1103a8feaSLen Brown /* 2103a8feaSLen Brown * turbostat -- show CPU frequency and C-state residency 3103a8feaSLen Brown * on modern Intel turbo-capable processors. 4103a8feaSLen Brown * 5e23da037SLen Brown * Copyright (c) 2012 Intel Corporation. 6103a8feaSLen Brown * Len Brown <len.brown@intel.com> 7103a8feaSLen Brown * 8103a8feaSLen Brown * This program is free software; you can redistribute it and/or modify it 9103a8feaSLen Brown * under the terms and conditions of the GNU General Public License, 10103a8feaSLen Brown * version 2, as published by the Free Software Foundation. 11103a8feaSLen Brown * 12103a8feaSLen Brown * This program is distributed in the hope it will be useful, but WITHOUT 13103a8feaSLen Brown * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14103a8feaSLen Brown * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 15103a8feaSLen Brown * more details. 16103a8feaSLen Brown * 17103a8feaSLen Brown * You should have received a copy of the GNU General Public License along with 18103a8feaSLen Brown * this program; if not, write to the Free Software Foundation, Inc., 19103a8feaSLen Brown * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. 20103a8feaSLen Brown */ 21103a8feaSLen Brown 2288c3281fSLen Brown #define _GNU_SOURCE 23103a8feaSLen Brown #include <stdio.h> 24103a8feaSLen Brown #include <unistd.h> 25103a8feaSLen Brown #include <sys/types.h> 26103a8feaSLen Brown #include <sys/wait.h> 27103a8feaSLen Brown #include <sys/stat.h> 28103a8feaSLen Brown #include <sys/resource.h> 29103a8feaSLen Brown #include <fcntl.h> 30103a8feaSLen Brown #include <signal.h> 31103a8feaSLen Brown #include <sys/time.h> 32103a8feaSLen Brown #include <stdlib.h> 33103a8feaSLen Brown #include <dirent.h> 34103a8feaSLen Brown #include <string.h> 35103a8feaSLen Brown #include <ctype.h> 3688c3281fSLen Brown #include <sched.h> 37103a8feaSLen Brown 38103a8feaSLen Brown #define MSR_NEHALEM_PLATFORM_INFO 0xCE 39103a8feaSLen Brown #define MSR_NEHALEM_TURBO_RATIO_LIMIT 0x1AD 406574a5d5SLen Brown #define MSR_IVT_TURBO_RATIO_LIMIT 0x1AE 41103a8feaSLen Brown #define MSR_APERF 0xE8 42103a8feaSLen Brown #define MSR_MPERF 0xE7 43103a8feaSLen Brown #define MSR_PKG_C2_RESIDENCY 0x60D /* SNB only */ 44103a8feaSLen Brown #define MSR_PKG_C3_RESIDENCY 0x3F8 45103a8feaSLen Brown #define MSR_PKG_C6_RESIDENCY 0x3F9 46103a8feaSLen Brown #define MSR_PKG_C7_RESIDENCY 0x3FA /* SNB only */ 47103a8feaSLen Brown #define MSR_CORE_C3_RESIDENCY 0x3FC 48103a8feaSLen Brown #define MSR_CORE_C6_RESIDENCY 0x3FD 49103a8feaSLen Brown #define MSR_CORE_C7_RESIDENCY 0x3FE /* SNB only */ 50103a8feaSLen Brown 51103a8feaSLen Brown char *proc_stat = "/proc/stat"; 52103a8feaSLen Brown unsigned int interval_sec = 5; /* set with -i interval_sec */ 53103a8feaSLen Brown unsigned int verbose; /* set with -v */ 54e23da037SLen Brown unsigned int summary_only; /* set with -s */ 55103a8feaSLen Brown unsigned int skip_c0; 56103a8feaSLen Brown unsigned int skip_c1; 57103a8feaSLen Brown unsigned int do_nhm_cstates; 58103a8feaSLen Brown unsigned int do_snb_cstates; 59103a8feaSLen Brown unsigned int has_aperf; 60103a8feaSLen Brown unsigned int units = 1000000000; /* Ghz etc */ 61103a8feaSLen Brown unsigned int genuine_intel; 62103a8feaSLen Brown unsigned int has_invariant_tsc; 63103a8feaSLen Brown unsigned int do_nehalem_platform_info; 64103a8feaSLen Brown unsigned int do_nehalem_turbo_ratio_limit; 656574a5d5SLen Brown unsigned int do_ivt_turbo_ratio_limit; 66103a8feaSLen Brown unsigned int extra_msr_offset; 67103a8feaSLen Brown double bclk; 68103a8feaSLen Brown unsigned int show_pkg; 69103a8feaSLen Brown unsigned int show_core; 70103a8feaSLen Brown unsigned int show_cpu; 71c98d5d94SLen Brown unsigned int show_pkg_only; 72c98d5d94SLen Brown unsigned int show_core_only; 73c98d5d94SLen Brown char *output_buffer, *outp; 74103a8feaSLen Brown 75103a8feaSLen Brown int aperf_mperf_unstable; 76103a8feaSLen Brown int backwards_count; 77103a8feaSLen Brown char *progname; 78103a8feaSLen Brown 79c98d5d94SLen Brown cpu_set_t *cpu_present_set, *cpu_affinity_set; 80c98d5d94SLen Brown size_t cpu_present_setsize, cpu_affinity_setsize; 81c98d5d94SLen Brown 82c98d5d94SLen Brown struct thread_data { 83c98d5d94SLen Brown unsigned long long tsc; 84c98d5d94SLen Brown unsigned long long aperf; 85c98d5d94SLen Brown unsigned long long mperf; 86c98d5d94SLen Brown unsigned long long c1; /* derived */ 87c98d5d94SLen Brown unsigned long long extra_msr; 88c98d5d94SLen Brown unsigned int cpu_id; 89c98d5d94SLen Brown unsigned int flags; 90c98d5d94SLen Brown #define CPU_IS_FIRST_THREAD_IN_CORE 0x2 91c98d5d94SLen Brown #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4 92c98d5d94SLen Brown } *thread_even, *thread_odd; 93c98d5d94SLen Brown 94c98d5d94SLen Brown struct core_data { 95c98d5d94SLen Brown unsigned long long c3; 96c98d5d94SLen Brown unsigned long long c6; 97c98d5d94SLen Brown unsigned long long c7; 98c98d5d94SLen Brown unsigned int core_id; 99c98d5d94SLen Brown } *core_even, *core_odd; 100c98d5d94SLen Brown 101c98d5d94SLen Brown struct pkg_data { 102c98d5d94SLen Brown unsigned long long pc2; 103c98d5d94SLen Brown unsigned long long pc3; 104c98d5d94SLen Brown unsigned long long pc6; 105c98d5d94SLen Brown unsigned long long pc7; 106c98d5d94SLen Brown unsigned int package_id; 107c98d5d94SLen Brown } *package_even, *package_odd; 108c98d5d94SLen Brown 109c98d5d94SLen Brown #define ODD_COUNTERS thread_odd, core_odd, package_odd 110c98d5d94SLen Brown #define EVEN_COUNTERS thread_even, core_even, package_even 111c98d5d94SLen Brown 112c98d5d94SLen Brown #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \ 113c98d5d94SLen Brown (thread_base + (pkg_no) * topo.num_cores_per_pkg * \ 114c98d5d94SLen Brown topo.num_threads_per_core + \ 115c98d5d94SLen Brown (core_no) * topo.num_threads_per_core + (thread_no)) 116c98d5d94SLen Brown #define GET_CORE(core_base, core_no, pkg_no) \ 117c98d5d94SLen Brown (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no)) 118c98d5d94SLen Brown #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no) 119c98d5d94SLen Brown 120c98d5d94SLen Brown struct system_summary { 121c98d5d94SLen Brown struct thread_data threads; 122c98d5d94SLen Brown struct core_data cores; 123c98d5d94SLen Brown struct pkg_data packages; 124c98d5d94SLen Brown } sum, average; 125c98d5d94SLen Brown 126c98d5d94SLen Brown 127c98d5d94SLen Brown struct topo_params { 128c98d5d94SLen Brown int num_packages; 129103a8feaSLen Brown int num_cpus; 130c98d5d94SLen Brown int num_cores; 131c98d5d94SLen Brown int max_cpu_num; 132c98d5d94SLen Brown int num_cores_per_pkg; 133c98d5d94SLen Brown int num_threads_per_core; 134c98d5d94SLen Brown } topo; 135103a8feaSLen Brown 136c98d5d94SLen Brown struct timeval tv_even, tv_odd, tv_delta; 137103a8feaSLen Brown 138c98d5d94SLen Brown void setup_all_buffers(void); 139103a8feaSLen Brown 140c98d5d94SLen Brown int cpu_is_not_present(int cpu) 141d15cf7c1SLen Brown { 142c98d5d94SLen Brown return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set); 143c98d5d94SLen Brown } 144c98d5d94SLen Brown /* 145c98d5d94SLen Brown * run func(thread, core, package) in topology order 146c98d5d94SLen Brown * skip non-present cpus 147c98d5d94SLen Brown */ 148c98d5d94SLen Brown 149c98d5d94SLen Brown int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *), 150c98d5d94SLen Brown struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base) 151c98d5d94SLen Brown { 152c98d5d94SLen Brown int retval, pkg_no, core_no, thread_no; 153c98d5d94SLen Brown 154c98d5d94SLen Brown for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) { 155c98d5d94SLen Brown for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) { 156c98d5d94SLen Brown for (thread_no = 0; thread_no < 157c98d5d94SLen Brown topo.num_threads_per_core; ++thread_no) { 158c98d5d94SLen Brown struct thread_data *t; 159c98d5d94SLen Brown struct core_data *c; 160c98d5d94SLen Brown struct pkg_data *p; 161c98d5d94SLen Brown 162c98d5d94SLen Brown t = GET_THREAD(thread_base, thread_no, core_no, pkg_no); 163c98d5d94SLen Brown 164c98d5d94SLen Brown if (cpu_is_not_present(t->cpu_id)) 165c98d5d94SLen Brown continue; 166c98d5d94SLen Brown 167c98d5d94SLen Brown c = GET_CORE(core_base, core_no, pkg_no); 168c98d5d94SLen Brown p = GET_PKG(pkg_base, pkg_no); 169c98d5d94SLen Brown 170c98d5d94SLen Brown retval = func(t, c, p); 171c98d5d94SLen Brown if (retval) 172c98d5d94SLen Brown return retval; 173c98d5d94SLen Brown } 174c98d5d94SLen Brown } 175c98d5d94SLen Brown } 176d15cf7c1SLen Brown return 0; 177d15cf7c1SLen Brown } 178d15cf7c1SLen Brown 17988c3281fSLen Brown int cpu_migrate(int cpu) 18088c3281fSLen Brown { 181c98d5d94SLen Brown CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set); 182c98d5d94SLen Brown CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set); 183c98d5d94SLen Brown if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) 18488c3281fSLen Brown return -1; 18588c3281fSLen Brown else 18688c3281fSLen Brown return 0; 18788c3281fSLen Brown } 18888c3281fSLen Brown 18915aaa346SLen Brown int get_msr(int cpu, off_t offset, unsigned long long *msr) 190103a8feaSLen Brown { 191103a8feaSLen Brown ssize_t retval; 192103a8feaSLen Brown char pathname[32]; 193103a8feaSLen Brown int fd; 194103a8feaSLen Brown 195103a8feaSLen Brown sprintf(pathname, "/dev/cpu/%d/msr", cpu); 196103a8feaSLen Brown fd = open(pathname, O_RDONLY); 19715aaa346SLen Brown if (fd < 0) 19815aaa346SLen Brown return -1; 199103a8feaSLen Brown 20015aaa346SLen Brown retval = pread(fd, msr, sizeof *msr, offset); 201103a8feaSLen Brown close(fd); 20215aaa346SLen Brown 20315aaa346SLen Brown if (retval != sizeof *msr) 20415aaa346SLen Brown return -1; 20515aaa346SLen Brown 20615aaa346SLen Brown return 0; 207103a8feaSLen Brown } 208103a8feaSLen Brown 209a829eb4dSLen Brown void print_header(void) 210103a8feaSLen Brown { 211103a8feaSLen Brown if (show_pkg) 212c98d5d94SLen Brown outp += sprintf(outp, "pk"); 213e23da037SLen Brown if (show_pkg) 214c98d5d94SLen Brown outp += sprintf(outp, " "); 215103a8feaSLen Brown if (show_core) 216c98d5d94SLen Brown outp += sprintf(outp, "cor"); 217103a8feaSLen Brown if (show_cpu) 218c98d5d94SLen Brown outp += sprintf(outp, " CPU"); 219e23da037SLen Brown if (show_pkg || show_core || show_cpu) 220c98d5d94SLen Brown outp += sprintf(outp, " "); 221103a8feaSLen Brown if (do_nhm_cstates) 222c98d5d94SLen Brown outp += sprintf(outp, " %%c0"); 223103a8feaSLen Brown if (has_aperf) 224c98d5d94SLen Brown outp += sprintf(outp, " GHz"); 225c98d5d94SLen Brown outp += sprintf(outp, " TSC"); 226*130ff304SLen Brown if (extra_msr_offset) 227*130ff304SLen Brown outp += sprintf(outp, " MSR 0x%04X", extra_msr_offset); 228103a8feaSLen Brown if (do_nhm_cstates) 229c98d5d94SLen Brown outp += sprintf(outp, " %%c1"); 230103a8feaSLen Brown if (do_nhm_cstates) 231c98d5d94SLen Brown outp += sprintf(outp, " %%c3"); 232103a8feaSLen Brown if (do_nhm_cstates) 233c98d5d94SLen Brown outp += sprintf(outp, " %%c6"); 234103a8feaSLen Brown if (do_snb_cstates) 235c98d5d94SLen Brown outp += sprintf(outp, " %%c7"); 236103a8feaSLen Brown if (do_snb_cstates) 237c98d5d94SLen Brown outp += sprintf(outp, " %%pc2"); 238103a8feaSLen Brown if (do_nhm_cstates) 239c98d5d94SLen Brown outp += sprintf(outp, " %%pc3"); 240103a8feaSLen Brown if (do_nhm_cstates) 241c98d5d94SLen Brown outp += sprintf(outp, " %%pc6"); 242103a8feaSLen Brown if (do_snb_cstates) 243c98d5d94SLen Brown outp += sprintf(outp, " %%pc7"); 244103a8feaSLen Brown 245c98d5d94SLen Brown outp += sprintf(outp, "\n"); 246103a8feaSLen Brown } 247103a8feaSLen Brown 248c98d5d94SLen Brown int dump_counters(struct thread_data *t, struct core_data *c, 249c98d5d94SLen Brown struct pkg_data *p) 250103a8feaSLen Brown { 251c98d5d94SLen Brown fprintf(stderr, "t %p, c %p, p %p\n", t, c, p); 252c98d5d94SLen Brown 253c98d5d94SLen Brown if (t) { 254c98d5d94SLen Brown fprintf(stderr, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags); 255c98d5d94SLen Brown fprintf(stderr, "TSC: %016llX\n", t->tsc); 256c98d5d94SLen Brown fprintf(stderr, "aperf: %016llX\n", t->aperf); 257c98d5d94SLen Brown fprintf(stderr, "mperf: %016llX\n", t->mperf); 258c98d5d94SLen Brown fprintf(stderr, "c1: %016llX\n", t->c1); 259c98d5d94SLen Brown fprintf(stderr, "msr0x%x: %016llX\n", 260c98d5d94SLen Brown extra_msr_offset, t->extra_msr); 261103a8feaSLen Brown } 262103a8feaSLen Brown 263c98d5d94SLen Brown if (c) { 264c98d5d94SLen Brown fprintf(stderr, "core: %d\n", c->core_id); 265c98d5d94SLen Brown fprintf(stderr, "c3: %016llX\n", c->c3); 266c98d5d94SLen Brown fprintf(stderr, "c6: %016llX\n", c->c6); 267c98d5d94SLen Brown fprintf(stderr, "c7: %016llX\n", c->c7); 268c98d5d94SLen Brown } 269103a8feaSLen Brown 270c98d5d94SLen Brown if (p) { 271c98d5d94SLen Brown fprintf(stderr, "package: %d\n", p->package_id); 272c98d5d94SLen Brown fprintf(stderr, "pc2: %016llX\n", p->pc2); 273c98d5d94SLen Brown fprintf(stderr, "pc3: %016llX\n", p->pc3); 274c98d5d94SLen Brown fprintf(stderr, "pc6: %016llX\n", p->pc6); 275c98d5d94SLen Brown fprintf(stderr, "pc7: %016llX\n", p->pc7); 276c98d5d94SLen Brown } 277c98d5d94SLen Brown return 0; 278103a8feaSLen Brown } 279103a8feaSLen Brown 280e23da037SLen Brown /* 281e23da037SLen Brown * column formatting convention & formats 282e23da037SLen Brown * package: "pk" 2 columns %2d 283e23da037SLen Brown * core: "cor" 3 columns %3d 284e23da037SLen Brown * CPU: "CPU" 3 columns %3d 285e23da037SLen Brown * GHz: "GHz" 3 columns %3.2 286e23da037SLen Brown * TSC: "TSC" 3 columns %3.2 287e23da037SLen Brown * percentage " %pc3" %6.2 288e23da037SLen Brown */ 289c98d5d94SLen Brown int format_counters(struct thread_data *t, struct core_data *c, 290c98d5d94SLen Brown struct pkg_data *p) 291103a8feaSLen Brown { 292103a8feaSLen Brown double interval_float; 293103a8feaSLen Brown 294c98d5d94SLen Brown /* if showing only 1st thread in core and this isn't one, bail out */ 295c98d5d94SLen Brown if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 296c98d5d94SLen Brown return 0; 297c98d5d94SLen Brown 298c98d5d94SLen Brown /* if showing only 1st thread in pkg and this isn't one, bail out */ 299c98d5d94SLen Brown if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 300c98d5d94SLen Brown return 0; 301c98d5d94SLen Brown 302103a8feaSLen Brown interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0; 303103a8feaSLen Brown 304c98d5d94SLen Brown /* topo columns, print blanks on 1st (average) line */ 305c98d5d94SLen Brown if (t == &average.threads) { 306103a8feaSLen Brown if (show_pkg) 307c98d5d94SLen Brown outp += sprintf(outp, " "); 308e23da037SLen Brown if (show_pkg && show_core) 309c98d5d94SLen Brown outp += sprintf(outp, " "); 310103a8feaSLen Brown if (show_core) 311c98d5d94SLen Brown outp += sprintf(outp, " "); 312103a8feaSLen Brown if (show_cpu) 313c98d5d94SLen Brown outp += sprintf(outp, " " " "); 314103a8feaSLen Brown } else { 315c98d5d94SLen Brown if (show_pkg) { 316c98d5d94SLen Brown if (p) 317c98d5d94SLen Brown outp += sprintf(outp, "%2d", p->package_id); 318c98d5d94SLen Brown else 319c98d5d94SLen Brown outp += sprintf(outp, " "); 320c98d5d94SLen Brown } 321e23da037SLen Brown if (show_pkg && show_core) 322c98d5d94SLen Brown outp += sprintf(outp, " "); 323c98d5d94SLen Brown if (show_core) { 324c98d5d94SLen Brown if (c) 325c98d5d94SLen Brown outp += sprintf(outp, "%3d", c->core_id); 326c98d5d94SLen Brown else 327c98d5d94SLen Brown outp += sprintf(outp, " "); 328c98d5d94SLen Brown } 329103a8feaSLen Brown if (show_cpu) 330c98d5d94SLen Brown outp += sprintf(outp, " %3d", t->cpu_id); 331103a8feaSLen Brown } 332103a8feaSLen Brown 333103a8feaSLen Brown /* %c0 */ 334103a8feaSLen Brown if (do_nhm_cstates) { 335e23da037SLen Brown if (show_pkg || show_core || show_cpu) 336c98d5d94SLen Brown outp += sprintf(outp, " "); 337103a8feaSLen Brown if (!skip_c0) 338c98d5d94SLen Brown outp += sprintf(outp, "%6.2f", 100.0 * t->mperf/t->tsc); 339103a8feaSLen Brown else 340c98d5d94SLen Brown outp += sprintf(outp, " ****"); 341103a8feaSLen Brown } 342103a8feaSLen Brown 343103a8feaSLen Brown /* GHz */ 344103a8feaSLen Brown if (has_aperf) { 345103a8feaSLen Brown if (!aperf_mperf_unstable) { 346c98d5d94SLen Brown outp += sprintf(outp, " %3.2f", 347c98d5d94SLen Brown 1.0 * t->tsc / units * t->aperf / 348c98d5d94SLen Brown t->mperf / interval_float); 349103a8feaSLen Brown } else { 350c98d5d94SLen Brown if (t->aperf > t->tsc || t->mperf > t->tsc) { 351c98d5d94SLen Brown outp += sprintf(outp, " ***"); 352103a8feaSLen Brown } else { 353c98d5d94SLen Brown outp += sprintf(outp, "%3.1f*", 354c98d5d94SLen Brown 1.0 * t->tsc / 355c98d5d94SLen Brown units * t->aperf / 356c98d5d94SLen Brown t->mperf / interval_float); 357103a8feaSLen Brown } 358103a8feaSLen Brown } 359103a8feaSLen Brown } 360103a8feaSLen Brown 361103a8feaSLen Brown /* TSC */ 362c98d5d94SLen Brown outp += sprintf(outp, "%5.2f", 1.0 * t->tsc/units/interval_float); 363103a8feaSLen Brown 364*130ff304SLen Brown /* MSR */ 365*130ff304SLen Brown if (extra_msr_offset) 366*130ff304SLen Brown outp += sprintf(outp, " 0x%016llx", t->extra_msr); 367*130ff304SLen Brown 368103a8feaSLen Brown if (do_nhm_cstates) { 369103a8feaSLen Brown if (!skip_c1) 370c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * t->c1/t->tsc); 371103a8feaSLen Brown else 372c98d5d94SLen Brown outp += sprintf(outp, " ****"); 373103a8feaSLen Brown } 374c98d5d94SLen Brown 375c98d5d94SLen Brown /* print per-core data only for 1st thread in core */ 376c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 377c98d5d94SLen Brown goto done; 378c98d5d94SLen Brown 379103a8feaSLen Brown if (do_nhm_cstates) 380c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * c->c3/t->tsc); 381103a8feaSLen Brown if (do_nhm_cstates) 382c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * c->c6/t->tsc); 383103a8feaSLen Brown if (do_snb_cstates) 384c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * c->c7/t->tsc); 385c98d5d94SLen Brown 386c98d5d94SLen Brown /* print per-package data only for 1st core in package */ 387c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 388c98d5d94SLen Brown goto done; 389c98d5d94SLen Brown 390103a8feaSLen Brown if (do_snb_cstates) 391c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc2/t->tsc); 392103a8feaSLen Brown if (do_nhm_cstates) 393c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc3/t->tsc); 394103a8feaSLen Brown if (do_nhm_cstates) 395c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc6/t->tsc); 396103a8feaSLen Brown if (do_snb_cstates) 397c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc7/t->tsc); 398c98d5d94SLen Brown done: 399c98d5d94SLen Brown outp += sprintf(outp, "\n"); 400c98d5d94SLen Brown 401c98d5d94SLen Brown return 0; 402103a8feaSLen Brown } 403103a8feaSLen Brown 404c98d5d94SLen Brown void flush_stdout() 405103a8feaSLen Brown { 406c98d5d94SLen Brown fputs(output_buffer, stdout); 407c98d5d94SLen Brown outp = output_buffer; 408c98d5d94SLen Brown } 409c98d5d94SLen Brown void flush_stderr() 410c98d5d94SLen Brown { 411c98d5d94SLen Brown fputs(output_buffer, stderr); 412c98d5d94SLen Brown outp = output_buffer; 413c98d5d94SLen Brown } 414c98d5d94SLen Brown void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 415c98d5d94SLen Brown { 416e23da037SLen Brown static int printed; 417103a8feaSLen Brown 418e23da037SLen Brown if (!printed || !summary_only) 419103a8feaSLen Brown print_header(); 420103a8feaSLen Brown 421c98d5d94SLen Brown if (topo.num_cpus > 1) 422c98d5d94SLen Brown format_counters(&average.threads, &average.cores, 423c98d5d94SLen Brown &average.packages); 424103a8feaSLen Brown 425e23da037SLen Brown printed = 1; 426e23da037SLen Brown 427e23da037SLen Brown if (summary_only) 428e23da037SLen Brown return; 429e23da037SLen Brown 430c98d5d94SLen Brown for_all_cpus(format_counters, t, c, p); 431103a8feaSLen Brown } 432103a8feaSLen Brown 433c98d5d94SLen Brown void 434c98d5d94SLen Brown delta_package(struct pkg_data *new, struct pkg_data *old) 435103a8feaSLen Brown { 436c98d5d94SLen Brown old->pc2 = new->pc2 - old->pc2; 437c98d5d94SLen Brown old->pc3 = new->pc3 - old->pc3; 438c98d5d94SLen Brown old->pc6 = new->pc6 - old->pc6; 439c98d5d94SLen Brown old->pc7 = new->pc7 - old->pc7; 440103a8feaSLen Brown } 441103a8feaSLen Brown 442c98d5d94SLen Brown void 443c98d5d94SLen Brown delta_core(struct core_data *new, struct core_data *old) 444c98d5d94SLen Brown { 445c98d5d94SLen Brown old->c3 = new->c3 - old->c3; 446c98d5d94SLen Brown old->c6 = new->c6 - old->c6; 447c98d5d94SLen Brown old->c7 = new->c7 - old->c7; 448103a8feaSLen Brown } 449c98d5d94SLen Brown 450c3ae331dSLen Brown /* 451c3ae331dSLen Brown * old = new - old 452c3ae331dSLen Brown */ 453c98d5d94SLen Brown void 454c98d5d94SLen Brown delta_thread(struct thread_data *new, struct thread_data *old, 455c98d5d94SLen Brown struct core_data *core_delta) 456c98d5d94SLen Brown { 457c98d5d94SLen Brown old->tsc = new->tsc - old->tsc; 458c98d5d94SLen Brown 459103a8feaSLen Brown /* check for TSC < 1 Mcycles over interval */ 460c98d5d94SLen Brown if (old->tsc < (1000 * 1000)) { 461c98d5d94SLen Brown fprintf(stderr, "Insanely slow TSC rate, TSC stops in idle?\n"); 462c98d5d94SLen Brown fprintf(stderr, "You can disable all c-states by booting with \"idle=poll\"\n"); 463c98d5d94SLen Brown fprintf(stderr, "or just the deep ones with \"processor.max_cstate=1\"\n"); 464103a8feaSLen Brown exit(-3); 465103a8feaSLen Brown } 466103a8feaSLen Brown 467c98d5d94SLen Brown old->c1 = new->c1 - old->c1; 468c98d5d94SLen Brown 469c98d5d94SLen Brown if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) { 470c98d5d94SLen Brown old->aperf = new->aperf - old->aperf; 471c98d5d94SLen Brown old->mperf = new->mperf - old->mperf; 472c98d5d94SLen Brown } else { 473c98d5d94SLen Brown 474103a8feaSLen Brown if (!aperf_mperf_unstable) { 475103a8feaSLen Brown fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname); 476103a8feaSLen Brown fprintf(stderr, "* Frequency results do not cover entire interval *\n"); 477103a8feaSLen Brown fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n"); 478103a8feaSLen Brown 479103a8feaSLen Brown aperf_mperf_unstable = 1; 480103a8feaSLen Brown } 481103a8feaSLen Brown /* 482103a8feaSLen Brown * mperf delta is likely a huge "positive" number 483103a8feaSLen Brown * can not use it for calculating c0 time 484103a8feaSLen Brown */ 485103a8feaSLen Brown skip_c0 = 1; 486103a8feaSLen Brown skip_c1 = 1; 487103a8feaSLen Brown } 488103a8feaSLen Brown 489c98d5d94SLen Brown 490103a8feaSLen Brown /* 491c3ae331dSLen Brown * As counter collection is not atomic, 492c3ae331dSLen Brown * it is possible for mperf's non-halted cycles + idle states 493103a8feaSLen Brown * to exceed TSC's all cycles: show c1 = 0% in that case. 494103a8feaSLen Brown */ 495c3ae331dSLen Brown if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc) 496c98d5d94SLen Brown old->c1 = 0; 497c98d5d94SLen Brown else { 498c98d5d94SLen Brown /* normal case, derive c1 */ 499c98d5d94SLen Brown old->c1 = old->tsc - old->mperf - core_delta->c3 500c98d5d94SLen Brown - core_delta->c6 - core_delta->c7; 501c98d5d94SLen Brown } 502c3ae331dSLen Brown 503c98d5d94SLen Brown if (old->mperf == 0) { 504c3ae331dSLen Brown if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id); 505c98d5d94SLen Brown old->mperf = 1; /* divide by 0 protection */ 506c98d5d94SLen Brown } 507103a8feaSLen Brown 508103a8feaSLen Brown /* 509103a8feaSLen Brown * for "extra msr", just copy the latest w/o subtracting 510103a8feaSLen Brown */ 511c98d5d94SLen Brown old->extra_msr = new->extra_msr; 512103a8feaSLen Brown } 513c98d5d94SLen Brown 514c98d5d94SLen Brown int delta_cpu(struct thread_data *t, struct core_data *c, 515c98d5d94SLen Brown struct pkg_data *p, struct thread_data *t2, 516c98d5d94SLen Brown struct core_data *c2, struct pkg_data *p2) 517c98d5d94SLen Brown { 518c98d5d94SLen Brown /* calculate core delta only for 1st thread in core */ 519c98d5d94SLen Brown if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE) 520c98d5d94SLen Brown delta_core(c, c2); 521c98d5d94SLen Brown 522c98d5d94SLen Brown /* always calculate thread delta */ 523c98d5d94SLen Brown delta_thread(t, t2, c2); /* c2 is core delta */ 524c98d5d94SLen Brown 525c98d5d94SLen Brown /* calculate package delta only for 1st core in package */ 526c98d5d94SLen Brown if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) 527c98d5d94SLen Brown delta_package(p, p2); 528c98d5d94SLen Brown 529103a8feaSLen Brown return 0; 530103a8feaSLen Brown } 531103a8feaSLen Brown 532c98d5d94SLen Brown void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 533103a8feaSLen Brown { 534c98d5d94SLen Brown t->tsc = 0; 535c98d5d94SLen Brown t->aperf = 0; 536c98d5d94SLen Brown t->mperf = 0; 537c98d5d94SLen Brown t->c1 = 0; 538103a8feaSLen Brown 539c98d5d94SLen Brown /* tells format_counters to dump all fields from this set */ 540c98d5d94SLen Brown t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE; 541c98d5d94SLen Brown 542c98d5d94SLen Brown c->c3 = 0; 543c98d5d94SLen Brown c->c6 = 0; 544c98d5d94SLen Brown c->c7 = 0; 545c98d5d94SLen Brown 546c98d5d94SLen Brown p->pc2 = 0; 547c98d5d94SLen Brown p->pc3 = 0; 548c98d5d94SLen Brown p->pc6 = 0; 549c98d5d94SLen Brown p->pc7 = 0; 550103a8feaSLen Brown } 551c98d5d94SLen Brown int sum_counters(struct thread_data *t, struct core_data *c, 552c98d5d94SLen Brown struct pkg_data *p) 553103a8feaSLen Brown { 554c98d5d94SLen Brown average.threads.tsc += t->tsc; 555c98d5d94SLen Brown average.threads.aperf += t->aperf; 556c98d5d94SLen Brown average.threads.mperf += t->mperf; 557c98d5d94SLen Brown average.threads.c1 += t->c1; 55815aaa346SLen Brown 559c98d5d94SLen Brown /* sum per-core values only for 1st thread in core */ 560c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 561c98d5d94SLen Brown return 0; 562c98d5d94SLen Brown 563c98d5d94SLen Brown average.cores.c3 += c->c3; 564c98d5d94SLen Brown average.cores.c6 += c->c6; 565c98d5d94SLen Brown average.cores.c7 += c->c7; 566c98d5d94SLen Brown 567c98d5d94SLen Brown /* sum per-pkg values only for 1st core in pkg */ 568c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 569c98d5d94SLen Brown return 0; 570c98d5d94SLen Brown 571c98d5d94SLen Brown average.packages.pc2 += p->pc2; 572c98d5d94SLen Brown average.packages.pc3 += p->pc3; 573c98d5d94SLen Brown average.packages.pc6 += p->pc6; 574c98d5d94SLen Brown average.packages.pc7 += p->pc7; 575c98d5d94SLen Brown 576c98d5d94SLen Brown return 0; 577c98d5d94SLen Brown } 578c98d5d94SLen Brown /* 579c98d5d94SLen Brown * sum the counters for all cpus in the system 580c98d5d94SLen Brown * compute the weighted average 581c98d5d94SLen Brown */ 582c98d5d94SLen Brown void compute_average(struct thread_data *t, struct core_data *c, 583c98d5d94SLen Brown struct pkg_data *p) 584c98d5d94SLen Brown { 585c98d5d94SLen Brown clear_counters(&average.threads, &average.cores, &average.packages); 586c98d5d94SLen Brown 587c98d5d94SLen Brown for_all_cpus(sum_counters, t, c, p); 588c98d5d94SLen Brown 589c98d5d94SLen Brown average.threads.tsc /= topo.num_cpus; 590c98d5d94SLen Brown average.threads.aperf /= topo.num_cpus; 591c98d5d94SLen Brown average.threads.mperf /= topo.num_cpus; 592c98d5d94SLen Brown average.threads.c1 /= topo.num_cpus; 593c98d5d94SLen Brown 594c98d5d94SLen Brown average.cores.c3 /= topo.num_cores; 595c98d5d94SLen Brown average.cores.c6 /= topo.num_cores; 596c98d5d94SLen Brown average.cores.c7 /= topo.num_cores; 597c98d5d94SLen Brown 598c98d5d94SLen Brown average.packages.pc2 /= topo.num_packages; 599c98d5d94SLen Brown average.packages.pc3 /= topo.num_packages; 600c98d5d94SLen Brown average.packages.pc6 /= topo.num_packages; 601c98d5d94SLen Brown average.packages.pc7 /= topo.num_packages; 602c98d5d94SLen Brown } 603c98d5d94SLen Brown 604c98d5d94SLen Brown static unsigned long long rdtsc(void) 605c98d5d94SLen Brown { 606c98d5d94SLen Brown unsigned int low, high; 607c98d5d94SLen Brown 608c98d5d94SLen Brown asm volatile("rdtsc" : "=a" (low), "=d" (high)); 609c98d5d94SLen Brown 610c98d5d94SLen Brown return low | ((unsigned long long)high) << 32; 611c98d5d94SLen Brown } 612c98d5d94SLen Brown 613c98d5d94SLen Brown 614c98d5d94SLen Brown /* 615c98d5d94SLen Brown * get_counters(...) 616c98d5d94SLen Brown * migrate to cpu 617c98d5d94SLen Brown * acquire and record local counters for that cpu 618c98d5d94SLen Brown */ 619c98d5d94SLen Brown int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 620c98d5d94SLen Brown { 621c98d5d94SLen Brown int cpu = t->cpu_id; 622c98d5d94SLen Brown 623c98d5d94SLen Brown if (cpu_migrate(cpu)) 62415aaa346SLen Brown return -1; 62515aaa346SLen Brown 626c98d5d94SLen Brown t->tsc = rdtsc(); /* we are running on local CPU of interest */ 62715aaa346SLen Brown 62815aaa346SLen Brown if (has_aperf) { 629c98d5d94SLen Brown if (get_msr(cpu, MSR_APERF, &t->aperf)) 630c98d5d94SLen Brown return -3; 631c98d5d94SLen Brown if (get_msr(cpu, MSR_MPERF, &t->mperf)) 632c98d5d94SLen Brown return -4; 63388c3281fSLen Brown } 63488c3281fSLen Brown 635c98d5d94SLen Brown if (extra_msr_offset) 636c98d5d94SLen Brown if (get_msr(cpu, extra_msr_offset, &t->extra_msr)) 637c98d5d94SLen Brown return -5; 638c98d5d94SLen Brown 639c98d5d94SLen Brown /* collect core counters only for 1st thread in core */ 640c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 641c98d5d94SLen Brown return 0; 642c98d5d94SLen Brown 64315aaa346SLen Brown if (do_nhm_cstates) { 644c98d5d94SLen Brown if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3)) 645c98d5d94SLen Brown return -6; 646c98d5d94SLen Brown if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6)) 647c98d5d94SLen Brown return -7; 648103a8feaSLen Brown } 64915aaa346SLen Brown 65015aaa346SLen Brown if (do_snb_cstates) 651c98d5d94SLen Brown if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7)) 652c98d5d94SLen Brown return -8; 653c98d5d94SLen Brown 654c98d5d94SLen Brown /* collect package counters only for 1st core in package */ 655c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 656c98d5d94SLen Brown return 0; 65715aaa346SLen Brown 65815aaa346SLen Brown if (do_nhm_cstates) { 659c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3)) 660c98d5d94SLen Brown return -9; 661c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6)) 662c98d5d94SLen Brown return -10; 66315aaa346SLen Brown } 66415aaa346SLen Brown if (do_snb_cstates) { 665c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2)) 666c98d5d94SLen Brown return -11; 667c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7)) 668c98d5d94SLen Brown return -12; 66915aaa346SLen Brown } 67015aaa346SLen Brown return 0; 671103a8feaSLen Brown } 672103a8feaSLen Brown 673c98d5d94SLen Brown void print_verbose_header(void) 674103a8feaSLen Brown { 675103a8feaSLen Brown unsigned long long msr; 676103a8feaSLen Brown unsigned int ratio; 677103a8feaSLen Brown 678103a8feaSLen Brown if (!do_nehalem_platform_info) 679103a8feaSLen Brown return; 680103a8feaSLen Brown 68115aaa346SLen Brown get_msr(0, MSR_NEHALEM_PLATFORM_INFO, &msr); 682103a8feaSLen Brown 6836574a5d5SLen Brown if (verbose > 1) 6846574a5d5SLen Brown fprintf(stderr, "MSR_NEHALEM_PLATFORM_INFO: 0x%llx\n", msr); 6856574a5d5SLen Brown 686103a8feaSLen Brown ratio = (msr >> 40) & 0xFF; 687103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n", 688103a8feaSLen Brown ratio, bclk, ratio * bclk); 689103a8feaSLen Brown 690103a8feaSLen Brown ratio = (msr >> 8) & 0xFF; 691103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n", 692103a8feaSLen Brown ratio, bclk, ratio * bclk); 693103a8feaSLen Brown 6946574a5d5SLen Brown if (!do_ivt_turbo_ratio_limit) 6956574a5d5SLen Brown goto print_nhm_turbo_ratio_limits; 6966574a5d5SLen Brown 6976574a5d5SLen Brown get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr); 6986574a5d5SLen Brown 699103a8feaSLen Brown if (verbose > 1) 7006574a5d5SLen Brown fprintf(stderr, "MSR_IVT_TURBO_RATIO_LIMIT: 0x%llx\n", msr); 7016574a5d5SLen Brown 7026574a5d5SLen Brown ratio = (msr >> 56) & 0xFF; 7036574a5d5SLen Brown if (ratio) 7046574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n", 7056574a5d5SLen Brown ratio, bclk, ratio * bclk); 7066574a5d5SLen Brown 7076574a5d5SLen Brown ratio = (msr >> 48) & 0xFF; 7086574a5d5SLen Brown if (ratio) 7096574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n", 7106574a5d5SLen Brown ratio, bclk, ratio * bclk); 7116574a5d5SLen Brown 7126574a5d5SLen Brown ratio = (msr >> 40) & 0xFF; 7136574a5d5SLen Brown if (ratio) 7146574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n", 7156574a5d5SLen Brown ratio, bclk, ratio * bclk); 7166574a5d5SLen Brown 7176574a5d5SLen Brown ratio = (msr >> 32) & 0xFF; 7186574a5d5SLen Brown if (ratio) 7196574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n", 7206574a5d5SLen Brown ratio, bclk, ratio * bclk); 7216574a5d5SLen Brown 7226574a5d5SLen Brown ratio = (msr >> 24) & 0xFF; 7236574a5d5SLen Brown if (ratio) 7246574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n", 7256574a5d5SLen Brown ratio, bclk, ratio * bclk); 7266574a5d5SLen Brown 7276574a5d5SLen Brown ratio = (msr >> 16) & 0xFF; 7286574a5d5SLen Brown if (ratio) 7296574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n", 7306574a5d5SLen Brown ratio, bclk, ratio * bclk); 7316574a5d5SLen Brown 7326574a5d5SLen Brown ratio = (msr >> 8) & 0xFF; 7336574a5d5SLen Brown if (ratio) 7346574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n", 7356574a5d5SLen Brown ratio, bclk, ratio * bclk); 7366574a5d5SLen Brown 7376574a5d5SLen Brown ratio = (msr >> 0) & 0xFF; 7386574a5d5SLen Brown if (ratio) 7396574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n", 7406574a5d5SLen Brown ratio, bclk, ratio * bclk); 7416574a5d5SLen Brown 7426574a5d5SLen Brown print_nhm_turbo_ratio_limits: 743103a8feaSLen Brown 744103a8feaSLen Brown if (!do_nehalem_turbo_ratio_limit) 745103a8feaSLen Brown return; 746103a8feaSLen Brown 74715aaa346SLen Brown get_msr(0, MSR_NEHALEM_TURBO_RATIO_LIMIT, &msr); 748103a8feaSLen Brown 7496574a5d5SLen Brown if (verbose > 1) 7506574a5d5SLen Brown fprintf(stderr, "MSR_NEHALEM_TURBO_RATIO_LIMIT: 0x%llx\n", msr); 7516574a5d5SLen Brown 7526574a5d5SLen Brown ratio = (msr >> 56) & 0xFF; 7536574a5d5SLen Brown if (ratio) 7546574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n", 7556574a5d5SLen Brown ratio, bclk, ratio * bclk); 7566574a5d5SLen Brown 7576574a5d5SLen Brown ratio = (msr >> 48) & 0xFF; 7586574a5d5SLen Brown if (ratio) 7596574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n", 7606574a5d5SLen Brown ratio, bclk, ratio * bclk); 7616574a5d5SLen Brown 7626574a5d5SLen Brown ratio = (msr >> 40) & 0xFF; 7636574a5d5SLen Brown if (ratio) 7646574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n", 7656574a5d5SLen Brown ratio, bclk, ratio * bclk); 7666574a5d5SLen Brown 7676574a5d5SLen Brown ratio = (msr >> 32) & 0xFF; 7686574a5d5SLen Brown if (ratio) 7696574a5d5SLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n", 7706574a5d5SLen Brown ratio, bclk, ratio * bclk); 7716574a5d5SLen Brown 772103a8feaSLen Brown ratio = (msr >> 24) & 0xFF; 773103a8feaSLen Brown if (ratio) 774103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n", 775103a8feaSLen Brown ratio, bclk, ratio * bclk); 776103a8feaSLen Brown 777103a8feaSLen Brown ratio = (msr >> 16) & 0xFF; 778103a8feaSLen Brown if (ratio) 779103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n", 780103a8feaSLen Brown ratio, bclk, ratio * bclk); 781103a8feaSLen Brown 782103a8feaSLen Brown ratio = (msr >> 8) & 0xFF; 783103a8feaSLen Brown if (ratio) 784103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n", 785103a8feaSLen Brown ratio, bclk, ratio * bclk); 786103a8feaSLen Brown 787103a8feaSLen Brown ratio = (msr >> 0) & 0xFF; 788103a8feaSLen Brown if (ratio) 789103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n", 790103a8feaSLen Brown ratio, bclk, ratio * bclk); 791103a8feaSLen Brown } 792103a8feaSLen Brown 793c98d5d94SLen Brown void free_all_buffers(void) 794103a8feaSLen Brown { 795c98d5d94SLen Brown CPU_FREE(cpu_present_set); 796c98d5d94SLen Brown cpu_present_set = NULL; 797c98d5d94SLen Brown cpu_present_set = 0; 798103a8feaSLen Brown 799c98d5d94SLen Brown CPU_FREE(cpu_affinity_set); 800c98d5d94SLen Brown cpu_affinity_set = NULL; 801c98d5d94SLen Brown cpu_affinity_setsize = 0; 802103a8feaSLen Brown 803c98d5d94SLen Brown free(thread_even); 804c98d5d94SLen Brown free(core_even); 805c98d5d94SLen Brown free(package_even); 806103a8feaSLen Brown 807c98d5d94SLen Brown thread_even = NULL; 808c98d5d94SLen Brown core_even = NULL; 809c98d5d94SLen Brown package_even = NULL; 810103a8feaSLen Brown 811c98d5d94SLen Brown free(thread_odd); 812c98d5d94SLen Brown free(core_odd); 813c98d5d94SLen Brown free(package_odd); 814103a8feaSLen Brown 815c98d5d94SLen Brown thread_odd = NULL; 816c98d5d94SLen Brown core_odd = NULL; 817c98d5d94SLen Brown package_odd = NULL; 818103a8feaSLen Brown 819c98d5d94SLen Brown free(output_buffer); 820c98d5d94SLen Brown output_buffer = NULL; 821c98d5d94SLen Brown outp = NULL; 822103a8feaSLen Brown } 823103a8feaSLen Brown 824103a8feaSLen Brown /* 825c98d5d94SLen Brown * cpu_is_first_sibling_in_core(cpu) 826c98d5d94SLen Brown * return 1 if given CPU is 1st HT sibling in the core 827103a8feaSLen Brown */ 828c98d5d94SLen Brown int cpu_is_first_sibling_in_core(int cpu) 829103a8feaSLen Brown { 830c98d5d94SLen Brown char path[64]; 831c98d5d94SLen Brown FILE *filep; 832c98d5d94SLen Brown int first_cpu; 833103a8feaSLen Brown 834c98d5d94SLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu); 835c98d5d94SLen Brown filep = fopen(path, "r"); 836c98d5d94SLen Brown if (filep == NULL) { 837c98d5d94SLen Brown perror(path); 838103a8feaSLen Brown exit(1); 839103a8feaSLen Brown } 840c98d5d94SLen Brown fscanf(filep, "%d", &first_cpu); 841c98d5d94SLen Brown fclose(filep); 842c98d5d94SLen Brown return (cpu == first_cpu); 843c98d5d94SLen Brown } 844103a8feaSLen Brown 845c98d5d94SLen Brown /* 846c98d5d94SLen Brown * cpu_is_first_core_in_package(cpu) 847c98d5d94SLen Brown * return 1 if given CPU is 1st core in package 848c98d5d94SLen Brown */ 849c98d5d94SLen Brown int cpu_is_first_core_in_package(int cpu) 850c98d5d94SLen Brown { 851c98d5d94SLen Brown char path[64]; 852c98d5d94SLen Brown FILE *filep; 853c98d5d94SLen Brown int first_cpu; 854c98d5d94SLen Brown 855c98d5d94SLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu); 856c98d5d94SLen Brown filep = fopen(path, "r"); 857c98d5d94SLen Brown if (filep == NULL) { 858c98d5d94SLen Brown perror(path); 859103a8feaSLen Brown exit(1); 860103a8feaSLen Brown } 861c98d5d94SLen Brown fscanf(filep, "%d", &first_cpu); 862c98d5d94SLen Brown fclose(filep); 863c98d5d94SLen Brown return (cpu == first_cpu); 864103a8feaSLen Brown } 865103a8feaSLen Brown 866103a8feaSLen Brown int get_physical_package_id(int cpu) 867103a8feaSLen Brown { 868c98d5d94SLen Brown char path[80]; 869103a8feaSLen Brown FILE *filep; 870103a8feaSLen Brown int pkg; 871103a8feaSLen Brown 872103a8feaSLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu); 873103a8feaSLen Brown filep = fopen(path, "r"); 874103a8feaSLen Brown if (filep == NULL) { 875103a8feaSLen Brown perror(path); 876103a8feaSLen Brown exit(1); 877103a8feaSLen Brown } 878103a8feaSLen Brown fscanf(filep, "%d", &pkg); 879103a8feaSLen Brown fclose(filep); 880103a8feaSLen Brown return pkg; 881103a8feaSLen Brown } 882103a8feaSLen Brown 883103a8feaSLen Brown int get_core_id(int cpu) 884103a8feaSLen Brown { 885c98d5d94SLen Brown char path[80]; 886103a8feaSLen Brown FILE *filep; 887103a8feaSLen Brown int core; 888103a8feaSLen Brown 889103a8feaSLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_id", cpu); 890103a8feaSLen Brown filep = fopen(path, "r"); 891103a8feaSLen Brown if (filep == NULL) { 892103a8feaSLen Brown perror(path); 893103a8feaSLen Brown exit(1); 894103a8feaSLen Brown } 895103a8feaSLen Brown fscanf(filep, "%d", &core); 896103a8feaSLen Brown fclose(filep); 897103a8feaSLen Brown return core; 898103a8feaSLen Brown } 899103a8feaSLen Brown 900c98d5d94SLen Brown int get_num_ht_siblings(int cpu) 901c98d5d94SLen Brown { 902c98d5d94SLen Brown char path[80]; 903c98d5d94SLen Brown FILE *filep; 904c98d5d94SLen Brown int sib1, sib2; 905c98d5d94SLen Brown int matches; 906c98d5d94SLen Brown char character; 907c98d5d94SLen Brown 908c98d5d94SLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu); 909c98d5d94SLen Brown filep = fopen(path, "r"); 910c98d5d94SLen Brown if (filep == NULL) { 911c98d5d94SLen Brown perror(path); 912c98d5d94SLen Brown exit(1); 913c98d5d94SLen Brown } 914103a8feaSLen Brown /* 915c98d5d94SLen Brown * file format: 916c98d5d94SLen Brown * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4) 917c98d5d94SLen Brown * otherwinse 1 sibling (self). 918c98d5d94SLen Brown */ 919c98d5d94SLen Brown matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2); 920c98d5d94SLen Brown 921c98d5d94SLen Brown fclose(filep); 922c98d5d94SLen Brown 923c98d5d94SLen Brown if (matches == 3) 924c98d5d94SLen Brown return 2; 925c98d5d94SLen Brown else 926c98d5d94SLen Brown return 1; 927c98d5d94SLen Brown } 928c98d5d94SLen Brown 929c98d5d94SLen Brown /* 930c98d5d94SLen Brown * run func(thread, core, package) in topology order 931c98d5d94SLen Brown * skip non-present cpus 932103a8feaSLen Brown */ 933103a8feaSLen Brown 934c98d5d94SLen Brown int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *, 935c98d5d94SLen Brown struct pkg_data *, struct thread_data *, struct core_data *, 936c98d5d94SLen Brown struct pkg_data *), struct thread_data *thread_base, 937c98d5d94SLen Brown struct core_data *core_base, struct pkg_data *pkg_base, 938c98d5d94SLen Brown struct thread_data *thread_base2, struct core_data *core_base2, 939c98d5d94SLen Brown struct pkg_data *pkg_base2) 940c98d5d94SLen Brown { 941c98d5d94SLen Brown int retval, pkg_no, core_no, thread_no; 942c98d5d94SLen Brown 943c98d5d94SLen Brown for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) { 944c98d5d94SLen Brown for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) { 945c98d5d94SLen Brown for (thread_no = 0; thread_no < 946c98d5d94SLen Brown topo.num_threads_per_core; ++thread_no) { 947c98d5d94SLen Brown struct thread_data *t, *t2; 948c98d5d94SLen Brown struct core_data *c, *c2; 949c98d5d94SLen Brown struct pkg_data *p, *p2; 950c98d5d94SLen Brown 951c98d5d94SLen Brown t = GET_THREAD(thread_base, thread_no, core_no, pkg_no); 952c98d5d94SLen Brown 953c98d5d94SLen Brown if (cpu_is_not_present(t->cpu_id)) 954c98d5d94SLen Brown continue; 955c98d5d94SLen Brown 956c98d5d94SLen Brown t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no); 957c98d5d94SLen Brown 958c98d5d94SLen Brown c = GET_CORE(core_base, core_no, pkg_no); 959c98d5d94SLen Brown c2 = GET_CORE(core_base2, core_no, pkg_no); 960c98d5d94SLen Brown 961c98d5d94SLen Brown p = GET_PKG(pkg_base, pkg_no); 962c98d5d94SLen Brown p2 = GET_PKG(pkg_base2, pkg_no); 963c98d5d94SLen Brown 964c98d5d94SLen Brown retval = func(t, c, p, t2, c2, p2); 965c98d5d94SLen Brown if (retval) 966c98d5d94SLen Brown return retval; 967c98d5d94SLen Brown } 968c98d5d94SLen Brown } 969c98d5d94SLen Brown } 970c98d5d94SLen Brown return 0; 971c98d5d94SLen Brown } 972c98d5d94SLen Brown 973c98d5d94SLen Brown /* 974c98d5d94SLen Brown * run func(cpu) on every cpu in /proc/stat 975c98d5d94SLen Brown * return max_cpu number 976c98d5d94SLen Brown */ 977c98d5d94SLen Brown int for_all_proc_cpus(int (func)(int)) 978103a8feaSLen Brown { 979103a8feaSLen Brown FILE *fp; 980c98d5d94SLen Brown int cpu_num; 981103a8feaSLen Brown int retval; 982103a8feaSLen Brown 983103a8feaSLen Brown fp = fopen(proc_stat, "r"); 984103a8feaSLen Brown if (fp == NULL) { 985103a8feaSLen Brown perror(proc_stat); 986103a8feaSLen Brown exit(1); 987103a8feaSLen Brown } 988103a8feaSLen Brown 989103a8feaSLen Brown retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n"); 990103a8feaSLen Brown if (retval != 0) { 991103a8feaSLen Brown perror("/proc/stat format"); 992103a8feaSLen Brown exit(1); 993103a8feaSLen Brown } 994103a8feaSLen Brown 995c98d5d94SLen Brown while (1) { 996c98d5d94SLen Brown retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num); 997103a8feaSLen Brown if (retval != 1) 998103a8feaSLen Brown break; 999103a8feaSLen Brown 1000c98d5d94SLen Brown retval = func(cpu_num); 1001c98d5d94SLen Brown if (retval) { 1002c98d5d94SLen Brown fclose(fp); 1003c98d5d94SLen Brown return(retval); 1004c98d5d94SLen Brown } 1005103a8feaSLen Brown } 1006103a8feaSLen Brown fclose(fp); 1007c98d5d94SLen Brown return 0; 1008103a8feaSLen Brown } 1009103a8feaSLen Brown 1010103a8feaSLen Brown void re_initialize(void) 1011103a8feaSLen Brown { 1012c98d5d94SLen Brown free_all_buffers(); 1013c98d5d94SLen Brown setup_all_buffers(); 1014c98d5d94SLen Brown printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus); 1015103a8feaSLen Brown } 1016103a8feaSLen Brown 1017c98d5d94SLen Brown 1018103a8feaSLen Brown /* 1019c98d5d94SLen Brown * count_cpus() 1020c98d5d94SLen Brown * remember the last one seen, it will be the max 1021103a8feaSLen Brown */ 1022c98d5d94SLen Brown int count_cpus(int cpu) 1023103a8feaSLen Brown { 1024c98d5d94SLen Brown if (topo.max_cpu_num < cpu) 1025c98d5d94SLen Brown topo.max_cpu_num = cpu; 1026103a8feaSLen Brown 1027c98d5d94SLen Brown topo.num_cpus += 1; 1028c98d5d94SLen Brown return 0; 1029103a8feaSLen Brown } 1030c98d5d94SLen Brown int mark_cpu_present(int cpu) 1031c98d5d94SLen Brown { 1032c98d5d94SLen Brown CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set); 103315aaa346SLen Brown return 0; 1034103a8feaSLen Brown } 1035103a8feaSLen Brown 1036103a8feaSLen Brown void turbostat_loop() 1037103a8feaSLen Brown { 1038c98d5d94SLen Brown int retval; 1039c98d5d94SLen Brown 1040103a8feaSLen Brown restart: 1041c98d5d94SLen Brown retval = for_all_cpus(get_counters, EVEN_COUNTERS); 1042c98d5d94SLen Brown if (retval) { 1043c98d5d94SLen Brown re_initialize(); 1044c98d5d94SLen Brown goto restart; 1045c98d5d94SLen Brown } 1046103a8feaSLen Brown gettimeofday(&tv_even, (struct timezone *)NULL); 1047103a8feaSLen Brown 1048103a8feaSLen Brown while (1) { 1049c98d5d94SLen Brown if (for_all_proc_cpus(cpu_is_not_present)) { 1050103a8feaSLen Brown re_initialize(); 1051103a8feaSLen Brown goto restart; 1052103a8feaSLen Brown } 1053103a8feaSLen Brown sleep(interval_sec); 1054c98d5d94SLen Brown retval = for_all_cpus(get_counters, ODD_COUNTERS); 1055c98d5d94SLen Brown if (retval) { 105615aaa346SLen Brown re_initialize(); 105715aaa346SLen Brown goto restart; 105815aaa346SLen Brown } 1059103a8feaSLen Brown gettimeofday(&tv_odd, (struct timezone *)NULL); 1060103a8feaSLen Brown timersub(&tv_odd, &tv_even, &tv_delta); 1061c98d5d94SLen Brown for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS); 1062c98d5d94SLen Brown compute_average(EVEN_COUNTERS); 1063c98d5d94SLen Brown format_all_counters(EVEN_COUNTERS); 1064c98d5d94SLen Brown flush_stdout(); 106515aaa346SLen Brown sleep(interval_sec); 1066c98d5d94SLen Brown retval = for_all_cpus(get_counters, EVEN_COUNTERS); 1067c98d5d94SLen Brown if (retval) { 1068103a8feaSLen Brown re_initialize(); 1069103a8feaSLen Brown goto restart; 1070103a8feaSLen Brown } 1071103a8feaSLen Brown gettimeofday(&tv_even, (struct timezone *)NULL); 1072103a8feaSLen Brown timersub(&tv_even, &tv_odd, &tv_delta); 1073c98d5d94SLen Brown for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS); 1074c98d5d94SLen Brown compute_average(ODD_COUNTERS); 1075c98d5d94SLen Brown format_all_counters(ODD_COUNTERS); 1076c98d5d94SLen Brown flush_stdout(); 1077103a8feaSLen Brown } 1078103a8feaSLen Brown } 1079103a8feaSLen Brown 1080103a8feaSLen Brown void check_dev_msr() 1081103a8feaSLen Brown { 1082103a8feaSLen Brown struct stat sb; 1083103a8feaSLen Brown 1084103a8feaSLen Brown if (stat("/dev/cpu/0/msr", &sb)) { 1085103a8feaSLen Brown fprintf(stderr, "no /dev/cpu/0/msr\n"); 1086103a8feaSLen Brown fprintf(stderr, "Try \"# modprobe msr\"\n"); 1087103a8feaSLen Brown exit(-5); 1088103a8feaSLen Brown } 1089103a8feaSLen Brown } 1090103a8feaSLen Brown 1091103a8feaSLen Brown void check_super_user() 1092103a8feaSLen Brown { 1093103a8feaSLen Brown if (getuid() != 0) { 1094103a8feaSLen Brown fprintf(stderr, "must be root\n"); 1095103a8feaSLen Brown exit(-6); 1096103a8feaSLen Brown } 1097103a8feaSLen Brown } 1098103a8feaSLen Brown 1099103a8feaSLen Brown int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model) 1100103a8feaSLen Brown { 1101103a8feaSLen Brown if (!genuine_intel) 1102103a8feaSLen Brown return 0; 1103103a8feaSLen Brown 1104103a8feaSLen Brown if (family != 6) 1105103a8feaSLen Brown return 0; 1106103a8feaSLen Brown 1107103a8feaSLen Brown switch (model) { 1108103a8feaSLen Brown case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */ 1109103a8feaSLen Brown case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */ 1110103a8feaSLen Brown case 0x1F: /* Core i7 and i5 Processor - Nehalem */ 1111103a8feaSLen Brown case 0x25: /* Westmere Client - Clarkdale, Arrandale */ 1112103a8feaSLen Brown case 0x2C: /* Westmere EP - Gulftown */ 1113103a8feaSLen Brown case 0x2A: /* SNB */ 1114103a8feaSLen Brown case 0x2D: /* SNB Xeon */ 1115553575f1SLen Brown case 0x3A: /* IVB */ 11161300651bSLen Brown case 0x3E: /* IVB Xeon */ 1117103a8feaSLen Brown return 1; 1118103a8feaSLen Brown case 0x2E: /* Nehalem-EX Xeon - Beckton */ 1119103a8feaSLen Brown case 0x2F: /* Westmere-EX Xeon - Eagleton */ 1120103a8feaSLen Brown default: 1121103a8feaSLen Brown return 0; 1122103a8feaSLen Brown } 1123103a8feaSLen Brown } 11246574a5d5SLen Brown int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model) 11256574a5d5SLen Brown { 11266574a5d5SLen Brown if (!genuine_intel) 11276574a5d5SLen Brown return 0; 11286574a5d5SLen Brown 11296574a5d5SLen Brown if (family != 6) 11306574a5d5SLen Brown return 0; 11316574a5d5SLen Brown 11326574a5d5SLen Brown switch (model) { 11336574a5d5SLen Brown case 0x3E: /* IVB Xeon */ 11346574a5d5SLen Brown return 1; 11356574a5d5SLen Brown default: 11366574a5d5SLen Brown return 0; 11376574a5d5SLen Brown } 11386574a5d5SLen Brown } 11396574a5d5SLen Brown 1140103a8feaSLen Brown 1141103a8feaSLen Brown int is_snb(unsigned int family, unsigned int model) 1142103a8feaSLen Brown { 1143103a8feaSLen Brown if (!genuine_intel) 1144103a8feaSLen Brown return 0; 1145103a8feaSLen Brown 1146103a8feaSLen Brown switch (model) { 1147103a8feaSLen Brown case 0x2A: 1148103a8feaSLen Brown case 0x2D: 1149650a37f3SLen Brown case 0x3A: /* IVB */ 11501300651bSLen Brown case 0x3E: /* IVB Xeon */ 1151103a8feaSLen Brown return 1; 1152103a8feaSLen Brown } 1153103a8feaSLen Brown return 0; 1154103a8feaSLen Brown } 1155103a8feaSLen Brown 1156103a8feaSLen Brown double discover_bclk(unsigned int family, unsigned int model) 1157103a8feaSLen Brown { 1158103a8feaSLen Brown if (is_snb(family, model)) 1159103a8feaSLen Brown return 100.00; 1160103a8feaSLen Brown else 1161103a8feaSLen Brown return 133.33; 1162103a8feaSLen Brown } 1163103a8feaSLen Brown 1164103a8feaSLen Brown void check_cpuid() 1165103a8feaSLen Brown { 1166103a8feaSLen Brown unsigned int eax, ebx, ecx, edx, max_level; 1167103a8feaSLen Brown unsigned int fms, family, model, stepping; 1168103a8feaSLen Brown 1169103a8feaSLen Brown eax = ebx = ecx = edx = 0; 1170103a8feaSLen Brown 1171103a8feaSLen Brown asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0)); 1172103a8feaSLen Brown 1173103a8feaSLen Brown if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e) 1174103a8feaSLen Brown genuine_intel = 1; 1175103a8feaSLen Brown 1176103a8feaSLen Brown if (verbose) 1177103a8feaSLen Brown fprintf(stderr, "%.4s%.4s%.4s ", 1178103a8feaSLen Brown (char *)&ebx, (char *)&edx, (char *)&ecx); 1179103a8feaSLen Brown 1180103a8feaSLen Brown asm("cpuid" : "=a" (fms), "=c" (ecx), "=d" (edx) : "a" (1) : "ebx"); 1181103a8feaSLen Brown family = (fms >> 8) & 0xf; 1182103a8feaSLen Brown model = (fms >> 4) & 0xf; 1183103a8feaSLen Brown stepping = fms & 0xf; 1184103a8feaSLen Brown if (family == 6 || family == 0xf) 1185103a8feaSLen Brown model += ((fms >> 16) & 0xf) << 4; 1186103a8feaSLen Brown 1187103a8feaSLen Brown if (verbose) 1188103a8feaSLen Brown fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n", 1189103a8feaSLen Brown max_level, family, model, stepping, family, model, stepping); 1190103a8feaSLen Brown 1191103a8feaSLen Brown if (!(edx & (1 << 5))) { 1192103a8feaSLen Brown fprintf(stderr, "CPUID: no MSR\n"); 1193103a8feaSLen Brown exit(1); 1194103a8feaSLen Brown } 1195103a8feaSLen Brown 1196103a8feaSLen Brown /* 1197103a8feaSLen Brown * check max extended function levels of CPUID. 1198103a8feaSLen Brown * This is needed to check for invariant TSC. 1199103a8feaSLen Brown * This check is valid for both Intel and AMD. 1200103a8feaSLen Brown */ 1201103a8feaSLen Brown ebx = ecx = edx = 0; 1202103a8feaSLen Brown asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000000)); 1203103a8feaSLen Brown 1204103a8feaSLen Brown if (max_level < 0x80000007) { 1205103a8feaSLen Brown fprintf(stderr, "CPUID: no invariant TSC (max_level 0x%x)\n", max_level); 1206103a8feaSLen Brown exit(1); 1207103a8feaSLen Brown } 1208103a8feaSLen Brown 1209103a8feaSLen Brown /* 1210103a8feaSLen Brown * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8 1211103a8feaSLen Brown * this check is valid for both Intel and AMD 1212103a8feaSLen Brown */ 1213103a8feaSLen Brown asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000007)); 12148209e054SThomas Renninger has_invariant_tsc = edx & (1 << 8); 1215103a8feaSLen Brown 1216103a8feaSLen Brown if (!has_invariant_tsc) { 1217103a8feaSLen Brown fprintf(stderr, "No invariant TSC\n"); 1218103a8feaSLen Brown exit(1); 1219103a8feaSLen Brown } 1220103a8feaSLen Brown 1221103a8feaSLen Brown /* 1222103a8feaSLen Brown * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0 1223103a8feaSLen Brown * this check is valid for both Intel and AMD 1224103a8feaSLen Brown */ 1225103a8feaSLen Brown 1226103a8feaSLen Brown asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x6)); 12278209e054SThomas Renninger has_aperf = ecx & (1 << 0); 1228103a8feaSLen Brown if (!has_aperf) { 1229103a8feaSLen Brown fprintf(stderr, "No APERF MSR\n"); 1230103a8feaSLen Brown exit(1); 1231103a8feaSLen Brown } 1232103a8feaSLen Brown 1233103a8feaSLen Brown do_nehalem_platform_info = genuine_intel && has_invariant_tsc; 1234103a8feaSLen Brown do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */ 1235103a8feaSLen Brown do_snb_cstates = is_snb(family, model); 1236103a8feaSLen Brown bclk = discover_bclk(family, model); 1237103a8feaSLen Brown 1238103a8feaSLen Brown do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model); 12396574a5d5SLen Brown do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model); 1240103a8feaSLen Brown } 1241103a8feaSLen Brown 1242103a8feaSLen Brown 1243103a8feaSLen Brown void usage() 1244103a8feaSLen Brown { 1245103a8feaSLen Brown fprintf(stderr, "%s: [-v] [-M MSR#] [-i interval_sec | command ...]\n", 1246103a8feaSLen Brown progname); 1247103a8feaSLen Brown exit(1); 1248103a8feaSLen Brown } 1249103a8feaSLen Brown 1250103a8feaSLen Brown 1251103a8feaSLen Brown /* 1252103a8feaSLen Brown * in /dev/cpu/ return success for names that are numbers 1253103a8feaSLen Brown * ie. filter out ".", "..", "microcode". 1254103a8feaSLen Brown */ 1255103a8feaSLen Brown int dir_filter(const struct dirent *dirp) 1256103a8feaSLen Brown { 1257103a8feaSLen Brown if (isdigit(dirp->d_name[0])) 1258103a8feaSLen Brown return 1; 1259103a8feaSLen Brown else 1260103a8feaSLen Brown return 0; 1261103a8feaSLen Brown } 1262103a8feaSLen Brown 1263103a8feaSLen Brown int open_dev_cpu_msr(int dummy1) 1264103a8feaSLen Brown { 1265103a8feaSLen Brown return 0; 1266103a8feaSLen Brown } 1267103a8feaSLen Brown 1268c98d5d94SLen Brown void topology_probe() 1269c98d5d94SLen Brown { 1270c98d5d94SLen Brown int i; 1271c98d5d94SLen Brown int max_core_id = 0; 1272c98d5d94SLen Brown int max_package_id = 0; 1273c98d5d94SLen Brown int max_siblings = 0; 1274c98d5d94SLen Brown struct cpu_topology { 1275c98d5d94SLen Brown int core_id; 1276c98d5d94SLen Brown int physical_package_id; 1277c98d5d94SLen Brown } *cpus; 1278c98d5d94SLen Brown 1279c98d5d94SLen Brown /* Initialize num_cpus, max_cpu_num */ 1280c98d5d94SLen Brown topo.num_cpus = 0; 1281c98d5d94SLen Brown topo.max_cpu_num = 0; 1282c98d5d94SLen Brown for_all_proc_cpus(count_cpus); 1283c98d5d94SLen Brown if (!summary_only && topo.num_cpus > 1) 1284c98d5d94SLen Brown show_cpu = 1; 1285c98d5d94SLen Brown 1286c98d5d94SLen Brown if (verbose > 1) 1287c98d5d94SLen Brown fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num); 1288c98d5d94SLen Brown 1289c98d5d94SLen Brown cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology)); 1290c98d5d94SLen Brown if (cpus == NULL) { 1291c98d5d94SLen Brown perror("calloc cpus"); 1292c98d5d94SLen Brown exit(1); 1293c98d5d94SLen Brown } 1294c98d5d94SLen Brown 1295c98d5d94SLen Brown /* 1296c98d5d94SLen Brown * Allocate and initialize cpu_present_set 1297c98d5d94SLen Brown */ 1298c98d5d94SLen Brown cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1)); 1299c98d5d94SLen Brown if (cpu_present_set == NULL) { 1300c98d5d94SLen Brown perror("CPU_ALLOC"); 1301c98d5d94SLen Brown exit(3); 1302c98d5d94SLen Brown } 1303c98d5d94SLen Brown cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1)); 1304c98d5d94SLen Brown CPU_ZERO_S(cpu_present_setsize, cpu_present_set); 1305c98d5d94SLen Brown for_all_proc_cpus(mark_cpu_present); 1306c98d5d94SLen Brown 1307c98d5d94SLen Brown /* 1308c98d5d94SLen Brown * Allocate and initialize cpu_affinity_set 1309c98d5d94SLen Brown */ 1310c98d5d94SLen Brown cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1)); 1311c98d5d94SLen Brown if (cpu_affinity_set == NULL) { 1312c98d5d94SLen Brown perror("CPU_ALLOC"); 1313c98d5d94SLen Brown exit(3); 1314c98d5d94SLen Brown } 1315c98d5d94SLen Brown cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1)); 1316c98d5d94SLen Brown CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set); 1317c98d5d94SLen Brown 1318c98d5d94SLen Brown 1319c98d5d94SLen Brown /* 1320c98d5d94SLen Brown * For online cpus 1321c98d5d94SLen Brown * find max_core_id, max_package_id 1322c98d5d94SLen Brown */ 1323c98d5d94SLen Brown for (i = 0; i <= topo.max_cpu_num; ++i) { 1324c98d5d94SLen Brown int siblings; 1325c98d5d94SLen Brown 1326c98d5d94SLen Brown if (cpu_is_not_present(i)) { 1327c98d5d94SLen Brown if (verbose > 1) 1328c98d5d94SLen Brown fprintf(stderr, "cpu%d NOT PRESENT\n", i); 1329c98d5d94SLen Brown continue; 1330c98d5d94SLen Brown } 1331c98d5d94SLen Brown cpus[i].core_id = get_core_id(i); 1332c98d5d94SLen Brown if (cpus[i].core_id > max_core_id) 1333c98d5d94SLen Brown max_core_id = cpus[i].core_id; 1334c98d5d94SLen Brown 1335c98d5d94SLen Brown cpus[i].physical_package_id = get_physical_package_id(i); 1336c98d5d94SLen Brown if (cpus[i].physical_package_id > max_package_id) 1337c98d5d94SLen Brown max_package_id = cpus[i].physical_package_id; 1338c98d5d94SLen Brown 1339c98d5d94SLen Brown siblings = get_num_ht_siblings(i); 1340c98d5d94SLen Brown if (siblings > max_siblings) 1341c98d5d94SLen Brown max_siblings = siblings; 1342c98d5d94SLen Brown if (verbose > 1) 1343c98d5d94SLen Brown fprintf(stderr, "cpu %d pkg %d core %d\n", 1344c98d5d94SLen Brown i, cpus[i].physical_package_id, cpus[i].core_id); 1345c98d5d94SLen Brown } 1346c98d5d94SLen Brown topo.num_cores_per_pkg = max_core_id + 1; 1347c98d5d94SLen Brown if (verbose > 1) 1348c98d5d94SLen Brown fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n", 1349c98d5d94SLen Brown max_core_id, topo.num_cores_per_pkg); 1350c98d5d94SLen Brown if (!summary_only && topo.num_cores_per_pkg > 1) 1351c98d5d94SLen Brown show_core = 1; 1352c98d5d94SLen Brown 1353c98d5d94SLen Brown topo.num_packages = max_package_id + 1; 1354c98d5d94SLen Brown if (verbose > 1) 1355c98d5d94SLen Brown fprintf(stderr, "max_package_id %d, sizing for %d packages\n", 1356c98d5d94SLen Brown max_package_id, topo.num_packages); 1357c98d5d94SLen Brown if (!summary_only && topo.num_packages > 1) 1358c98d5d94SLen Brown show_pkg = 1; 1359c98d5d94SLen Brown 1360c98d5d94SLen Brown topo.num_threads_per_core = max_siblings; 1361c98d5d94SLen Brown if (verbose > 1) 1362c98d5d94SLen Brown fprintf(stderr, "max_siblings %d\n", max_siblings); 1363c98d5d94SLen Brown 1364c98d5d94SLen Brown free(cpus); 1365c98d5d94SLen Brown } 1366c98d5d94SLen Brown 1367c98d5d94SLen Brown void 1368c98d5d94SLen Brown allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p) 1369c98d5d94SLen Brown { 1370c98d5d94SLen Brown int i; 1371c98d5d94SLen Brown 1372c98d5d94SLen Brown *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg * 1373c98d5d94SLen Brown topo.num_packages, sizeof(struct thread_data)); 1374c98d5d94SLen Brown if (*t == NULL) 1375c98d5d94SLen Brown goto error; 1376c98d5d94SLen Brown 1377c98d5d94SLen Brown for (i = 0; i < topo.num_threads_per_core * 1378c98d5d94SLen Brown topo.num_cores_per_pkg * topo.num_packages; i++) 1379c98d5d94SLen Brown (*t)[i].cpu_id = -1; 1380c98d5d94SLen Brown 1381c98d5d94SLen Brown *c = calloc(topo.num_cores_per_pkg * topo.num_packages, 1382c98d5d94SLen Brown sizeof(struct core_data)); 1383c98d5d94SLen Brown if (*c == NULL) 1384c98d5d94SLen Brown goto error; 1385c98d5d94SLen Brown 1386c98d5d94SLen Brown for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++) 1387c98d5d94SLen Brown (*c)[i].core_id = -1; 1388c98d5d94SLen Brown 1389c98d5d94SLen Brown *p = calloc(topo.num_packages, sizeof(struct pkg_data)); 1390c98d5d94SLen Brown if (*p == NULL) 1391c98d5d94SLen Brown goto error; 1392c98d5d94SLen Brown 1393c98d5d94SLen Brown for (i = 0; i < topo.num_packages; i++) 1394c98d5d94SLen Brown (*p)[i].package_id = i; 1395c98d5d94SLen Brown 1396c98d5d94SLen Brown return; 1397c98d5d94SLen Brown error: 1398c98d5d94SLen Brown perror("calloc counters"); 1399c98d5d94SLen Brown exit(1); 1400c98d5d94SLen Brown } 1401c98d5d94SLen Brown /* 1402c98d5d94SLen Brown * init_counter() 1403c98d5d94SLen Brown * 1404c98d5d94SLen Brown * set cpu_id, core_num, pkg_num 1405c98d5d94SLen Brown * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE 1406c98d5d94SLen Brown * 1407c98d5d94SLen Brown * increment topo.num_cores when 1st core in pkg seen 1408c98d5d94SLen Brown */ 1409c98d5d94SLen Brown void init_counter(struct thread_data *thread_base, struct core_data *core_base, 1410c98d5d94SLen Brown struct pkg_data *pkg_base, int thread_num, int core_num, 1411c98d5d94SLen Brown int pkg_num, int cpu_id) 1412c98d5d94SLen Brown { 1413c98d5d94SLen Brown struct thread_data *t; 1414c98d5d94SLen Brown struct core_data *c; 1415c98d5d94SLen Brown struct pkg_data *p; 1416c98d5d94SLen Brown 1417c98d5d94SLen Brown t = GET_THREAD(thread_base, thread_num, core_num, pkg_num); 1418c98d5d94SLen Brown c = GET_CORE(core_base, core_num, pkg_num); 1419c98d5d94SLen Brown p = GET_PKG(pkg_base, pkg_num); 1420c98d5d94SLen Brown 1421c98d5d94SLen Brown t->cpu_id = cpu_id; 1422c98d5d94SLen Brown if (thread_num == 0) { 1423c98d5d94SLen Brown t->flags |= CPU_IS_FIRST_THREAD_IN_CORE; 1424c98d5d94SLen Brown if (cpu_is_first_core_in_package(cpu_id)) 1425c98d5d94SLen Brown t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE; 1426c98d5d94SLen Brown } 1427c98d5d94SLen Brown 1428c98d5d94SLen Brown c->core_id = core_num; 1429c98d5d94SLen Brown p->package_id = pkg_num; 1430c98d5d94SLen Brown } 1431c98d5d94SLen Brown 1432c98d5d94SLen Brown 1433c98d5d94SLen Brown int initialize_counters(int cpu_id) 1434c98d5d94SLen Brown { 1435c98d5d94SLen Brown int my_thread_id, my_core_id, my_package_id; 1436c98d5d94SLen Brown 1437c98d5d94SLen Brown my_package_id = get_physical_package_id(cpu_id); 1438c98d5d94SLen Brown my_core_id = get_core_id(cpu_id); 1439c98d5d94SLen Brown 1440c98d5d94SLen Brown if (cpu_is_first_sibling_in_core(cpu_id)) { 1441c98d5d94SLen Brown my_thread_id = 0; 1442c98d5d94SLen Brown topo.num_cores++; 1443c98d5d94SLen Brown } else { 1444c98d5d94SLen Brown my_thread_id = 1; 1445c98d5d94SLen Brown } 1446c98d5d94SLen Brown 1447c98d5d94SLen Brown init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id); 1448c98d5d94SLen Brown init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id); 1449c98d5d94SLen Brown return 0; 1450c98d5d94SLen Brown } 1451c98d5d94SLen Brown 1452c98d5d94SLen Brown void allocate_output_buffer() 1453c98d5d94SLen Brown { 1454c98d5d94SLen Brown output_buffer = calloc(1, (1 + topo.num_cpus) * 128); 1455c98d5d94SLen Brown outp = output_buffer; 1456c98d5d94SLen Brown if (outp == NULL) { 1457c98d5d94SLen Brown perror("calloc"); 1458c98d5d94SLen Brown exit(-1); 1459c98d5d94SLen Brown } 1460c98d5d94SLen Brown } 1461c98d5d94SLen Brown 1462c98d5d94SLen Brown void setup_all_buffers(void) 1463c98d5d94SLen Brown { 1464c98d5d94SLen Brown topology_probe(); 1465c98d5d94SLen Brown allocate_counters(&thread_even, &core_even, &package_even); 1466c98d5d94SLen Brown allocate_counters(&thread_odd, &core_odd, &package_odd); 1467c98d5d94SLen Brown allocate_output_buffer(); 1468c98d5d94SLen Brown for_all_proc_cpus(initialize_counters); 1469c98d5d94SLen Brown } 1470103a8feaSLen Brown void turbostat_init() 1471103a8feaSLen Brown { 1472103a8feaSLen Brown check_cpuid(); 1473103a8feaSLen Brown 1474103a8feaSLen Brown check_dev_msr(); 1475103a8feaSLen Brown check_super_user(); 1476103a8feaSLen Brown 1477c98d5d94SLen Brown setup_all_buffers(); 1478103a8feaSLen Brown 1479103a8feaSLen Brown if (verbose) 1480c98d5d94SLen Brown print_verbose_header(); 1481103a8feaSLen Brown } 1482103a8feaSLen Brown 1483103a8feaSLen Brown int fork_it(char **argv) 1484103a8feaSLen Brown { 1485103a8feaSLen Brown pid_t child_pid; 1486d15cf7c1SLen Brown 1487c98d5d94SLen Brown for_all_cpus(get_counters, EVEN_COUNTERS); 1488d15cf7c1SLen Brown /* clear affinity side-effect of get_counters() */ 1489d15cf7c1SLen Brown sched_setaffinity(0, cpu_present_setsize, cpu_present_set); 1490103a8feaSLen Brown gettimeofday(&tv_even, (struct timezone *)NULL); 1491103a8feaSLen Brown 1492103a8feaSLen Brown child_pid = fork(); 1493103a8feaSLen Brown if (!child_pid) { 1494103a8feaSLen Brown /* child */ 1495103a8feaSLen Brown execvp(argv[0], argv); 1496103a8feaSLen Brown } else { 1497103a8feaSLen Brown int status; 1498103a8feaSLen Brown 1499103a8feaSLen Brown /* parent */ 1500103a8feaSLen Brown if (child_pid == -1) { 1501103a8feaSLen Brown perror("fork"); 1502103a8feaSLen Brown exit(1); 1503103a8feaSLen Brown } 1504103a8feaSLen Brown 1505103a8feaSLen Brown signal(SIGINT, SIG_IGN); 1506103a8feaSLen Brown signal(SIGQUIT, SIG_IGN); 1507103a8feaSLen Brown if (waitpid(child_pid, &status, 0) == -1) { 1508103a8feaSLen Brown perror("wait"); 1509103a8feaSLen Brown exit(1); 1510103a8feaSLen Brown } 1511103a8feaSLen Brown } 1512c98d5d94SLen Brown /* 1513c98d5d94SLen Brown * n.b. fork_it() does not check for errors from for_all_cpus() 1514c98d5d94SLen Brown * because re-starting is problematic when forking 1515c98d5d94SLen Brown */ 1516c98d5d94SLen Brown for_all_cpus(get_counters, ODD_COUNTERS); 1517103a8feaSLen Brown gettimeofday(&tv_odd, (struct timezone *)NULL); 1518103a8feaSLen Brown timersub(&tv_odd, &tv_even, &tv_delta); 1519c98d5d94SLen Brown for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS); 1520c98d5d94SLen Brown compute_average(EVEN_COUNTERS); 1521c98d5d94SLen Brown format_all_counters(EVEN_COUNTERS); 1522c98d5d94SLen Brown flush_stderr(); 1523103a8feaSLen Brown 15246eab04a8SJustin P. Mattock fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0); 1525103a8feaSLen Brown 1526103a8feaSLen Brown return 0; 1527103a8feaSLen Brown } 1528103a8feaSLen Brown 1529103a8feaSLen Brown void cmdline(int argc, char **argv) 1530103a8feaSLen Brown { 1531103a8feaSLen Brown int opt; 1532103a8feaSLen Brown 1533103a8feaSLen Brown progname = argv[0]; 1534103a8feaSLen Brown 1535c98d5d94SLen Brown while ((opt = getopt(argc, argv, "+cpsvi:M:")) != -1) { 1536103a8feaSLen Brown switch (opt) { 1537c98d5d94SLen Brown case 'c': 1538c98d5d94SLen Brown show_core_only++; 1539c98d5d94SLen Brown break; 1540c98d5d94SLen Brown case 'p': 1541c98d5d94SLen Brown show_pkg_only++; 1542c98d5d94SLen Brown break; 1543e23da037SLen Brown case 's': 1544e23da037SLen Brown summary_only++; 1545e23da037SLen Brown break; 1546103a8feaSLen Brown case 'v': 1547103a8feaSLen Brown verbose++; 1548103a8feaSLen Brown break; 1549103a8feaSLen Brown case 'i': 1550103a8feaSLen Brown interval_sec = atoi(optarg); 1551103a8feaSLen Brown break; 1552103a8feaSLen Brown case 'M': 1553103a8feaSLen Brown sscanf(optarg, "%x", &extra_msr_offset); 1554103a8feaSLen Brown if (verbose > 1) 1555103a8feaSLen Brown fprintf(stderr, "MSR 0x%X\n", extra_msr_offset); 1556103a8feaSLen Brown break; 1557103a8feaSLen Brown default: 1558103a8feaSLen Brown usage(); 1559103a8feaSLen Brown } 1560103a8feaSLen Brown } 1561103a8feaSLen Brown } 1562103a8feaSLen Brown 1563103a8feaSLen Brown int main(int argc, char **argv) 1564103a8feaSLen Brown { 1565103a8feaSLen Brown cmdline(argc, argv); 1566103a8feaSLen Brown 1567103a8feaSLen Brown if (verbose > 1) 1568c98d5d94SLen Brown fprintf(stderr, "turbostat v2.0 May 16, 2012" 1569103a8feaSLen Brown " - Len Brown <lenb@kernel.org>\n"); 1570103a8feaSLen Brown 1571103a8feaSLen Brown turbostat_init(); 1572103a8feaSLen Brown 1573103a8feaSLen Brown /* 1574103a8feaSLen Brown * if any params left, it must be a command to fork 1575103a8feaSLen Brown */ 1576103a8feaSLen Brown if (argc - optind) 1577103a8feaSLen Brown return fork_it(argv + optind); 1578103a8feaSLen Brown else 1579103a8feaSLen Brown turbostat_loop(); 1580103a8feaSLen Brown 1581103a8feaSLen Brown return 0; 1582103a8feaSLen Brown } 1583