages.energy_cores += p->energy_cores;
1510         average.packages.energy_gfx += p->energy_gfx;
1511
1512         average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
1513         average.packages.gfx_mhz = p->gfx_mhz;
1514
1515         average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
1516
1517         average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
1518         average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
1519
1520         for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1521                 if (mp->format == FORMAT_RAW)
1522                         continue;
1523                 average.packages.counter[i] += p->counter[i];
1524         }
1525         return 0;
1526 }
1527 /*
1528  * sum the counters for all cpus in the system
1529  * compute the weighted average
1530  */
1531 void compute_average(struct thread_data *t, struct core_data *c,
1532         struct pkg_data *p)
1533 {
1534         int i;
1535         struct msr_counter *mp;
1536
1537         clear_counters(&average.threads, &average.cores, &average.packages);
1538
1539         for_all_cpus(sum_counters, t, c, p);
1540
1541         average.threads.tsc /= topo.num_cpus;
1542         average.threads.aperf /= topo.num_cpus;
1543         average.threads.mperf /= topo.num_cpus;
1544         average.threads.c1 /= topo.num_cpus;
1545
1546         if (average.threads.irq_count > 9999999)
1547                 sums_need_wide_columns = 1;
1548
1549         average.cores.c3 /= topo.num_cores;
1550         average.cores.c6 /= topo.num_cores;
1551         average.cores.c7 /= topo.num_cores;
1552         average.cores.mc6_us /= topo.num_cores;
1553
1554         if (DO_BIC(BIC_Totl_c0))
1555                 average.packages.pkg_wtd_core_c0 /= topo.num_packages;
1556         if (DO_BIC(BIC_Any_c0))
1557                 average.packages.pkg_any_core_c0 /= topo.num_packages;
1558         if (DO_BIC(BIC_GFX_c0))
1559                 average.packages.pkg_any_gfxe_c0 /= topo.num_packages;
1560         if (DO_BIC(BIC_CPUGFX))
1561                 average.packages.pkg_both_core_gfxe_c0 /= topo.num_packages;
1562
1563         average.packages.pc2 /= topo.num_packages;
1564         if (DO_BIC(BIC_Pkgpc3))
1565                 average.packages.pc3 /= topo.num_packages;
1566         if (DO_BIC(BIC_Pkgpc6))
1567                 average.packages.pc6 /= topo.num_packages;
1568         if (DO_BIC(BIC_Pkgpc7))
1569                 average.packages.pc7 /= topo.num_packages;
1570
1571         average.packages.pc8 /= topo.num_packages;
1572         average.packages.pc9 /= topo.num_packages;
1573         average.packages.pc10 /= topo.num_packages;
1574
1575         for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1576                 if (mp->format == FORMAT_RAW)
1577                         continue;
1578                 if (mp->type == COUNTER_ITEMS) {
1579                         if (average.threads.counter[i] > 9999999)
1580                                 sums_need_wide_columns = 1;
1581                         continue;
1582                 }
1583                 average.threads.counter[i] /= topo.num_cpus;
1584         }
1585         for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1586                 if (mp->format == FORMAT_RAW)
1587                         continue;
1588                 if (mp->type == COUNTER_ITEMS) {
1589                         if (average.cores.counter[i] > 9999999)
1590                                 sums_need_wide_columns = 1;
1591                 }
1592                 average.cores.counter[i] /= topo.num_cores;
1593         }
1594         for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1595                 if (mp->format == FORMAT_RAW)
1596                         continue;
1597                 if (mp->type == COUNTER_ITEMS) {
1598                         if (average.packages.counter[i] > 9999999)
1599                                 sums_need_wide_columns = 1;
1600                 }
1601                 average.packages.counter[i] /= topo.num_packages;
1602         }
1603 }
1604
1605 static unsigned long long rdtsc(void)
1606 {
1607         unsigned int low, high;
1608
1609         asm volatile("rdtsc" : "=a" (low), "=d" (high));
1610
1611         return low | ((unsigned long long)high) << 32;
1612 }
1613
1614 /*
1615  * Open a file, and exit on failure
1616  */
1617 FILE *fopen_or_die(const char *path, const char *mode)
1618 {
1619         FILE *filep = fopen(path, mode);
1620
1621         if (!filep)
1622                 err(1, "%s: open failed", path);
1623         return filep;
1624 }
1625 /*
1626  * snapshot_sysfs_counter()
1627  *
1628  * return snapshot of given counter
1629  */
1630 unsigned long long snapshot_sysfs_counter(char *path)
1631 {
1632         FILE *fp;
1633         int retval;
1634         unsigned long long counter;
1635
1636         fp = fopen_or_die(path, "r");
1637
1638         retval = fscanf(fp, "%lld", &counter);
1639         if (retval != 1)
1640                 err(1, "snapshot_sysfs_counter(%s)", path);
1641
1642         fclose(fp);
1643
1644         return counter;
1645 }
1646
1647 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp)
1648 {
1649         if (mp->msr_num != 0) {
1650                 if (get_msr(cpu, mp->msr_num, counterp))
1651                         return -1;
1652         } else {
1653                 char path[128 + PATH_BYTES];
1654
1655                 if (mp->flags & SYSFS_PERCPU) {
1656                         sprintf(path, "/sys/devices/system/cpu/cpu%d/%s",
1657                                  cpu, mp->path);
1658
1659                         *counterp = snapshot_sysfs_counter(path);
1660                 } else {
1661                         *counterp = snapshot_sysfs_counter(mp->path);
1662                 }
1663         }
1664
1665         return 0;
1666 }
1667
1668 void get_apic_id(struct thread_data *t)
1669 {
1670         unsigned int eax, ebx, ecx, edx, max_level;
1671
1672         eax = ebx = ecx = edx = 0;
1673
1674         if (!genuine_intel)
1675                 return;
1676
1677         __cpuid(0, max_level, ebx, ecx, edx);
1678
1679         __cpuid(1, eax, ebx, ecx, edx);
1680         t->apic_id = (ebx >> 24) & 0xf;
1681
1682         if (max_level < 0xb)
1683                 return;
1684
1685         if (!DO_BIC(BIC_X2APIC))
1686                 return;
1687
1688         ecx = 0;
1689         __cpuid(0xb, eax, ebx, ecx, edx);
1690         t->x2apic_id = edx;
1691
1692         if (debug && (t->apic_id != t->x2apic_id))
1693                 fprintf(outf, "cpu%d: apic 0x%x x2apic 0x%x\n", t->cpu_id, t->apic_id, t->x2apic_id);
1694 }
1695
1696 /*
1697  * get_counters(...)
1698  * migrate to cpu
1699  * acquire and record local counters for that cpu
1700  */
1701 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1702 {
1703         int cpu = t->cpu_id;
1704         unsigned long long msr;
1705         int aperf_mperf_retry_count = 0;
1706         struct msr_counter *mp;
1707         int i;
1708
1709         gettimeofday(&t->tv_begin, (struct timezone *)NULL);
1710
1711         if (cpu_migrate(cpu)) {
1712                 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
1713                 return -1;
1714         }
1715
1716         if (first_counter_read)
1717                 get_apic_id(t);
1718 retry:
1719         t->tsc = rdtsc();       /* we are running on local CPU of interest */
1720
1721         if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz)) {
1722                 unsigned long long tsc_before, tsc_between, tsc_after, aperf_time, mperf_time;
1723
1724                 /*
1725                  * The TSC, APERF and MPERF must be read together for
1726                  * APERF/MPERF and MPERF/TSC to give accurate results.
1727                  *
1728                  * Unfortunately, APERF and MPERF are read by
1729                  * individual system call, so delays may occur
1730                  * between them.  If the time to read them
1731                  * varies by a large amount, we re-read them.
1732                  */
1733
1734                 /*
1735                  * This initial dummy APERF read has been seen to
1736                  * reduce jitter in the subsequent reads.
1737                  */
1738
1739                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1740                         return -3;
1741
1742                 t->tsc = rdtsc();       /* re-read close to APERF */
1743
1744                 tsc_before = t->tsc;
1745
1746                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1747                         return -3;
1748
1749                 tsc_between = rdtsc();
1750
1751                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
1752                         return -4;
1753
1754                 tsc_after = rdtsc();
1755
1756                 aperf_time = tsc_between - tsc_before;
1757                 mperf_time = tsc_after - tsc_between;
1758
1759                 /*
1760                  * If the system call latency to read APERF and MPERF
1761                  * differ by more than 2x, then try again.
1762                  */
1763                 if ((aperf_time > (2 * mperf_time)) || (mperf_time > (2 * aperf_time))) {
1764                         aperf_mperf_retry_count++;
1765                         if (aperf_mperf_retry_count < 5)
1766                                 goto retry;
1767                         else
1768                                 warnx("cpu%d jitter %lld %lld",
1769                                         cpu, aperf_time, mperf_time);
1770                 }
1771                 aperf_mperf_retry_count = 0;
1772
1773                 t->aperf = t->aperf * aperf_mperf_multiplier;
1774                 t->mperf = t->mperf * aperf_mperf_multiplier;
1775         }
1776
1777         if (DO_BIC(BIC_IRQ))
1778                 t->irq_count = irqs_per_cpu[cpu];
1779         if (DO_BIC(BIC_SMI)) {
1780                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
1781                         return -5;
1782                 t->smi_count = msr & 0xFFFFFFFF;
1783         }
1784         if (DO_BIC(BIC_CPU_c1) && use_c1_residency_msr) {
1785                 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
1786                         return -6;
1787         }
1788
1789         for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1790                 if (get_mp(cpu, mp, &t->counter[i]))
1791                         return -10;
1792         }
1793
1794         /* collect core counters only for 1st thread in core */
1795         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1796                 goto done;
1797
1798         if (DO_BIC(BIC_CPU_c3) && !do_slm_cstates && !do_knl_cstates && !do_cnl_cstates) {
1799                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
1800                         return -6;
1801         }
1802
1803         if (DO_BIC(BIC_CPU_c6) && !do_knl_cstates) {
1804                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
1805                         return -7;
1806         } else if (do_knl_cstates) {
1807                 if (get_msr(cpu, MSR_KNL_CORE_C6_RESIDENCY, &c->c6))
1808                         return -7;
1809         }
1810
1811         if (DO_BIC(BIC_CPU_c7))
1812                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
1813                         return -8;
1814
1815         if (DO_BIC(BIC_Mod_c6))
1816                 if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
1817                         return -8;
1818
1819         if (DO_BIC(BIC_CoreTmp)) {
1820                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1821                         return -9;
1822                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1823         }
1824
1825         for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1826                 if (get_mp(cpu, mp, &c->counter[i]))
1827                         return -10;
1828         }
1829
1830         /* collect package counters only for 1st core in package */
1831         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1832                 goto done;
1833
1834         if (DO_BIC(BIC_Totl_c0)) {
1835                 if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
1836                         return -10;
1837         }
1838         if (DO_BIC(BIC_Any_c0)) {
1839                 if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
1840                         return -11;
1841         }
1842         if (DO_BIC(BIC_GFX_c0)) {
1843                 if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
1844                         return -12;
1845         }
1846         if (DO_BIC(BIC_CPUGFX)) {
1847                 if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
1848                         return -13;
1849         }
1850         if (DO_BIC(BIC_Pkgpc3))
1851                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
1852                         return -9;
1853         if (DO_BIC(BIC_Pkgpc6)) {
1854                 if (do_slm_cstates) {
1855                         if (get_msr(cpu, MSR_ATOM_PKG_C6_RESIDENCY, &p->pc6))
1856                                 return -10;
1857                 } else {
1858                         if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
1859                                 return -10;
1860                 }
1861         }
1862
1863         if (DO_BIC(BIC_Pkgpc2))
1864                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
1865                         return -11;
1866         if (DO_BIC(BIC_Pkgpc7))
1867                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1868                         return -12;
1869         if (DO_BIC(BIC_Pkgpc8))
1870                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1871                         return -13;
1872         if (DO_BIC(BIC_Pkgpc9))
1873                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1874                         return -13;
1875         if (DO_BIC(BIC_Pkgpc10))
1876                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1877                         return -13;
1878
1879         if (DO_BIC(BIC_CPU_LPI))
1880                 p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
1881         if (DO_BIC(BIC_SYS_LPI))
1882                 p->sys_lpi = cpuidle_cur_sys_lpi_us;
1883
1884         if (do_rapl & RAPL_PKG) {
1885                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1886                         return -13;
1887                 p->energy_pkg = msr & 0xFFFFFFFF;
1888         }
1889         if (do_rapl & RAPL_CORES_ENERGY_STATUS) {
1890                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1891                         return -14;
1892                 p->energy_cores = msr & 0xFFFFFFFF;
1893         }
1894         if (do_rapl & RAPL_DRAM) {
1895                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1896                         return -15;
1897                 p->energy_dram = msr & 0xFFFFFFFF;
1898         }
1899         if (do_rapl & RAPL_GFX) {
1900                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1901                         return -16;
1902                 p->energy_gfx = msr & 0xFFFFFFFF;
1903         }
1904         if (do_rapl & RAPL_PKG_PERF_STATUS) {
1905                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1906                         return -16;
1907                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1908         }
1909         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1910                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1911                         return -16;
1912                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1913         }
1914         if (DO_BIC(BIC_PkgTmp)) {
1915                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1916                         return -17;
1917                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1918         }
1919
1920         if (DO_BIC(BIC_GFX_rc6))
1921                 p->gfx_rc6_ms = gfx_cur_rc6_ms;
1922
1923         if (DO_BIC(BIC_GFXMHz))
1924                 p->gfx_mhz = gfx_cur_mhz;
1925
1926         for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1927                 if (get_mp(cpu, mp, &p->counter[i]))
1928                         return -10;
1929         }
1930 done:
1931         gettimeofday(&t->tv_end, (struct timezone *)NULL);
1932
1933         return 0;
1934 }
1935
1936 /*
1937  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1938  * If you change the values, note they are used both in comparisons
1939  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1940  */
1941
1942 #define PCLUKN 0 /* Unknown */
1943 #define PCLRSV 1 /* Reserved */
1944 #define PCL__0 2 /* PC0 */
1945 #define PCL__1 3 /* PC1 */
1946 #define PCL__2 4 /* PC2 */
1947 #define PCL__3 5 /* PC3 */
1948 #define PCL__4 6 /* PC4 */
1949 #define PCL__6 7 /* PC6 */
1950 #define PCL_6N 8 /* PC6 No Retention */
1951 #define PCL_6R 9 /* PC6 Retention */
1952 #define PCL__7 10 /* PC7 */
1953 #define PCL_7S 11 /* PC7 Shrink */
1954 #define PCL__8 12 /* PC8 */
1955 #define PCL__9 13 /* PC9 */
1956 #define PCLUNL 14 /* Unlimited */
1957
1958 int pkg_cstate_limit = PCLUKN;
1959 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1960         "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "unlimited"};
1961
1962 int nhm_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1963 int snb_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1964 int hsw_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1965 int slv_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7};
1966 int amt_pkg_cstate_limits[16] = {PCLUNL, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1967 int phi_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1968 int bxt_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1969 int skx_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1970
1971
1972 static void
1973 calculate_tsc_tweak()
1974 {
1975         tsc_tweak = base_hz / tsc_hz;
1976 }
1977
1978 static void
1979 dump_nhm_platform_info(void)
1980 {
1981         unsigned long long msr;
1982         unsigned int ratio;
1983
1984         get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
1985
1986         fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
1987
1988         ratio = (msr >> 40) & 0xFF;
1989         fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n",
1990                 ratio, bclk, ratio * bclk);
1991
1992         ratio = (msr >> 8) & 0xFF;
1993         fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n",
1994                 ratio, bclk, ratio * bclk);
1995
1996         get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
1997         fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1998                 base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
1999
2000         return;
2001 }
2002
2003 static void
2004 dump_hsw_turbo_ratio_limits(void)
2005 {
2006         unsigned long long msr;
2007         unsigned int ratio;
2008
2009         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
2010
2011         fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
2012
2013         ratio = (msr >> 8) & 0xFF;
2014         if (ratio)
2015                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n",
2016                         ratio, bclk, ratio * bclk);
2017
2018         ratio = (msr >> 0) & 0xFF;
2019         if (ratio)
2020                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n",
2021                         ratio, bclk, ratio * bclk);
2022         return;
2023 }
2024
2025 static void
2026 dump_ivt_turbo_ratio_limits(void)
2027 {
2028         unsigned long long msr;
2029         unsigned int ratio;
2030
2031         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
2032
2033         fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
2034
2035         ratio = (msr >> 56) & 0xFF;
2036         if (ratio)
2037                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n",
2038                         ratio, bclk, ratio * bclk);
2039
2040         ratio = (msr >> 48) & 0xFF;
2041         if (ratio)
2042                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n",
2043                         ratio, bclk, ratio * bclk);
2044
2045         ratio = (msr >> 40) & 0xFF;
2046         if (ratio)
2047                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n",
2048                         ratio, bclk, ratio * bclk);
2049
2050         ratio = (msr >> 32) & 0xFF;
2051         if (ratio)
2052                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n",
2053                         ratio, bclk, ratio * bclk);
2054
2055         ratio = (msr >> 24) & 0xFF;
2056         if (ratio)
2057                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n",
2058                         ratio, bclk, ratio * bclk);
2059
2060         ratio = (msr >> 16) & 0xFF;
2061         if (ratio)
2062                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n",
2063                         ratio, bclk, ratio * bclk);
2064
2065         ratio = (msr >> 8) & 0xFF;
2066         if (ratio)
2067                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n",
2068                         ratio, bclk, ratio * bclk);
2069
2070         ratio = (msr >> 0) & 0xFF;
2071         if (ratio)
2072                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n",
2073                         ratio, bclk, ratio * bclk);
2074         return;
2075 }
2076 int has_turbo_ratio_group_limits(int family, int model)
2077 {
2078
2079         if (!genuine_intel)
2080                 return 0;
2081
2082         switch (model) {
2083         case INTEL_FAM6_ATOM_GOLDMONT:
2084         case INTEL_FAM6_SKYLAKE_X:
2085         case INTEL_FAM6_ATOM_DENVERTON:
2086                 return 1;
2087         }
2088         return 0;
2089 }
2090
2091 static void
2092 dump_turbo_ratio_limits(int family, int model)
2093 {
2094         unsigned long long msr, core_counts;
2095         unsigned int ratio, group_size;
2096
2097         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
2098         fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
2099
2100         if (has_turbo_ratio_group_limits(family, model)) {
2101                 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
2102                 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, core_counts);
2103         } else {
2104                 core_counts = 0x0807060504030201;
2105         }
2106
2107         ratio = (msr >> 56) & 0xFF;
2108         group_size = (core_counts >> 56) & 0xFF;
2109         if (ratio)
2110                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2111                         ratio, bclk, ratio * bclk, group_size);
2112
2113         ratio = (msr >> 48) & 0xFF;
2114         group_size = (core_counts >> 48) & 0xFF;
2115         if (ratio)
2116                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2117                         ratio, bclk, ratio * bclk, group_size);
2118
2119         ratio = (msr >> 40) & 0xFF;
2120         group_size = (core_counts >> 40) & 0xFF;
2121         if (ratio)
2122                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2123                         ratio, bclk, ratio * bclk, group_size);
2124
2125         ratio = (msr >> 32) & 0xFF;
2126         group_size = (core_counts >> 32) & 0xFF;
2127         if (ratio)
2128                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2129                         ratio, bclk, ratio * bclk, group_size);
2130
2131         ratio = (msr >> 24) & 0xFF;
2132         group_size = (core_counts >> 24) & 0xFF;
2133         if (ratio)
2134                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2135                         ratio, bclk, ratio * bclk, group_size);
2136
2137         ratio = (msr >> 16) & 0xFF;
2138         group_size = (core_counts >> 16) & 0xFF;
2139         if (ratio)
2140                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2141                         ratio, bclk, ratio * bclk, group_size);
2142
2143         ratio = (msr >> 8) & 0xFF;
2144         group_size = (core_counts >> 8) & 0xFF;
2145         if (ratio)
2146                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2147                         ratio, bclk, ratio * bclk, group_size);
2148
2149         ratio = (msr >> 0) & 0xFF;
2150         group_size = (core_counts >> 0) & 0xFF;
2151         if (ratio)
2152                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2153                         ratio, bclk, ratio * bclk, group_size);
2154         return;
2155 }
2156
2157 static void
2158 dump_atom_turbo_ratio_limits(void)
2159 {
2160         unsigned long long msr;
2161         unsigned int ratio;
2162
2163         get_msr(base_cpu, MSR_ATOM_CORE_RATIOS, &msr);
2164         fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
2165
2166         ratio = (msr >> 0) & 0x3F;
2167         if (ratio)
2168                 fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n",
2169                         ratio, bclk, ratio * bclk);
2170
2171         ratio = (msr >> 8) & 0x3F;
2172         if (ratio)
2173                 fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n",
2174                         ratio, bclk, ratio * bclk);
2175
2176         ratio = (msr >> 16) & 0x3F;
2177         if (ratio)
2178                 fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n",
2179                         ratio, bclk, ratio * bclk);
2180
2181         get_msr(base_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
2182         fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
2183
2184         ratio = (msr >> 24) & 0x3F;
2185         if (ratio)
2186                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n",
2187                         ratio, bclk, ratio * bclk);
2188
2189         ratio = (msr >> 16) & 0x3F;
2190         if (ratio)
2191                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n",
2192                         ratio, bclk, ratio * bclk);
2193
2194         ratio = (msr >> 8) & 0x3F;
2195         if (ratio)
2196                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n",
2197                         ratio, bclk, ratio * bclk);
2198
2199         ratio = (msr >> 0) & 0x3F;
2200         if (ratio)
2201                 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n",
2202                         ratio, bclk, ratio * bclk);
2203 }
2204
2205 static void
2206 dump_knl_turbo_ratio_limits(void)
2207 {
2208         const unsigned int buckets_no = 7;
2209
2210         unsigned long long msr;
2211         int delta_cores, delta_ratio;
2212         int i, b_nr;
2213         unsigned int cores[buckets_no];
2214         unsigned int ratio[buckets_no];
2215
2216         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
2217
2218         fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n",
2219                 base_cpu, msr);
2220
2221         /**
2222          * Turbo encoding in KNL is as follows:
2223          * [0] -- Reserved
2224          * [7:1] -- Base value of number of active cores of bucket 1.
2225          * [15:8] -- Base value of freq ratio of bucket 1.
2226          * [20:16] -- +ve delta of number of active cores of bucket 2.
2227          * i.e. active cores of bucket 2 =
2228          * active cores of bucket 1 + delta
2229          * [23:21] -- Negative delta of freq ratio of bucket 2.
2230          * i.e. freq ratio of bucket 2 =
2231          * freq ratio of bucket 1 - delta
2232          * [28:24]-- +ve delta of number of active cores of bucket 3.
2233          * [31:29]-- -ve delta of freq ratio of bucket 3.
2234          * [36:32]-- +ve delta of number of active cores of bucket 4.
2235          * [39:37]-- -ve delta of freq ratio of bucket 4.
2236          * [44:40]-- +ve delta of number of active cores of bucket 5.
2237          * [47:45]-- -ve delta of freq ratio of bucket 5.
2238          * [52:48]-- +ve delta of number of active cores of bucket 6.
2239          * [55:53]-- -ve delta of freq ratio of bucket 6.
2240          * [60:56]-- +ve delta of number of active cores of bucket 7.
2241          * [63:61]-- -ve delta of freq ratio of bucket 7.
2242          */
2243
2244         b_nr = 0;
2245         cores[b_nr] = (msr & 0xFF) >> 1;
2246         ratio[b_nr] = (msr >> 8) & 0xFF;
2247
2248         for (i = 16; i < 64; i += 8) {
2249                 delta_cores = (msr >> i) & 0x1F;
2250                 delta_ratio = (msr >> (i + 5)) & 0x7;
2251
2252                 cores[b_nr + 1] = cores[b_nr] + delta_cores;
2253                 ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
2254                 b_nr++;
2255         }
2256
2257         for (i = buckets_no - 1; i >= 0; i--)
2258                 if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
2259                         fprintf(outf,
2260                                 "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2261                                 ratio[i], bclk, ratio[i] * bclk, cores[i]);
2262 }
2263
2264 static void
2265 dump_nhm_cst_cfg(void)
2266 {
2267         unsigned long long msr;
2268
2269         get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
2270
2271         fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", base_cpu, msr);
2272
2273         fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
2274                 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
2275                 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
2276                 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
2277                 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
2278                 (msr & (1 << 15)) ? "" : "UN",
2279                 (unsigned int)msr & 0xF,
2280                 pkg_cstate_limit_strings[pkg_cstate_limit]);
2281
2282 #define AUTOMATIC_CSTATE_CONVERSION             (1UL << 16)
2283         if (has_automatic_cstate_conversion) {
2284                 fprintf(outf, ", automatic c-state conversion=%s",
2285                         (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
2286         }
2287
2288         fprintf(outf, ")\n");
2289
2290         return;
2291 }
2292
2293 static void
2294 dump_config_tdp(void)
2295 {
2296         unsigned long long msr;
2297
2298         get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
2299         fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
2300         fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
2301
2302         get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
2303         fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
2304         if (msr) {
2305                 fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
2306                 fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
2307                 fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
2308                 fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
2309         }
2310         fprintf(outf, ")\n");
2311
2312         get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
2313         fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
2314         if (msr) {
2315                 fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
2316                 fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
2317                 fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
2318                 fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
2319         }
2320         fprintf(outf, ")\n");
2321
2322         get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
2323         fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
2324         if ((msr) & 0x3)
2325                 fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
2326         fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
2327         fprintf(outf, ")\n");
2328
2329         get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
2330         fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
2331         fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
2332         fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
2333         fprintf(outf, ")\n");
2334 }
2335
2336 unsigned int irtl_time_units[] = {1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
2337
2338 void print_irtl(void)
2339 {
2340         unsigned long long msr;
2341
2342         get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
2343         fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
2344         fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2345                 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2346
2347         get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
2348         fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
2349         fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2350                 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2351
2352         get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
2353         fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
2354         fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2355                 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2356
2357         if (!do_irtl_hsw)
2358                 return;
2359
2360         get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
2361         fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
2362         fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2363                 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2364
2365         get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
2366         fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
2367         fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2368                 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2369
2370         get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
2371         fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
2372         fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2373                 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2374
2375 }
2376 void free_fd_percpu(void)
2377 {
2378         int i;
2379
2380         for (i = 0; i < topo.max_cpu_num + 1; ++i) {
2381                 if (fd_percpu[i] != 0)
2382                         close(fd_percpu[i]);
2383         }
2384
2385         free(fd_percpu);
2386 }
2387
2388 void free_all_buffers(void)
2389 {
2390         int i;
2391
2392         CPU_FREE(cpu_present_set);
2393         cpu_present_set = NULL;
2394         cpu_present_setsize = 0;
2395
2396         CPU_FREE(cpu_affinity_set);
2397         cpu_affinity_set = NULL;
2398         cpu_affinity_setsize = 0;
2399
2400         free(thread_even);
2401         free(core_even);
2402         free(package_even);
2403
2404         thread_even = NULL;
2405         core_even = NULL;
2406         package_even = NULL;
2407
2408         free(thread_odd);
2409         free(core_odd);
2410         free(package_odd);
2411
2412         thread_odd = NULL;
2413         core_odd = NULL;
2414         package_odd = NULL;
2415
2416         free(output_buffer);
2417         output_buffer = NULL;
2418         outp = NULL;
2419
2420         free_fd_percpu();
2421
2422         free(irq_column_2_cpu);
2423         free(irqs_per_cpu);
2424
2425         for (i = 0; i <= topo.max_cpu_num; ++i) {
2426                 if (cpus[i].put_ids)
2427                         CPU_FREE(cpus[i].put_ids);
2428         }
2429         free(cpus);
2430 }
2431
2432
2433 /*
2434  * Parse a file containing a single int.
2435  */
2436 int parse_int_file(const char *fmt, ...)
2437 {
2438         va_list args;
2439         char path[PATH_MAX];
2440         FILE *filep;
2441         int value;
2442
2443         va_start(args, fmt);
2444         vsnprintf(path, sizeof(path), fmt, args);
2445         va_end(args);
2446         filep = fopen_or_die(path, "r");
2447         if (fscanf(filep, "%d", &value) != 1)
2448                 err(1, "%s: failed to parse number from file", path);
2449         fclose(filep);
2450         return value;
2451 }
2452
2453 /*
2454  * cpu_is_first_core_in_package(cpu)
2455  * return 1 if given CPU is 1st core in package
2456  */
2457 int cpu_is_first_core_in_package(int cpu)
2458 {
2459         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
2460 }
2461
2462 int get_physical_package_id(int cpu)
2463 {
2464         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
2465 }
2466
2467 int get_core_id(int cpu)
2468 {
2469         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
2470 }
2471
2472 void set_node_data(void)
2473 {
2474         int pkg, node, lnode, cpu, cpux;
2475         int cpu_count;
2476
2477         /* initialize logical_node_id */
2478         for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu)
2479                 cpus[cpu].logical_node_id = -1;
2480
2481         cpu_count = 0;
2482         for (pkg = 0; pkg < topo.num_packages; pkg++) {
2483                 lnode = 0;
2484                 for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
2485                         if (cpus[cpu].physical_package_id != pkg)
2486                                 continue;
2487                         /* find a cpu with an unset logical_node_id */
2488                         if (cpus[cpu].logical_node_id != -1)
2489                                 continue;
2490                         cpus[cpu].logical_node_id = lnode;
2491                         node = cpus[cpu].physical_node_id;
2492                         cpu_count++;
2493                         /*
2494                          * find all matching cpus on this pkg and set
2495                          * the logical_node_id
2496                          */
2497                         for (cpux = cpu; cpux <= topo.max_cpu_num; cpux++) {
2498                                 if ((cpus[cpux].physical_package_id == pkg) &&
2499                                    (cpus[cpux].physical_node_id == node)) {
2500                                         cpus[cpux].logical_node_id = lnode;
2501                                         cpu_count++;
2502                                 }
2503                         }
2504                         lnode++;
2505                         if (lnode > topo.nodes_per_pkg)
2506                                 topo.nodes_per_pkg = lnode;
2507                 }
2508                 if (cpu_count >= topo.max_cpu_num)
2509                         break;
2510         }
2511 }
2512
2513 int get_physical_node_id(struct cpu_topology *thiscpu)
2514 {
2515         char path[80];
2516         FILE *filep;
2517         int i;
2518         int cpu = thiscpu->logical_cpu_id;
2519
2520         for (i = 0; i <= topo.max_cpu_num; i++) {
2521                 sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist",
2522                         cpu, i);
2523                 filep = fopen(path, "r");
2524                 if (!filep)
2525                         continue;
2526                 fclose(filep);
2527                 return i;
2528         }
2529         return -1;
2530 }
2531
2532 int get_thread_siblings(struct cpu_topology *thiscpu)
2533 {
2534         char path[80], character;
2535         FILE *filep;
2536         unsigned long map;
2537         int so, shift, sib_core;
2538         int cpu = thiscpu->logical_cpu_id;
2539         int offset = topo.max_cpu_num + 1;
2540         size_t size;
2541         int thread_id = 0;
2542
2543         thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
2544         if (thiscpu->thread_id < 0)
2545                 thiscpu->thread_id = thread_id++;
2546         if (!thiscpu->put_ids)
2547                 return -1;
2548
2549         size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2550         CPU_ZERO_S(size, thiscpu->put_ids);
2551
2552         sprintf(path,
2553                 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", cpu);
2554         filep = fopen_or_die(path, "r");
2555         do {
2556                 offset -= BITMASK_SIZE;
2557                 fscanf(filep, "%lx%c", &map, &character);
2558                 for (shift = 0; shift < BITMASK_SIZE; shift++) {
2559                         if ((map >> shift) & 0x1) {
2560                                 so = shift + offset;
2561                                 sib_core = get_core_id(so);
2562                                 if (sib_core == thiscpu->physical_core_id) {
2563                                         CPU_SET_S(so, size, thiscpu->put_ids);
2564                                         if ((so != cpu) &&
2565                                             (cpus[so].thread_id < 0))
2566                                                 cpus[so].thread_id =
2567                                                                     thread_id++;
2568                                 }
2569                         }
2570                 }
2571         } while (!strncmp(&character, ",", 1));
2572         fclose(filep);
2573
2574         return CPU_COUNT_S(size, thiscpu->put_ids);
2575 }
2576
2577 /*
2578  * run func(thread, core, package) in topology order
2579  * skip non-present cpus
2580  */
2581
2582 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
2583         struct pkg_data *, struct thread_data *, struct core_data *,
2584         struct pkg_data *), struct thread_data *thread_base,
2585         struct core_data *core_base, struct pkg_data *pkg_base,
2586         struct thread_data *thread_base2, struct core_data *core_base2,
2587         struct pkg_data *pkg_base2)
2588 {
2589         int retval, pkg_no, node_no, core_no, thread_no;
2590
2591         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
2592                 for (node_no = 0; node_no < topo.nodes_per_pkg; ++node_no) {
2593                         for (core_no = 0; core_no < topo.cores_per_node;
2594                              ++core_no) {
2595                                 for (thread_no = 0; thread_no <
2596                                         topo.threads_per_core; ++thread_no) {
2597                                         struct thread_data *t, *t2;
2598                                         struct core_data *c, *c2;
2599                                         struct pkg_data *p, *p2;
2600
2601                                         t = GET_THREAD(thread_base, thread_no,
2602                                                        core_no, node_no,
2603                                                        pkg_no);
2604
2605                                         if (cpu_is_not_present(t->cpu_id))
2606                                                 continue;
2607
2608                                         t2 = GET_THREAD(thread_base2, thread_no,
2609                                                         core_no, node_no,
2610                                                         pkg_no);
2611
2612                                         c = GET_CORE(core_base, core_no,
2613                                                      node_no, pkg_no);
2614                                         c2 = GET_CORE(core_base2, core_no,
2615                                                       node_no,
2616                                                       pkg_no);
2617
2618                                         p = GET_PKG(pkg_base, pkg_no);
2619                                         p2 = GET_PKG(pkg_base2, pkg_no);
2620
2621                                         retval = func(t, c, p, t2, c2, p2);
2622                                         if (retval)
2623                                                 return retval;
2624                                 }
2625                         }
2626                 }
2627         }
2628         return 0;
2629 }
2630
2631 /*
2632  * run func(cpu) on every cpu in /proc/stat
2633  * return max_cpu number
2634  */
2635 int for_all_proc_cpus(int (func)(int))
2636 {
2637         FILE *fp;
2638         int cpu_num;
2639         int retval;
2640
2641         fp = fopen_or_die(proc_stat, "r");
2642
2643         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
2644         if (retval != 0)
2645                 err(1, "%s: failed to parse format", proc_stat);
2646
2647         while (1) {
2648                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
2649                 if (retval != 1)
2650                         break;
2651
2652                 retval = func(cpu_num);
2653                 if (retval) {
2654                         fclose(fp);
2655                         return(retval);
2656                 }
2657         }
2658         fclose(fp);
2659         return 0;
2660 }
2661
2662 void re_initialize(void)
2663 {
2664         free_all_buffers();
2665         setup_all_buffers();
2666         printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
2667 }
2668
2669 void set_max_cpu_num(void)
2670 {
2671         FILE *filep;
2672         unsigned long dummy;
2673
2674         topo.max_cpu_num = 0;
2675         filep = fopen_or_die(
2676                         "/sys/devices/system/cpu/cpu0/topology/thread_siblings",
2677                         "r");
2678         while (fscanf(filep, "%lx,", &dummy) == 1)
2679                 topo.max_cpu_num += BITMASK_SIZE;
2680         fclose(filep);
2681         topo.max_cpu_num--; /* 0 based */
2682 }
2683
2684 /*
2685  * count_cpus()
2686  * remember the last one seen, it will be the max
2687  */
2688 int count_cpus(int cpu)
2689 {
2690         topo.num_cpus++;
2691         return 0;
2692 }
2693 int mark_cpu_present(int cpu)
2694 {
2695         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
2696         return 0;
2697 }
2698
2699 int init_thread_id(int cpu)
2700 {
2701         cpus[cpu].thread_id = -1;
2702         return 0;
2703 }
2704
2705 /*
2706  * snapshot_proc_interrupts()
2707  *
2708  * read and record summary of /proc/interrupts
2709  *
2710  * return 1 if config change requires a restart, else return 0
2711  */
2712 int snapshot_proc_interrupts(void)
2713 {
2714         static FILE *fp;
2715         int column, retval;
2716
2717         if (fp == NULL)
2718                 fp = fopen_or_die("/proc/interrupts", "r");
2719         else
2720                 rewind(fp);
2721
2722         /* read 1st line of /proc/interrupts to get cpu* name for each column */
2723         for (column = 0; column < topo.num_cpus; ++column) {
2724                 int cpu_number;
2725
2726                 retval = fscanf(fp, " CPU%d", &cpu_number);
2727                 if (retval != 1)
2728                         break;
2729
2730                 if (cpu_number > topo.max_cpu_num) {
2731                         warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
2732                         return 1;
2733                 }
2734
2735                 irq_column_2_cpu[column] = cpu_number;
2736                 irqs_per_cpu[cpu_number] = 0;
2737         }
2738
2739         /* read /proc/interrupt count lines and sum up irqs per cpu */
2740         while (1) {
2741                 int column;
2742                 char buf[64];
2743
2744                 retval = fscanf(fp, " %s:", buf);       /* flush irq# "N:" */
2745                 if (retval != 1)
2746                         break;
2747
2748                 /* read the count per cpu */
2749                 for (column = 0; column < topo.num_cpus; ++column) {
2750
2751                         int cpu_number, irq_count;
2752
2753                         retval = fscanf(fp, " %d", &irq_count);
2754                         if (retval != 1)
2755                                 break;
2756
2757                         cpu_number = irq_column_2_cpu[column];
2758                         irqs_per_cpu[cpu_number] += irq_count;
2759
2760                 }
2761
2762                 while (getc(fp) != '\n')
2763                         ;       /* flush interrupt description */
2764
2765         }
2766         return 0;
2767 }
2768 /*
2769  * snapshot_gfx_rc6_ms()
2770  *
2771  * record snapshot of
2772  * /sys/class/drm/card0/power/rc6_residency_ms
2773  *
2774  * return 1 if config change requires a restart, else return 0
2775  */
2776 int snapshot_gfx_rc6_ms(void)
2777 {
2778         FILE *fp;
2779         int retval;
2780
2781         fp = fopen_or_die("/sys/class/drm/card0/power/rc6_residency_ms", "r");
2782
2783         retval = fscanf(fp, "%lld", &gfx_cur_rc6_ms);
2784         if (retval != 1)
2785                 err(1, "GFX rc6");
2786
2787         fclose(fp);
2788
2789         return 0;
2790 }
2791 /*
2792  * snapshot_gfx_mhz()
2793  *
2794  * record snapshot of
2795  * /sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz
2796  *
2797  * return 1 if config change requires a restart, else return 0
2798  */
2799 int snapshot_gfx_mhz(void)
2800 {
2801         static FILE *fp;
2802         int retval;
2803
2804         if (fp == NULL)
2805                 fp = fopen_or_die("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", "r");
2806         else {
2807                 rewind(fp);
2808                 fflush(fp);
2809         }
2810
2811         retval = fscanf(fp, "%d", &gfx_cur_mhz);
2812         if (retval != 1)
2813                 err(1, "GFX MHz");
2814
2815         return 0;
2816 }
2817
2818 /*
2819  * snapshot_cpu_lpi()
2820  *
2821  * record snapshot of
2822  * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
2823  *
2824  * return 1 if config change requires a restart, else return 0
2825  */
2826 int snapshot_cpu_lpi_us(void)
2827 {
2828         FILE *fp;
2829         int retval;
2830
2831         fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
2832
2833         retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
2834         if (retval != 1)
2835                 err(1, "CPU LPI");
2836
2837         fclose(fp);
2838
2839         return 0;
2840 }
2841 /*
2842  * snapshot_sys_lpi()
2843  *
2844  * record snapshot of
2845  * /sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us
2846  *
2847  * return 1 if config change requires a restart, else return 0
2848  */
2849 int snapshot_sys_lpi_us(void)
2850 {
2851         FILE *fp;
2852         int retval;
2853
2854         fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us", "r");
2855
2856         retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
2857         if (retval != 1)
2858                 err(1, "SYS LPI");
2859
2860         fclose(fp);
2861
2862         return 0;
2863 }
2864 /*
2865  * snapshot /proc and /sys files
2866  *
2867  * return 1 if configuration restart needed, else return 0
2868  */
2869 int snapshot_proc_sysfs_files(void)
2870 {
2871         if (DO_BIC(BIC_IRQ))
2872                 if (snapshot_proc_interrupts())
2873                         return 1;
2874
2875         if (DO_BIC(BIC_GFX_rc6))
2876                 snapshot_gfx_rc6_ms();
2877
2878         if (DO_BIC(BIC_GFXMHz))
2879                 snapshot_gfx_mhz();
2880
2881         if (DO_BIC(BIC_CPU_LPI))
2882                 snapshot_cpu_lpi_us();
2883
2884         if (DO_BIC(BIC_SYS_LPI))
2885                 snapshot_sys_lpi_us();
2886
2887         return 0;
2888 }
2889
2890 int exit_requested;
2891
2892 static void signal_handler (int signal)
2893 {
2894         switch (signal) {
2895         case SIGINT:
2896                 exit_requested = 1;
2897                 if (debug)
2898                         fprintf(stderr, " SIGINT\n");
2899                 break;
2900         case SIGUSR1:
2901                 if (debug > 1)
2902                         fprintf(stderr, "SIGUSR1\n");
2903                 break;
2904         }
2905         /* make sure this manually-invoked interval is at least 1ms long */
2906         nanosleep(&one_msec, NULL);
2907 }
2908
2909 void setup_signal_handler(void)
2910 {
2911         struct sigaction sa;
2912
2913         memset(&sa, 0, sizeof(sa));
2914
2915         sa.sa_handler = &signal_handler;
2916
2917         if (sigaction(SIGINT, &sa, NULL) < 0)
2918                 err(1, "sigaction SIGINT");
2919         if (sigaction(SIGUSR1, &sa, NULL) < 0)
2920                 err(1, "sigaction SIGUSR1");
2921 }
2922
2923 void do_sleep(void)
2924 {
2925         struct timeval select_timeout;
2926         fd_set readfds;
2927         int retval;
2928
2929         FD_ZERO(&readfds);
2930         FD_SET(0, &readfds);
2931
2932         if (!isatty(fileno(stdin))) {
2933                 nanosleep(&interval_ts, NULL);
2934                 return;
2935         }
2936
2937         select_timeout = interval_tv;
2938         retval = select(1, &readfds, NULL, NULL, &select_timeout);
2939
2940         if (retval == 1) {
2941                 switch (getc(stdin)) {
2942                 case 'q':
2943                         exit_requested = 1;
2944                         break;
2945                 }
2946                 /* make sure this manually-invoked interval is at least 1ms long */
2947                 nanosleep(&one_msec, NULL);
2948         }
2949 }
2950
2951
2952 void turbostat_loop()
2953 {
2954         int retval;
2955         int restarted = 0;
2956         int done_iters = 0;
2957
2958         setup_signal_handler();
2959
2960 restart:
2961         restarted++;
2962
2963         snapshot_proc_sysfs_files();
2964         retval = for_all_cpus(get_counters, EVEN_COUNTERS);
2965         first_counter_read = 0;
2966         if (retval < -1) {
2967                 exit(retval);
2968         } else if (retval == -1) {
2969                 if (restarted > 1) {
2970                         exit(retval);
2971                 }
2972                 re_initialize();
2973                 goto restart;
2974         }
2975         restarted = 0;
2976         done_iters = 0;
2977         gettimeofday(&tv_even, (struct timezone *)NULL);
2978
2979         while (1) {
2980                 if (for_all_proc_cpus(cpu_is_not_present)) {
2981                         re_initialize();
2982                         goto restart;
2983                 }
2984                 do_sleep();
2985                 if (snapshot_proc_sysfs_files())
2986                         goto restart;
2987                 retval = for_all_cpus(get_counters, ODD_COUNTERS);
2988                 if (retval < -1) {
2989                         exit(retval);
2990                 } else if (retval == -1) {
2991                         re_initialize();
2992                         goto restart;
2993                 }
2994                 gettimeofday(&tv_odd, (struct timezone *)NULL);
2995                 timersub(&tv_odd, &tv_even, &tv_delta);
2996                 if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
2997                         re_initialize();
2998                         goto restart;
2999                 }
3000                 compute_average(EVEN_COUNTERS);
3001                 format_all_counters(EVEN_COUNTERS);
3002                 flush_output_stdout();
3003                 if (exit_requested)
3004                         break;
3005                 if (num_iterations && ++done_iters >= num_iterations)
3006                         break;
3007                 do_sleep();
3008                 if (snapshot_proc_sysfs_files())
3009                         goto restart;
3010                 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
3011                 if (retval < -1) {
3012                         exit(retval);
3013                 } else if (retval == -1) {
3014                         re_initialize();
3015                         goto restart;
3016                 }
3017                 gettimeofday(&tv_even, (struct timezone *)NULL);
3018                 timersub(&tv_even, &tv_odd, &tv_delta);
3019                 if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
3020                         re_initialize();
3021                         goto restart;
3022                 }
3023                 compute_average(ODD_COUNTERS);
3024                 format_all_counters(ODD_COUNTERS);
3025                 flush_output_stdout();
3026                 if (exit_requested)
3027                         break;
3028                 if (num_iterations && ++done_iters >= num_iterations)
3029                         break;
3030         }
3031 }
3032
3033 void check_dev_msr()
3034 {
3035         struct stat sb;
3036         char pathname[32];
3037
3038         sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
3039         if (stat(pathname, &sb))
3040                 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
3041                         err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
3042 }
3043
3044 void check_permissions()
3045 {
3046         struct __user_cap_header_struct cap_header_data;
3047         cap_user_header_t cap_header = &cap_header_data;
3048         struct __user_cap_data_struct cap_data_data;
3049         cap_user_data_t cap_data = &cap_data_data;
3050         extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
3051         int do_exit = 0;
3052         char pathname[32];
3053
3054         /* check for CAP_SYS_RAWIO */
3055         cap_header->pid = getpid();
3056         cap_header->version = _LINUX_CAPABILITY_VERSION;
3057         if (capget(cap_header, cap_data) < 0)
3058                 err(-6, "capget(2) failed");
3059
3060         if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
3061                 do_exit++;
3062                 warnx("capget(CAP_SYS_RAWIO) failed,"
3063                         " try \"# setcap cap_sys_rawio=ep %s\"", progname);
3064         }
3065
3066         /* test file permissions */
3067         sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
3068         if (euidaccess(pathname, R_OK)) {
3069                 do_exit++;
3070                 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
3071         }
3072
3073         /* if all else fails, thell them to be root */
3074         if (do_exit)
3075                 if (getuid() != 0)
3076                         warnx("... or simply run as root");
3077
3078         if (do_exit)
3079                 exit(-6);
3080 }
3081
3082 /*
3083  * NHM adds support for additional MSRs:
3084  *
3085  * MSR_SMI_COUNT                   0x00000034
3086  *
3087  * MSR_PLATFORM_INFO               0x000000ce
3088  * MSR_PKG_CST_CONFIG_CONTROL     0x000000e2
3089  *
3090  * MSR_MISC_PWR_MGMT               0x000001aa
3091  *
3092  * MSR_PKG_C3_RESIDENCY            0x000003f8
3093  * MSR_PKG_C6_RESIDENCY            0x000003f9
3094  * MSR_CORE_C3_RESIDENCY           0x000003fc
3095  * MSR_CORE_C6_RESIDENCY           0x000003fd
3096  *
3097  * Side effect:
3098  * sets global pkg_cstate_limit to decode MSR_PKG_CST_CONFIG_CONTROL
3099  * sets has_misc_feature_control
3100  */
3101 int probe_nhm_msrs(unsigned int family, unsigned int model)
3102 {
3103         unsigned long long msr;
3104         unsigned int base_ratio;
3105         int *pkg_cstate_limits;
3106
3107         if (!genuine_intel)
3108                 return 0;
3109
3110         if (family != 6)
3111                 return 0;
3112
3113         bclk = discover_bclk(family, model);
3114
3115         switch (model) {
3116         case INTEL_FAM6_NEHALEM_EP:     /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
3117         case INTEL_FAM6_NEHALEM:        /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
3118         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
3119         case INTEL_FAM6_WESTMERE:       /* Westmere Client - Clarkdale, Arrandale */
3120         case INTEL_FAM6_WESTMERE_EP:    /* Westmere EP - Gulftown */
3121         case INTEL_FAM6_NEHALEM_EX:     /* Nehalem-EX Xeon - Beckton */
3122         case INTEL_FAM6_WESTMERE_EX:    /* Westmere-EX Xeon - Eagleton */
3123                 pkg_cstate_limits = nhm_pkg_cstate_limits;
3124                 break;
3125         case INTEL_FAM6_SANDYBRIDGE:    /* SNB */
3126         case INTEL_FAM6_SANDYBRIDGE_X:  /* SNB Xeon */
3127         case INTEL_FAM6_IVYBRIDGE:      /* IVB */
3128         case INTEL_FAM6_IVYBRIDGE_X:    /* IVB Xeon */
3129                 pkg_cstate_limits = snb_pkg_cstate_limits;
3130                 has_misc_feature_control = 1;
3131                 break;
3132         case INTEL_FAM6_HASWELL_CORE:   /* HSW */
3133         case INTEL_FAM6_HASWELL_X:      /* HSX */
3134         case INTEL_FAM6_HASWELL_ULT:    /* HSW */
3135         case INTEL_FAM6_HASWELL_GT3E:   /* HSW */
3136         case INTEL_FAM6_BROADWELL_CORE: /* BDW */
3137         case INTEL_FAM6_BROADWELL_GT3E: /* BDW */
3138         case INTEL_FAM6_BROADWELL_X:    /* BDX */
3139         case INTEL_FAM6_BROADWELL_XEON_D:       /* BDX-DE */
3140         case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
3141         case INTEL_FAM6_SKYLAKE_DESKTOP:        /* SKL */
3142         case INTEL_FAM6_KABYLAKE_MOBILE:        /* KBL */
3143         case INTEL_FAM6_KABYLAKE_DESKTOP:       /* KBL */
3144         case INTEL_FAM6_CANNONLAKE_MOBILE:      /* CNL */
3145                 pkg_cstate_limits = hsw_pkg_cstate_limits;
3146                 has_misc_feature_control = 1;
3147                 break;
3148         case INTEL_FAM6_SKYLAKE_X:      /* SKX */
3149                 pkg_cstate_limits = skx_pkg_cstate_limits;
3150                 has_misc_feature_control = 1;
3151                 break;
3152         case INTEL_FAM6_ATOM_SILVERMONT1:       /* BYT */
3153                 no_MSR_MISC_PWR_MGMT = 1;
3154         case INTEL_FAM6_ATOM_SILVERMONT2:       /* AVN */
3155                 pkg_cstate_limits = slv_pkg_cstate_limits;
3156                 break;
3157         case INTEL_FAM6_ATOM_AIRMONT:   /* AMT */
3158                 pkg_cstate_limits = amt_pkg_cstate_limits;
3159                 no_MSR_MISC_PWR_MGMT = 1;
3160                 break;
3161         case INTEL_FAM6_XEON_PHI_KNL:   /* PHI */
3162         case INTEL_FAM6_XEON_PHI_KNM:
3163                 pkg_cstate_limits = phi_pkg_cstate_limits;
3164                 break;
3165         case INTEL_FAM6_ATOM_GOLDMONT:  /* BXT */
3166         case INTEL_FAM6_ATOM_GEMINI_LAKE:
3167         case INTEL_FAM6_ATOM_DENVERTON: /* DNV */
3168                 pkg_cstate_limits = bxt_pkg_cstate_limits;
3169                 break;
3170         default:
3171                 return 0;
3172         }
3173         get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
3174         pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
3175
3176         get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
3177         base_ratio = (msr >> 8) & 0xFF;
3178
3179         base_hz = base_ratio * bclk * 1000000;
3180         has_base_hz = 1;
3181         return 1;