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:
42 #include "utils_time.h"
44 #include <asm/msr-index.h>
46 #ifdef HAVE_SYS_CAPABILITY_H
47 #include <sys/capability.h>
48 #endif /* HAVE_SYS_CAPABILITY_H */
50 #define PLUGIN_NAME "turbostat"
53 * This tool uses the Model-Specific Registers (MSRs) present on Intel
55 * The general description each of these registers, depending on the
57 * can be found in the IntelĀ® 64 and IA-32 Architectures Software Developer
59 * Volume 3 Chapter 35.
63 * If set, aperf_mperf_unstable disables a/mperf based stats.
64 * This includes: C0 & C1 states, frequency
66 * This value is automatically set if mperf or aperf go backward
68 static _Bool aperf_mperf_unstable;
71 * Bitmask of the list of core C states supported by the processor.
72 * Currently supported C-states (by this plugin): 3, 6, 7
74 static unsigned int do_core_cstate;
75 static unsigned int config_core_cstate;
76 static _Bool apply_config_core_cstate;
79 * Bitmask of the list of pacages C states supported by the processor.
80 * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
82 static unsigned int do_pkg_cstate;
83 static unsigned int config_pkg_cstate;
84 static _Bool apply_config_pkg_cstate;
87 * Boolean indicating if the processor supports 'I/O System-Management Interrupt
91 static _Bool config_smi;
92 static _Bool apply_config_smi;
95 * Boolean indicating if the processor supports 'Digital temperature sensor'
96 * This feature enables the monitoring of the temperature of each core
98 * This feature has two limitations:
99 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature
101 * - Temperatures above the tcc_activation_temp are not recorded
104 static _Bool config_dts;
105 static _Bool apply_config_dts;
108 * Boolean indicating if the processor supports 'Package thermal management'
109 * This feature allows the monitoring of the temperature of each package
111 * This feature has two limitations:
112 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature
114 * - Temperatures above the tcc_activation_temp are not recorded
117 static _Bool config_ptm;
118 static _Bool apply_config_ptm;
121 * Thermal Control Circuit Activation Temperature as configured by the user.
122 * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
123 * and should only be used if the automated detection fails.
125 static unsigned int tcc_activation_temp;
127 static unsigned int do_rapl;
128 static unsigned int config_rapl;
129 static _Bool apply_config_rapl;
130 static double rapl_energy_units;
132 #define RAPL_PKG (1 << 0)
133 /* 0x610 MSR_PKG_POWER_LIMIT */
134 /* 0x611 MSR_PKG_ENERGY_STATUS */
135 #define RAPL_DRAM (1 << 1)
136 /* 0x618 MSR_DRAM_POWER_LIMIT */
137 /* 0x619 MSR_DRAM_ENERGY_STATUS */
138 /* 0x61c MSR_DRAM_POWER_INFO */
139 #define RAPL_CORES (1 << 2)
140 /* 0x638 MSR_PP0_POWER_LIMIT */
141 /* 0x639 MSR_PP0_ENERGY_STATUS */
143 #define RAPL_GFX (1 << 3)
144 /* 0x640 MSR_PP1_POWER_LIMIT */
145 /* 0x641 MSR_PP1_ENERGY_STATUS */
146 /* 0x642 MSR_PP1_POLICY */
147 #define TJMAX_DEFAULT 100
149 static cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
150 static size_t cpu_present_setsize, cpu_affinity_setsize,
151 cpu_saved_affinity_setsize;
153 static struct thread_data {
154 unsigned long long tsc;
155 unsigned long long aperf;
156 unsigned long long mperf;
157 unsigned long long c1;
158 unsigned int smi_count;
161 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
162 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
163 } * thread_delta, *thread_even, *thread_odd;
165 static struct core_data {
166 unsigned long long c3;
167 unsigned long long c6;
168 unsigned long long c7;
169 unsigned int core_temp_c;
170 unsigned int core_id;
171 } * core_delta, *core_even, *core_odd;
173 static struct pkg_data {
174 unsigned long long pc2;
175 unsigned long long pc3;
176 unsigned long long pc6;
177 unsigned long long pc7;
178 unsigned long long pc8;
179 unsigned long long pc9;
180 unsigned long long pc10;
181 unsigned int package_id;
182 uint32_t energy_pkg; /* MSR_PKG_ENERGY_STATUS */
183 uint32_t energy_dram; /* MSR_DRAM_ENERGY_STATUS */
184 uint32_t energy_cores; /* MSR_PP0_ENERGY_STATUS */
185 uint32_t energy_gfx; /* MSR_PP1_ENERGY_STATUS */
186 unsigned int tcc_activation_temp;
187 unsigned int pkg_temp_c;
188 } * package_delta, *package_even, *package_odd;
190 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
191 #define ODD_COUNTERS thread_odd, core_odd, package_odd
192 #define EVEN_COUNTERS thread_even, core_even, package_even
193 static _Bool is_even = 1;
195 static _Bool allocated = 0;
196 static _Bool initialized = 0;
198 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
199 (thread_base + (pkg_no)*topology.num_cores * topology.num_threads + \
200 (core_no)*topology.num_threads + (thread_no))
201 #define GET_CORE(core_base, core_no, pkg_no) \
202 (core_base + (pkg_no)*topology.num_cores + (core_no))
203 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
205 struct cpu_topology {
206 unsigned int package_id;
207 unsigned int core_id;
208 _Bool first_core_in_package;
209 _Bool first_thread_in_core;
212 static struct topology {
213 unsigned int max_cpu_id;
214 unsigned int num_packages;
215 unsigned int num_cores;
216 unsigned int num_threads;
217 struct cpu_topology *cpus;
220 static cdtime_t time_even, time_odd, time_delta;
222 static const char *config_keys[] = {
225 "SystemManagementInterrupt",
226 "DigitalTemperatureSensor",
227 "PackageThermalManagement",
229 "RunningAveragePowerLimit",
231 static const int config_keys_num = STATIC_ARRAY_SIZE(config_keys);
233 /*****************************
234 * MSR Manipulation helpers *
235 *****************************/
238 * Open a MSR device for reading
239 * Can change the scheduling affinity of the current process if multiple_read is
242 static int __attribute__((warn_unused_result))
243 open_msr(unsigned int cpu, _Bool multiple_read) {
248 * If we need to do multiple read, let's migrate to the CPU
249 * Otherwise, we would lose time calling functions on another CPU
251 * If we are not yet initialized (cpu_affinity_setsize = 0),
252 * we need to skip this optimisation.
254 if (multiple_read && cpu_affinity_setsize) {
255 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
256 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
257 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
258 ERROR("turbostat plugin: Could not migrate to CPU %d", cpu);
263 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
264 fd = open(pathname, O_RDONLY);
266 ERROR("turbostat plugin: failed to open %s", pathname);
273 * Read a single MSR from an open file descriptor
275 static int __attribute__((warn_unused_result))
276 read_msr(int fd, off_t offset, unsigned long long *msr) {
279 retval = pread(fd, msr, sizeof *msr, offset);
281 if (retval != sizeof *msr) {
282 ERROR("turbostat plugin: MSR offset 0x%llx read failed",
283 (unsigned long long)offset);
290 * Open a MSR device for reading, read the value asked for and close it.
291 * This call will not affect the scheduling affinity of this thread.
293 static ssize_t __attribute__((warn_unused_result))
294 get_msr(unsigned int cpu, off_t offset, unsigned long long *msr) {
298 fd = open_msr(cpu, 0);
301 retval = read_msr(fd, offset, msr);
306 /********************************
307 * Raw data acquisition (1 CPU) *
308 ********************************/
311 * Read every data avalaible for a single CPU
313 * Core data is shared for all threads in one core: extracted only for the first
315 * Package data is shared for all core in one package: extracted only for the
316 * first thread of the first core
318 * Side effect: migrates to the targeted CPU
320 static int __attribute__((warn_unused_result))
321 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p) {
322 unsigned int cpu = t->cpu_id;
323 unsigned long long msr;
327 msr_fd = open_msr(cpu, 1);
331 #define READ_MSR(msr, dst) \
333 if (read_msr(msr_fd, msr, dst)) { \
334 ERROR("turbostat plugin: Unable to read " #msr); \
340 READ_MSR(MSR_IA32_TSC, &t->tsc);
342 READ_MSR(MSR_IA32_APERF, &t->aperf);
343 READ_MSR(MSR_IA32_MPERF, &t->mperf);
346 READ_MSR(MSR_SMI_COUNT, &msr);
347 t->smi_count = msr & 0xFFFFFFFF;
350 /* collect core counters only for 1st thread in core */
351 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
356 if (do_core_cstate & (1 << 3))
357 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
358 if (do_core_cstate & (1 << 6))
359 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
360 if (do_core_cstate & (1 << 7))
361 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
364 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
365 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
368 /* collect package counters only for 1st core in package */
369 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
374 if (do_pkg_cstate & (1 << 2))
375 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
376 if (do_pkg_cstate & (1 << 3))
377 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
378 if (do_pkg_cstate & (1 << 6))
379 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
380 if (do_pkg_cstate & (1 << 7))
381 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
382 if (do_pkg_cstate & (1 << 8))
383 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
384 if (do_pkg_cstate & (1 << 9))
385 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
386 if (do_pkg_cstate & (1 << 10))
387 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
389 if (do_rapl & RAPL_PKG) {
390 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
391 p->energy_pkg = msr & 0xFFFFFFFF;
393 if (do_rapl & RAPL_CORES) {
394 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
395 p->energy_cores = msr & 0xFFFFFFFF;
397 if (do_rapl & RAPL_DRAM) {
398 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
399 p->energy_dram = msr & 0xFFFFFFFF;
401 if (do_rapl & RAPL_GFX) {
402 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
403 p->energy_gfx = msr & 0xFFFFFFFF;
406 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
407 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
415 /**********************************
416 * Evaluating the changes (1 CPU) *
417 **********************************/
420 * Extract the evolution old->new in delta at a package level
421 * (some are not new-delta, e.g. temperature)
423 static inline void delta_package(struct pkg_data *delta,
424 const struct pkg_data *new,
425 const struct pkg_data *old) {
426 delta->pc2 = new->pc2 - old->pc2;
427 delta->pc3 = new->pc3 - old->pc3;
428 delta->pc6 = new->pc6 - old->pc6;
429 delta->pc7 = new->pc7 - old->pc7;
430 delta->pc8 = new->pc8 - old->pc8;
431 delta->pc9 = new->pc9 - old->pc9;
432 delta->pc10 = new->pc10 - old->pc10;
433 delta->pkg_temp_c = new->pkg_temp_c;
435 delta->energy_pkg = new->energy_pkg - old->energy_pkg;
436 delta->energy_cores = new->energy_cores - old->energy_cores;
437 delta->energy_gfx = new->energy_gfx - old->energy_gfx;
438 delta->energy_dram = new->energy_dram - old->energy_dram;
442 * Extract the evolution old->new in delta at a core level
443 * (some are not new-delta, e.g. temperature)
445 static inline void delta_core(struct core_data *delta,
446 const struct core_data *new,
447 const struct core_data *old) {
448 delta->c3 = new->c3 - old->c3;
449 delta->c6 = new->c6 - old->c6;
450 delta->c7 = new->c7 - old->c7;
451 delta->core_temp_c = new->core_temp_c;
455 * Extract the evolution old->new in delta at a package level
456 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
458 static inline int __attribute__((warn_unused_result))
459 delta_thread(struct thread_data *delta, const struct thread_data *new,
460 const struct thread_data *old, const struct core_data *cdelta) {
461 delta->tsc = new->tsc - old->tsc;
463 /* check for TSC < 1 Mcycles over interval */
464 if (delta->tsc < (1000 * 1000)) {
465 WARNING("turbostat plugin: Insanely slow TSC rate, TSC stops "
466 "in idle? You can disable all c-states by booting with"
467 " 'idle=poll' or just the deep ones with"
468 " 'processor.max_cstate=1'");
472 delta->c1 = new->c1 - old->c1;
474 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
475 delta->aperf = new->aperf - old->aperf;
476 delta->mperf = new->mperf - old->mperf;
478 if (!aperf_mperf_unstable) {
479 WARNING("turbostat plugin: APERF or MPERF went "
480 "backwards. Frequency results do not cover "
481 "the entire interval. Fix this by running "
482 "Linux-2.6.30 or later.");
484 aperf_mperf_unstable = 1;
489 * As counter collection is not atomic,
490 * it is possible for mperf's non-halted cycles + idle states
491 * to exceed TSC's all cycles: show c1 = 0% in that case.
493 if ((delta->mperf + cdelta->c3 + cdelta->c6 + cdelta->c7) > delta->tsc)
496 /* normal case, derive c1 */
498 delta->tsc - delta->mperf - cdelta->c3 - cdelta->c6 - cdelta->c7;
501 if (delta->mperf == 0) {
502 WARNING("turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
503 delta->mperf = 1; /* divide by 0 protection */
507 delta->smi_count = new->smi_count - old->smi_count;
512 /**********************************
513 * Submitting the results (1 CPU) *
514 **********************************/
517 * Submit one gauge value
519 static void turbostat_submit(const char *plugin_instance, const char *type,
520 const char *type_instance, gauge_t value) {
521 value_list_t vl = VALUE_LIST_INIT;
527 sstrncpy(vl.host, hostname_g, sizeof(vl.host));
528 sstrncpy(vl.plugin, PLUGIN_NAME, sizeof(vl.plugin));
529 if (plugin_instance != NULL)
530 sstrncpy(vl.plugin_instance, plugin_instance, sizeof(vl.plugin_instance));
531 sstrncpy(vl.type, type, sizeof(vl.type));
532 if (type_instance != NULL)
533 sstrncpy(vl.type_instance, type_instance, sizeof(vl.type_instance));
535 plugin_dispatch_values(&vl);
539 * Submit every data for a single CPU
541 * Core data is shared for all threads in one core: submitted only for the first
543 * Package data is shared for all core in one package: submitted only for the
544 * first thread of the first core
546 static int submit_counters(struct thread_data *t, struct core_data *c,
547 struct pkg_data *p) {
548 char name[DATA_MAX_NAME_LEN];
549 double interval_float;
551 interval_float = CDTIME_T_TO_DOUBLE(time_delta);
553 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
555 if (!aperf_mperf_unstable)
556 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf / t->tsc);
557 if (!aperf_mperf_unstable)
558 turbostat_submit(name, "percent", "c1", 100.0 * t->c1 / t->tsc);
560 turbostat_submit(name, "frequency", "average",
561 1.0 / 1000000 * t->aperf / interval_float);
563 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
564 turbostat_submit(name, "frequency", "busy", 1.0 * t->tsc / 1000000 *
565 t->aperf / t->mperf /
568 /* Sanity check (should stay stable) */
569 turbostat_submit(name, "gauge", "TSC",
570 1.0 * t->tsc / 1000000 / interval_float);
574 turbostat_submit(name, "count", NULL, t->smi_count);
576 /* submit per-core data only for 1st thread in core */
577 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
580 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
582 if (do_core_cstate & (1 << 3))
583 turbostat_submit(name, "percent", "c3", 100.0 * c->c3 / t->tsc);
584 if (do_core_cstate & (1 << 6))
585 turbostat_submit(name, "percent", "c6", 100.0 * c->c6 / t->tsc);
586 if (do_core_cstate & (1 << 7))
587 turbostat_submit(name, "percent", "c7", 100.0 * c->c7 / t->tsc);
590 turbostat_submit(name, "temperature", NULL, c->core_temp_c);
592 /* submit per-package data only for 1st core in package */
593 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
596 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
599 turbostat_submit(name, "temperature", NULL, p->pkg_temp_c);
601 if (do_pkg_cstate & (1 << 2))
602 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2 / t->tsc);
603 if (do_pkg_cstate & (1 << 3))
604 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3 / t->tsc);
605 if (do_pkg_cstate & (1 << 6))
606 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6 / t->tsc);
607 if (do_pkg_cstate & (1 << 7))
608 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7 / t->tsc);
609 if (do_pkg_cstate & (1 << 8))
610 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8 / t->tsc);
611 if (do_pkg_cstate & (1 << 9))
612 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9 / t->tsc);
613 if (do_pkg_cstate & (1 << 10))
614 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10 / t->tsc);
617 if (do_rapl & RAPL_PKG)
618 turbostat_submit(name, "power", "pkg",
619 p->energy_pkg * rapl_energy_units / interval_float);
620 if (do_rapl & RAPL_CORES)
621 turbostat_submit(name, "power", "cores",
622 p->energy_cores * rapl_energy_units / interval_float);
623 if (do_rapl & RAPL_GFX)
624 turbostat_submit(name, "power", "GFX",
625 p->energy_gfx * rapl_energy_units / interval_float);
626 if (do_rapl & RAPL_DRAM)
627 turbostat_submit(name, "power", "DRAM",
628 p->energy_dram * rapl_energy_units / interval_float);
634 /**********************************
635 * Looping function over all CPUs *
636 **********************************/
639 * Check if a given cpu id is in our compiled list of existing CPUs
641 static int cpu_is_not_present(unsigned int cpu) {
642 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
646 * Loop on all CPUs in topological order
648 * Skip non-present cpus
649 * Return the error code at the first error or 0
651 static int __attribute__((warn_unused_result))
652 for_all_cpus(int(func)(struct thread_data *, struct core_data *,
654 struct thread_data *thread_base, struct core_data *core_base,
655 struct pkg_data *pkg_base) {
658 for (unsigned int pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
659 for (unsigned int core_no = 0; core_no < topology.num_cores; ++core_no) {
660 for (unsigned int thread_no = 0; thread_no < topology.num_threads;
662 struct thread_data *t;
666 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
668 if (cpu_is_not_present(t->cpu_id))
671 c = GET_CORE(core_base, core_no, pkg_no);
672 p = GET_PKG(pkg_base, pkg_no);
674 retval = func(t, c, p);
684 * Dedicated loop: Extract every data evolution for all CPU
686 * Skip non-present cpus
687 * Return the error code at the first error or 0
689 * Core data is shared for all threads in one core: extracted only for the first
691 * Package data is shared for all core in one package: extracted only for the
692 * first thread of the first core
694 static int __attribute__((warn_unused_result))
695 for_all_cpus_delta(const struct thread_data *thread_new_base,
696 const struct core_data *core_new_base,
697 const struct pkg_data *pkg_new_base,
698 const struct thread_data *thread_old_base,
699 const struct core_data *core_old_base,
700 const struct pkg_data *pkg_old_base) {
703 for (unsigned int pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
704 for (unsigned int core_no = 0; core_no < topology.num_cores; ++core_no) {
705 for (unsigned int thread_no = 0; thread_no < topology.num_threads;
707 struct thread_data *t_delta;
708 const struct thread_data *t_old, *t_new;
709 struct core_data *c_delta;
711 /* Get correct pointers for threads */
712 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
713 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
714 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
716 /* Skip threads that disappeared */
717 if (cpu_is_not_present(t_delta->cpu_id))
720 /* c_delta is always required for delta_thread */
721 c_delta = GET_CORE(core_delta, core_no, pkg_no);
723 /* calculate core delta only for 1st thread in core */
724 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
725 const struct core_data *c_old, *c_new;
727 c_new = GET_CORE(core_new_base, core_no, pkg_no);
728 c_old = GET_CORE(core_old_base, core_no, pkg_no);
730 delta_core(c_delta, c_new, c_old);
733 /* Always calculate thread delta */
734 retval = delta_thread(t_delta, t_new, t_old, c_delta);
738 /* calculate package delta only for 1st core in package */
739 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
740 struct pkg_data *p_delta;
741 const struct pkg_data *p_old, *p_new;
743 p_delta = GET_PKG(package_delta, pkg_no);
744 p_new = GET_PKG(pkg_new_base, pkg_no);
745 p_old = GET_PKG(pkg_old_base, pkg_no);
747 delta_package(p_delta, p_new, p_old);
760 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
761 * the Thermal Control Circuit (TCC) activates.
762 * This is usually equal to tjMax.
764 * Older processors do not have this MSR, so there we guess,
765 * but also allow conficuration over-ride with "TCCActivationTemp".
767 * Several MSR temperature values are in units of degrees-C
768 * below this value, including the Digital Thermal Sensor (DTS),
769 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
771 static int __attribute__((warn_unused_result))
772 set_temperature_target(struct thread_data *t, struct core_data *c,
773 struct pkg_data *p) {
774 unsigned long long msr;
775 unsigned int target_c_local;
777 /* tcc_activation_temp is used only for dts or ptm */
778 if (!(do_dts || do_ptm))
781 /* this is a per-package concept */
782 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) ||
783 !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
786 if (tcc_activation_temp != 0) {
787 p->tcc_activation_temp = tcc_activation_temp;
791 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
794 target_c_local = (msr >> 16) & 0xFF;
799 p->tcc_activation_temp = target_c_local;
804 p->tcc_activation_temp = TJMAX_DEFAULT;
805 WARNING("turbostat plugin: cpu%d: Guessing tjMax %d C,"
806 " Please use TCCActivationTemp to specify it.",
807 t->cpu_id, p->tcc_activation_temp);
813 * Identify the functionality of the CPU
815 static int __attribute__((warn_unused_result)) probe_cpu(void) {
816 unsigned int eax, ebx, ecx, edx, max_level;
817 unsigned int fms, family, model;
820 * - EAX: Maximum Input Value for Basic CPUID Information
821 * - EBX: "Genu" (0x756e6547)
822 * - EDX: "ineI" (0x49656e69)
823 * - ECX: "ntel" (0x6c65746e)
825 max_level = ebx = ecx = edx = 0;
826 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
827 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
828 ERROR("turbostat plugin: Unsupported CPU (not Intel)");
833 * - EAX: Version Information: Type, Family, Model, and Stepping ID
836 * + 12-13: Processor type
837 * + 16-19: Extended Model ID
838 * + 20-27: Extended Family ID
839 * - EDX: Feature Information:
840 * + 5: Support for MSR read/write operations
842 fms = ebx = ecx = edx = 0;
843 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
844 family = (fms >> 8) & 0xf;
845 model = (fms >> 4) & 0xf;
847 family += (fms >> 20) & 0xf;
848 if (family == 6 || family == 0xf)
849 model += ((fms >> 16) & 0xf) << 4;
850 if (!(edx & (1 << 5))) {
851 ERROR("turbostat plugin: Unsupported CPU (no MSR support)");
858 * + 0: Digital temperature sensor is supported if set
859 * + 6: Package thermal management is supported if set
861 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and
863 * + 3: The processor supports performance-energy bias preference if set.
864 * It also implies the presence of a new architectural MSR called
865 * IA32_ENERGY_PERF_BIAS
867 * This check is valid for both Intel and AMD
869 eax = ebx = ecx = edx = 0;
870 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
871 do_dts = eax & (1 << 0);
872 do_ptm = eax & (1 << 6);
873 if (!(ecx & (1 << 0))) {
874 ERROR("turbostat plugin: Unsupported CPU (No APERF)");
879 * Enable or disable C states depending on the model and family
887 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
893 do_core_cstate = (1 << 1) | (1 << 6);
894 do_pkg_cstate = (1 << 6);
897 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
898 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper
900 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
901 case 0x2E: /* Nehalem-EX Xeon - Beckton */
903 do_core_cstate = (1 << 3) | (1 << 6);
904 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
907 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
908 case 0x2C: /* Westmere EP - Gulftown */
909 case 0x2F: /* Westmere-EX Xeon - Eagleton */
911 do_core_cstate = (1 << 3) | (1 << 6);
912 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
916 case 0x2D: /* SNB Xeon */
918 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
919 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
923 case 0x3E: /* IVB Xeon */
925 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
926 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
933 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
934 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
938 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
939 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) |
940 (1 << 9) | (1 << 10);
944 case 0x56: /* BDX-DE */
946 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
947 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
951 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
952 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) |
953 (1 << 9) | (1 << 10);
968 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_GFX;
972 case 0x56: /* BDX-DE */
973 do_rapl = RAPL_PKG | RAPL_DRAM;
975 case 0x2D: /* SNB Xeon */
976 case 0x3E: /* IVB Xeon */
977 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_DRAM;
981 do_rapl = RAPL_PKG | RAPL_CORES;
987 ERROR("turbostat plugin: Unsupported CPU (family: %#x, "
993 /* Override detected values with configuration */
994 if (apply_config_core_cstate)
995 do_core_cstate = config_core_cstate;
996 if (apply_config_pkg_cstate)
997 do_pkg_cstate = config_pkg_cstate;
998 if (apply_config_smi)
1000 if (apply_config_dts)
1001 do_dts = config_dts;
1002 if (apply_config_ptm)
1003 do_ptm = config_ptm;
1004 if (apply_config_rapl)
1005 do_rapl = config_rapl;
1008 unsigned long long msr;
1009 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1013 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1015 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1021 /********************
1022 * Topology Probing *
1023 ********************/
1026 * Read a single int from a file.
1028 static int __attribute__((format(printf, 1, 2)))
1029 parse_int_file(const char *fmt, ...) {
1031 char path[PATH_MAX];
1035 va_start(args, fmt);
1036 len = vsnprintf(path, sizeof(path), fmt, args);
1038 if (len < 0 || len >= PATH_MAX) {
1039 ERROR("turbostat plugin: path truncated: '%s'", path);
1043 filep = fopen(path, "r");
1045 ERROR("turbostat plugin: Failed to open '%s'", path);
1048 if (fscanf(filep, "%d", &value) != 1) {
1049 ERROR("turbostat plugin: Failed to parse number from '%s'", path);
1057 static int get_threads_on_core(unsigned int cpu) {
1064 ssnprintf(path, sizeof(path),
1065 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1066 filep = fopen(path, "r");
1068 ERROR("turbostat plugin: Failed to open '%s'", path);
1073 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1074 * otherwinse 1 sibling (self).
1076 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1087 * run func(cpu) on every cpu in /proc/stat
1088 * return max_cpu number
1090 static int __attribute__((warn_unused_result))
1091 for_all_proc_cpus(int(func)(unsigned int)) {
1093 unsigned int cpu_num;
1096 fp = fopen("/proc/stat", "r");
1098 ERROR("turbostat plugin: Failed to open /proc/stat");
1102 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1104 ERROR("turbostat plugin: Failed to parse /proc/stat");
1111 fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1115 retval = func(cpu_num);
1126 * Update the stored topology.max_cpu_id
1128 static int update_max_cpu_id(unsigned int cpu) {
1129 if (topology.max_cpu_id < cpu)
1130 topology.max_cpu_id = cpu;
1134 static int mark_cpu_present(unsigned int cpu) {
1135 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1139 static int __attribute__((warn_unused_result))
1140 allocate_cpu_set(cpu_set_t **set, size_t *size) {
1141 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1143 ERROR("turbostat plugin: Unable to allocate CPU state");
1146 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1147 CPU_ZERO_S(*size, *set);
1152 * Build a local representation of the cpu distribution
1154 static int __attribute__((warn_unused_result)) topology_probe(void) {
1156 unsigned int max_package_id, max_core_id, max_threads;
1157 max_package_id = max_core_id = max_threads = 0;
1159 /* Clean topology */
1160 free(topology.cpus);
1161 memset(&topology, 0, sizeof(topology));
1163 ret = for_all_proc_cpus(update_max_cpu_id);
1168 calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1169 if (topology.cpus == NULL) {
1170 ERROR("turbostat plugin: Unable to allocate memory for CPU topology");
1174 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1177 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1180 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1184 ret = for_all_proc_cpus(mark_cpu_present);
1190 * find max_core_id, max_package_id
1192 for (unsigned int i = 0; i <= topology.max_cpu_id; ++i) {
1193 unsigned int num_threads;
1194 struct cpu_topology *cpu = &topology.cpus[i];
1196 if (cpu_is_not_present(i)) {
1197 WARNING("turbostat plugin: cpu%d NOT PRESENT", i);
1201 ret = parse_int_file(
1202 "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1206 cpu->package_id = (unsigned int)ret;
1207 if (cpu->package_id > max_package_id)
1208 max_package_id = cpu->package_id;
1210 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1214 cpu->core_id = (unsigned int)ret;
1215 if (cpu->core_id > max_core_id)
1216 max_core_id = cpu->core_id;
1217 ret = parse_int_file(
1218 "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1221 else if ((unsigned int)ret == i)
1222 cpu->first_core_in_package = 1;
1224 ret = get_threads_on_core(i);
1228 num_threads = (unsigned int)ret;
1229 if (num_threads > max_threads)
1230 max_threads = num_threads;
1231 ret = parse_int_file(
1232 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1235 else if ((unsigned int)ret == i)
1236 cpu->first_thread_in_core = 1;
1238 DEBUG("turbostat plugin: cpu %d pkg %d core %d\n", i, cpu->package_id,
1241 /* Num is max + 1 (need to count 0) */
1242 topology.num_packages = max_package_id + 1;
1243 topology.num_cores = max_core_id + 1;
1244 topology.num_threads = max_threads;
1248 free(topology.cpus);
1252 /************************
1253 * Main alloc/init/free *
1254 ************************/
1256 static int allocate_counters(struct thread_data **threads,
1257 struct core_data **cores,
1258 struct pkg_data **packages) {
1259 unsigned int total_threads, total_cores;
1261 if ((topology.num_threads == 0) || (topology.num_cores == 0) ||
1262 (topology.num_packages == 0)) {
1264 "turbostat plugin: Invalid topology: %u threads, %u cores, %u packages",
1265 topology.num_threads, topology.num_cores, topology.num_packages);
1270 topology.num_threads * topology.num_cores * topology.num_packages;
1271 *threads = calloc(total_threads, sizeof(struct thread_data));
1272 if (*threads == NULL) {
1273 ERROR("turbostat plugin: calloc failed");
1277 for (unsigned int i = 0; i < total_threads; ++i)
1278 (*threads)[i].cpu_id = topology.max_cpu_id + 1;
1280 total_cores = topology.num_cores * topology.num_packages;
1281 *cores = calloc(total_cores, sizeof(struct core_data));
1282 if (*cores == NULL) {
1283 ERROR("turbostat plugin: calloc failed");
1288 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1289 if (*packages == NULL) {
1290 ERROR("turbostat plugin: calloc failed");
1299 static void init_counter(struct thread_data *thread_base,
1300 struct core_data *core_base, struct pkg_data *pkg_base,
1301 unsigned int cpu_id) {
1302 struct thread_data *t;
1303 struct core_data *c;
1305 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1307 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id,
1309 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1310 p = GET_PKG(pkg_base, cpu->package_id);
1313 if (cpu->first_thread_in_core)
1314 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1315 if (cpu->first_core_in_package)
1316 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1318 c->core_id = cpu->core_id;
1319 p->package_id = cpu->package_id;
1322 static void initialize_counters(void) {
1323 for (unsigned int cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1324 if (cpu_is_not_present(cpu_id))
1326 init_counter(EVEN_COUNTERS, cpu_id);
1327 init_counter(ODD_COUNTERS, cpu_id);
1328 init_counter(DELTA_COUNTERS, cpu_id);
1332 static void free_all_buffers(void) {
1336 CPU_FREE(cpu_present_set);
1337 cpu_present_set = NULL;
1338 cpu_present_setsize = 0;
1340 CPU_FREE(cpu_affinity_set);
1341 cpu_affinity_set = NULL;
1342 cpu_affinity_setsize = 0;
1344 CPU_FREE(cpu_saved_affinity_set);
1345 cpu_saved_affinity_set = NULL;
1346 cpu_saved_affinity_setsize = 0;
1354 package_even = NULL;
1366 free(package_delta);
1368 thread_delta = NULL;
1370 package_delta = NULL;
1373 /**********************
1374 * Collectd functions *
1375 **********************/
1377 #define DO_OR_GOTO_ERR(something) \
1379 ret = (something); \
1384 static int setup_all_buffers(void) {
1387 DO_OR_GOTO_ERR(topology_probe());
1388 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1389 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1390 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1391 initialize_counters();
1392 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1393 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1402 static int turbostat_read(void) {
1406 if ((ret = setup_all_buffers()) < 0)
1410 if (for_all_proc_cpus(cpu_is_not_present)) {
1412 if ((ret = setup_all_buffers()) < 0)
1414 if (for_all_proc_cpus(cpu_is_not_present)) {
1415 ERROR("turbostat plugin: CPU appeared just after "
1421 /* Saving the scheduling affinity, as it will be modified by get_counters */
1422 if (sched_getaffinity(0, cpu_saved_affinity_setsize,
1423 cpu_saved_affinity_set) != 0) {
1424 ERROR("turbostat plugin: Unable to save the CPU affinity");
1429 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1431 time_even = cdtime();
1439 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1441 time_odd = cdtime();
1443 time_delta = time_odd - time_even;
1444 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1446 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1449 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1451 time_even = cdtime();
1453 time_delta = time_even - time_odd;
1454 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1456 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1462 * Let's restore the affinity
1463 * This might fail if the number of CPU changed, but we can't do anything in
1466 (void)sched_setaffinity(0, cpu_saved_affinity_setsize,
1467 cpu_saved_affinity_set);
1471 static int check_permissions(void) {
1473 if (getuid() == 0) {
1474 /* We have everything we need */
1476 #if !defined(HAVE_SYS_CAPABILITY_H) && !defined(CAP_SYS_RAWIO)
1478 ERROR("turbostat plugin: Initialization failed: this plugin "
1479 "requires collectd to run as root");
1482 #else /* HAVE_SYS_CAPABILITY_H && CAP_SYS_RAWIO */
1487 if (check_capability(CAP_SYS_RAWIO) != 0) {
1488 WARNING("turbostat plugin: Collectd doesn't have the "
1489 "CAP_SYS_RAWIO capability. If you don't want to run "
1490 "collectd as root, try running \"setcap "
1491 "cap_sys_rawio=ep\" on collectd binary");
1495 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1496 WARNING("turbostat plugin: Collectd cannot open "
1497 "/dev/cpu/0/msr. If you don't want to run collectd as "
1498 "root, you need to change the ownership (chown) and "
1499 "permissions on /dev/cpu/*/msr to allow such access");
1504 ERROR("turbostat plugin: Initialization failed: this plugin "
1505 "requires collectd to either to run as root or give "
1506 "collectd a special capability (CAP_SYS_RAWIO) and read "
1507 "access to /dev/cpu/*/msr (see previous warnings)");
1509 #endif /* HAVE_SYS_CAPABILITY_H && CAP_SYS_RAWIO */
1512 static int turbostat_init(void) {
1516 if (stat("/dev/cpu/0/msr", &sb)) {
1517 ERROR("turbostat plugin: Initialization failed: /dev/cpu/0/msr "
1518 "does not exist while the CPU supports MSR. You may be "
1519 "missing the corresponding kernel module, please try '# "
1524 DO_OR_GOTO_ERR(check_permissions());
1526 DO_OR_GOTO_ERR(probe_cpu());
1528 DO_OR_GOTO_ERR(setup_all_buffers());
1530 plugin_register_read(PLUGIN_NAME, turbostat_read);
1538 static int turbostat_config(const char *key, const char *value) {
1539 long unsigned int tmp_val;
1542 if (strcasecmp("CoreCstates", key) == 0) {
1543 tmp_val = strtoul(value, &end, 0);
1544 if (*end != '\0' || tmp_val > UINT_MAX) {
1545 ERROR("turbostat plugin: Invalid CoreCstates '%s'", value);
1548 config_core_cstate = (unsigned int)tmp_val;
1549 apply_config_core_cstate = 1;
1550 } else if (strcasecmp("PackageCstates", key) == 0) {
1551 tmp_val = strtoul(value, &end, 0);
1552 if (*end != '\0' || tmp_val > UINT_MAX) {
1553 ERROR("turbostat plugin: Invalid PackageCstates '%s'", value);
1556 config_pkg_cstate = (unsigned int)tmp_val;
1557 apply_config_pkg_cstate = 1;
1558 } else if (strcasecmp("SystemManagementInterrupt", key) == 0) {
1559 config_smi = IS_TRUE(value);
1560 apply_config_smi = 1;
1561 } else if (strcasecmp("DigitalTemperatureSensor", key) == 0) {
1562 config_dts = IS_TRUE(value);
1563 apply_config_dts = 1;
1564 } else if (strcasecmp("PackageThermalManagement", key) == 0) {
1565 config_ptm = IS_TRUE(value);
1566 apply_config_ptm = 1;
1567 } else if (strcasecmp("RunningAveragePowerLimit", key) == 0) {
1568 tmp_val = strtoul(value, &end, 0);
1569 if (*end != '\0' || tmp_val > UINT_MAX) {
1570 ERROR("turbostat plugin: Invalid RunningAveragePowerLimit '%s'", value);
1573 config_rapl = (unsigned int)tmp_val;
1574 apply_config_rapl = 1;
1575 } else if (strcasecmp("TCCActivationTemp", key) == 0) {
1576 tmp_val = strtoul(value, &end, 0);
1577 if (*end != '\0' || tmp_val > UINT_MAX) {
1578 ERROR("turbostat plugin: Invalid TCCActivationTemp '%s'", value);
1581 tcc_activation_temp = (unsigned int)tmp_val;
1583 ERROR("turbostat plugin: Invalid configuration option '%s'", key);
1589 void module_register(void) {
1590 plugin_register_init(PLUGIN_NAME, turbostat_init);
1591 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys,