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_TSC 0x10 39103a8feaSLen Brown #define MSR_NEHALEM_PLATFORM_INFO 0xCE 40103a8feaSLen Brown #define MSR_NEHALEM_TURBO_RATIO_LIMIT 0x1AD 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; 65103a8feaSLen Brown unsigned int extra_msr_offset; 66103a8feaSLen Brown double bclk; 67103a8feaSLen Brown unsigned int show_pkg; 68103a8feaSLen Brown unsigned int show_core; 69103a8feaSLen Brown unsigned int show_cpu; 70c98d5d94SLen Brown unsigned int show_pkg_only; 71c98d5d94SLen Brown unsigned int show_core_only; 72c98d5d94SLen Brown char *output_buffer, *outp; 73103a8feaSLen Brown 74103a8feaSLen Brown int aperf_mperf_unstable; 75103a8feaSLen Brown int backwards_count; 76103a8feaSLen Brown char *progname; 77103a8feaSLen Brown 78c98d5d94SLen Brown cpu_set_t *cpu_present_set, *cpu_affinity_set; 79c98d5d94SLen Brown size_t cpu_present_setsize, cpu_affinity_setsize; 80c98d5d94SLen Brown 81c98d5d94SLen Brown struct thread_data { 82c98d5d94SLen Brown unsigned long long tsc; 83c98d5d94SLen Brown unsigned long long aperf; 84c98d5d94SLen Brown unsigned long long mperf; 85c98d5d94SLen Brown unsigned long long c1; /* derived */ 86c98d5d94SLen Brown unsigned long long extra_msr; 87c98d5d94SLen Brown unsigned int cpu_id; 88c98d5d94SLen Brown unsigned int flags; 89c98d5d94SLen Brown #define CPU_IS_FIRST_THREAD_IN_CORE 0x2 90c98d5d94SLen Brown #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4 91c98d5d94SLen Brown } *thread_even, *thread_odd; 92c98d5d94SLen Brown 93c98d5d94SLen Brown struct core_data { 94c98d5d94SLen Brown unsigned long long c3; 95c98d5d94SLen Brown unsigned long long c6; 96c98d5d94SLen Brown unsigned long long c7; 97c98d5d94SLen Brown unsigned int core_id; 98c98d5d94SLen Brown } *core_even, *core_odd; 99c98d5d94SLen Brown 100c98d5d94SLen Brown struct pkg_data { 101c98d5d94SLen Brown unsigned long long pc2; 102c98d5d94SLen Brown unsigned long long pc3; 103c98d5d94SLen Brown unsigned long long pc6; 104c98d5d94SLen Brown unsigned long long pc7; 105c98d5d94SLen Brown unsigned int package_id; 106c98d5d94SLen Brown } *package_even, *package_odd; 107c98d5d94SLen Brown 108c98d5d94SLen Brown #define ODD_COUNTERS thread_odd, core_odd, package_odd 109c98d5d94SLen Brown #define EVEN_COUNTERS thread_even, core_even, package_even 110c98d5d94SLen Brown 111c98d5d94SLen Brown #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \ 112c98d5d94SLen Brown (thread_base + (pkg_no) * topo.num_cores_per_pkg * \ 113c98d5d94SLen Brown topo.num_threads_per_core + \ 114c98d5d94SLen Brown (core_no) * topo.num_threads_per_core + (thread_no)) 115c98d5d94SLen Brown #define GET_CORE(core_base, core_no, pkg_no) \ 116c98d5d94SLen Brown (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no)) 117c98d5d94SLen Brown #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no) 118c98d5d94SLen Brown 119c98d5d94SLen Brown struct system_summary { 120c98d5d94SLen Brown struct thread_data threads; 121c98d5d94SLen Brown struct core_data cores; 122c98d5d94SLen Brown struct pkg_data packages; 123c98d5d94SLen Brown } sum, average; 124c98d5d94SLen Brown 125c98d5d94SLen Brown 126c98d5d94SLen Brown struct topo_params { 127c98d5d94SLen Brown int num_packages; 128103a8feaSLen Brown int num_cpus; 129c98d5d94SLen Brown int num_cores; 130c98d5d94SLen Brown int max_cpu_num; 131c98d5d94SLen Brown int num_cores_per_pkg; 132c98d5d94SLen Brown int num_threads_per_core; 133c98d5d94SLen Brown } topo; 134103a8feaSLen Brown 135c98d5d94SLen Brown struct timeval tv_even, tv_odd, tv_delta; 136103a8feaSLen Brown 137c98d5d94SLen Brown void setup_all_buffers(void); 138103a8feaSLen Brown 139c98d5d94SLen Brown int cpu_is_not_present(int cpu) 140d15cf7c1SLen Brown { 141c98d5d94SLen Brown return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set); 142c98d5d94SLen Brown } 143c98d5d94SLen Brown /* 144c98d5d94SLen Brown * run func(thread, core, package) in topology order 145c98d5d94SLen Brown * skip non-present cpus 146c98d5d94SLen Brown */ 147c98d5d94SLen Brown 148c98d5d94SLen Brown int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *), 149c98d5d94SLen Brown struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base) 150c98d5d94SLen Brown { 151c98d5d94SLen Brown int retval, pkg_no, core_no, thread_no; 152c98d5d94SLen Brown 153c98d5d94SLen Brown for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) { 154c98d5d94SLen Brown for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) { 155c98d5d94SLen Brown for (thread_no = 0; thread_no < 156c98d5d94SLen Brown topo.num_threads_per_core; ++thread_no) { 157c98d5d94SLen Brown struct thread_data *t; 158c98d5d94SLen Brown struct core_data *c; 159c98d5d94SLen Brown struct pkg_data *p; 160c98d5d94SLen Brown 161c98d5d94SLen Brown t = GET_THREAD(thread_base, thread_no, core_no, pkg_no); 162c98d5d94SLen Brown 163c98d5d94SLen Brown if (cpu_is_not_present(t->cpu_id)) 164c98d5d94SLen Brown continue; 165c98d5d94SLen Brown 166c98d5d94SLen Brown c = GET_CORE(core_base, core_no, pkg_no); 167c98d5d94SLen Brown p = GET_PKG(pkg_base, pkg_no); 168c98d5d94SLen Brown 169c98d5d94SLen Brown retval = func(t, c, p); 170c98d5d94SLen Brown if (retval) 171c98d5d94SLen Brown return retval; 172c98d5d94SLen Brown } 173c98d5d94SLen Brown } 174c98d5d94SLen Brown } 175d15cf7c1SLen Brown return 0; 176d15cf7c1SLen Brown } 177d15cf7c1SLen Brown 17888c3281fSLen Brown int cpu_migrate(int cpu) 17988c3281fSLen Brown { 180c98d5d94SLen Brown CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set); 181c98d5d94SLen Brown CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set); 182c98d5d94SLen Brown if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) 18388c3281fSLen Brown return -1; 18488c3281fSLen Brown else 18588c3281fSLen Brown return 0; 18688c3281fSLen Brown } 18788c3281fSLen Brown 18815aaa346SLen Brown int get_msr(int cpu, off_t offset, unsigned long long *msr) 189103a8feaSLen Brown { 190103a8feaSLen Brown ssize_t retval; 191103a8feaSLen Brown char pathname[32]; 192103a8feaSLen Brown int fd; 193103a8feaSLen Brown 194103a8feaSLen Brown sprintf(pathname, "/dev/cpu/%d/msr", cpu); 195103a8feaSLen Brown fd = open(pathname, O_RDONLY); 19615aaa346SLen Brown if (fd < 0) 19715aaa346SLen Brown return -1; 198103a8feaSLen Brown 19915aaa346SLen Brown retval = pread(fd, msr, sizeof *msr, offset); 200103a8feaSLen Brown close(fd); 20115aaa346SLen Brown 20215aaa346SLen Brown if (retval != sizeof *msr) 20315aaa346SLen Brown return -1; 20415aaa346SLen Brown 20515aaa346SLen Brown return 0; 206103a8feaSLen Brown } 207103a8feaSLen Brown 208a829eb4dSLen Brown void print_header(void) 209103a8feaSLen Brown { 210103a8feaSLen Brown if (show_pkg) 211c98d5d94SLen Brown outp += sprintf(outp, "pk"); 212e23da037SLen Brown if (show_pkg) 213c98d5d94SLen Brown outp += sprintf(outp, " "); 214103a8feaSLen Brown if (show_core) 215c98d5d94SLen Brown outp += sprintf(outp, "cor"); 216103a8feaSLen Brown if (show_cpu) 217c98d5d94SLen Brown outp += sprintf(outp, " CPU"); 218e23da037SLen Brown if (show_pkg || show_core || show_cpu) 219c98d5d94SLen Brown outp += sprintf(outp, " "); 220103a8feaSLen Brown if (do_nhm_cstates) 221c98d5d94SLen Brown outp += sprintf(outp, " %%c0"); 222103a8feaSLen Brown if (has_aperf) 223c98d5d94SLen Brown outp += sprintf(outp, " GHz"); 224c98d5d94SLen Brown outp += sprintf(outp, " TSC"); 225103a8feaSLen Brown if (do_nhm_cstates) 226c98d5d94SLen Brown outp += sprintf(outp, " %%c1"); 227103a8feaSLen Brown if (do_nhm_cstates) 228c98d5d94SLen Brown outp += sprintf(outp, " %%c3"); 229103a8feaSLen Brown if (do_nhm_cstates) 230c98d5d94SLen Brown outp += sprintf(outp, " %%c6"); 231103a8feaSLen Brown if (do_snb_cstates) 232c98d5d94SLen Brown outp += sprintf(outp, " %%c7"); 233103a8feaSLen Brown if (do_snb_cstates) 234c98d5d94SLen Brown outp += sprintf(outp, " %%pc2"); 235103a8feaSLen Brown if (do_nhm_cstates) 236c98d5d94SLen Brown outp += sprintf(outp, " %%pc3"); 237103a8feaSLen Brown if (do_nhm_cstates) 238c98d5d94SLen Brown outp += sprintf(outp, " %%pc6"); 239103a8feaSLen Brown if (do_snb_cstates) 240c98d5d94SLen Brown outp += sprintf(outp, " %%pc7"); 241103a8feaSLen Brown if (extra_msr_offset) 242c98d5d94SLen Brown outp += sprintf(outp, " MSR 0x%x ", extra_msr_offset); 243103a8feaSLen Brown 244c98d5d94SLen Brown outp += sprintf(outp, "\n"); 245103a8feaSLen Brown } 246103a8feaSLen Brown 247c98d5d94SLen Brown int dump_counters(struct thread_data *t, struct core_data *c, 248c98d5d94SLen Brown struct pkg_data *p) 249103a8feaSLen Brown { 250c98d5d94SLen Brown fprintf(stderr, "t %p, c %p, p %p\n", t, c, p); 251c98d5d94SLen Brown 252c98d5d94SLen Brown if (t) { 253c98d5d94SLen Brown fprintf(stderr, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags); 254c98d5d94SLen Brown fprintf(stderr, "TSC: %016llX\n", t->tsc); 255c98d5d94SLen Brown fprintf(stderr, "aperf: %016llX\n", t->aperf); 256c98d5d94SLen Brown fprintf(stderr, "mperf: %016llX\n", t->mperf); 257c98d5d94SLen Brown fprintf(stderr, "c1: %016llX\n", t->c1); 258c98d5d94SLen Brown fprintf(stderr, "msr0x%x: %016llX\n", 259c98d5d94SLen Brown extra_msr_offset, t->extra_msr); 260103a8feaSLen Brown } 261103a8feaSLen Brown 262c98d5d94SLen Brown if (c) { 263c98d5d94SLen Brown fprintf(stderr, "core: %d\n", c->core_id); 264c98d5d94SLen Brown fprintf(stderr, "c3: %016llX\n", c->c3); 265c98d5d94SLen Brown fprintf(stderr, "c6: %016llX\n", c->c6); 266c98d5d94SLen Brown fprintf(stderr, "c7: %016llX\n", c->c7); 267c98d5d94SLen Brown } 268103a8feaSLen Brown 269c98d5d94SLen Brown if (p) { 270c98d5d94SLen Brown fprintf(stderr, "package: %d\n", p->package_id); 271c98d5d94SLen Brown fprintf(stderr, "pc2: %016llX\n", p->pc2); 272c98d5d94SLen Brown fprintf(stderr, "pc3: %016llX\n", p->pc3); 273c98d5d94SLen Brown fprintf(stderr, "pc6: %016llX\n", p->pc6); 274c98d5d94SLen Brown fprintf(stderr, "pc7: %016llX\n", p->pc7); 275c98d5d94SLen Brown } 276c98d5d94SLen Brown return 0; 277103a8feaSLen Brown } 278103a8feaSLen Brown 279e23da037SLen Brown /* 280e23da037SLen Brown * column formatting convention & formats 281e23da037SLen Brown * package: "pk" 2 columns %2d 282e23da037SLen Brown * core: "cor" 3 columns %3d 283e23da037SLen Brown * CPU: "CPU" 3 columns %3d 284e23da037SLen Brown * GHz: "GHz" 3 columns %3.2 285e23da037SLen Brown * TSC: "TSC" 3 columns %3.2 286e23da037SLen Brown * percentage " %pc3" %6.2 287e23da037SLen Brown */ 288c98d5d94SLen Brown int format_counters(struct thread_data *t, struct core_data *c, 289c98d5d94SLen Brown struct pkg_data *p) 290103a8feaSLen Brown { 291103a8feaSLen Brown double interval_float; 292103a8feaSLen Brown 293c98d5d94SLen Brown /* if showing only 1st thread in core and this isn't one, bail out */ 294c98d5d94SLen Brown if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 295c98d5d94SLen Brown return 0; 296c98d5d94SLen Brown 297c98d5d94SLen Brown /* if showing only 1st thread in pkg and this isn't one, bail out */ 298c98d5d94SLen Brown if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 299c98d5d94SLen Brown return 0; 300c98d5d94SLen Brown 301103a8feaSLen Brown interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0; 302103a8feaSLen Brown 303c98d5d94SLen Brown /* topo columns, print blanks on 1st (average) line */ 304c98d5d94SLen Brown if (t == &average.threads) { 305103a8feaSLen Brown if (show_pkg) 306c98d5d94SLen Brown outp += sprintf(outp, " "); 307e23da037SLen Brown if (show_pkg && show_core) 308c98d5d94SLen Brown outp += sprintf(outp, " "); 309103a8feaSLen Brown if (show_core) 310c98d5d94SLen Brown outp += sprintf(outp, " "); 311103a8feaSLen Brown if (show_cpu) 312c98d5d94SLen Brown outp += sprintf(outp, " " " "); 313103a8feaSLen Brown } else { 314c98d5d94SLen Brown if (show_pkg) { 315c98d5d94SLen Brown if (p) 316c98d5d94SLen Brown outp += sprintf(outp, "%2d", p->package_id); 317c98d5d94SLen Brown else 318c98d5d94SLen Brown outp += sprintf(outp, " "); 319c98d5d94SLen Brown } 320e23da037SLen Brown if (show_pkg && show_core) 321c98d5d94SLen Brown outp += sprintf(outp, " "); 322c98d5d94SLen Brown if (show_core) { 323c98d5d94SLen Brown if (c) 324c98d5d94SLen Brown outp += sprintf(outp, "%3d", c->core_id); 325c98d5d94SLen Brown else 326c98d5d94SLen Brown outp += sprintf(outp, " "); 327c98d5d94SLen Brown } 328103a8feaSLen Brown if (show_cpu) 329c98d5d94SLen Brown outp += sprintf(outp, " %3d", t->cpu_id); 330103a8feaSLen Brown } 331103a8feaSLen Brown 332103a8feaSLen Brown /* %c0 */ 333103a8feaSLen Brown if (do_nhm_cstates) { 334e23da037SLen Brown if (show_pkg || show_core || show_cpu) 335c98d5d94SLen Brown outp += sprintf(outp, " "); 336103a8feaSLen Brown if (!skip_c0) 337c98d5d94SLen Brown outp += sprintf(outp, "%6.2f", 100.0 * t->mperf/t->tsc); 338103a8feaSLen Brown else 339c98d5d94SLen Brown outp += sprintf(outp, " ****"); 340103a8feaSLen Brown } 341103a8feaSLen Brown 342103a8feaSLen Brown /* GHz */ 343103a8feaSLen Brown if (has_aperf) { 344103a8feaSLen Brown if (!aperf_mperf_unstable) { 345c98d5d94SLen Brown outp += sprintf(outp, " %3.2f", 346c98d5d94SLen Brown 1.0 * t->tsc / units * t->aperf / 347c98d5d94SLen Brown t->mperf / interval_float); 348103a8feaSLen Brown } else { 349c98d5d94SLen Brown if (t->aperf > t->tsc || t->mperf > t->tsc) { 350c98d5d94SLen Brown outp += sprintf(outp, " ***"); 351103a8feaSLen Brown } else { 352c98d5d94SLen Brown outp += sprintf(outp, "%3.1f*", 353c98d5d94SLen Brown 1.0 * t->tsc / 354c98d5d94SLen Brown units * t->aperf / 355c98d5d94SLen Brown t->mperf / interval_float); 356103a8feaSLen Brown } 357103a8feaSLen Brown } 358103a8feaSLen Brown } 359103a8feaSLen Brown 360103a8feaSLen Brown /* TSC */ 361c98d5d94SLen Brown outp += sprintf(outp, "%5.2f", 1.0 * t->tsc/units/interval_float); 362103a8feaSLen Brown 363103a8feaSLen Brown if (do_nhm_cstates) { 364103a8feaSLen Brown if (!skip_c1) 365c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * t->c1/t->tsc); 366103a8feaSLen Brown else 367c98d5d94SLen Brown outp += sprintf(outp, " ****"); 368103a8feaSLen Brown } 369c98d5d94SLen Brown 370c98d5d94SLen Brown /* print per-core data only for 1st thread in core */ 371c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 372c98d5d94SLen Brown goto done; 373c98d5d94SLen Brown 374103a8feaSLen Brown if (do_nhm_cstates) 375c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * c->c3/t->tsc); 376103a8feaSLen Brown if (do_nhm_cstates) 377c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * c->c6/t->tsc); 378103a8feaSLen Brown if (do_snb_cstates) 379c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * c->c7/t->tsc); 380c98d5d94SLen Brown 381c98d5d94SLen Brown /* print per-package data only for 1st core in package */ 382c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 383c98d5d94SLen Brown goto done; 384c98d5d94SLen Brown 385103a8feaSLen Brown if (do_snb_cstates) 386c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc2/t->tsc); 387103a8feaSLen Brown if (do_nhm_cstates) 388c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc3/t->tsc); 389103a8feaSLen Brown if (do_nhm_cstates) 390c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc6/t->tsc); 391103a8feaSLen Brown if (do_snb_cstates) 392c98d5d94SLen Brown outp += sprintf(outp, " %6.2f", 100.0 * p->pc7/t->tsc); 393c98d5d94SLen Brown done: 394103a8feaSLen Brown if (extra_msr_offset) 395c98d5d94SLen Brown outp += sprintf(outp, " 0x%016llx", t->extra_msr); 396c98d5d94SLen Brown outp += sprintf(outp, "\n"); 397c98d5d94SLen Brown 398c98d5d94SLen Brown return 0; 399103a8feaSLen Brown } 400103a8feaSLen Brown 401c98d5d94SLen Brown void flush_stdout() 402103a8feaSLen Brown { 403c98d5d94SLen Brown fputs(output_buffer, stdout); 404c98d5d94SLen Brown outp = output_buffer; 405c98d5d94SLen Brown } 406c98d5d94SLen Brown void flush_stderr() 407c98d5d94SLen Brown { 408c98d5d94SLen Brown fputs(output_buffer, stderr); 409c98d5d94SLen Brown outp = output_buffer; 410c98d5d94SLen Brown } 411c98d5d94SLen Brown void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 412c98d5d94SLen Brown { 413e23da037SLen Brown static int printed; 414103a8feaSLen Brown 415e23da037SLen Brown if (!printed || !summary_only) 416103a8feaSLen Brown print_header(); 417103a8feaSLen Brown 418c98d5d94SLen Brown if (topo.num_cpus > 1) 419c98d5d94SLen Brown format_counters(&average.threads, &average.cores, 420c98d5d94SLen Brown &average.packages); 421103a8feaSLen Brown 422e23da037SLen Brown printed = 1; 423e23da037SLen Brown 424e23da037SLen Brown if (summary_only) 425e23da037SLen Brown return; 426e23da037SLen Brown 427c98d5d94SLen Brown for_all_cpus(format_counters, t, c, p); 428103a8feaSLen Brown } 429103a8feaSLen Brown 430c98d5d94SLen Brown void 431c98d5d94SLen Brown delta_package(struct pkg_data *new, struct pkg_data *old) 432103a8feaSLen Brown { 433c98d5d94SLen Brown old->pc2 = new->pc2 - old->pc2; 434c98d5d94SLen Brown old->pc3 = new->pc3 - old->pc3; 435c98d5d94SLen Brown old->pc6 = new->pc6 - old->pc6; 436c98d5d94SLen Brown old->pc7 = new->pc7 - old->pc7; 437103a8feaSLen Brown } 438103a8feaSLen Brown 439c98d5d94SLen Brown void 440c98d5d94SLen Brown delta_core(struct core_data *new, struct core_data *old) 441c98d5d94SLen Brown { 442c98d5d94SLen Brown old->c3 = new->c3 - old->c3; 443c98d5d94SLen Brown old->c6 = new->c6 - old->c6; 444c98d5d94SLen Brown old->c7 = new->c7 - old->c7; 445103a8feaSLen Brown } 446c98d5d94SLen Brown 447*c3ae331dSLen Brown /* 448*c3ae331dSLen Brown * old = new - old 449*c3ae331dSLen Brown */ 450c98d5d94SLen Brown void 451c98d5d94SLen Brown delta_thread(struct thread_data *new, struct thread_data *old, 452c98d5d94SLen Brown struct core_data *core_delta) 453c98d5d94SLen Brown { 454c98d5d94SLen Brown old->tsc = new->tsc - old->tsc; 455c98d5d94SLen Brown 456103a8feaSLen Brown /* check for TSC < 1 Mcycles over interval */ 457c98d5d94SLen Brown if (old->tsc < (1000 * 1000)) { 458c98d5d94SLen Brown fprintf(stderr, "Insanely slow TSC rate, TSC stops in idle?\n"); 459c98d5d94SLen Brown fprintf(stderr, "You can disable all c-states by booting with \"idle=poll\"\n"); 460c98d5d94SLen Brown fprintf(stderr, "or just the deep ones with \"processor.max_cstate=1\"\n"); 461103a8feaSLen Brown exit(-3); 462103a8feaSLen Brown } 463103a8feaSLen Brown 464c98d5d94SLen Brown old->c1 = new->c1 - old->c1; 465c98d5d94SLen Brown 466c98d5d94SLen Brown if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) { 467c98d5d94SLen Brown old->aperf = new->aperf - old->aperf; 468c98d5d94SLen Brown old->mperf = new->mperf - old->mperf; 469c98d5d94SLen Brown } else { 470c98d5d94SLen Brown 471103a8feaSLen Brown if (!aperf_mperf_unstable) { 472103a8feaSLen Brown fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname); 473103a8feaSLen Brown fprintf(stderr, "* Frequency results do not cover entire interval *\n"); 474103a8feaSLen Brown fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n"); 475103a8feaSLen Brown 476103a8feaSLen Brown aperf_mperf_unstable = 1; 477103a8feaSLen Brown } 478103a8feaSLen Brown /* 479103a8feaSLen Brown * mperf delta is likely a huge "positive" number 480103a8feaSLen Brown * can not use it for calculating c0 time 481103a8feaSLen Brown */ 482103a8feaSLen Brown skip_c0 = 1; 483103a8feaSLen Brown skip_c1 = 1; 484103a8feaSLen Brown } 485103a8feaSLen Brown 486c98d5d94SLen Brown 487103a8feaSLen Brown /* 488*c3ae331dSLen Brown * As counter collection is not atomic, 489*c3ae331dSLen Brown * it is possible for mperf's non-halted cycles + idle states 490103a8feaSLen Brown * to exceed TSC's all cycles: show c1 = 0% in that case. 491103a8feaSLen Brown */ 492*c3ae331dSLen Brown if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc) 493c98d5d94SLen Brown old->c1 = 0; 494c98d5d94SLen Brown else { 495c98d5d94SLen Brown /* normal case, derive c1 */ 496c98d5d94SLen Brown old->c1 = old->tsc - old->mperf - core_delta->c3 497c98d5d94SLen Brown - core_delta->c6 - core_delta->c7; 498c98d5d94SLen Brown } 499*c3ae331dSLen Brown 500c98d5d94SLen Brown if (old->mperf == 0) { 501*c3ae331dSLen Brown if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id); 502c98d5d94SLen Brown old->mperf = 1; /* divide by 0 protection */ 503c98d5d94SLen Brown } 504103a8feaSLen Brown 505103a8feaSLen Brown /* 506103a8feaSLen Brown * for "extra msr", just copy the latest w/o subtracting 507103a8feaSLen Brown */ 508c98d5d94SLen Brown old->extra_msr = new->extra_msr; 509103a8feaSLen Brown } 510c98d5d94SLen Brown 511c98d5d94SLen Brown int delta_cpu(struct thread_data *t, struct core_data *c, 512c98d5d94SLen Brown struct pkg_data *p, struct thread_data *t2, 513c98d5d94SLen Brown struct core_data *c2, struct pkg_data *p2) 514c98d5d94SLen Brown { 515c98d5d94SLen Brown /* calculate core delta only for 1st thread in core */ 516c98d5d94SLen Brown if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE) 517c98d5d94SLen Brown delta_core(c, c2); 518c98d5d94SLen Brown 519c98d5d94SLen Brown /* always calculate thread delta */ 520c98d5d94SLen Brown delta_thread(t, t2, c2); /* c2 is core delta */ 521c98d5d94SLen Brown 522c98d5d94SLen Brown /* calculate package delta only for 1st core in package */ 523c98d5d94SLen Brown if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) 524c98d5d94SLen Brown delta_package(p, p2); 525c98d5d94SLen Brown 526103a8feaSLen Brown return 0; 527103a8feaSLen Brown } 528103a8feaSLen Brown 529c98d5d94SLen Brown void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 530103a8feaSLen Brown { 531c98d5d94SLen Brown t->tsc = 0; 532c98d5d94SLen Brown t->aperf = 0; 533c98d5d94SLen Brown t->mperf = 0; 534c98d5d94SLen Brown t->c1 = 0; 535103a8feaSLen Brown 536c98d5d94SLen Brown /* tells format_counters to dump all fields from this set */ 537c98d5d94SLen Brown t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE; 538c98d5d94SLen Brown 539c98d5d94SLen Brown c->c3 = 0; 540c98d5d94SLen Brown c->c6 = 0; 541c98d5d94SLen Brown c->c7 = 0; 542c98d5d94SLen Brown 543c98d5d94SLen Brown p->pc2 = 0; 544c98d5d94SLen Brown p->pc3 = 0; 545c98d5d94SLen Brown p->pc6 = 0; 546c98d5d94SLen Brown p->pc7 = 0; 547103a8feaSLen Brown } 548c98d5d94SLen Brown int sum_counters(struct thread_data *t, struct core_data *c, 549c98d5d94SLen Brown struct pkg_data *p) 550103a8feaSLen Brown { 551c98d5d94SLen Brown average.threads.tsc += t->tsc; 552c98d5d94SLen Brown average.threads.aperf += t->aperf; 553c98d5d94SLen Brown average.threads.mperf += t->mperf; 554c98d5d94SLen Brown average.threads.c1 += t->c1; 55515aaa346SLen Brown 556c98d5d94SLen Brown /* sum per-core values only for 1st thread in core */ 557c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 558c98d5d94SLen Brown return 0; 559c98d5d94SLen Brown 560c98d5d94SLen Brown average.cores.c3 += c->c3; 561c98d5d94SLen Brown average.cores.c6 += c->c6; 562c98d5d94SLen Brown average.cores.c7 += c->c7; 563c98d5d94SLen Brown 564c98d5d94SLen Brown /* sum per-pkg values only for 1st core in pkg */ 565c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 566c98d5d94SLen Brown return 0; 567c98d5d94SLen Brown 568c98d5d94SLen Brown average.packages.pc2 += p->pc2; 569c98d5d94SLen Brown average.packages.pc3 += p->pc3; 570c98d5d94SLen Brown average.packages.pc6 += p->pc6; 571c98d5d94SLen Brown average.packages.pc7 += p->pc7; 572c98d5d94SLen Brown 573c98d5d94SLen Brown return 0; 574c98d5d94SLen Brown } 575c98d5d94SLen Brown /* 576c98d5d94SLen Brown * sum the counters for all cpus in the system 577c98d5d94SLen Brown * compute the weighted average 578c98d5d94SLen Brown */ 579c98d5d94SLen Brown void compute_average(struct thread_data *t, struct core_data *c, 580c98d5d94SLen Brown struct pkg_data *p) 581c98d5d94SLen Brown { 582c98d5d94SLen Brown clear_counters(&average.threads, &average.cores, &average.packages); 583c98d5d94SLen Brown 584c98d5d94SLen Brown for_all_cpus(sum_counters, t, c, p); 585c98d5d94SLen Brown 586c98d5d94SLen Brown average.threads.tsc /= topo.num_cpus; 587c98d5d94SLen Brown average.threads.aperf /= topo.num_cpus; 588c98d5d94SLen Brown average.threads.mperf /= topo.num_cpus; 589c98d5d94SLen Brown average.threads.c1 /= topo.num_cpus; 590c98d5d94SLen Brown 591c98d5d94SLen Brown average.cores.c3 /= topo.num_cores; 592c98d5d94SLen Brown average.cores.c6 /= topo.num_cores; 593c98d5d94SLen Brown average.cores.c7 /= topo.num_cores; 594c98d5d94SLen Brown 595c98d5d94SLen Brown average.packages.pc2 /= topo.num_packages; 596c98d5d94SLen Brown average.packages.pc3 /= topo.num_packages; 597c98d5d94SLen Brown average.packages.pc6 /= topo.num_packages; 598c98d5d94SLen Brown average.packages.pc7 /= topo.num_packages; 599c98d5d94SLen Brown } 600c98d5d94SLen Brown 601c98d5d94SLen Brown static unsigned long long rdtsc(void) 602c98d5d94SLen Brown { 603c98d5d94SLen Brown unsigned int low, high; 604c98d5d94SLen Brown 605c98d5d94SLen Brown asm volatile("rdtsc" : "=a" (low), "=d" (high)); 606c98d5d94SLen Brown 607c98d5d94SLen Brown return low | ((unsigned long long)high) << 32; 608c98d5d94SLen Brown } 609c98d5d94SLen Brown 610c98d5d94SLen Brown 611c98d5d94SLen Brown /* 612c98d5d94SLen Brown * get_counters(...) 613c98d5d94SLen Brown * migrate to cpu 614c98d5d94SLen Brown * acquire and record local counters for that cpu 615c98d5d94SLen Brown */ 616c98d5d94SLen Brown int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) 617c98d5d94SLen Brown { 618c98d5d94SLen Brown int cpu = t->cpu_id; 619c98d5d94SLen Brown 620c98d5d94SLen Brown if (cpu_migrate(cpu)) 62115aaa346SLen Brown return -1; 62215aaa346SLen Brown 623c98d5d94SLen Brown t->tsc = rdtsc(); /* we are running on local CPU of interest */ 62415aaa346SLen Brown 62515aaa346SLen Brown if (has_aperf) { 626c98d5d94SLen Brown if (get_msr(cpu, MSR_APERF, &t->aperf)) 627c98d5d94SLen Brown return -3; 628c98d5d94SLen Brown if (get_msr(cpu, MSR_MPERF, &t->mperf)) 629c98d5d94SLen Brown return -4; 63088c3281fSLen Brown } 63188c3281fSLen Brown 632c98d5d94SLen Brown if (extra_msr_offset) 633c98d5d94SLen Brown if (get_msr(cpu, extra_msr_offset, &t->extra_msr)) 634c98d5d94SLen Brown return -5; 635c98d5d94SLen Brown 636c98d5d94SLen Brown /* collect core counters only for 1st thread in core */ 637c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 638c98d5d94SLen Brown return 0; 639c98d5d94SLen Brown 64015aaa346SLen Brown if (do_nhm_cstates) { 641c98d5d94SLen Brown if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3)) 642c98d5d94SLen Brown return -6; 643c98d5d94SLen Brown if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6)) 644c98d5d94SLen Brown return -7; 645103a8feaSLen Brown } 64615aaa346SLen Brown 64715aaa346SLen Brown if (do_snb_cstates) 648c98d5d94SLen Brown if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7)) 649c98d5d94SLen Brown return -8; 650c98d5d94SLen Brown 651c98d5d94SLen Brown /* collect package counters only for 1st core in package */ 652c98d5d94SLen Brown if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) 653c98d5d94SLen Brown return 0; 65415aaa346SLen Brown 65515aaa346SLen Brown if (do_nhm_cstates) { 656c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3)) 657c98d5d94SLen Brown return -9; 658c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6)) 659c98d5d94SLen Brown return -10; 66015aaa346SLen Brown } 66115aaa346SLen Brown if (do_snb_cstates) { 662c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2)) 663c98d5d94SLen Brown return -11; 664c98d5d94SLen Brown if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7)) 665c98d5d94SLen Brown return -12; 66615aaa346SLen Brown } 66715aaa346SLen Brown return 0; 668103a8feaSLen Brown } 669103a8feaSLen Brown 670c98d5d94SLen Brown void print_verbose_header(void) 671103a8feaSLen Brown { 672103a8feaSLen Brown unsigned long long msr; 673103a8feaSLen Brown unsigned int ratio; 674103a8feaSLen Brown 675103a8feaSLen Brown if (!do_nehalem_platform_info) 676103a8feaSLen Brown return; 677103a8feaSLen Brown 67815aaa346SLen Brown get_msr(0, MSR_NEHALEM_PLATFORM_INFO, &msr); 679103a8feaSLen Brown 680103a8feaSLen Brown ratio = (msr >> 40) & 0xFF; 681103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n", 682103a8feaSLen Brown ratio, bclk, ratio * bclk); 683103a8feaSLen Brown 684103a8feaSLen Brown ratio = (msr >> 8) & 0xFF; 685103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n", 686103a8feaSLen Brown ratio, bclk, ratio * bclk); 687103a8feaSLen Brown 688103a8feaSLen Brown if (verbose > 1) 689103a8feaSLen Brown fprintf(stderr, "MSR_NEHALEM_PLATFORM_INFO: 0x%llx\n", msr); 690103a8feaSLen Brown 691103a8feaSLen Brown if (!do_nehalem_turbo_ratio_limit) 692103a8feaSLen Brown return; 693103a8feaSLen Brown 69415aaa346SLen Brown get_msr(0, MSR_NEHALEM_TURBO_RATIO_LIMIT, &msr); 695103a8feaSLen Brown 696103a8feaSLen Brown ratio = (msr >> 24) & 0xFF; 697103a8feaSLen Brown if (ratio) 698103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n", 699103a8feaSLen Brown ratio, bclk, ratio * bclk); 700103a8feaSLen Brown 701103a8feaSLen Brown ratio = (msr >> 16) & 0xFF; 702103a8feaSLen Brown if (ratio) 703103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n", 704103a8feaSLen Brown ratio, bclk, ratio * bclk); 705103a8feaSLen Brown 706103a8feaSLen Brown ratio = (msr >> 8) & 0xFF; 707103a8feaSLen Brown if (ratio) 708103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n", 709103a8feaSLen Brown ratio, bclk, ratio * bclk); 710103a8feaSLen Brown 711103a8feaSLen Brown ratio = (msr >> 0) & 0xFF; 712103a8feaSLen Brown if (ratio) 713103a8feaSLen Brown fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n", 714103a8feaSLen Brown ratio, bclk, ratio * bclk); 715103a8feaSLen Brown 716103a8feaSLen Brown } 717103a8feaSLen Brown 718c98d5d94SLen Brown void free_all_buffers(void) 719103a8feaSLen Brown { 720c98d5d94SLen Brown CPU_FREE(cpu_present_set); 721c98d5d94SLen Brown cpu_present_set = NULL; 722c98d5d94SLen Brown cpu_present_set = 0; 723103a8feaSLen Brown 724c98d5d94SLen Brown CPU_FREE(cpu_affinity_set); 725c98d5d94SLen Brown cpu_affinity_set = NULL; 726c98d5d94SLen Brown cpu_affinity_setsize = 0; 727103a8feaSLen Brown 728c98d5d94SLen Brown free(thread_even); 729c98d5d94SLen Brown free(core_even); 730c98d5d94SLen Brown free(package_even); 731103a8feaSLen Brown 732c98d5d94SLen Brown thread_even = NULL; 733c98d5d94SLen Brown core_even = NULL; 734c98d5d94SLen Brown package_even = NULL; 735103a8feaSLen Brown 736c98d5d94SLen Brown free(thread_odd); 737c98d5d94SLen Brown free(core_odd); 738c98d5d94SLen Brown free(package_odd); 739103a8feaSLen Brown 740c98d5d94SLen Brown thread_odd = NULL; 741c98d5d94SLen Brown core_odd = NULL; 742c98d5d94SLen Brown package_odd = NULL; 743103a8feaSLen Brown 744c98d5d94SLen Brown free(output_buffer); 745c98d5d94SLen Brown output_buffer = NULL; 746c98d5d94SLen Brown outp = NULL; 747103a8feaSLen Brown } 748103a8feaSLen Brown 749103a8feaSLen Brown /* 750c98d5d94SLen Brown * cpu_is_first_sibling_in_core(cpu) 751c98d5d94SLen Brown * return 1 if given CPU is 1st HT sibling in the core 752103a8feaSLen Brown */ 753c98d5d94SLen Brown int cpu_is_first_sibling_in_core(int cpu) 754103a8feaSLen Brown { 755c98d5d94SLen Brown char path[64]; 756c98d5d94SLen Brown FILE *filep; 757c98d5d94SLen Brown int first_cpu; 758103a8feaSLen Brown 759c98d5d94SLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu); 760c98d5d94SLen Brown filep = fopen(path, "r"); 761c98d5d94SLen Brown if (filep == NULL) { 762c98d5d94SLen Brown perror(path); 763103a8feaSLen Brown exit(1); 764103a8feaSLen Brown } 765c98d5d94SLen Brown fscanf(filep, "%d", &first_cpu); 766c98d5d94SLen Brown fclose(filep); 767c98d5d94SLen Brown return (cpu == first_cpu); 768c98d5d94SLen Brown } 769103a8feaSLen Brown 770c98d5d94SLen Brown /* 771c98d5d94SLen Brown * cpu_is_first_core_in_package(cpu) 772c98d5d94SLen Brown * return 1 if given CPU is 1st core in package 773c98d5d94SLen Brown */ 774c98d5d94SLen Brown int cpu_is_first_core_in_package(int cpu) 775c98d5d94SLen Brown { 776c98d5d94SLen Brown char path[64]; 777c98d5d94SLen Brown FILE *filep; 778c98d5d94SLen Brown int first_cpu; 779c98d5d94SLen Brown 780c98d5d94SLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu); 781c98d5d94SLen Brown filep = fopen(path, "r"); 782c98d5d94SLen Brown if (filep == NULL) { 783c98d5d94SLen Brown perror(path); 784103a8feaSLen Brown exit(1); 785103a8feaSLen Brown } 786c98d5d94SLen Brown fscanf(filep, "%d", &first_cpu); 787c98d5d94SLen Brown fclose(filep); 788c98d5d94SLen Brown return (cpu == first_cpu); 789103a8feaSLen Brown } 790103a8feaSLen Brown 791103a8feaSLen Brown int get_physical_package_id(int cpu) 792103a8feaSLen Brown { 793c98d5d94SLen Brown char path[80]; 794103a8feaSLen Brown FILE *filep; 795103a8feaSLen Brown int pkg; 796103a8feaSLen Brown 797103a8feaSLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu); 798103a8feaSLen Brown filep = fopen(path, "r"); 799103a8feaSLen Brown if (filep == NULL) { 800103a8feaSLen Brown perror(path); 801103a8feaSLen Brown exit(1); 802103a8feaSLen Brown } 803103a8feaSLen Brown fscanf(filep, "%d", &pkg); 804103a8feaSLen Brown fclose(filep); 805103a8feaSLen Brown return pkg; 806103a8feaSLen Brown } 807103a8feaSLen Brown 808103a8feaSLen Brown int get_core_id(int cpu) 809103a8feaSLen Brown { 810c98d5d94SLen Brown char path[80]; 811103a8feaSLen Brown FILE *filep; 812103a8feaSLen Brown int core; 813103a8feaSLen Brown 814103a8feaSLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_id", cpu); 815103a8feaSLen Brown filep = fopen(path, "r"); 816103a8feaSLen Brown if (filep == NULL) { 817103a8feaSLen Brown perror(path); 818103a8feaSLen Brown exit(1); 819103a8feaSLen Brown } 820103a8feaSLen Brown fscanf(filep, "%d", &core); 821103a8feaSLen Brown fclose(filep); 822103a8feaSLen Brown return core; 823103a8feaSLen Brown } 824103a8feaSLen Brown 825c98d5d94SLen Brown int get_num_ht_siblings(int cpu) 826c98d5d94SLen Brown { 827c98d5d94SLen Brown char path[80]; 828c98d5d94SLen Brown FILE *filep; 829c98d5d94SLen Brown int sib1, sib2; 830c98d5d94SLen Brown int matches; 831c98d5d94SLen Brown char character; 832c98d5d94SLen Brown 833c98d5d94SLen Brown sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu); 834c98d5d94SLen Brown filep = fopen(path, "r"); 835c98d5d94SLen Brown if (filep == NULL) { 836c98d5d94SLen Brown perror(path); 837c98d5d94SLen Brown exit(1); 838c98d5d94SLen Brown } 839103a8feaSLen Brown /* 840c98d5d94SLen Brown * file format: 841c98d5d94SLen Brown * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4) 842c98d5d94SLen Brown * otherwinse 1 sibling (self). 843c98d5d94SLen Brown */ 844c98d5d94SLen Brown matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2); 845c98d5d94SLen Brown 846c98d5d94SLen Brown fclose(filep); 847c98d5d94SLen Brown 848c98d5d94SLen Brown if (matches == 3) 849c98d5d94SLen Brown return 2; 850c98d5d94SLen Brown else 851c98d5d94SLen Brown return 1; 852c98d5d94SLen Brown } 853c98d5d94SLen Brown 854c98d5d94SLen Brown /* 855c98d5d94SLen Brown * run func(thread, core, package) in topology order 856c98d5d94SLen Brown * skip non-present cpus 857103a8feaSLen Brown */ 858103a8feaSLen Brown 859c98d5d94SLen Brown int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *, 860c98d5d94SLen Brown struct pkg_data *, struct thread_data *, struct core_data *, 861c98d5d94SLen Brown struct pkg_data *), struct thread_data *thread_base, 862c98d5d94SLen Brown struct core_data *core_base, struct pkg_data *pkg_base, 863c98d5d94SLen Brown struct thread_data *thread_base2, struct core_data *core_base2, 864c98d5d94SLen Brown struct pkg_data *pkg_base2) 865c98d5d94SLen Brown { 866c98d5d94SLen Brown int retval, pkg_no, core_no, thread_no; 867c98d5d94SLen Brown 868c98d5d94SLen Brown for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) { 869c98d5d94SLen Brown for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) { 870c98d5d94SLen Brown for (thread_no = 0; thread_no < 871c98d5d94SLen Brown topo.num_threads_per_core; ++thread_no) { 872c98d5d94SLen Brown struct thread_data *t, *t2; 873c98d5d94SLen Brown struct core_data *c, *c2; 874c98d5d94SLen Brown struct pkg_data *p, *p2; 875c98d5d94SLen Brown 876c98d5d94SLen Brown t = GET_THREAD(thread_base, thread_no, core_no, pkg_no); 877c98d5d94SLen Brown 878c98d5d94SLen Brown if (cpu_is_not_present(t->cpu_id)) 879c98d5d94SLen Brown continue; 880c98d5d94SLen Brown 881c98d5d94SLen Brown t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no); 882c98d5d94SLen Brown 883c98d5d94SLen Brown c = GET_CORE(core_base, core_no, pkg_no); 884c98d5d94SLen Brown c2 = GET_CORE(core_base2, core_no, pkg_no); 885c98d5d94SLen Brown 886c98d5d94SLen Brown p = GET_PKG(pkg_base, pkg_no); 887c98d5d94SLen Brown p2 = GET_PKG(pkg_base2, pkg_no); 888c98d5d94SLen Brown 889c98d5d94SLen Brown retval = func(t, c, p, t2, c2, p2); 890c98d5d94SLen Brown if (retval) 891c98d5d94SLen Brown return retval; 892c98d5d94SLen Brown } 893c98d5d94SLen Brown } 894c98d5d94SLen Brown } 895c98d5d94SLen Brown return 0; 896c98d5d94SLen Brown } 897c98d5d94SLen Brown 898c98d5d94SLen Brown /* 899c98d5d94SLen Brown * run func(cpu) on every cpu in /proc/stat 900c98d5d94SLen Brown * return max_cpu number 901c98d5d94SLen Brown */ 902c98d5d94SLen Brown int for_all_proc_cpus(int (func)(int)) 903103a8feaSLen Brown { 904103a8feaSLen Brown FILE *fp; 905c98d5d94SLen Brown int cpu_num; 906103a8feaSLen Brown int retval; 907103a8feaSLen Brown 908103a8feaSLen Brown fp = fopen(proc_stat, "r"); 909103a8feaSLen Brown if (fp == NULL) { 910103a8feaSLen Brown perror(proc_stat); 911103a8feaSLen Brown exit(1); 912103a8feaSLen Brown } 913103a8feaSLen Brown 914103a8feaSLen Brown retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n"); 915103a8feaSLen Brown if (retval != 0) { 916103a8feaSLen Brown perror("/proc/stat format"); 917103a8feaSLen Brown exit(1); 918103a8feaSLen Brown } 919103a8feaSLen Brown 920c98d5d94SLen Brown while (1) { 921c98d5d94SLen Brown retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num); 922103a8feaSLen Brown if (retval != 1) 923103a8feaSLen Brown break; 924103a8feaSLen Brown 925c98d5d94SLen Brown retval = func(cpu_num); 926c98d5d94SLen Brown if (retval) { 927c98d5d94SLen Brown fclose(fp); 928c98d5d94SLen Brown return(retval); 929c98d5d94SLen Brown } 930103a8feaSLen Brown } 931103a8feaSLen Brown fclose(fp); 932c98d5d94SLen Brown return 0; 933103a8feaSLen Brown } 934103a8feaSLen Brown 935103a8feaSLen Brown void re_initialize(void) 936103a8feaSLen Brown { 937c98d5d94SLen Brown free_all_buffers(); 938c98d5d94SLen Brown setup_all_buffers(); 939c98d5d94SLen Brown printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus); 940103a8feaSLen Brown } 941103a8feaSLen Brown 942c98d5d94SLen Brown 943103a8feaSLen Brown /* 944c98d5d94SLen Brown * count_cpus() 945c98d5d94SLen Brown * remember the last one seen, it will be the max 946103a8feaSLen Brown */ 947c98d5d94SLen Brown int count_cpus(int cpu) 948103a8feaSLen Brown { 949c98d5d94SLen Brown if (topo.max_cpu_num < cpu) 950c98d5d94SLen Brown topo.max_cpu_num = cpu; 951103a8feaSLen Brown 952c98d5d94SLen Brown topo.num_cpus += 1; 953c98d5d94SLen Brown return 0; 954103a8feaSLen Brown } 955c98d5d94SLen Brown int mark_cpu_present(int cpu) 956c98d5d94SLen Brown { 957c98d5d94SLen Brown CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set); 95815aaa346SLen Brown return 0; 959103a8feaSLen Brown } 960103a8feaSLen Brown 961103a8feaSLen Brown void turbostat_loop() 962103a8feaSLen Brown { 963c98d5d94SLen Brown int retval; 964c98d5d94SLen Brown 965103a8feaSLen Brown restart: 966c98d5d94SLen Brown retval = for_all_cpus(get_counters, EVEN_COUNTERS); 967c98d5d94SLen Brown if (retval) { 968c98d5d94SLen Brown re_initialize(); 969c98d5d94SLen Brown goto restart; 970c98d5d94SLen Brown } 971103a8feaSLen Brown gettimeofday(&tv_even, (struct timezone *)NULL); 972103a8feaSLen Brown 973103a8feaSLen Brown while (1) { 974c98d5d94SLen Brown if (for_all_proc_cpus(cpu_is_not_present)) { 975103a8feaSLen Brown re_initialize(); 976103a8feaSLen Brown goto restart; 977103a8feaSLen Brown } 978103a8feaSLen Brown sleep(interval_sec); 979c98d5d94SLen Brown retval = for_all_cpus(get_counters, ODD_COUNTERS); 980c98d5d94SLen Brown if (retval) { 98115aaa346SLen Brown re_initialize(); 98215aaa346SLen Brown goto restart; 98315aaa346SLen Brown } 984103a8feaSLen Brown gettimeofday(&tv_odd, (struct timezone *)NULL); 985103a8feaSLen Brown timersub(&tv_odd, &tv_even, &tv_delta); 986c98d5d94SLen Brown for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS); 987c98d5d94SLen Brown compute_average(EVEN_COUNTERS); 988c98d5d94SLen Brown format_all_counters(EVEN_COUNTERS); 989c98d5d94SLen Brown flush_stdout(); 99015aaa346SLen Brown sleep(interval_sec); 991c98d5d94SLen Brown retval = for_all_cpus(get_counters, EVEN_COUNTERS); 992c98d5d94SLen Brown if (retval) { 993103a8feaSLen Brown re_initialize(); 994103a8feaSLen Brown goto restart; 995103a8feaSLen Brown } 996103a8feaSLen Brown gettimeofday(&tv_even, (struct timezone *)NULL); 997103a8feaSLen Brown timersub(&tv_even, &tv_odd, &tv_delta); 998c98d5d94SLen Brown for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS); 999c98d5d94SLen Brown compute_average(ODD_COUNTERS); 1000c98d5d94SLen Brown format_all_counters(ODD_COUNTERS); 1001c98d5d94SLen Brown flush_stdout(); 1002103a8feaSLen Brown } 1003103a8feaSLen Brown } 1004103a8feaSLen Brown 1005103a8feaSLen Brown void check_dev_msr() 1006103a8feaSLen Brown { 1007103a8feaSLen Brown struct stat sb; 1008103a8feaSLen Brown 1009103a8feaSLen Brown if (stat("/dev/cpu/0/msr", &sb)) { 1010103a8feaSLen Brown fprintf(stderr, "no /dev/cpu/0/msr\n"); 1011103a8feaSLen Brown fprintf(stderr, "Try \"# modprobe msr\"\n"); 1012103a8feaSLen Brown exit(-5); 1013103a8feaSLen Brown } 1014103a8feaSLen Brown } 1015103a8feaSLen Brown 1016103a8feaSLen Brown void check_super_user() 1017103a8feaSLen Brown { 1018103a8feaSLen Brown if (getuid() != 0) { 1019103a8feaSLen Brown fprintf(stderr, "must be root\n"); 1020103a8feaSLen Brown exit(-6); 1021103a8feaSLen Brown } 1022103a8feaSLen Brown } 1023103a8feaSLen Brown 1024103a8feaSLen Brown int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model) 1025103a8feaSLen Brown { 1026103a8feaSLen Brown if (!genuine_intel) 1027103a8feaSLen Brown return 0; 1028103a8feaSLen Brown 1029103a8feaSLen Brown if (family != 6) 1030103a8feaSLen Brown return 0; 1031103a8feaSLen Brown 1032103a8feaSLen Brown switch (model) { 1033103a8feaSLen Brown case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */ 1034103a8feaSLen Brown case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */ 1035103a8feaSLen Brown case 0x1F: /* Core i7 and i5 Processor - Nehalem */ 1036103a8feaSLen Brown case 0x25: /* Westmere Client - Clarkdale, Arrandale */ 1037103a8feaSLen Brown case 0x2C: /* Westmere EP - Gulftown */ 1038103a8feaSLen Brown case 0x2A: /* SNB */ 1039103a8feaSLen Brown case 0x2D: /* SNB Xeon */ 1040553575f1SLen Brown case 0x3A: /* IVB */ 1041553575f1SLen Brown case 0x3D: /* IVB Xeon */ 1042103a8feaSLen Brown return 1; 1043103a8feaSLen Brown case 0x2E: /* Nehalem-EX Xeon - Beckton */ 1044103a8feaSLen Brown case 0x2F: /* Westmere-EX Xeon - Eagleton */ 1045103a8feaSLen Brown default: 1046103a8feaSLen Brown return 0; 1047103a8feaSLen Brown } 1048103a8feaSLen Brown } 1049103a8feaSLen Brown 1050103a8feaSLen Brown int is_snb(unsigned int family, unsigned int model) 1051103a8feaSLen Brown { 1052103a8feaSLen Brown if (!genuine_intel) 1053103a8feaSLen Brown return 0; 1054103a8feaSLen Brown 1055103a8feaSLen Brown switch (model) { 1056103a8feaSLen Brown case 0x2A: 1057103a8feaSLen Brown case 0x2D: 1058650a37f3SLen Brown case 0x3A: /* IVB */ 1059650a37f3SLen Brown case 0x3D: /* IVB Xeon */ 1060103a8feaSLen Brown return 1; 1061103a8feaSLen Brown } 1062103a8feaSLen Brown return 0; 1063103a8feaSLen Brown } 1064103a8feaSLen Brown 1065103a8feaSLen Brown double discover_bclk(unsigned int family, unsigned int model) 1066103a8feaSLen Brown { 1067103a8feaSLen Brown if (is_snb(family, model)) 1068103a8feaSLen Brown return 100.00; 1069103a8feaSLen Brown else 1070103a8feaSLen Brown return 133.33; 1071103a8feaSLen Brown } 1072103a8feaSLen Brown 1073103a8feaSLen Brown void check_cpuid() 1074103a8feaSLen Brown { 1075103a8feaSLen Brown unsigned int eax, ebx, ecx, edx, max_level; 1076103a8feaSLen Brown unsigned int fms, family, model, stepping; 1077103a8feaSLen Brown 1078103a8feaSLen Brown eax = ebx = ecx = edx = 0; 1079103a8feaSLen Brown 1080103a8feaSLen Brown asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0)); 1081103a8feaSLen Brown 1082103a8feaSLen Brown if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e) 1083103a8feaSLen Brown genuine_intel = 1; 1084103a8feaSLen Brown 1085103a8feaSLen Brown if (verbose) 1086103a8feaSLen Brown fprintf(stderr, "%.4s%.4s%.4s ", 1087103a8feaSLen Brown (char *)&ebx, (char *)&edx, (char *)&ecx); 1088103a8feaSLen Brown 1089103a8feaSLen Brown asm("cpuid" : "=a" (fms), "=c" (ecx), "=d" (edx) : "a" (1) : "ebx"); 1090103a8feaSLen Brown family = (fms >> 8) & 0xf; 1091103a8feaSLen Brown model = (fms >> 4) & 0xf; 1092103a8feaSLen Brown stepping = fms & 0xf; 1093103a8feaSLen Brown if (family == 6 || family == 0xf) 1094103a8feaSLen Brown model += ((fms >> 16) & 0xf) << 4; 1095103a8feaSLen Brown 1096103a8feaSLen Brown if (verbose) 1097103a8feaSLen Brown fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n", 1098103a8feaSLen Brown max_level, family, model, stepping, family, model, stepping); 1099103a8feaSLen Brown 1100103a8feaSLen Brown if (!(edx & (1 << 5))) { 1101103a8feaSLen Brown fprintf(stderr, "CPUID: no MSR\n"); 1102103a8feaSLen Brown exit(1); 1103103a8feaSLen Brown } 1104103a8feaSLen Brown 1105103a8feaSLen Brown /* 1106103a8feaSLen Brown * check max extended function levels of CPUID. 1107103a8feaSLen Brown * This is needed to check for invariant TSC. 1108103a8feaSLen Brown * This check is valid for both Intel and AMD. 1109103a8feaSLen Brown */ 1110103a8feaSLen Brown ebx = ecx = edx = 0; 1111103a8feaSLen Brown asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000000)); 1112103a8feaSLen Brown 1113103a8feaSLen Brown if (max_level < 0x80000007) { 1114103a8feaSLen Brown fprintf(stderr, "CPUID: no invariant TSC (max_level 0x%x)\n", max_level); 1115103a8feaSLen Brown exit(1); 1116103a8feaSLen Brown } 1117103a8feaSLen Brown 1118103a8feaSLen Brown /* 1119103a8feaSLen Brown * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8 1120103a8feaSLen Brown * this check is valid for both Intel and AMD 1121103a8feaSLen Brown */ 1122103a8feaSLen Brown asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000007)); 11238209e054SThomas Renninger has_invariant_tsc = edx & (1 << 8); 1124103a8feaSLen Brown 1125103a8feaSLen Brown if (!has_invariant_tsc) { 1126103a8feaSLen Brown fprintf(stderr, "No invariant TSC\n"); 1127103a8feaSLen Brown exit(1); 1128103a8feaSLen Brown } 1129103a8feaSLen Brown 1130103a8feaSLen Brown /* 1131103a8feaSLen Brown * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0 1132103a8feaSLen Brown * this check is valid for both Intel and AMD 1133103a8feaSLen Brown */ 1134103a8feaSLen Brown 1135103a8feaSLen Brown asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x6)); 11368209e054SThomas Renninger has_aperf = ecx & (1 << 0); 1137103a8feaSLen Brown if (!has_aperf) { 1138103a8feaSLen Brown fprintf(stderr, "No APERF MSR\n"); 1139103a8feaSLen Brown exit(1); 1140103a8feaSLen Brown } 1141103a8feaSLen Brown 1142103a8feaSLen Brown do_nehalem_platform_info = genuine_intel && has_invariant_tsc; 1143103a8feaSLen Brown do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */ 1144103a8feaSLen Brown do_snb_cstates = is_snb(family, model); 1145103a8feaSLen Brown bclk = discover_bclk(family, model); 1146103a8feaSLen Brown 1147103a8feaSLen Brown do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model); 1148103a8feaSLen Brown } 1149103a8feaSLen Brown 1150103a8feaSLen Brown 1151103a8feaSLen Brown void usage() 1152103a8feaSLen Brown { 1153103a8feaSLen Brown fprintf(stderr, "%s: [-v] [-M MSR#] [-i interval_sec | command ...]\n", 1154103a8feaSLen Brown progname); 1155103a8feaSLen Brown exit(1); 1156103a8feaSLen Brown } 1157103a8feaSLen Brown 1158103a8feaSLen Brown 1159103a8feaSLen Brown /* 1160103a8feaSLen Brown * in /dev/cpu/ return success for names that are numbers 1161103a8feaSLen Brown * ie. filter out ".", "..", "microcode". 1162103a8feaSLen Brown */ 1163103a8feaSLen Brown int dir_filter(const struct dirent *dirp) 1164103a8feaSLen Brown { 1165103a8feaSLen Brown if (isdigit(dirp->d_name[0])) 1166103a8feaSLen Brown return 1; 1167103a8feaSLen Brown else 1168103a8feaSLen Brown return 0; 1169103a8feaSLen Brown } 1170103a8feaSLen Brown 1171103a8feaSLen Brown int open_dev_cpu_msr(int dummy1) 1172103a8feaSLen Brown { 1173103a8feaSLen Brown return 0; 1174103a8feaSLen Brown } 1175103a8feaSLen Brown 1176c98d5d94SLen Brown void topology_probe() 1177c98d5d94SLen Brown { 1178c98d5d94SLen Brown int i; 1179c98d5d94SLen Brown int max_core_id = 0; 1180c98d5d94SLen Brown int max_package_id = 0; 1181c98d5d94SLen Brown int max_siblings = 0; 1182c98d5d94SLen Brown struct cpu_topology { 1183c98d5d94SLen Brown int core_id; 1184c98d5d94SLen Brown int physical_package_id; 1185c98d5d94SLen Brown } *cpus; 1186c98d5d94SLen Brown 1187c98d5d94SLen Brown /* Initialize num_cpus, max_cpu_num */ 1188c98d5d94SLen Brown topo.num_cpus = 0; 1189c98d5d94SLen Brown topo.max_cpu_num = 0; 1190c98d5d94SLen Brown for_all_proc_cpus(count_cpus); 1191c98d5d94SLen Brown if (!summary_only && topo.num_cpus > 1) 1192c98d5d94SLen Brown show_cpu = 1; 1193c98d5d94SLen Brown 1194c98d5d94SLen Brown if (verbose > 1) 1195c98d5d94SLen Brown fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num); 1196c98d5d94SLen Brown 1197c98d5d94SLen Brown cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology)); 1198c98d5d94SLen Brown if (cpus == NULL) { 1199c98d5d94SLen Brown perror("calloc cpus"); 1200c98d5d94SLen Brown exit(1); 1201c98d5d94SLen Brown } 1202c98d5d94SLen Brown 1203c98d5d94SLen Brown /* 1204c98d5d94SLen Brown * Allocate and initialize cpu_present_set 1205c98d5d94SLen Brown */ 1206c98d5d94SLen Brown cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1)); 1207c98d5d94SLen Brown if (cpu_present_set == NULL) { 1208c98d5d94SLen Brown perror("CPU_ALLOC"); 1209c98d5d94SLen Brown exit(3); 1210c98d5d94SLen Brown } 1211c98d5d94SLen Brown cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1)); 1212c98d5d94SLen Brown CPU_ZERO_S(cpu_present_setsize, cpu_present_set); 1213c98d5d94SLen Brown for_all_proc_cpus(mark_cpu_present); 1214c98d5d94SLen Brown 1215c98d5d94SLen Brown /* 1216c98d5d94SLen Brown * Allocate and initialize cpu_affinity_set 1217c98d5d94SLen Brown */ 1218c98d5d94SLen Brown cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1)); 1219c98d5d94SLen Brown if (cpu_affinity_set == NULL) { 1220c98d5d94SLen Brown perror("CPU_ALLOC"); 1221c98d5d94SLen Brown exit(3); 1222c98d5d94SLen Brown } 1223c98d5d94SLen Brown cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1)); 1224c98d5d94SLen Brown CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set); 1225c98d5d94SLen Brown 1226c98d5d94SLen Brown 1227c98d5d94SLen Brown /* 1228c98d5d94SLen Brown * For online cpus 1229c98d5d94SLen Brown * find max_core_id, max_package_id 1230c98d5d94SLen Brown */ 1231c98d5d94SLen Brown for (i = 0; i <= topo.max_cpu_num; ++i) { 1232c98d5d94SLen Brown int siblings; 1233c98d5d94SLen Brown 1234c98d5d94SLen Brown if (cpu_is_not_present(i)) { 1235c98d5d94SLen Brown if (verbose > 1) 1236c98d5d94SLen Brown fprintf(stderr, "cpu%d NOT PRESENT\n", i); 1237c98d5d94SLen Brown continue; 1238c98d5d94SLen Brown } 1239c98d5d94SLen Brown cpus[i].core_id = get_core_id(i); 1240c98d5d94SLen Brown if (cpus[i].core_id > max_core_id) 1241c98d5d94SLen Brown max_core_id = cpus[i].core_id; 1242c98d5d94SLen Brown 1243c98d5d94SLen Brown cpus[i].physical_package_id = get_physical_package_id(i); 1244c98d5d94SLen Brown if (cpus[i].physical_package_id > max_package_id) 1245c98d5d94SLen Brown max_package_id = cpus[i].physical_package_id; 1246c98d5d94SLen Brown 1247c98d5d94SLen Brown siblings = get_num_ht_siblings(i); 1248c98d5d94SLen Brown if (siblings > max_siblings) 1249c98d5d94SLen Brown max_siblings = siblings; 1250c98d5d94SLen Brown if (verbose > 1) 1251c98d5d94SLen Brown fprintf(stderr, "cpu %d pkg %d core %d\n", 1252c98d5d94SLen Brown i, cpus[i].physical_package_id, cpus[i].core_id); 1253c98d5d94SLen Brown } 1254c98d5d94SLen Brown topo.num_cores_per_pkg = max_core_id + 1; 1255c98d5d94SLen Brown if (verbose > 1) 1256c98d5d94SLen Brown fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n", 1257c98d5d94SLen Brown max_core_id, topo.num_cores_per_pkg); 1258c98d5d94SLen Brown if (!summary_only && topo.num_cores_per_pkg > 1) 1259c98d5d94SLen Brown show_core = 1; 1260c98d5d94SLen Brown 1261c98d5d94SLen Brown topo.num_packages = max_package_id + 1; 1262c98d5d94SLen Brown if (verbose > 1) 1263c98d5d94SLen Brown fprintf(stderr, "max_package_id %d, sizing for %d packages\n", 1264c98d5d94SLen Brown max_package_id, topo.num_packages); 1265c98d5d94SLen Brown if (!summary_only && topo.num_packages > 1) 1266c98d5d94SLen Brown show_pkg = 1; 1267c98d5d94SLen Brown 1268c98d5d94SLen Brown topo.num_threads_per_core = max_siblings; 1269c98d5d94SLen Brown if (verbose > 1) 1270c98d5d94SLen Brown fprintf(stderr, "max_siblings %d\n", max_siblings); 1271c98d5d94SLen Brown 1272c98d5d94SLen Brown free(cpus); 1273c98d5d94SLen Brown } 1274c98d5d94SLen Brown 1275c98d5d94SLen Brown void 1276c98d5d94SLen Brown allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p) 1277c98d5d94SLen Brown { 1278c98d5d94SLen Brown int i; 1279c98d5d94SLen Brown 1280c98d5d94SLen Brown *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg * 1281c98d5d94SLen Brown topo.num_packages, sizeof(struct thread_data)); 1282c98d5d94SLen Brown if (*t == NULL) 1283c98d5d94SLen Brown goto error; 1284c98d5d94SLen Brown 1285c98d5d94SLen Brown for (i = 0; i < topo.num_threads_per_core * 1286c98d5d94SLen Brown topo.num_cores_per_pkg * topo.num_packages; i++) 1287c98d5d94SLen Brown (*t)[i].cpu_id = -1; 1288c98d5d94SLen Brown 1289c98d5d94SLen Brown *c = calloc(topo.num_cores_per_pkg * topo.num_packages, 1290c98d5d94SLen Brown sizeof(struct core_data)); 1291c98d5d94SLen Brown if (*c == NULL) 1292c98d5d94SLen Brown goto error; 1293c98d5d94SLen Brown 1294c98d5d94SLen Brown for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++) 1295c98d5d94SLen Brown (*c)[i].core_id = -1; 1296c98d5d94SLen Brown 1297c98d5d94SLen Brown *p = calloc(topo.num_packages, sizeof(struct pkg_data)); 1298c98d5d94SLen Brown if (*p == NULL) 1299c98d5d94SLen Brown goto error; 1300c98d5d94SLen Brown 1301c98d5d94SLen Brown for (i = 0; i < topo.num_packages; i++) 1302c98d5d94SLen Brown (*p)[i].package_id = i; 1303c98d5d94SLen Brown 1304c98d5d94SLen Brown return; 1305c98d5d94SLen Brown error: 1306c98d5d94SLen Brown perror("calloc counters"); 1307c98d5d94SLen Brown exit(1); 1308c98d5d94SLen Brown } 1309c98d5d94SLen Brown /* 1310c98d5d94SLen Brown * init_counter() 1311c98d5d94SLen Brown * 1312c98d5d94SLen Brown * set cpu_id, core_num, pkg_num 1313c98d5d94SLen Brown * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE 1314c98d5d94SLen Brown * 1315c98d5d94SLen Brown * increment topo.num_cores when 1st core in pkg seen 1316c98d5d94SLen Brown */ 1317c98d5d94SLen Brown void init_counter(struct thread_data *thread_base, struct core_data *core_base, 1318c98d5d94SLen Brown struct pkg_data *pkg_base, int thread_num, int core_num, 1319c98d5d94SLen Brown int pkg_num, int cpu_id) 1320c98d5d94SLen Brown { 1321c98d5d94SLen Brown struct thread_data *t; 1322c98d5d94SLen Brown struct core_data *c; 1323c98d5d94SLen Brown struct pkg_data *p; 1324c98d5d94SLen Brown 1325c98d5d94SLen Brown t = GET_THREAD(thread_base, thread_num, core_num, pkg_num); 1326c98d5d94SLen Brown c = GET_CORE(core_base, core_num, pkg_num); 1327c98d5d94SLen Brown p = GET_PKG(pkg_base, pkg_num); 1328c98d5d94SLen Brown 1329c98d5d94SLen Brown t->cpu_id = cpu_id; 1330c98d5d94SLen Brown if (thread_num == 0) { 1331c98d5d94SLen Brown t->flags |= CPU_IS_FIRST_THREAD_IN_CORE; 1332c98d5d94SLen Brown if (cpu_is_first_core_in_package(cpu_id)) 1333c98d5d94SLen Brown t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE; 1334c98d5d94SLen Brown } 1335c98d5d94SLen Brown 1336c98d5d94SLen Brown c->core_id = core_num; 1337c98d5d94SLen Brown p->package_id = pkg_num; 1338c98d5d94SLen Brown } 1339c98d5d94SLen Brown 1340c98d5d94SLen Brown 1341c98d5d94SLen Brown int initialize_counters(int cpu_id) 1342c98d5d94SLen Brown { 1343c98d5d94SLen Brown int my_thread_id, my_core_id, my_package_id; 1344c98d5d94SLen Brown 1345c98d5d94SLen Brown my_package_id = get_physical_package_id(cpu_id); 1346c98d5d94SLen Brown my_core_id = get_core_id(cpu_id); 1347c98d5d94SLen Brown 1348c98d5d94SLen Brown if (cpu_is_first_sibling_in_core(cpu_id)) { 1349c98d5d94SLen Brown my_thread_id = 0; 1350c98d5d94SLen Brown topo.num_cores++; 1351c98d5d94SLen Brown } else { 1352c98d5d94SLen Brown my_thread_id = 1; 1353c98d5d94SLen Brown } 1354c98d5d94SLen Brown 1355c98d5d94SLen Brown init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id); 1356c98d5d94SLen Brown init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id); 1357c98d5d94SLen Brown return 0; 1358c98d5d94SLen Brown } 1359c98d5d94SLen Brown 1360c98d5d94SLen Brown void allocate_output_buffer() 1361c98d5d94SLen Brown { 1362c98d5d94SLen Brown output_buffer = calloc(1, (1 + topo.num_cpus) * 128); 1363c98d5d94SLen Brown outp = output_buffer; 1364c98d5d94SLen Brown if (outp == NULL) { 1365c98d5d94SLen Brown perror("calloc"); 1366c98d5d94SLen Brown exit(-1); 1367c98d5d94SLen Brown } 1368c98d5d94SLen Brown } 1369c98d5d94SLen Brown 1370c98d5d94SLen Brown void setup_all_buffers(void) 1371c98d5d94SLen Brown { 1372c98d5d94SLen Brown topology_probe(); 1373c98d5d94SLen Brown allocate_counters(&thread_even, &core_even, &package_even); 1374c98d5d94SLen Brown allocate_counters(&thread_odd, &core_odd, &package_odd); 1375c98d5d94SLen Brown allocate_output_buffer(); 1376c98d5d94SLen Brown for_all_proc_cpus(initialize_counters); 1377c98d5d94SLen Brown } 1378103a8feaSLen Brown void turbostat_init() 1379103a8feaSLen Brown { 1380103a8feaSLen Brown check_cpuid(); 1381103a8feaSLen Brown 1382103a8feaSLen Brown check_dev_msr(); 1383103a8feaSLen Brown check_super_user(); 1384103a8feaSLen Brown 1385c98d5d94SLen Brown setup_all_buffers(); 1386103a8feaSLen Brown 1387103a8feaSLen Brown if (verbose) 1388c98d5d94SLen Brown print_verbose_header(); 1389103a8feaSLen Brown } 1390103a8feaSLen Brown 1391103a8feaSLen Brown int fork_it(char **argv) 1392103a8feaSLen Brown { 1393103a8feaSLen Brown pid_t child_pid; 1394d15cf7c1SLen Brown 1395c98d5d94SLen Brown for_all_cpus(get_counters, EVEN_COUNTERS); 1396d15cf7c1SLen Brown /* clear affinity side-effect of get_counters() */ 1397d15cf7c1SLen Brown sched_setaffinity(0, cpu_present_setsize, cpu_present_set); 1398103a8feaSLen Brown gettimeofday(&tv_even, (struct timezone *)NULL); 1399103a8feaSLen Brown 1400103a8feaSLen Brown child_pid = fork(); 1401103a8feaSLen Brown if (!child_pid) { 1402103a8feaSLen Brown /* child */ 1403103a8feaSLen Brown execvp(argv[0], argv); 1404103a8feaSLen Brown } else { 1405103a8feaSLen Brown int status; 1406103a8feaSLen Brown 1407103a8feaSLen Brown /* parent */ 1408103a8feaSLen Brown if (child_pid == -1) { 1409103a8feaSLen Brown perror("fork"); 1410103a8feaSLen Brown exit(1); 1411103a8feaSLen Brown } 1412103a8feaSLen Brown 1413103a8feaSLen Brown signal(SIGINT, SIG_IGN); 1414103a8feaSLen Brown signal(SIGQUIT, SIG_IGN); 1415103a8feaSLen Brown if (waitpid(child_pid, &status, 0) == -1) { 1416103a8feaSLen Brown perror("wait"); 1417103a8feaSLen Brown exit(1); 1418103a8feaSLen Brown } 1419103a8feaSLen Brown } 1420c98d5d94SLen Brown /* 1421c98d5d94SLen Brown * n.b. fork_it() does not check for errors from for_all_cpus() 1422c98d5d94SLen Brown * because re-starting is problematic when forking 1423c98d5d94SLen Brown */ 1424c98d5d94SLen Brown for_all_cpus(get_counters, ODD_COUNTERS); 1425103a8feaSLen Brown gettimeofday(&tv_odd, (struct timezone *)NULL); 1426103a8feaSLen Brown timersub(&tv_odd, &tv_even, &tv_delta); 1427c98d5d94SLen Brown for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS); 1428c98d5d94SLen Brown compute_average(EVEN_COUNTERS); 1429c98d5d94SLen Brown format_all_counters(EVEN_COUNTERS); 1430c98d5d94SLen Brown flush_stderr(); 1431103a8feaSLen Brown 14326eab04a8SJustin P. Mattock fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0); 1433103a8feaSLen Brown 1434103a8feaSLen Brown return 0; 1435103a8feaSLen Brown } 1436103a8feaSLen Brown 1437103a8feaSLen Brown void cmdline(int argc, char **argv) 1438103a8feaSLen Brown { 1439103a8feaSLen Brown int opt; 1440103a8feaSLen Brown 1441103a8feaSLen Brown progname = argv[0]; 1442103a8feaSLen Brown 1443c98d5d94SLen Brown while ((opt = getopt(argc, argv, "+cpsvi:M:")) != -1) { 1444103a8feaSLen Brown switch (opt) { 1445c98d5d94SLen Brown case 'c': 1446c98d5d94SLen Brown show_core_only++; 1447c98d5d94SLen Brown break; 1448c98d5d94SLen Brown case 'p': 1449c98d5d94SLen Brown show_pkg_only++; 1450c98d5d94SLen Brown break; 1451e23da037SLen Brown case 's': 1452e23da037SLen Brown summary_only++; 1453e23da037SLen Brown break; 1454103a8feaSLen Brown case 'v': 1455103a8feaSLen Brown verbose++; 1456103a8feaSLen Brown break; 1457103a8feaSLen Brown case 'i': 1458103a8feaSLen Brown interval_sec = atoi(optarg); 1459103a8feaSLen Brown break; 1460103a8feaSLen Brown case 'M': 1461103a8feaSLen Brown sscanf(optarg, "%x", &extra_msr_offset); 1462103a8feaSLen Brown if (verbose > 1) 1463103a8feaSLen Brown fprintf(stderr, "MSR 0x%X\n", extra_msr_offset); 1464103a8feaSLen Brown break; 1465103a8feaSLen Brown default: 1466103a8feaSLen Brown usage(); 1467103a8feaSLen Brown } 1468103a8feaSLen Brown } 1469103a8feaSLen Brown } 1470103a8feaSLen Brown 1471103a8feaSLen Brown int main(int argc, char **argv) 1472103a8feaSLen Brown { 1473103a8feaSLen Brown cmdline(argc, argv); 1474103a8feaSLen Brown 1475103a8feaSLen Brown if (verbose > 1) 1476c98d5d94SLen Brown fprintf(stderr, "turbostat v2.0 May 16, 2012" 1477103a8feaSLen Brown " - Len Brown <lenb@kernel.org>\n"); 1478103a8feaSLen Brown 1479103a8feaSLen Brown turbostat_init(); 1480103a8feaSLen Brown 1481103a8feaSLen Brown /* 1482103a8feaSLen Brown * if any params left, it must be a command to fork 1483103a8feaSLen Brown */ 1484103a8feaSLen Brown if (argc - optind) 1485103a8feaSLen Brown return fork_it(argv + optind); 1486103a8feaSLen Brown else 1487103a8feaSLen Brown turbostat_loop(); 1488103a8feaSLen Brown 1489103a8feaSLen Brown return 0; 1490103a8feaSLen Brown } 1491