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>
43 #include <linux/capability.h>
46 char *proc_stat = "/proc/stat";
49 struct timespec interval_ts = {5, 0};
51 unsigned int rapl_joules;
52 unsigned int summary_only;
53 unsigned int dump_only;
56 unsigned int do_nhm_cstates;
57 unsigned int do_snb_cstates;
58 unsigned int do_knl_cstates;
63 unsigned int do_c8_c9_c10;
64 unsigned int do_skl_residency;
65 unsigned int do_slm_cstates;
66 unsigned int use_c1_residency_msr;
67 unsigned int has_aperf;
69 unsigned int do_irtl_snb;
70 unsigned int do_irtl_hsw;
71 unsigned int units = 1000000; /* MHz etc */
72 unsigned int genuine_intel;
73 unsigned int has_invariant_tsc;
74 unsigned int do_nhm_platform_info;
75 unsigned int extra_msr_offset32;
76 unsigned int extra_msr_offset64;
77 unsigned int extra_delta_offset32;
78 unsigned int extra_delta_offset64;
79 unsigned int aperf_mperf_multiplier = 1;
84 unsigned int has_base_hz;
85 double tsc_tweak = 1.0;
86 unsigned int show_pkg;
87 unsigned int show_core;
88 unsigned int show_cpu;
89 unsigned int show_pkg_only;
90 unsigned int show_core_only;
91 char *output_buffer, *outp;
95 unsigned int do_gfx_rc6_ms;
96 unsigned long long gfx_cur_rc6_ms;
97 unsigned int do_gfx_mhz;
98 unsigned int gfx_cur_mhz;
99 unsigned int tcc_activation_temp;
100 unsigned int tcc_activation_temp_override;
101 double rapl_power_units, rapl_time_units;
102 double rapl_dram_energy_units, rapl_energy_units;
103 double rapl_joule_counter_range;
104 unsigned int do_core_perf_limit_reasons;
105 unsigned int do_gfx_perf_limit_reasons;
106 unsigned int do_ring_perf_limit_reasons;
107 unsigned int crystal_hz;
108 unsigned long long tsc_hz;
110 double discover_bclk(unsigned int family, unsigned int model);
111 unsigned int has_hwp; /* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
112 /* IA32_HWP_REQUEST, IA32_HWP_STATUS */
113 unsigned int has_hwp_notify; /* IA32_HWP_INTERRUPT */
114 unsigned int has_hwp_activity_window; /* IA32_HWP_REQUEST[bits 41:32] */
115 unsigned int has_hwp_epp; /* IA32_HWP_REQUEST[bits 31:24] */
116 unsigned int has_hwp_pkg; /* IA32_HWP_REQUEST_PKG */
118 #define RAPL_PKG (1 << 0)
119 /* 0x610 MSR_PKG_POWER_LIMIT */
120 /* 0x611 MSR_PKG_ENERGY_STATUS */
121 #define RAPL_PKG_PERF_STATUS (1 << 1)
122 /* 0x613 MSR_PKG_PERF_STATUS */
123 #define RAPL_PKG_POWER_INFO (1 << 2)
124 /* 0x614 MSR_PKG_POWER_INFO */
126 #define RAPL_DRAM (1 << 3)
127 /* 0x618 MSR_DRAM_POWER_LIMIT */
128 /* 0x619 MSR_DRAM_ENERGY_STATUS */
129 #define RAPL_DRAM_PERF_STATUS (1 << 4)
130 /* 0x61b MSR_DRAM_PERF_STATUS */
131 #define RAPL_DRAM_POWER_INFO (1 << 5)
132 /* 0x61c MSR_DRAM_POWER_INFO */
134 #define RAPL_CORES (1 << 6)
135 /* 0x638 MSR_PP0_POWER_LIMIT */
136 /* 0x639 MSR_PP0_ENERGY_STATUS */
137 #define RAPL_CORE_POLICY (1 << 7)
138 /* 0x63a MSR_PP0_POLICY */
140 #define RAPL_GFX (1 << 8)
141 /* 0x640 MSR_PP1_POWER_LIMIT */
142 /* 0x641 MSR_PP1_ENERGY_STATUS */
143 /* 0x642 MSR_PP1_POLICY */
144 #define TJMAX_DEFAULT 100
146 #define MAX(a, b) ((a) > (b) ? (a) : (b))
148 int aperf_mperf_unstable;
152 cpu_set_t *cpu_present_set, *cpu_affinity_set;
153 size_t cpu_present_setsize, cpu_affinity_setsize;
156 unsigned long long tsc;
157 unsigned long long aperf;
158 unsigned long long mperf;
159 unsigned long long c1;
160 unsigned long long extra_msr64;
161 unsigned long long extra_delta64;
162 unsigned long long extra_msr32;
163 unsigned long long extra_delta32;
164 unsigned int irq_count;
165 unsigned int smi_count;
168 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
169 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
170 } *thread_even, *thread_odd;
173 unsigned long long c3;
174 unsigned long long c6;
175 unsigned long long c7;
176 unsigned int core_temp_c;
177 unsigned int core_id;
178 } *core_even, *core_odd;
181 unsigned long long pc2;
182 unsigned long long pc3;
183 unsigned long long pc6;
184 unsigned long long pc7;
185 unsigned long long pc8;
186 unsigned long long pc9;
187 unsigned long long pc10;
188 unsigned long long pkg_wtd_core_c0;
189 unsigned long long pkg_any_core_c0;
190 unsigned long long pkg_any_gfxe_c0;
191 unsigned long long pkg_both_core_gfxe_c0;
192 long long gfx_rc6_ms;
193 unsigned int gfx_mhz;
194 unsigned int package_id;
195 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
196 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
197 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
198 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
199 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
200 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
201 unsigned int pkg_temp_c;
203 } *package_even, *package_odd;
205 #define ODD_COUNTERS thread_odd, core_odd, package_odd
206 #define EVEN_COUNTERS thread_even, core_even, package_even
208 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
209 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
210 topo.num_threads_per_core + \
211 (core_no) * topo.num_threads_per_core + (thread_no))
212 #define GET_CORE(core_base, core_no, pkg_no) \
213 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
214 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
216 struct system_summary {
217 struct thread_data threads;
218 struct core_data cores;
219 struct pkg_data packages;
228 int num_cores_per_pkg;
229 int num_threads_per_core;
232 struct timeval tv_even, tv_odd, tv_delta;
234 int *irq_column_2_cpu; /* /proc/interrupts column numbers */
235 int *irqs_per_cpu; /* indexed by cpu_num */
237 void setup_all_buffers(void);
239 int cpu_is_not_present(int cpu)
241 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
244 * run func(thread, core, package) in topology order
245 * skip non-present cpus
248 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
249 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
251 int retval, pkg_no, core_no, thread_no;
253 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
254 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
255 for (thread_no = 0; thread_no <
256 topo.num_threads_per_core; ++thread_no) {
257 struct thread_data *t;
261 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
263 if (cpu_is_not_present(t->cpu_id))
266 c = GET_CORE(core_base, core_no, pkg_no);
267 p = GET_PKG(pkg_base, pkg_no);
269 retval = func(t, c, p);
278 int cpu_migrate(int cpu)
280 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
281 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
282 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
287 int get_msr_fd(int cpu)
297 sprintf(pathname, "/dev/cpu/%d/msr", cpu);
298 fd = open(pathname, O_RDONLY);
300 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
307 int get_msr(int cpu, off_t offset, unsigned long long *msr)
311 retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
313 if (retval != sizeof *msr)
314 err(-1, "msr %d offset 0x%llx read failed", cpu, (unsigned long long)offset);
320 * Example Format w/ field column widths:
322 * Package Core CPU Avg_MHz Bzy_MHz TSC_MHz IRQ SMI Busy% CPU_%c1 CPU_%c3 CPU_%c6 CPU_%c7 CoreTmp PkgTmp GFXMHz Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 PkgWatt CorWatt GFXWatt
323 * 12345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
326 void print_header(void)
329 outp += sprintf(outp, " Package");
331 outp += sprintf(outp, " Core");
333 outp += sprintf(outp, " CPU");
335 outp += sprintf(outp, " Avg_MHz");
337 outp += sprintf(outp, " Busy%%");
339 outp += sprintf(outp, " Bzy_MHz");
340 outp += sprintf(outp, " TSC_MHz");
342 if (extra_delta_offset32)
343 outp += sprintf(outp, " count 0x%03X", extra_delta_offset32);
344 if (extra_delta_offset64)
345 outp += sprintf(outp, " COUNT 0x%03X", extra_delta_offset64);
346 if (extra_msr_offset32)
347 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset32);
348 if (extra_msr_offset64)
349 outp += sprintf(outp, " MSR 0x%03X", extra_msr_offset64);
355 outp += sprintf(outp, " IRQ");
357 outp += sprintf(outp, " SMI");
360 outp += sprintf(outp, " CPU%%c1");
361 if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
362 outp += sprintf(outp, " CPU%%c3");
364 outp += sprintf(outp, " CPU%%c6");
366 outp += sprintf(outp, " CPU%%c7");
369 outp += sprintf(outp, " CoreTmp");
371 outp += sprintf(outp, " PkgTmp");
374 outp += sprintf(outp, " GFX%%rc6");
377 outp += sprintf(outp, " GFXMHz");
379 if (do_skl_residency) {
380 outp += sprintf(outp, " Totl%%C0");
381 outp += sprintf(outp, " Any%%C0");
382 outp += sprintf(outp, " GFX%%C0");
383 outp += sprintf(outp, " CPUGFX%%");
387 outp += sprintf(outp, " Pkg%%pc2");
389 outp += sprintf(outp, " Pkg%%pc3");
391 outp += sprintf(outp, " Pkg%%pc6");
393 outp += sprintf(outp, " Pkg%%pc7");
395 outp += sprintf(outp, " Pkg%%pc8");
396 outp += sprintf(outp, " Pkg%%pc9");
397 outp += sprintf(outp, " Pk%%pc10");
400 if (do_rapl && !rapl_joules) {
401 if (do_rapl & RAPL_PKG)
402 outp += sprintf(outp, " PkgWatt");
403 if (do_rapl & RAPL_CORES)
404 outp += sprintf(outp, " CorWatt");
405 if (do_rapl & RAPL_GFX)
406 outp += sprintf(outp, " GFXWatt");
407 if (do_rapl & RAPL_DRAM)
408 outp += sprintf(outp, " RAMWatt");
409 if (do_rapl & RAPL_PKG_PERF_STATUS)
410 outp += sprintf(outp, " PKG_%%");
411 if (do_rapl & RAPL_DRAM_PERF_STATUS)
412 outp += sprintf(outp, " RAM_%%");
413 } else if (do_rapl && rapl_joules) {
414 if (do_rapl & RAPL_PKG)
415 outp += sprintf(outp, " Pkg_J");
416 if (do_rapl & RAPL_CORES)
417 outp += sprintf(outp, " Cor_J");
418 if (do_rapl & RAPL_GFX)
419 outp += sprintf(outp, " GFX_J");
420 if (do_rapl & RAPL_DRAM)
421 outp += sprintf(outp, " RAM_J");
422 if (do_rapl & RAPL_PKG_PERF_STATUS)
423 outp += sprintf(outp, " PKG_%%");
424 if (do_rapl & RAPL_DRAM_PERF_STATUS)
425 outp += sprintf(outp, " RAM_%%");
426 outp += sprintf(outp, " time");
430 outp += sprintf(outp, "\n");
433 int dump_counters(struct thread_data *t, struct core_data *c,
436 outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
439 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
440 t->cpu_id, t->flags);
441 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
442 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
443 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
444 outp += sprintf(outp, "c1: %016llX\n", t->c1);
445 outp += sprintf(outp, "msr0x%x: %08llX\n",
446 extra_delta_offset32, t->extra_delta32);
447 outp += sprintf(outp, "msr0x%x: %016llX\n",
448 extra_delta_offset64, t->extra_delta64);
449 outp += sprintf(outp, "msr0x%x: %08llX\n",
450 extra_msr_offset32, t->extra_msr32);
451 outp += sprintf(outp, "msr0x%x: %016llX\n",
452 extra_msr_offset64, t->extra_msr64);
454 outp += sprintf(outp, "IRQ: %08X\n", t->irq_count);
456 outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
460 outp += sprintf(outp, "core: %d\n", c->core_id);
461 outp += sprintf(outp, "c3: %016llX\n", c->c3);
462 outp += sprintf(outp, "c6: %016llX\n", c->c6);
463 outp += sprintf(outp, "c7: %016llX\n", c->c7);
464 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
468 outp += sprintf(outp, "package: %d\n", p->package_id);
470 outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
471 outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
472 outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
473 outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
475 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
477 outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
479 outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
481 outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
482 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
483 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
484 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
485 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
486 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
487 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
488 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
489 outp += sprintf(outp, "Throttle PKG: %0X\n",
490 p->rapl_pkg_perf_status);
491 outp += sprintf(outp, "Throttle RAM: %0X\n",
492 p->rapl_dram_perf_status);
493 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
496 outp += sprintf(outp, "\n");
502 * column formatting convention & formats
504 int format_counters(struct thread_data *t, struct core_data *c,
507 double interval_float;
510 /* if showing only 1st thread in core and this isn't one, bail out */
511 if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
514 /* if showing only 1st thread in pkg and this isn't one, bail out */
515 if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
518 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
520 /* topo columns, print blanks on 1st (average) line */
521 if (t == &average.threads) {
523 outp += sprintf(outp, " -");
525 outp += sprintf(outp, " -");
527 outp += sprintf(outp, " -");
531 outp += sprintf(outp, "%8d", p->package_id);
533 outp += sprintf(outp, " -");
537 outp += sprintf(outp, "%8d", c->core_id);
539 outp += sprintf(outp, " -");
542 outp += sprintf(outp, "%8d", t->cpu_id);
547 outp += sprintf(outp, "%8.0f",
548 1.0 / units * t->aperf / interval_float);
553 outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc/tsc_tweak);
555 outp += sprintf(outp, "********");
561 outp += sprintf(outp, "%8.0f", base_hz / units * t->aperf / t->mperf);
563 outp += sprintf(outp, "%8.0f",
564 1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
568 outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
571 if (extra_delta_offset32)
572 outp += sprintf(outp, " %11llu", t->extra_delta32);
575 if (extra_delta_offset64)
576 outp += sprintf(outp, " %11llu", t->extra_delta64);
578 if (extra_msr_offset32)
579 outp += sprintf(outp, " 0x%08llx", t->extra_msr32);
582 if (extra_msr_offset64)
583 outp += sprintf(outp, " 0x%016llx", t->extra_msr64);
590 outp += sprintf(outp, "%8d", t->irq_count);
594 outp += sprintf(outp, "%8d", t->smi_count);
596 if (do_nhm_cstates) {
598 outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
600 outp += sprintf(outp, "********");
603 /* print per-core data only for 1st thread in core */
604 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
607 if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
608 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
610 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
612 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
615 outp += sprintf(outp, "%8d", c->core_temp_c);
617 /* print per-package data only for 1st core in package */
618 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
623 outp += sprintf(outp, "%8d", p->pkg_temp_c);
627 if (p->gfx_rc6_ms == -1) { /* detect counter reset */
628 outp += sprintf(outp, " ***.**");
630 outp += sprintf(outp, "%8.2f",
631 p->gfx_rc6_ms / 10.0 / interval_float);
637 outp += sprintf(outp, "%8d", p->gfx_mhz);
639 /* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
640 if (do_skl_residency) {
641 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_wtd_core_c0/t->tsc);
642 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_any_core_c0/t->tsc);
643 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_any_gfxe_c0/t->tsc);
644 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_both_core_gfxe_c0/t->tsc);
648 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
650 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
652 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
654 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
656 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
657 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
658 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
662 * If measurement interval exceeds minimum RAPL Joule Counter range,
663 * indicate that results are suspect by printing "**" in fraction place.
665 if (interval_float < rapl_joule_counter_range)
670 if (do_rapl && !rapl_joules) {
671 if (do_rapl & RAPL_PKG)
672 outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
673 if (do_rapl & RAPL_CORES)
674 outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
675 if (do_rapl & RAPL_GFX)
676 outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
677 if (do_rapl & RAPL_DRAM)
678 outp += sprintf(outp, fmt8, p->energy_dram * rapl_dram_energy_units / interval_float);
679 if (do_rapl & RAPL_PKG_PERF_STATUS)
680 outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
681 if (do_rapl & RAPL_DRAM_PERF_STATUS)
682 outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
683 } else if (do_rapl && rapl_joules) {
684 if (do_rapl & RAPL_PKG)
685 outp += sprintf(outp, fmt8,
686 p->energy_pkg * rapl_energy_units);
687 if (do_rapl & RAPL_CORES)
688 outp += sprintf(outp, fmt8,
689 p->energy_cores * rapl_energy_units);
690 if (do_rapl & RAPL_GFX)
691 outp += sprintf(outp, fmt8,
692 p->energy_gfx * rapl_energy_units);
693 if (do_rapl & RAPL_DRAM)
694 outp += sprintf(outp, fmt8,
695 p->energy_dram * rapl_dram_energy_units);
696 if (do_rapl & RAPL_PKG_PERF_STATUS)
697 outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
698 if (do_rapl & RAPL_DRAM_PERF_STATUS)
699 outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
701 outp += sprintf(outp, fmt8, interval_float);
704 outp += sprintf(outp, "\n");
709 void flush_output_stdout(void)
718 fputs(output_buffer, filep);
721 outp = output_buffer;
723 void flush_output_stderr(void)
725 fputs(output_buffer, outf);
727 outp = output_buffer;
729 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
733 if (!printed || !summary_only)
736 if (topo.num_cpus > 1)
737 format_counters(&average.threads, &average.cores,
745 for_all_cpus(format_counters, t, c, p);
748 #define DELTA_WRAP32(new, old) \
752 old = 0x100000000 + new - old; \
756 delta_package(struct pkg_data *new, struct pkg_data *old)
759 if (do_skl_residency) {
760 old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
761 old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
762 old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
763 old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
765 old->pc2 = new->pc2 - old->pc2;
767 old->pc3 = new->pc3 - old->pc3;
769 old->pc6 = new->pc6 - old->pc6;
771 old->pc7 = new->pc7 - old->pc7;
772 old->pc8 = new->pc8 - old->pc8;
773 old->pc9 = new->pc9 - old->pc9;
774 old->pc10 = new->pc10 - old->pc10;
775 old->pkg_temp_c = new->pkg_temp_c;
777 /* flag an error when rc6 counter resets/wraps */
778 if (old->gfx_rc6_ms > new->gfx_rc6_ms)
779 old->gfx_rc6_ms = -1;
781 old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
783 old->gfx_mhz = new->gfx_mhz;
785 DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
786 DELTA_WRAP32(new->energy_cores, old->energy_cores);
787 DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
788 DELTA_WRAP32(new->energy_dram, old->energy_dram);
789 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
790 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
794 delta_core(struct core_data *new, struct core_data *old)
796 old->c3 = new->c3 - old->c3;
797 old->c6 = new->c6 - old->c6;
798 old->c7 = new->c7 - old->c7;
799 old->core_temp_c = new->core_temp_c;
806 delta_thread(struct thread_data *new, struct thread_data *old,
807 struct core_data *core_delta)
809 old->tsc = new->tsc - old->tsc;
811 /* check for TSC < 1 Mcycles over interval */
812 if (old->tsc < (1000 * 1000))
813 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
814 "You can disable all c-states by booting with \"idle=poll\"\n"
815 "or just the deep ones with \"processor.max_cstate=1\"");
817 old->c1 = new->c1 - old->c1;
820 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
821 old->aperf = new->aperf - old->aperf;
822 old->mperf = new->mperf - old->mperf;
825 if (!aperf_mperf_unstable) {
826 fprintf(outf, "%s: APERF or MPERF went backwards *\n", progname);
827 fprintf(outf, "* Frequency results do not cover entire interval *\n");
828 fprintf(outf, "* fix this by running Linux-2.6.30 or later *\n");
830 aperf_mperf_unstable = 1;
833 * mperf delta is likely a huge "positive" number
834 * can not use it for calculating c0 time
842 if (use_c1_residency_msr) {
844 * Some models have a dedicated C1 residency MSR,
845 * which should be more accurate than the derivation below.
849 * As counter collection is not atomic,
850 * it is possible for mperf's non-halted cycles + idle states
851 * to exceed TSC's all cycles: show c1 = 0% in that case.
853 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
856 /* normal case, derive c1 */
857 old->c1 = old->tsc - old->mperf - core_delta->c3
858 - core_delta->c6 - core_delta->c7;
862 if (old->mperf == 0) {
864 fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
865 old->mperf = 1; /* divide by 0 protection */
868 old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
869 old->extra_delta32 &= 0xFFFFFFFF;
871 old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
874 * Extra MSR is just a snapshot, simply copy latest w/o subtracting
876 old->extra_msr32 = new->extra_msr32;
877 old->extra_msr64 = new->extra_msr64;
880 old->irq_count = new->irq_count - old->irq_count;
883 old->smi_count = new->smi_count - old->smi_count;
886 int delta_cpu(struct thread_data *t, struct core_data *c,
887 struct pkg_data *p, struct thread_data *t2,
888 struct core_data *c2, struct pkg_data *p2)
890 /* calculate core delta only for 1st thread in core */
891 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
894 /* always calculate thread delta */
895 delta_thread(t, t2, c2); /* c2 is core delta */
897 /* calculate package delta only for 1st core in package */
898 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
899 delta_package(p, p2);
904 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
911 t->extra_delta32 = 0;
912 t->extra_delta64 = 0;
917 /* tells format_counters to dump all fields from this set */
918 t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
925 p->pkg_wtd_core_c0 = 0;
926 p->pkg_any_core_c0 = 0;
927 p->pkg_any_gfxe_c0 = 0;
928 p->pkg_both_core_gfxe_c0 = 0;
945 p->rapl_pkg_perf_status = 0;
946 p->rapl_dram_perf_status = 0;
952 int sum_counters(struct thread_data *t, struct core_data *c,
955 average.threads.tsc += t->tsc;
956 average.threads.aperf += t->aperf;
957 average.threads.mperf += t->mperf;
958 average.threads.c1 += t->c1;
960 average.threads.extra_delta32 += t->extra_delta32;
961 average.threads.extra_delta64 += t->extra_delta64;
963 average.threads.irq_count += t->irq_count;
964 average.threads.smi_count += t->smi_count;
966 /* sum per-core values only for 1st thread in core */
967 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
970 average.cores.c3 += c->c3;
971 average.cores.c6 += c->c6;
972 average.cores.c7 += c->c7;
974 average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
976 /* sum per-pkg values only for 1st core in pkg */
977 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
980 if (do_skl_residency) {
981 average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
982 average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
983 average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
984 average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
987 average.packages.pc2 += p->pc2;
989 average.packages.pc3 += p->pc3;
991 average.packages.pc6 += p->pc6;
993 average.packages.pc7 += p->pc7;
994 average.packages.pc8 += p->pc8;
995 average.packages.pc9 += p->pc9;
996 average.packages.pc10 += p->pc10;
998 average.packages.energy_pkg += p->energy_pkg;
999 average.packages.energy_dram += p->energy_dram;
1000 average.packages.energy_cores += p->energy_cores;
1001 average.packages.energy_gfx += p->energy_gfx;
1003 average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
1004 average.packages.gfx_mhz = p->gfx_mhz;
1006 average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
1008 average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
1009 average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
1013 * sum the counters for all cpus in the system
1014 * compute the weighted average
1016 void compute_average(struct thread_data *t, struct core_data *c,
1019 clear_counters(&average.threads, &average.cores, &average.packages);
1021 for_all_cpus(sum_counters, t, c, p);
1023 average.threads.tsc /= topo.num_cpus;
1024 average.threads.aperf /= topo.num_cpus;
1025 average.threads.mperf /= topo.num_cpus;
1026 average.threads.c1 /= topo.num_cpus;
1028 average.threads.extra_delta32 /= topo.num_cpus;
1029 average.threads.extra_delta32 &= 0xFFFFFFFF;
1031 average.threads.extra_delta64 /= topo.num_cpus;
1033 average.cores.c3 /= topo.num_cores;
1034 average.cores.c6 /= topo.num_cores;
1035 average.cores.c7 /= topo.num_cores;
1037 if (do_skl_residency) {
1038 average.packages.pkg_wtd_core_c0 /= topo.num_packages;
1039 average.packages.pkg_any_core_c0 /= topo.num_packages;
1040 average.packages.pkg_any_gfxe_c0 /= topo.num_packages;
1041 average.packages.pkg_both_core_gfxe_c0 /= topo.num_packages;
1044 average.packages.pc2 /= topo.num_packages;
1046 average.packages.pc3 /= topo.num_packages;
1048 average.packages.pc6 /= topo.num_packages;
1050 average.packages.pc7 /= topo.num_packages;
1052 average.packages.pc8 /= topo.num_packages;
1053 average.packages.pc9 /= topo.num_packages;
1054 average.packages.pc10 /= topo.num_packages;
1057 static unsigned long long rdtsc(void)
1059 unsigned int low, high;
1061 asm volatile("rdtsc" : "=a" (low), "=d" (high));
1063 return low | ((unsigned long long)high) << 32;
1069 * acquire and record local counters for that cpu
1071 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1073 int cpu = t->cpu_id;
1074 unsigned long long msr;
1075 int aperf_mperf_retry_count = 0;
1077 if (cpu_migrate(cpu)) {
1078 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
1083 t->tsc = rdtsc(); /* we are running on local CPU of interest */
1086 unsigned long long tsc_before, tsc_between, tsc_after, aperf_time, mperf_time;
1089 * The TSC, APERF and MPERF must be read together for
1090 * APERF/MPERF and MPERF/TSC to give accurate results.
1092 * Unfortunately, APERF and MPERF are read by
1093 * individual system call, so delays may occur
1094 * between them. If the time to read them
1095 * varies by a large amount, we re-read them.
1099 * This initial dummy APERF read has been seen to
1100 * reduce jitter in the subsequent reads.
1103 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1106 t->tsc = rdtsc(); /* re-read close to APERF */
1108 tsc_before = t->tsc;
1110 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1113 tsc_between = rdtsc();
1115 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
1118 tsc_after = rdtsc();
1120 aperf_time = tsc_between - tsc_before;
1121 mperf_time = tsc_after - tsc_between;
1124 * If the system call latency to read APERF and MPERF
1125 * differ by more than 2x, then try again.
1127 if ((aperf_time > (2 * mperf_time)) || (mperf_time > (2 * aperf_time))) {
1128 aperf_mperf_retry_count++;
1129 if (aperf_mperf_retry_count < 5)
1132 warnx("cpu%d jitter %lld %lld",
1133 cpu, aperf_time, mperf_time);
1135 aperf_mperf_retry_count = 0;
1137 t->aperf = t->aperf * aperf_mperf_multiplier;
1138 t->mperf = t->mperf * aperf_mperf_multiplier;
1142 t->irq_count = irqs_per_cpu[cpu];
1144 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
1146 t->smi_count = msr & 0xFFFFFFFF;
1148 if (extra_delta_offset32) {
1149 if (get_msr(cpu, extra_delta_offset32, &msr))
1151 t->extra_delta32 = msr & 0xFFFFFFFF;
1154 if (extra_delta_offset64)
1155 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
1158 if (extra_msr_offset32) {
1159 if (get_msr(cpu, extra_msr_offset32, &msr))
1161 t->extra_msr32 = msr & 0xFFFFFFFF;
1164 if (extra_msr_offset64)
1165 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
1168 if (use_c1_residency_msr) {
1169 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
1173 /* collect core counters only for 1st thread in core */
1174 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1177 if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates) {
1178 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
1182 if (do_nhm_cstates && !do_knl_cstates) {
1183 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
1185 } else if (do_knl_cstates) {
1186 if (get_msr(cpu, MSR_KNL_CORE_C6_RESIDENCY, &c->c6))
1191 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
1195 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1197 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1201 /* collect package counters only for 1st core in package */
1202 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1205 if (do_skl_residency) {
1206 if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
1208 if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
1210 if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
1212 if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
1216 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
1219 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
1222 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
1225 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1228 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1230 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1232 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1235 if (do_rapl & RAPL_PKG) {
1236 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1238 p->energy_pkg = msr & 0xFFFFFFFF;
1240 if (do_rapl & RAPL_CORES) {
1241 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1243 p->energy_cores = msr & 0xFFFFFFFF;
1245 if (do_rapl & RAPL_DRAM) {
1246 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1248 p->energy_dram = msr & 0xFFFFFFFF;
1250 if (do_rapl & RAPL_GFX) {
1251 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1253 p->energy_gfx = msr & 0xFFFFFFFF;
1255 if (do_rapl & RAPL_PKG_PERF_STATUS) {
1256 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1258 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1260 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1261 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1263 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1266 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1268 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1272 p->gfx_rc6_ms = gfx_cur_rc6_ms;
1275 p->gfx_mhz = gfx_cur_mhz;
1281 * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1282 * If you change the values, note they are used both in comparisons
1283 * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1286 #define PCLUKN 0 /* Unknown */
1287 #define PCLRSV 1 /* Reserved */
1288 #define PCL__0 2 /* PC0 */
1289 #define PCL__1 3 /* PC1 */
1290 #define PCL__2 4 /* PC2 */
1291 #define PCL__3 5 /* PC3 */
1292 #define PCL__4 6 /* PC4 */
1293 #define PCL__6 7 /* PC6 */
1294 #define PCL_6N 8 /* PC6 No Retention */
1295 #define PCL_6R 9 /* PC6 Retention */
1296 #define PCL__7 10 /* PC7 */
1297 #define PCL_7S 11 /* PC7 Shrink */
1298 #define PCL__8 12 /* PC8 */
1299 #define PCL__9 13 /* PC9 */
1300 #define PCLUNL 14 /* Unlimited */
1302 int pkg_cstate_limit = PCLUKN;
1303 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1304 "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "unlimited"};
1306 int nhm_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1307 int snb_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1308 int hsw_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1309 int slv_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1310 int amt_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1311 int phi_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1312 int bxt_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1316 calculate_tsc_tweak()
1318 tsc_tweak = base_hz / tsc_hz;
1322 dump_nhm_platform_info(void)
1324 unsigned long long msr;
1327 get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
1329 fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
1331 ratio = (msr >> 40) & 0xFF;
1332 fprintf(outf, "%d * %.0f = %.0f MHz max efficiency frequency\n",
1333 ratio, bclk, ratio * bclk);
1335 ratio = (msr >> 8) & 0xFF;
1336 fprintf(outf, "%d * %.0f = %.0f MHz base frequency\n",
1337 ratio, bclk, ratio * bclk);
1339 get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
1340 fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1341 base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
1347 dump_hsw_turbo_ratio_limits(void)
1349 unsigned long long msr;
1352 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
1354 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
1356 ratio = (msr >> 8) & 0xFF;
1358 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 18 active cores\n",
1359 ratio, bclk, ratio * bclk);
1361 ratio = (msr >> 0) & 0xFF;
1363 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 17 active cores\n",
1364 ratio, bclk, ratio * bclk);
1369 dump_ivt_turbo_ratio_limits(void)
1371 unsigned long long msr;
1374 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
1376 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
1378 ratio = (msr >> 56) & 0xFF;
1380 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1381 ratio, bclk, ratio * bclk);
1383 ratio = (msr >> 48) & 0xFF;
1385 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1386 ratio, bclk, ratio * bclk);
1388 ratio = (msr >> 40) & 0xFF;
1390 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1391 ratio, bclk, ratio * bclk);
1393 ratio = (msr >> 32) & 0xFF;
1395 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1396 ratio, bclk, ratio * bclk);
1398 ratio = (msr >> 24) & 0xFF;
1400 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1401 ratio, bclk, ratio * bclk);
1403 ratio = (msr >> 16) & 0xFF;
1405 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1406 ratio, bclk, ratio * bclk);
1408 ratio = (msr >> 8) & 0xFF;
1410 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1411 ratio, bclk, ratio * bclk);
1413 ratio = (msr >> 0) & 0xFF;
1415 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1416 ratio, bclk, ratio * bclk);
1421 dump_nhm_turbo_ratio_limits(void)
1423 unsigned long long msr;
1426 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
1428 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
1430 ratio = (msr >> 56) & 0xFF;
1432 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1433 ratio, bclk, ratio * bclk);
1435 ratio = (msr >> 48) & 0xFF;
1437 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1438 ratio, bclk, ratio * bclk);
1440 ratio = (msr >> 40) & 0xFF;
1442 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1443 ratio, bclk, ratio * bclk);
1445 ratio = (msr >> 32) & 0xFF;
1447 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1448 ratio, bclk, ratio * bclk);
1450 ratio = (msr >> 24) & 0xFF;
1452 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1453 ratio, bclk, ratio * bclk);
1455 ratio = (msr >> 16) & 0xFF;
1457 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1458 ratio, bclk, ratio * bclk);
1460 ratio = (msr >> 8) & 0xFF;
1462 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1463 ratio, bclk, ratio * bclk);
1465 ratio = (msr >> 0) & 0xFF;
1467 fprintf(outf, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1468 ratio, bclk, ratio * bclk);
1473 dump_knl_turbo_ratio_limits(void)
1475 const unsigned int buckets_no = 7;
1477 unsigned long long msr;
1478 int delta_cores, delta_ratio;
1480 unsigned int cores[buckets_no];
1481 unsigned int ratio[buckets_no];
1483 get_msr(base_cpu, MSR_NHM_TURBO_RATIO_LIMIT, &msr);
1485 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n",
1489 * Turbo encoding in KNL is as follows:
1491 * [7:1] -- Base value of number of active cores of bucket 1.
1492 * [15:8] -- Base value of freq ratio of bucket 1.
1493 * [20:16] -- +ve delta of number of active cores of bucket 2.
1494 * i.e. active cores of bucket 2 =
1495 * active cores of bucket 1 + delta
1496 * [23:21] -- Negative delta of freq ratio of bucket 2.
1497 * i.e. freq ratio of bucket 2 =
1498 * freq ratio of bucket 1 - delta
1499 * [28:24]-- +ve delta of number of active cores of bucket 3.
1500 * [31:29]-- -ve delta of freq ratio of bucket 3.
1501 * [36:32]-- +ve delta of number of active cores of bucket 4.
1502 * [39:37]-- -ve delta of freq ratio of bucket 4.
1503 * [44:40]-- +ve delta of number of active cores of bucket 5.
1504 * [47:45]-- -ve delta of freq ratio of bucket 5.
1505 * [52:48]-- +ve delta of number of active cores of bucket 6.
1506 * [55:53]-- -ve delta of freq ratio of bucket 6.
1507 * [60:56]-- +ve delta of number of active cores of bucket 7.
1508 * [63:61]-- -ve delta of freq ratio of bucket 7.
1512 cores[b_nr] = (msr & 0xFF) >> 1;
1513 ratio[b_nr] = (msr >> 8) & 0xFF;
1515 for (i = 16; i < 64; i += 8) {
1516 delta_cores = (msr >> i) & 0x1F;
1517 delta_ratio = (msr >> (i + 5)) & 0x7;
1519 cores[b_nr + 1] = cores[b_nr] + delta_cores;
1520 ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
1524 for (i = buckets_no - 1; i >= 0; i--)
1525 if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
1527 "%d * %.0f = %.0f MHz max turbo %d active cores\n",
1528 ratio[i], bclk, ratio[i] * bclk, cores[i]);
1532 dump_nhm_cst_cfg(void)
1534 unsigned long long msr;
1536 get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1538 #define SNB_C1_AUTO_UNDEMOTE (1UL << 27)
1539 #define SNB_C3_AUTO_UNDEMOTE (1UL << 28)
1541 fprintf(outf, "cpu%d: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", base_cpu, msr);
1543 fprintf(outf, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: %s)\n",
1544 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1545 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1546 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1547 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1548 (msr & (1 << 15)) ? "" : "UN",
1549 (unsigned int)msr & 0xF,
1550 pkg_cstate_limit_strings[pkg_cstate_limit]);
1555 dump_config_tdp(void)
1557 unsigned long long msr;
1559 get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
1560 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
1561 fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
1563 get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
1564 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
1566 fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
1567 fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
1568 fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
1569 fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
1571 fprintf(outf, ")\n");
1573 get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
1574 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
1576 fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
1577 fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
1578 fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
1579 fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
1581 fprintf(outf, ")\n");
1583 get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
1584 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
1586 fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
1587 fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
1588 fprintf(outf, ")\n");
1590 get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
1591 fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
1592 fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
1593 fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
1594 fprintf(outf, ")\n");
1597 unsigned int irtl_time_units[] = {1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
1599 void print_irtl(void)
1601 unsigned long long msr;
1603 get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
1604 fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
1605 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1606 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1608 get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
1609 fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
1610 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1611 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1613 get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
1614 fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
1615 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1616 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1621 get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
1622 fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
1623 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1624 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1626 get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
1627 fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
1628 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1629 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1631 get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
1632 fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
1633 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1634 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1637 void free_fd_percpu(void)
1641 for (i = 0; i < topo.max_cpu_num; ++i) {
1642 if (fd_percpu[i] != 0)
1643 close(fd_percpu[i]);
1649 void free_all_buffers(void)
1651 CPU_FREE(cpu_present_set);
1652 cpu_present_set = NULL;
1653 cpu_present_setsize = 0;
1655 CPU_FREE(cpu_affinity_set);
1656 cpu_affinity_set = NULL;
1657 cpu_affinity_setsize = 0;
1665 package_even = NULL;
1675 free(output_buffer);
1676 output_buffer = NULL;
1681 free(irq_column_2_cpu);
1686 * Open a file, and exit on failure
1688 FILE *fopen_or_die(const char *path, const char *mode)
1690 FILE *filep = fopen(path, mode);
1692 err(1, "%s: open failed", path);
1697 * Parse a file containing a single int.
1699 int parse_int_file(const char *fmt, ...)
1702 char path[PATH_MAX];
1706 va_start(args, fmt);
1707 vsnprintf(path, sizeof(path), fmt, args);
1709 filep = fopen_or_die(path, "r");
1710 if (fscanf(filep, "%d", &value) != 1)
1711 err(1, "%s: failed to parse number from file", path);
1717 * get_cpu_position_in_core(cpu)
1718 * return the position of the CPU among its HT siblings in the core
1719 * return -1 if the sibling is not in list
1721 int get_cpu_position_in_core(int cpu)
1730 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list",
1732 filep = fopen(path, "r");
1733 if (filep == NULL) {
1738 for (i = 0; i < topo.num_threads_per_core; i++) {
1739 fscanf(filep, "%d", &this_cpu);
1740 if (this_cpu == cpu) {
1745 /* Account for no separator after last thread*/
1746 if (i != (topo.num_threads_per_core - 1))
1747 fscanf(filep, "%c", &character);
1755 * cpu_is_first_core_in_package(cpu)
1756 * return 1 if given CPU is 1st core in package
1758 int cpu_is_first_core_in_package(int cpu)
1760 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1763 int get_physical_package_id(int cpu)
1765 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1768 int get_core_id(int cpu)
1770 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1773 int get_num_ht_siblings(int cpu)
1783 sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1784 filep = fopen_or_die(path, "r");
1788 * A ',' separated or '-' separated set of numbers
1789 * (eg 1-2 or 1,3,4,5)
1791 fscanf(filep, "%d%c\n", &sib1, &character);
1792 fseek(filep, 0, SEEK_SET);
1793 fgets(str, 100, filep);
1794 ch = strchr(str, character);
1795 while (ch != NULL) {
1797 ch = strchr(ch+1, character);
1805 * run func(thread, core, package) in topology order
1806 * skip non-present cpus
1809 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1810 struct pkg_data *, struct thread_data *, struct core_data *,
1811 struct pkg_data *), struct thread_data *thread_base,
1812 struct core_data *core_base, struct pkg_data *pkg_base,
1813 struct thread_data *thread_base2, struct core_data *core_base2,
1814 struct pkg_data *pkg_base2)
1816 int retval, pkg_no, core_no, thread_no;
1818 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1819 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1820 for (thread_no = 0; thread_no <
1821 topo.num_threads_per_core; ++thread_no) {
1822 struct thread_data *t, *t2;
1823 struct core_data *c, *c2;
1824 struct pkg_data *p, *p2;
1826 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1828 if (cpu_is_not_present(t->cpu_id))
1831 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1833 c = GET_CORE(core_base, core_no, pkg_no);
1834 c2 = GET_CORE(core_base2, core_no, pkg_no);
1836 p = GET_PKG(pkg_base, pkg_no);
1837 p2 = GET_PKG(pkg_base2, pkg_no);
1839 retval = func(t, c, p, t2, c2, p2);
1849 * run func(cpu) on every cpu in /proc/stat
1850 * return max_cpu number
1852 int for_all_proc_cpus(int (func)(int))
1858 fp = fopen_or_die(proc_stat, "r");
1860 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1862 err(1, "%s: failed to parse format", proc_stat);
1865 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1869 retval = func(cpu_num);
1879 void re_initialize(void)
1882 setup_all_buffers();
1883 printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1889 * remember the last one seen, it will be the max
1891 int count_cpus(int cpu)
1893 if (topo.max_cpu_num < cpu)
1894 topo.max_cpu_num = cpu;
1899 int mark_cpu_present(int cpu)
1901 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1906 * snapshot_proc_interrupts()
1908 * read and record summary of /proc/interrupts
1910 * return 1 if config change requires a restart, else return 0
1912 int snapshot_proc_interrupts(void)
1918 fp = fopen_or_die("/proc/interrupts", "r");
1922 /* read 1st line of /proc/interrupts to get cpu* name for each column */
1923 for (column = 0; column < topo.num_cpus; ++column) {
1926 retval = fscanf(fp, " CPU%d", &cpu_number);
1930 if (cpu_number > topo.max_cpu_num) {
1931 warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
1935 irq_column_2_cpu[column] = cpu_number;
1936 irqs_per_cpu[cpu_number] = 0;
1939 /* read /proc/interrupt count lines and sum up irqs per cpu */
1944 retval = fscanf(fp, " %s:", buf); /* flush irq# "N:" */
1948 /* read the count per cpu */
1949 for (column = 0; column < topo.num_cpus; ++column) {
1951 int cpu_number, irq_count;
1953 retval = fscanf(fp, " %d", &irq_count);
1957 cpu_number = irq_column_2_cpu[column];
1958 irqs_per_cpu[cpu_number] += irq_count;
1962 while (getc(fp) != '\n')
1963 ; /* flush interrupt description */
1969 * snapshot_gfx_rc6_ms()
1971 * record snapshot of
1972 * /sys/class/drm/card0/power/rc6_residency_ms
1974 * return 1 if config change requires a restart, else return 0
1976 int snapshot_gfx_rc6_ms(void)
1981 fp = fopen_or_die("/sys/class/drm/card0/power/rc6_residency_ms", "r");
1983 retval = fscanf(fp, "%lld", &gfx_cur_rc6_ms);
1992 * snapshot_gfx_mhz()
1994 * record snapshot of
1995 * /sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz
1997 * return 1 if config change requires a restart, else return 0
1999 int snapshot_gfx_mhz(void)
2005 fp = fopen_or_die("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", "r");
2009 retval = fscanf(fp, "%d", &gfx_cur_mhz);
2017 * snapshot /proc and /sys files
2019 * return 1 if configuration restart needed, else return 0
2021 int snapshot_proc_sysfs_files(void)
2023 if (snapshot_proc_interrupts())
2027 snapshot_gfx_rc6_ms();
2035 void turbostat_loop()
2043 snapshot_proc_sysfs_files();
2044 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
2047 } else if (retval == -1) {
2048 if (restarted > 1) {
2055 gettimeofday(&tv_even, (struct timezone *)NULL);
2058 if (for_all_proc_cpus(cpu_is_not_present)) {
2062 nanosleep(&interval_ts, NULL);
2063 if (snapshot_proc_sysfs_files())
2065 retval = for_all_cpus(get_counters, ODD_COUNTERS);
2068 } else if (retval == -1) {
2072 gettimeofday(&tv_odd, (struct timezone *)NULL);
2073 timersub(&tv_odd, &tv_even, &tv_delta);
2074 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
2075 compute_average(EVEN_COUNTERS);
2076 format_all_counters(EVEN_COUNTERS);
2077 flush_output_stdout();
2078 nanosleep(&interval_ts, NULL);
2079 if (snapshot_proc_sysfs_files())
2081 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
2084 } else if (retval == -1) {
2088 gettimeofday(&tv_even, (struct timezone *)NULL);
2089 timersub(&tv_even, &tv_odd, &tv_delta);
2090 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
2091 compute_average(ODD_COUNTERS);
2092 format_all_counters(ODD_COUNTERS);
2093 flush_output_stdout();
2097 void check_dev_msr()
2102 sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
2103 if (stat(pathname, &sb))
2104 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
2105 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
2108 void check_permissions()
2110 struct __user_cap_header_struct cap_header_data;
2111 cap_user_header_t cap_header = &cap_header_data;
2112 struct __user_cap_data_struct cap_data_data;
2113 cap_user_data_t cap_data = &cap_data_data;
2114 extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
2118 /* check for CAP_SYS_RAWIO */
2119 cap_header->pid = getpid();
2120 cap_header->version = _LINUX_CAPABILITY_VERSION;
2121 if (capget(cap_header, cap_data) < 0)
2122 err(-6, "capget(2) failed");
2124 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
2126 warnx("capget(CAP_SYS_RAWIO) failed,"
2127 " try \"# setcap cap_sys_rawio=ep %s\"", progname);
2130 /* test file permissions */
2131 sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
2132 if (euidaccess(pathname, R_OK)) {
2134 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
2137 /* if all else fails, thell them to be root */
2140 warnx("... or simply run as root");
2147 * NHM adds support for additional MSRs:
2149 * MSR_SMI_COUNT 0x00000034
2151 * MSR_PLATFORM_INFO 0x000000ce
2152 * MSR_NHM_SNB_PKG_CST_CFG_CTL 0x000000e2
2154 * MSR_PKG_C3_RESIDENCY 0x000003f8
2155 * MSR_PKG_C6_RESIDENCY 0x000003f9
2156 * MSR_CORE_C3_RESIDENCY 0x000003fc
2157 * MSR_CORE_C6_RESIDENCY 0x000003fd
2160 * sets global pkg_cstate_limit to decode MSR_NHM_SNB_PKG_CST_CFG_CTL
2162 int probe_nhm_msrs(unsigned int family, unsigned int model)
2164 unsigned long long msr;
2165 unsigned int base_ratio;
2166 int *pkg_cstate_limits;
2174 bclk = discover_bclk(family, model);
2177 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
2178 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
2179 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
2180 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
2181 case 0x2C: /* Westmere EP - Gulftown */
2182 case 0x2E: /* Nehalem-EX Xeon - Beckton */
2183 case 0x2F: /* Westmere-EX Xeon - Eagleton */
2184 pkg_cstate_limits = nhm_pkg_cstate_limits;
2186 case 0x2A: /* SNB */
2187 case 0x2D: /* SNB Xeon */
2188 case 0x3A: /* IVB */
2189 case 0x3E: /* IVB Xeon */
2190 pkg_cstate_limits = snb_pkg_cstate_limits;
2192 case 0x3C: /* HSW */
2193 case 0x3F: /* HSX */
2194 case 0x45: /* HSW */
2195 case 0x46: /* HSW */
2196 case 0x3D: /* BDW */
2197 case 0x47: /* BDW */
2198 case 0x4F: /* BDX */
2199 case 0x56: /* BDX-DE */
2200 case 0x4E: /* SKL */
2201 case 0x5E: /* SKL */
2202 case 0x8E: /* KBL */
2203 case 0x9E: /* KBL */
2204 case 0x55: /* SKX */
2205 pkg_cstate_limits = hsw_pkg_cstate_limits;
2207 case 0x37: /* BYT */
2208 case 0x4D: /* AVN */
2209 pkg_cstate_limits = slv_pkg_cstate_limits;
2211 case 0x4C: /* AMT */
2212 pkg_cstate_limits = amt_pkg_cstate_limits;
2214 case 0x57: /* PHI */
2215 pkg_cstate_limits = phi_pkg_cstate_limits;
2217 case 0x5C: /* BXT */
2218 pkg_cstate_limits = bxt_pkg_cstate_limits;
2223 get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
2224 pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
2226 get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
2227 base_ratio = (msr >> 8) & 0xFF;
2229 base_hz = base_ratio * bclk * 1000000;
2233 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
2236 /* Nehalem compatible, but do not include turbo-ratio limit support */
2237 case 0x2E: /* Nehalem-EX Xeon - Beckton */
2238 case 0x2F: /* Westmere-EX Xeon - Eagleton */
2239 case 0x57: /* PHI - Knights Landing (different MSR definition) */
2245 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
2254 case 0x3E: /* IVB Xeon */
2255 case 0x3F: /* HSW Xeon */
2261 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
2270 case 0x3F: /* HSW Xeon */
2277 int has_knl_turbo_ratio_limit(unsigned int family, unsigned int model)
2286 case 0x57: /* Knights Landing */
2292 int has_config_tdp(unsigned int family, unsigned int model)
2301 case 0x3A: /* IVB */
2302 case 0x3C: /* HSW */
2303 case 0x3F: /* HSX */
2304 case 0x45: /* HSW */
2305 case 0x46: /* HSW */
2306 case 0x3D: /* BDW */
2307 case 0x47: /* BDW */
2308 case 0x4F: /* BDX */
2309 case 0x56: /* BDX-DE */
2310 case 0x4E: /* SKL */
2311 case 0x5E: /* SKL */
2312 case 0x8E: /* KBL */
2313 case 0x9E: /* KBL */
2314 case 0x55: /* SKX */
2316 case 0x57: /* Knights Landing */
2324 dump_cstate_pstate_config_info(int family, int model)
2326 if (!do_nhm_platform_info)
2329 dump_nhm_platform_info();
2331 if (has_hsw_turbo_ratio_limit(family, model))
2332 dump_hsw_turbo_ratio_limits();
2334 if (has_ivt_turbo_ratio_limit(family, model))
2335 dump_ivt_turbo_ratio_limits();
2337 if (has_nhm_turbo_ratio_limit(family, model))
2338 dump_nhm_turbo_ratio_limits();
2340 if (has_knl_turbo_ratio_limit(family, model))
2341 dump_knl_turbo_ratio_limits();
2343 if (has_config_tdp(family, model))
2352 * Decode the ENERGY_PERF_BIAS MSR
2354 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2356 unsigned long long msr;
2365 /* EPB is per-package */
2366 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2369 if (cpu_migrate(cpu)) {
2370 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2374 if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
2377 switch (msr & 0xF) {
2378 case ENERGY_PERF_BIAS_PERFORMANCE:
2379 epb_string = "performance";
2381 case ENERGY_PERF_BIAS_NORMAL:
2382 epb_string = "balanced";
2384 case ENERGY_PERF_BIAS_POWERSAVE:
2385 epb_string = "powersave";
2388 epb_string = "custom";
2391 fprintf(outf, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
2397 * Decode the MSR_HWP_CAPABILITIES
2399 int print_hwp(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2401 unsigned long long msr;
2409 /* MSR_HWP_CAPABILITIES is per-package */
2410 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2413 if (cpu_migrate(cpu)) {
2414 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2418 if (get_msr(cpu, MSR_PM_ENABLE, &msr))
2421 fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n",
2422 cpu, msr, (msr & (1 << 0)) ? "" : "No-");
2424 /* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
2425 if ((msr & (1 << 0)) == 0)
2428 if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
2431 fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
2432 "(high 0x%x guar 0x%x eff 0x%x low 0x%x)\n",
2434 (unsigned int)HWP_HIGHEST_PERF(msr),
2435 (unsigned int)HWP_GUARANTEED_PERF(msr),
2436 (unsigned int)HWP_MOSTEFFICIENT_PERF(msr),
2437 (unsigned int)HWP_LOWEST_PERF(msr));
2439 if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
2442 fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
2443 "(min 0x%x max 0x%x des 0x%x epp 0x%x window 0x%x pkg 0x%x)\n",
2445 (unsigned int)(((msr) >> 0) & 0xff),
2446 (unsigned int)(((msr) >> 8) & 0xff),
2447 (unsigned int)(((msr) >> 16) & 0xff),
2448 (unsigned int)(((msr) >> 24) & 0xff),
2449 (unsigned int)(((msr) >> 32) & 0xff3),
2450 (unsigned int)(((msr) >> 42) & 0x1));
2453 if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
2456 fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
2457 "(min 0x%x max 0x%x des 0x%x epp 0x%x window 0x%x)\n",
2459 (unsigned int)(((msr) >> 0) & 0xff),
2460 (unsigned int)(((msr) >> 8) & 0xff),
2461 (unsigned int)(((msr) >> 16) & 0xff),
2462 (unsigned int)(((msr) >> 24) & 0xff),
2463 (unsigned int)(((msr) >> 32) & 0xff3));
2465 if (has_hwp_notify) {
2466 if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
2469 fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
2470 "(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
2472 ((msr) & 0x1) ? "EN" : "Dis",
2473 ((msr) & 0x2) ? "EN" : "Dis");
2475 if (get_msr(cpu, MSR_HWP_STATUS, &msr))
2478 fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
2479 "(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
2481 ((msr) & 0x1) ? "" : "No-",
2482 ((msr) & 0x2) ? "" : "No-");
2488 * print_perf_limit()
2490 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2492 unsigned long long msr;
2498 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2501 if (cpu_migrate(cpu)) {
2502 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2506 if (do_core_perf_limit_reasons) {
2507 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
2508 fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2509 fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
2510 (msr & 1 << 15) ? "bit15, " : "",
2511 (msr & 1 << 14) ? "bit14, " : "",
2512 (msr & 1 << 13) ? "Transitions, " : "",
2513 (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
2514 (msr & 1 << 11) ? "PkgPwrL2, " : "",
2515 (msr & 1 << 10) ? "PkgPwrL1, " : "",
2516 (msr & 1 << 9) ? "CorePwr, " : "",
2517 (msr & 1 << 8) ? "Amps, " : "",
2518 (msr & 1 << 6) ? "VR-Therm, " : "",
2519 (msr & 1 << 5) ? "Auto-HWP, " : "",
2520 (msr & 1 << 4) ? "Graphics, " : "",
2521 (msr & 1 << 2) ? "bit2, " : "",
2522 (msr & 1 << 1) ? "ThermStatus, " : "",
2523 (msr & 1 << 0) ? "PROCHOT, " : "");
2524 fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
2525 (msr & 1 << 31) ? "bit31, " : "",
2526 (msr & 1 << 30) ? "bit30, " : "",
2527 (msr & 1 << 29) ? "Transitions, " : "",
2528 (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
2529 (msr & 1 << 27) ? "PkgPwrL2, " : "",
2530 (msr & 1 << 26) ? "PkgPwrL1, " : "",
2531 (msr & 1 << 25) ? "CorePwr, " : "",
2532 (msr & 1 << 24) ? "Amps, " : "",
2533 (msr & 1 << 22) ? "VR-Therm, " : "",
2534 (msr & 1 << 21) ? "Auto-HWP, " : "",
2535 (msr & 1 << 20) ? "Graphics, " : "",
2536 (msr & 1 << 18) ? "bit18, " : "",
2537 (msr & 1 << 17) ? "ThermStatus, " : "",
2538 (msr & 1 << 16) ? "PROCHOT, " : "");
2541 if (do_gfx_perf_limit_reasons) {
2542 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
2543 fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2544 fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
2545 (msr & 1 << 0) ? "PROCHOT, " : "",
2546 (msr & 1 << 1) ? "ThermStatus, " : "",
2547 (msr & 1 << 4) ? "Graphics, " : "",
2548 (msr & 1 << 6) ? "VR-Therm, " : "",
2549 (msr & 1 << 8) ? "Amps, " : "",
2550 (msr & 1 << 9) ? "GFXPwr, " : "",
2551 (msr & 1 << 10) ? "PkgPwrL1, " : "",
2552 (msr & 1 << 11) ? "PkgPwrL2, " : "");
2553 fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
2554 (msr & 1 << 16) ? "PROCHOT, " : "",
2555 (msr & 1 << 17) ? "ThermStatus, " : "",
2556 (msr & 1 << 20) ? "Graphics, " : "",
2557 (msr & 1 << 22) ? "VR-Therm, " : "",
2558 (msr & 1 << 24) ? "Amps, " : "",
2559 (msr & 1 << 25) ? "GFXPwr, " : "",
2560 (msr & 1 << 26) ? "PkgPwrL1, " : "",
2561 (msr & 1 << 27) ? "PkgPwrL2, " : "");
2563 if (do_ring_perf_limit_reasons) {
2564 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
2565 fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2566 fprintf(outf, " (Active: %s%s%s%s%s%s)",
2567 (msr & 1 << 0) ? "PROCHOT, " : "",
2568 (msr & 1 << 1) ? "ThermStatus, " : "",
2569 (msr & 1 << 6) ? "VR-Therm, " : "",
2570 (msr & 1 << 8) ? "Amps, " : "",
2571 (msr & 1 << 10) ? "PkgPwrL1, " : "",
2572 (msr & 1 << 11) ? "PkgPwrL2, " : "");
2573 fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
2574 (msr & 1 << 16) ? "PROCHOT, " : "",
2575 (msr & 1 << 17) ? "ThermStatus, " : "",
2576 (msr & 1 << 22) ? "VR-Therm, " : "",
2577 (msr & 1 << 24) ? "Amps, " : "",
2578 (msr & 1 << 26) ? "PkgPwrL1, " : "",
2579 (msr & 1 << 27) ? "PkgPwrL2, " : "");
2584 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
2585 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
2587 double get_tdp(int model)
2589 unsigned long long msr;
2591 if (do_rapl & RAPL_PKG_POWER_INFO)
2592 if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
2593 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
2605 * rapl_dram_energy_units_probe()
2606 * Energy units are either hard-coded, or come from RAPL Energy Unit MSR.
2609 rapl_dram_energy_units_probe(int model, double rapl_energy_units)
2611 /* only called for genuine_intel, family 6 */
2614 case 0x3F: /* HSX */
2615 case 0x4F: /* BDX */
2616 case 0x56: /* BDX-DE */
2617 case 0x57: /* KNL */
2618 return (rapl_dram_energy_units = 15.3 / 1000000);
2620 return (rapl_energy_units);
2628 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
2630 void rapl_probe(unsigned int family, unsigned int model)
2632 unsigned long long msr;
2633 unsigned int time_unit;
2645 case 0x3C: /* HSW */
2646 case 0x45: /* HSW */
2647 case 0x46: /* HSW */
2648 case 0x3D: /* BDW */
2649 case 0x47: /* BDW */
2650 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
2652 case 0x5C: /* BXT */
2653 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO;
2655 case 0x4E: /* SKL */
2656 case 0x5E: /* SKL */
2657 case 0x8E: /* KBL */
2658 case 0x9E: /* KBL */
2659 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2661 case 0x3F: /* HSX */
2662 case 0x4F: /* BDX */
2663 case 0x56: /* BDX-DE */
2664 case 0x55: /* SKX */
2665 case 0x57: /* KNL */
2666 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2670 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
2672 case 0x37: /* BYT */
2673 case 0x4D: /* AVN */
2674 do_rapl = RAPL_PKG | RAPL_CORES ;
2680 /* units on package 0, verify later other packages match */
2681 if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
2684 rapl_power_units = 1.0 / (1 << (msr & 0xF));
2686 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
2688 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
2690 rapl_dram_energy_units = rapl_dram_energy_units_probe(model, rapl_energy_units);
2692 time_unit = msr >> 16 & 0xF;
2696 rapl_time_units = 1.0 / (1 << (time_unit));
2698 tdp = get_tdp(model);
2700 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
2702 fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
2707 void perf_limit_reasons_probe(int family, int model)
2716 case 0x3C: /* HSW */
2717 case 0x45: /* HSW */
2718 case 0x46: /* HSW */
2719 do_gfx_perf_limit_reasons = 1;
2720 case 0x3F: /* HSX */
2721 do_core_perf_limit_reasons = 1;
2722 do_ring_perf_limit_reasons = 1;
2728 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2730 unsigned long long msr;
2734 if (!(do_dts || do_ptm))
2739 /* DTS is per-core, no need to print for each thread */
2740 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
2743 if (cpu_migrate(cpu)) {
2744 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2748 if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
2749 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
2752 dts = (msr >> 16) & 0x7F;
2753 fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
2754 cpu, msr, tcc_activation_temp - dts);
2757 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
2760 dts = (msr >> 16) & 0x7F;
2761 dts2 = (msr >> 8) & 0x7F;
2762 fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2763 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2769 unsigned int resolution;
2771 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
2774 dts = (msr >> 16) & 0x7F;
2775 resolution = (msr >> 27) & 0xF;
2776 fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
2777 cpu, msr, tcc_activation_temp - dts, resolution);
2780 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
2783 dts = (msr >> 16) & 0x7F;
2784 dts2 = (msr >> 8) & 0x7F;
2785 fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2786 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2793 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
2795 fprintf(outf, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
2797 ((msr >> 15) & 1) ? "EN" : "DIS",
2798 ((msr >> 0) & 0x7FFF) * rapl_power_units,
2799 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
2800 (((msr >> 16) & 1) ? "EN" : "DIS"));
2805 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2807 unsigned long long msr;
2813 /* RAPL counters are per package, so print only for 1st thread/package */
2814 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2818 if (cpu_migrate(cpu)) {
2819 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2823 if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
2827 fprintf(outf, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
2828 "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
2829 rapl_power_units, rapl_energy_units, rapl_time_units);
2831 if (do_rapl & RAPL_PKG_POWER_INFO) {
2833 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
2837 fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2839 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2840 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2841 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2842 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2845 if (do_rapl & RAPL_PKG) {
2847 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
2850 fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
2851 cpu, msr, (msr >> 63) & 1 ? "": "UN");
2853 print_power_limit_msr(cpu, msr, "PKG Limit #1");
2854 fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
2856 ((msr >> 47) & 1) ? "EN" : "DIS",
2857 ((msr >> 32) & 0x7FFF) * rapl_power_units,
2858 (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
2859 ((msr >> 48) & 1) ? "EN" : "DIS");
2862 if (do_rapl & RAPL_DRAM_POWER_INFO) {
2863 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
2866 fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2868 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2869 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2870 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2871 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2873 if (do_rapl & RAPL_DRAM) {
2874 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
2876 fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
2877 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2879 print_power_limit_msr(cpu, msr, "DRAM Limit");
2881 if (do_rapl & RAPL_CORE_POLICY) {
2883 if (get_msr(cpu, MSR_PP0_POLICY, &msr))
2886 fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
2889 if (do_rapl & RAPL_CORES) {
2892 if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
2894 fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2895 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2896 print_power_limit_msr(cpu, msr, "Cores Limit");
2899 if (do_rapl & RAPL_GFX) {
2901 if (get_msr(cpu, MSR_PP1_POLICY, &msr))
2904 fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
2906 if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
2908 fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2909 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2910 print_power_limit_msr(cpu, msr, "GFX Limit");
2917 * SNB adds support for additional MSRs:
2919 * MSR_PKG_C7_RESIDENCY 0x000003fa
2920 * MSR_CORE_C7_RESIDENCY 0x000003fe
2921 * MSR_PKG_C2_RESIDENCY 0x0000060d
2924 int has_snb_msrs(unsigned int family, unsigned int model)
2932 case 0x3A: /* IVB */
2933 case 0x3E: /* IVB Xeon */
2934 case 0x3C: /* HSW */
2935 case 0x3F: /* HSW */
2936 case 0x45: /* HSW */
2937 case 0x46: /* HSW */
2938 case 0x3D: /* BDW */
2939 case 0x47: /* BDW */
2940 case 0x4F: /* BDX */
2941 case 0x56: /* BDX-DE */
2942 case 0x4E: /* SKL */
2943 case 0x5E: /* SKL */
2944 case 0x8E: /* KBL */
2945 case 0x9E: /* KBL */
2946 case 0x55: /* SKX */
2947 case 0x5C: /* BXT */
2954 * HSW adds support for additional MSRs:
2956 * MSR_PKG_C8_RESIDENCY 0x00000630
2957 * MSR_PKG_C9_RESIDENCY 0x00000631
2958 * MSR_PKG_C10_RESIDENCY 0x00000632
2960 * MSR_PKGC8_IRTL 0x00000633
2961 * MSR_PKGC9_IRTL 0x00000634
2962 * MSR_PKGC10_IRTL 0x00000635
2965 int has_hsw_msrs(unsigned int family, unsigned int model)
2971 case 0x45: /* HSW */
2972 case 0x3D: /* BDW */
2973 case 0x4E: /* SKL */
2974 case 0x5E: /* SKL */
2975 case 0x8E: /* KBL */
2976 case 0x9E: /* KBL */
2977 case 0x5C: /* BXT */
2984 * SKL adds support for additional MSRS:
2986 * MSR_PKG_WEIGHTED_CORE_C0_RES 0x00000658
2987 * MSR_PKG_ANY_CORE_C0_RES 0x00000659
2988 * MSR_PKG_ANY_GFXE_C0_RES 0x0000065A
2989 * MSR_PKG_BOTH_CORE_GFXE_C0_RES 0x0000065B
2991 int has_skl_msrs(unsigned int family, unsigned int model)
2997 case 0x4E: /* SKL */
2998 case 0x5E: /* SKL */
2999 case 0x8E: /* KBL */
3000 case 0x9E: /* KBL */
3008 int is_slm(unsigned int family, unsigned int model)
3013 case 0x37: /* BYT */
3014 case 0x4D: /* AVN */
3020 int is_knl(unsigned int family, unsigned int model)
3025 case 0x57: /* KNL */
3031 unsigned int get_aperf_mperf_multiplier(unsigned int family, unsigned int model)
3033 if (is_knl(family, model))
3038 #define SLM_BCLK_FREQS 5
3039 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
3041 double slm_bclk(void)
3043 unsigned long long msr = 3;
3047 if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
3048 fprintf(outf, "SLM BCLK: unknown\n");
3051 if (i >= SLM_BCLK_FREQS) {
3052 fprintf(outf, "SLM BCLK[%d] invalid\n", i);
3055 freq = slm_freq_table[i];
3057 fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
3062 double discover_bclk(unsigned int family, unsigned int model)
3064 if (has_snb_msrs(family, model) || is_knl(family, model))
3066 else if (is_slm(family, model))
3073 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
3074 * the Thermal Control Circuit (TCC) activates.
3075 * This is usually equal to tjMax.
3077 * Older processors do not have this MSR, so there we guess,
3078 * but also allow cmdline over-ride with -T.
3080 * Several MSR temperature values are in units of degrees-C
3081 * below this value, including the Digital Thermal Sensor (DTS),
3082 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
3084 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3086 unsigned long long msr;
3087 unsigned int target_c_local;
3090 /* tcc_activation_temp is used only for dts or ptm */
3091 if (!(do_dts || do_ptm))
3094 /* this is a per-package concept */
3095 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
3099 if (cpu_migrate(cpu)) {
3100 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
3104 if (tcc_activation_temp_override != 0) {
3105 tcc_activation_temp = tcc_activation_temp_override;
3106 fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n",
3107 cpu, tcc_activation_temp);
3111 /* Temperature Target MSR is Nehalem and newer only */
3112 if (!do_nhm_platform_info)
3115 if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
3118 target_c_local = (msr >> 16) & 0xFF;
3121 fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
3122 cpu, msr, target_c_local);
3124 if (!target_c_local)
3127 tcc_activation_temp = target_c_local;
3132 tcc_activation_temp = TJMAX_DEFAULT;
3133 fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
3134 cpu, tcc_activation_temp);
3139 void decode_feature_control_msr(void)
3141 unsigned long long msr;
3143 if (!get_msr(base_cpu, MSR_IA32_FEATURE_CONTROL, &msr))
3144 fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
3146 msr & FEATURE_CONTROL_LOCKED ? "" : "UN-",
3147 msr & (1 << 18) ? "SGX" : "");
3150 void decode_misc_enable_msr(void)
3152 unsigned long long msr;
3154 if (!get_msr(base_cpu, MSR_IA32_MISC_ENABLE, &msr))
3155 fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%s %s %s)\n",
3157 msr & (1 << 3) ? "TCC" : "",
3158 msr & (1 << 16) ? "EIST" : "",
3159 msr & (1 << 18) ? "MONITOR" : "");
3163 * Decode MSR_MISC_PWR_MGMT
3165 * Decode the bits according to the Nehalem documentation
3166 * bit[0] seems to continue to have same meaning going forward
3169 void decode_misc_pwr_mgmt_msr(void)
3171 unsigned long long msr;
3173 if (!do_nhm_platform_info)
3176 if (!get_msr(base_cpu, MSR_MISC_PWR_MGMT, &msr))
3177 fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB)\n",
3179 msr & (1 << 0) ? "DIS" : "EN",
3180 msr & (1 << 1) ? "EN" : "DIS");
3183 void process_cpuid()
3185 unsigned int eax, ebx, ecx, edx, max_level, max_extended_level;
3186 unsigned int fms, family, model, stepping;
3188 eax = ebx = ecx = edx = 0;
3190 __cpuid(0, max_level, ebx, ecx, edx);
3192 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
3196 fprintf(outf, "CPUID(0): %.4s%.4s%.4s ",
3197 (char *)&ebx, (char *)&edx, (char *)&ecx);
3199 __cpuid(1, fms, ebx, ecx, edx);
3200 family = (fms >> 8) & 0xf;
3201 model = (fms >> 4) & 0xf;
3202 stepping = fms & 0xf;
3203 if (family == 6 || family == 0xf)
3204 model += ((fms >> 16) & 0xf) << 4;
3207 fprintf(outf, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
3208 max_level, family, model, stepping, family, model, stepping);
3209 fprintf(outf, "CPUID(1): %s %s %s %s %s %s %s %s %s\n",
3210 ecx & (1 << 0) ? "SSE3" : "-",
3211 ecx & (1 << 3) ? "MONITOR" : "-",
3212 ecx & (1 << 6) ? "SMX" : "-",
3213 ecx & (1 << 7) ? "EIST" : "-",
3214 ecx & (1 << 8) ? "TM2" : "-",
3215 edx & (1 << 4) ? "TSC" : "-",
3216 edx & (1 << 5) ? "MSR" : "-",
3217 edx & (1 << 22) ? "ACPI-TM" : "-",
3218 edx & (1 << 29) ? "TM" : "-");
3221 if (!(edx & (1 << 5)))
3222 errx(1, "CPUID: no MSR");
3225 * check max extended function levels of CPUID.
3226 * This is needed to check for invariant TSC.
3227 * This check is valid for both Intel and AMD.
3229 ebx = ecx = edx = 0;
3230 __cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
3232 if (max_extended_level >= 0x80000007) {
3235 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
3236 * this check is valid for both Intel and AMD
3238 __cpuid(0x80000007, eax, ebx, ecx, edx);
3239 has_invariant_tsc = edx & (1 << 8);
3243 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
3244 * this check is valid for both Intel and AMD
3247 __cpuid(0x6, eax, ebx, ecx, edx);
3248 has_aperf = ecx & (1 << 0);
3249 do_dts = eax & (1 << 0);
3250 do_ptm = eax & (1 << 6);
3251 has_hwp = eax & (1 << 7);
3252 has_hwp_notify = eax & (1 << 8);
3253 has_hwp_activity_window = eax & (1 << 9);
3254 has_hwp_epp = eax & (1 << 10);
3255 has_hwp_pkg = eax & (1 << 11);
3256 has_epb = ecx & (1 << 3);
3259 fprintf(outf, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sHWP, "
3260 "%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
3261 has_aperf ? "" : "No-",
3262 do_dts ? "" : "No-",
3263 do_ptm ? "" : "No-",
3264 has_hwp ? "" : "No-",
3265 has_hwp_notify ? "" : "No-",
3266 has_hwp_activity_window ? "" : "No-",
3267 has_hwp_epp ? "" : "No-",
3268 has_hwp_pkg ? "" : "No-",
3269 has_epb ? "" : "No-");
3272 decode_misc_enable_msr();
3274 if (max_level >= 0x7 && debug) {
3279 __cpuid_count(0x7, 0, eax, ebx, ecx, edx);
3281 has_sgx = ebx & (1 << 2);
3282 fprintf(outf, "CPUID(7): %sSGX\n", has_sgx ? "" : "No-");
3285 decode_feature_control_msr();
3288 if (max_level >= 0x15) {
3289 unsigned int eax_crystal;
3290 unsigned int ebx_tsc;
3293 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
3295 eax_crystal = ebx_tsc = crystal_hz = edx = 0;
3296 __cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
3300 if (debug && (ebx != 0))
3301 fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
3302 eax_crystal, ebx_tsc, crystal_hz);
3304 if (crystal_hz == 0)
3306 case 0x4E: /* SKL */
3307 case 0x5E: /* SKL */
3308 case 0x8E: /* KBL */
3309 case 0x9E: /* KBL */
3310 crystal_hz = 24000000; /* 24.0 MHz */
3312 case 0x55: /* SKX */
3313 crystal_hz = 25000000; /* 25.0 MHz */
3315 case 0x5C: /* BXT */
3316 crystal_hz = 19200000; /* 19.2 MHz */
3323 tsc_hz = (unsigned long long) crystal_hz * ebx_tsc / eax_crystal;
3325 fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
3326 tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
3330 if (max_level >= 0x16) {
3331 unsigned int base_mhz, max_mhz, bus_mhz, edx;
3334 * CPUID 16H Base MHz, Max MHz, Bus MHz
3336 base_mhz = max_mhz = bus_mhz = edx = 0;
3338 __cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
3340 fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
3341 base_mhz, max_mhz, bus_mhz);
3345 aperf_mperf_multiplier = get_aperf_mperf_multiplier(family, model);
3347 do_nhm_platform_info = do_nhm_cstates = do_smi = probe_nhm_msrs(family, model);
3348 do_snb_cstates = has_snb_msrs(family, model);
3349 do_irtl_snb = has_snb_msrs(family, model);
3350 do_pc2 = do_snb_cstates && (pkg_cstate_limit >= PCL__2);
3351 do_pc3 = (pkg_cstate_limit >= PCL__3);
3352 do_pc6 = (pkg_cstate_limit >= PCL__6);
3353 do_pc7 = do_snb_cstates && (pkg_cstate_limit >= PCL__7);
3354 do_c8_c9_c10 = has_hsw_msrs(family, model);
3355 do_irtl_hsw = has_hsw_msrs(family, model);
3356 do_skl_residency = has_skl_msrs(family, model);
3357 do_slm_cstates = is_slm(family, model);
3358 do_knl_cstates = is_knl(family, model);
3361 decode_misc_pwr_mgmt_msr();
3363 rapl_probe(family, model);
3364 perf_limit_reasons_probe(family, model);
3367 dump_cstate_pstate_config_info(family, model);
3369 if (has_skl_msrs(family, model))
3370 calculate_tsc_tweak();
3372 do_gfx_rc6_ms = !access("/sys/class/drm/card0/power/rc6_residency_ms", R_OK);
3374 do_gfx_mhz = !access("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", R_OK);
3382 "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
3384 "Turbostat forks the specified COMMAND and prints statistics\n"
3385 "when COMMAND completes.\n"
3386 "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
3387 "to print statistics, until interrupted.\n"
3388 "--debug run in \"debug\" mode\n"
3389 "--interval sec Override default 5-second measurement interval\n"
3390 "--help print this help message\n"
3391 "--counter msr print 32-bit counter at address \"msr\"\n"
3392 "--Counter msr print 64-bit Counter at address \"msr\"\n"
3393 "--out file create or truncate \"file\" for all output\n"
3394 "--msr msr print 32-bit value at address \"msr\"\n"
3395 "--MSR msr print 64-bit Value at address \"msr\"\n"
3396 "--version print version information\n"
3398 "For more help, run \"man turbostat\"\n");
3403 * in /dev/cpu/ return success for names that are numbers
3404 * ie. filter out ".", "..", "microcode".
3406 int dir_filter(const struct dirent *dirp)
3408 if (isdigit(dirp->d_name[0]))
3414 int open_dev_cpu_msr(int dummy1)
3419 void topology_probe()
3422 int max_core_id = 0;
3423 int max_package_id = 0;
3424 int max_siblings = 0;
3425 struct cpu_topology {
3427 int physical_package_id;
3430 /* Initialize num_cpus, max_cpu_num */
3432 topo.max_cpu_num = 0;
3433 for_all_proc_cpus(count_cpus);
3434 if (!summary_only && topo.num_cpus > 1)
3438 fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
3440 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
3442 err(1, "calloc cpus");
3445 * Allocate and initialize cpu_present_set
3447 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
3448 if (cpu_present_set == NULL)
3449 err(3, "CPU_ALLOC");
3450 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
3451 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
3452 for_all_proc_cpus(mark_cpu_present);
3455 * Allocate and initialize cpu_affinity_set
3457 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
3458 if (cpu_affinity_set == NULL)
3459 err(3, "CPU_ALLOC");
3460 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
3461 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
3466 * find max_core_id, max_package_id
3468 for (i = 0; i <= topo.max_cpu_num; ++i) {
3471 if (cpu_is_not_present(i)) {
3473 fprintf(outf, "cpu%d NOT PRESENT\n", i);
3476 cpus[i].core_id = get_core_id(i);
3477 if (cpus[i].core_id > max_core_id)
3478 max_core_id = cpus[i].core_id;
3480 cpus[i].physical_package_id = get_physical_package_id(i);
3481 if (cpus[i].physical_package_id > max_package_id)
3482 max_package_id = cpus[i].physical_package_id;
3484 siblings = get_num_ht_siblings(i);
3485 if (siblings > max_siblings)
3486 max_siblings = siblings;
3488 fprintf(outf, "cpu %d pkg %d core %d\n",
3489 i, cpus[i].physical_package_id, cpus[i].core_id);
3491 topo.num_cores_per_pkg = max_core_id + 1;
3493 fprintf(outf, "max_core_id %d, sizing for %d cores per package\n",
3494 max_core_id, topo.num_cores_per_pkg);
3495 if (debug && !summary_only && topo.num_cores_per_pkg > 1)
3498 topo.num_packages = max_package_id + 1;
3500 fprintf(outf, "max_package_id %d, sizing for %d packages\n",
3501 max_package_id, topo.num_packages);
3502 if (debug && !summary_only && topo.num_packages > 1)
3505 topo.num_threads_per_core = max_siblings;
3507 fprintf(outf, "max_siblings %d\n", max_siblings);
3513 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
3517 *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
3518 topo.num_packages, sizeof(struct thread_data));
3522 for (i = 0; i < topo.num_threads_per_core *
3523 topo.num_cores_per_pkg * topo.num_packages; i++)
3524 (*t)[i].cpu_id = -1;
3526 *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
3527 sizeof(struct core_data));
3531 for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
3532 (*c)[i].core_id = -1;
3534 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
3538 for (i = 0; i < topo.num_packages; i++)
3539 (*p)[i].package_id = i;
3543 err(1, "calloc counters");
3548 * set cpu_id, core_num, pkg_num
3549 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
3551 * increment topo.num_cores when 1st core in pkg seen
3553 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
3554 struct pkg_data *pkg_base, int thread_num, int core_num,
3555 int pkg_num, int cpu_id)
3557 struct thread_data *t;
3558 struct core_data *c;
3561 t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
3562 c = GET_CORE(core_base, core_num, pkg_num);
3563 p = GET_PKG(pkg_base, pkg_num);
3566 if (thread_num == 0) {
3567 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
3568 if (cpu_is_first_core_in_package(cpu_id))
3569 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
3572 c->core_id = core_num;
3573 p->package_id = pkg_num;
3577 int initialize_counters(int cpu_id)
3579 int my_thread_id, my_core_id, my_package_id;
3581 my_package_id = get_physical_package_id(cpu_id);
3582 my_core_id = get_core_id(cpu_id);
3583 my_thread_id = get_cpu_position_in_core(cpu_id);
3587 init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
3588 init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
3592 void allocate_output_buffer()
3594 output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
3595 outp = output_buffer;
3597 err(-1, "calloc output buffer");
3599 void allocate_fd_percpu(void)
3601 fd_percpu = calloc(topo.max_cpu_num, sizeof(int));
3602 if (fd_percpu == NULL)
3603 err(-1, "calloc fd_percpu");
3605 void allocate_irq_buffers(void)
3607 irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
3608 if (irq_column_2_cpu == NULL)
3609 err(-1, "calloc %d", topo.num_cpus);
3611 irqs_per_cpu = calloc(topo.max_cpu_num, sizeof(int));
3612 if (irqs_per_cpu == NULL)
3613 err(-1, "calloc %d", topo.max_cpu_num);
3615 void setup_all_buffers(void)
3618 allocate_irq_buffers();
3619 allocate_fd_percpu();
3620 allocate_counters(&thread_even, &core_even, &package_even);
3621 allocate_counters(&thread_odd, &core_odd, &package_odd);
3622 allocate_output_buffer();
3623 for_all_proc_cpus(initialize_counters);
3626 void set_base_cpu(void)
3628 base_cpu = sched_getcpu();
3630 err(-ENODEV, "No valid cpus found");
3633 fprintf(outf, "base_cpu = %d\n", base_cpu);
3636 void turbostat_init()
3638 setup_all_buffers();
3641 check_permissions();
3646 for_all_cpus(print_hwp, ODD_COUNTERS);
3649 for_all_cpus(print_epb, ODD_COUNTERS);
3652 for_all_cpus(print_perf_limit, ODD_COUNTERS);
3655 for_all_cpus(print_rapl, ODD_COUNTERS);
3657 for_all_cpus(set_temperature_target, ODD_COUNTERS);
3660 for_all_cpus(print_thermal, ODD_COUNTERS);
3662 if (debug && do_irtl_snb)
3666 int fork_it(char **argv)
3671 status = for_all_cpus(get_counters, EVEN_COUNTERS);
3674 /* clear affinity side-effect of get_counters() */
3675 sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
3676 gettimeofday(&tv_even, (struct timezone *)NULL);
3681 execvp(argv[0], argv);
3685 if (child_pid == -1)
3688 signal(SIGINT, SIG_IGN);
3689 signal(SIGQUIT, SIG_IGN);
3690 if (waitpid(child_pid, &status, 0) == -1)
3691 err(status, "waitpid");
3694 * n.b. fork_it() does not check for errors from for_all_cpus()
3695 * because re-starting is problematic when forking
3697 for_all_cpus(get_counters, ODD_COUNTERS);
3698 gettimeofday(&tv_odd, (struct timezone *)NULL);
3699 timersub(&tv_odd, &tv_even, &tv_delta);
3700 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
3701 compute_average(EVEN_COUNTERS);
3702 format_all_counters(EVEN_COUNTERS);
3704 fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
3706 flush_output_stderr();
3711 int get_and_dump_counters(void)
3715 status = for_all_cpus(get_counters, ODD_COUNTERS);
3719 status = for_all_cpus(dump_counters, ODD_COUNTERS);
3723 flush_output_stdout();
3728 void print_version() {
3729 fprintf(outf, "turbostat version 4.12 5 Apr 2016"
3730 " - Len Brown <lenb@kernel.org>\n");
3733 void cmdline(int argc, char **argv)
3736 int option_index = 0;
3737 static struct option long_options[] = {
3738 {"Counter", required_argument, 0, 'C'},
3739 {"counter", required_argument, 0, 'c'},
3740 {"Dump", no_argument, 0, 'D'},
3741 {"debug", no_argument, 0, 'd'},
3742 {"interval", required_argument, 0, 'i'},
3743 {"help", no_argument, 0, 'h'},
3744 {"Joules", no_argument, 0, 'J'},
3745 {"MSR", required_argument, 0, 'M'},
3746 {"msr", required_argument, 0, 'm'},
3747 {"out", required_argument, 0, 'o'},
3748 {"Package", no_argument, 0, 'p'},
3749 {"processor", no_argument, 0, 'p'},
3750 {"Summary", no_argument, 0, 'S'},
3751 {"TCC", required_argument, 0, 'T'},
3752 {"version", no_argument, 0, 'v' },
3758 while ((opt = getopt_long_only(argc, argv, "+C:c:Ddhi:JM:m:o:PpST:v",
3759 long_options, &option_index)) != -1) {
3762 sscanf(optarg, "%x", &extra_delta_offset64);
3765 sscanf(optarg, "%x", &extra_delta_offset32);
3779 double interval = strtod(optarg, NULL);
3781 if (interval < 0.001) {
3782 fprintf(outf, "interval %f seconds is too small\n",
3787 interval_ts.tv_sec = interval;
3788 interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
3795 sscanf(optarg, "%x", &extra_msr_offset64);
3798 sscanf(optarg, "%x", &extra_msr_offset32);
3801 outf = fopen_or_die(optarg, "w");
3813 tcc_activation_temp_override = atoi(optarg);
3823 int main(int argc, char **argv)
3827 cmdline(argc, argv);
3834 /* dump counters and exit */
3836 return get_and_dump_counters();
3839 * if any params left, it must be a command to fork
3842 return fork_it(argv + optind);