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";
47 struct timespec interval_ts
= {5, 0};
49 unsigned int rapl_joules
;
50 unsigned int summary_only
;
51 unsigned int dump_only
;
54 unsigned int do_nhm_cstates
;
55 unsigned int do_snb_cstates
;
56 unsigned int do_knl_cstates
;
61 unsigned int do_c8_c9_c10
;
62 unsigned int do_skl_residency
;
63 unsigned int do_slm_cstates
;
64 unsigned int use_c1_residency_msr
;
65 unsigned int has_aperf
;
67 unsigned int units
= 1000000; /* MHz etc */
68 unsigned int genuine_intel
;
69 unsigned int has_invariant_tsc
;
70 unsigned int do_nhm_platform_info
;
71 unsigned int extra_msr_offset32
;
72 unsigned int extra_msr_offset64
;
73 unsigned int extra_delta_offset32
;
74 unsigned int extra_delta_offset64
;
75 unsigned int aperf_mperf_multiplier
= 1;
79 unsigned int has_base_hz
;
80 double tsc_tweak
= 1.0;
81 unsigned int show_pkg
;
82 unsigned int show_core
;
83 unsigned int show_cpu
;
84 unsigned int show_pkg_only
;
85 unsigned int show_core_only
;
86 char *output_buffer
, *outp
;
90 unsigned int tcc_activation_temp
;
91 unsigned int tcc_activation_temp_override
;
92 double rapl_power_units
, rapl_time_units
;
93 double rapl_dram_energy_units
, rapl_energy_units
;
94 double rapl_joule_counter_range
;
95 unsigned int do_core_perf_limit_reasons
;
96 unsigned int do_gfx_perf_limit_reasons
;
97 unsigned int do_ring_perf_limit_reasons
;
98 unsigned int crystal_hz
;
99 unsigned long long tsc_hz
;
101 double discover_bclk(unsigned int family
, unsigned int model
);
102 unsigned int has_hwp
; /* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
103 /* IA32_HWP_REQUEST, IA32_HWP_STATUS */
104 unsigned int has_hwp_notify
; /* IA32_HWP_INTERRUPT */
105 unsigned int has_hwp_activity_window
; /* IA32_HWP_REQUEST[bits 41:32] */
106 unsigned int has_hwp_epp
; /* IA32_HWP_REQUEST[bits 31:24] */
107 unsigned int has_hwp_pkg
; /* IA32_HWP_REQUEST_PKG */
109 #define RAPL_PKG (1 << 0)
110 /* 0x610 MSR_PKG_POWER_LIMIT */
111 /* 0x611 MSR_PKG_ENERGY_STATUS */
112 #define RAPL_PKG_PERF_STATUS (1 << 1)
113 /* 0x613 MSR_PKG_PERF_STATUS */
114 #define RAPL_PKG_POWER_INFO (1 << 2)
115 /* 0x614 MSR_PKG_POWER_INFO */
117 #define RAPL_DRAM (1 << 3)
118 /* 0x618 MSR_DRAM_POWER_LIMIT */
119 /* 0x619 MSR_DRAM_ENERGY_STATUS */
120 #define RAPL_DRAM_PERF_STATUS (1 << 4)
121 /* 0x61b MSR_DRAM_PERF_STATUS */
122 #define RAPL_DRAM_POWER_INFO (1 << 5)
123 /* 0x61c MSR_DRAM_POWER_INFO */
125 #define RAPL_CORES (1 << 6)
126 /* 0x638 MSR_PP0_POWER_LIMIT */
127 /* 0x639 MSR_PP0_ENERGY_STATUS */
128 #define RAPL_CORE_POLICY (1 << 7)
129 /* 0x63a MSR_PP0_POLICY */
131 #define RAPL_GFX (1 << 8)
132 /* 0x640 MSR_PP1_POWER_LIMIT */
133 /* 0x641 MSR_PP1_ENERGY_STATUS */
134 /* 0x642 MSR_PP1_POLICY */
135 #define TJMAX_DEFAULT 100
137 #define MAX(a, b) ((a) > (b) ? (a) : (b))
139 int aperf_mperf_unstable
;
143 cpu_set_t
*cpu_present_set
, *cpu_affinity_set
;
144 size_t cpu_present_setsize
, cpu_affinity_setsize
;
147 unsigned long long tsc
;
148 unsigned long long aperf
;
149 unsigned long long mperf
;
150 unsigned long long c1
;
151 unsigned long long extra_msr64
;
152 unsigned long long extra_delta64
;
153 unsigned long long extra_msr32
;
154 unsigned long long extra_delta32
;
155 unsigned int smi_count
;
158 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
159 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
160 } *thread_even
, *thread_odd
;
163 unsigned long long c3
;
164 unsigned long long c6
;
165 unsigned long long c7
;
166 unsigned int core_temp_c
;
167 unsigned int core_id
;
168 } *core_even
, *core_odd
;
171 unsigned long long pc2
;
172 unsigned long long pc3
;
173 unsigned long long pc6
;
174 unsigned long long pc7
;
175 unsigned long long pc8
;
176 unsigned long long pc9
;
177 unsigned long long pc10
;
178 unsigned long long pkg_wtd_core_c0
;
179 unsigned long long pkg_any_core_c0
;
180 unsigned long long pkg_any_gfxe_c0
;
181 unsigned long long pkg_both_core_gfxe_c0
;
182 unsigned int package_id
;
183 unsigned int energy_pkg
; /* MSR_PKG_ENERGY_STATUS */
184 unsigned int energy_dram
; /* MSR_DRAM_ENERGY_STATUS */
185 unsigned int energy_cores
; /* MSR_PP0_ENERGY_STATUS */
186 unsigned int energy_gfx
; /* MSR_PP1_ENERGY_STATUS */
187 unsigned int rapl_pkg_perf_status
; /* MSR_PKG_PERF_STATUS */
188 unsigned int rapl_dram_perf_status
; /* MSR_DRAM_PERF_STATUS */
189 unsigned int pkg_temp_c
;
191 } *package_even
, *package_odd
;
193 #define ODD_COUNTERS thread_odd, core_odd, package_odd
194 #define EVEN_COUNTERS thread_even, core_even, package_even
196 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
197 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
198 topo.num_threads_per_core + \
199 (core_no) * topo.num_threads_per_core + (thread_no))
200 #define GET_CORE(core_base, core_no, pkg_no) \
201 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
202 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
204 struct system_summary
{
205 struct thread_data threads
;
206 struct core_data cores
;
207 struct pkg_data packages
;
216 int num_cores_per_pkg
;
217 int num_threads_per_core
;
220 struct timeval tv_even
, tv_odd
, tv_delta
;
222 void setup_all_buffers(void);
224 int cpu_is_not_present(int cpu
)
226 return !CPU_ISSET_S(cpu
, cpu_present_setsize
, cpu_present_set
);
229 * run func(thread, core, package) in topology order
230 * skip non-present cpus
233 int for_all_cpus(int (func
)(struct thread_data
*, struct core_data
*, struct pkg_data
*),
234 struct thread_data
*thread_base
, struct core_data
*core_base
, struct pkg_data
*pkg_base
)
236 int retval
, pkg_no
, core_no
, thread_no
;
238 for (pkg_no
= 0; pkg_no
< topo
.num_packages
; ++pkg_no
) {
239 for (core_no
= 0; core_no
< topo
.num_cores_per_pkg
; ++core_no
) {
240 for (thread_no
= 0; thread_no
<
241 topo
.num_threads_per_core
; ++thread_no
) {
242 struct thread_data
*t
;
246 t
= GET_THREAD(thread_base
, thread_no
, core_no
, pkg_no
);
248 if (cpu_is_not_present(t
->cpu_id
))
251 c
= GET_CORE(core_base
, core_no
, pkg_no
);
252 p
= GET_PKG(pkg_base
, pkg_no
);
254 retval
= func(t
, c
, p
);
263 int cpu_migrate(int cpu
)
265 CPU_ZERO_S(cpu_affinity_setsize
, cpu_affinity_set
);
266 CPU_SET_S(cpu
, cpu_affinity_setsize
, cpu_affinity_set
);
267 if (sched_setaffinity(0, cpu_affinity_setsize
, cpu_affinity_set
) == -1)
273 int get_msr(int cpu
, off_t offset
, unsigned long long *msr
)
279 sprintf(pathname
, "/dev/cpu/%d/msr", cpu
);
280 fd
= open(pathname
, O_RDONLY
);
282 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname
);
284 retval
= pread(fd
, msr
, sizeof *msr
, offset
);
287 if (retval
!= sizeof *msr
)
288 err(-1, "%s offset 0x%llx read failed", pathname
, (unsigned long long)offset
);
294 * Example Format w/ field column widths:
296 * 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
297 * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
300 void print_header(void)
303 outp
+= sprintf(outp
, " Package");
305 outp
+= sprintf(outp
, " Core");
307 outp
+= sprintf(outp
, " CPU");
309 outp
+= sprintf(outp
, " Avg_MHz");
311 outp
+= sprintf(outp
, " %%Busy");
313 outp
+= sprintf(outp
, " Bzy_MHz");
314 outp
+= sprintf(outp
, " TSC_MHz");
316 if (extra_delta_offset32
)
317 outp
+= sprintf(outp
, " count 0x%03X", extra_delta_offset32
);
318 if (extra_delta_offset64
)
319 outp
+= sprintf(outp
, " COUNT 0x%03X", extra_delta_offset64
);
320 if (extra_msr_offset32
)
321 outp
+= sprintf(outp
, " MSR 0x%03X", extra_msr_offset32
);
322 if (extra_msr_offset64
)
323 outp
+= sprintf(outp
, " MSR 0x%03X", extra_msr_offset64
);
329 outp
+= sprintf(outp
, " SMI");
332 outp
+= sprintf(outp
, " CPU%%c1");
333 if (do_nhm_cstates
&& !do_slm_cstates
&& !do_knl_cstates
)
334 outp
+= sprintf(outp
, " CPU%%c3");
336 outp
+= sprintf(outp
, " CPU%%c6");
338 outp
+= sprintf(outp
, " CPU%%c7");
341 outp
+= sprintf(outp
, " CoreTmp");
343 outp
+= sprintf(outp
, " PkgTmp");
345 if (do_skl_residency
) {
346 outp
+= sprintf(outp
, " Totl%%C0");
347 outp
+= sprintf(outp
, " Any%%C0");
348 outp
+= sprintf(outp
, " GFX%%C0");
349 outp
+= sprintf(outp
, " CPUGFX%%");
353 outp
+= sprintf(outp
, " Pkg%%pc2");
355 outp
+= sprintf(outp
, " Pkg%%pc3");
357 outp
+= sprintf(outp
, " Pkg%%pc6");
359 outp
+= sprintf(outp
, " Pkg%%pc7");
361 outp
+= sprintf(outp
, " Pkg%%pc8");
362 outp
+= sprintf(outp
, " Pkg%%pc9");
363 outp
+= sprintf(outp
, " Pk%%pc10");
366 if (do_rapl
&& !rapl_joules
) {
367 if (do_rapl
& RAPL_PKG
)
368 outp
+= sprintf(outp
, " PkgWatt");
369 if (do_rapl
& RAPL_CORES
)
370 outp
+= sprintf(outp
, " CorWatt");
371 if (do_rapl
& RAPL_GFX
)
372 outp
+= sprintf(outp
, " GFXWatt");
373 if (do_rapl
& RAPL_DRAM
)
374 outp
+= sprintf(outp
, " RAMWatt");
375 if (do_rapl
& RAPL_PKG_PERF_STATUS
)
376 outp
+= sprintf(outp
, " PKG_%%");
377 if (do_rapl
& RAPL_DRAM_PERF_STATUS
)
378 outp
+= sprintf(outp
, " RAM_%%");
379 } else if (do_rapl
&& rapl_joules
) {
380 if (do_rapl
& RAPL_PKG
)
381 outp
+= sprintf(outp
, " Pkg_J");
382 if (do_rapl
& RAPL_CORES
)
383 outp
+= sprintf(outp
, " Cor_J");
384 if (do_rapl
& RAPL_GFX
)
385 outp
+= sprintf(outp
, " GFX_J");
386 if (do_rapl
& RAPL_DRAM
)
387 outp
+= sprintf(outp
, " RAM_J");
388 if (do_rapl
& RAPL_PKG_PERF_STATUS
)
389 outp
+= sprintf(outp
, " PKG_%%");
390 if (do_rapl
& RAPL_DRAM_PERF_STATUS
)
391 outp
+= sprintf(outp
, " RAM_%%");
392 outp
+= sprintf(outp
, " time");
396 outp
+= sprintf(outp
, "\n");
399 int dump_counters(struct thread_data
*t
, struct core_data
*c
,
402 outp
+= sprintf(outp
, "t %p, c %p, p %p\n", t
, c
, p
);
405 outp
+= sprintf(outp
, "CPU: %d flags 0x%x\n",
406 t
->cpu_id
, t
->flags
);
407 outp
+= sprintf(outp
, "TSC: %016llX\n", t
->tsc
);
408 outp
+= sprintf(outp
, "aperf: %016llX\n", t
->aperf
);
409 outp
+= sprintf(outp
, "mperf: %016llX\n", t
->mperf
);
410 outp
+= sprintf(outp
, "c1: %016llX\n", t
->c1
);
411 outp
+= sprintf(outp
, "msr0x%x: %08llX\n",
412 extra_delta_offset32
, t
->extra_delta32
);
413 outp
+= sprintf(outp
, "msr0x%x: %016llX\n",
414 extra_delta_offset64
, t
->extra_delta64
);
415 outp
+= sprintf(outp
, "msr0x%x: %08llX\n",
416 extra_msr_offset32
, t
->extra_msr32
);
417 outp
+= sprintf(outp
, "msr0x%x: %016llX\n",
418 extra_msr_offset64
, t
->extra_msr64
);
420 outp
+= sprintf(outp
, "SMI: %08X\n", t
->smi_count
);
424 outp
+= sprintf(outp
, "core: %d\n", c
->core_id
);
425 outp
+= sprintf(outp
, "c3: %016llX\n", c
->c3
);
426 outp
+= sprintf(outp
, "c6: %016llX\n", c
->c6
);
427 outp
+= sprintf(outp
, "c7: %016llX\n", c
->c7
);
428 outp
+= sprintf(outp
, "DTS: %dC\n", c
->core_temp_c
);
432 outp
+= sprintf(outp
, "package: %d\n", p
->package_id
);
434 outp
+= sprintf(outp
, "Weighted cores: %016llX\n", p
->pkg_wtd_core_c0
);
435 outp
+= sprintf(outp
, "Any cores: %016llX\n", p
->pkg_any_core_c0
);
436 outp
+= sprintf(outp
, "Any GFX: %016llX\n", p
->pkg_any_gfxe_c0
);
437 outp
+= sprintf(outp
, "CPU + GFX: %016llX\n", p
->pkg_both_core_gfxe_c0
);
439 outp
+= sprintf(outp
, "pc2: %016llX\n", p
->pc2
);
441 outp
+= sprintf(outp
, "pc3: %016llX\n", p
->pc3
);
443 outp
+= sprintf(outp
, "pc6: %016llX\n", p
->pc6
);
445 outp
+= sprintf(outp
, "pc7: %016llX\n", p
->pc7
);
446 outp
+= sprintf(outp
, "pc8: %016llX\n", p
->pc8
);
447 outp
+= sprintf(outp
, "pc9: %016llX\n", p
->pc9
);
448 outp
+= sprintf(outp
, "pc10: %016llX\n", p
->pc10
);
449 outp
+= sprintf(outp
, "Joules PKG: %0X\n", p
->energy_pkg
);
450 outp
+= sprintf(outp
, "Joules COR: %0X\n", p
->energy_cores
);
451 outp
+= sprintf(outp
, "Joules GFX: %0X\n", p
->energy_gfx
);
452 outp
+= sprintf(outp
, "Joules RAM: %0X\n", p
->energy_dram
);
453 outp
+= sprintf(outp
, "Throttle PKG: %0X\n",
454 p
->rapl_pkg_perf_status
);
455 outp
+= sprintf(outp
, "Throttle RAM: %0X\n",
456 p
->rapl_dram_perf_status
);
457 outp
+= sprintf(outp
, "PTM: %dC\n", p
->pkg_temp_c
);
460 outp
+= sprintf(outp
, "\n");
466 * column formatting convention & formats
468 int format_counters(struct thread_data
*t
, struct core_data
*c
,
471 double interval_float
;
474 /* if showing only 1st thread in core and this isn't one, bail out */
475 if (show_core_only
&& !(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
))
478 /* if showing only 1st thread in pkg and this isn't one, bail out */
479 if (show_pkg_only
&& !(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
482 interval_float
= tv_delta
.tv_sec
+ tv_delta
.tv_usec
/1000000.0;
484 /* topo columns, print blanks on 1st (average) line */
485 if (t
== &average
.threads
) {
487 outp
+= sprintf(outp
, " -");
489 outp
+= sprintf(outp
, " -");
491 outp
+= sprintf(outp
, " -");
495 outp
+= sprintf(outp
, "%8d", p
->package_id
);
497 outp
+= sprintf(outp
, " -");
501 outp
+= sprintf(outp
, "%8d", c
->core_id
);
503 outp
+= sprintf(outp
, " -");
506 outp
+= sprintf(outp
, "%8d", t
->cpu_id
);
511 outp
+= sprintf(outp
, "%8.0f",
512 1.0 / units
* t
->aperf
/ interval_float
);
517 outp
+= sprintf(outp
, "%8.2f", 100.0 * t
->mperf
/t
->tsc
/tsc_tweak
);
519 outp
+= sprintf(outp
, "********");
525 outp
+= sprintf(outp
, "%8.0f", base_hz
/ units
* t
->aperf
/ t
->mperf
);
527 outp
+= sprintf(outp
, "%8.0f",
528 1.0 * t
->tsc
/ units
* t
->aperf
/ t
->mperf
/ interval_float
);
532 outp
+= sprintf(outp
, "%8.0f", 1.0 * t
->tsc
/units
/interval_float
);
535 if (extra_delta_offset32
)
536 outp
+= sprintf(outp
, " %11llu", t
->extra_delta32
);
539 if (extra_delta_offset64
)
540 outp
+= sprintf(outp
, " %11llu", t
->extra_delta64
);
542 if (extra_msr_offset32
)
543 outp
+= sprintf(outp
, " 0x%08llx", t
->extra_msr32
);
546 if (extra_msr_offset64
)
547 outp
+= sprintf(outp
, " 0x%016llx", t
->extra_msr64
);
554 outp
+= sprintf(outp
, "%8d", t
->smi_count
);
556 if (do_nhm_cstates
) {
558 outp
+= sprintf(outp
, "%8.2f", 100.0 * t
->c1
/t
->tsc
);
560 outp
+= sprintf(outp
, "********");
563 /* print per-core data only for 1st thread in core */
564 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
))
567 if (do_nhm_cstates
&& !do_slm_cstates
&& !do_knl_cstates
)
568 outp
+= sprintf(outp
, "%8.2f", 100.0 * c
->c3
/t
->tsc
);
570 outp
+= sprintf(outp
, "%8.2f", 100.0 * c
->c6
/t
->tsc
);
572 outp
+= sprintf(outp
, "%8.2f", 100.0 * c
->c7
/t
->tsc
);
575 outp
+= sprintf(outp
, "%8d", c
->core_temp_c
);
577 /* print per-package data only for 1st core in package */
578 if (!(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
583 outp
+= sprintf(outp
, "%8d", p
->pkg_temp_c
);
585 /* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
586 if (do_skl_residency
) {
587 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pkg_wtd_core_c0
/t
->tsc
);
588 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pkg_any_core_c0
/t
->tsc
);
589 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pkg_any_gfxe_c0
/t
->tsc
);
590 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pkg_both_core_gfxe_c0
/t
->tsc
);
594 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pc2
/t
->tsc
);
596 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pc3
/t
->tsc
);
598 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pc6
/t
->tsc
);
600 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pc7
/t
->tsc
);
602 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pc8
/t
->tsc
);
603 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pc9
/t
->tsc
);
604 outp
+= sprintf(outp
, "%8.2f", 100.0 * p
->pc10
/t
->tsc
);
608 * If measurement interval exceeds minimum RAPL Joule Counter range,
609 * indicate that results are suspect by printing "**" in fraction place.
611 if (interval_float
< rapl_joule_counter_range
)
616 if (do_rapl
&& !rapl_joules
) {
617 if (do_rapl
& RAPL_PKG
)
618 outp
+= sprintf(outp
, fmt8
, p
->energy_pkg
* rapl_energy_units
/ interval_float
);
619 if (do_rapl
& RAPL_CORES
)
620 outp
+= sprintf(outp
, fmt8
, p
->energy_cores
* rapl_energy_units
/ interval_float
);
621 if (do_rapl
& RAPL_GFX
)
622 outp
+= sprintf(outp
, fmt8
, p
->energy_gfx
* rapl_energy_units
/ interval_float
);
623 if (do_rapl
& RAPL_DRAM
)
624 outp
+= sprintf(outp
, fmt8
, p
->energy_dram
* rapl_dram_energy_units
/ interval_float
);
625 if (do_rapl
& RAPL_PKG_PERF_STATUS
)
626 outp
+= sprintf(outp
, fmt8
, 100.0 * p
->rapl_pkg_perf_status
* rapl_time_units
/ interval_float
);
627 if (do_rapl
& RAPL_DRAM_PERF_STATUS
)
628 outp
+= sprintf(outp
, fmt8
, 100.0 * p
->rapl_dram_perf_status
* rapl_time_units
/ interval_float
);
629 } else if (do_rapl
&& rapl_joules
) {
630 if (do_rapl
& RAPL_PKG
)
631 outp
+= sprintf(outp
, fmt8
,
632 p
->energy_pkg
* rapl_energy_units
);
633 if (do_rapl
& RAPL_CORES
)
634 outp
+= sprintf(outp
, fmt8
,
635 p
->energy_cores
* rapl_energy_units
);
636 if (do_rapl
& RAPL_GFX
)
637 outp
+= sprintf(outp
, fmt8
,
638 p
->energy_gfx
* rapl_energy_units
);
639 if (do_rapl
& RAPL_DRAM
)
640 outp
+= sprintf(outp
, fmt8
,
641 p
->energy_dram
* rapl_dram_energy_units
);
642 if (do_rapl
& RAPL_PKG_PERF_STATUS
)
643 outp
+= sprintf(outp
, fmt8
, 100.0 * p
->rapl_pkg_perf_status
* rapl_time_units
/ interval_float
);
644 if (do_rapl
& RAPL_DRAM_PERF_STATUS
)
645 outp
+= sprintf(outp
, fmt8
, 100.0 * p
->rapl_dram_perf_status
* rapl_time_units
/ interval_float
);
647 outp
+= sprintf(outp
, fmt8
, interval_float
);
650 outp
+= sprintf(outp
, "\n");
657 fputs(output_buffer
, stdout
);
659 outp
= output_buffer
;
663 fputs(output_buffer
, stderr
);
664 outp
= output_buffer
;
666 void format_all_counters(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
670 if (!printed
|| !summary_only
)
673 if (topo
.num_cpus
> 1)
674 format_counters(&average
.threads
, &average
.cores
,
682 for_all_cpus(format_counters
, t
, c
, p
);
685 #define DELTA_WRAP32(new, old) \
689 old = 0x100000000 + new - old; \
693 delta_package(struct pkg_data
*new, struct pkg_data
*old
)
696 if (do_skl_residency
) {
697 old
->pkg_wtd_core_c0
= new->pkg_wtd_core_c0
- old
->pkg_wtd_core_c0
;
698 old
->pkg_any_core_c0
= new->pkg_any_core_c0
- old
->pkg_any_core_c0
;
699 old
->pkg_any_gfxe_c0
= new->pkg_any_gfxe_c0
- old
->pkg_any_gfxe_c0
;
700 old
->pkg_both_core_gfxe_c0
= new->pkg_both_core_gfxe_c0
- old
->pkg_both_core_gfxe_c0
;
702 old
->pc2
= new->pc2
- old
->pc2
;
704 old
->pc3
= new->pc3
- old
->pc3
;
706 old
->pc6
= new->pc6
- old
->pc6
;
708 old
->pc7
= new->pc7
- old
->pc7
;
709 old
->pc8
= new->pc8
- old
->pc8
;
710 old
->pc9
= new->pc9
- old
->pc9
;
711 old
->pc10
= new->pc10
- old
->pc10
;
712 old
->pkg_temp_c
= new->pkg_temp_c
;
714 DELTA_WRAP32(new->energy_pkg
, old
->energy_pkg
);
715 DELTA_WRAP32(new->energy_cores
, old
->energy_cores
);
716 DELTA_WRAP32(new->energy_gfx
, old
->energy_gfx
);
717 DELTA_WRAP32(new->energy_dram
, old
->energy_dram
);
718 DELTA_WRAP32(new->rapl_pkg_perf_status
, old
->rapl_pkg_perf_status
);
719 DELTA_WRAP32(new->rapl_dram_perf_status
, old
->rapl_dram_perf_status
);
723 delta_core(struct core_data
*new, struct core_data
*old
)
725 old
->c3
= new->c3
- old
->c3
;
726 old
->c6
= new->c6
- old
->c6
;
727 old
->c7
= new->c7
- old
->c7
;
728 old
->core_temp_c
= new->core_temp_c
;
735 delta_thread(struct thread_data
*new, struct thread_data
*old
,
736 struct core_data
*core_delta
)
738 old
->tsc
= new->tsc
- old
->tsc
;
740 /* check for TSC < 1 Mcycles over interval */
741 if (old
->tsc
< (1000 * 1000))
742 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
743 "You can disable all c-states by booting with \"idle=poll\"\n"
744 "or just the deep ones with \"processor.max_cstate=1\"");
746 old
->c1
= new->c1
- old
->c1
;
749 if ((new->aperf
> old
->aperf
) && (new->mperf
> old
->mperf
)) {
750 old
->aperf
= new->aperf
- old
->aperf
;
751 old
->mperf
= new->mperf
- old
->mperf
;
754 if (!aperf_mperf_unstable
) {
755 fprintf(stderr
, "%s: APERF or MPERF went backwards *\n", progname
);
756 fprintf(stderr
, "* Frequency results do not cover entire interval *\n");
757 fprintf(stderr
, "* fix this by running Linux-2.6.30 or later *\n");
759 aperf_mperf_unstable
= 1;
762 * mperf delta is likely a huge "positive" number
763 * can not use it for calculating c0 time
771 if (use_c1_residency_msr
) {
773 * Some models have a dedicated C1 residency MSR,
774 * which should be more accurate than the derivation below.
778 * As counter collection is not atomic,
779 * it is possible for mperf's non-halted cycles + idle states
780 * to exceed TSC's all cycles: show c1 = 0% in that case.
782 if ((old
->mperf
+ core_delta
->c3
+ core_delta
->c6
+ core_delta
->c7
) > old
->tsc
)
785 /* normal case, derive c1 */
786 old
->c1
= old
->tsc
- old
->mperf
- core_delta
->c3
787 - core_delta
->c6
- core_delta
->c7
;
791 if (old
->mperf
== 0) {
792 if (debug
> 1) fprintf(stderr
, "cpu%d MPERF 0!\n", old
->cpu_id
);
793 old
->mperf
= 1; /* divide by 0 protection */
796 old
->extra_delta32
= new->extra_delta32
- old
->extra_delta32
;
797 old
->extra_delta32
&= 0xFFFFFFFF;
799 old
->extra_delta64
= new->extra_delta64
- old
->extra_delta64
;
802 * Extra MSR is just a snapshot, simply copy latest w/o subtracting
804 old
->extra_msr32
= new->extra_msr32
;
805 old
->extra_msr64
= new->extra_msr64
;
808 old
->smi_count
= new->smi_count
- old
->smi_count
;
811 int delta_cpu(struct thread_data
*t
, struct core_data
*c
,
812 struct pkg_data
*p
, struct thread_data
*t2
,
813 struct core_data
*c2
, struct pkg_data
*p2
)
815 /* calculate core delta only for 1st thread in core */
816 if (t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
)
819 /* always calculate thread delta */
820 delta_thread(t
, t2
, c2
); /* c2 is core delta */
822 /* calculate package delta only for 1st core in package */
823 if (t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
)
824 delta_package(p
, p2
);
829 void clear_counters(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
837 t
->extra_delta32
= 0;
838 t
->extra_delta64
= 0;
840 /* tells format_counters to dump all fields from this set */
841 t
->flags
= CPU_IS_FIRST_THREAD_IN_CORE
| CPU_IS_FIRST_CORE_IN_PACKAGE
;
848 p
->pkg_wtd_core_c0
= 0;
849 p
->pkg_any_core_c0
= 0;
850 p
->pkg_any_gfxe_c0
= 0;
851 p
->pkg_both_core_gfxe_c0
= 0;
868 p
->rapl_pkg_perf_status
= 0;
869 p
->rapl_dram_perf_status
= 0;
872 int sum_counters(struct thread_data
*t
, struct core_data
*c
,
875 average
.threads
.tsc
+= t
->tsc
;
876 average
.threads
.aperf
+= t
->aperf
;
877 average
.threads
.mperf
+= t
->mperf
;
878 average
.threads
.c1
+= t
->c1
;
880 average
.threads
.extra_delta32
+= t
->extra_delta32
;
881 average
.threads
.extra_delta64
+= t
->extra_delta64
;
883 /* sum per-core values only for 1st thread in core */
884 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
))
887 average
.cores
.c3
+= c
->c3
;
888 average
.cores
.c6
+= c
->c6
;
889 average
.cores
.c7
+= c
->c7
;
891 average
.cores
.core_temp_c
= MAX(average
.cores
.core_temp_c
, c
->core_temp_c
);
893 /* sum per-pkg values only for 1st core in pkg */
894 if (!(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
897 if (do_skl_residency
) {
898 average
.packages
.pkg_wtd_core_c0
+= p
->pkg_wtd_core_c0
;
899 average
.packages
.pkg_any_core_c0
+= p
->pkg_any_core_c0
;
900 average
.packages
.pkg_any_gfxe_c0
+= p
->pkg_any_gfxe_c0
;
901 average
.packages
.pkg_both_core_gfxe_c0
+= p
->pkg_both_core_gfxe_c0
;
904 average
.packages
.pc2
+= p
->pc2
;
906 average
.packages
.pc3
+= p
->pc3
;
908 average
.packages
.pc6
+= p
->pc6
;
910 average
.packages
.pc7
+= p
->pc7
;
911 average
.packages
.pc8
+= p
->pc8
;
912 average
.packages
.pc9
+= p
->pc9
;
913 average
.packages
.pc10
+= p
->pc10
;
915 average
.packages
.energy_pkg
+= p
->energy_pkg
;
916 average
.packages
.energy_dram
+= p
->energy_dram
;
917 average
.packages
.energy_cores
+= p
->energy_cores
;
918 average
.packages
.energy_gfx
+= p
->energy_gfx
;
920 average
.packages
.pkg_temp_c
= MAX(average
.packages
.pkg_temp_c
, p
->pkg_temp_c
);
922 average
.packages
.rapl_pkg_perf_status
+= p
->rapl_pkg_perf_status
;
923 average
.packages
.rapl_dram_perf_status
+= p
->rapl_dram_perf_status
;
927 * sum the counters for all cpus in the system
928 * compute the weighted average
930 void compute_average(struct thread_data
*t
, struct core_data
*c
,
933 clear_counters(&average
.threads
, &average
.cores
, &average
.packages
);
935 for_all_cpus(sum_counters
, t
, c
, p
);
937 average
.threads
.tsc
/= topo
.num_cpus
;
938 average
.threads
.aperf
/= topo
.num_cpus
;
939 average
.threads
.mperf
/= topo
.num_cpus
;
940 average
.threads
.c1
/= topo
.num_cpus
;
942 average
.threads
.extra_delta32
/= topo
.num_cpus
;
943 average
.threads
.extra_delta32
&= 0xFFFFFFFF;
945 average
.threads
.extra_delta64
/= topo
.num_cpus
;
947 average
.cores
.c3
/= topo
.num_cores
;
948 average
.cores
.c6
/= topo
.num_cores
;
949 average
.cores
.c7
/= topo
.num_cores
;
951 if (do_skl_residency
) {
952 average
.packages
.pkg_wtd_core_c0
/= topo
.num_packages
;
953 average
.packages
.pkg_any_core_c0
/= topo
.num_packages
;
954 average
.packages
.pkg_any_gfxe_c0
/= topo
.num_packages
;
955 average
.packages
.pkg_both_core_gfxe_c0
/= topo
.num_packages
;
958 average
.packages
.pc2
/= topo
.num_packages
;
960 average
.packages
.pc3
/= topo
.num_packages
;
962 average
.packages
.pc6
/= topo
.num_packages
;
964 average
.packages
.pc7
/= topo
.num_packages
;
966 average
.packages
.pc8
/= topo
.num_packages
;
967 average
.packages
.pc9
/= topo
.num_packages
;
968 average
.packages
.pc10
/= topo
.num_packages
;
971 static unsigned long long rdtsc(void)
973 unsigned int low
, high
;
975 asm volatile("rdtsc" : "=a" (low
), "=d" (high
));
977 return low
| ((unsigned long long)high
) << 32;
984 * acquire and record local counters for that cpu
986 int get_counters(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
989 unsigned long long msr
;
991 if (cpu_migrate(cpu
)) {
992 fprintf(stderr
, "Could not migrate to CPU %d\n", cpu
);
996 t
->tsc
= rdtsc(); /* we are running on local CPU of interest */
999 if (get_msr(cpu
, MSR_IA32_APERF
, &t
->aperf
))
1001 if (get_msr(cpu
, MSR_IA32_MPERF
, &t
->mperf
))
1003 t
->aperf
= t
->aperf
* aperf_mperf_multiplier
;
1004 t
->mperf
= t
->mperf
* aperf_mperf_multiplier
;
1008 if (get_msr(cpu
, MSR_SMI_COUNT
, &msr
))
1010 t
->smi_count
= msr
& 0xFFFFFFFF;
1012 if (extra_delta_offset32
) {
1013 if (get_msr(cpu
, extra_delta_offset32
, &msr
))
1015 t
->extra_delta32
= msr
& 0xFFFFFFFF;
1018 if (extra_delta_offset64
)
1019 if (get_msr(cpu
, extra_delta_offset64
, &t
->extra_delta64
))
1022 if (extra_msr_offset32
) {
1023 if (get_msr(cpu
, extra_msr_offset32
, &msr
))
1025 t
->extra_msr32
= msr
& 0xFFFFFFFF;
1028 if (extra_msr_offset64
)
1029 if (get_msr(cpu
, extra_msr_offset64
, &t
->extra_msr64
))
1032 if (use_c1_residency_msr
) {
1033 if (get_msr(cpu
, MSR_CORE_C1_RES
, &t
->c1
))
1037 /* collect core counters only for 1st thread in core */
1038 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
))
1041 if (do_nhm_cstates
&& !do_slm_cstates
&& !do_knl_cstates
) {
1042 if (get_msr(cpu
, MSR_CORE_C3_RESIDENCY
, &c
->c3
))
1046 if (do_nhm_cstates
&& !do_knl_cstates
) {
1047 if (get_msr(cpu
, MSR_CORE_C6_RESIDENCY
, &c
->c6
))
1049 } else if (do_knl_cstates
) {
1050 if (get_msr(cpu
, MSR_KNL_CORE_C6_RESIDENCY
, &c
->c6
))
1055 if (get_msr(cpu
, MSR_CORE_C7_RESIDENCY
, &c
->c7
))
1059 if (get_msr(cpu
, MSR_IA32_THERM_STATUS
, &msr
))
1061 c
->core_temp_c
= tcc_activation_temp
- ((msr
>> 16) & 0x7F);
1065 /* collect package counters only for 1st core in package */
1066 if (!(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
1069 if (do_skl_residency
) {
1070 if (get_msr(cpu
, MSR_PKG_WEIGHTED_CORE_C0_RES
, &p
->pkg_wtd_core_c0
))
1072 if (get_msr(cpu
, MSR_PKG_ANY_CORE_C0_RES
, &p
->pkg_any_core_c0
))
1074 if (get_msr(cpu
, MSR_PKG_ANY_GFXE_C0_RES
, &p
->pkg_any_gfxe_c0
))
1076 if (get_msr(cpu
, MSR_PKG_BOTH_CORE_GFXE_C0_RES
, &p
->pkg_both_core_gfxe_c0
))
1080 if (get_msr(cpu
, MSR_PKG_C3_RESIDENCY
, &p
->pc3
))
1083 if (get_msr(cpu
, MSR_PKG_C6_RESIDENCY
, &p
->pc6
))
1086 if (get_msr(cpu
, MSR_PKG_C2_RESIDENCY
, &p
->pc2
))
1089 if (get_msr(cpu
, MSR_PKG_C7_RESIDENCY
, &p
->pc7
))
1092 if (get_msr(cpu
, MSR_PKG_C8_RESIDENCY
, &p
->pc8
))
1094 if (get_msr(cpu
, MSR_PKG_C9_RESIDENCY
, &p
->pc9
))
1096 if (get_msr(cpu
, MSR_PKG_C10_RESIDENCY
, &p
->pc10
))
1099 if (do_rapl
& RAPL_PKG
) {
1100 if (get_msr(cpu
, MSR_PKG_ENERGY_STATUS
, &msr
))
1102 p
->energy_pkg
= msr
& 0xFFFFFFFF;
1104 if (do_rapl
& RAPL_CORES
) {
1105 if (get_msr(cpu
, MSR_PP0_ENERGY_STATUS
, &msr
))
1107 p
->energy_cores
= msr
& 0xFFFFFFFF;
1109 if (do_rapl
& RAPL_DRAM
) {
1110 if (get_msr(cpu
, MSR_DRAM_ENERGY_STATUS
, &msr
))
1112 p
->energy_dram
= msr
& 0xFFFFFFFF;
1114 if (do_rapl
& RAPL_GFX
) {
1115 if (get_msr(cpu
, MSR_PP1_ENERGY_STATUS
, &msr
))
1117 p
->energy_gfx
= msr
& 0xFFFFFFFF;
1119 if (do_rapl
& RAPL_PKG_PERF_STATUS
) {
1120 if (get_msr(cpu
, MSR_PKG_PERF_STATUS
, &msr
))
1122 p
->rapl_pkg_perf_status
= msr
& 0xFFFFFFFF;
1124 if (do_rapl
& RAPL_DRAM_PERF_STATUS
) {
1125 if (get_msr(cpu
, MSR_DRAM_PERF_STATUS
, &msr
))
1127 p
->rapl_dram_perf_status
= msr
& 0xFFFFFFFF;
1130 if (get_msr(cpu
, MSR_IA32_PACKAGE_THERM_STATUS
, &msr
))
1132 p
->pkg_temp_c
= tcc_activation_temp
- ((msr
>> 16) & 0x7F);
1138 * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1139 * If you change the values, note they are used both in comparisons
1140 * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1143 #define PCLUKN 0 /* Unknown */
1144 #define PCLRSV 1 /* Reserved */
1145 #define PCL__0 2 /* PC0 */
1146 #define PCL__1 3 /* PC1 */
1147 #define PCL__2 4 /* PC2 */
1148 #define PCL__3 5 /* PC3 */
1149 #define PCL__4 6 /* PC4 */
1150 #define PCL__6 7 /* PC6 */
1151 #define PCL_6N 8 /* PC6 No Retention */
1152 #define PCL_6R 9 /* PC6 Retention */
1153 #define PCL__7 10 /* PC7 */
1154 #define PCL_7S 11 /* PC7 Shrink */
1155 #define PCL__8 12 /* PC8 */
1156 #define PCL__9 13 /* PC9 */
1157 #define PCLUNL 14 /* Unlimited */
1159 int pkg_cstate_limit
= PCLUKN
;
1160 char *pkg_cstate_limit_strings
[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1161 "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "unlimited"};
1163 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
};
1164 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
};
1165 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
};
1166 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
};
1167 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
};
1168 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
};
1172 calculate_tsc_tweak()
1174 tsc_tweak
= base_hz
/ tsc_hz
;
1178 dump_nhm_platform_info(void)
1180 unsigned long long msr
;
1183 get_msr(base_cpu
, MSR_PLATFORM_INFO
, &msr
);
1185 fprintf(stderr
, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu
, msr
);
1187 ratio
= (msr
>> 40) & 0xFF;
1188 fprintf(stderr
, "%d * %.0f = %.0f MHz max efficiency frequency\n",
1189 ratio
, bclk
, ratio
* bclk
);
1191 ratio
= (msr
>> 8) & 0xFF;
1192 fprintf(stderr
, "%d * %.0f = %.0f MHz base frequency\n",
1193 ratio
, bclk
, ratio
* bclk
);
1195 get_msr(base_cpu
, MSR_IA32_POWER_CTL
, &msr
);
1196 fprintf(stderr
, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1197 base_cpu
, msr
, msr
& 0x2 ? "EN" : "DIS");
1203 dump_hsw_turbo_ratio_limits(void)
1205 unsigned long long msr
;
1208 get_msr(base_cpu
, MSR_TURBO_RATIO_LIMIT2
, &msr
);
1210 fprintf(stderr
, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu
, msr
);
1212 ratio
= (msr
>> 8) & 0xFF;
1214 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 18 active cores\n",
1215 ratio
, bclk
, ratio
* bclk
);
1217 ratio
= (msr
>> 0) & 0xFF;
1219 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 17 active cores\n",
1220 ratio
, bclk
, ratio
* bclk
);
1225 dump_ivt_turbo_ratio_limits(void)
1227 unsigned long long msr
;
1230 get_msr(base_cpu
, MSR_TURBO_RATIO_LIMIT1
, &msr
);
1232 fprintf(stderr
, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu
, msr
);
1234 ratio
= (msr
>> 56) & 0xFF;
1236 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1237 ratio
, bclk
, ratio
* bclk
);
1239 ratio
= (msr
>> 48) & 0xFF;
1241 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1242 ratio
, bclk
, ratio
* bclk
);
1244 ratio
= (msr
>> 40) & 0xFF;
1246 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1247 ratio
, bclk
, ratio
* bclk
);
1249 ratio
= (msr
>> 32) & 0xFF;
1251 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1252 ratio
, bclk
, ratio
* bclk
);
1254 ratio
= (msr
>> 24) & 0xFF;
1256 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1257 ratio
, bclk
, ratio
* bclk
);
1259 ratio
= (msr
>> 16) & 0xFF;
1261 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1262 ratio
, bclk
, ratio
* bclk
);
1264 ratio
= (msr
>> 8) & 0xFF;
1266 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1267 ratio
, bclk
, ratio
* bclk
);
1269 ratio
= (msr
>> 0) & 0xFF;
1271 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1272 ratio
, bclk
, ratio
* bclk
);
1277 dump_nhm_turbo_ratio_limits(void)
1279 unsigned long long msr
;
1282 get_msr(base_cpu
, MSR_TURBO_RATIO_LIMIT
, &msr
);
1284 fprintf(stderr
, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu
, msr
);
1286 ratio
= (msr
>> 56) & 0xFF;
1288 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1289 ratio
, bclk
, ratio
* bclk
);
1291 ratio
= (msr
>> 48) & 0xFF;
1293 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1294 ratio
, bclk
, ratio
* bclk
);
1296 ratio
= (msr
>> 40) & 0xFF;
1298 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1299 ratio
, bclk
, ratio
* bclk
);
1301 ratio
= (msr
>> 32) & 0xFF;
1303 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1304 ratio
, bclk
, ratio
* bclk
);
1306 ratio
= (msr
>> 24) & 0xFF;
1308 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1309 ratio
, bclk
, ratio
* bclk
);
1311 ratio
= (msr
>> 16) & 0xFF;
1313 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1314 ratio
, bclk
, ratio
* bclk
);
1316 ratio
= (msr
>> 8) & 0xFF;
1318 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1319 ratio
, bclk
, ratio
* bclk
);
1321 ratio
= (msr
>> 0) & 0xFF;
1323 fprintf(stderr
, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1324 ratio
, bclk
, ratio
* bclk
);
1329 dump_knl_turbo_ratio_limits(void)
1331 const unsigned int buckets_no
= 7;
1333 unsigned long long msr
;
1334 int delta_cores
, delta_ratio
;
1336 unsigned int cores
[buckets_no
];
1337 unsigned int ratio
[buckets_no
];
1339 get_msr(base_cpu
, MSR_NHM_TURBO_RATIO_LIMIT
, &msr
);
1341 fprintf(stderr
, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n",
1345 * Turbo encoding in KNL is as follows:
1347 * [7:1] -- Base value of number of active cores of bucket 1.
1348 * [15:8] -- Base value of freq ratio of bucket 1.
1349 * [20:16] -- +ve delta of number of active cores of bucket 2.
1350 * i.e. active cores of bucket 2 =
1351 * active cores of bucket 1 + delta
1352 * [23:21] -- Negative delta of freq ratio of bucket 2.
1353 * i.e. freq ratio of bucket 2 =
1354 * freq ratio of bucket 1 - delta
1355 * [28:24]-- +ve delta of number of active cores of bucket 3.
1356 * [31:29]-- -ve delta of freq ratio of bucket 3.
1357 * [36:32]-- +ve delta of number of active cores of bucket 4.
1358 * [39:37]-- -ve delta of freq ratio of bucket 4.
1359 * [44:40]-- +ve delta of number of active cores of bucket 5.
1360 * [47:45]-- -ve delta of freq ratio of bucket 5.
1361 * [52:48]-- +ve delta of number of active cores of bucket 6.
1362 * [55:53]-- -ve delta of freq ratio of bucket 6.
1363 * [60:56]-- +ve delta of number of active cores of bucket 7.
1364 * [63:61]-- -ve delta of freq ratio of bucket 7.
1368 cores
[b_nr
] = (msr
& 0xFF) >> 1;
1369 ratio
[b_nr
] = (msr
>> 8) & 0xFF;
1371 for (i
= 16; i
< 64; i
+= 8) {
1372 delta_cores
= (msr
>> i
) & 0x1F;
1373 delta_ratio
= (msr
>> (i
+ 5)) & 0x7;
1375 cores
[b_nr
+ 1] = cores
[b_nr
] + delta_cores
;
1376 ratio
[b_nr
+ 1] = ratio
[b_nr
] - delta_ratio
;
1380 for (i
= buckets_no
- 1; i
>= 0; i
--)
1381 if (i
> 0 ? ratio
[i
] != ratio
[i
- 1] : 1)
1383 "%d * %.0f = %.0f MHz max turbo %d active cores\n",
1384 ratio
[i
], bclk
, ratio
[i
] * bclk
, cores
[i
]);
1388 dump_nhm_cst_cfg(void)
1390 unsigned long long msr
;
1392 get_msr(base_cpu
, MSR_NHM_SNB_PKG_CST_CFG_CTL
, &msr
);
1394 #define SNB_C1_AUTO_UNDEMOTE (1UL << 27)
1395 #define SNB_C3_AUTO_UNDEMOTE (1UL << 28)
1397 fprintf(stderr
, "cpu%d: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", base_cpu
, msr
);
1399 fprintf(stderr
, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: %s)\n",
1400 (msr
& SNB_C3_AUTO_UNDEMOTE
) ? "UNdemote-C3, " : "",
1401 (msr
& SNB_C1_AUTO_UNDEMOTE
) ? "UNdemote-C1, " : "",
1402 (msr
& NHM_C3_AUTO_DEMOTE
) ? "demote-C3, " : "",
1403 (msr
& NHM_C1_AUTO_DEMOTE
) ? "demote-C1, " : "",
1404 (msr
& (1 << 15)) ? "" : "UN",
1405 (unsigned int)msr
& 7,
1406 pkg_cstate_limit_strings
[pkg_cstate_limit
]);
1411 dump_config_tdp(void)
1413 unsigned long long msr
;
1415 get_msr(base_cpu
, MSR_CONFIG_TDP_NOMINAL
, &msr
);
1416 fprintf(stderr
, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu
, msr
);
1417 fprintf(stderr
, " (base_ratio=%d)\n", (unsigned int)msr
& 0xEF);
1419 get_msr(base_cpu
, MSR_CONFIG_TDP_LEVEL_1
, &msr
);
1420 fprintf(stderr
, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu
, msr
);
1422 fprintf(stderr
, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr
>> 48) & 0xEFFF);
1423 fprintf(stderr
, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr
>> 32) & 0xEFFF);
1424 fprintf(stderr
, "LVL1_RATIO=%d ", (unsigned int)(msr
>> 16) & 0xEF);
1425 fprintf(stderr
, "PKG_TDP_LVL1=%d", (unsigned int)(msr
) & 0xEFFF);
1427 fprintf(stderr
, ")\n");
1429 get_msr(base_cpu
, MSR_CONFIG_TDP_LEVEL_2
, &msr
);
1430 fprintf(stderr
, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu
, msr
);
1432 fprintf(stderr
, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr
>> 48) & 0xEFFF);
1433 fprintf(stderr
, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr
>> 32) & 0xEFFF);
1434 fprintf(stderr
, "LVL2_RATIO=%d ", (unsigned int)(msr
>> 16) & 0xEF);
1435 fprintf(stderr
, "PKG_TDP_LVL2=%d", (unsigned int)(msr
) & 0xEFFF);
1437 fprintf(stderr
, ")\n");
1439 get_msr(base_cpu
, MSR_CONFIG_TDP_CONTROL
, &msr
);
1440 fprintf(stderr
, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu
, msr
);
1442 fprintf(stderr
, "TDP_LEVEL=%d ", (unsigned int)(msr
) & 0x3);
1443 fprintf(stderr
, " lock=%d", (unsigned int)(msr
>> 31) & 1);
1444 fprintf(stderr
, ")\n");
1446 get_msr(base_cpu
, MSR_TURBO_ACTIVATION_RATIO
, &msr
);
1447 fprintf(stderr
, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu
, msr
);
1448 fprintf(stderr
, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr
) & 0x7F);
1449 fprintf(stderr
, " lock=%d", (unsigned int)(msr
>> 31) & 1);
1450 fprintf(stderr
, ")\n");
1453 void free_all_buffers(void)
1455 CPU_FREE(cpu_present_set
);
1456 cpu_present_set
= NULL
;
1457 cpu_present_set
= 0;
1459 CPU_FREE(cpu_affinity_set
);
1460 cpu_affinity_set
= NULL
;
1461 cpu_affinity_setsize
= 0;
1469 package_even
= NULL
;
1479 free(output_buffer
);
1480 output_buffer
= NULL
;
1485 * Open a file, and exit on failure
1487 FILE *fopen_or_die(const char *path
, const char *mode
)
1489 FILE *filep
= fopen(path
, "r");
1491 err(1, "%s: open failed", path
);
1496 * Parse a file containing a single int.
1498 int parse_int_file(const char *fmt
, ...)
1501 char path
[PATH_MAX
];
1505 va_start(args
, fmt
);
1506 vsnprintf(path
, sizeof(path
), fmt
, args
);
1508 filep
= fopen_or_die(path
, "r");
1509 if (fscanf(filep
, "%d", &value
) != 1)
1510 err(1, "%s: failed to parse number from file", path
);
1516 * get_cpu_position_in_core(cpu)
1517 * return the position of the CPU among its HT siblings in the core
1518 * return -1 if the sibling is not in list
1520 int get_cpu_position_in_core(int cpu
)
1529 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list",
1531 filep
= fopen(path
, "r");
1532 if (filep
== NULL
) {
1537 for (i
= 0; i
< topo
.num_threads_per_core
; i
++) {
1538 fscanf(filep
, "%d", &this_cpu
);
1539 if (this_cpu
== cpu
) {
1544 /* Account for no separator after last thread*/
1545 if (i
!= (topo
.num_threads_per_core
- 1))
1546 fscanf(filep
, "%c", &character
);
1554 * cpu_is_first_core_in_package(cpu)
1555 * return 1 if given CPU is 1st core in package
1557 int cpu_is_first_core_in_package(int cpu
)
1559 return cpu
== parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu
);
1562 int get_physical_package_id(int cpu
)
1564 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu
);
1567 int get_core_id(int cpu
)
1569 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu
);
1572 int get_num_ht_siblings(int cpu
)
1582 sprintf(path
, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu
);
1583 filep
= fopen_or_die(path
, "r");
1587 * A ',' separated or '-' separated set of numbers
1588 * (eg 1-2 or 1,3,4,5)
1590 fscanf(filep
, "%d%c\n", &sib1
, &character
);
1591 fseek(filep
, 0, SEEK_SET
);
1592 fgets(str
, 100, filep
);
1593 ch
= strchr(str
, character
);
1594 while (ch
!= NULL
) {
1596 ch
= strchr(ch
+1, character
);
1604 * run func(thread, core, package) in topology order
1605 * skip non-present cpus
1608 int for_all_cpus_2(int (func
)(struct thread_data
*, struct core_data
*,
1609 struct pkg_data
*, struct thread_data
*, struct core_data
*,
1610 struct pkg_data
*), struct thread_data
*thread_base
,
1611 struct core_data
*core_base
, struct pkg_data
*pkg_base
,
1612 struct thread_data
*thread_base2
, struct core_data
*core_base2
,
1613 struct pkg_data
*pkg_base2
)
1615 int retval
, pkg_no
, core_no
, thread_no
;
1617 for (pkg_no
= 0; pkg_no
< topo
.num_packages
; ++pkg_no
) {
1618 for (core_no
= 0; core_no
< topo
.num_cores_per_pkg
; ++core_no
) {
1619 for (thread_no
= 0; thread_no
<
1620 topo
.num_threads_per_core
; ++thread_no
) {
1621 struct thread_data
*t
, *t2
;
1622 struct core_data
*c
, *c2
;
1623 struct pkg_data
*p
, *p2
;
1625 t
= GET_THREAD(thread_base
, thread_no
, core_no
, pkg_no
);
1627 if (cpu_is_not_present(t
->cpu_id
))
1630 t2
= GET_THREAD(thread_base2
, thread_no
, core_no
, pkg_no
);
1632 c
= GET_CORE(core_base
, core_no
, pkg_no
);
1633 c2
= GET_CORE(core_base2
, core_no
, pkg_no
);
1635 p
= GET_PKG(pkg_base
, pkg_no
);
1636 p2
= GET_PKG(pkg_base2
, pkg_no
);
1638 retval
= func(t
, c
, p
, t2
, c2
, p2
);
1648 * run func(cpu) on every cpu in /proc/stat
1649 * return max_cpu number
1651 int for_all_proc_cpus(int (func
)(int))
1657 fp
= fopen_or_die(proc_stat
, "r");
1659 retval
= fscanf(fp
, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1661 err(1, "%s: failed to parse format", proc_stat
);
1664 retval
= fscanf(fp
, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num
);
1668 retval
= func(cpu_num
);
1678 void re_initialize(void)
1681 setup_all_buffers();
1682 printf("turbostat: re-initialized with num_cpus %d\n", topo
.num_cpus
);
1688 * remember the last one seen, it will be the max
1690 int count_cpus(int cpu
)
1692 if (topo
.max_cpu_num
< cpu
)
1693 topo
.max_cpu_num
= cpu
;
1698 int mark_cpu_present(int cpu
)
1700 CPU_SET_S(cpu
, cpu_present_setsize
, cpu_present_set
);
1704 void turbostat_loop()
1712 retval
= for_all_cpus(get_counters
, EVEN_COUNTERS
);
1715 } else if (retval
== -1) {
1716 if (restarted
> 1) {
1723 gettimeofday(&tv_even
, (struct timezone
*)NULL
);
1726 if (for_all_proc_cpus(cpu_is_not_present
)) {
1730 nanosleep(&interval_ts
, NULL
);
1731 retval
= for_all_cpus(get_counters
, ODD_COUNTERS
);
1734 } else if (retval
== -1) {
1738 gettimeofday(&tv_odd
, (struct timezone
*)NULL
);
1739 timersub(&tv_odd
, &tv_even
, &tv_delta
);
1740 for_all_cpus_2(delta_cpu
, ODD_COUNTERS
, EVEN_COUNTERS
);
1741 compute_average(EVEN_COUNTERS
);
1742 format_all_counters(EVEN_COUNTERS
);
1744 nanosleep(&interval_ts
, NULL
);
1745 retval
= for_all_cpus(get_counters
, EVEN_COUNTERS
);
1748 } else if (retval
== -1) {
1752 gettimeofday(&tv_even
, (struct timezone
*)NULL
);
1753 timersub(&tv_even
, &tv_odd
, &tv_delta
);
1754 for_all_cpus_2(delta_cpu
, EVEN_COUNTERS
, ODD_COUNTERS
);
1755 compute_average(ODD_COUNTERS
);
1756 format_all_counters(ODD_COUNTERS
);
1761 void check_dev_msr()
1766 sprintf(pathname
, "/dev/cpu/%d/msr", base_cpu
);
1767 if (stat(pathname
, &sb
))
1768 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
1769 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
1772 void check_permissions()
1774 struct __user_cap_header_struct cap_header_data
;
1775 cap_user_header_t cap_header
= &cap_header_data
;
1776 struct __user_cap_data_struct cap_data_data
;
1777 cap_user_data_t cap_data
= &cap_data_data
;
1778 extern int capget(cap_user_header_t hdrp
, cap_user_data_t datap
);
1782 /* check for CAP_SYS_RAWIO */
1783 cap_header
->pid
= getpid();
1784 cap_header
->version
= _LINUX_CAPABILITY_VERSION
;
1785 if (capget(cap_header
, cap_data
) < 0)
1786 err(-6, "capget(2) failed");
1788 if ((cap_data
->effective
& (1 << CAP_SYS_RAWIO
)) == 0) {
1790 warnx("capget(CAP_SYS_RAWIO) failed,"
1791 " try \"# setcap cap_sys_rawio=ep %s\"", progname
);
1794 /* test file permissions */
1795 sprintf(pathname
, "/dev/cpu/%d/msr", base_cpu
);
1796 if (euidaccess(pathname
, R_OK
)) {
1798 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1801 /* if all else fails, thell them to be root */
1804 warnx("... or simply run as root");
1811 * NHM adds support for additional MSRs:
1813 * MSR_SMI_COUNT 0x00000034
1815 * MSR_PLATFORM_INFO 0x000000ce
1816 * MSR_NHM_SNB_PKG_CST_CFG_CTL 0x000000e2
1818 * MSR_PKG_C3_RESIDENCY 0x000003f8
1819 * MSR_PKG_C6_RESIDENCY 0x000003f9
1820 * MSR_CORE_C3_RESIDENCY 0x000003fc
1821 * MSR_CORE_C6_RESIDENCY 0x000003fd
1824 * sets global pkg_cstate_limit to decode MSR_NHM_SNB_PKG_CST_CFG_CTL
1826 int probe_nhm_msrs(unsigned int family
, unsigned int model
)
1828 unsigned long long msr
;
1829 unsigned int base_ratio
;
1830 int *pkg_cstate_limits
;
1838 bclk
= discover_bclk(family
, model
);
1841 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1842 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1843 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1844 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1845 case 0x2C: /* Westmere EP - Gulftown */
1846 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1847 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1848 pkg_cstate_limits
= nhm_pkg_cstate_limits
;
1850 case 0x2A: /* SNB */
1851 case 0x2D: /* SNB Xeon */
1852 case 0x3A: /* IVB */
1853 case 0x3E: /* IVB Xeon */
1854 pkg_cstate_limits
= snb_pkg_cstate_limits
;
1856 case 0x3C: /* HSW */
1857 case 0x3F: /* HSX */
1858 case 0x45: /* HSW */
1859 case 0x46: /* HSW */
1860 case 0x3D: /* BDW */
1861 case 0x47: /* BDW */
1862 case 0x4F: /* BDX */
1863 case 0x56: /* BDX-DE */
1864 case 0x4E: /* SKL */
1865 case 0x5E: /* SKL */
1866 pkg_cstate_limits
= hsw_pkg_cstate_limits
;
1868 case 0x37: /* BYT */
1869 case 0x4D: /* AVN */
1870 pkg_cstate_limits
= slv_pkg_cstate_limits
;
1872 case 0x4C: /* AMT */
1873 pkg_cstate_limits
= amt_pkg_cstate_limits
;
1875 case 0x57: /* PHI */
1876 pkg_cstate_limits
= phi_pkg_cstate_limits
;
1881 get_msr(base_cpu
, MSR_NHM_SNB_PKG_CST_CFG_CTL
, &msr
);
1882 pkg_cstate_limit
= pkg_cstate_limits
[msr
& 0xF];
1884 get_msr(base_cpu
, MSR_PLATFORM_INFO
, &msr
);
1885 base_ratio
= (msr
>> 8) & 0xFF;
1887 base_hz
= base_ratio
* bclk
* 1000000;
1891 int has_nhm_turbo_ratio_limit(unsigned int family
, unsigned int model
)
1894 /* Nehalem compatible, but do not include turbo-ratio limit support */
1895 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1896 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1897 case 0x57: /* PHI - Knights Landing (different MSR definition) */
1903 int has_ivt_turbo_ratio_limit(unsigned int family
, unsigned int model
)
1912 case 0x3E: /* IVB Xeon */
1913 case 0x3F: /* HSW Xeon */
1919 int has_hsw_turbo_ratio_limit(unsigned int family
, unsigned int model
)
1928 case 0x3F: /* HSW Xeon */
1935 int has_knl_turbo_ratio_limit(unsigned int family
, unsigned int model
)
1944 case 0x57: /* Knights Landing */
1950 int has_config_tdp(unsigned int family
, unsigned int model
)
1959 case 0x3A: /* IVB */
1960 case 0x3C: /* HSW */
1961 case 0x3F: /* HSX */
1962 case 0x45: /* HSW */
1963 case 0x46: /* HSW */
1964 case 0x3D: /* BDW */
1965 case 0x47: /* BDW */
1966 case 0x4F: /* BDX */
1967 case 0x56: /* BDX-DE */
1968 case 0x4E: /* SKL */
1969 case 0x5E: /* SKL */
1971 case 0x57: /* Knights Landing */
1979 dump_cstate_pstate_config_info(family
, model
)
1981 if (!do_nhm_platform_info
)
1984 dump_nhm_platform_info();
1986 if (has_hsw_turbo_ratio_limit(family
, model
))
1987 dump_hsw_turbo_ratio_limits();
1989 if (has_ivt_turbo_ratio_limit(family
, model
))
1990 dump_ivt_turbo_ratio_limits();
1992 if (has_nhm_turbo_ratio_limit(family
, model
))
1993 dump_nhm_turbo_ratio_limits();
1995 if (has_knl_turbo_ratio_limit(family
, model
))
1996 dump_knl_turbo_ratio_limits();
1998 if (has_config_tdp(family
, model
))
2007 * Decode the ENERGY_PERF_BIAS MSR
2009 int print_epb(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
2011 unsigned long long msr
;
2020 /* EPB is per-package */
2021 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
) || !(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
2024 if (cpu_migrate(cpu
)) {
2025 fprintf(stderr
, "Could not migrate to CPU %d\n", cpu
);
2029 if (get_msr(cpu
, MSR_IA32_ENERGY_PERF_BIAS
, &msr
))
2032 switch (msr
& 0xF) {
2033 case ENERGY_PERF_BIAS_PERFORMANCE
:
2034 epb_string
= "performance";
2036 case ENERGY_PERF_BIAS_NORMAL
:
2037 epb_string
= "balanced";
2039 case ENERGY_PERF_BIAS_POWERSAVE
:
2040 epb_string
= "powersave";
2043 epb_string
= "custom";
2046 fprintf(stderr
, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu
, msr
, epb_string
);
2052 * Decode the MSR_HWP_CAPABILITIES
2054 int print_hwp(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
2056 unsigned long long msr
;
2064 /* MSR_HWP_CAPABILITIES is per-package */
2065 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
) || !(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
2068 if (cpu_migrate(cpu
)) {
2069 fprintf(stderr
, "Could not migrate to CPU %d\n", cpu
);
2073 if (get_msr(cpu
, MSR_PM_ENABLE
, &msr
))
2076 fprintf(stderr
, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n",
2077 cpu
, msr
, (msr
& (1 << 0)) ? "" : "No-");
2079 /* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
2080 if ((msr
& (1 << 0)) == 0)
2083 if (get_msr(cpu
, MSR_HWP_CAPABILITIES
, &msr
))
2086 fprintf(stderr
, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
2087 "(high 0x%x guar 0x%x eff 0x%x low 0x%x)\n",
2089 (unsigned int)HWP_HIGHEST_PERF(msr
),
2090 (unsigned int)HWP_GUARANTEED_PERF(msr
),
2091 (unsigned int)HWP_MOSTEFFICIENT_PERF(msr
),
2092 (unsigned int)HWP_LOWEST_PERF(msr
));
2094 if (get_msr(cpu
, MSR_HWP_REQUEST
, &msr
))
2097 fprintf(stderr
, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
2098 "(min 0x%x max 0x%x des 0x%x epp 0x%x window 0x%x pkg 0x%x)\n",
2100 (unsigned int)(((msr
) >> 0) & 0xff),
2101 (unsigned int)(((msr
) >> 8) & 0xff),
2102 (unsigned int)(((msr
) >> 16) & 0xff),
2103 (unsigned int)(((msr
) >> 24) & 0xff),
2104 (unsigned int)(((msr
) >> 32) & 0xff3),
2105 (unsigned int)(((msr
) >> 42) & 0x1));
2108 if (get_msr(cpu
, MSR_HWP_REQUEST_PKG
, &msr
))
2111 fprintf(stderr
, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
2112 "(min 0x%x max 0x%x des 0x%x epp 0x%x window 0x%x)\n",
2114 (unsigned int)(((msr
) >> 0) & 0xff),
2115 (unsigned int)(((msr
) >> 8) & 0xff),
2116 (unsigned int)(((msr
) >> 16) & 0xff),
2117 (unsigned int)(((msr
) >> 24) & 0xff),
2118 (unsigned int)(((msr
) >> 32) & 0xff3));
2120 if (has_hwp_notify
) {
2121 if (get_msr(cpu
, MSR_HWP_INTERRUPT
, &msr
))
2124 fprintf(stderr
, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
2125 "(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
2127 ((msr
) & 0x1) ? "EN" : "Dis",
2128 ((msr
) & 0x2) ? "EN" : "Dis");
2130 if (get_msr(cpu
, MSR_HWP_STATUS
, &msr
))
2133 fprintf(stderr
, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
2134 "(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
2136 ((msr
) & 0x1) ? "" : "No-",
2137 ((msr
) & 0x2) ? "" : "No-");
2143 * print_perf_limit()
2145 int print_perf_limit(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
2147 unsigned long long msr
;
2153 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
) || !(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
2156 if (cpu_migrate(cpu
)) {
2157 fprintf(stderr
, "Could not migrate to CPU %d\n", cpu
);
2161 if (do_core_perf_limit_reasons
) {
2162 get_msr(cpu
, MSR_CORE_PERF_LIMIT_REASONS
, &msr
);
2163 fprintf(stderr
, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu
, msr
);
2164 fprintf(stderr
, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
2165 (msr
& 1 << 15) ? "bit15, " : "",
2166 (msr
& 1 << 14) ? "bit14, " : "",
2167 (msr
& 1 << 13) ? "Transitions, " : "",
2168 (msr
& 1 << 12) ? "MultiCoreTurbo, " : "",
2169 (msr
& 1 << 11) ? "PkgPwrL2, " : "",
2170 (msr
& 1 << 10) ? "PkgPwrL1, " : "",
2171 (msr
& 1 << 9) ? "CorePwr, " : "",
2172 (msr
& 1 << 8) ? "Amps, " : "",
2173 (msr
& 1 << 6) ? "VR-Therm, " : "",
2174 (msr
& 1 << 5) ? "Auto-HWP, " : "",
2175 (msr
& 1 << 4) ? "Graphics, " : "",
2176 (msr
& 1 << 2) ? "bit2, " : "",
2177 (msr
& 1 << 1) ? "ThermStatus, " : "",
2178 (msr
& 1 << 0) ? "PROCHOT, " : "");
2179 fprintf(stderr
, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
2180 (msr
& 1 << 31) ? "bit31, " : "",
2181 (msr
& 1 << 30) ? "bit30, " : "",
2182 (msr
& 1 << 29) ? "Transitions, " : "",
2183 (msr
& 1 << 28) ? "MultiCoreTurbo, " : "",
2184 (msr
& 1 << 27) ? "PkgPwrL2, " : "",
2185 (msr
& 1 << 26) ? "PkgPwrL1, " : "",
2186 (msr
& 1 << 25) ? "CorePwr, " : "",
2187 (msr
& 1 << 24) ? "Amps, " : "",
2188 (msr
& 1 << 22) ? "VR-Therm, " : "",
2189 (msr
& 1 << 21) ? "Auto-HWP, " : "",
2190 (msr
& 1 << 20) ? "Graphics, " : "",
2191 (msr
& 1 << 18) ? "bit18, " : "",
2192 (msr
& 1 << 17) ? "ThermStatus, " : "",
2193 (msr
& 1 << 16) ? "PROCHOT, " : "");
2196 if (do_gfx_perf_limit_reasons
) {
2197 get_msr(cpu
, MSR_GFX_PERF_LIMIT_REASONS
, &msr
);
2198 fprintf(stderr
, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu
, msr
);
2199 fprintf(stderr
, " (Active: %s%s%s%s%s%s%s%s)",
2200 (msr
& 1 << 0) ? "PROCHOT, " : "",
2201 (msr
& 1 << 1) ? "ThermStatus, " : "",
2202 (msr
& 1 << 4) ? "Graphics, " : "",
2203 (msr
& 1 << 6) ? "VR-Therm, " : "",
2204 (msr
& 1 << 8) ? "Amps, " : "",
2205 (msr
& 1 << 9) ? "GFXPwr, " : "",
2206 (msr
& 1 << 10) ? "PkgPwrL1, " : "",
2207 (msr
& 1 << 11) ? "PkgPwrL2, " : "");
2208 fprintf(stderr
, " (Logged: %s%s%s%s%s%s%s%s)\n",
2209 (msr
& 1 << 16) ? "PROCHOT, " : "",
2210 (msr
& 1 << 17) ? "ThermStatus, " : "",
2211 (msr
& 1 << 20) ? "Graphics, " : "",
2212 (msr
& 1 << 22) ? "VR-Therm, " : "",
2213 (msr
& 1 << 24) ? "Amps, " : "",
2214 (msr
& 1 << 25) ? "GFXPwr, " : "",
2215 (msr
& 1 << 26) ? "PkgPwrL1, " : "",
2216 (msr
& 1 << 27) ? "PkgPwrL2, " : "");
2218 if (do_ring_perf_limit_reasons
) {
2219 get_msr(cpu
, MSR_RING_PERF_LIMIT_REASONS
, &msr
);
2220 fprintf(stderr
, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu
, msr
);
2221 fprintf(stderr
, " (Active: %s%s%s%s%s%s)",
2222 (msr
& 1 << 0) ? "PROCHOT, " : "",
2223 (msr
& 1 << 1) ? "ThermStatus, " : "",
2224 (msr
& 1 << 6) ? "VR-Therm, " : "",
2225 (msr
& 1 << 8) ? "Amps, " : "",
2226 (msr
& 1 << 10) ? "PkgPwrL1, " : "",
2227 (msr
& 1 << 11) ? "PkgPwrL2, " : "");
2228 fprintf(stderr
, " (Logged: %s%s%s%s%s%s)\n",
2229 (msr
& 1 << 16) ? "PROCHOT, " : "",
2230 (msr
& 1 << 17) ? "ThermStatus, " : "",
2231 (msr
& 1 << 22) ? "VR-Therm, " : "",
2232 (msr
& 1 << 24) ? "Amps, " : "",
2233 (msr
& 1 << 26) ? "PkgPwrL1, " : "",
2234 (msr
& 1 << 27) ? "PkgPwrL2, " : "");
2239 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
2240 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
2242 double get_tdp(model
)
2244 unsigned long long msr
;
2246 if (do_rapl
& RAPL_PKG_POWER_INFO
)
2247 if (!get_msr(base_cpu
, MSR_PKG_POWER_INFO
, &msr
))
2248 return ((msr
>> 0) & RAPL_POWER_GRANULARITY
) * rapl_power_units
;
2260 * rapl_dram_energy_units_probe()
2261 * Energy units are either hard-coded, or come from RAPL Energy Unit MSR.
2264 rapl_dram_energy_units_probe(int model
, double rapl_energy_units
)
2266 /* only called for genuine_intel, family 6 */
2269 case 0x3F: /* HSX */
2270 case 0x4F: /* BDX */
2271 case 0x56: /* BDX-DE */
2272 case 0x57: /* KNL */
2273 return (rapl_dram_energy_units
= 15.3 / 1000000);
2275 return (rapl_energy_units
);
2283 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
2285 void rapl_probe(unsigned int family
, unsigned int model
)
2287 unsigned long long msr
;
2288 unsigned int time_unit
;
2300 case 0x3C: /* HSW */
2301 case 0x45: /* HSW */
2302 case 0x46: /* HSW */
2303 case 0x3D: /* BDW */
2304 case 0x47: /* BDW */
2305 do_rapl
= RAPL_PKG
| RAPL_CORES
| RAPL_CORE_POLICY
| RAPL_GFX
| RAPL_PKG_POWER_INFO
;
2307 case 0x4E: /* SKL */
2308 case 0x5E: /* SKL */
2309 do_rapl
= RAPL_PKG
| RAPL_DRAM
| RAPL_DRAM_PERF_STATUS
| RAPL_PKG_PERF_STATUS
| RAPL_PKG_POWER_INFO
;
2311 case 0x3F: /* HSX */
2312 case 0x4F: /* BDX */
2313 case 0x56: /* BDX-DE */
2314 case 0x57: /* KNL */
2315 do_rapl
= RAPL_PKG
| RAPL_DRAM
| RAPL_DRAM_POWER_INFO
| RAPL_DRAM_PERF_STATUS
| RAPL_PKG_PERF_STATUS
| RAPL_PKG_POWER_INFO
;
2319 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
;
2321 case 0x37: /* BYT */
2322 case 0x4D: /* AVN */
2323 do_rapl
= RAPL_PKG
| RAPL_CORES
;
2329 /* units on package 0, verify later other packages match */
2330 if (get_msr(base_cpu
, MSR_RAPL_POWER_UNIT
, &msr
))
2333 rapl_power_units
= 1.0 / (1 << (msr
& 0xF));
2335 rapl_energy_units
= 1.0 * (1 << (msr
>> 8 & 0x1F)) / 1000000;
2337 rapl_energy_units
= 1.0 / (1 << (msr
>> 8 & 0x1F));
2339 rapl_dram_energy_units
= rapl_dram_energy_units_probe(model
, rapl_energy_units
);
2341 time_unit
= msr
>> 16 & 0xF;
2345 rapl_time_units
= 1.0 / (1 << (time_unit
));
2347 tdp
= get_tdp(model
);
2349 rapl_joule_counter_range
= 0xFFFFFFFF * rapl_energy_units
/ tdp
;
2351 fprintf(stderr
, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range
, tdp
);
2356 void perf_limit_reasons_probe(family
, model
)
2365 case 0x3C: /* HSW */
2366 case 0x45: /* HSW */
2367 case 0x46: /* HSW */
2368 do_gfx_perf_limit_reasons
= 1;
2369 case 0x3F: /* HSX */
2370 do_core_perf_limit_reasons
= 1;
2371 do_ring_perf_limit_reasons
= 1;
2377 int print_thermal(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
2379 unsigned long long msr
;
2383 if (!(do_dts
|| do_ptm
))
2388 /* DTS is per-core, no need to print for each thread */
2389 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
))
2392 if (cpu_migrate(cpu
)) {
2393 fprintf(stderr
, "Could not migrate to CPU %d\n", cpu
);
2397 if (do_ptm
&& (t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
)) {
2398 if (get_msr(cpu
, MSR_IA32_PACKAGE_THERM_STATUS
, &msr
))
2401 dts
= (msr
>> 16) & 0x7F;
2402 fprintf(stderr
, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
2403 cpu
, msr
, tcc_activation_temp
- dts
);
2406 if (get_msr(cpu
, MSR_IA32_PACKAGE_THERM_INTERRUPT
, &msr
))
2409 dts
= (msr
>> 16) & 0x7F;
2410 dts2
= (msr
>> 8) & 0x7F;
2411 fprintf(stderr
, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2412 cpu
, msr
, tcc_activation_temp
- dts
, tcc_activation_temp
- dts2
);
2418 unsigned int resolution
;
2420 if (get_msr(cpu
, MSR_IA32_THERM_STATUS
, &msr
))
2423 dts
= (msr
>> 16) & 0x7F;
2424 resolution
= (msr
>> 27) & 0xF;
2425 fprintf(stderr
, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
2426 cpu
, msr
, tcc_activation_temp
- dts
, resolution
);
2429 if (get_msr(cpu
, MSR_IA32_THERM_INTERRUPT
, &msr
))
2432 dts
= (msr
>> 16) & 0x7F;
2433 dts2
= (msr
>> 8) & 0x7F;
2434 fprintf(stderr
, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2435 cpu
, msr
, tcc_activation_temp
- dts
, tcc_activation_temp
- dts2
);
2442 void print_power_limit_msr(int cpu
, unsigned long long msr
, char *label
)
2444 fprintf(stderr
, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
2446 ((msr
>> 15) & 1) ? "EN" : "DIS",
2447 ((msr
>> 0) & 0x7FFF) * rapl_power_units
,
2448 (1.0 + (((msr
>> 22) & 0x3)/4.0)) * (1 << ((msr
>> 17) & 0x1F)) * rapl_time_units
,
2449 (((msr
>> 16) & 1) ? "EN" : "DIS"));
2454 int print_rapl(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
2456 unsigned long long msr
;
2462 /* RAPL counters are per package, so print only for 1st thread/package */
2463 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
) || !(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
2467 if (cpu_migrate(cpu
)) {
2468 fprintf(stderr
, "Could not migrate to CPU %d\n", cpu
);
2472 if (get_msr(cpu
, MSR_RAPL_POWER_UNIT
, &msr
))
2476 fprintf(stderr
, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
2477 "(%f Watts, %f Joules, %f sec.)\n", cpu
, msr
,
2478 rapl_power_units
, rapl_energy_units
, rapl_time_units
);
2480 if (do_rapl
& RAPL_PKG_POWER_INFO
) {
2482 if (get_msr(cpu
, MSR_PKG_POWER_INFO
, &msr
))
2486 fprintf(stderr
, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2488 ((msr
>> 0) & RAPL_POWER_GRANULARITY
) * rapl_power_units
,
2489 ((msr
>> 16) & RAPL_POWER_GRANULARITY
) * rapl_power_units
,
2490 ((msr
>> 32) & RAPL_POWER_GRANULARITY
) * rapl_power_units
,
2491 ((msr
>> 48) & RAPL_TIME_GRANULARITY
) * rapl_time_units
);
2494 if (do_rapl
& RAPL_PKG
) {
2496 if (get_msr(cpu
, MSR_PKG_POWER_LIMIT
, &msr
))
2499 fprintf(stderr
, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
2500 cpu
, msr
, (msr
>> 63) & 1 ? "": "UN");
2502 print_power_limit_msr(cpu
, msr
, "PKG Limit #1");
2503 fprintf(stderr
, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
2505 ((msr
>> 47) & 1) ? "EN" : "DIS",
2506 ((msr
>> 32) & 0x7FFF) * rapl_power_units
,
2507 (1.0 + (((msr
>> 54) & 0x3)/4.0)) * (1 << ((msr
>> 49) & 0x1F)) * rapl_time_units
,
2508 ((msr
>> 48) & 1) ? "EN" : "DIS");
2511 if (do_rapl
& RAPL_DRAM_POWER_INFO
) {
2512 if (get_msr(cpu
, MSR_DRAM_POWER_INFO
, &msr
))
2515 fprintf(stderr
, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2517 ((msr
>> 0) & RAPL_POWER_GRANULARITY
) * rapl_power_units
,
2518 ((msr
>> 16) & RAPL_POWER_GRANULARITY
) * rapl_power_units
,
2519 ((msr
>> 32) & RAPL_POWER_GRANULARITY
) * rapl_power_units
,
2520 ((msr
>> 48) & RAPL_TIME_GRANULARITY
) * rapl_time_units
);
2522 if (do_rapl
& RAPL_DRAM
) {
2523 if (get_msr(cpu
, MSR_DRAM_POWER_LIMIT
, &msr
))
2525 fprintf(stderr
, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
2526 cpu
, msr
, (msr
>> 31) & 1 ? "": "UN");
2528 print_power_limit_msr(cpu
, msr
, "DRAM Limit");
2530 if (do_rapl
& RAPL_CORE_POLICY
) {
2532 if (get_msr(cpu
, MSR_PP0_POLICY
, &msr
))
2535 fprintf(stderr
, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu
, msr
& 0xF);
2538 if (do_rapl
& RAPL_CORES
) {
2541 if (get_msr(cpu
, MSR_PP0_POWER_LIMIT
, &msr
))
2543 fprintf(stderr
, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2544 cpu
, msr
, (msr
>> 31) & 1 ? "": "UN");
2545 print_power_limit_msr(cpu
, msr
, "Cores Limit");
2548 if (do_rapl
& RAPL_GFX
) {
2550 if (get_msr(cpu
, MSR_PP1_POLICY
, &msr
))
2553 fprintf(stderr
, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu
, msr
& 0xF);
2555 if (get_msr(cpu
, MSR_PP1_POWER_LIMIT
, &msr
))
2557 fprintf(stderr
, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2558 cpu
, msr
, (msr
>> 31) & 1 ? "": "UN");
2559 print_power_limit_msr(cpu
, msr
, "GFX Limit");
2566 * SNB adds support for additional MSRs:
2568 * MSR_PKG_C7_RESIDENCY 0x000003fa
2569 * MSR_CORE_C7_RESIDENCY 0x000003fe
2570 * MSR_PKG_C2_RESIDENCY 0x0000060d
2573 int has_snb_msrs(unsigned int family
, unsigned int model
)
2581 case 0x3A: /* IVB */
2582 case 0x3E: /* IVB Xeon */
2583 case 0x3C: /* HSW */
2584 case 0x3F: /* HSW */
2585 case 0x45: /* HSW */
2586 case 0x46: /* HSW */
2587 case 0x3D: /* BDW */
2588 case 0x47: /* BDW */
2589 case 0x4F: /* BDX */
2590 case 0x56: /* BDX-DE */
2591 case 0x4E: /* SKL */
2592 case 0x5E: /* SKL */
2599 * HSW adds support for additional MSRs:
2601 * MSR_PKG_C8_RESIDENCY 0x00000630
2602 * MSR_PKG_C9_RESIDENCY 0x00000631
2603 * MSR_PKG_C10_RESIDENCY 0x00000632
2605 int has_hsw_msrs(unsigned int family
, unsigned int model
)
2611 case 0x45: /* HSW */
2612 case 0x3D: /* BDW */
2613 case 0x4E: /* SKL */
2614 case 0x5E: /* SKL */
2621 * SKL adds support for additional MSRS:
2623 * MSR_PKG_WEIGHTED_CORE_C0_RES 0x00000658
2624 * MSR_PKG_ANY_CORE_C0_RES 0x00000659
2625 * MSR_PKG_ANY_GFXE_C0_RES 0x0000065A
2626 * MSR_PKG_BOTH_CORE_GFXE_C0_RES 0x0000065B
2628 int has_skl_msrs(unsigned int family
, unsigned int model
)
2634 case 0x4E: /* SKL */
2635 case 0x5E: /* SKL */
2643 int is_slm(unsigned int family
, unsigned int model
)
2648 case 0x37: /* BYT */
2649 case 0x4D: /* AVN */
2655 int is_knl(unsigned int family
, unsigned int model
)
2660 case 0x57: /* KNL */
2666 unsigned int get_aperf_mperf_multiplier(unsigned int family
, unsigned int model
)
2668 if (is_knl(family
, model
))
2673 #define SLM_BCLK_FREQS 5
2674 double slm_freq_table
[SLM_BCLK_FREQS
] = { 83.3, 100.0, 133.3, 116.7, 80.0};
2676 double slm_bclk(void)
2678 unsigned long long msr
= 3;
2682 if (get_msr(base_cpu
, MSR_FSB_FREQ
, &msr
))
2683 fprintf(stderr
, "SLM BCLK: unknown\n");
2686 if (i
>= SLM_BCLK_FREQS
) {
2687 fprintf(stderr
, "SLM BCLK[%d] invalid\n", i
);
2690 freq
= slm_freq_table
[i
];
2692 fprintf(stderr
, "SLM BCLK: %.1f Mhz\n", freq
);
2697 double discover_bclk(unsigned int family
, unsigned int model
)
2699 if (has_snb_msrs(family
, model
) || is_knl(family
, model
))
2701 else if (is_slm(family
, model
))
2708 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
2709 * the Thermal Control Circuit (TCC) activates.
2710 * This is usually equal to tjMax.
2712 * Older processors do not have this MSR, so there we guess,
2713 * but also allow cmdline over-ride with -T.
2715 * Several MSR temperature values are in units of degrees-C
2716 * below this value, including the Digital Thermal Sensor (DTS),
2717 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
2719 int set_temperature_target(struct thread_data
*t
, struct core_data
*c
, struct pkg_data
*p
)
2721 unsigned long long msr
;
2722 unsigned int target_c_local
;
2725 /* tcc_activation_temp is used only for dts or ptm */
2726 if (!(do_dts
|| do_ptm
))
2729 /* this is a per-package concept */
2730 if (!(t
->flags
& CPU_IS_FIRST_THREAD_IN_CORE
) || !(t
->flags
& CPU_IS_FIRST_CORE_IN_PACKAGE
))
2734 if (cpu_migrate(cpu
)) {
2735 fprintf(stderr
, "Could not migrate to CPU %d\n", cpu
);
2739 if (tcc_activation_temp_override
!= 0) {
2740 tcc_activation_temp
= tcc_activation_temp_override
;
2741 fprintf(stderr
, "cpu%d: Using cmdline TCC Target (%d C)\n",
2742 cpu
, tcc_activation_temp
);
2746 /* Temperature Target MSR is Nehalem and newer only */
2747 if (!do_nhm_platform_info
)
2750 if (get_msr(base_cpu
, MSR_IA32_TEMPERATURE_TARGET
, &msr
))
2753 target_c_local
= (msr
>> 16) & 0xFF;
2756 fprintf(stderr
, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2757 cpu
, msr
, target_c_local
);
2759 if (!target_c_local
)
2762 tcc_activation_temp
= target_c_local
;
2767 tcc_activation_temp
= TJMAX_DEFAULT
;
2768 fprintf(stderr
, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2769 cpu
, tcc_activation_temp
);
2774 void decode_misc_enable_msr(void)
2776 unsigned long long msr
;
2778 if (!get_msr(base_cpu
, MSR_IA32_MISC_ENABLE
, &msr
))
2779 fprintf(stderr
, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%s %s %s)\n",
2781 msr
& (1 << 3) ? "TCC" : "",
2782 msr
& (1 << 16) ? "EIST" : "",
2783 msr
& (1 << 18) ? "MONITOR" : "");
2787 * Decode MSR_MISC_PWR_MGMT
2789 * Decode the bits according to the Nehalem documentation
2790 * bit[0] seems to continue to have same meaning going forward
2793 void decode_misc_pwr_mgmt_msr(void)
2795 unsigned long long msr
;
2797 if (!do_nhm_platform_info
)
2800 if (!get_msr(base_cpu
, MSR_MISC_PWR_MGMT
, &msr
))
2801 fprintf(stderr
, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB)\n",
2803 msr
& (1 << 0) ? "DIS" : "EN",
2804 msr
& (1 << 1) ? "EN" : "DIS");
2807 void process_cpuid()
2809 unsigned int eax
, ebx
, ecx
, edx
, max_level
, max_extended_level
;
2810 unsigned int fms
, family
, model
, stepping
;
2812 eax
= ebx
= ecx
= edx
= 0;
2814 __get_cpuid(0, &max_level
, &ebx
, &ecx
, &edx
);
2816 if (ebx
== 0x756e6547 && edx
== 0x49656e69 && ecx
== 0x6c65746e)
2820 fprintf(stderr
, "CPUID(0): %.4s%.4s%.4s ",
2821 (char *)&ebx
, (char *)&edx
, (char *)&ecx
);
2823 __get_cpuid(1, &fms
, &ebx
, &ecx
, &edx
);
2824 family
= (fms
>> 8) & 0xf;
2825 model
= (fms
>> 4) & 0xf;
2826 stepping
= fms
& 0xf;
2827 if (family
== 6 || family
== 0xf)
2828 model
+= ((fms
>> 16) & 0xf) << 4;
2831 fprintf(stderr
, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2832 max_level
, family
, model
, stepping
, family
, model
, stepping
);
2833 fprintf(stderr
, "CPUID(1): %s %s %s %s %s %s %s %s\n",
2834 ecx
& (1 << 0) ? "SSE3" : "-",
2835 ecx
& (1 << 3) ? "MONITOR" : "-",
2836 ecx
& (1 << 7) ? "EIST" : "-",
2837 ecx
& (1 << 8) ? "TM2" : "-",
2838 edx
& (1 << 4) ? "TSC" : "-",
2839 edx
& (1 << 5) ? "MSR" : "-",
2840 edx
& (1 << 22) ? "ACPI-TM" : "-",
2841 edx
& (1 << 29) ? "TM" : "-");
2844 if (!(edx
& (1 << 5)))
2845 errx(1, "CPUID: no MSR");
2848 * check max extended function levels of CPUID.
2849 * This is needed to check for invariant TSC.
2850 * This check is valid for both Intel and AMD.
2852 ebx
= ecx
= edx
= 0;
2853 __get_cpuid(0x80000000, &max_extended_level
, &ebx
, &ecx
, &edx
);
2855 if (max_extended_level
>= 0x80000007) {
2858 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2859 * this check is valid for both Intel and AMD
2861 __get_cpuid(0x80000007, &eax
, &ebx
, &ecx
, &edx
);
2862 has_invariant_tsc
= edx
& (1 << 8);
2866 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2867 * this check is valid for both Intel and AMD
2870 __get_cpuid(0x6, &eax
, &ebx
, &ecx
, &edx
);
2871 has_aperf
= ecx
& (1 << 0);
2872 do_dts
= eax
& (1 << 0);
2873 do_ptm
= eax
& (1 << 6);
2874 has_hwp
= eax
& (1 << 7);
2875 has_hwp_notify
= eax
& (1 << 8);
2876 has_hwp_activity_window
= eax
& (1 << 9);
2877 has_hwp_epp
= eax
& (1 << 10);
2878 has_hwp_pkg
= eax
& (1 << 11);
2879 has_epb
= ecx
& (1 << 3);
2882 fprintf(stderr
, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sHWP, "
2883 "%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
2884 has_aperf
? "" : "No-",
2885 do_dts
? "" : "No-",
2886 do_ptm
? "" : "No-",
2887 has_hwp
? "" : "No-",
2888 has_hwp_notify
? "" : "No-",
2889 has_hwp_activity_window
? "" : "No-",
2890 has_hwp_epp
? "" : "No-",
2891 has_hwp_pkg
? "" : "No-",
2892 has_epb
? "" : "No-");
2895 decode_misc_enable_msr();
2897 if (max_level
>= 0x15) {
2898 unsigned int eax_crystal
;
2899 unsigned int ebx_tsc
;
2902 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
2904 eax_crystal
= ebx_tsc
= crystal_hz
= edx
= 0;
2905 __get_cpuid(0x15, &eax_crystal
, &ebx_tsc
, &crystal_hz
, &edx
);
2909 if (debug
&& (ebx
!= 0))
2910 fprintf(stderr
, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
2911 eax_crystal
, ebx_tsc
, crystal_hz
);
2913 if (crystal_hz
== 0)
2915 case 0x4E: /* SKL */
2916 case 0x5E: /* SKL */
2917 crystal_hz
= 24000000; /* 24 MHz */
2924 tsc_hz
= (unsigned long long) crystal_hz
* ebx_tsc
/ eax_crystal
;
2926 fprintf(stderr
, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
2927 tsc_hz
/ 1000000, crystal_hz
, ebx_tsc
, eax_crystal
);
2931 if (max_level
>= 0x16) {
2932 unsigned int base_mhz
, max_mhz
, bus_mhz
, edx
;
2935 * CPUID 16H Base MHz, Max MHz, Bus MHz
2937 base_mhz
= max_mhz
= bus_mhz
= edx
= 0;
2939 __get_cpuid(0x16, &base_mhz
, &max_mhz
, &bus_mhz
, &edx
);
2941 fprintf(stderr
, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
2942 base_mhz
, max_mhz
, bus_mhz
);
2946 aperf_mperf_multiplier
= get_aperf_mperf_multiplier(family
, model
);
2948 do_nhm_platform_info
= do_nhm_cstates
= do_smi
= probe_nhm_msrs(family
, model
);
2949 do_snb_cstates
= has_snb_msrs(family
, model
);
2950 do_pc2
= do_snb_cstates
&& (pkg_cstate_limit
>= PCL__2
);
2951 do_pc3
= (pkg_cstate_limit
>= PCL__3
);
2952 do_pc6
= (pkg_cstate_limit
>= PCL__6
);
2953 do_pc7
= do_snb_cstates
&& (pkg_cstate_limit
>= PCL__7
);
2954 do_c8_c9_c10
= has_hsw_msrs(family
, model
);
2955 do_skl_residency
= has_skl_msrs(family
, model
);
2956 do_slm_cstates
= is_slm(family
, model
);
2957 do_knl_cstates
= is_knl(family
, model
);
2960 decode_misc_pwr_mgmt_msr();
2962 rapl_probe(family
, model
);
2963 perf_limit_reasons_probe(family
, model
);
2966 dump_cstate_pstate_config_info();
2968 if (has_skl_msrs(family
, model
))
2969 calculate_tsc_tweak();
2977 "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2979 "Turbostat forks the specified COMMAND and prints statistics\n"
2980 "when COMMAND completes.\n"
2981 "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2982 "to print statistics, until interrupted.\n"
2983 "--debug run in \"debug\" mode\n"
2984 "--interval sec Override default 5-second measurement interval\n"
2985 "--help print this help message\n"
2986 "--counter msr print 32-bit counter at address \"msr\"\n"
2987 "--Counter msr print 64-bit Counter at address \"msr\"\n"
2988 "--msr msr print 32-bit value at address \"msr\"\n"
2989 "--MSR msr print 64-bit Value at address \"msr\"\n"
2990 "--version print version information\n"
2992 "For more help, run \"man turbostat\"\n");
2997 * in /dev/cpu/ return success for names that are numbers
2998 * ie. filter out ".", "..", "microcode".
3000 int dir_filter(const struct dirent
*dirp
)
3002 if (isdigit(dirp
->d_name
[0]))
3008 int open_dev_cpu_msr(int dummy1
)
3013 void topology_probe()
3016 int max_core_id
= 0;
3017 int max_package_id
= 0;
3018 int max_siblings
= 0;
3019 struct cpu_topology
{
3021 int physical_package_id
;
3024 /* Initialize num_cpus, max_cpu_num */
3026 topo
.max_cpu_num
= 0;
3027 for_all_proc_cpus(count_cpus
);
3028 if (!summary_only
&& topo
.num_cpus
> 1)
3032 fprintf(stderr
, "num_cpus %d max_cpu_num %d\n", topo
.num_cpus
, topo
.max_cpu_num
);
3034 cpus
= calloc(1, (topo
.max_cpu_num
+ 1) * sizeof(struct cpu_topology
));
3036 err(1, "calloc cpus");
3039 * Allocate and initialize cpu_present_set
3041 cpu_present_set
= CPU_ALLOC((topo
.max_cpu_num
+ 1));
3042 if (cpu_present_set
== NULL
)
3043 err(3, "CPU_ALLOC");
3044 cpu_present_setsize
= CPU_ALLOC_SIZE((topo
.max_cpu_num
+ 1));
3045 CPU_ZERO_S(cpu_present_setsize
, cpu_present_set
);
3046 for_all_proc_cpus(mark_cpu_present
);
3049 * Allocate and initialize cpu_affinity_set
3051 cpu_affinity_set
= CPU_ALLOC((topo
.max_cpu_num
+ 1));
3052 if (cpu_affinity_set
== NULL
)
3053 err(3, "CPU_ALLOC");
3054 cpu_affinity_setsize
= CPU_ALLOC_SIZE((topo
.max_cpu_num
+ 1));
3055 CPU_ZERO_S(cpu_affinity_setsize
, cpu_affinity_set
);
3060 * find max_core_id, max_package_id
3062 for (i
= 0; i
<= topo
.max_cpu_num
; ++i
) {
3065 if (cpu_is_not_present(i
)) {
3067 fprintf(stderr
, "cpu%d NOT PRESENT\n", i
);
3070 cpus
[i
].core_id
= get_core_id(i
);
3071 if (cpus
[i
].core_id
> max_core_id
)
3072 max_core_id
= cpus
[i
].core_id
;
3074 cpus
[i
].physical_package_id
= get_physical_package_id(i
);
3075 if (cpus
[i
].physical_package_id
> max_package_id
)
3076 max_package_id
= cpus
[i
].physical_package_id
;
3078 siblings
= get_num_ht_siblings(i
);
3079 if (siblings
> max_siblings
)
3080 max_siblings
= siblings
;
3082 fprintf(stderr
, "cpu %d pkg %d core %d\n",
3083 i
, cpus
[i
].physical_package_id
, cpus
[i
].core_id
);
3085 topo
.num_cores_per_pkg
= max_core_id
+ 1;
3087 fprintf(stderr
, "max_core_id %d, sizing for %d cores per package\n",
3088 max_core_id
, topo
.num_cores_per_pkg
);
3089 if (debug
&& !summary_only
&& topo
.num_cores_per_pkg
> 1)
3092 topo
.num_packages
= max_package_id
+ 1;
3094 fprintf(stderr
, "max_package_id %d, sizing for %d packages\n",
3095 max_package_id
, topo
.num_packages
);
3096 if (debug
&& !summary_only
&& topo
.num_packages
> 1)
3099 topo
.num_threads_per_core
= max_siblings
;
3101 fprintf(stderr
, "max_siblings %d\n", max_siblings
);
3107 allocate_counters(struct thread_data
**t
, struct core_data
**c
, struct pkg_data
**p
)
3111 *t
= calloc(topo
.num_threads_per_core
* topo
.num_cores_per_pkg
*
3112 topo
.num_packages
, sizeof(struct thread_data
));
3116 for (i
= 0; i
< topo
.num_threads_per_core
*
3117 topo
.num_cores_per_pkg
* topo
.num_packages
; i
++)
3118 (*t
)[i
].cpu_id
= -1;
3120 *c
= calloc(topo
.num_cores_per_pkg
* topo
.num_packages
,
3121 sizeof(struct core_data
));
3125 for (i
= 0; i
< topo
.num_cores_per_pkg
* topo
.num_packages
; i
++)
3126 (*c
)[i
].core_id
= -1;
3128 *p
= calloc(topo
.num_packages
, sizeof(struct pkg_data
));
3132 for (i
= 0; i
< topo
.num_packages
; i
++)
3133 (*p
)[i
].package_id
= i
;
3137 err(1, "calloc counters");
3142 * set cpu_id, core_num, pkg_num
3143 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
3145 * increment topo.num_cores when 1st core in pkg seen
3147 void init_counter(struct thread_data
*thread_base
, struct core_data
*core_base
,
3148 struct pkg_data
*pkg_base
, int thread_num
, int core_num
,
3149 int pkg_num
, int cpu_id
)
3151 struct thread_data
*t
;
3152 struct core_data
*c
;
3155 t
= GET_THREAD(thread_base
, thread_num
, core_num
, pkg_num
);
3156 c
= GET_CORE(core_base
, core_num
, pkg_num
);
3157 p
= GET_PKG(pkg_base
, pkg_num
);
3160 if (thread_num
== 0) {
3161 t
->flags
|= CPU_IS_FIRST_THREAD_IN_CORE
;
3162 if (cpu_is_first_core_in_package(cpu_id
))
3163 t
->flags
|= CPU_IS_FIRST_CORE_IN_PACKAGE
;
3166 c
->core_id
= core_num
;
3167 p
->package_id
= pkg_num
;
3171 int initialize_counters(int cpu_id
)
3173 int my_thread_id
, my_core_id
, my_package_id
;
3175 my_package_id
= get_physical_package_id(cpu_id
);
3176 my_core_id
= get_core_id(cpu_id
);
3177 my_thread_id
= get_cpu_position_in_core(cpu_id
);
3181 init_counter(EVEN_COUNTERS
, my_thread_id
, my_core_id
, my_package_id
, cpu_id
);
3182 init_counter(ODD_COUNTERS
, my_thread_id
, my_core_id
, my_package_id
, cpu_id
);
3186 void allocate_output_buffer()
3188 output_buffer
= calloc(1, (1 + topo
.num_cpus
) * 1024);
3189 outp
= output_buffer
;
3191 err(-1, "calloc output buffer");
3194 void setup_all_buffers(void)
3197 allocate_counters(&thread_even
, &core_even
, &package_even
);
3198 allocate_counters(&thread_odd
, &core_odd
, &package_odd
);
3199 allocate_output_buffer();
3200 for_all_proc_cpus(initialize_counters
);
3203 void set_base_cpu(void)
3205 base_cpu
= sched_getcpu();
3207 err(-ENODEV
, "No valid cpus found");
3210 fprintf(stderr
, "base_cpu = %d\n", base_cpu
);
3213 void turbostat_init()
3215 setup_all_buffers();
3218 check_permissions();
3223 for_all_cpus(print_hwp
, ODD_COUNTERS
);
3226 for_all_cpus(print_epb
, ODD_COUNTERS
);
3229 for_all_cpus(print_perf_limit
, ODD_COUNTERS
);
3232 for_all_cpus(print_rapl
, ODD_COUNTERS
);
3234 for_all_cpus(set_temperature_target
, ODD_COUNTERS
);
3237 for_all_cpus(print_thermal
, ODD_COUNTERS
);
3240 int fork_it(char **argv
)
3245 status
= for_all_cpus(get_counters
, EVEN_COUNTERS
);
3248 /* clear affinity side-effect of get_counters() */
3249 sched_setaffinity(0, cpu_present_setsize
, cpu_present_set
);
3250 gettimeofday(&tv_even
, (struct timezone
*)NULL
);
3255 execvp(argv
[0], argv
);
3259 if (child_pid
== -1)
3262 signal(SIGINT
, SIG_IGN
);
3263 signal(SIGQUIT
, SIG_IGN
);
3264 if (waitpid(child_pid
, &status
, 0) == -1)
3265 err(status
, "waitpid");
3268 * n.b. fork_it() does not check for errors from for_all_cpus()
3269 * because re-starting is problematic when forking
3271 for_all_cpus(get_counters
, ODD_COUNTERS
);
3272 gettimeofday(&tv_odd
, (struct timezone
*)NULL
);
3273 timersub(&tv_odd
, &tv_even
, &tv_delta
);
3274 for_all_cpus_2(delta_cpu
, ODD_COUNTERS
, EVEN_COUNTERS
);
3275 compute_average(EVEN_COUNTERS
);
3276 format_all_counters(EVEN_COUNTERS
);
3279 fprintf(stderr
, "%.6f sec\n", tv_delta
.tv_sec
+ tv_delta
.tv_usec
/1000000.0);
3284 int get_and_dump_counters(void)
3288 status
= for_all_cpus(get_counters
, ODD_COUNTERS
);
3292 status
= for_all_cpus(dump_counters
, ODD_COUNTERS
);
3301 void print_version() {
3302 fprintf(stderr
, "turbostat version 4.10 10 Dec, 2015"
3303 " - Len Brown <lenb@kernel.org>\n");
3306 void cmdline(int argc
, char **argv
)
3309 int option_index
= 0;
3310 static struct option long_options
[] = {
3311 {"Counter", required_argument
, 0, 'C'},
3312 {"counter", required_argument
, 0, 'c'},
3313 {"Dump", no_argument
, 0, 'D'},
3314 {"debug", no_argument
, 0, 'd'},
3315 {"interval", required_argument
, 0, 'i'},
3316 {"help", no_argument
, 0, 'h'},
3317 {"Joules", no_argument
, 0, 'J'},
3318 {"MSR", required_argument
, 0, 'M'},
3319 {"msr", required_argument
, 0, 'm'},
3320 {"Package", no_argument
, 0, 'p'},
3321 {"processor", no_argument
, 0, 'p'},
3322 {"Summary", no_argument
, 0, 'S'},
3323 {"TCC", required_argument
, 0, 'T'},
3324 {"version", no_argument
, 0, 'v' },
3330 while ((opt
= getopt_long_only(argc
, argv
, "+C:c:Ddhi:JM:m:PpST:v",
3331 long_options
, &option_index
)) != -1) {
3334 sscanf(optarg
, "%x", &extra_delta_offset64
);
3337 sscanf(optarg
, "%x", &extra_delta_offset32
);
3351 double interval
= strtod(optarg
, NULL
);
3353 if (interval
< 0.001) {
3354 fprintf(stderr
, "interval %f seconds is too small\n",
3359 interval_ts
.tv_sec
= interval
;
3360 interval_ts
.tv_nsec
= (interval
- interval_ts
.tv_sec
) * 1000000000;
3367 sscanf(optarg
, "%x", &extra_msr_offset64
);
3370 sscanf(optarg
, "%x", &extra_msr_offset32
);
3382 tcc_activation_temp_override
= atoi(optarg
);
3392 int main(int argc
, char **argv
)
3394 cmdline(argc
, argv
);
3401 /* dump counters and exit */
3403 return get_and_dump_counters();
3406 * if any params left, it must be a command to fork
3409 return fork_it(argv
+ optind
);