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 * _GNU_SOURCE is required because of the following functions:
38 #include <asm/msr-index.h>
43 #include <sys/types.h>
46 #include <sys/resource.h>
61 #define PLUGIN_NAME "turbostat"
63 static const char *proc_stat = "/proc/stat";
64 static unsigned int skip_c0;
65 static unsigned int skip_c1;
66 static unsigned int do_nhm_cstates;
67 static unsigned int do_snb_cstates;
68 static unsigned int do_c8_c9_c10;
69 static unsigned int do_slm_cstates;
70 static unsigned int has_aperf;
71 static unsigned int has_epb;
72 static unsigned int units = 1000000000; /* Ghz etc */
73 static unsigned int genuine_intel;
74 static unsigned int has_invariant_tsc;
75 static unsigned int do_nehalem_platform_info;
77 static unsigned int do_rapl;
78 static unsigned int do_dts;
79 static unsigned int do_ptm;
80 static unsigned int tcc_activation_temp;
81 static unsigned int tcc_activation_temp_override;
82 static double rapl_power_units, rapl_energy_units, rapl_time_units;
83 static double rapl_joule_counter_range;
85 #define RAPL_PKG (1 << 0)
86 /* 0x610 MSR_PKG_POWER_LIMIT */
87 /* 0x611 MSR_PKG_ENERGY_STATUS */
88 #define RAPL_PKG_PERF_STATUS (1 << 1)
89 /* 0x613 MSR_PKG_PERF_STATUS */
90 #define RAPL_PKG_POWER_INFO (1 << 2)
91 /* 0x614 MSR_PKG_POWER_INFO */
93 #define RAPL_DRAM (1 << 3)
94 /* 0x618 MSR_DRAM_POWER_LIMIT */
95 /* 0x619 MSR_DRAM_ENERGY_STATUS */
96 /* 0x61c MSR_DRAM_POWER_INFO */
97 #define RAPL_DRAM_PERF_STATUS (1 << 4)
98 /* 0x61b MSR_DRAM_PERF_STATUS */
100 #define RAPL_CORES (1 << 5)
101 /* 0x638 MSR_PP0_POWER_LIMIT */
102 /* 0x639 MSR_PP0_ENERGY_STATUS */
103 #define RAPL_CORE_POLICY (1 << 6)
104 /* 0x63a MSR_PP0_POLICY */
107 #define RAPL_GFX (1 << 7)
108 /* 0x640 MSR_PP1_POWER_LIMIT */
109 /* 0x641 MSR_PP1_ENERGY_STATUS */
110 /* 0x642 MSR_PP1_POLICY */
111 #define TJMAX_DEFAULT 100
113 int aperf_mperf_unstable;
117 cpu_set_t *cpu_present_set, *cpu_affinity_set;
118 size_t cpu_present_setsize, cpu_affinity_setsize;
121 unsigned long long tsc;
122 unsigned long long aperf;
123 unsigned long long mperf;
124 unsigned long long c1;
125 unsigned int smi_count;
128 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
129 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
130 } *thread_even, *thread_odd;
133 unsigned long long c3;
134 unsigned long long c6;
135 unsigned long long c7;
136 unsigned int core_temp_c;
137 unsigned int core_id;
138 } *core_even, *core_odd;
141 unsigned long long pc2;
142 unsigned long long pc3;
143 unsigned long long pc6;
144 unsigned long long pc7;
145 unsigned long long pc8;
146 unsigned long long pc9;
147 unsigned long long pc10;
148 unsigned int package_id;
149 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
150 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
151 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
152 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
153 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
154 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
155 unsigned int pkg_temp_c;
157 } *package_even, *package_odd;
159 #define ODD_COUNTERS thread_odd, core_odd, package_odd
160 #define EVEN_COUNTERS thread_even, core_even, package_even
161 static _Bool is_even = 1;
163 static _Bool allocated = 0;
164 static _Bool initialized = 0;
166 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
167 (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
168 topo.num_threads_per_core + \
169 (core_no) * topo.num_threads_per_core + (thread_no))
170 #define GET_CORE(core_base, core_no, pkg_no) \
171 (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
172 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
179 int num_cores_per_pkg;
180 int num_threads_per_core;
183 struct timeval tv_even, tv_odd, tv_delta;
191 ERR_MSR_CORE_C3_RESIDENCY,
192 ERR_MSR_CORE_C6_RESIDENCY,
193 ERR_MSR_CORE_C7_RESIDENCY,
194 ERR_MSR_IA32_THERM_STATUS,
195 ERR_MSR_PKG_C3_RESIDENCY,
196 ERR_MSR_PKG_C6_RESIDENCY,
197 ERR_MSR_PKG_C2_RESIDENCY,
198 ERR_MSR_PKG_C7_RESIDENCY,
199 ERR_MSR_PKG_C8_RESIDENCY,
200 ERR_MSR_PKG_C9_RESIDENCY,
201 ERR_MSR_PKG_C10_RESIDENCY,
202 ERR_MSR_PKG_ENERGY_STATUS,
203 ERR_MSR_PP0_ENERGY_STATUS,
204 ERR_MSR_DRAM_ENERGY_STATUS,
205 ERR_MSR_PP1_ENERGY_STATUS,
206 ERR_MSR_PKG_PERF_STATUS,
207 ERR_MSR_DRAM_PERF_STATUS,
208 ERR_MSR_IA32_PACKAGE_THERM_STATUS,
212 ERR_CANT_READ_NUMBER,
213 ERR_CANT_READ_PROC_STAT,
214 ERR_NO_INVARIANT_TSC,
221 static int setup_all_buffers(void);
224 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 static int __attribute__((warn_unused_result))
234 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
235 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
237 int retval, pkg_no, core_no, thread_no;
239 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
240 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
241 for (thread_no = 0; thread_no <
242 topo.num_threads_per_core; ++thread_no) {
243 struct thread_data *t;
247 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
249 if (cpu_is_not_present(t->cpu_id))
252 c = GET_CORE(core_base, core_no, pkg_no);
253 p = GET_PKG(pkg_base, pkg_no);
255 retval = func(t, c, p);
264 static int __attribute__((warn_unused_result))
267 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
268 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
269 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
270 return -ERR_CPU_MIGRATE;
275 static int __attribute__((warn_unused_result))
276 get_msr(int cpu, off_t offset, unsigned long long *msr)
282 ssnprintf(pathname, 32, "/dev/cpu/%d/msr", cpu);
283 fd = open(pathname, O_RDONLY);
287 retval = pread(fd, msr, sizeof *msr, offset);
290 if (retval != sizeof *msr) {
291 ERROR ("%s offset 0x%llx read failed\n", pathname, (unsigned long long)offset);
298 #define DELTA_WRAP32(new, old) \
302 old = 0x100000000 + new - old; \
306 delta_package(struct pkg_data *new, struct pkg_data *old)
308 old->pc2 = new->pc2 - old->pc2;
309 old->pc3 = new->pc3 - old->pc3;
310 old->pc6 = new->pc6 - old->pc6;
311 old->pc7 = new->pc7 - old->pc7;
312 old->pc8 = new->pc8 - old->pc8;
313 old->pc9 = new->pc9 - old->pc9;
314 old->pc10 = new->pc10 - old->pc10;
315 old->pkg_temp_c = new->pkg_temp_c;
317 DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
318 DELTA_WRAP32(new->energy_cores, old->energy_cores);
319 DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
320 DELTA_WRAP32(new->energy_dram, old->energy_dram);
321 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
322 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
326 delta_core(struct core_data *new, struct core_data *old)
328 old->c3 = new->c3 - old->c3;
329 old->c6 = new->c6 - old->c6;
330 old->c7 = new->c7 - old->c7;
331 old->core_temp_c = new->core_temp_c;
337 static int __attribute__((warn_unused_result))
338 delta_thread(struct thread_data *new, struct thread_data *old,
339 struct core_data *core_delta)
341 old->tsc = new->tsc - old->tsc;
343 /* check for TSC < 1 Mcycles over interval */
344 if (old->tsc < (1000 * 1000)) {
345 WARNING("Insanely slow TSC rate, TSC stops in idle?\n"
346 "You can disable all c-states by booting with \"idle=poll\"\n"
347 "or just the deep ones with \"processor.max_cstate=1\"");
351 old->c1 = new->c1 - old->c1;
353 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
354 old->aperf = new->aperf - old->aperf;
355 old->mperf = new->mperf - old->mperf;
358 if (!aperf_mperf_unstable) {
359 WARNING("%s: APERF or MPERF went backwards *\n", progname);
360 WARNING("* Frequency results do not cover entire interval *\n");
361 WARNING("* fix this by running Linux-2.6.30 or later *\n");
363 aperf_mperf_unstable = 1;
366 * mperf delta is likely a huge "positive" number
367 * can not use it for calculating c0 time
375 * As counter collection is not atomic,
376 * it is possible for mperf's non-halted cycles + idle states
377 * to exceed TSC's all cycles: show c1 = 0% in that case.
379 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
382 /* normal case, derive c1 */
383 old->c1 = old->tsc - old->mperf - core_delta->c3
384 - core_delta->c6 - core_delta->c7;
387 if (old->mperf == 0) {
388 WARNING("cpu%d MPERF 0!\n", old->cpu_id);
389 old->mperf = 1; /* divide by 0 protection */
393 old->smi_count = new->smi_count - old->smi_count;
398 static int __attribute__((warn_unused_result))
399 delta_cpu(struct thread_data *t, struct core_data *c,
400 struct pkg_data *p, struct thread_data *t2,
401 struct core_data *c2, struct pkg_data *p2)
405 /* calculate core delta only for 1st thread in core */
406 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
409 /* always calculate thread delta */
410 ret = delta_thread(t, t2, c2); /* c2 is core delta */
414 /* calculate package delta only for 1st core in package */
415 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
416 delta_package(p, p2);
421 static unsigned long long
424 unsigned int low, high;
426 asm volatile("rdtsc" : "=a" (low), "=d" (high));
428 return low | ((unsigned long long)high) << 32;
435 * acquire and record local counters for that cpu
437 static int __attribute__((warn_unused_result))
438 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
441 unsigned long long msr;
443 if (cpu_migrate(cpu)) {
444 WARNING("Could not migrate to CPU %d\n", cpu);
445 return -ERR_CPU_MIGRATE;
448 t->tsc = rdtsc(); /* we are running on local CPU of interest */
451 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
452 return -ERR_MSR_IA32_APERF;
453 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
454 return -ERR_MSR_IA32_MPERF;
458 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
459 return -ERR_MSR_SMI_COUNT;
460 t->smi_count = msr & 0xFFFFFFFF;
463 /* collect core counters only for 1st thread in core */
464 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
467 if (do_nhm_cstates && !do_slm_cstates) {
468 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
469 return -ERR_MSR_CORE_C3_RESIDENCY;
472 if (do_nhm_cstates) {
473 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
474 return -ERR_MSR_CORE_C6_RESIDENCY;
478 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
479 return -ERR_MSR_CORE_C7_RESIDENCY;
482 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
483 return -ERR_MSR_IA32_THERM_STATUS;
484 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
488 /* collect package counters only for 1st core in package */
489 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
492 if (do_nhm_cstates && !do_slm_cstates) {
493 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
494 return -ERR_MSR_PKG_C3_RESIDENCY;
495 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
496 return -ERR_MSR_PKG_C6_RESIDENCY;
498 if (do_snb_cstates) {
499 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
500 return -ERR_MSR_PKG_C2_RESIDENCY;
501 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
502 return -ERR_MSR_PKG_C7_RESIDENCY;
505 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
506 return -ERR_MSR_PKG_C8_RESIDENCY;
507 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
508 return -ERR_MSR_PKG_C9_RESIDENCY;
509 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
510 return -ERR_MSR_PKG_C10_RESIDENCY;
512 if (do_rapl & RAPL_PKG) {
513 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
514 return -ERR_MSR_PKG_ENERGY_STATUS;
515 p->energy_pkg = msr & 0xFFFFFFFF;
517 if (do_rapl & RAPL_CORES) {
518 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
519 return MSR_PP0_ENERGY_STATUS;
520 p->energy_cores = msr & 0xFFFFFFFF;
522 if (do_rapl & RAPL_DRAM) {
523 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
524 return -ERR_MSR_DRAM_ENERGY_STATUS;
525 p->energy_dram = msr & 0xFFFFFFFF;
527 if (do_rapl & RAPL_GFX) {
528 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
529 return -ERR_MSR_PP1_ENERGY_STATUS;
530 p->energy_gfx = msr & 0xFFFFFFFF;
532 if (do_rapl & RAPL_PKG_PERF_STATUS) {
533 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
534 return -ERR_MSR_PKG_PERF_STATUS;
535 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
537 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
538 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
539 return -ERR_MSR_DRAM_PERF_STATUS;
540 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
543 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
544 return -ERR_MSR_IA32_PACKAGE_THERM_STATUS;
545 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
551 free_all_buffers(void)
556 CPU_FREE(cpu_present_set);
557 cpu_present_set = NULL;
560 CPU_FREE(cpu_affinity_set);
561 cpu_affinity_set = NULL;
562 cpu_affinity_setsize = 0;
582 * Parse a file containing a single int.
584 static int __attribute__ ((format(printf,1,2)))
585 parse_int_file(const char *fmt, ...)
593 vsnprintf(path, sizeof(path), fmt, args);
595 filep = fopen(path, "r");
597 ERROR("%s: open failed", path);
598 return -ERR_CANT_OPEN_FILE;
600 if (fscanf(filep, "%d", &value) != 1) {
601 ERROR("%s: failed to parse number from file", path);
602 return -ERR_CANT_READ_NUMBER;
609 * cpu_is_first_sibling_in_core(cpu)
610 * return 1 if given CPU is 1st HT sibling in the core
613 cpu_is_first_sibling_in_core(int cpu)
615 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
619 * cpu_is_first_core_in_package(cpu)
620 * return 1 if given CPU is 1st core in package
623 cpu_is_first_core_in_package(int cpu)
625 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
629 get_physical_package_id(int cpu)
631 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
637 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
641 get_num_ht_siblings(int cpu)
649 ssnprintf(path, 80, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
650 filep = fopen(path, "r");
652 ERROR("%s: open failed", path);
653 return -ERR_CANT_OPEN_FILE;
657 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
658 * otherwinse 1 sibling (self).
660 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
671 * run func(thread, core, package) in topology order
672 * skip non-present cpus
676 static int __attribute__((warn_unused_result))
677 for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
678 struct pkg_data *, struct thread_data *, struct core_data *,
679 struct pkg_data *), struct thread_data *thread_base,
680 struct core_data *core_base, struct pkg_data *pkg_base,
681 struct thread_data *thread_base2, struct core_data *core_base2,
682 struct pkg_data *pkg_base2)
684 int retval, pkg_no, core_no, thread_no;
686 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
687 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
688 for (thread_no = 0; thread_no <
689 topo.num_threads_per_core; ++thread_no) {
690 struct thread_data *t, *t2;
691 struct core_data *c, *c2;
692 struct pkg_data *p, *p2;
694 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
696 if (cpu_is_not_present(t->cpu_id))
699 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
701 c = GET_CORE(core_base, core_no, pkg_no);
702 c2 = GET_CORE(core_base2, core_no, pkg_no);
704 p = GET_PKG(pkg_base, pkg_no);
705 p2 = GET_PKG(pkg_base2, pkg_no);
707 retval = func(t, c, p, t2, c2, p2);
717 * run func(cpu) on every cpu in /proc/stat
718 * return max_cpu number
720 static int __attribute__((warn_unused_result))
721 for_all_proc_cpus(int (func)(int))
727 fp = fopen(proc_stat, "r");
729 ERROR("%s: open failed", proc_stat);
730 return -ERR_CANT_OPEN_FILE;
733 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
735 ERROR("%s: failed to parse format", proc_stat);
736 return -ERR_CANT_READ_PROC_STAT;
740 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
744 retval = func(cpu_num);
756 * remember the last one seen, it will be the max
761 if (topo.max_cpu_num < cpu)
762 topo.max_cpu_num = cpu;
768 mark_cpu_present(int cpu)
770 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
776 turbostat_submit (const char *plugin_instance,
777 const char *type, const char *type_instance,
780 value_list_t vl = VALUE_LIST_INIT;
786 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
787 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
788 if (plugin_instance != NULL)
789 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
790 sstrncpy (vl.type, type, sizeof (vl.type));
791 if (type_instance != NULL)
792 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
794 plugin_dispatch_values (&vl);
798 * column formatting convention & formats
799 * package: "pk" 2 columns %2d
800 * core: "cor" 3 columns %3d
801 * CPU: "CPU" 3 columns %3d
806 * GHz: "GHz" 3 columns %3.2
807 * TSC: "TSC" 3 columns %3.2
808 * SMI: "SMI" 4 columns %4d
809 * percentage " %pc3" %6.2
810 * Perf Status percentage: %5.2
811 * "CTMP" 4 columns %4d
815 submit_counters(struct thread_data *t, struct core_data *c,
819 double interval_float;
821 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
823 snprintf(name, NAME_LEN, "cpu%02d", t->cpu_id);
825 if (do_nhm_cstates) {
827 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
829 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
833 if (has_aperf && ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc))))
834 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
838 turbostat_submit(NULL, "current", name, t->smi_count);
840 /* print per-core data only for 1st thread in core */
841 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
844 snprintf(name, NAME_LEN, "core%02d", c->core_id);
846 if (do_nhm_cstates && !do_slm_cstates)
847 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
849 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
851 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
854 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
856 /* print per-package data only for 1st core in package */
857 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
860 snprintf(name, NAME_LEN, "pkg%02d", p->package_id);
863 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
866 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
867 if (do_nhm_cstates && !do_slm_cstates)
868 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
869 if (do_nhm_cstates && !do_slm_cstates)
870 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
872 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
874 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
875 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
876 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
880 if (do_rapl & RAPL_PKG)
881 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
882 if (do_rapl & RAPL_CORES)
883 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
884 if (do_rapl & RAPL_GFX)
885 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
886 if (do_rapl & RAPL_DRAM)
887 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
894 turbostat_read(user_data_t * not_used)
899 if ((ret = setup_all_buffers()) < 0)
903 if (for_all_proc_cpus(cpu_is_not_present)) {
905 if ((ret = setup_all_buffers()) < 0)
907 if (for_all_proc_cpus(cpu_is_not_present))
908 return -ERR_CPU_NOT_PRESENT;
912 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
914 gettimeofday(&tv_even, (struct timezone *)NULL);
921 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
923 gettimeofday(&tv_odd, (struct timezone *)NULL);
925 timersub(&tv_odd, &tv_even, &tv_delta);
926 if ((ret = for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) < 0)
928 if ((ret = for_all_cpus(submit_counters, EVEN_COUNTERS)) < 0)
931 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
933 gettimeofday(&tv_even, (struct timezone *)NULL);
935 timersub(&tv_even, &tv_odd, &tv_delta);
936 if ((ret = for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) < 0)
938 if ((ret = for_all_cpus(submit_counters, ODD_COUNTERS)) < 0)
944 static int __attribute__((warn_unused_result))
949 if (stat("/dev/cpu/0/msr", &sb)) {
950 ERROR("no /dev/cpu/0/msr\n"
951 "Try \"# modprobe msr\"");
957 static int __attribute__((warn_unused_result))
961 ERROR("must be root");
962 return -ERR_NOT_ROOT;
968 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
969 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
972 get_tdp(unsigned int model)
974 unsigned long long msr;
976 if (do_rapl & RAPL_PKG_POWER_INFO)
977 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
978 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
993 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
996 rapl_probe(unsigned int family, unsigned int model)
998 unsigned long long msr;
999 unsigned int time_unit;
1011 case 0x3C: /* HSW */
1012 case 0x45: /* HSW */
1013 case 0x46: /* HSW */
1014 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1016 case 0x3F: /* HSX */
1017 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1021 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1023 case 0x37: /* BYT */
1024 case 0x4D: /* AVN */
1025 do_rapl = RAPL_PKG | RAPL_CORES ;
1031 /* units on package 0, verify later other packages match */
1032 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1035 rapl_power_units = 1.0 / (1 << (msr & 0xF));
1037 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1039 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1041 time_unit = msr >> 16 & 0xF;
1045 rapl_time_units = 1.0 / (1 << (time_unit));
1047 tdp = get_tdp(model);
1049 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1051 // fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1057 is_snb(unsigned int family, unsigned int model)
1065 case 0x3A: /* IVB */
1066 case 0x3E: /* IVB Xeon */
1067 case 0x3C: /* HSW */
1068 case 0x3F: /* HSW */
1069 case 0x45: /* HSW */
1070 case 0x46: /* HSW */
1077 has_c8_c9_c10(unsigned int family, unsigned int model)
1091 is_slm(unsigned int family, unsigned int model)
1096 case 0x37: /* BYT */
1097 case 0x4D: /* AVN */
1104 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1105 * the Thermal Control Circuit (TCC) activates.
1106 * This is usually equal to tjMax.
1108 * Older processors do not have this MSR, so there we guess,
1109 * but also allow cmdline over-ride with -T.
1111 * Several MSR temperature values are in units of degrees-C
1112 * below this value, including the Digital Thermal Sensor (DTS),
1113 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1115 static int __attribute__((warn_unused_result))
1116 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1118 unsigned long long msr;
1119 unsigned int target_c_local;
1122 /* tcc_activation_temp is used only for dts or ptm */
1123 if (!(do_dts || do_ptm))
1126 /* this is a per-package concept */
1127 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1131 if (cpu_migrate(cpu)) {
1132 ERROR("Could not migrate to CPU %d\n", cpu);
1133 return -ERR_CPU_MIGRATE;
1136 if (tcc_activation_temp_override != 0) {
1137 tcc_activation_temp = tcc_activation_temp_override;
1138 ERROR("cpu%d: Using cmdline TCC Target (%d C)\n",
1139 cpu, tcc_activation_temp);
1143 /* Temperature Target MSR is Nehalem and newer only */
1144 if (!do_nehalem_platform_info)
1147 if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
1150 target_c_local = (msr >> 16) & 0x7F;
1152 if (target_c_local < 85 || target_c_local > 127)
1155 tcc_activation_temp = target_c_local;
1160 tcc_activation_temp = TJMAX_DEFAULT;
1161 WARNING("cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1162 cpu, tcc_activation_temp);
1167 static int __attribute__((warn_unused_result))
1170 unsigned int eax, ebx, ecx, edx, max_level;
1171 unsigned int fms, family, model;
1173 eax = ebx = ecx = edx = 0;
1175 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1177 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1181 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1182 family = (fms >> 8) & 0xf;
1183 model = (fms >> 4) & 0xf;
1184 if (family == 6 || family == 0xf)
1185 model += ((fms >> 16) & 0xf) << 4;
1187 if (!(edx & (1 << 5))) {
1188 ERROR("CPUID: no MSR");
1193 * check max extended function levels of CPUID.
1194 * This is needed to check for invariant TSC.
1195 * This check is valid for both Intel and AMD.
1197 ebx = ecx = edx = 0;
1198 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1200 if (max_level < 0x80000007) {
1201 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1202 return -ERR_NO_INVARIANT_TSC;
1206 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1207 * this check is valid for both Intel and AMD
1209 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1210 has_invariant_tsc = edx & (1 << 8);
1212 if (!has_invariant_tsc) {
1213 ERROR("No invariant TSC");
1214 return -ERR_NO_INVARIANT_TSC;
1218 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1219 * this check is valid for both Intel and AMD
1222 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1223 has_aperf = ecx & (1 << 0);
1224 do_dts = eax & (1 << 0);
1225 do_ptm = eax & (1 << 6);
1226 has_epb = ecx & (1 << 3);
1230 return -ERR_NO_APERF;
1233 do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1234 do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */
1235 do_smi = do_nhm_cstates;
1236 do_snb_cstates = is_snb(family, model);
1237 do_c8_c9_c10 = has_c8_c9_c10(family, model);
1238 do_slm_cstates = is_slm(family, model);
1240 rapl_probe(family, model);
1247 static int __attribute__((warn_unused_result))
1252 int max_core_id = 0;
1253 int max_package_id = 0;
1254 int max_siblings = 0;
1255 struct cpu_topology {
1257 int physical_package_id;
1260 /* Initialize num_cpus, max_cpu_num */
1262 topo.max_cpu_num = 0;
1263 ret = for_all_proc_cpus(count_cpus);
1267 DEBUG("num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1269 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
1271 ERROR("calloc cpus");
1276 * Allocate and initialize cpu_present_set
1278 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1279 if (cpu_present_set == NULL) {
1282 return -ERR_CPU_ALLOC;
1284 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1285 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1286 ret = for_all_proc_cpus(mark_cpu_present);
1293 * Allocate and initialize cpu_affinity_set
1295 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1296 if (cpu_affinity_set == NULL) {
1299 return -ERR_CPU_ALLOC;
1301 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1302 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1307 * find max_core_id, max_package_id
1309 for (i = 0; i <= topo.max_cpu_num; ++i) {
1312 if (cpu_is_not_present(i)) {
1314 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1317 cpus[i].core_id = get_core_id(i);
1318 if (cpus[i].core_id < 0)
1319 return cpus[i].core_id;
1320 if (cpus[i].core_id > max_core_id)
1321 max_core_id = cpus[i].core_id;
1323 cpus[i].physical_package_id = get_physical_package_id(i);
1324 if (cpus[i].physical_package_id < 0)
1325 return cpus[i].physical_package_id;
1326 if (cpus[i].physical_package_id > max_package_id)
1327 max_package_id = cpus[i].physical_package_id;
1329 siblings = get_num_ht_siblings(i);
1332 if (siblings > max_siblings)
1333 max_siblings = siblings;
1334 DEBUG("cpu %d pkg %d core %d\n",
1335 i, cpus[i].physical_package_id, cpus[i].core_id);
1337 topo.num_cores_per_pkg = max_core_id + 1;
1338 DEBUG("max_core_id %d, sizing for %d cores per package\n",
1339 max_core_id, topo.num_cores_per_pkg);
1341 topo.num_packages = max_package_id + 1;
1342 DEBUG("max_package_id %d, sizing for %d packages\n",
1343 max_package_id, topo.num_packages);
1345 topo.num_threads_per_core = max_siblings;
1346 DEBUG("max_siblings %d\n", max_siblings);
1353 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1357 *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1358 topo.num_packages, sizeof(struct thread_data));
1362 for (i = 0; i < topo.num_threads_per_core *
1363 topo.num_cores_per_pkg * topo.num_packages; i++)
1364 (*t)[i].cpu_id = -1;
1366 *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1367 sizeof(struct core_data));
1371 for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1372 (*c)[i].core_id = -1;
1374 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
1378 for (i = 0; i < topo.num_packages; i++)
1379 (*p)[i].package_id = i;
1383 ERROR("calloc counters");
1389 * set cpu_id, core_num, pkg_num
1390 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1392 * increment topo.num_cores when 1st core in pkg seen
1395 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1396 struct pkg_data *pkg_base, int thread_num, int core_num,
1397 int pkg_num, int cpu_id)
1400 struct thread_data *t;
1401 struct core_data *c;
1404 t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1405 c = GET_CORE(core_base, core_num, pkg_num);
1406 p = GET_PKG(pkg_base, pkg_num);
1409 if (thread_num == 0) {
1410 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1411 if ((ret = cpu_is_first_core_in_package(cpu_id)) < 0) {
1413 } else if (ret != 0) {
1414 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1418 c->core_id = core_num;
1419 p->package_id = pkg_num;
1426 initialize_counters(int cpu_id)
1428 int my_thread_id, my_core_id, my_package_id;
1431 my_package_id = get_physical_package_id(cpu_id);
1432 if (my_package_id < 0)
1433 return my_package_id;
1434 my_core_id = get_core_id(cpu_id);
1438 if ((ret = cpu_is_first_sibling_in_core(cpu_id)) < 0) {
1440 } else if (ret != 0) {
1447 ret = init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1450 ret = init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1456 #define DO_OR_GOTO_ERR(something) \
1458 ret = (something); \
1463 static int setup_all_buffers(void)
1467 DO_OR_GOTO_ERR(topology_probe());
1468 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1469 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1470 DO_OR_GOTO_ERR(for_all_proc_cpus(initialize_counters));
1480 turbostat_init(void)
1484 DO_OR_GOTO_ERR(check_cpuid());
1485 DO_OR_GOTO_ERR(check_dev_msr());
1486 DO_OR_GOTO_ERR(check_super_user());
1487 DO_OR_GOTO_ERR(setup_all_buffers());
1488 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1490 plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1498 void module_register(void);
1499 void module_register(void)
1501 plugin_register_init(PLUGIN_NAME, turbostat_init);