2 * turbostat -- Log CPU frequency and C-state residency
3 * on modern Intel turbo-capable processors for collectd.
5 * Based on the 'turbostat' tool of the Linux kernel, found at
6 * linux/tools/power/x86/turbostat/turbostat.c:
8 * Copyright (c) 2013 Intel Corporation.
9 * Len Brown <len.brown@intel.com>
11 * This program is free software; you can redistribute it and/or modify it
12 * under the terms and conditions of the GNU General Public License,
13 * version 2, as published by the Free Software Foundation.
15 * This program is distributed in the hope it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
20 * You should have received a copy of the GNU General Public License along with
21 * this program; if not, write to the Free Software Foundation, Inc.,
22 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
24 * Ported to collectd by Vincent Brillault <git@lerya.net>
28 #include <asm/msr-index.h>
34 #include <sys/types.h>
37 #include <sys/resource.h>
52 #define PLUGIN_NAME "turbostat"
54 static const char *proc_stat = "/proc/stat";
55 static unsigned int interval_sec = 5; /* set with -i interval_sec */
56 static unsigned int skip_c0;
57 static unsigned int skip_c1;
58 static unsigned int do_nhm_cstates;
59 static unsigned int do_snb_cstates;
60 static unsigned int do_c8_c9_c10;
61 static unsigned int do_slm_cstates;
62 static unsigned int has_aperf;
63 static unsigned int has_epb;
64 static unsigned int units = 1000000000; /* Ghz etc */
65 static unsigned int genuine_intel;
66 static unsigned int has_invariant_tsc;
67 static unsigned int do_nehalem_platform_info;
69 static unsigned int show_pkg;
70 static unsigned int show_core;
71 static unsigned int show_cpu;
72 static unsigned int do_rapl;
73 static unsigned int do_dts;
74 static unsigned int do_ptm;
75 static unsigned int tcc_activation_temp;
76 static unsigned int tcc_activation_temp_override;
77 static double rapl_power_units, rapl_energy_units, rapl_time_units;
78 static double rapl_joule_counter_range;
80 #define RAPL_PKG (1 << 0)
81 /* 0x610 MSR_PKG_POWER_LIMIT */
82 /* 0x611 MSR_PKG_ENERGY_STATUS */
83 #define RAPL_PKG_PERF_STATUS (1 << 1)
84 /* 0x613 MSR_PKG_PERF_STATUS */
85 #define RAPL_PKG_POWER_INFO (1 << 2)
86 /* 0x614 MSR_PKG_POWER_INFO */
88 #define RAPL_DRAM (1 << 3)
89 /* 0x618 MSR_DRAM_POWER_LIMIT */
90 /* 0x619 MSR_DRAM_ENERGY_STATUS */
91 /* 0x61c MSR_DRAM_POWER_INFO */
92 #define RAPL_DRAM_PERF_STATUS (1 << 4)
93 /* 0x61b MSR_DRAM_PERF_STATUS */
95 #define RAPL_CORES (1 << 5)
96 /* 0x638 MSR_PP0_POWER_LIMIT */
97 /* 0x639 MSR_PP0_ENERGY_STATUS */
98 #define RAPL_CORE_POLICY (1 << 6)
99 /* 0x63a MSR_PP0_POLICY */
102 #define RAPL_GFX (1 << 7)
103 /* 0x640 MSR_PP1_POWER_LIMIT */
104 /* 0x641 MSR_PP1_ENERGY_STATUS */
105 /* 0x642 MSR_PP1_POLICY */
106 #define TJMAX_DEFAULT 100
108 int aperf_mperf_unstable;
112 cpu_set_t *cpu_present_set, *cpu_affinity_set;
113 size_t cpu_present_setsize, cpu_affinity_setsize;
116 unsigned long long tsc;
117 unsigned long long aperf;
118 unsigned long long mperf;
119 unsigned long long c1;
120 unsigned int smi_count;
123 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
124 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
125 } *thread_even, *thread_odd;
128 unsigned long long c3;
129 unsigned long long c6;
130 unsigned long long c7;
131 unsigned int core_temp_c;
132 unsigned int core_id;
133 } *core_even, *core_odd;
136 unsigned long long pc2;
137 unsigned long long pc3;
138 unsigned long long pc6;
139 unsigned long long pc7;
140 unsigned long long pc8;
141 unsigned long long pc9;
142 unsigned long long pc10;
143 unsigned int package_id;
144 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
145 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
146 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
147 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
148 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
149 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
150 unsigned int pkg_temp_c;
152 } *package_even, *package_odd;
154 #define ODD_COUNTERS thread_odd, core_odd, package_odd
155 #define EVEN_COUNTERS thread_even, core_even, package_even
156 static bool is_even = true;
158 static bool allocated = false;
159 static bool initialized = false;
161 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
162 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
163 topo.num_threads_per_core + \
164 (core_no) * topo.num_threads_per_core + (thread_no))
165 #define GET_CORE(core_base, core_no, pkg_no) \
166 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
167 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
174 int num_cores_per_pkg;
175 int num_threads_per_core;
178 struct timeval tv_even, tv_odd, tv_delta;
186 ERR_MSR_CORE_C3_RESIDENCY,
187 ERR_MSR_CORE_C6_RESIDENCY,
188 ERR_MSR_CORE_C7_RESIDENCY,
189 ERR_MSR_IA32_THERM_STATUS,
190 ERR_MSR_PKG_C3_RESIDENCY,
191 ERR_MSR_PKG_C6_RESIDENCY,
192 ERR_MSR_PKG_C2_RESIDENCY,
193 ERR_MSR_PKG_C7_RESIDENCY,
194 ERR_MSR_PKG_C8_RESIDENCY,
195 ERR_MSR_PKG_C9_RESIDENCY,
196 ERR_MSR_PKG_C10_RESIDENCY,
197 ERR_MSR_PKG_ENERGY_STATUS,
198 ERR_MSR_PP0_ENERGY_STATUS,
199 ERR_MSR_DRAM_ENERGY_STATUS,
200 ERR_MSR_PP1_ENERGY_STATUS,
201 ERR_MSR_PKG_PERF_STATUS,
202 ERR_MSR_DRAM_PERF_STATUS,
203 ERR_MSR_IA32_PACKAGE_THERM_STATUS,
207 ERR_CANT_READ_NUMBER,
208 ERR_CANT_READ_PROC_STAT,
209 ERR_NO_INVARIANT_TSC,
216 #define __must_check __attribute__((warn_unused_result))
218 static int setup_all_buffers(void);
221 cpu_is_not_present(int cpu)
223 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
226 * run func(thread, core, package) in topology order
227 * skip non-present cpus
230 static int __must_check
231 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
232 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
234 int retval, pkg_no, core_no, thread_no;
236 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
237 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
238 for (thread_no = 0; thread_no <
239 topo.num_threads_per_core; ++thread_no) {
240 struct thread_data *t;
244 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
246 if (cpu_is_not_present(t->cpu_id))
249 c = GET_CORE(core_base, core_no, pkg_no);
250 p = GET_PKG(pkg_base, pkg_no);
252 retval = func(t, c, p);
261 static int __must_check
264 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
265 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
266 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
267 return -ERR_CPU_MIGRATE;
272 static int __must_check
273 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);
284 retval = pread(fd, msr, sizeof *msr, offset);
287 if (retval != sizeof *msr) {
288 ERROR ("%s offset 0x%llx read failed\n", pathname, (unsigned long long)offset);
295 #define DELTA_WRAP32(new, old) \
299 old = 0x100000000 + new - old; \
303 delta_package(struct pkg_data *new, struct pkg_data *old)
305 old->pc2 = new->pc2 - old->pc2;
306 old->pc3 = new->pc3 - old->pc3;
307 old->pc6 = new->pc6 - old->pc6;
308 old->pc7 = new->pc7 - old->pc7;
309 old->pc8 = new->pc8 - old->pc8;
310 old->pc9 = new->pc9 - old->pc9;
311 old->pc10 = new->pc10 - old->pc10;
312 old->pkg_temp_c = new->pkg_temp_c;
314 DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
315 DELTA_WRAP32(new->energy_cores, old->energy_cores);
316 DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
317 DELTA_WRAP32(new->energy_dram, old->energy_dram);
318 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
319 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
323 delta_core(struct core_data *new, struct core_data *old)
325 old->c3 = new->c3 - old->c3;
326 old->c6 = new->c6 - old->c6;
327 old->c7 = new->c7 - old->c7;
328 old->core_temp_c = new->core_temp_c;
334 static int __must_check
335 delta_thread(struct thread_data *new, struct thread_data *old,
336 struct core_data *core_delta)
338 old->tsc = new->tsc - old->tsc;
340 /* check for TSC < 1 Mcycles over interval */
341 if (old->tsc < (1000 * 1000)) {
342 WARNING("Insanely slow TSC rate, TSC stops in idle?\n"
343 "You can disable all c-states by booting with \"idle=poll\"\n"
344 "or just the deep ones with \"processor.max_cstate=1\"");
348 old->c1 = new->c1 - old->c1;
350 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
351 old->aperf = new->aperf - old->aperf;
352 old->mperf = new->mperf - old->mperf;
355 if (!aperf_mperf_unstable) {
356 WARNING("%s: APERF or MPERF went backwards *\n", progname);
357 WARNING("* Frequency results do not cover entire interval *\n");
358 WARNING("* fix this by running Linux-2.6.30 or later *\n");
360 aperf_mperf_unstable = 1;
363 * mperf delta is likely a huge "positive" number
364 * can not use it for calculating c0 time
372 * As counter collection is not atomic,
373 * it is possible for mperf's non-halted cycles + idle states
374 * to exceed TSC's all cycles: show c1 = 0% in that case.
376 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
379 /* normal case, derive c1 */
380 old->c1 = old->tsc - old->mperf - core_delta->c3
381 - core_delta->c6 - core_delta->c7;
384 if (old->mperf == 0) {
385 WARNING("cpu%d MPERF 0!\n", old->cpu_id);
386 old->mperf = 1; /* divide by 0 protection */
390 old->smi_count = new->smi_count - old->smi_count;
395 static int __must_check
396 delta_cpu(struct thread_data *t, struct core_data *c,
397 struct pkg_data *p, struct thread_data *t2,
398 struct core_data *c2, struct pkg_data *p2)
402 /* calculate core delta only for 1st thread in core */
403 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
406 /* always calculate thread delta */
407 ret = delta_thread(t, t2, c2); /* c2 is core delta */
411 /* calculate package delta only for 1st core in package */
412 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
413 delta_package(p, p2);
418 static unsigned long long
421 unsigned int low, high;
423 asm volatile("rdtsc" : "=a" (low), "=d" (high));
425 return low | ((unsigned long long)high) << 32;
432 * acquire and record local counters for that cpu
434 static int __must_check
435 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
438 unsigned long long msr;
440 if (cpu_migrate(cpu)) {
441 WARNING("Could not migrate to CPU %d\n", cpu);
442 return -ERR_CPU_MIGRATE;
445 t->tsc = rdtsc(); /* we are running on local CPU of interest */
448 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
449 return -ERR_MSR_IA32_APERF;
450 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
451 return -ERR_MSR_IA32_MPERF;
455 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
456 return -ERR_MSR_SMI_COUNT;
457 t->smi_count = msr & 0xFFFFFFFF;
460 /* collect core counters only for 1st thread in core */
461 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
464 if (do_nhm_cstates && !do_slm_cstates) {
465 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
466 return -ERR_MSR_CORE_C3_RESIDENCY;
469 if (do_nhm_cstates) {
470 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
471 return -ERR_MSR_CORE_C6_RESIDENCY;
475 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
476 return -ERR_MSR_CORE_C7_RESIDENCY;
479 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
480 return -ERR_MSR_IA32_THERM_STATUS;
481 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
485 /* collect package counters only for 1st core in package */
486 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
489 if (do_nhm_cstates && !do_slm_cstates) {
490 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
491 return -ERR_MSR_PKG_C3_RESIDENCY;
492 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
493 return -ERR_MSR_PKG_C6_RESIDENCY;
495 if (do_snb_cstates) {
496 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
497 return -ERR_MSR_PKG_C2_RESIDENCY;
498 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
499 return -ERR_MSR_PKG_C7_RESIDENCY;
502 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
503 return -ERR_MSR_PKG_C8_RESIDENCY;
504 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
505 return -ERR_MSR_PKG_C9_RESIDENCY;
506 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
507 return -ERR_MSR_PKG_C10_RESIDENCY;
509 if (do_rapl & RAPL_PKG) {
510 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
511 return -ERR_MSR_PKG_ENERGY_STATUS;
512 p->energy_pkg = msr & 0xFFFFFFFF;
514 if (do_rapl & RAPL_CORES) {
515 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
516 return MSR_PP0_ENERGY_STATUS;
517 p->energy_cores = msr & 0xFFFFFFFF;
519 if (do_rapl & RAPL_DRAM) {
520 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
521 return -ERR_MSR_DRAM_ENERGY_STATUS;
522 p->energy_dram = msr & 0xFFFFFFFF;
524 if (do_rapl & RAPL_GFX) {
525 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
526 return -ERR_MSR_PP1_ENERGY_STATUS;
527 p->energy_gfx = msr & 0xFFFFFFFF;
529 if (do_rapl & RAPL_PKG_PERF_STATUS) {
530 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
531 return -ERR_MSR_PKG_PERF_STATUS;
532 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
534 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
535 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
536 return -ERR_MSR_DRAM_PERF_STATUS;
537 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
540 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
541 return -ERR_MSR_IA32_PACKAGE_THERM_STATUS;
542 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
548 free_all_buffers(void)
553 CPU_FREE(cpu_present_set);
554 cpu_present_set = NULL;
557 CPU_FREE(cpu_affinity_set);
558 cpu_affinity_set = NULL;
559 cpu_affinity_setsize = 0;
579 * Parse a file containing a single int.
582 parse_int_file(const char *fmt, ...)
590 vsnprintf(path, sizeof(path), fmt, args);
592 filep = fopen(path, "r");
594 ERROR("%s: open failed", path);
595 return -ERR_CANT_OPEN_FILE;
597 if (fscanf(filep, "%d", &value) != 1) {
598 ERROR("%s: failed to parse number from file", path);
599 return -ERR_CANT_READ_NUMBER;
606 * cpu_is_first_sibling_in_core(cpu)
607 * return 1 if given CPU is 1st HT sibling in the core
610 cpu_is_first_sibling_in_core(int cpu)
612 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
616 * cpu_is_first_core_in_package(cpu)
617 * return 1 if given CPU is 1st core in package
620 cpu_is_first_core_in_package(int cpu)
622 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
626 get_physical_package_id(int cpu)
628 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
634 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
638 get_num_ht_siblings(int cpu)
646 sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
647 filep = fopen(path, "r");
649 ERROR("%s: open failed", path);
650 return -ERR_CANT_OPEN_FILE;
654 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
655 * otherwinse 1 sibling (self).
657 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
668 * run func(thread, core, package) in topology order
669 * skip non-present cpus
673 static int __must_check
674 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
675 struct pkg_data *, struct thread_data *, struct core_data *,
676 struct pkg_data *), struct thread_data *thread_base,
677 struct core_data *core_base, struct pkg_data *pkg_base,
678 struct thread_data *thread_base2, struct core_data *core_base2,
679 struct pkg_data *pkg_base2)
681 int retval, pkg_no, core_no, thread_no;
683 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
684 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
685 for (thread_no = 0; thread_no <
686 topo.num_threads_per_core; ++thread_no) {
687 struct thread_data *t, *t2;
688 struct core_data *c, *c2;
689 struct pkg_data *p, *p2;
691 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
693 if (cpu_is_not_present(t->cpu_id))
696 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
698 c = GET_CORE(core_base, core_no, pkg_no);
699 c2 = GET_CORE(core_base2, core_no, pkg_no);
701 p = GET_PKG(pkg_base, pkg_no);
702 p2 = GET_PKG(pkg_base2, pkg_no);
704 retval = func(t, c, p, t2, c2, p2);
714 * run func(cpu) on every cpu in /proc/stat
715 * return max_cpu number
717 static int __must_check
718 for_all_proc_cpus(int (func)(int))
724 fp = fopen(proc_stat, "r");
726 ERROR("%s: open failed", proc_stat);
727 return -ERR_CANT_OPEN_FILE;
730 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
732 ERROR("%s: failed to parse format", proc_stat);
733 return -ERR_CANT_READ_PROC_STAT;
737 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
741 retval = func(cpu_num);
753 * remember the last one seen, it will be the max
758 if (topo.max_cpu_num < cpu)
759 topo.max_cpu_num = cpu;
765 mark_cpu_present(int cpu)
767 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
773 turbostat_submit (const char *plugin_instance,
774 const char *type, const char *type_instance,
777 value_list_t vl = VALUE_LIST_INIT;
783 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
784 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
785 if (plugin_instance != NULL)
786 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
787 sstrncpy (vl.type, type, sizeof (vl.type));
788 if (type_instance != NULL)
789 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
791 plugin_dispatch_values (&vl);
795 * column formatting convention & formats
796 * package: "pk" 2 columns %2d
797 * core: "cor" 3 columns %3d
798 * CPU: "CPU" 3 columns %3d
803 * GHz: "GHz" 3 columns %3.2
804 * TSC: "TSC" 3 columns %3.2
805 * SMI: "SMI" 4 columns %4d
806 * percentage " %pc3" %6.2
807 * Perf Status percentage: %5.2
808 * "CTMP" 4 columns %4d
812 submit_counters(struct thread_data *t, struct core_data *c,
816 double interval_float;
818 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
820 snprintf(name, NAME_LEN, "cpu%02d", t->cpu_id);
822 if (do_nhm_cstates) {
824 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
826 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
830 if (has_aperf && ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc))))
831 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
835 turbostat_submit(NULL, "current", name, t->smi_count);
837 /* print per-core data only for 1st thread in core */
838 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
841 snprintf(name, NAME_LEN, "core%02d", c->core_id);
843 if (do_nhm_cstates && !do_slm_cstates)
844 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
846 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
848 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
851 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
853 /* print per-package data only for 1st core in package */
854 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
857 snprintf(name, NAME_LEN, "pkg%02d", p->package_id);
860 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
863 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
864 if (do_nhm_cstates && !do_slm_cstates)
865 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
866 if (do_nhm_cstates && !do_slm_cstates)
867 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
869 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
871 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
872 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
873 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
877 if (do_rapl & RAPL_PKG)
878 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
879 if (do_rapl & RAPL_CORES)
880 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
881 if (do_rapl & RAPL_GFX)
882 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
883 if (do_rapl & RAPL_DRAM)
884 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
891 turbostat_read(user_data_t * not_used)
896 if ((ret = setup_all_buffers()) < 0)
900 if (for_all_proc_cpus(cpu_is_not_present)) {
902 if ((ret = setup_all_buffers()) < 0)
904 if (for_all_proc_cpus(cpu_is_not_present))
905 return -ERR_CPU_NOT_PRESENT;
909 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
911 gettimeofday(&tv_even, (struct timezone *)NULL);
918 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
920 gettimeofday(&tv_odd, (struct timezone *)NULL);
922 timersub(&tv_odd, &tv_even, &tv_delta);
923 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
925 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
928 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
930 gettimeofday(&tv_even, (struct timezone *)NULL);
932 timersub(&tv_even, &tv_odd, &tv_delta);
933 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
935 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
941 static int __must_check
946 if (stat("/dev/cpu/0/msr", &sb)) {
947 ERROR("no /dev/cpu/0/msr\n"
948 "Try \"# modprobe msr\"");
954 static int __must_check
958 ERROR("must be root");
959 return -ERR_NOT_ROOT;
965 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
966 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
969 get_tdp(unsigned int model)
971 unsigned long long msr;
973 if (do_rapl & RAPL_PKG_POWER_INFO)
974 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
975 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
990 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
993 rapl_probe(unsigned int family, unsigned int model)
995 unsigned long long msr;
996 unsigned int time_unit;
1008 case 0x3C: /* HSW */
1009 case 0x45: /* HSW */
1010 case 0x46: /* HSW */
1011 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1013 case 0x3F: /* HSX */
1014 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1018 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1020 case 0x37: /* BYT */
1021 case 0x4D: /* AVN */
1022 do_rapl = RAPL_PKG | RAPL_CORES ;
1028 /* units on package 0, verify later other packages match */
1029 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1032 rapl_power_units = 1.0 / (1 << (msr & 0xF));
1034 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1036 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1038 time_unit = msr >> 16 & 0xF;
1042 rapl_time_units = 1.0 / (1 << (time_unit));
1044 tdp = get_tdp(model);
1046 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1048 // fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1054 is_snb(unsigned int family, unsigned int model)
1062 case 0x3A: /* IVB */
1063 case 0x3E: /* IVB Xeon */
1064 case 0x3C: /* HSW */
1065 case 0x3F: /* HSW */
1066 case 0x45: /* HSW */
1067 case 0x46: /* HSW */
1074 has_c8_c9_c10(unsigned int family, unsigned int model)
1088 is_slm(unsigned int family, unsigned int model)
1093 case 0x37: /* BYT */
1094 case 0x4D: /* AVN */
1101 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1102 * the Thermal Control Circuit (TCC) activates.
1103 * This is usually equal to tjMax.
1105 * Older processors do not have this MSR, so there we guess,
1106 * but also allow cmdline over-ride with -T.
1108 * Several MSR temperature values are in units of degrees-C
1109 * below this value, including the Digital Thermal Sensor (DTS),
1110 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1112 static int __must_check
1113 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1115 unsigned long long msr;
1116 unsigned int target_c_local;
1119 /* tcc_activation_temp is used only for dts or ptm */
1120 if (!(do_dts || do_ptm))
1123 /* this is a per-package concept */
1124 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1128 if (cpu_migrate(cpu)) {
1129 ERROR("Could not migrate to CPU %d\n", cpu);
1130 return -ERR_CPU_MIGRATE;
1133 if (tcc_activation_temp_override != 0) {
1134 tcc_activation_temp = tcc_activation_temp_override;
1135 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
1136 cpu, tcc_activation_temp);
1140 /* Temperature Target MSR is Nehalem and newer only */
1141 if (!do_nehalem_platform_info)
1144 if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
1147 target_c_local = (msr >> 16) & 0x7F;
1149 if (target_c_local < 85 || target_c_local > 127)
1152 tcc_activation_temp = target_c_local;
1157 tcc_activation_temp = TJMAX_DEFAULT;
1158 WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1159 cpu, tcc_activation_temp);
1164 static int __must_check
1167 unsigned int eax, ebx, ecx, edx, max_level;
1168 unsigned int fms, family, model;
1170 eax = ebx = ecx = edx = 0;
1172 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1174 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1178 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1179 family = (fms >> 8) & 0xf;
1180 model = (fms >> 4) & 0xf;
1181 if (family == 6 || family == 0xf)
1182 model += ((fms >> 16) & 0xf) << 4;
1184 if (!(edx & (1 << 5))) {
1185 ERROR("CPUID: no MSR");
1190 * check max extended function levels of CPUID.
1191 * This is needed to check for invariant TSC.
1192 * This check is valid for both Intel and AMD.
1194 ebx = ecx = edx = 0;
1195 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1197 if (max_level < 0x80000007) {
1198 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1199 return -ERR_NO_INVARIANT_TSC;
1203 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1204 * this check is valid for both Intel and AMD
1206 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1207 has_invariant_tsc = edx & (1 << 8);
1209 if (!has_invariant_tsc) {
1210 ERROR("No invariant TSC");
1211 return -ERR_NO_INVARIANT_TSC;
1215 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1216 * this check is valid for both Intel and AMD
1219 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1220 has_aperf = ecx & (1 << 0);
1221 do_dts = eax & (1 << 0);
1222 do_ptm = eax & (1 << 6);
1223 has_epb = ecx & (1 << 3);
1227 return -ERR_NO_APERF;
1230 do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1231 do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */
1232 do_smi = do_nhm_cstates;
1233 do_snb_cstates = is_snb(family, model);
1234 do_c8_c9_c10 = has_c8_c9_c10(family, model);
1235 do_slm_cstates = is_slm(family, model);
1237 rapl_probe(family, model);
1244 static int __must_check
1249 int max_core_id = 0;
1250 int max_package_id = 0;
1251 int max_siblings = 0;
1252 struct cpu_topology {
1254 int physical_package_id;
1257 /* Initialize num_cpus, max_cpu_num */
1259 topo.max_cpu_num = 0;
1260 ret = for_all_proc_cpus(count_cpus);
1263 if (topo.num_cpus > 1)
1266 DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1268 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
1270 ERROR("calloc cpus");
1275 * Allocate and initialize cpu_present_set
1277 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1278 if (cpu_present_set == NULL) {
1281 return -ERR_CPU_ALLOC;
1283 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1284 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1285 ret = for_all_proc_cpus(mark_cpu_present);
1292 * Allocate and initialize cpu_affinity_set
1294 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1295 if (cpu_affinity_set == NULL) {
1298 return -ERR_CPU_ALLOC;
1300 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1301 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1306 * find max_core_id, max_package_id
1308 for (i = 0; i <= topo.max_cpu_num; ++i) {
1311 if (cpu_is_not_present(i)) {
1313 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1316 cpus[i].core_id = get_core_id(i);
1317 if (cpus[i].core_id < 0)
1318 return cpus[i].core_id;
1319 if (cpus[i].core_id > max_core_id)
1320 max_core_id = cpus[i].core_id;
1322 cpus[i].physical_package_id = get_physical_package_id(i);
1323 if (cpus[i].physical_package_id < 0)
1324 return cpus[i].physical_package_id;
1325 if (cpus[i].physical_package_id > max_package_id)
1326 max_package_id = cpus[i].physical_package_id;
1328 siblings = get_num_ht_siblings(i);
1331 if (siblings > max_siblings)
1332 max_siblings = siblings;
1333 DEBUG("cpu %d pkg %d core %d\n",
1334 i, cpus[i].physical_package_id, cpus[i].core_id);
1336 topo.num_cores_per_pkg = max_core_id + 1;
1337 DEBUG("max_core_id %d, sizing for %d cores per package\n",
1338 max_core_id, topo.num_cores_per_pkg);
1339 if (topo.num_cores_per_pkg > 1)
1342 topo.num_packages = max_package_id + 1;
1343 DEBUG("max_package_id %d, sizing for %d packages\n",
1344 max_package_id, topo.num_packages);
1345 if (topo.num_packages > 1)
1348 topo.num_threads_per_core = max_siblings;
1349 DEBUG("max_siblings %d\n", max_siblings);
1356 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1360 *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1361 topo.num_packages, sizeof(struct thread_data));
1365 for (i = 0; i < topo.num_threads_per_core *
1366 topo.num_cores_per_pkg * topo.num_packages; i++)
1367 (*t)[i].cpu_id = -1;
1369 *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1370 sizeof(struct core_data));
1374 for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1375 (*c)[i].core_id = -1;
1377 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1381 for (i = 0; i < topo.num_packages; i++)
1382 (*p)[i].package_id = i;
1386 ERROR("calloc counters");
1392 * set cpu_id, core_num, pkg_num
1393 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1395 * increment topo.num_cores when 1st core in pkg seen
1398 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1399 struct pkg_data *pkg_base, int thread_num, int core_num,
1400 int pkg_num, int cpu_id)
1403 struct thread_data *t;
1404 struct core_data *c;
1407 t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1408 c = GET_CORE(core_base, core_num, pkg_num);
1409 p = GET_PKG(pkg_base, pkg_num);
1412 if (thread_num == 0) {
1413 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1414 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1416 } else if (ret != 0) {
1417 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1421 c->core_id = core_num;
1422 p->package_id = pkg_num;
1429 initialize_counters(int cpu_id)
1431 int my_thread_id, my_core_id, my_package_id;
1434 my_package_id = get_physical_package_id(cpu_id);
1435 if (my_package_id < 0)
1436 return my_package_id;
1437 my_core_id = get_core_id(cpu_id);
1441 if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1443 } else if (ret != 0) {
1450 ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1453 ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1459 #define DO_OR_GOTO_ERR(something) \
1461 ret = (something); \
1466 static int setup_all_buffers(void)
1470 DO_OR_GOTO_ERR(topology_probe());
1471 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1472 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1473 DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1483 turbostat_init(void)
1488 DO_OR_GOTO_ERR(check_cpuid());
1489 DO_OR_GOTO_ERR(check_dev_msr());
1490 DO_OR_GOTO_ERR(check_super_user());
1491 DO_OR_GOTO_ERR(setup_all_buffers());
1492 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1494 ts.tv_sec = interval_sec;
1497 plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, &ts, NULL);
1505 static const char *config_keys[] =
1509 static int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1512 turbostat_config(const char *key, const char *value)
1514 if (strcasecmp("Interval", key) == 0)
1515 interval_sec = atoi(value);
1521 void module_register(void);
1522 void module_register(void)
1524 plugin_register_init(PLUGIN_NAME, turbostat_init);
1525 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);