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>
42 #include <linux/capability.h>
45 char *proc_stat = "/proc/stat";
46 unsigned int interval_sec = 5;
48 unsigned int rapl_joules;
49 unsigned int summary_only;
50 unsigned int dump_only;
53 unsigned int do_nhm_cstates;
54 unsigned int do_snb_cstates;
59 unsigned int do_c8_c9_c10;
60 unsigned int do_slm_cstates;
61 unsigned int use_c1_residency_msr;
62 unsigned int has_aperf;
64 unsigned int units = 1000000; /* MHz etc */
65 unsigned int genuine_intel;
66 unsigned int has_invariant_tsc;
67 unsigned int do_nhm_platform_info;
68 unsigned int extra_msr_offset32;
69 unsigned int extra_msr_offset64;
70 unsigned int extra_delta_offset32;
71 unsigned int extra_delta_offset64;
74 unsigned int show_pkg;
75 unsigned int show_core;
76 unsigned int show_cpu;
77 unsigned int show_pkg_only;
78 unsigned int show_core_only;
79 char *output_buffer, *outp;
83 unsigned int tcc_activation_temp;
84 unsigned int tcc_activation_temp_override;
85 double rapl_power_units, rapl_energy_units, rapl_time_units;
86 double rapl_joule_counter_range;
87 unsigned int do_core_perf_limit_reasons;
88 unsigned int do_gfx_perf_limit_reasons;
89 unsigned int do_ring_perf_limit_reasons;
91 #define RAPL_PKG (1 << 0)
92 /* 0x610 MSR_PKG_POWER_LIMIT */
93 /* 0x611 MSR_PKG_ENERGY_STATUS */
94 #define RAPL_PKG_PERF_STATUS (1 << 1)
95 /* 0x613 MSR_PKG_PERF_STATUS */
96 #define RAPL_PKG_POWER_INFO (1 << 2)
97 /* 0x614 MSR_PKG_POWER_INFO */
99 #define RAPL_DRAM (1 << 3)
100 /* 0x618 MSR_DRAM_POWER_LIMIT */
101 /* 0x619 MSR_DRAM_ENERGY_STATUS */
102 /* 0x61c MSR_DRAM_POWER_INFO */
103 #define RAPL_DRAM_PERF_STATUS (1 << 4)
104 /* 0x61b MSR_DRAM_PERF_STATUS */
106 #define RAPL_CORES (1 << 5)
107 /* 0x638 MSR_PP0_POWER_LIMIT */
108 /* 0x639 MSR_PP0_ENERGY_STATUS */
109 #define RAPL_CORE_POLICY (1 << 6)
110 /* 0x63a MSR_PP0_POLICY */
113 #define RAPL_GFX (1 << 7)
114 /* 0x640 MSR_PP1_POWER_LIMIT */
115 /* 0x641 MSR_PP1_ENERGY_STATUS */
116 /* 0x642 MSR_PP1_POLICY */
117 #define TJMAX_DEFAULT 100
119 #define MAX(a, b) ((a) > (b) ? (a) : (b))
121 int aperf_mperf_unstable;
125 cpu_set_t *cpu_present_set, *cpu_affinity_set;
126 size_t cpu_present_setsize, cpu_affinity_setsize;
129 unsigned long long tsc;
130 unsigned long long aperf;
131 unsigned long long mperf;
132 unsigned long long c1;
133 unsigned long long extra_msr64;
134 unsigned long long extra_delta64;
135 unsigned long long extra_msr32;
136 unsigned long long extra_delta32;
137 unsigned int smi_count;
140 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
141 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
142 } *thread_even, *thread_odd;
145 unsigned long long c3;
146 unsigned long long c6;
147 unsigned long long c7;
148 unsigned int core_temp_c;
149 unsigned int core_id;
150 } *core_even, *core_odd;
153 unsigned long long pc2;
154 unsigned long long pc3;
155 unsigned long long pc6;
156 unsigned long long pc7;
157 unsigned long long pc8;
158 unsigned long long pc9;
159 unsigned long long pc10;
160 unsigned int package_id;
161 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
162 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
163 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
164 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
165 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
166 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
167 unsigned int pkg_temp_c;
169 } *package_even, *package_odd;
171 #define ODD_COUNTERS thread_odd, core_odd, package_odd
172 #define EVEN_COUNTERS thread_even, core_even, package_even
174 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
175 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
176 topo.num_threads_per_core + \
177 (core_no) * topo.num_threads_per_core + (thread_no))
178 #define GET_CORE(core_base, core_no, pkg_no) \
179 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
180 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
182 struct system_summary {
183 struct thread_data threads;
184 struct core_data cores;
185 struct pkg_data packages;
194 int num_cores_per_pkg;
195 int num_threads_per_core;
198 struct timeval tv_even, tv_odd, tv_delta;
200 void setup_all_buffers(void);
202 int cpu_is_not_present(int cpu)
204 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
207 * run func(thread, core, package) in topology order
208 * skip non-present cpus
211 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
212 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
214 int retval, pkg_no, core_no, thread_no;
216 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
217 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
218 for (thread_no = 0; thread_no <
219 topo.num_threads_per_core; ++thread_no) {
220 struct thread_data *t;
224 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
226 if (cpu_is_not_present(t->cpu_id))
229 c = GET_CORE(core_base, core_no, pkg_no);
230 p = GET_PKG(pkg_base, pkg_no);
232 retval = func(t, c, p);
241 int cpu_migrate(int cpu)
243 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
244 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
245 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
251 int get_msr(int cpu, off_t offset, unsigned long long *msr)
257 sprintf(pathname, "/dev/cpu/%d/msr", cpu);
258 fd = open(pathname, O_RDONLY);
260 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
262 retval = pread(fd, msr, sizeof *msr, offset);
265 if (retval != sizeof *msr)
266 err(-1, "%s offset 0x%llx read failed", pathname, (unsigned long long)offset);
272 * Example Format w/ field column widths:
274 * 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
275 * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
278 void print_header(void)
281 outp += sprintf(outp, " Package");
283 outp += sprintf(outp, " Core");
285 outp += sprintf(outp, " CPU");
287 outp += sprintf(outp, " Avg_MHz");
289 outp += sprintf(outp, " %%Busy");
291 outp += sprintf(outp, " Bzy_MHz");
292 outp += sprintf(outp, " TSC_MHz");
294 if (extra_delta_offset32)
295 outp += sprintf(outp, " count 0x%03X", extra_delta_offset32);
296 if (extra_delta_offset64)
297 outp += sprintf(outp, " COUNT 0x%03X", extra_delta_offset64);
298 if (extra_msr_offset32)
299 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset32);
300 if (extra_msr_offset64)
301 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset64);
307 outp += sprintf(outp, " SMI");
310 outp += sprintf(outp, " CPU%%c1");
311 if (do_nhm_cstates && !do_slm_cstates)
312 outp += sprintf(outp, " CPU%%c3");
314 outp += sprintf(outp, " CPU%%c6");
316 outp += sprintf(outp, " CPU%%c7");
319 outp += sprintf(outp, " CoreTmp");
321 outp += sprintf(outp, " PkgTmp");
324 outp += sprintf(outp, " Pkg%%pc2");
326 outp += sprintf(outp, " Pkg%%pc3");
328 outp += sprintf(outp, " Pkg%%pc6");
330 outp += sprintf(outp, " Pkg%%pc7");
332 outp += sprintf(outp, " Pkg%%pc8");
333 outp += sprintf(outp, " Pkg%%pc9");
334 outp += sprintf(outp, " Pk%%pc10");
337 if (do_rapl && !rapl_joules) {
338 if (do_rapl & RAPL_PKG)
339 outp += sprintf(outp, " PkgWatt");
340 if (do_rapl & RAPL_CORES)
341 outp += sprintf(outp, " CorWatt");
342 if (do_rapl & RAPL_GFX)
343 outp += sprintf(outp, " GFXWatt");
344 if (do_rapl & RAPL_DRAM)
345 outp += sprintf(outp, " RAMWatt");
346 if (do_rapl & RAPL_PKG_PERF_STATUS)
347 outp += sprintf(outp, " PKG_%%");
348 if (do_rapl & RAPL_DRAM_PERF_STATUS)
349 outp += sprintf(outp, " RAM_%%");
350 } else if (do_rapl && rapl_joules) {
351 if (do_rapl & RAPL_PKG)
352 outp += sprintf(outp, " Pkg_J");
353 if (do_rapl & RAPL_CORES)
354 outp += sprintf(outp, " Cor_J");
355 if (do_rapl & RAPL_GFX)
356 outp += sprintf(outp, " GFX_J");
357 if (do_rapl & RAPL_DRAM)
358 outp += sprintf(outp, " RAM_W");
359 if (do_rapl & RAPL_PKG_PERF_STATUS)
360 outp += sprintf(outp, " PKG_%%");
361 if (do_rapl & RAPL_DRAM_PERF_STATUS)
362 outp += sprintf(outp, " RAM_%%");
363 outp += sprintf(outp, " time");
367 outp += sprintf(outp, "\n");
370 int dump_counters(struct thread_data *t, struct core_data *c,
373 outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
376 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
377 t->cpu_id, t->flags);
378 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
379 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
380 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
381 outp += sprintf(outp, "c1: %016llX\n", t->c1);
382 outp += sprintf(outp, "msr0x%x: %08llX\n",
383 extra_delta_offset32, t->extra_delta32);
384 outp += sprintf(outp, "msr0x%x: %016llX\n",
385 extra_delta_offset64, t->extra_delta64);
386 outp += sprintf(outp, "msr0x%x: %08llX\n",
387 extra_msr_offset32, t->extra_msr32);
388 outp += sprintf(outp, "msr0x%x: %016llX\n",
389 extra_msr_offset64, t->extra_msr64);
391 outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
395 outp += sprintf(outp, "core: %d\n", c->core_id);
396 outp += sprintf(outp, "c3: %016llX\n", c->c3);
397 outp += sprintf(outp, "c6: %016llX\n", c->c6);
398 outp += sprintf(outp, "c7: %016llX\n", c->c7);
399 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
403 outp += sprintf(outp, "package: %d\n", p->package_id);
404 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
406 outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
408 outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
410 outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
411 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
412 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
413 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
414 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
415 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
416 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
417 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
418 outp += sprintf(outp, "Throttle PKG: %0X\n",
419 p->rapl_pkg_perf_status);
420 outp += sprintf(outp, "Throttle RAM: %0X\n",
421 p->rapl_dram_perf_status);
422 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
425 outp += sprintf(outp, "\n");
431 * column formatting convention & formats
433 int format_counters(struct thread_data *t, struct core_data *c,
436 double interval_float;
439 /* if showing only 1st thread in core and this isn't one, bail out */
440 if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
443 /* if showing only 1st thread in pkg and this isn't one, bail out */
444 if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
447 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
449 /* topo columns, print blanks on 1st (average) line */
450 if (t == &average.threads) {
452 outp += sprintf(outp, " -");
454 outp += sprintf(outp, " -");
456 outp += sprintf(outp, " -");
460 outp += sprintf(outp, "%8d", p->package_id);
462 outp += sprintf(outp, " -");
466 outp += sprintf(outp, "%8d", c->core_id);
468 outp += sprintf(outp, " -");
471 outp += sprintf(outp, "%8d", t->cpu_id);
476 outp += sprintf(outp, "%8.0f",
477 1.0 / units * t->aperf / interval_float);
482 outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc);
484 outp += sprintf(outp, "********");
489 outp += sprintf(outp, "%8.0f",
490 1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
493 outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
496 if (extra_delta_offset32)
497 outp += sprintf(outp, " %11llu", t->extra_delta32);
500 if (extra_delta_offset64)
501 outp += sprintf(outp, " %11llu", t->extra_delta64);
503 if (extra_msr_offset32)
504 outp += sprintf(outp, " 0x%08llx", t->extra_msr32);
507 if (extra_msr_offset64)
508 outp += sprintf(outp, " 0x%016llx", t->extra_msr64);
515 outp += sprintf(outp, "%8d", t->smi_count);
517 if (do_nhm_cstates) {
519 outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
521 outp += sprintf(outp, "********");
524 /* print per-core data only for 1st thread in core */
525 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
528 if (do_nhm_cstates && !do_slm_cstates)
529 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
531 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
533 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
536 outp += sprintf(outp, "%8d", c->core_temp_c);
538 /* print per-package data only for 1st core in package */
539 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
543 outp += sprintf(outp, "%8d", p->pkg_temp_c);
546 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
548 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
550 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
552 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
554 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
555 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
556 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
560 * If measurement interval exceeds minimum RAPL Joule Counter range,
561 * indicate that results are suspect by printing "**" in fraction place.
563 if (interval_float < rapl_joule_counter_range)
568 if (do_rapl && !rapl_joules) {
569 if (do_rapl & RAPL_PKG)
570 outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
571 if (do_rapl & RAPL_CORES)
572 outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
573 if (do_rapl & RAPL_GFX)
574 outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
575 if (do_rapl & RAPL_DRAM)
576 outp += sprintf(outp, fmt8, p->energy_dram * rapl_energy_units / interval_float);
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 } else if (do_rapl && rapl_joules) {
582 if (do_rapl & RAPL_PKG)
583 outp += sprintf(outp, fmt8,
584 p->energy_pkg * rapl_energy_units);
585 if (do_rapl & RAPL_CORES)
586 outp += sprintf(outp, fmt8,
587 p->energy_cores * rapl_energy_units);
588 if (do_rapl & RAPL_GFX)
589 outp += sprintf(outp, fmt8,
590 p->energy_gfx * rapl_energy_units);
591 if (do_rapl & RAPL_DRAM)
592 outp += sprintf(outp, fmt8,
593 p->energy_dram * rapl_energy_units);
594 if (do_rapl & RAPL_PKG_PERF_STATUS)
595 outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
596 if (do_rapl & RAPL_DRAM_PERF_STATUS)
597 outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
599 outp += sprintf(outp, fmt8, interval_float);
602 outp += sprintf(outp, "\n");
609 fputs(output_buffer, stdout);
611 outp = output_buffer;
615 fputs(output_buffer, stderr);
616 outp = output_buffer;
618 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
622 if (!printed || !summary_only)
625 if (topo.num_cpus > 1)
626 format_counters(&average.threads, &average.cores,
634 for_all_cpus(format_counters, t, c, p);
637 #define DELTA_WRAP32(new, old) \
641 old = 0x100000000 + new - old; \
645 delta_package(struct pkg_data *new, struct pkg_data *old)
647 old->pc2 = new->pc2 - old->pc2;
649 old->pc3 = new->pc3 - old->pc3;
651 old->pc6 = new->pc6 - old->pc6;
653 old->pc7 = new->pc7 - old->pc7;
654 old->pc8 = new->pc8 - old->pc8;
655 old->pc9 = new->pc9 - old->pc9;
656 old->pc10 = new->pc10 - old->pc10;
657 old->pkg_temp_c = new->pkg_temp_c;
659 DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
660 DELTA_WRAP32(new->energy_cores, old->energy_cores);
661 DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
662 DELTA_WRAP32(new->energy_dram, old->energy_dram);
663 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
664 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
668 delta_core(struct core_data *new, struct core_data *old)
670 old->c3 = new->c3 - old->c3;
671 old->c6 = new->c6 - old->c6;
672 old->c7 = new->c7 - old->c7;
673 old->core_temp_c = new->core_temp_c;
680 delta_thread(struct thread_data *new, struct thread_data *old,
681 struct core_data *core_delta)
683 old->tsc = new->tsc - old->tsc;
685 /* check for TSC < 1 Mcycles over interval */
686 if (old->tsc < (1000 * 1000))
687 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
688 "You can disable all c-states by booting with \"idle=poll\"\n"
689 "or just the deep ones with \"processor.max_cstate=1\"");
691 old->c1 = new->c1 - old->c1;
694 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
695 old->aperf = new->aperf - old->aperf;
696 old->mperf = new->mperf - old->mperf;
699 if (!aperf_mperf_unstable) {
700 fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
701 fprintf(stderr, "* Frequency results do not cover entire interval *\n");
702 fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
704 aperf_mperf_unstable = 1;
707 * mperf delta is likely a huge "positive" number
708 * can not use it for calculating c0 time
716 if (use_c1_residency_msr) {
718 * Some models have a dedicated C1 residency MSR,
719 * which should be more accurate than the derivation below.
723 * As counter collection is not atomic,
724 * it is possible for mperf's non-halted cycles + idle states
725 * to exceed TSC's all cycles: show c1 = 0% in that case.
727 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
730 /* normal case, derive c1 */
731 old->c1 = old->tsc - old->mperf - core_delta->c3
732 - core_delta->c6 - core_delta->c7;
736 if (old->mperf == 0) {
737 if (debug > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
738 old->mperf = 1; /* divide by 0 protection */
741 old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
742 old->extra_delta32 &= 0xFFFFFFFF;
744 old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
747 * Extra MSR is just a snapshot, simply copy latest w/o subtracting
749 old->extra_msr32 = new->extra_msr32;
750 old->extra_msr64 = new->extra_msr64;
753 old->smi_count = new->smi_count - old->smi_count;
756 int delta_cpu(struct thread_data *t, struct core_data *c,
757 struct pkg_data *p, struct thread_data *t2,
758 struct core_data *c2, struct pkg_data *p2)
760 /* calculate core delta only for 1st thread in core */
761 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
764 /* always calculate thread delta */
765 delta_thread(t, t2, c2); /* c2 is core delta */
767 /* calculate package delta only for 1st core in package */
768 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
769 delta_package(p, p2);
774 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
782 t->extra_delta32 = 0;
783 t->extra_delta64 = 0;
785 /* tells format_counters to dump all fields from this set */
786 t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
808 p->rapl_pkg_perf_status = 0;
809 p->rapl_dram_perf_status = 0;
812 int sum_counters(struct thread_data *t, struct core_data *c,
815 average.threads.tsc += t->tsc;
816 average.threads.aperf += t->aperf;
817 average.threads.mperf += t->mperf;
818 average.threads.c1 += t->c1;
820 average.threads.extra_delta32 += t->extra_delta32;
821 average.threads.extra_delta64 += t->extra_delta64;
823 /* sum per-core values only for 1st thread in core */
824 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
827 average.cores.c3 += c->c3;
828 average.cores.c6 += c->c6;
829 average.cores.c7 += c->c7;
831 average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
833 /* sum per-pkg values only for 1st core in pkg */
834 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
837 average.packages.pc2 += p->pc2;
839 average.packages.pc3 += p->pc3;
841 average.packages.pc6 += p->pc6;
843 average.packages.pc7 += p->pc7;
844 average.packages.pc8 += p->pc8;
845 average.packages.pc9 += p->pc9;
846 average.packages.pc10 += p->pc10;
848 average.packages.energy_pkg += p->energy_pkg;
849 average.packages.energy_dram += p->energy_dram;
850 average.packages.energy_cores += p->energy_cores;
851 average.packages.energy_gfx += p->energy_gfx;
853 average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
855 average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
856 average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
860 * sum the counters for all cpus in the system
861 * compute the weighted average
863 void compute_average(struct thread_data *t, struct core_data *c,
866 clear_counters(&average.threads, &average.cores, &average.packages);
868 for_all_cpus(sum_counters, t, c, p);
870 average.threads.tsc /= topo.num_cpus;
871 average.threads.aperf /= topo.num_cpus;
872 average.threads.mperf /= topo.num_cpus;
873 average.threads.c1 /= topo.num_cpus;
875 average.threads.extra_delta32 /= topo.num_cpus;
876 average.threads.extra_delta32 &= 0xFFFFFFFF;
878 average.threads.extra_delta64 /= topo.num_cpus;
880 average.cores.c3 /= topo.num_cores;
881 average.cores.c6 /= topo.num_cores;
882 average.cores.c7 /= topo.num_cores;
884 average.packages.pc2 /= topo.num_packages;
886 average.packages.pc3 /= topo.num_packages;
888 average.packages.pc6 /= topo.num_packages;
890 average.packages.pc7 /= topo.num_packages;
892 average.packages.pc8 /= topo.num_packages;
893 average.packages.pc9 /= topo.num_packages;
894 average.packages.pc10 /= topo.num_packages;
897 static unsigned long long rdtsc(void)
899 unsigned int low, high;
901 asm volatile("rdtsc" : "=a" (low), "=d" (high));
903 return low | ((unsigned long long)high) << 32;
910 * acquire and record local counters for that cpu
912 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
915 unsigned long long msr;
917 if (cpu_migrate(cpu)) {
918 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
922 t->tsc = rdtsc(); /* we are running on local CPU of interest */
925 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
927 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
932 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
934 t->smi_count = msr & 0xFFFFFFFF;
936 if (extra_delta_offset32) {
937 if (get_msr(cpu, extra_delta_offset32, &msr))
939 t->extra_delta32 = msr & 0xFFFFFFFF;
942 if (extra_delta_offset64)
943 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
946 if (extra_msr_offset32) {
947 if (get_msr(cpu, extra_msr_offset32, &msr))
949 t->extra_msr32 = msr & 0xFFFFFFFF;
952 if (extra_msr_offset64)
953 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
956 if (use_c1_residency_msr) {
957 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
961 /* collect core counters only for 1st thread in core */
962 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
965 if (do_nhm_cstates && !do_slm_cstates) {
966 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
970 if (do_nhm_cstates) {
971 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
976 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
980 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
982 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
986 /* collect package counters only for 1st core in package */
987 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
991 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
994 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
997 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
1000 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1003 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1005 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1007 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1010 if (do_rapl & RAPL_PKG) {
1011 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1013 p->energy_pkg = msr & 0xFFFFFFFF;
1015 if (do_rapl & RAPL_CORES) {
1016 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1018 p->energy_cores = msr & 0xFFFFFFFF;
1020 if (do_rapl & RAPL_DRAM) {
1021 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1023 p->energy_dram = msr & 0xFFFFFFFF;
1025 if (do_rapl & RAPL_GFX) {
1026 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1028 p->energy_gfx = msr & 0xFFFFFFFF;
1030 if (do_rapl & RAPL_PKG_PERF_STATUS) {
1031 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1033 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1035 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1036 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1038 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1041 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1043 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1049 * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1050 * If you change the values, note they are used both in comparisons
1051 * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1054 #define PCLUKN 0 /* Unknown */
1055 #define PCLRSV 1 /* Reserved */
1056 #define PCL__0 2 /* PC0 */
1057 #define PCL__1 3 /* PC1 */
1058 #define PCL__2 4 /* PC2 */
1059 #define PCL__3 5 /* PC3 */
1060 #define PCL__4 6 /* PC4 */
1061 #define PCL__6 7 /* PC6 */
1062 #define PCL_6N 8 /* PC6 No Retention */
1063 #define PCL_6R 9 /* PC6 Retention */
1064 #define PCL__7 10 /* PC7 */
1065 #define PCL_7S 11 /* PC7 Shrink */
1066 #define PCLUNL 12 /* Unlimited */
1068 int pkg_cstate_limit = PCLUKN;
1069 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1070 "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "unlimited"};
1072 int nhm_pkg_cstate_limits[8] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL};
1073 int snb_pkg_cstate_limits[8] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL};
1074 int hsw_pkg_cstate_limits[8] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCLRSV, PCLUNL};
1075 int slv_pkg_cstate_limits[8] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7};
1076 int amt_pkg_cstate_limits[8] = {PCL__0, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7};
1077 int phi_pkg_cstate_limits[8] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL};
1080 dump_nhm_platform_info(void)
1082 unsigned long long msr;
1085 get_msr(0, MSR_NHM_PLATFORM_INFO, &msr);
1087 fprintf(stderr, "cpu0: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", msr);
1089 ratio = (msr >> 40) & 0xFF;
1090 fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency frequency\n",
1091 ratio, bclk, ratio * bclk);
1093 ratio = (msr >> 8) & 0xFF;
1094 fprintf(stderr, "%d * %.0f = %.0f MHz base frequency\n",
1095 ratio, bclk, ratio * bclk);
1097 get_msr(0, MSR_IA32_POWER_CTL, &msr);
1098 fprintf(stderr, "cpu0: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1099 msr, msr & 0x2 ? "EN" : "DIS");
1105 dump_hsw_turbo_ratio_limits(void)
1107 unsigned long long msr;
1110 get_msr(0, MSR_TURBO_RATIO_LIMIT2, &msr);
1112 fprintf(stderr, "cpu0: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", msr);
1114 ratio = (msr >> 8) & 0xFF;
1116 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 18 active cores\n",
1117 ratio, bclk, ratio * bclk);
1119 ratio = (msr >> 0) & 0xFF;
1121 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 17 active cores\n",
1122 ratio, bclk, ratio * bclk);
1127 dump_ivt_turbo_ratio_limits(void)
1129 unsigned long long msr;
1132 get_msr(0, MSR_TURBO_RATIO_LIMIT1, &msr);
1134 fprintf(stderr, "cpu0: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", msr);
1136 ratio = (msr >> 56) & 0xFF;
1138 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1139 ratio, bclk, ratio * bclk);
1141 ratio = (msr >> 48) & 0xFF;
1143 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1144 ratio, bclk, ratio * bclk);
1146 ratio = (msr >> 40) & 0xFF;
1148 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1149 ratio, bclk, ratio * bclk);
1151 ratio = (msr >> 32) & 0xFF;
1153 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1154 ratio, bclk, ratio * bclk);
1156 ratio = (msr >> 24) & 0xFF;
1158 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1159 ratio, bclk, ratio * bclk);
1161 ratio = (msr >> 16) & 0xFF;
1163 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1164 ratio, bclk, ratio * bclk);
1166 ratio = (msr >> 8) & 0xFF;
1168 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1169 ratio, bclk, ratio * bclk);
1171 ratio = (msr >> 0) & 0xFF;
1173 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1174 ratio, bclk, ratio * bclk);
1179 dump_nhm_turbo_ratio_limits(void)
1181 unsigned long long msr;
1184 get_msr(0, MSR_TURBO_RATIO_LIMIT, &msr);
1186 fprintf(stderr, "cpu0: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1188 ratio = (msr >> 56) & 0xFF;
1190 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1191 ratio, bclk, ratio * bclk);
1193 ratio = (msr >> 48) & 0xFF;
1195 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1196 ratio, bclk, ratio * bclk);
1198 ratio = (msr >> 40) & 0xFF;
1200 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1201 ratio, bclk, ratio * bclk);
1203 ratio = (msr >> 32) & 0xFF;
1205 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1206 ratio, bclk, ratio * bclk);
1208 ratio = (msr >> 24) & 0xFF;
1210 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1211 ratio, bclk, ratio * bclk);
1213 ratio = (msr >> 16) & 0xFF;
1215 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1216 ratio, bclk, ratio * bclk);
1218 ratio = (msr >> 8) & 0xFF;
1220 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1221 ratio, bclk, ratio * bclk);
1223 ratio = (msr >> 0) & 0xFF;
1225 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1226 ratio, bclk, ratio * bclk);
1231 dump_nhm_cst_cfg(void)
1233 unsigned long long msr;
1235 get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1237 #define SNB_C1_AUTO_UNDEMOTE (1UL << 27)
1238 #define SNB_C3_AUTO_UNDEMOTE (1UL << 28)
1240 fprintf(stderr, "cpu0: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", msr);
1242 fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: %s)\n",
1243 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1244 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1245 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1246 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1247 (msr & (1 << 15)) ? "" : "UN",
1248 (unsigned int)msr & 7,
1249 pkg_cstate_limit_strings[pkg_cstate_limit]);
1253 void free_all_buffers(void)
1255 CPU_FREE(cpu_present_set);
1256 cpu_present_set = NULL;
1257 cpu_present_set = 0;
1259 CPU_FREE(cpu_affinity_set);
1260 cpu_affinity_set = NULL;
1261 cpu_affinity_setsize = 0;
1269 package_even = NULL;
1279 free(output_buffer);
1280 output_buffer = NULL;
1285 * Open a file, and exit on failure
1287 FILE *fopen_or_die(const char *path, const char *mode)
1289 FILE *filep = fopen(path, "r");
1291 err(1, "%s: open failed", path);
1296 * Parse a file containing a single int.
1298 int parse_int_file(const char *fmt, ...)
1301 char path[PATH_MAX];
1305 va_start(args, fmt);
1306 vsnprintf(path, sizeof(path), fmt, args);
1308 filep = fopen_or_die(path, "r");
1309 if (fscanf(filep, "%d", &value) != 1)
1310 err(1, "%s: failed to parse number from file", path);
1316 * cpu_is_first_sibling_in_core(cpu)
1317 * return 1 if given CPU is 1st HT sibling in the core
1319 int cpu_is_first_sibling_in_core(int cpu)
1321 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1325 * cpu_is_first_core_in_package(cpu)
1326 * return 1 if given CPU is 1st core in package
1328 int cpu_is_first_core_in_package(int cpu)
1330 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1333 int get_physical_package_id(int cpu)
1335 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1338 int get_core_id(int cpu)
1340 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1343 int get_num_ht_siblings(int cpu)
1351 sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1352 filep = fopen_or_die(path, "r");
1355 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1356 * otherwinse 1 sibling (self).
1358 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1369 * run func(thread, core, package) in topology order
1370 * skip non-present cpus
1373 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1374 struct pkg_data *, struct thread_data *, struct core_data *,
1375 struct pkg_data *), struct thread_data *thread_base,
1376 struct core_data *core_base, struct pkg_data *pkg_base,
1377 struct thread_data *thread_base2, struct core_data *core_base2,
1378 struct pkg_data *pkg_base2)
1380 int retval, pkg_no, core_no, thread_no;
1382 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1383 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1384 for (thread_no = 0; thread_no <
1385 topo.num_threads_per_core; ++thread_no) {
1386 struct thread_data *t, *t2;
1387 struct core_data *c, *c2;
1388 struct pkg_data *p, *p2;
1390 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1392 if (cpu_is_not_present(t->cpu_id))
1395 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1397 c = GET_CORE(core_base, core_no, pkg_no);
1398 c2 = GET_CORE(core_base2, core_no, pkg_no);
1400 p = GET_PKG(pkg_base, pkg_no);
1401 p2 = GET_PKG(pkg_base2, pkg_no);
1403 retval = func(t, c, p, t2, c2, p2);
1413 * run func(cpu) on every cpu in /proc/stat
1414 * return max_cpu number
1416 int for_all_proc_cpus(int (func)(int))
1422 fp = fopen_or_die(proc_stat, "r");
1424 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1426 err(1, "%s: failed to parse format", proc_stat);
1429 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1433 retval = func(cpu_num);
1443 void re_initialize(void)
1446 setup_all_buffers();
1447 printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1453 * remember the last one seen, it will be the max
1455 int count_cpus(int cpu)
1457 if (topo.max_cpu_num < cpu)
1458 topo.max_cpu_num = cpu;
1463 int mark_cpu_present(int cpu)
1465 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1469 void turbostat_loop()
1477 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1480 } else if (retval == -1) {
1481 if (restarted > 1) {
1488 gettimeofday(&tv_even, (struct timezone *)NULL);
1491 if (for_all_proc_cpus(cpu_is_not_present)) {
1495 sleep(interval_sec);
1496 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1499 } else if (retval == -1) {
1503 gettimeofday(&tv_odd, (struct timezone *)NULL);
1504 timersub(&tv_odd, &tv_even, &tv_delta);
1505 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1506 compute_average(EVEN_COUNTERS);
1507 format_all_counters(EVEN_COUNTERS);
1509 sleep(interval_sec);
1510 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1513 } else if (retval == -1) {
1517 gettimeofday(&tv_even, (struct timezone *)NULL);
1518 timersub(&tv_even, &tv_odd, &tv_delta);
1519 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1520 compute_average(ODD_COUNTERS);
1521 format_all_counters(ODD_COUNTERS);
1526 void check_dev_msr()
1530 if (stat("/dev/cpu/0/msr", &sb))
1531 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
1532 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
1535 void check_permissions()
1537 struct __user_cap_header_struct cap_header_data;
1538 cap_user_header_t cap_header = &cap_header_data;
1539 struct __user_cap_data_struct cap_data_data;
1540 cap_user_data_t cap_data = &cap_data_data;
1541 extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
1544 /* check for CAP_SYS_RAWIO */
1545 cap_header->pid = getpid();
1546 cap_header->version = _LINUX_CAPABILITY_VERSION;
1547 if (capget(cap_header, cap_data) < 0)
1548 err(-6, "capget(2) failed");
1550 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1552 warnx("capget(CAP_SYS_RAWIO) failed,"
1553 " try \"# setcap cap_sys_rawio=ep %s\"", progname);
1556 /* test file permissions */
1557 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1559 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1562 /* if all else fails, thell them to be root */
1565 warnx("... or simply run as root");
1572 * NHM adds support for additional MSRs:
1574 * MSR_SMI_COUNT 0x00000034
1576 * MSR_NHM_PLATFORM_INFO 0x000000ce
1577 * MSR_NHM_SNB_PKG_CST_CFG_CTL 0x000000e2
1579 * MSR_PKG_C3_RESIDENCY 0x000003f8
1580 * MSR_PKG_C6_RESIDENCY 0x000003f9
1581 * MSR_CORE_C3_RESIDENCY 0x000003fc
1582 * MSR_CORE_C6_RESIDENCY 0x000003fd
1585 * sets global pkg_cstate_limit to decode MSR_NHM_SNB_PKG_CST_CFG_CTL
1587 int probe_nhm_msrs(unsigned int family, unsigned int model)
1589 unsigned long long msr;
1590 int *pkg_cstate_limits;
1599 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1600 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1601 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1602 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1603 case 0x2C: /* Westmere EP - Gulftown */
1604 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1605 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1606 pkg_cstate_limits = nhm_pkg_cstate_limits;
1608 case 0x2A: /* SNB */
1609 case 0x2D: /* SNB Xeon */
1610 case 0x3A: /* IVB */
1611 case 0x3E: /* IVB Xeon */
1612 pkg_cstate_limits = snb_pkg_cstate_limits;
1614 case 0x3C: /* HSW */
1615 case 0x3F: /* HSX */
1616 case 0x45: /* HSW */
1617 case 0x46: /* HSW */
1618 case 0x3D: /* BDW */
1619 case 0x47: /* BDW */
1620 case 0x4F: /* BDX */
1621 case 0x56: /* BDX-DE */
1622 pkg_cstate_limits = hsw_pkg_cstate_limits;
1624 case 0x37: /* BYT */
1625 case 0x4D: /* AVN */
1626 pkg_cstate_limits = slv_pkg_cstate_limits;
1628 case 0x4C: /* AMT */
1629 pkg_cstate_limits = amt_pkg_cstate_limits;
1631 case 0x57: /* PHI */
1632 pkg_cstate_limits = phi_pkg_cstate_limits;
1637 get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1639 pkg_cstate_limit = pkg_cstate_limits[msr & 0x7];
1643 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
1646 /* Nehalem compatible, but do not include turbo-ratio limit support */
1647 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1648 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1654 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1663 case 0x3E: /* IVB Xeon */
1664 case 0x3F: /* HSW Xeon */
1670 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
1679 case 0x3F: /* HSW Xeon */
1687 dump_cstate_pstate_config_info(family, model)
1689 if (!do_nhm_platform_info)
1692 dump_nhm_platform_info();
1694 if (has_hsw_turbo_ratio_limit(family, model))
1695 dump_hsw_turbo_ratio_limits();
1697 if (has_ivt_turbo_ratio_limit(family, model))
1698 dump_ivt_turbo_ratio_limits();
1700 if (has_nhm_turbo_ratio_limit(family, model))
1701 dump_nhm_turbo_ratio_limits();
1709 * Decode the ENERGY_PERF_BIAS MSR
1711 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1713 unsigned long long msr;
1722 /* EPB is per-package */
1723 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1726 if (cpu_migrate(cpu)) {
1727 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1731 if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
1734 switch (msr & 0x7) {
1735 case ENERGY_PERF_BIAS_PERFORMANCE:
1736 epb_string = "performance";
1738 case ENERGY_PERF_BIAS_NORMAL:
1739 epb_string = "balanced";
1741 case ENERGY_PERF_BIAS_POWERSAVE:
1742 epb_string = "powersave";
1745 epb_string = "custom";
1748 fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
1754 * print_perf_limit()
1756 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1758 unsigned long long msr;
1764 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1767 if (cpu_migrate(cpu)) {
1768 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1772 if (do_core_perf_limit_reasons) {
1773 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
1774 fprintf(stderr, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1775 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
1776 (msr & 1 << 15) ? "bit15, " : "",
1777 (msr & 1 << 14) ? "bit14, " : "",
1778 (msr & 1 << 13) ? "Transitions, " : "",
1779 (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
1780 (msr & 1 << 11) ? "PkgPwrL2, " : "",
1781 (msr & 1 << 10) ? "PkgPwrL1, " : "",
1782 (msr & 1 << 9) ? "CorePwr, " : "",
1783 (msr & 1 << 8) ? "Amps, " : "",
1784 (msr & 1 << 6) ? "VR-Therm, " : "",
1785 (msr & 1 << 5) ? "Auto-HWP, " : "",
1786 (msr & 1 << 4) ? "Graphics, " : "",
1787 (msr & 1 << 2) ? "bit2, " : "",
1788 (msr & 1 << 1) ? "ThermStatus, " : "",
1789 (msr & 1 << 0) ? "PROCHOT, " : "");
1790 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
1791 (msr & 1 << 31) ? "bit31, " : "",
1792 (msr & 1 << 30) ? "bit30, " : "",
1793 (msr & 1 << 29) ? "Transitions, " : "",
1794 (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
1795 (msr & 1 << 27) ? "PkgPwrL2, " : "",
1796 (msr & 1 << 26) ? "PkgPwrL1, " : "",
1797 (msr & 1 << 25) ? "CorePwr, " : "",
1798 (msr & 1 << 24) ? "Amps, " : "",
1799 (msr & 1 << 22) ? "VR-Therm, " : "",
1800 (msr & 1 << 21) ? "Auto-HWP, " : "",
1801 (msr & 1 << 20) ? "Graphics, " : "",
1802 (msr & 1 << 18) ? "bit18, " : "",
1803 (msr & 1 << 17) ? "ThermStatus, " : "",
1804 (msr & 1 << 16) ? "PROCHOT, " : "");
1807 if (do_gfx_perf_limit_reasons) {
1808 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
1809 fprintf(stderr, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1810 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s)",
1811 (msr & 1 << 0) ? "PROCHOT, " : "",
1812 (msr & 1 << 1) ? "ThermStatus, " : "",
1813 (msr & 1 << 4) ? "Graphics, " : "",
1814 (msr & 1 << 6) ? "VR-Therm, " : "",
1815 (msr & 1 << 8) ? "Amps, " : "",
1816 (msr & 1 << 9) ? "GFXPwr, " : "",
1817 (msr & 1 << 10) ? "PkgPwrL1, " : "",
1818 (msr & 1 << 11) ? "PkgPwrL2, " : "");
1819 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s)\n",
1820 (msr & 1 << 16) ? "PROCHOT, " : "",
1821 (msr & 1 << 17) ? "ThermStatus, " : "",
1822 (msr & 1 << 20) ? "Graphics, " : "",
1823 (msr & 1 << 22) ? "VR-Therm, " : "",
1824 (msr & 1 << 24) ? "Amps, " : "",
1825 (msr & 1 << 25) ? "GFXPwr, " : "",
1826 (msr & 1 << 26) ? "PkgPwrL1, " : "",
1827 (msr & 1 << 27) ? "PkgPwrL2, " : "");
1829 if (do_ring_perf_limit_reasons) {
1830 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
1831 fprintf(stderr, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
1832 fprintf(stderr, " (Active: %s%s%s%s%s%s)",
1833 (msr & 1 << 0) ? "PROCHOT, " : "",
1834 (msr & 1 << 1) ? "ThermStatus, " : "",
1835 (msr & 1 << 6) ? "VR-Therm, " : "",
1836 (msr & 1 << 8) ? "Amps, " : "",
1837 (msr & 1 << 10) ? "PkgPwrL1, " : "",
1838 (msr & 1 << 11) ? "PkgPwrL2, " : "");
1839 fprintf(stderr, " (Logged: %s%s%s%s%s%s)\n",
1840 (msr & 1 << 16) ? "PROCHOT, " : "",
1841 (msr & 1 << 17) ? "ThermStatus, " : "",
1842 (msr & 1 << 22) ? "VR-Therm, " : "",
1843 (msr & 1 << 24) ? "Amps, " : "",
1844 (msr & 1 << 26) ? "PkgPwrL1, " : "",
1845 (msr & 1 << 27) ? "PkgPwrL2, " : "");
1850 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
1851 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
1853 double get_tdp(model)
1855 unsigned long long msr;
1857 if (do_rapl & RAPL_PKG_POWER_INFO)
1858 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
1859 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
1874 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
1876 void rapl_probe(unsigned int family, unsigned int model)
1878 unsigned long long msr;
1879 unsigned int time_unit;
1891 case 0x3C: /* HSW */
1892 case 0x45: /* HSW */
1893 case 0x46: /* HSW */
1894 case 0x3D: /* BDW */
1895 case 0x47: /* BDW */
1896 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1898 case 0x3F: /* HSX */
1899 case 0x4F: /* BDX */
1900 case 0x56: /* BDX-DE */
1901 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1905 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1907 case 0x37: /* BYT */
1908 case 0x4D: /* AVN */
1909 do_rapl = RAPL_PKG | RAPL_CORES ;
1915 /* units on package 0, verify later other packages match */
1916 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1919 rapl_power_units = 1.0 / (1 << (msr & 0xF));
1921 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1923 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1925 time_unit = msr >> 16 & 0xF;
1929 rapl_time_units = 1.0 / (1 << (time_unit));
1931 tdp = get_tdp(model);
1933 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1935 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1940 void perf_limit_reasons_probe(family, model)
1949 case 0x3C: /* HSW */
1950 case 0x45: /* HSW */
1951 case 0x46: /* HSW */
1952 do_gfx_perf_limit_reasons = 1;
1953 case 0x3F: /* HSX */
1954 do_core_perf_limit_reasons = 1;
1955 do_ring_perf_limit_reasons = 1;
1961 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1963 unsigned long long msr;
1967 if (!(do_dts || do_ptm))
1972 /* DTS is per-core, no need to print for each thread */
1973 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1976 if (cpu_migrate(cpu)) {
1977 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1981 if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
1982 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1985 dts = (msr >> 16) & 0x7F;
1986 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
1987 cpu, msr, tcc_activation_temp - dts);
1990 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
1993 dts = (msr >> 16) & 0x7F;
1994 dts2 = (msr >> 8) & 0x7F;
1995 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1996 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2002 unsigned int resolution;
2004 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
2007 dts = (msr >> 16) & 0x7F;
2008 resolution = (msr >> 27) & 0xF;
2009 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
2010 cpu, msr, tcc_activation_temp - dts, resolution);
2013 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
2016 dts = (msr >> 16) & 0x7F;
2017 dts2 = (msr >> 8) & 0x7F;
2018 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2019 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2026 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
2028 fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
2030 ((msr >> 15) & 1) ? "EN" : "DIS",
2031 ((msr >> 0) & 0x7FFF) * rapl_power_units,
2032 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
2033 (((msr >> 16) & 1) ? "EN" : "DIS"));
2038 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2040 unsigned long long msr;
2046 /* RAPL counters are per package, so print only for 1st thread/package */
2047 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2051 if (cpu_migrate(cpu)) {
2052 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2056 if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
2060 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
2061 "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
2062 rapl_power_units, rapl_energy_units, rapl_time_units);
2064 if (do_rapl & RAPL_PKG_POWER_INFO) {
2066 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
2070 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2072 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2073 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2074 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2075 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2078 if (do_rapl & RAPL_PKG) {
2080 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
2083 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
2084 cpu, msr, (msr >> 63) & 1 ? "": "UN");
2086 print_power_limit_msr(cpu, msr, "PKG Limit #1");
2087 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
2089 ((msr >> 47) & 1) ? "EN" : "DIS",
2090 ((msr >> 32) & 0x7FFF) * rapl_power_units,
2091 (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
2092 ((msr >> 48) & 1) ? "EN" : "DIS");
2095 if (do_rapl & RAPL_DRAM) {
2096 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
2100 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2102 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2103 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2104 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2105 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2108 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
2110 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
2111 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2113 print_power_limit_msr(cpu, msr, "DRAM Limit");
2115 if (do_rapl & RAPL_CORE_POLICY) {
2117 if (get_msr(cpu, MSR_PP0_POLICY, &msr))
2120 fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
2123 if (do_rapl & RAPL_CORES) {
2126 if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
2128 fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2129 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2130 print_power_limit_msr(cpu, msr, "Cores Limit");
2133 if (do_rapl & RAPL_GFX) {
2135 if (get_msr(cpu, MSR_PP1_POLICY, &msr))
2138 fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
2140 if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
2142 fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2143 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2144 print_power_limit_msr(cpu, msr, "GFX Limit");
2151 * SNB adds support for additional MSRs:
2153 * MSR_PKG_C7_RESIDENCY 0x000003fa
2154 * MSR_CORE_C7_RESIDENCY 0x000003fe
2155 * MSR_PKG_C2_RESIDENCY 0x0000060d
2158 int has_snb_msrs(unsigned int family, unsigned int model)
2166 case 0x3A: /* IVB */
2167 case 0x3E: /* IVB Xeon */
2168 case 0x3C: /* HSW */
2169 case 0x3F: /* HSW */
2170 case 0x45: /* HSW */
2171 case 0x46: /* HSW */
2172 case 0x3D: /* BDW */
2173 case 0x47: /* BDW */
2174 case 0x4F: /* BDX */
2175 case 0x56: /* BDX-DE */
2182 * HSW adds support for additional MSRs:
2184 * MSR_PKG_C8_RESIDENCY 0x00000630
2185 * MSR_PKG_C9_RESIDENCY 0x00000631
2186 * MSR_PKG_C10_RESIDENCY 0x00000632
2188 int has_hsw_msrs(unsigned int family, unsigned int model)
2194 case 0x45: /* HSW */
2195 case 0x3D: /* BDW */
2202 int is_slm(unsigned int family, unsigned int model)
2207 case 0x37: /* BYT */
2208 case 0x4D: /* AVN */
2214 #define SLM_BCLK_FREQS 5
2215 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
2217 double slm_bclk(void)
2219 unsigned long long msr = 3;
2223 if (get_msr(0, MSR_FSB_FREQ, &msr))
2224 fprintf(stderr, "SLM BCLK: unknown\n");
2227 if (i >= SLM_BCLK_FREQS) {
2228 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
2231 freq = slm_freq_table[i];
2233 fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
2238 double discover_bclk(unsigned int family, unsigned int model)
2240 if (has_snb_msrs(family, model))
2242 else if (is_slm(family, model))
2249 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
2250 * the Thermal Control Circuit (TCC) activates.
2251 * This is usually equal to tjMax.
2253 * Older processors do not have this MSR, so there we guess,
2254 * but also allow cmdline over-ride with -T.
2256 * Several MSR temperature values are in units of degrees-C
2257 * below this value, including the Digital Thermal Sensor (DTS),
2258 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
2260 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2262 unsigned long long msr;
2263 unsigned int target_c_local;
2266 /* tcc_activation_temp is used only for dts or ptm */
2267 if (!(do_dts || do_ptm))
2270 /* this is a per-package concept */
2271 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2275 if (cpu_migrate(cpu)) {
2276 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2280 if (tcc_activation_temp_override != 0) {
2281 tcc_activation_temp = tcc_activation_temp_override;
2282 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
2283 cpu, tcc_activation_temp);
2287 /* Temperature Target MSR is Nehalem and newer only */
2288 if (!do_nhm_platform_info)
2291 if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
2294 target_c_local = (msr >> 16) & 0xFF;
2297 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2298 cpu, msr, target_c_local);
2300 if (!target_c_local)
2303 tcc_activation_temp = target_c_local;
2308 tcc_activation_temp = TJMAX_DEFAULT;
2309 fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2310 cpu, tcc_activation_temp);
2314 void process_cpuid()
2316 unsigned int eax, ebx, ecx, edx, max_level;
2317 unsigned int fms, family, model, stepping;
2319 eax = ebx = ecx = edx = 0;
2321 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2323 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2327 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2328 (char *)&ebx, (char *)&edx, (char *)&ecx);
2330 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2331 family = (fms >> 8) & 0xf;
2332 model = (fms >> 4) & 0xf;
2333 stepping = fms & 0xf;
2334 if (family == 6 || family == 0xf)
2335 model += ((fms >> 16) & 0xf) << 4;
2338 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2339 max_level, family, model, stepping, family, model, stepping);
2341 if (!(edx & (1 << 5)))
2342 errx(1, "CPUID: no MSR");
2345 * check max extended function levels of CPUID.
2346 * This is needed to check for invariant TSC.
2347 * This check is valid for both Intel and AMD.
2349 ebx = ecx = edx = 0;
2350 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2352 if (max_level >= 0x80000007) {
2355 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2356 * this check is valid for both Intel and AMD
2358 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2359 has_invariant_tsc = edx & (1 << 8);
2363 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2364 * this check is valid for both Intel and AMD
2367 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2368 has_aperf = ecx & (1 << 0);
2369 do_dts = eax & (1 << 0);
2370 do_ptm = eax & (1 << 6);
2371 has_epb = ecx & (1 << 3);
2374 fprintf(stderr, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sEPB\n",
2375 has_aperf ? "" : "No ",
2376 do_dts ? "" : "No ",
2377 do_ptm ? "" : "No ",
2378 has_epb ? "" : "No ");
2380 do_nhm_platform_info = do_nhm_cstates = do_smi = probe_nhm_msrs(family, model);
2381 do_snb_cstates = has_snb_msrs(family, model);
2382 do_pc2 = do_snb_cstates && (pkg_cstate_limit >= PCL__2);
2383 do_pc3 = (pkg_cstate_limit >= PCL__3);
2384 do_pc6 = (pkg_cstate_limit >= PCL__6);
2385 do_pc7 = do_snb_cstates && (pkg_cstate_limit >= PCL__7);
2386 do_c8_c9_c10 = has_hsw_msrs(family, model);
2387 do_slm_cstates = is_slm(family, model);
2388 bclk = discover_bclk(family, model);
2390 rapl_probe(family, model);
2391 perf_limit_reasons_probe(family, model);
2394 dump_cstate_pstate_config_info();
2402 "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2404 "Turbostat forks the specified COMMAND and prints statistics\n"
2405 "when COMMAND completes.\n"
2406 "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2407 "to print statistics, until interrupted.\n"
2408 "--debug run in \"debug\" mode\n"
2409 "--interval sec Override default 5-second measurement interval\n"
2410 "--help print this help message\n"
2411 "--counter msr print 32-bit counter at address \"msr\"\n"
2412 "--Counter msr print 64-bit Counter at address \"msr\"\n"
2413 "--msr msr print 32-bit value at address \"msr\"\n"
2414 "--MSR msr print 64-bit Value at address \"msr\"\n"
2415 "--version print version information\n"
2417 "For more help, run \"man turbostat\"\n");
2422 * in /dev/cpu/ return success for names that are numbers
2423 * ie. filter out ".", "..", "microcode".
2425 int dir_filter(const struct dirent *dirp)
2427 if (isdigit(dirp->d_name[0]))
2433 int open_dev_cpu_msr(int dummy1)
2438 void topology_probe()
2441 int max_core_id = 0;
2442 int max_package_id = 0;
2443 int max_siblings = 0;
2444 struct cpu_topology {
2446 int physical_package_id;
2449 /* Initialize num_cpus, max_cpu_num */
2451 topo.max_cpu_num = 0;
2452 for_all_proc_cpus(count_cpus);
2453 if (!summary_only && topo.num_cpus > 1)
2457 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2459 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
2461 err(1, "calloc cpus");
2464 * Allocate and initialize cpu_present_set
2466 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2467 if (cpu_present_set == NULL)
2468 err(3, "CPU_ALLOC");
2469 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2470 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2471 for_all_proc_cpus(mark_cpu_present);
2474 * Allocate and initialize cpu_affinity_set
2476 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2477 if (cpu_affinity_set == NULL)
2478 err(3, "CPU_ALLOC");
2479 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2480 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2485 * find max_core_id, max_package_id
2487 for (i = 0; i <= topo.max_cpu_num; ++i) {
2490 if (cpu_is_not_present(i)) {
2492 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2495 cpus[i].core_id = get_core_id(i);
2496 if (cpus[i].core_id > max_core_id)
2497 max_core_id = cpus[i].core_id;
2499 cpus[i].physical_package_id = get_physical_package_id(i);
2500 if (cpus[i].physical_package_id > max_package_id)
2501 max_package_id = cpus[i].physical_package_id;
2503 siblings = get_num_ht_siblings(i);
2504 if (siblings > max_siblings)
2505 max_siblings = siblings;
2507 fprintf(stderr, "cpu %d pkg %d core %d\n",
2508 i, cpus[i].physical_package_id, cpus[i].core_id);
2510 topo.num_cores_per_pkg = max_core_id + 1;
2512 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2513 max_core_id, topo.num_cores_per_pkg);
2514 if (debug && !summary_only && topo.num_cores_per_pkg > 1)
2517 topo.num_packages = max_package_id + 1;
2519 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2520 max_package_id, topo.num_packages);
2521 if (debug && !summary_only && topo.num_packages > 1)
2524 topo.num_threads_per_core = max_siblings;
2526 fprintf(stderr, "max_siblings %d\n", max_siblings);
2532 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2536 *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2537 topo.num_packages, sizeof(struct thread_data));
2541 for (i = 0; i < topo.num_threads_per_core *
2542 topo.num_cores_per_pkg * topo.num_packages; i++)
2543 (*t)[i].cpu_id = -1;
2545 *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2546 sizeof(struct core_data));
2550 for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2551 (*c)[i].core_id = -1;
2553 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2557 for (i = 0; i < topo.num_packages; i++)
2558 (*p)[i].package_id = i;
2562 err(1, "calloc counters");
2567 * set cpu_id, core_num, pkg_num
2568 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2570 * increment topo.num_cores when 1st core in pkg seen
2572 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2573 struct pkg_data *pkg_base, int thread_num, int core_num,
2574 int pkg_num, int cpu_id)
2576 struct thread_data *t;
2577 struct core_data *c;
2580 t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2581 c = GET_CORE(core_base, core_num, pkg_num);
2582 p = GET_PKG(pkg_base, pkg_num);
2585 if (thread_num == 0) {
2586 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2587 if (cpu_is_first_core_in_package(cpu_id))
2588 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2591 c->core_id = core_num;
2592 p->package_id = pkg_num;
2596 int initialize_counters(int cpu_id)
2598 int my_thread_id, my_core_id, my_package_id;
2600 my_package_id = get_physical_package_id(cpu_id);
2601 my_core_id = get_core_id(cpu_id);
2603 if (cpu_is_first_sibling_in_core(cpu_id)) {
2610 init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2611 init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2615 void allocate_output_buffer()
2617 output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
2618 outp = output_buffer;
2620 err(-1, "calloc output buffer");
2623 void setup_all_buffers(void)
2626 allocate_counters(&thread_even, &core_even, &package_even);
2627 allocate_counters(&thread_odd, &core_odd, &package_odd);
2628 allocate_output_buffer();
2629 for_all_proc_cpus(initialize_counters);
2632 void turbostat_init()
2635 check_permissions();
2638 setup_all_buffers();
2641 for_all_cpus(print_epb, ODD_COUNTERS);
2644 for_all_cpus(print_perf_limit, ODD_COUNTERS);
2647 for_all_cpus(print_rapl, ODD_COUNTERS);
2649 for_all_cpus(set_temperature_target, ODD_COUNTERS);
2652 for_all_cpus(print_thermal, ODD_COUNTERS);
2655 int fork_it(char **argv)
2660 status = for_all_cpus(get_counters, EVEN_COUNTERS);
2663 /* clear affinity side-effect of get_counters() */
2664 sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
2665 gettimeofday(&tv_even, (struct timezone *)NULL);
2670 execvp(argv[0], argv);
2674 if (child_pid == -1)
2677 signal(SIGINT, SIG_IGN);
2678 signal(SIGQUIT, SIG_IGN);
2679 if (waitpid(child_pid, &status, 0) == -1)
2680 err(status, "waitpid");
2683 * n.b. fork_it() does not check for errors from for_all_cpus()
2684 * because re-starting is problematic when forking
2686 for_all_cpus(get_counters, ODD_COUNTERS);
2687 gettimeofday(&tv_odd, (struct timezone *)NULL);
2688 timersub(&tv_odd, &tv_even, &tv_delta);
2689 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
2690 compute_average(EVEN_COUNTERS);
2691 format_all_counters(EVEN_COUNTERS);
2694 fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
2699 int get_and_dump_counters(void)
2703 status = for_all_cpus(get_counters, ODD_COUNTERS);
2707 status = for_all_cpus(dump_counters, ODD_COUNTERS);
2716 void print_version() {
2717 fprintf(stderr, "turbostat version 4.3 24 Mar, 2015"
2718 " - Len Brown <lenb@kernel.org>\n");
2721 void cmdline(int argc, char **argv)
2724 int option_index = 0;
2725 static struct option long_options[] = {
2726 {"Counter", required_argument, 0, 'C'},
2727 {"counter", required_argument, 0, 'c'},
2728 {"Dump", no_argument, 0, 'D'},
2729 {"debug", no_argument, 0, 'd'},
2730 {"interval", required_argument, 0, 'i'},
2731 {"help", no_argument, 0, 'h'},
2732 {"Joules", no_argument, 0, 'J'},
2733 {"MSR", required_argument, 0, 'M'},
2734 {"msr", required_argument, 0, 'm'},
2735 {"Package", no_argument, 0, 'p'},
2736 {"processor", no_argument, 0, 'p'},
2737 {"Summary", no_argument, 0, 'S'},
2738 {"TCC", required_argument, 0, 'T'},
2739 {"version", no_argument, 0, 'v' },
2745 while ((opt = getopt_long_only(argc, argv, "C:c:Ddhi:JM:m:PpST:v",
2746 long_options, &option_index)) != -1) {
2749 sscanf(optarg, "%x", &extra_delta_offset64);
2752 sscanf(optarg, "%x", &extra_delta_offset32);
2765 interval_sec = atoi(optarg);
2771 sscanf(optarg, "%x", &extra_msr_offset64);
2774 sscanf(optarg, "%x", &extra_msr_offset32);
2786 tcc_activation_temp_override = atoi(optarg);
2796 int main(int argc, char **argv)
2798 cmdline(argc, argv);
2805 /* dump counters and exit */
2807 return get_and_dump_counters();
2810 * if any params left, it must be a command to fork
2813 return fork_it(argv + optind);