2 * turbostat -- show CPU frequency and C-state residency
3 * on modern Intel turbo-capable processors.
5 * Copyright (c) 2013 Intel Corporation.
6 * Len Brown <len.brown@intel.com>
8 * This program is free software; you can redistribute it and/or modify it
9 * under the terms and conditions of the GNU General Public License,
10 * version 2, as published by the Free Software Foundation.
12 * This program is distributed in the hope it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
17 * You should have received a copy of the GNU General Public License along with
18 * this program; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
28 #include <sys/types.h>
31 #include <sys/resource.h>
41 #include <linux/capability.h>
44 char *proc_stat = "/proc/stat";
45 unsigned int interval_sec = 5; /* set with -i interval_sec */
46 unsigned int verbose; /* set with -v */
47 unsigned int rapl_verbose; /* set with -R */
48 unsigned int rapl_joules; /* set with -J */
49 unsigned int thermal_verbose; /* set with -T */
50 unsigned int summary_only; /* set with -S */
51 unsigned int dump_only; /* set with -s */
54 unsigned int do_nhm_cstates;
55 unsigned int do_snb_cstates;
56 unsigned int do_c8_c9_c10;
57 unsigned int do_slm_cstates;
58 unsigned int use_c1_residency_msr;
59 unsigned int has_aperf;
61 unsigned int units = 1000000; /* MHz etc */
62 unsigned int genuine_intel;
63 unsigned int has_invariant_tsc;
64 unsigned int do_nehalem_platform_info;
65 unsigned int do_nehalem_turbo_ratio_limit;
66 unsigned int do_ivt_turbo_ratio_limit;
67 unsigned int extra_msr_offset32;
68 unsigned int extra_msr_offset64;
69 unsigned int extra_delta_offset32;
70 unsigned int extra_delta_offset64;
73 unsigned int show_pkg;
74 unsigned int show_core;
75 unsigned int show_cpu;
76 unsigned int show_pkg_only;
77 unsigned int show_core_only;
78 char *output_buffer, *outp;
82 unsigned int tcc_activation_temp;
83 unsigned int tcc_activation_temp_override;
84 double rapl_power_units, rapl_energy_units, rapl_time_units;
85 double rapl_joule_counter_range;
87 #define RAPL_PKG (1 << 0)
88 /* 0x610 MSR_PKG_POWER_LIMIT */
89 /* 0x611 MSR_PKG_ENERGY_STATUS */
90 #define RAPL_PKG_PERF_STATUS (1 << 1)
91 /* 0x613 MSR_PKG_PERF_STATUS */
92 #define RAPL_PKG_POWER_INFO (1 << 2)
93 /* 0x614 MSR_PKG_POWER_INFO */
95 #define RAPL_DRAM (1 << 3)
96 /* 0x618 MSR_DRAM_POWER_LIMIT */
97 /* 0x619 MSR_DRAM_ENERGY_STATUS */
98 /* 0x61c MSR_DRAM_POWER_INFO */
99 #define RAPL_DRAM_PERF_STATUS (1 << 4)
100 /* 0x61b MSR_DRAM_PERF_STATUS */
102 #define RAPL_CORES (1 << 5)
103 /* 0x638 MSR_PP0_POWER_LIMIT */
104 /* 0x639 MSR_PP0_ENERGY_STATUS */
105 #define RAPL_CORE_POLICY (1 << 6)
106 /* 0x63a MSR_PP0_POLICY */
109 #define RAPL_GFX (1 << 7)
110 /* 0x640 MSR_PP1_POWER_LIMIT */
111 /* 0x641 MSR_PP1_ENERGY_STATUS */
112 /* 0x642 MSR_PP1_POLICY */
113 #define TJMAX_DEFAULT 100
115 #define MAX(a, b) ((a) > (b) ? (a) : (b))
117 int aperf_mperf_unstable;
121 cpu_set_t *cpu_present_set, *cpu_affinity_set;
122 size_t cpu_present_setsize, cpu_affinity_setsize;
125 unsigned long long tsc;
126 unsigned long long aperf;
127 unsigned long long mperf;
128 unsigned long long c1;
129 unsigned long long extra_msr64;
130 unsigned long long extra_delta64;
131 unsigned long long extra_msr32;
132 unsigned long long extra_delta32;
133 unsigned int smi_count;
136 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
137 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
138 } *thread_even, *thread_odd;
141 unsigned long long c3;
142 unsigned long long c6;
143 unsigned long long c7;
144 unsigned int core_temp_c;
145 unsigned int core_id;
146 } *core_even, *core_odd;
149 unsigned long long pc2;
150 unsigned long long pc3;
151 unsigned long long pc6;
152 unsigned long long pc7;
153 unsigned long long pc8;
154 unsigned long long pc9;
155 unsigned long long pc10;
156 unsigned int package_id;
157 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
158 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
159 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
160 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
161 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
162 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
163 unsigned int pkg_temp_c;
165 } *package_even, *package_odd;
167 #define ODD_COUNTERS thread_odd, core_odd, package_odd
168 #define EVEN_COUNTERS thread_even, core_even, package_even
170 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
171 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
172 topo.num_threads_per_core + \
173 (core_no) * topo.num_threads_per_core + (thread_no))
174 #define GET_CORE(core_base, core_no, pkg_no) \
175 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
176 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
178 struct system_summary {
179 struct thread_data threads;
180 struct core_data cores;
181 struct pkg_data packages;
190 int num_cores_per_pkg;
191 int num_threads_per_core;
194 struct timeval tv_even, tv_odd, tv_delta;
196 void setup_all_buffers(void);
198 int cpu_is_not_present(int cpu)
200 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
203 * run func(thread, core, package) in topology order
204 * skip non-present cpus
207 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
208 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
210 int retval, pkg_no, core_no, thread_no;
212 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
213 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
214 for (thread_no = 0; thread_no <
215 topo.num_threads_per_core; ++thread_no) {
216 struct thread_data *t;
220 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
222 if (cpu_is_not_present(t->cpu_id))
225 c = GET_CORE(core_base, core_no, pkg_no);
226 p = GET_PKG(pkg_base, pkg_no);
228 retval = func(t, c, p);
237 int cpu_migrate(int cpu)
239 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
240 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
241 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
247 int get_msr(int cpu, off_t offset, unsigned long long *msr)
253 sprintf(pathname, "/dev/cpu/%d/msr", cpu);
254 fd = open(pathname, O_RDONLY);
256 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
258 retval = pread(fd, msr, sizeof *msr, offset);
261 if (retval != sizeof *msr)
262 err(-1, "%s offset 0x%llx read failed", pathname, (unsigned long long)offset);
268 * Example Format w/ field column widths:
270 * Package Core CPU Avg_MHz Bzy_MHz TSC_MHz SMI %Busy CPU_%c1 CPU_%c3 CPU_%c6 CPU_%c7 CoreTmp PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 PkgWatt CorWatt GFXWatt
271 * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
274 void print_header(void)
277 outp += sprintf(outp, " Package");
279 outp += sprintf(outp, " Core");
281 outp += sprintf(outp, " CPU");
283 outp += sprintf(outp, " Avg_MHz");
285 outp += sprintf(outp, " %%Busy");
287 outp += sprintf(outp, " Bzy_MHz");
288 outp += sprintf(outp, " TSC_MHz");
290 outp += sprintf(outp, " SMI");
291 if (extra_delta_offset32)
292 outp += sprintf(outp, " count 0x%03X", extra_delta_offset32);
293 if (extra_delta_offset64)
294 outp += sprintf(outp, " COUNT 0x%03X", extra_delta_offset64);
295 if (extra_msr_offset32)
296 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset32);
297 if (extra_msr_offset64)
298 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset64);
300 outp += sprintf(outp, " CPU%%c1");
301 if (do_nhm_cstates && !do_slm_cstates)
302 outp += sprintf(outp, " CPU%%c3");
304 outp += sprintf(outp, " CPU%%c6");
306 outp += sprintf(outp, " CPU%%c7");
309 outp += sprintf(outp, " CoreTmp");
311 outp += sprintf(outp, " PkgTmp");
314 outp += sprintf(outp, " Pkg%%pc2");
315 if (do_nhm_cstates && !do_slm_cstates)
316 outp += sprintf(outp, " Pkg%%pc3");
317 if (do_nhm_cstates && !do_slm_cstates)
318 outp += sprintf(outp, " Pkg%%pc6");
320 outp += sprintf(outp, " Pkg%%pc7");
322 outp += sprintf(outp, " Pkg%%pc8");
323 outp += sprintf(outp, " Pkg%%pc9");
324 outp += sprintf(outp, " Pk%%pc10");
327 if (do_rapl && !rapl_joules) {
328 if (do_rapl & RAPL_PKG)
329 outp += sprintf(outp, " PkgWatt");
330 if (do_rapl & RAPL_CORES)
331 outp += sprintf(outp, " CorWatt");
332 if (do_rapl & RAPL_GFX)
333 outp += sprintf(outp, " GFXWatt");
334 if (do_rapl & RAPL_DRAM)
335 outp += sprintf(outp, " RAMWatt");
336 if (do_rapl & RAPL_PKG_PERF_STATUS)
337 outp += sprintf(outp, " PKG_%%");
338 if (do_rapl & RAPL_DRAM_PERF_STATUS)
339 outp += sprintf(outp, " RAM_%%");
341 if (do_rapl & RAPL_PKG)
342 outp += sprintf(outp, " Pkg_J");
343 if (do_rapl & RAPL_CORES)
344 outp += sprintf(outp, " Cor_J");
345 if (do_rapl & RAPL_GFX)
346 outp += sprintf(outp, " GFX_J");
347 if (do_rapl & RAPL_DRAM)
348 outp += sprintf(outp, " RAM_W");
349 if (do_rapl & RAPL_PKG_PERF_STATUS)
350 outp += sprintf(outp, " PKG_%%");
351 if (do_rapl & RAPL_DRAM_PERF_STATUS)
352 outp += sprintf(outp, " RAM_%%");
353 outp += sprintf(outp, " time");
356 outp += sprintf(outp, "\n");
359 int dump_counters(struct thread_data *t, struct core_data *c,
362 outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
365 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
366 t->cpu_id, t->flags);
367 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
368 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
369 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
370 outp += sprintf(outp, "c1: %016llX\n", t->c1);
371 outp += sprintf(outp, "msr0x%x: %08llX\n",
372 extra_delta_offset32, t->extra_delta32);
373 outp += sprintf(outp, "msr0x%x: %016llX\n",
374 extra_delta_offset64, t->extra_delta64);
375 outp += sprintf(outp, "msr0x%x: %08llX\n",
376 extra_msr_offset32, t->extra_msr32);
377 outp += sprintf(outp, "msr0x%x: %016llX\n",
378 extra_msr_offset64, t->extra_msr64);
380 outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
384 outp += sprintf(outp, "core: %d\n", c->core_id);
385 outp += sprintf(outp, "c3: %016llX\n", c->c3);
386 outp += sprintf(outp, "c6: %016llX\n", c->c6);
387 outp += sprintf(outp, "c7: %016llX\n", c->c7);
388 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
392 outp += sprintf(outp, "package: %d\n", p->package_id);
393 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
394 outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
395 outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
396 outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
397 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
398 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
399 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
400 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
401 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
402 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
403 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
404 outp += sprintf(outp, "Throttle PKG: %0X\n",
405 p->rapl_pkg_perf_status);
406 outp += sprintf(outp, "Throttle RAM: %0X\n",
407 p->rapl_dram_perf_status);
408 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
411 outp += sprintf(outp, "\n");
417 * column formatting convention & formats
419 int format_counters(struct thread_data *t, struct core_data *c,
422 double interval_float;
425 /* if showing only 1st thread in core and this isn't one, bail out */
426 if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
429 /* if showing only 1st thread in pkg and this isn't one, bail out */
430 if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
433 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
435 /* topo columns, print blanks on 1st (average) line */
436 if (t == &average.threads) {
438 outp += sprintf(outp, " -");
440 outp += sprintf(outp, " -");
442 outp += sprintf(outp, " -");
446 outp += sprintf(outp, "%8d", p->package_id);
448 outp += sprintf(outp, " -");
452 outp += sprintf(outp, "%8d", c->core_id);
454 outp += sprintf(outp, " -");
457 outp += sprintf(outp, "%8d", t->cpu_id);
462 outp += sprintf(outp, "%8.0f",
463 1.0 / units * t->aperf / interval_float);
466 if (do_nhm_cstates) {
468 outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc);
470 outp += sprintf(outp, "********");
475 outp += sprintf(outp, "%8.0f",
476 1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
479 outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
483 outp += sprintf(outp, "%8d", t->smi_count);
486 if (extra_delta_offset32)
487 outp += sprintf(outp, " %11llu", t->extra_delta32);
490 if (extra_delta_offset64)
491 outp += sprintf(outp, " %11llu", t->extra_delta64);
493 if (extra_msr_offset32)
494 outp += sprintf(outp, " 0x%08llx", t->extra_msr32);
497 if (extra_msr_offset64)
498 outp += sprintf(outp, " 0x%016llx", t->extra_msr64);
500 if (do_nhm_cstates) {
502 outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
504 outp += sprintf(outp, "********");
507 /* print per-core data only for 1st thread in core */
508 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
511 if (do_nhm_cstates && !do_slm_cstates)
512 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
514 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
516 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
519 outp += sprintf(outp, "%8d", c->core_temp_c);
521 /* print per-package data only for 1st core in package */
522 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
526 outp += sprintf(outp, "%8d", p->pkg_temp_c);
529 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
530 if (do_nhm_cstates && !do_slm_cstates)
531 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
532 if (do_nhm_cstates && !do_slm_cstates)
533 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
535 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
537 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
538 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
539 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
543 * If measurement interval exceeds minimum RAPL Joule Counter range,
544 * indicate that results are suspect by printing "**" in fraction place.
546 if (interval_float < rapl_joule_counter_range)
551 if (do_rapl && !rapl_joules) {
552 if (do_rapl & RAPL_PKG)
553 outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
554 if (do_rapl & RAPL_CORES)
555 outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
556 if (do_rapl & RAPL_GFX)
557 outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
558 if (do_rapl & RAPL_DRAM)
559 outp += sprintf(outp, fmt8, p->energy_dram * rapl_energy_units / interval_float);
560 if (do_rapl & RAPL_PKG_PERF_STATUS)
561 outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
562 if (do_rapl & RAPL_DRAM_PERF_STATUS)
563 outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
565 if (do_rapl & RAPL_PKG)
566 outp += sprintf(outp, fmt8,
567 p->energy_pkg * rapl_energy_units);
568 if (do_rapl & RAPL_CORES)
569 outp += sprintf(outp, fmt8,
570 p->energy_cores * rapl_energy_units);
571 if (do_rapl & RAPL_GFX)
572 outp += sprintf(outp, fmt8,
573 p->energy_gfx * rapl_energy_units);
574 if (do_rapl & RAPL_DRAM)
575 outp += sprintf(outp, fmt8,
576 p->energy_dram * rapl_energy_units);
577 if (do_rapl & RAPL_PKG_PERF_STATUS)
578 outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
579 if (do_rapl & RAPL_DRAM_PERF_STATUS)
580 outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
581 outp += sprintf(outp, fmt8, interval_float);
585 outp += sprintf(outp, "\n");
592 fputs(output_buffer, stdout);
594 outp = output_buffer;
598 fputs(output_buffer, stderr);
599 outp = output_buffer;
601 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
605 if (!printed || !summary_only)
608 if (topo.num_cpus > 1)
609 format_counters(&average.threads, &average.cores,
617 for_all_cpus(format_counters, t, c, p);
620 #define DELTA_WRAP32(new, old) \
624 old = 0x100000000 + new - old; \
628 delta_package(struct pkg_data *new, struct pkg_data *old)
630 old->pc2 = new->pc2 - old->pc2;
631 old->pc3 = new->pc3 - old->pc3;
632 old->pc6 = new->pc6 - old->pc6;
633 old->pc7 = new->pc7 - old->pc7;
634 old->pc8 = new->pc8 - old->pc8;
635 old->pc9 = new->pc9 - old->pc9;
636 old->pc10 = new->pc10 - old->pc10;
637 old->pkg_temp_c = new->pkg_temp_c;
639 DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
640 DELTA_WRAP32(new->energy_cores, old->energy_cores);
641 DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
642 DELTA_WRAP32(new->energy_dram, old->energy_dram);
643 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
644 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
648 delta_core(struct core_data *new, struct core_data *old)
650 old->c3 = new->c3 - old->c3;
651 old->c6 = new->c6 - old->c6;
652 old->c7 = new->c7 - old->c7;
653 old->core_temp_c = new->core_temp_c;
660 delta_thread(struct thread_data *new, struct thread_data *old,
661 struct core_data *core_delta)
663 old->tsc = new->tsc - old->tsc;
665 /* check for TSC < 1 Mcycles over interval */
666 if (old->tsc < (1000 * 1000))
667 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
668 "You can disable all c-states by booting with \"idle=poll\"\n"
669 "or just the deep ones with \"processor.max_cstate=1\"");
671 old->c1 = new->c1 - old->c1;
673 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
674 old->aperf = new->aperf - old->aperf;
675 old->mperf = new->mperf - old->mperf;
678 if (!aperf_mperf_unstable) {
679 fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
680 fprintf(stderr, "* Frequency results do not cover entire interval *\n");
681 fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
683 aperf_mperf_unstable = 1;
686 * mperf delta is likely a huge "positive" number
687 * can not use it for calculating c0 time
694 if (use_c1_residency_msr) {
696 * Some models have a dedicated C1 residency MSR,
697 * which should be more accurate than the derivation below.
701 * As counter collection is not atomic,
702 * it is possible for mperf's non-halted cycles + idle states
703 * to exceed TSC's all cycles: show c1 = 0% in that case.
705 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
708 /* normal case, derive c1 */
709 old->c1 = old->tsc - old->mperf - core_delta->c3
710 - core_delta->c6 - core_delta->c7;
714 if (old->mperf == 0) {
715 if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
716 old->mperf = 1; /* divide by 0 protection */
719 old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
720 old->extra_delta32 &= 0xFFFFFFFF;
722 old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
725 * Extra MSR is just a snapshot, simply copy latest w/o subtracting
727 old->extra_msr32 = new->extra_msr32;
728 old->extra_msr64 = new->extra_msr64;
731 old->smi_count = new->smi_count - old->smi_count;
734 int delta_cpu(struct thread_data *t, struct core_data *c,
735 struct pkg_data *p, struct thread_data *t2,
736 struct core_data *c2, struct pkg_data *p2)
738 /* calculate core delta only for 1st thread in core */
739 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
742 /* always calculate thread delta */
743 delta_thread(t, t2, c2); /* c2 is core delta */
745 /* calculate package delta only for 1st core in package */
746 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
747 delta_package(p, p2);
752 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
760 t->extra_delta32 = 0;
761 t->extra_delta64 = 0;
763 /* tells format_counters to dump all fields from this set */
764 t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
783 p->rapl_pkg_perf_status = 0;
784 p->rapl_dram_perf_status = 0;
787 int sum_counters(struct thread_data *t, struct core_data *c,
790 average.threads.tsc += t->tsc;
791 average.threads.aperf += t->aperf;
792 average.threads.mperf += t->mperf;
793 average.threads.c1 += t->c1;
795 average.threads.extra_delta32 += t->extra_delta32;
796 average.threads.extra_delta64 += t->extra_delta64;
798 /* sum per-core values only for 1st thread in core */
799 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
802 average.cores.c3 += c->c3;
803 average.cores.c6 += c->c6;
804 average.cores.c7 += c->c7;
806 average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
808 /* sum per-pkg values only for 1st core in pkg */
809 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
812 average.packages.pc2 += p->pc2;
813 average.packages.pc3 += p->pc3;
814 average.packages.pc6 += p->pc6;
815 average.packages.pc7 += p->pc7;
816 average.packages.pc8 += p->pc8;
817 average.packages.pc9 += p->pc9;
818 average.packages.pc10 += p->pc10;
820 average.packages.energy_pkg += p->energy_pkg;
821 average.packages.energy_dram += p->energy_dram;
822 average.packages.energy_cores += p->energy_cores;
823 average.packages.energy_gfx += p->energy_gfx;
825 average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
827 average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
828 average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
832 * sum the counters for all cpus in the system
833 * compute the weighted average
835 void compute_average(struct thread_data *t, struct core_data *c,
838 clear_counters(&average.threads, &average.cores, &average.packages);
840 for_all_cpus(sum_counters, t, c, p);
842 average.threads.tsc /= topo.num_cpus;
843 average.threads.aperf /= topo.num_cpus;
844 average.threads.mperf /= topo.num_cpus;
845 average.threads.c1 /= topo.num_cpus;
847 average.threads.extra_delta32 /= topo.num_cpus;
848 average.threads.extra_delta32 &= 0xFFFFFFFF;
850 average.threads.extra_delta64 /= topo.num_cpus;
852 average.cores.c3 /= topo.num_cores;
853 average.cores.c6 /= topo.num_cores;
854 average.cores.c7 /= topo.num_cores;
856 average.packages.pc2 /= topo.num_packages;
857 average.packages.pc3 /= topo.num_packages;
858 average.packages.pc6 /= topo.num_packages;
859 average.packages.pc7 /= topo.num_packages;
861 average.packages.pc8 /= topo.num_packages;
862 average.packages.pc9 /= topo.num_packages;
863 average.packages.pc10 /= topo.num_packages;
866 static unsigned long long rdtsc(void)
868 unsigned int low, high;
870 asm volatile("rdtsc" : "=a" (low), "=d" (high));
872 return low | ((unsigned long long)high) << 32;
879 * acquire and record local counters for that cpu
881 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
884 unsigned long long msr;
886 if (cpu_migrate(cpu)) {
887 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
891 t->tsc = rdtsc(); /* we are running on local CPU of interest */
894 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
896 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
901 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
903 t->smi_count = msr & 0xFFFFFFFF;
905 if (extra_delta_offset32) {
906 if (get_msr(cpu, extra_delta_offset32, &msr))
908 t->extra_delta32 = msr & 0xFFFFFFFF;
911 if (extra_delta_offset64)
912 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
915 if (extra_msr_offset32) {
916 if (get_msr(cpu, extra_msr_offset32, &msr))
918 t->extra_msr32 = msr & 0xFFFFFFFF;
921 if (extra_msr_offset64)
922 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
925 if (use_c1_residency_msr) {
926 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
930 /* collect core counters only for 1st thread in core */
931 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
934 if (do_nhm_cstates && !do_slm_cstates) {
935 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
939 if (do_nhm_cstates) {
940 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
945 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
949 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
951 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
955 /* collect package counters only for 1st core in package */
956 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
959 if (do_nhm_cstates && !do_slm_cstates) {
960 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
962 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
965 if (do_snb_cstates) {
966 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
968 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
972 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
974 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
976 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
979 if (do_rapl & RAPL_PKG) {
980 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
982 p->energy_pkg = msr & 0xFFFFFFFF;
984 if (do_rapl & RAPL_CORES) {
985 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
987 p->energy_cores = msr & 0xFFFFFFFF;
989 if (do_rapl & RAPL_DRAM) {
990 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
992 p->energy_dram = msr & 0xFFFFFFFF;
994 if (do_rapl & RAPL_GFX) {
995 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
997 p->energy_gfx = msr & 0xFFFFFFFF;
999 if (do_rapl & RAPL_PKG_PERF_STATUS) {
1000 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1002 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1004 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1005 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1007 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1010 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1012 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1017 void print_verbose_header(void)
1019 unsigned long long msr;
1022 if (!do_nehalem_platform_info)
1025 get_msr(0, MSR_NHM_PLATFORM_INFO, &msr);
1027 fprintf(stderr, "cpu0: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", msr);
1029 ratio = (msr >> 40) & 0xFF;
1030 fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
1031 ratio, bclk, ratio * bclk);
1033 ratio = (msr >> 8) & 0xFF;
1034 fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
1035 ratio, bclk, ratio * bclk);
1037 get_msr(0, MSR_IA32_POWER_CTL, &msr);
1038 fprintf(stderr, "cpu0: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1039 msr, msr & 0x2 ? "EN" : "DIS");
1041 if (!do_ivt_turbo_ratio_limit)
1042 goto print_nhm_turbo_ratio_limits;
1044 get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr);
1046 fprintf(stderr, "cpu0: MSR_IVT_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1048 ratio = (msr >> 56) & 0xFF;
1050 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1051 ratio, bclk, ratio * bclk);
1053 ratio = (msr >> 48) & 0xFF;
1055 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1056 ratio, bclk, ratio * bclk);
1058 ratio = (msr >> 40) & 0xFF;
1060 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1061 ratio, bclk, ratio * bclk);
1063 ratio = (msr >> 32) & 0xFF;
1065 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1066 ratio, bclk, ratio * bclk);
1068 ratio = (msr >> 24) & 0xFF;
1070 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1071 ratio, bclk, ratio * bclk);
1073 ratio = (msr >> 16) & 0xFF;
1075 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1076 ratio, bclk, ratio * bclk);
1078 ratio = (msr >> 8) & 0xFF;
1080 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1081 ratio, bclk, ratio * bclk);
1083 ratio = (msr >> 0) & 0xFF;
1085 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1086 ratio, bclk, ratio * bclk);
1088 print_nhm_turbo_ratio_limits:
1089 get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1091 #define SNB_C1_AUTO_UNDEMOTE (1UL << 27)
1092 #define SNB_C3_AUTO_UNDEMOTE (1UL << 28)
1094 fprintf(stderr, "cpu0: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", msr);
1096 fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: ",
1097 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1098 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1099 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1100 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1101 (msr & (1 << 15)) ? "" : "UN",
1102 (unsigned int)msr & 7);
1107 fprintf(stderr, do_slm_cstates ? "no pkg states" : "pc0");
1110 fprintf(stderr, do_slm_cstates ? "no pkg states" : do_snb_cstates ? "pc2" : "pc0");
1113 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc6-noret" : "pc3");
1116 fprintf(stderr, do_slm_cstates ? "invalid" : "pc6");
1119 fprintf(stderr, do_slm_cstates ? "pc4" : "pc7");
1122 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc7s" : "invalid");
1125 fprintf(stderr, do_slm_cstates ? "pc6" : "invalid");
1128 fprintf(stderr, do_slm_cstates ? "pc7" : "unlimited");
1131 fprintf(stderr, "invalid");
1133 fprintf(stderr, ")\n");
1135 if (!do_nehalem_turbo_ratio_limit)
1138 get_msr(0, MSR_NHM_TURBO_RATIO_LIMIT, &msr);
1140 fprintf(stderr, "cpu0: MSR_NHM_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1142 ratio = (msr >> 56) & 0xFF;
1144 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1145 ratio, bclk, ratio * bclk);
1147 ratio = (msr >> 48) & 0xFF;
1149 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1150 ratio, bclk, ratio * bclk);
1152 ratio = (msr >> 40) & 0xFF;
1154 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1155 ratio, bclk, ratio * bclk);
1157 ratio = (msr >> 32) & 0xFF;
1159 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1160 ratio, bclk, ratio * bclk);
1162 ratio = (msr >> 24) & 0xFF;
1164 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1165 ratio, bclk, ratio * bclk);
1167 ratio = (msr >> 16) & 0xFF;
1169 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1170 ratio, bclk, ratio * bclk);
1172 ratio = (msr >> 8) & 0xFF;
1174 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1175 ratio, bclk, ratio * bclk);
1177 ratio = (msr >> 0) & 0xFF;
1179 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1180 ratio, bclk, ratio * bclk);
1183 void free_all_buffers(void)
1185 CPU_FREE(cpu_present_set);
1186 cpu_present_set = NULL;
1187 cpu_present_set = 0;
1189 CPU_FREE(cpu_affinity_set);
1190 cpu_affinity_set = NULL;
1191 cpu_affinity_setsize = 0;
1199 package_even = NULL;
1209 free(output_buffer);
1210 output_buffer = NULL;
1215 * Open a file, and exit on failure
1217 FILE *fopen_or_die(const char *path, const char *mode)
1219 FILE *filep = fopen(path, "r");
1221 err(1, "%s: open failed", path);
1226 * Parse a file containing a single int.
1228 int parse_int_file(const char *fmt, ...)
1231 char path[PATH_MAX];
1235 va_start(args, fmt);
1236 vsnprintf(path, sizeof(path), fmt, args);
1238 filep = fopen_or_die(path, "r");
1239 if (fscanf(filep, "%d", &value) != 1)
1240 err(1, "%s: failed to parse number from file", path);
1246 * cpu_is_first_sibling_in_core(cpu)
1247 * return 1 if given CPU is 1st HT sibling in the core
1249 int cpu_is_first_sibling_in_core(int cpu)
1251 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1255 * cpu_is_first_core_in_package(cpu)
1256 * return 1 if given CPU is 1st core in package
1258 int cpu_is_first_core_in_package(int cpu)
1260 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1263 int get_physical_package_id(int cpu)
1265 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1268 int get_core_id(int cpu)
1270 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1273 int get_num_ht_siblings(int cpu)
1281 sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1282 filep = fopen_or_die(path, "r");
1285 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1286 * otherwinse 1 sibling (self).
1288 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1299 * run func(thread, core, package) in topology order
1300 * skip non-present cpus
1303 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1304 struct pkg_data *, struct thread_data *, struct core_data *,
1305 struct pkg_data *), struct thread_data *thread_base,
1306 struct core_data *core_base, struct pkg_data *pkg_base,
1307 struct thread_data *thread_base2, struct core_data *core_base2,
1308 struct pkg_data *pkg_base2)
1310 int retval, pkg_no, core_no, thread_no;
1312 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1313 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1314 for (thread_no = 0; thread_no <
1315 topo.num_threads_per_core; ++thread_no) {
1316 struct thread_data *t, *t2;
1317 struct core_data *c, *c2;
1318 struct pkg_data *p, *p2;
1320 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1322 if (cpu_is_not_present(t->cpu_id))
1325 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1327 c = GET_CORE(core_base, core_no, pkg_no);
1328 c2 = GET_CORE(core_base2, core_no, pkg_no);
1330 p = GET_PKG(pkg_base, pkg_no);
1331 p2 = GET_PKG(pkg_base2, pkg_no);
1333 retval = func(t, c, p, t2, c2, p2);
1343 * run func(cpu) on every cpu in /proc/stat
1344 * return max_cpu number
1346 int for_all_proc_cpus(int (func)(int))
1352 fp = fopen_or_die(proc_stat, "r");
1354 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1356 err(1, "%s: failed to parse format", proc_stat);
1359 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1363 retval = func(cpu_num);
1373 void re_initialize(void)
1376 setup_all_buffers();
1377 printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1383 * remember the last one seen, it will be the max
1385 int count_cpus(int cpu)
1387 if (topo.max_cpu_num < cpu)
1388 topo.max_cpu_num = cpu;
1393 int mark_cpu_present(int cpu)
1395 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1399 void turbostat_loop()
1407 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1410 } else if (retval == -1) {
1411 if (restarted > 1) {
1418 gettimeofday(&tv_even, (struct timezone *)NULL);
1421 if (for_all_proc_cpus(cpu_is_not_present)) {
1425 sleep(interval_sec);
1426 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1429 } else if (retval == -1) {
1433 gettimeofday(&tv_odd, (struct timezone *)NULL);
1434 timersub(&tv_odd, &tv_even, &tv_delta);
1435 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1436 compute_average(EVEN_COUNTERS);
1437 format_all_counters(EVEN_COUNTERS);
1439 sleep(interval_sec);
1440 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1443 } else if (retval == -1) {
1447 gettimeofday(&tv_even, (struct timezone *)NULL);
1448 timersub(&tv_even, &tv_odd, &tv_delta);
1449 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1450 compute_average(ODD_COUNTERS);
1451 format_all_counters(ODD_COUNTERS);
1456 void check_dev_msr()
1460 if (stat("/dev/cpu/0/msr", &sb))
1461 err(-5, "no /dev/cpu/0/msr\n"
1462 "Try \"# modprobe msr\"");
1465 void check_permissions()
1467 struct __user_cap_header_struct cap_header_data;
1468 cap_user_header_t cap_header = &cap_header_data;
1469 struct __user_cap_data_struct cap_data_data;
1470 cap_user_data_t cap_data = &cap_data_data;
1471 extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
1474 /* check for CAP_SYS_RAWIO */
1475 cap_header->pid = getpid();
1476 cap_header->version = _LINUX_CAPABILITY_VERSION;
1477 if (capget(cap_header, cap_data) < 0)
1478 err(-6, "capget(2) failed");
1480 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1482 warnx("capget(CAP_SYS_RAWIO) failed,"
1483 " try \"# setcap cap_sys_rawio=ep %s\"", progname);
1486 /* test file permissions */
1487 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1489 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1492 /* if all else fails, thell them to be root */
1495 warnx("Or simply run as root");
1501 int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model)
1510 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1511 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1512 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1513 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1514 case 0x2C: /* Westmere EP - Gulftown */
1515 case 0x2A: /* SNB */
1516 case 0x2D: /* SNB Xeon */
1517 case 0x3A: /* IVB */
1518 case 0x3E: /* IVB Xeon */
1519 case 0x3C: /* HSW */
1520 case 0x3F: /* HSX */
1521 case 0x45: /* HSW */
1522 case 0x46: /* HSW */
1523 case 0x37: /* BYT */
1524 case 0x4D: /* AVN */
1525 case 0x3D: /* BDW */
1526 case 0x4F: /* BDX */
1527 case 0x56: /* BDX-DE */
1529 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1530 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1535 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1544 case 0x3E: /* IVB Xeon */
1553 * Decode the ENERGY_PERF_BIAS MSR
1555 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1557 unsigned long long msr;
1566 /* EPB is per-package */
1567 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1570 if (cpu_migrate(cpu)) {
1571 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1575 if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
1578 switch (msr & 0x7) {
1579 case ENERGY_PERF_BIAS_PERFORMANCE:
1580 epb_string = "performance";
1582 case ENERGY_PERF_BIAS_NORMAL:
1583 epb_string = "balanced";
1585 case ENERGY_PERF_BIAS_POWERSAVE:
1586 epb_string = "powersave";
1589 epb_string = "custom";
1592 fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
1597 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
1598 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
1600 double get_tdp(model)
1602 unsigned long long msr;
1604 if (do_rapl & RAPL_PKG_POWER_INFO)
1605 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
1606 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
1621 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
1623 void rapl_probe(unsigned int family, unsigned int model)
1625 unsigned long long msr;
1626 unsigned int time_unit;
1638 case 0x3C: /* HSW */
1639 case 0x45: /* HSW */
1640 case 0x46: /* HSW */
1641 case 0x3D: /* BDW */
1642 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1644 case 0x3F: /* HSX */
1645 case 0x4F: /* BDX */
1646 case 0x56: /* BDX-DE */
1647 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1651 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1653 case 0x37: /* BYT */
1654 case 0x4D: /* AVN */
1655 do_rapl = RAPL_PKG | RAPL_CORES ;
1661 /* units on package 0, verify later other packages match */
1662 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1665 rapl_power_units = 1.0 / (1 << (msr & 0xF));
1667 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1669 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1671 time_unit = msr >> 16 & 0xF;
1675 rapl_time_units = 1.0 / (1 << (time_unit));
1677 tdp = get_tdp(model);
1679 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1681 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1686 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1688 unsigned long long msr;
1692 if (!(do_dts || do_ptm))
1697 /* DTS is per-core, no need to print for each thread */
1698 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1701 if (cpu_migrate(cpu)) {
1702 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1706 if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
1707 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1710 dts = (msr >> 16) & 0x7F;
1711 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
1712 cpu, msr, tcc_activation_temp - dts);
1715 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
1718 dts = (msr >> 16) & 0x7F;
1719 dts2 = (msr >> 8) & 0x7F;
1720 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1721 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1727 unsigned int resolution;
1729 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1732 dts = (msr >> 16) & 0x7F;
1733 resolution = (msr >> 27) & 0xF;
1734 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
1735 cpu, msr, tcc_activation_temp - dts, resolution);
1738 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
1741 dts = (msr >> 16) & 0x7F;
1742 dts2 = (msr >> 8) & 0x7F;
1743 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1744 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1751 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
1753 fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
1755 ((msr >> 15) & 1) ? "EN" : "DIS",
1756 ((msr >> 0) & 0x7FFF) * rapl_power_units,
1757 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
1758 (((msr >> 16) & 1) ? "EN" : "DIS"));
1763 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1765 unsigned long long msr;
1771 /* RAPL counters are per package, so print only for 1st thread/package */
1772 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1776 if (cpu_migrate(cpu)) {
1777 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1781 if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
1785 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
1786 "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
1787 rapl_power_units, rapl_energy_units, rapl_time_units);
1789 if (do_rapl & RAPL_PKG_POWER_INFO) {
1791 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
1795 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1797 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1798 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1799 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1800 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1803 if (do_rapl & RAPL_PKG) {
1805 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
1808 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
1809 cpu, msr, (msr >> 63) & 1 ? "": "UN");
1811 print_power_limit_msr(cpu, msr, "PKG Limit #1");
1812 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
1814 ((msr >> 47) & 1) ? "EN" : "DIS",
1815 ((msr >> 32) & 0x7FFF) * rapl_power_units,
1816 (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
1817 ((msr >> 48) & 1) ? "EN" : "DIS");
1820 if (do_rapl & RAPL_DRAM) {
1821 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
1825 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1827 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1828 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1829 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1830 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1833 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
1835 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
1836 cpu, msr, (msr >> 31) & 1 ? "": "UN");
1838 print_power_limit_msr(cpu, msr, "DRAM Limit");
1840 if (do_rapl & RAPL_CORE_POLICY) {
1842 if (get_msr(cpu, MSR_PP0_POLICY, &msr))
1845 fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
1848 if (do_rapl & RAPL_CORES) {
1851 if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
1853 fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
1854 cpu, msr, (msr >> 31) & 1 ? "": "UN");
1855 print_power_limit_msr(cpu, msr, "Cores Limit");
1858 if (do_rapl & RAPL_GFX) {
1860 if (get_msr(cpu, MSR_PP1_POLICY, &msr))
1863 fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
1865 if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
1867 fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
1868 cpu, msr, (msr >> 31) & 1 ? "": "UN");
1869 print_power_limit_msr(cpu, msr, "GFX Limit");
1876 int is_snb(unsigned int family, unsigned int model)
1884 case 0x3A: /* IVB */
1885 case 0x3E: /* IVB Xeon */
1886 case 0x3C: /* HSW */
1887 case 0x3F: /* HSW */
1888 case 0x45: /* HSW */
1889 case 0x46: /* HSW */
1890 case 0x3D: /* BDW */
1891 case 0x4F: /* BDX */
1892 case 0x56: /* BDX-DE */
1898 int has_c8_c9_c10(unsigned int family, unsigned int model)
1904 case 0x45: /* HSW */
1905 case 0x3D: /* BDW */
1912 int is_slm(unsigned int family, unsigned int model)
1917 case 0x37: /* BYT */
1918 case 0x4D: /* AVN */
1924 #define SLM_BCLK_FREQS 5
1925 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
1927 double slm_bclk(void)
1929 unsigned long long msr = 3;
1933 if (get_msr(0, MSR_FSB_FREQ, &msr))
1934 fprintf(stderr, "SLM BCLK: unknown\n");
1937 if (i >= SLM_BCLK_FREQS) {
1938 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
1941 freq = slm_freq_table[i];
1943 fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
1948 double discover_bclk(unsigned int family, unsigned int model)
1950 if (is_snb(family, model))
1952 else if (is_slm(family, model))
1959 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1960 * the Thermal Control Circuit (TCC) activates.
1961 * This is usually equal to tjMax.
1963 * Older processors do not have this MSR, so there we guess,
1964 * but also allow cmdline over-ride with -T.
1966 * Several MSR temperature values are in units of degrees-C
1967 * below this value, including the Digital Thermal Sensor (DTS),
1968 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1970 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1972 unsigned long long msr;
1973 unsigned int target_c_local;
1976 /* tcc_activation_temp is used only for dts or ptm */
1977 if (!(do_dts || do_ptm))
1980 /* this is a per-package concept */
1981 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1985 if (cpu_migrate(cpu)) {
1986 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1990 if (tcc_activation_temp_override != 0) {
1991 tcc_activation_temp = tcc_activation_temp_override;
1992 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
1993 cpu, tcc_activation_temp);
1997 /* Temperature Target MSR is Nehalem and newer only */
1998 if (!do_nehalem_platform_info)
2001 if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
2004 target_c_local = (msr >> 16) & 0xFF;
2007 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2008 cpu, msr, target_c_local);
2010 if (!target_c_local)
2013 tcc_activation_temp = target_c_local;
2018 tcc_activation_temp = TJMAX_DEFAULT;
2019 fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2020 cpu, tcc_activation_temp);
2026 unsigned int eax, ebx, ecx, edx, max_level;
2027 unsigned int fms, family, model, stepping;
2029 eax = ebx = ecx = edx = 0;
2031 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2033 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2037 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2038 (char *)&ebx, (char *)&edx, (char *)&ecx);
2040 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2041 family = (fms >> 8) & 0xf;
2042 model = (fms >> 4) & 0xf;
2043 stepping = fms & 0xf;
2044 if (family == 6 || family == 0xf)
2045 model += ((fms >> 16) & 0xf) << 4;
2048 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2049 max_level, family, model, stepping, family, model, stepping);
2051 if (!(edx & (1 << 5)))
2052 errx(1, "CPUID: no MSR");
2055 * check max extended function levels of CPUID.
2056 * This is needed to check for invariant TSC.
2057 * This check is valid for both Intel and AMD.
2059 ebx = ecx = edx = 0;
2060 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2062 if (max_level < 0x80000007)
2063 errx(1, "CPUID: no invariant TSC (max_level 0x%x)", max_level);
2066 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2067 * this check is valid for both Intel and AMD
2069 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2070 has_invariant_tsc = edx & (1 << 8);
2072 if (!has_invariant_tsc)
2073 errx(1, "No invariant TSC");
2076 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2077 * this check is valid for both Intel and AMD
2080 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2081 has_aperf = ecx & (1 << 0);
2082 do_dts = eax & (1 << 0);
2083 do_ptm = eax & (1 << 6);
2084 has_epb = ecx & (1 << 3);
2087 fprintf(stderr, "CPUID(6): %s%s%s%s\n",
2088 has_aperf ? "APERF" : "No APERF!",
2089 do_dts ? ", DTS" : "",
2090 do_ptm ? ", PTM": "",
2091 has_epb ? ", EPB": "");
2094 errx(-1, "No APERF");
2096 do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
2097 do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */
2098 do_smi = do_nhm_cstates;
2099 do_snb_cstates = is_snb(family, model);
2100 do_c8_c9_c10 = has_c8_c9_c10(family, model);
2101 do_slm_cstates = is_slm(family, model);
2102 bclk = discover_bclk(family, model);
2104 do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model);
2105 do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model);
2106 rapl_probe(family, model);
2114 errx(1, "%s: [-v][-R][-T][-p|-P|-S][-c MSR#][-C MSR#][-m MSR#][-M MSR#][-i interval_sec | command ...]\n",
2120 * in /dev/cpu/ return success for names that are numbers
2121 * ie. filter out ".", "..", "microcode".
2123 int dir_filter(const struct dirent *dirp)
2125 if (isdigit(dirp->d_name[0]))
2131 int open_dev_cpu_msr(int dummy1)
2136 void topology_probe()
2139 int max_core_id = 0;
2140 int max_package_id = 0;
2141 int max_siblings = 0;
2142 struct cpu_topology {
2144 int physical_package_id;
2147 /* Initialize num_cpus, max_cpu_num */
2149 topo.max_cpu_num = 0;
2150 for_all_proc_cpus(count_cpus);
2151 if (!summary_only && topo.num_cpus > 1)
2155 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2157 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
2159 err(1, "calloc cpus");
2162 * Allocate and initialize cpu_present_set
2164 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2165 if (cpu_present_set == NULL)
2166 err(3, "CPU_ALLOC");
2167 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2168 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2169 for_all_proc_cpus(mark_cpu_present);
2172 * Allocate and initialize cpu_affinity_set
2174 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2175 if (cpu_affinity_set == NULL)
2176 err(3, "CPU_ALLOC");
2177 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2178 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2183 * find max_core_id, max_package_id
2185 for (i = 0; i <= topo.max_cpu_num; ++i) {
2188 if (cpu_is_not_present(i)) {
2190 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2193 cpus[i].core_id = get_core_id(i);
2194 if (cpus[i].core_id > max_core_id)
2195 max_core_id = cpus[i].core_id;
2197 cpus[i].physical_package_id = get_physical_package_id(i);
2198 if (cpus[i].physical_package_id > max_package_id)
2199 max_package_id = cpus[i].physical_package_id;
2201 siblings = get_num_ht_siblings(i);
2202 if (siblings > max_siblings)
2203 max_siblings = siblings;
2205 fprintf(stderr, "cpu %d pkg %d core %d\n",
2206 i, cpus[i].physical_package_id, cpus[i].core_id);
2208 topo.num_cores_per_pkg = max_core_id + 1;
2210 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2211 max_core_id, topo.num_cores_per_pkg);
2212 if (!summary_only && topo.num_cores_per_pkg > 1)
2215 topo.num_packages = max_package_id + 1;
2217 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2218 max_package_id, topo.num_packages);
2219 if (!summary_only && topo.num_packages > 1)
2222 topo.num_threads_per_core = max_siblings;
2224 fprintf(stderr, "max_siblings %d\n", max_siblings);
2230 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2234 *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2235 topo.num_packages, sizeof(struct thread_data));
2239 for (i = 0; i < topo.num_threads_per_core *
2240 topo.num_cores_per_pkg * topo.num_packages; i++)
2241 (*t)[i].cpu_id = -1;
2243 *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2244 sizeof(struct core_data));
2248 for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2249 (*c)[i].core_id = -1;
2251 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2255 for (i = 0; i < topo.num_packages; i++)
2256 (*p)[i].package_id = i;
2260 err(1, "calloc counters");
2265 * set cpu_id, core_num, pkg_num
2266 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2268 * increment topo.num_cores when 1st core in pkg seen
2270 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2271 struct pkg_data *pkg_base, int thread_num, int core_num,
2272 int pkg_num, int cpu_id)
2274 struct thread_data *t;
2275 struct core_data *c;
2278 t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2279 c = GET_CORE(core_base, core_num, pkg_num);
2280 p = GET_PKG(pkg_base, pkg_num);
2283 if (thread_num == 0) {
2284 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2285 if (cpu_is_first_core_in_package(cpu_id))
2286 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2289 c->core_id = core_num;
2290 p->package_id = pkg_num;
2294 int initialize_counters(int cpu_id)
2296 int my_thread_id, my_core_id, my_package_id;
2298 my_package_id = get_physical_package_id(cpu_id);
2299 my_core_id = get_core_id(cpu_id);
2301 if (cpu_is_first_sibling_in_core(cpu_id)) {
2308 init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2309 init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2313 void allocate_output_buffer()
2315 output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
2316 outp = output_buffer;
2318 err(-1, "calloc output buffer");
2321 void setup_all_buffers(void)
2324 allocate_counters(&thread_even, &core_even, &package_even);
2325 allocate_counters(&thread_odd, &core_odd, &package_odd);
2326 allocate_output_buffer();
2327 for_all_proc_cpus(initialize_counters);
2330 void turbostat_init()
2333 check_permissions();
2336 setup_all_buffers();
2339 print_verbose_header();
2342 for_all_cpus(print_epb, ODD_COUNTERS);
2345 for_all_cpus(print_rapl, ODD_COUNTERS);
2347 for_all_cpus(set_temperature_target, ODD_COUNTERS);
2350 for_all_cpus(print_thermal, ODD_COUNTERS);
2353 int fork_it(char **argv)
2358 status = for_all_cpus(get_counters, EVEN_COUNTERS);
2361 /* clear affinity side-effect of get_counters() */
2362 sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
2363 gettimeofday(&tv_even, (struct timezone *)NULL);
2368 execvp(argv[0], argv);
2372 if (child_pid == -1)
2375 signal(SIGINT, SIG_IGN);
2376 signal(SIGQUIT, SIG_IGN);
2377 if (waitpid(child_pid, &status, 0) == -1)
2378 err(status, "waitpid");
2381 * n.b. fork_it() does not check for errors from for_all_cpus()
2382 * because re-starting is problematic when forking
2384 for_all_cpus(get_counters, ODD_COUNTERS);
2385 gettimeofday(&tv_odd, (struct timezone *)NULL);
2386 timersub(&tv_odd, &tv_even, &tv_delta);
2387 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
2388 compute_average(EVEN_COUNTERS);
2389 format_all_counters(EVEN_COUNTERS);
2392 fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
2397 int get_and_dump_counters(void)
2401 status = for_all_cpus(get_counters, ODD_COUNTERS);
2405 status = for_all_cpus(dump_counters, ODD_COUNTERS);
2414 void cmdline(int argc, char **argv)
2420 while ((opt = getopt(argc, argv, "+pPsSvi:c:C:m:M:RJT:")) != -1) {
2438 interval_sec = atoi(optarg);
2441 sscanf(optarg, "%x", &extra_delta_offset32);
2444 sscanf(optarg, "%x", &extra_delta_offset64);
2447 sscanf(optarg, "%x", &extra_msr_offset32);
2450 sscanf(optarg, "%x", &extra_msr_offset64);
2456 tcc_activation_temp_override = atoi(optarg);
2468 int main(int argc, char **argv)
2470 cmdline(argc, argv);
2473 fprintf(stderr, "turbostat v3.8 14-Aug 2014"
2474 " - Len Brown <lenb@kernel.org>\n");
2478 /* dump counters and exit */
2480 return get_and_dump_counters();
2483 * if any params left, it must be a command to fork
2486 return fork_it(argv + optind);