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:
41 #include "utils_time.h"
43 #include <asm/msr-index.h>
48 #include <sys/types.h>
51 #include <sys/resource.h>
60 #include <sys/capability.h>
62 #define PLUGIN_NAME "turbostat"
65 * This tool uses the Model-Specific Registers (MSRs) present on Intel processors.
66 * The general description each of these registers, depending on the architecture,
67 * can be found in the IntelĀ® 64 and IA-32 Architectures Software Developer Manual,
68 * Volume 3 Chapter 35.
72 * If set, aperf_mperf_unstable disables a/mperf based stats.
73 * This includes: C0 & C1 states, frequency
75 * This value is automatically set if mperf or aperf go backward
77 static _Bool aperf_mperf_unstable;
80 * Bitmask of the list of core C states supported by the processor.
81 * Currently supported C-states (by this plugin): 3, 6, 7
83 static unsigned int do_core_cstate;
86 * Bitmask of the list of pacages C states supported by the processor.
87 * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
89 static unsigned int do_pkg_cstate;
92 * Boolean indicating if the processor supports 'Digital temperature sensor'
93 * This feature enables the monitoring of the temperature of each core
95 * This feature has two limitations:
96 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
97 * - Temperatures above the tcc_activation_temp are not recorded
102 * Boolean indicating if the processor supports 'Package thermal management'
103 * This feature allows the monitoring of the temperature of each package
105 * This feature has two limitations:
106 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
107 * - Temperatures above the tcc_activation_temp are not recorded
112 * Thermal Control Circuit Activation Temperature as configured by the user.
113 * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
114 * and should only be used if the automated detection fails.
116 static unsigned int tcc_activation_temp;
118 static unsigned int do_rapl;
119 static double rapl_energy_units;
121 #define RAPL_PKG (1 << 0)
122 /* 0x610 MSR_PKG_POWER_LIMIT */
123 /* 0x611 MSR_PKG_ENERGY_STATUS */
124 #define RAPL_PKG_PERF_STATUS (1 << 1)
125 /* 0x613 MSR_PKG_PERF_STATUS */
126 #define RAPL_PKG_POWER_INFO (1 << 2)
127 /* 0x614 MSR_PKG_POWER_INFO */
129 #define RAPL_DRAM (1 << 3)
130 /* 0x618 MSR_DRAM_POWER_LIMIT */
131 /* 0x619 MSR_DRAM_ENERGY_STATUS */
132 /* 0x61c MSR_DRAM_POWER_INFO */
133 #define RAPL_DRAM_PERF_STATUS (1 << 4)
134 /* 0x61b MSR_DRAM_PERF_STATUS */
136 #define RAPL_CORES (1 << 5)
137 /* 0x638 MSR_PP0_POWER_LIMIT */
138 /* 0x639 MSR_PP0_ENERGY_STATUS */
139 #define RAPL_CORE_POLICY (1 << 6)
140 /* 0x63a MSR_PP0_POLICY */
143 #define RAPL_GFX (1 << 7)
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, cpu_saved_affinity_setsize;
152 static struct thread_data {
153 unsigned long long tsc;
154 unsigned long long aperf;
155 unsigned long long mperf;
156 unsigned long long c1;
157 unsigned int smi_count;
160 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
161 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
162 } *thread_delta, *thread_even, *thread_odd;
164 static struct core_data {
165 unsigned long long c3;
166 unsigned long long c6;
167 unsigned long long c7;
168 unsigned int core_temp_c;
169 unsigned int core_id;
170 } *core_delta, *core_even, *core_odd;
172 static struct pkg_data {
173 unsigned long long pc2;
174 unsigned long long pc3;
175 unsigned long long pc6;
176 unsigned long long pc7;
177 unsigned long long pc8;
178 unsigned long long pc9;
179 unsigned long long pc10;
180 unsigned int package_id;
181 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
182 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
183 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
184 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
185 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
186 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
187 unsigned int tcc_activation_temp;
188 unsigned int pkg_temp_c;
189 } *package_delta, *package_even, *package_odd;
191 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
192 #define ODD_COUNTERS thread_odd, core_odd, package_odd
193 #define EVEN_COUNTERS thread_even, core_even, package_even
194 static _Bool is_even = 1;
196 static _Bool allocated = 0;
197 static _Bool initialized = 0;
199 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
201 (pkg_no) * topology.num_cores * topology.num_threads + \
202 (core_no) * topology.num_threads + \
204 #define GET_CORE(core_base, core_no, pkg_no) \
206 (pkg_no) * topology.num_cores + \
208 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
210 struct cpu_topology {
213 _Bool first_core_in_package;
214 _Bool first_thread_in_core;
217 static struct topology {
222 struct cpu_topology *cpus;
225 static cdtime_t time_even, time_odd, time_delta;
227 static const char *config_keys[] =
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 1
241 static int __attribute__((warn_unused_result))
242 open_msr(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)
280 retval = pread(fd, msr, sizeof *msr, offset);
282 if (retval != sizeof *msr) {
283 ERROR("Turbostat plugin: MSR offset 0x%llx read failed",
284 (unsigned long long)offset);
291 * Open a MSR device for reading, read the value asked for and close it.
292 * This call will not affect the scheduling affinity of this thread.
294 static int __attribute__((warn_unused_result))
295 get_msr(int cpu, off_t offset, unsigned long long *msr)
300 fd = open_msr(cpu, 0);
303 retval = read_msr(fd, offset, msr);
309 /********************************
310 * Raw data acquisition (1 CPU) *
311 ********************************/
314 * Read every data avalaible for a single CPU
316 * Core data is shared for all threads in one core: extracted only for the first thread
317 * Package data is shared for all core in one package: extracted only for the first thread of the first core
319 * Side effect: migrates to the targeted CPU
321 static int __attribute__((warn_unused_result))
322 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
325 unsigned long long msr;
329 msr_fd = open_msr(cpu, 1);
333 #define READ_MSR(msr, dst) \
335 if (read_msr(msr_fd, msr, dst)) { \
336 ERROR("Turbostat plugin: Unable to read " #msr); \
342 READ_MSR(MSR_IA32_TSC, &t->tsc);
344 READ_MSR(MSR_IA32_APERF, &t->aperf);
345 READ_MSR(MSR_IA32_MPERF, &t->mperf);
347 READ_MSR(MSR_SMI_COUNT, &msr);
348 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;
405 if (do_rapl & RAPL_PKG_PERF_STATUS) {
406 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
407 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
409 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
410 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
411 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
414 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
415 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
424 /**********************************
425 * Evaluating the changes (1 CPU) *
426 **********************************/
429 * Do delta = new - old on 32bits cyclique intergers
431 #define DELTA_WRAP32(delta, new, old) \
435 delta = 0x100000000 + new - old; \
439 * Extract the evolution old->new in delta at a package level
440 * (some are not new-delta, e.g. temperature)
443 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
445 delta->pc2 = new->pc2 - old->pc2;
446 delta->pc3 = new->pc3 - old->pc3;
447 delta->pc6 = new->pc6 - old->pc6;
448 delta->pc7 = new->pc7 - old->pc7;
449 delta->pc8 = new->pc8 - old->pc8;
450 delta->pc9 = new->pc9 - old->pc9;
451 delta->pc10 = new->pc10 - old->pc10;
452 delta->pkg_temp_c = new->pkg_temp_c;
454 DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
455 DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
456 DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
457 DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
458 DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
459 DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
463 * Extract the evolution old->new in delta at a core level
464 * (some are not new-delta, e.g. temperature)
467 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
469 delta->c3 = new->c3 - old->c3;
470 delta->c6 = new->c6 - old->c6;
471 delta->c7 = new->c7 - old->c7;
472 delta->core_temp_c = new->core_temp_c;
476 * Extract the evolution old->new in delta at a package level
477 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
479 static inline int __attribute__((warn_unused_result))
480 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
481 const struct core_data *core_delta)
483 delta->tsc = new->tsc - old->tsc;
485 /* check for TSC < 1 Mcycles over interval */
486 if (delta->tsc < (1000 * 1000)) {
487 WARNING("Turbostat plugin: Insanely slow TSC rate, TSC stops "
488 "in idle? You can disable all c-states by booting with"
489 " 'idle=poll' or just the deep ones with"
490 " 'processor.max_cstate=1'");
494 delta->c1 = new->c1 - old->c1;
496 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
497 delta->aperf = new->aperf - old->aperf;
498 delta->mperf = new->mperf - old->mperf;
500 if (!aperf_mperf_unstable) {
501 WARNING("Turbostat plugin: APERF or MPERF went "
502 "backwards. Frequency results do not cover "
503 "the entire interval. Fix this by running "
504 "Linux-2.6.30 or later.");
506 aperf_mperf_unstable = 1;
511 * As counter collection is not atomic,
512 * it is possible for mperf's non-halted cycles + idle states
513 * to exceed TSC's all cycles: show c1 = 0% in that case.
515 if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
518 /* normal case, derive c1 */
519 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
520 - core_delta->c6 - core_delta->c7;
523 if (delta->mperf == 0) {
524 WARNING("Turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
525 delta->mperf = 1; /* divide by 0 protection */
528 delta->smi_count = new->smi_count - old->smi_count;
533 /**********************************
534 * Submitting the results (1 CPU) *
535 **********************************/
538 * Submit one gauge value
541 turbostat_submit (const char *plugin_instance,
542 const char *type, const char *type_instance,
545 value_list_t vl = VALUE_LIST_INIT;
551 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
552 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
553 if (plugin_instance != NULL)
554 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
555 sstrncpy (vl.type, type, sizeof (vl.type));
556 if (type_instance != NULL)
557 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
559 plugin_dispatch_values (&vl);
563 * Submit every data for a single CPU
565 * Core data is shared for all threads in one core: submitted only for the first thread
566 * Package data is shared for all core in one package: submitted only for the first thread of the first core
569 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
571 char name[DATA_MAX_NAME_LEN];
572 double interval_float;
574 interval_float = CDTIME_T_TO_DOUBLE(time_delta);
576 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
578 if (!aperf_mperf_unstable)
579 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
580 if (!aperf_mperf_unstable)
581 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
583 turbostat_submit("Average", "frequency", name, 1.0 / 1000000 * t->aperf / interval_float);
585 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
586 turbostat_submit("Buzy", "frequency", name, 1.0 * t->tsc / 1000000 * t->aperf / t->mperf / interval_float);
589 turbostat_submit(NULL, "current", name, t->smi_count);
591 /* submit per-core data only for 1st thread in core */
592 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
595 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
597 if (do_core_cstate & (1 << 3))
598 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
599 if (do_core_cstate & (1 << 6))
600 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
601 if (do_core_cstate & (1 << 7))
602 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
605 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
607 /* submit per-package data only for 1st core in package */
608 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
611 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
614 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
616 if (do_pkg_cstate & (1 << 2))
617 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
618 if (do_pkg_cstate & (1 << 3))
619 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
620 if (do_pkg_cstate & (1 << 6))
621 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
622 if (do_pkg_cstate & (1 << 7))
623 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
624 if (do_pkg_cstate & (1 << 8))
625 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
626 if (do_pkg_cstate & (1 << 9))
627 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
628 if (do_pkg_cstate & (1 << 10))
629 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
632 if (do_rapl & RAPL_PKG)
633 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
634 if (do_rapl & RAPL_CORES)
635 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
636 if (do_rapl & RAPL_GFX)
637 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
638 if (do_rapl & RAPL_DRAM)
639 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
646 /**********************************
647 * Looping function over all CPUs *
648 **********************************/
651 * Check if a given cpu id is in our compiled list of existing CPUs
654 cpu_is_not_present(int cpu)
656 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
660 * Loop on all CPUs in topological order
662 * Skip non-present cpus
663 * Return the error code at the first error or 0
665 static int __attribute__((warn_unused_result))
666 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
667 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
669 int retval, pkg_no, core_no, thread_no;
671 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
672 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
673 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
674 struct thread_data *t;
678 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
680 if (cpu_is_not_present(t->cpu_id))
683 c = GET_CORE(core_base, core_no, pkg_no);
684 p = GET_PKG(pkg_base, pkg_no);
686 retval = func(t, c, p);
696 * Dedicated loop: Extract every data evolution for all CPU
698 * Skip non-present cpus
699 * Return the error code at the first error or 0
701 * Core data is shared for all threads in one core: extracted only for the first thread
702 * Package data is shared for all core in one package: extracted only for the first thread of the first core
704 static int __attribute__((warn_unused_result))
705 for_all_cpus_delta(const struct thread_data *thread_new_base, const struct core_data *core_new_base, const struct pkg_data *pkg_new_base,
706 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
708 int retval, pkg_no, core_no, thread_no;
710 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
711 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
712 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
713 struct thread_data *t_delta;
714 const struct thread_data *t_old, *t_new;
715 struct core_data *c_delta;
717 /* Get correct pointers for threads */
718 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
719 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
720 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
722 /* Skip threads that disappeared */
723 if (cpu_is_not_present(t_delta->cpu_id))
726 /* c_delta is always required for delta_thread */
727 c_delta = GET_CORE(core_delta, core_no, pkg_no);
729 /* calculate core delta only for 1st thread in core */
730 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
731 const struct core_data *c_old, *c_new;
733 c_new = GET_CORE(core_new_base, core_no, pkg_no);
734 c_old = GET_CORE(core_old_base, core_no, pkg_no);
736 delta_core(c_delta, c_new, c_old);
739 /* Always calculate thread delta */
740 retval = delta_thread(t_delta, t_new, t_old, c_delta);
744 /* calculate package delta only for 1st core in package */
745 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
746 struct pkg_data *p_delta;
747 const struct pkg_data *p_old, *p_new;
749 p_delta = GET_PKG(package_delta, pkg_no);
750 p_new = GET_PKG(pkg_new_base, pkg_no);
751 p_old = GET_PKG(pkg_old_base, pkg_no);
753 delta_package(p_delta, p_new, p_old);
767 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
768 * the Thermal Control Circuit (TCC) activates.
769 * This is usually equal to tjMax.
771 * Older processors do not have this MSR, so there we guess,
772 * but also allow conficuration over-ride with "TCCActivationTemp".
774 * Several MSR temperature values are in units of degrees-C
775 * below this value, including the Digital Thermal Sensor (DTS),
776 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
778 static int __attribute__((warn_unused_result))
779 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
781 unsigned long long msr;
782 unsigned int target_c_local;
784 /* tcc_activation_temp is used only for dts or ptm */
785 if (!(do_dts || do_ptm))
788 /* this is a per-package concept */
789 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
792 if (tcc_activation_temp != 0) {
793 p->tcc_activation_temp = tcc_activation_temp;
797 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
800 target_c_local = (msr >> 16) & 0xFF;
805 p->tcc_activation_temp = target_c_local;
810 p->tcc_activation_temp = TJMAX_DEFAULT;
811 WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
812 " Please use TCCActivationTemp to specify it.",
813 t->cpu_id, p->tcc_activation_temp);
819 * Identify the functionality of the CPU
821 static int __attribute__((warn_unused_result))
824 unsigned int eax, ebx, ecx, edx, max_level;
825 unsigned int fms, family, model;
828 * - EAX: Maximum Input Value for Basic CPUID Information
829 * - EBX: "Genu" (0x756e6547)
830 * - EDX: "ineI" (0x49656e69)
831 * - ECX: "ntel" (0x6c65746e)
833 max_level = ebx = ecx = edx = 0;
834 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
835 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
836 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
841 * - EAX: Version Information: Type, Family, Model, and Stepping ID
844 * + 12-13: Processor type
845 * + 16-19: Extended Model ID
846 * + 20-27: Extended Family ID
847 * - EDX: Feature Information:
848 * + 5: Support for MSR read/write operations
850 fms = ebx = ecx = edx = 0;
851 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
852 family = (fms >> 8) & 0xf;
853 model = (fms >> 4) & 0xf;
855 family += (fms >> 20) & 0xf;
856 if (family == 6 || family == 0xf)
857 model += ((fms >> 16) & 0xf) << 4;
858 if (!(edx & (1 << 5))) {
859 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
865 * - EAX: Maximum Input Value for Extended Function CPUID Information
867 * This allows us to verify if the CPUID(0x80000007) can be called
869 * This check is valid for both Intel and AMD.
871 max_level = ebx = ecx = edx = 0;
872 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
873 if (max_level < 0x80000007) {
874 ERROR("Turbostat plugin: Unsupported CPU (no invariant TSC, "
875 " Maximum Extended Function: 0x%x)", max_level);
882 * + 8: Invariant TSC available if set
884 * This check is valid for both Intel and AMD
886 eax = ebx = ecx = edx = 0;
887 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
888 if (!(edx & (1 << 8))) {
889 ERROR("Turbostat plugin: Unsupported CPU (No invariant TSC)");
896 * + 0: Digital temperature sensor is supported if set
897 * + 6: Package thermal management is supported if set
899 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
900 * + 3: The processor supports performance-energy bias preference if set.
901 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
903 * This check is valid for both Intel and AMD
905 eax = ebx = ecx = edx = 0;
906 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
907 do_dts = eax & (1 << 0);
908 do_ptm = eax & (1 << 6);
909 if (!(ecx & (1 << 0))) {
910 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
915 * Enable or disable C states depending on the model and family
922 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
930 do_core_cstate = (1 << 1) | (1 << 6);
931 do_pkg_cstate = (1 << 6);
934 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
935 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
936 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
937 case 0x2E: /* Nehalem-EX Xeon - Beckton */
938 do_core_cstate = (1 << 3) | (1 << 6);
939 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
942 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
943 case 0x2C: /* Westmere EP - Gulftown */
944 case 0x2F: /* Westmere-EX Xeon - Eagleton */
945 do_core_cstate = (1 << 3) | (1 << 6);
946 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
950 case 0x2D: /* SNB Xeon */
951 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
952 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
956 case 0x3E: /* IVB Xeon */
957 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
958 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
964 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
965 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
968 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
969 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
973 case 0x56: /* BDX-DE */
974 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
975 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
978 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
979 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
982 ERROR("Turbostat plugin: Unsupported CPU (family: %#x,"
983 " model: %#x)", family, model);
991 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
994 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
998 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1002 do_rapl = RAPL_PKG | RAPL_CORES;
1008 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
1009 "model: %#x)", family, model);
1014 unsigned long long msr;
1015 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1019 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1021 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1028 /********************
1029 * Topology Probing *
1030 ********************/
1033 * Read a single int from a file.
1035 static int __attribute__ ((format(printf,1,2)))
1036 parse_int_file(const char *fmt, ...)
1039 char path[PATH_MAX];
1043 va_start(args, fmt);
1044 len = vsnprintf(path, sizeof(path), fmt, args);
1046 if (len < 0 || len >= PATH_MAX) {
1047 ERROR("Turbostat plugin: path truncated: '%s'", path);
1051 filep = fopen(path, "r");
1053 ERROR("Turbostat plugin: Failed to open '%s'", path);
1056 if (fscanf(filep, "%d", &value) != 1) {
1057 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1065 get_threads_on_core(int cpu)
1073 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1074 filep = fopen(path, "r");
1076 ERROR("Turbostat plugin: Failed to open '%s'", path);
1081 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1082 * otherwinse 1 sibling (self).
1084 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1095 * run func(cpu) on every cpu in /proc/stat
1096 * return max_cpu number
1098 static int __attribute__((warn_unused_result))
1099 for_all_proc_cpus(int (func)(int))
1105 fp = fopen("/proc/stat", "r");
1107 ERROR("Turbostat plugin: Failed to open /proc/stat");
1111 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1113 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1119 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1123 retval = func(cpu_num);
1134 * Update the stored topology.max_cpu_id
1137 update_max_cpu_id(int cpu)
1139 if (topology.max_cpu_id < cpu)
1140 topology.max_cpu_id = cpu;
1145 mark_cpu_present(int cpu)
1147 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1151 static int __attribute__((warn_unused_result))
1152 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1153 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1155 ERROR("Turbostat plugin: Unable to allocate CPU state");
1158 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1159 CPU_ZERO_S(*size, *set);
1164 * Build a local representation of the cpu distribution
1166 static int __attribute__((warn_unused_result))
1171 int max_package_id, max_core_id, max_thread_id;
1172 max_package_id = max_core_id = max_thread_id = 0;
1174 /* Clean topology */
1175 free(topology.cpus);
1176 memset(&topology, 0, sizeof(topology));
1178 ret = for_all_proc_cpus(update_max_cpu_id);
1182 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1183 if (topology.cpus == NULL) {
1184 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1188 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1191 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1194 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1198 ret = for_all_proc_cpus(mark_cpu_present);
1204 * find max_core_id, max_package_id
1206 for (i = 0; i <= topology.max_cpu_id; ++i) {
1208 struct cpu_topology *cpu = &topology.cpus[i];
1210 if (cpu_is_not_present(i)) {
1211 WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1215 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1219 cpu->package_id = ret;
1220 if (cpu->package_id > max_package_id)
1221 max_package_id = cpu->package_id;
1223 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1228 if (cpu->core_id > max_core_id)
1229 max_core_id = cpu->core_id;
1230 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1234 cpu->first_core_in_package = 1;
1236 ret = get_threads_on_core(i);
1241 if (num_threads > max_thread_id)
1242 max_thread_id = num_threads;
1243 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1247 cpu->first_thread_in_core = 1;
1249 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1250 i, cpu->package_id, cpu->core_id);
1252 /* Num is max + 1 (need to count 0) */
1253 topology.num_packages = max_package_id + 1;
1254 topology.num_cores = max_core_id + 1;
1255 topology.num_threads = max_thread_id + 1;
1259 free(topology.cpus);
1264 /************************
1265 * Main alloc/init/free *
1266 ************************/
1269 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1272 int total_threads, total_cores;
1274 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1275 *threads = calloc(total_threads, sizeof(struct thread_data));
1276 if (*threads == NULL)
1279 for (i = 0; i < total_threads; ++i)
1280 (*threads)[i].cpu_id = -1;
1282 total_cores = topology.num_cores * topology.num_packages;
1283 *cores = calloc(total_cores, sizeof(struct core_data));
1285 goto err_clean_threads;
1287 for (i = 0; i < total_cores; ++i)
1288 (*cores)[i].core_id = -1;
1290 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1291 if (*packages == NULL)
1292 goto err_clean_cores;
1294 for (i = 0; i < topology.num_packages; i++)
1295 (*packages)[i].package_id = i;
1304 ERROR("Turbostat plugin: Failled to allocate memory for counters");
1309 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1310 struct pkg_data *pkg_base, int cpu_id)
1312 struct thread_data *t;
1313 struct core_data *c;
1315 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1317 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1318 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1319 p = GET_PKG(pkg_base, cpu->package_id);
1322 if (cpu->first_thread_in_core)
1323 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1324 if (cpu->first_core_in_package)
1325 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1327 c->core_id = cpu->core_id;
1328 p->package_id = cpu->package_id;
1332 initialize_counters(void)
1336 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1337 if (cpu_is_not_present(cpu_id))
1339 init_counter(EVEN_COUNTERS, cpu_id);
1340 init_counter(ODD_COUNTERS, cpu_id);
1341 init_counter(DELTA_COUNTERS, cpu_id);
1348 free_all_buffers(void)
1353 CPU_FREE(cpu_present_set);
1354 cpu_present_set = NULL;
1355 cpu_present_set = 0;
1357 CPU_FREE(cpu_affinity_set);
1358 cpu_affinity_set = NULL;
1359 cpu_affinity_setsize = 0;
1361 CPU_FREE(cpu_saved_affinity_set);
1362 cpu_saved_affinity_set = NULL;
1363 cpu_saved_affinity_setsize = 0;
1371 package_even = NULL;
1383 free(package_delta);
1385 thread_delta = NULL;
1387 package_delta = NULL;
1391 /**********************
1392 * Collectd functions *
1393 **********************/
1395 #define DO_OR_GOTO_ERR(something) \
1397 ret = (something); \
1402 static int setup_all_buffers(void)
1406 DO_OR_GOTO_ERR(topology_probe());
1407 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1408 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1409 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1410 initialize_counters();
1411 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1412 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1422 turbostat_read(void)
1427 if ((ret = setup_all_buffers()) < 0)
1431 if (for_all_proc_cpus(cpu_is_not_present)) {
1433 if ((ret = setup_all_buffers()) < 0)
1435 if (for_all_proc_cpus(cpu_is_not_present)) {
1436 ERROR("Turbostat plugin: CPU appeared just after "
1442 /* Saving the scheduling affinity, as it will be modified by get_counters */
1443 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1444 ERROR("Turbostat plugin: Unable to save the CPU affinity");
1449 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1451 time_even = cdtime();
1459 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1461 time_odd = cdtime();
1463 time_delta = time_odd - time_even;
1464 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1466 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1469 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1471 time_even = cdtime();
1473 time_delta = time_even - time_odd;
1474 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1476 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1482 * Let's restore the affinity
1483 * This might fail if the number of CPU changed, but we can't do anything in that case..
1485 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1490 check_permissions(void)
1492 struct __user_cap_header_struct cap_header_data;
1493 cap_user_header_t cap_header = &cap_header_data;
1494 struct __user_cap_data_struct cap_data_data;
1495 cap_user_data_t cap_data = &cap_data_data;
1498 if (getuid() == 0) {
1499 /* We have everything we need */
1503 /* check for CAP_SYS_RAWIO */
1504 cap_header->pid = getpid();
1505 cap_header->version = _LINUX_CAPABILITY_VERSION;
1506 if (capget(cap_header, cap_data) < 0) {
1507 ERROR("Turbostat plugin: capget failed");
1511 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1512 WARNING("Turbostat plugin: Collectd doesn't have the "
1513 "CAP_SYS_RAWIO capability. If you don't want to run "
1514 "collectd as root, try running \"setcap "
1515 "cap_sys_rawio=ep\" on collectd binary");
1519 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1520 WARNING("Turbostat plugin: Collectd cannot open"
1521 "/dev/cpu/0/msr. If you don't want to run collectd as "
1522 "root, you need to change the ownership (chown) and "
1523 "permissions on /dev/cpu/*/msr to allow such access");
1528 ERROR("Turbostat plugin: Initialization failed: this plugin "
1529 "requires collectd to either to run as root or give "
1530 "collectd a special capability (CAP_SYS_RAWIO) and read "
1531 "access to /dev/cpu/*/msr (see previous warnings)");
1536 turbostat_init(void)
1541 if (stat("/dev/cpu/0/msr", &sb)) {
1542 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1543 " does not exist while the CPU supports MSR. You may be "
1544 "missing the corresponding kernel module, please try '# "
1549 DO_OR_GOTO_ERR(check_permissions());
1551 DO_OR_GOTO_ERR(probe_cpu());
1553 DO_OR_GOTO_ERR(setup_all_buffers());
1555 plugin_register_read(PLUGIN_NAME, turbostat_read);
1564 turbostat_config(const char *key, const char *value)
1566 long unsigned int tmp_val;
1569 if (strcasecmp("TCCActivationTemp", key) == 0) {
1570 tmp_val = strtoul(value, &end, 0);
1571 if (*end != '\0' || tmp_val > UINT_MAX) {
1572 ERROR("Turbostat plugin: Invalid TCCActivationTemp '%s'",
1576 tcc_activation_temp = (unsigned int) tmp_val;
1578 ERROR("Turbostat plugin: Invalid configuration option '%s'",
1585 void module_register(void)
1587 plugin_register_init(PLUGIN_NAME, turbostat_init);
1588 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);