2 * turbostat -- Log CPU frequency and C-state residency
3 * on modern Intel turbo-capable processors for collectd.
5 * Based on the kernel tools by:
6 * Copyright (c) 2013 Intel Corporation.
7 * Len Brown <len.brown@intel.com>
9 * This program is free software; you can redistribute it and/or modify it
10 * under the terms and conditions of the GNU General Public License,
11 * version 2, as published by the Free Software Foundation.
13 * This program is distributed in the hope it will be useful, but WITHOUT
14 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
18 * You should have received a copy of the GNU General Public License along with
19 * this program; if not, write to the Free Software Foundation, Inc.,
20 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
22 * Ported to collectd by Vincent Brillault <git@lerya.net>
26 #include "msr-index.h"
32 #include <sys/types.h>
35 #include <sys/resource.h>
50 #define PLUGIN_NAME "turbostat"
52 static const char *proc_stat = "/proc/stat";
53 static unsigned int interval_sec = 5; /* set with -i interval_sec */
54 static unsigned int skip_c0;
55 static unsigned int skip_c1;
56 static unsigned int do_nhm_cstates;
57 static unsigned int do_snb_cstates;
58 static unsigned int do_c8_c9_c10;
59 static unsigned int do_slm_cstates;
60 static unsigned int has_aperf;
61 static unsigned int has_epb;
62 static unsigned int units = 1000000000; /* Ghz etc */
63 static unsigned int genuine_intel;
64 static unsigned int has_invariant_tsc;
65 static unsigned int do_nehalem_platform_info;
67 static unsigned int show_pkg;
68 static unsigned int show_core;
69 static unsigned int show_cpu;
70 static unsigned int do_rapl;
71 static unsigned int do_dts;
72 static unsigned int do_ptm;
73 static unsigned int tcc_activation_temp;
74 static unsigned int tcc_activation_temp_override;
75 static double rapl_power_units, rapl_energy_units, rapl_time_units;
76 static double rapl_joule_counter_range;
78 #define RAPL_PKG (1 << 0)
79 /* 0x610 MSR_PKG_POWER_LIMIT */
80 /* 0x611 MSR_PKG_ENERGY_STATUS */
81 #define RAPL_PKG_PERF_STATUS (1 << 1)
82 /* 0x613 MSR_PKG_PERF_STATUS */
83 #define RAPL_PKG_POWER_INFO (1 << 2)
84 /* 0x614 MSR_PKG_POWER_INFO */
86 #define RAPL_DRAM (1 << 3)
87 /* 0x618 MSR_DRAM_POWER_LIMIT */
88 /* 0x619 MSR_DRAM_ENERGY_STATUS */
89 /* 0x61c MSR_DRAM_POWER_INFO */
90 #define RAPL_DRAM_PERF_STATUS (1 << 4)
91 /* 0x61b MSR_DRAM_PERF_STATUS */
93 #define RAPL_CORES (1 << 5)
94 /* 0x638 MSR_PP0_POWER_LIMIT */
95 /* 0x639 MSR_PP0_ENERGY_STATUS */
96 #define RAPL_CORE_POLICY (1 << 6)
97 /* 0x63a MSR_PP0_POLICY */
100 #define RAPL_GFX (1 << 7)
101 /* 0x640 MSR_PP1_POWER_LIMIT */
102 /* 0x641 MSR_PP1_ENERGY_STATUS */
103 /* 0x642 MSR_PP1_POLICY */
104 #define TJMAX_DEFAULT 100
106 int aperf_mperf_unstable;
110 cpu_set_t *cpu_present_set, *cpu_affinity_set;
111 size_t cpu_present_setsize, cpu_affinity_setsize;
114 unsigned long long tsc;
115 unsigned long long aperf;
116 unsigned long long mperf;
117 unsigned long long c1;
118 unsigned int smi_count;
121 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
122 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
123 } *thread_even, *thread_odd;
126 unsigned long long c3;
127 unsigned long long c6;
128 unsigned long long c7;
129 unsigned int core_temp_c;
130 unsigned int core_id;
131 } *core_even, *core_odd;
134 unsigned long long pc2;
135 unsigned long long pc3;
136 unsigned long long pc6;
137 unsigned long long pc7;
138 unsigned long long pc8;
139 unsigned long long pc9;
140 unsigned long long pc10;
141 unsigned int package_id;
142 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
143 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
144 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
145 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
146 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
147 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
148 unsigned int pkg_temp_c;
150 } *package_even, *package_odd;
152 #define ODD_COUNTERS thread_odd, core_odd, package_odd
153 #define EVEN_COUNTERS thread_even, core_even, package_even
154 static bool is_even = true;
156 static bool allocated = false;
157 static bool initialized = false;
159 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
160 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
161 topo.num_threads_per_core + \
162 (core_no) * topo.num_threads_per_core + (thread_no))
163 #define GET_CORE(core_base, core_no, pkg_no) \
164 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
165 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
172 int num_cores_per_pkg;
173 int num_threads_per_core;
176 struct timeval tv_even, tv_odd, tv_delta;
184 ERR_MSR_CORE_C3_RESIDENCY,
185 ERR_MSR_CORE_C6_RESIDENCY,
186 ERR_MSR_CORE_C7_RESIDENCY,
187 ERR_MSR_IA32_THERM_STATUS,
188 ERR_MSR_PKG_C3_RESIDENCY,
189 ERR_MSR_PKG_C6_RESIDENCY,
190 ERR_MSR_PKG_C2_RESIDENCY,
191 ERR_MSR_PKG_C7_RESIDENCY,
192 ERR_MSR_PKG_C8_RESIDENCY,
193 ERR_MSR_PKG_C9_RESIDENCY,
194 ERR_MSR_PKG_C10_RESIDENCY,
195 ERR_MSR_PKG_ENERGY_STATUS,
196 ERR_MSR_PP0_ENERGY_STATUS,
197 ERR_MSR_DRAM_ENERGY_STATUS,
198 ERR_MSR_PP1_ENERGY_STATUS,
199 ERR_MSR_PKG_PERF_STATUS,
200 ERR_MSR_DRAM_PERF_STATUS,
201 ERR_MSR_IA32_PACKAGE_THERM_STATUS,
205 ERR_CANT_READ_NUMBER,
206 ERR_CANT_READ_PROC_STAT,
207 ERR_NO_INVARIANT_TSC,
214 #define __must_check __attribute__((warn_unused_result))
216 static int setup_all_buffers(void);
219 cpu_is_not_present(int cpu)
221 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
224 * run func(thread, core, package) in topology order
225 * skip non-present cpus
228 static int __must_check
229 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
230 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
232 int retval, pkg_no, core_no, thread_no;
234 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
235 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
236 for (thread_no = 0; thread_no <
237 topo.num_threads_per_core; ++thread_no) {
238 struct thread_data *t;
242 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
244 if (cpu_is_not_present(t->cpu_id))
247 c = GET_CORE(core_base, core_no, pkg_no);
248 p = GET_PKG(pkg_base, pkg_no);
250 retval = func(t, c, p);
259 static int __must_check
262 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
263 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
264 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
265 return -ERR_CPU_MIGRATE;
270 static int __must_check
271 get_msr(int cpu, off_t offset, unsigned long long *msr)
277 sprintf(pathname, "/dev/cpu/%d/msr", cpu);
278 fd = open(pathname, O_RDONLY);
282 retval = pread(fd, msr, sizeof *msr, offset);
285 if (retval != sizeof *msr) {
286 ERROR ("%s offset 0x%llx read failed\n", pathname, (unsigned long long)offset);
293 #define DELTA_WRAP32(new, old) \
297 old = 0x100000000 + new - old; \
301 delta_package(struct pkg_data *new, struct pkg_data *old)
303 old->pc2 = new->pc2 - old->pc2;
304 old->pc3 = new->pc3 - old->pc3;
305 old->pc6 = new->pc6 - old->pc6;
306 old->pc7 = new->pc7 - old->pc7;
307 old->pc8 = new->pc8 - old->pc8;
308 old->pc9 = new->pc9 - old->pc9;
309 old->pc10 = new->pc10 - old->pc10;
310 old->pkg_temp_c = new->pkg_temp_c;
312 DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
313 DELTA_WRAP32(new->energy_cores, old->energy_cores);
314 DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
315 DELTA_WRAP32(new->energy_dram, old->energy_dram);
316 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
317 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
321 delta_core(struct core_data *new, struct core_data *old)
323 old->c3 = new->c3 - old->c3;
324 old->c6 = new->c6 - old->c6;
325 old->c7 = new->c7 - old->c7;
326 old->core_temp_c = new->core_temp_c;
332 static int __must_check
333 delta_thread(struct thread_data *new, struct thread_data *old,
334 struct core_data *core_delta)
336 old->tsc = new->tsc - old->tsc;
338 /* check for TSC < 1 Mcycles over interval */
339 if (old->tsc < (1000 * 1000)) {
340 WARNING("Insanely slow TSC rate, TSC stops in idle?\n"
341 "You can disable all c-states by booting with \"idle=poll\"\n"
342 "or just the deep ones with \"processor.max_cstate=1\"");
346 old->c1 = new->c1 - old->c1;
348 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
349 old->aperf = new->aperf - old->aperf;
350 old->mperf = new->mperf - old->mperf;
353 if (!aperf_mperf_unstable) {
354 WARNING("%s: APERF or MPERF went backwards *\n", progname);
355 WARNING("* Frequency results do not cover entire interval *\n");
356 WARNING("* fix this by running Linux-2.6.30 or later *\n");
358 aperf_mperf_unstable = 1;
361 * mperf delta is likely a huge "positive" number
362 * can not use it for calculating c0 time
370 * As counter collection is not atomic,
371 * it is possible for mperf's non-halted cycles + idle states
372 * to exceed TSC's all cycles: show c1 = 0% in that case.
374 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
377 /* normal case, derive c1 */
378 old->c1 = old->tsc - old->mperf - core_delta->c3
379 - core_delta->c6 - core_delta->c7;
382 if (old->mperf == 0) {
383 WARNING("cpu%d MPERF 0!\n", old->cpu_id);
384 old->mperf = 1; /* divide by 0 protection */
388 old->smi_count = new->smi_count - old->smi_count;
393 static int __must_check
394 delta_cpu(struct thread_data *t, struct core_data *c,
395 struct pkg_data *p, struct thread_data *t2,
396 struct core_data *c2, struct pkg_data *p2)
400 /* calculate core delta only for 1st thread in core */
401 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
404 /* always calculate thread delta */
405 ret = delta_thread(t, t2, c2); /* c2 is core delta */
409 /* calculate package delta only for 1st core in package */
410 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
411 delta_package(p, p2);
416 static unsigned long long
419 unsigned int low, high;
421 asm volatile("rdtsc" : "=a" (low), "=d" (high));
423 return low | ((unsigned long long)high) << 32;
430 * acquire and record local counters for that cpu
432 static int __must_check
433 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
436 unsigned long long msr;
438 if (cpu_migrate(cpu)) {
439 WARNING("Could not migrate to CPU %d\n", cpu);
440 return -ERR_CPU_MIGRATE;
443 t->tsc = rdtsc(); /* we are running on local CPU of interest */
446 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
447 return -ERR_MSR_IA32_APERF;
448 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
449 return -ERR_MSR_IA32_MPERF;
453 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
454 return -ERR_MSR_SMI_COUNT;
455 t->smi_count = msr & 0xFFFFFFFF;
458 /* collect core counters only for 1st thread in core */
459 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
462 if (do_nhm_cstates && !do_slm_cstates) {
463 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
464 return -ERR_MSR_CORE_C3_RESIDENCY;
467 if (do_nhm_cstates) {
468 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
469 return -ERR_MSR_CORE_C6_RESIDENCY;
473 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
474 return -ERR_MSR_CORE_C7_RESIDENCY;
477 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
478 return -ERR_MSR_IA32_THERM_STATUS;
479 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
483 /* collect package counters only for 1st core in package */
484 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
487 if (do_nhm_cstates && !do_slm_cstates) {
488 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
489 return -ERR_MSR_PKG_C3_RESIDENCY;
490 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
491 return -ERR_MSR_PKG_C6_RESIDENCY;
493 if (do_snb_cstates) {
494 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
495 return -ERR_MSR_PKG_C2_RESIDENCY;
496 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
497 return -ERR_MSR_PKG_C7_RESIDENCY;
500 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
501 return -ERR_MSR_PKG_C8_RESIDENCY;
502 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
503 return -ERR_MSR_PKG_C9_RESIDENCY;
504 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
505 return -ERR_MSR_PKG_C10_RESIDENCY;
507 if (do_rapl & RAPL_PKG) {
508 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
509 return -ERR_MSR_PKG_ENERGY_STATUS;
510 p->energy_pkg = msr & 0xFFFFFFFF;
512 if (do_rapl & RAPL_CORES) {
513 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
514 return MSR_PP0_ENERGY_STATUS;
515 p->energy_cores = msr & 0xFFFFFFFF;
517 if (do_rapl & RAPL_DRAM) {
518 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
519 return -ERR_MSR_DRAM_ENERGY_STATUS;
520 p->energy_dram = msr & 0xFFFFFFFF;
522 if (do_rapl & RAPL_GFX) {
523 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
524 return -ERR_MSR_PP1_ENERGY_STATUS;
525 p->energy_gfx = msr & 0xFFFFFFFF;
527 if (do_rapl & RAPL_PKG_PERF_STATUS) {
528 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
529 return -ERR_MSR_PKG_PERF_STATUS;
530 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
532 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
533 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
534 return -ERR_MSR_DRAM_PERF_STATUS;
535 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
538 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
539 return -ERR_MSR_IA32_PACKAGE_THERM_STATUS;
540 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
546 free_all_buffers(void)
551 CPU_FREE(cpu_present_set);
552 cpu_present_set = NULL;
555 CPU_FREE(cpu_affinity_set);
556 cpu_affinity_set = NULL;
557 cpu_affinity_setsize = 0;
577 * Parse a file containing a single int.
580 parse_int_file(const char *fmt, ...)
588 vsnprintf(path, sizeof(path), fmt, args);
590 filep = fopen(path, "r");
592 ERROR("%s: open failed", path);
593 return -ERR_CANT_OPEN_FILE;
595 if (fscanf(filep, "%d", &value) != 1) {
596 ERROR("%s: failed to parse number from file", path);
597 return -ERR_CANT_READ_NUMBER;
604 * cpu_is_first_sibling_in_core(cpu)
605 * return 1 if given CPU is 1st HT sibling in the core
608 cpu_is_first_sibling_in_core(int cpu)
610 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
614 * cpu_is_first_core_in_package(cpu)
615 * return 1 if given CPU is 1st core in package
618 cpu_is_first_core_in_package(int cpu)
620 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
624 get_physical_package_id(int cpu)
626 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
632 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
636 get_num_ht_siblings(int cpu)
644 sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
645 filep = fopen(path, "r");
647 ERROR("%s: open failed", path);
648 return -ERR_CANT_OPEN_FILE;
652 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
653 * otherwinse 1 sibling (self).
655 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
666 * run func(thread, core, package) in topology order
667 * skip non-present cpus
671 static int __must_check
672 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
673 struct pkg_data *, struct thread_data *, struct core_data *,
674 struct pkg_data *), struct thread_data *thread_base,
675 struct core_data *core_base, struct pkg_data *pkg_base,
676 struct thread_data *thread_base2, struct core_data *core_base2,
677 struct pkg_data *pkg_base2)
679 int retval, pkg_no, core_no, thread_no;
681 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
682 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
683 for (thread_no = 0; thread_no <
684 topo.num_threads_per_core; ++thread_no) {
685 struct thread_data *t, *t2;
686 struct core_data *c, *c2;
687 struct pkg_data *p, *p2;
689 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
691 if (cpu_is_not_present(t->cpu_id))
694 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
696 c = GET_CORE(core_base, core_no, pkg_no);
697 c2 = GET_CORE(core_base2, core_no, pkg_no);
699 p = GET_PKG(pkg_base, pkg_no);
700 p2 = GET_PKG(pkg_base2, pkg_no);
702 retval = func(t, c, p, t2, c2, p2);
712 * run func(cpu) on every cpu in /proc/stat
713 * return max_cpu number
715 static int __must_check
716 for_all_proc_cpus(int (func)(int))
722 fp = fopen(proc_stat, "r");
724 ERROR("%s: open failed", proc_stat);
725 return -ERR_CANT_OPEN_FILE;
728 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
730 ERROR("%s: failed to parse format", proc_stat);
731 return -ERR_CANT_READ_PROC_STAT;
735 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
739 retval = func(cpu_num);
751 * remember the last one seen, it will be the max
756 if (topo.max_cpu_num < cpu)
757 topo.max_cpu_num = cpu;
763 mark_cpu_present(int cpu)
765 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
771 turbostat_submit (const char *plugin_instance,
772 const char *type, const char *type_instance,
775 value_list_t vl = VALUE_LIST_INIT;
781 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
782 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
783 if (plugin_instance != NULL)
784 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
785 sstrncpy (vl.type, type, sizeof (vl.type));
786 if (type_instance != NULL)
787 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
789 plugin_dispatch_values (&vl);
793 * column formatting convention & formats
794 * package: "pk" 2 columns %2d
795 * core: "cor" 3 columns %3d
796 * CPU: "CPU" 3 columns %3d
801 * GHz: "GHz" 3 columns %3.2
802 * TSC: "TSC" 3 columns %3.2
803 * SMI: "SMI" 4 columns %4d
804 * percentage " %pc3" %6.2
805 * Perf Status percentage: %5.2
806 * "CTMP" 4 columns %4d
810 submit_counters(struct thread_data *t, struct core_data *c,
814 double interval_float;
816 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
818 snprintf(name, NAME_LEN, "cpu%02d", t->cpu_id);
820 if (do_nhm_cstates) {
822 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
824 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
828 if (has_aperf && ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc))))
829 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
833 turbostat_submit(NULL, "current", name, t->smi_count);
835 /* print per-core data only for 1st thread in core */
836 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
839 snprintf(name, NAME_LEN, "core%02d", c->core_id);
841 if (do_nhm_cstates && !do_slm_cstates)
842 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
844 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
846 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
849 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
851 /* print per-package data only for 1st core in package */
852 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
855 snprintf(name, NAME_LEN, "pc%02d", p->package_id);
858 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
861 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
862 if (do_nhm_cstates && !do_slm_cstates)
863 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
864 if (do_nhm_cstates && !do_slm_cstates)
865 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
867 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
869 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
870 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
871 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
875 if (do_rapl & RAPL_PKG)
876 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
877 if (do_rapl & RAPL_CORES)
878 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
879 if (do_rapl & RAPL_GFX)
880 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
881 if (do_rapl & RAPL_DRAM)
882 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
889 turbostat_read(user_data_t * not_used)
894 if ((ret = setup_all_buffers()) < 0)
898 if (for_all_proc_cpus(cpu_is_not_present)) {
900 if ((ret = setup_all_buffers()) < 0)
902 if (for_all_proc_cpus(cpu_is_not_present))
903 return -ERR_CPU_NOT_PRESENT;
907 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
909 gettimeofday(&tv_even, (struct timezone *)NULL);
916 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
918 gettimeofday(&tv_odd, (struct timezone *)NULL);
920 timersub(&tv_odd, &tv_even, &tv_delta);
921 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
923 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
926 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
928 gettimeofday(&tv_even, (struct timezone *)NULL);
930 timersub(&tv_even, &tv_odd, &tv_delta);
931 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
933 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
939 static int __must_check
944 if (stat("/dev/cpu/0/msr", &sb)) {
945 ERROR("no /dev/cpu/0/msr\n"
946 "Try \"# modprobe msr\"");
952 static int __must_check
956 ERROR("must be root");
957 return -ERR_NOT_ROOT;
963 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
964 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
967 get_tdp(unsigned int model)
969 unsigned long long msr;
971 if (do_rapl & RAPL_PKG_POWER_INFO)
972 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
973 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
988 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
991 rapl_probe(unsigned int family, unsigned int model)
993 unsigned long long msr;
994 unsigned int time_unit;
1006 case 0x3C: /* HSW */
1007 case 0x45: /* HSW */
1008 case 0x46: /* HSW */
1009 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1011 case 0x3F: /* HSX */
1012 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1016 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1018 case 0x37: /* BYT */
1019 case 0x4D: /* AVN */
1020 do_rapl = RAPL_PKG | RAPL_CORES ;
1026 /* units on package 0, verify later other packages match */
1027 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1030 rapl_power_units = 1.0 / (1 << (msr & 0xF));
1032 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1034 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1036 time_unit = msr >> 16 & 0xF;
1040 rapl_time_units = 1.0 / (1 << (time_unit));
1042 tdp = get_tdp(model);
1044 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1046 // fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1052 is_snb(unsigned int family, unsigned int model)
1060 case 0x3A: /* IVB */
1061 case 0x3E: /* IVB Xeon */
1062 case 0x3C: /* HSW */
1063 case 0x3F: /* HSW */
1064 case 0x45: /* HSW */
1065 case 0x46: /* HSW */
1072 has_c8_c9_c10(unsigned int family, unsigned int model)
1086 is_slm(unsigned int family, unsigned int model)
1091 case 0x37: /* BYT */
1092 case 0x4D: /* AVN */
1099 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1100 * the Thermal Control Circuit (TCC) activates.
1101 * This is usually equal to tjMax.
1103 * Older processors do not have this MSR, so there we guess,
1104 * but also allow cmdline over-ride with -T.
1106 * Several MSR temperature values are in units of degrees-C
1107 * below this value, including the Digital Thermal Sensor (DTS),
1108 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1110 static int __must_check
1111 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1113 unsigned long long msr;
1114 unsigned int target_c_local;
1117 /* tcc_activation_temp is used only for dts or ptm */
1118 if (!(do_dts || do_ptm))
1121 /* this is a per-package concept */
1122 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1126 if (cpu_migrate(cpu)) {
1127 ERROR("Could not migrate to CPU %d\n", cpu);
1128 return -ERR_CPU_MIGRATE;
1131 if (tcc_activation_temp_override != 0) {
1132 tcc_activation_temp = tcc_activation_temp_override;
1133 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
1134 cpu, tcc_activation_temp);
1138 /* Temperature Target MSR is Nehalem and newer only */
1139 if (!do_nehalem_platform_info)
1142 if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
1145 target_c_local = (msr >> 16) & 0x7F;
1147 if (target_c_local < 85 || target_c_local > 127)
1150 tcc_activation_temp = target_c_local;
1155 tcc_activation_temp = TJMAX_DEFAULT;
1156 WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1157 cpu, tcc_activation_temp);
1162 static int __must_check
1165 unsigned int eax, ebx, ecx, edx, max_level;
1166 unsigned int fms, family, model;
1168 eax = ebx = ecx = edx = 0;
1170 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1172 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1176 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1177 family = (fms >> 8) & 0xf;
1178 model = (fms >> 4) & 0xf;
1179 if (family == 6 || family == 0xf)
1180 model += ((fms >> 16) & 0xf) << 4;
1182 if (!(edx & (1 << 5))) {
1183 ERROR("CPUID: no MSR");
1188 * check max extended function levels of CPUID.
1189 * This is needed to check for invariant TSC.
1190 * This check is valid for both Intel and AMD.
1192 ebx = ecx = edx = 0;
1193 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1195 if (max_level < 0x80000007) {
1196 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1197 return -ERR_NO_INVARIANT_TSC;
1201 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1202 * this check is valid for both Intel and AMD
1204 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1205 has_invariant_tsc = edx & (1 << 8);
1207 if (!has_invariant_tsc) {
1208 ERROR("No invariant TSC");
1209 return -ERR_NO_INVARIANT_TSC;
1213 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1214 * this check is valid for both Intel and AMD
1217 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1218 has_aperf = ecx & (1 << 0);
1219 do_dts = eax & (1 << 0);
1220 do_ptm = eax & (1 << 6);
1221 has_epb = ecx & (1 << 3);
1225 return -ERR_NO_APERF;
1228 do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1229 do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */
1230 do_smi = do_nhm_cstates;
1231 do_snb_cstates = is_snb(family, model);
1232 do_c8_c9_c10 = has_c8_c9_c10(family, model);
1233 do_slm_cstates = is_slm(family, model);
1235 rapl_probe(family, model);
1242 static int __must_check
1247 int max_core_id = 0;
1248 int max_package_id = 0;
1249 int max_siblings = 0;
1250 struct cpu_topology {
1252 int physical_package_id;
1255 /* Initialize num_cpus, max_cpu_num */
1257 topo.max_cpu_num = 0;
1258 ret = for_all_proc_cpus(count_cpus);
1261 if (topo.num_cpus > 1)
1264 DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1266 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
1268 ERROR("calloc cpus");
1273 * Allocate and initialize cpu_present_set
1275 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1276 if (cpu_present_set == NULL) {
1279 return -ERR_CPU_ALLOC;
1281 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1282 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1283 ret = for_all_proc_cpus(mark_cpu_present);
1290 * Allocate and initialize cpu_affinity_set
1292 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1293 if (cpu_affinity_set == NULL) {
1296 return -ERR_CPU_ALLOC;
1298 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1299 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1304 * find max_core_id, max_package_id
1306 for (i = 0; i <= topo.max_cpu_num; ++i) {
1309 if (cpu_is_not_present(i)) {
1311 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1314 cpus[i].core_id = get_core_id(i);
1315 if (cpus[i].core_id < 0)
1316 return cpus[i].core_id;
1317 if (cpus[i].core_id > max_core_id)
1318 max_core_id = cpus[i].core_id;
1320 cpus[i].physical_package_id = get_physical_package_id(i);
1321 if (cpus[i].physical_package_id < 0)
1322 return cpus[i].physical_package_id;
1323 if (cpus[i].physical_package_id > max_package_id)
1324 max_package_id = cpus[i].physical_package_id;
1326 siblings = get_num_ht_siblings(i);
1329 if (siblings > max_siblings)
1330 max_siblings = siblings;
1331 DEBUG("cpu %d pkg %d core %d\n",
1332 i, cpus[i].physical_package_id, cpus[i].core_id);
1334 topo.num_cores_per_pkg = max_core_id + 1;
1335 DEBUG("max_core_id %d, sizing for %d cores per package\n",
1336 max_core_id, topo.num_cores_per_pkg);
1337 if (topo.num_cores_per_pkg > 1)
1340 topo.num_packages = max_package_id + 1;
1341 DEBUG("max_package_id %d, sizing for %d packages\n",
1342 max_package_id, topo.num_packages);
1343 if (topo.num_packages > 1)
1346 topo.num_threads_per_core = max_siblings;
1347 DEBUG("max_siblings %d\n", max_siblings);
1354 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1358 *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1359 topo.num_packages, sizeof(struct thread_data));
1363 for (i = 0; i < topo.num_threads_per_core *
1364 topo.num_cores_per_pkg * topo.num_packages; i++)
1365 (*t)[i].cpu_id = -1;
1367 *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1368 sizeof(struct core_data));
1372 for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1373 (*c)[i].core_id = -1;
1375 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1379 for (i = 0; i < topo.num_packages; i++)
1380 (*p)[i].package_id = i;
1384 ERROR("calloc counters");
1390 * set cpu_id, core_num, pkg_num
1391 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1393 * increment topo.num_cores when 1st core in pkg seen
1396 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1397 struct pkg_data *pkg_base, int thread_num, int core_num,
1398 int pkg_num, int cpu_id)
1401 struct thread_data *t;
1402 struct core_data *c;
1405 t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1406 c = GET_CORE(core_base, core_num, pkg_num);
1407 p = GET_PKG(pkg_base, pkg_num);
1410 if (thread_num == 0) {
1411 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1412 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1414 } else if (ret != 0) {
1415 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1419 c->core_id = core_num;
1420 p->package_id = pkg_num;
1427 initialize_counters(int cpu_id)
1429 int my_thread_id, my_core_id, my_package_id;
1432 my_package_id = get_physical_package_id(cpu_id);
1433 if (my_package_id < 0)
1434 return my_package_id;
1435 my_core_id = get_core_id(cpu_id);
1439 if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1441 } else if (ret != 0) {
1448 ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1451 ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1457 #define DO_OR_GOTO_ERR(something) \
1459 ret = (something); \
1464 static int setup_all_buffers(void)
1468 DO_OR_GOTO_ERR(topology_probe());
1469 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1470 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1471 DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1481 turbostat_init(void)
1486 DO_OR_GOTO_ERR(check_cpuid());
1487 DO_OR_GOTO_ERR(check_dev_msr());
1488 DO_OR_GOTO_ERR(check_super_user());
1489 DO_OR_GOTO_ERR(setup_all_buffers());
1490 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1492 ts.tv_sec = interval_sec;
1495 plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, &ts, NULL);
1503 static const char *config_keys[] =
1507 static int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1510 turbostat_config(const char *key, const char *value)
1512 if (strcasecmp("Interval", key) == 0)
1513 interval_sec = atoi(value);
1519 void module_register(void);
1520 void module_register(void)
1522 plugin_register_init(PLUGIN_NAME, turbostat_init);
1523 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);