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>
61 #define PLUGIN_NAME "turbostat"
64 * This tool uses the Model-Specific Registers (MSRs) present on Intel processors.
65 * The general description each of these registers, depending on the architecture,
66 * can be found in the IntelĀ® 64 and IA-32 Architectures Software Developer Manual,
67 * Volume 3 Chapter 35.
71 * If set, aperf_mperf_unstable disables a/mperf based stats.
72 * This includes: C0 & C1 states, frequency
74 * This value is automatically set if mperf or aperf go backward
76 static _Bool aperf_mperf_unstable;
79 * Bitmask of the list of core C states supported by the processor.
80 * Currently supported C-states (by this plugin): 3, 6, 7
82 static unsigned int do_core_cstate;
85 * Bitmask of the list of pacages C states supported by the processor.
86 * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
88 static unsigned int do_pkg_cstate;
91 * Boolean indicating if the processor supports 'Digital temperature sensor'
92 * This feature enables the monitoring of the temperature of each core
94 * This feature has two limitations:
95 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
96 * - Temperatures above the tcc_activation_temp are not recorded
101 * Boolean indicating if the processor supports 'Package thermal management'
102 * This feature allows the monitoring of the temperature of each package
104 * This feature has two limitations:
105 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
106 * - Temperatures above the tcc_activation_temp are not recorded
111 * Thermal Control Circuit Activation Temperature as configured by the user.
112 * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
113 * and should only be used if the automated detection fails.
115 static unsigned int tcc_activation_temp;
117 static unsigned int do_rapl;
118 static double rapl_energy_units;
120 #define RAPL_PKG (1 << 0)
121 /* 0x610 MSR_PKG_POWER_LIMIT */
122 /* 0x611 MSR_PKG_ENERGY_STATUS */
123 #define RAPL_PKG_PERF_STATUS (1 << 1)
124 /* 0x613 MSR_PKG_PERF_STATUS */
125 #define RAPL_PKG_POWER_INFO (1 << 2)
126 /* 0x614 MSR_PKG_POWER_INFO */
128 #define RAPL_DRAM (1 << 3)
129 /* 0x618 MSR_DRAM_POWER_LIMIT */
130 /* 0x619 MSR_DRAM_ENERGY_STATUS */
131 /* 0x61c MSR_DRAM_POWER_INFO */
132 #define RAPL_DRAM_PERF_STATUS (1 << 4)
133 /* 0x61b MSR_DRAM_PERF_STATUS */
135 #define RAPL_CORES (1 << 5)
136 /* 0x638 MSR_PP0_POWER_LIMIT */
137 /* 0x639 MSR_PP0_ENERGY_STATUS */
138 #define RAPL_CORE_POLICY (1 << 6)
139 /* 0x63a MSR_PP0_POLICY */
142 #define RAPL_GFX (1 << 7)
143 /* 0x640 MSR_PP1_POWER_LIMIT */
144 /* 0x641 MSR_PP1_ENERGY_STATUS */
145 /* 0x642 MSR_PP1_POLICY */
146 #define TJMAX_DEFAULT 100
148 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
149 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
152 unsigned long long tsc;
153 unsigned long long aperf;
154 unsigned long long mperf;
155 unsigned long long c1;
156 unsigned int smi_count;
159 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
160 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
161 } *thread_delta, *thread_even, *thread_odd;
164 unsigned long long c3;
165 unsigned long long c6;
166 unsigned long long c7;
167 unsigned int core_temp_c;
168 unsigned int core_id;
169 } *core_delta, *core_even, *core_odd;
172 unsigned long long pc2;
173 unsigned long long pc3;
174 unsigned long long pc6;
175 unsigned long long pc7;
176 unsigned long long pc8;
177 unsigned long long pc9;
178 unsigned long long pc10;
179 unsigned int package_id;
180 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
181 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
182 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
183 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
184 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
185 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_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) \
200 (pkg_no) * topology.num_cores * topology.num_threads + \
201 (core_no) * topology.num_threads + \
203 #define GET_CORE(core_base, core_no, pkg_no) \
205 (pkg_no) * topology.num_cores + \
207 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
209 struct cpu_topology {
212 _Bool first_core_in_package;
213 _Bool first_thread_in_core;
221 struct cpu_topology *cpus;
224 cdtime_t time_even, time_odd, time_delta;
226 static const char *config_keys[] =
230 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
232 /*****************************
233 * MSR Manipulation helpers *
234 *****************************/
237 * Open a MSR device for reading
238 * Can change the scheduling affinity of the current process if multiple_read is 1
240 static int __attribute__((warn_unused_result))
241 open_msr(int cpu, _Bool multiple_read)
247 * If we need to do multiple read, let's migrate to the CPU
248 * Otherwise, we would lose time calling functions on another CPU
250 * If we are not yet initialized (cpu_affinity_setsize = 0),
251 * we need to skip this optimisation.
253 if (multiple_read && cpu_affinity_setsize) {
254 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
255 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
256 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
257 ERROR("Turbostat plugin: Could not migrate to CPU %d", cpu);
262 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
263 fd = open(pathname, O_RDONLY);
265 ERROR("Turbostat plugin: failed to open %s", pathname);
272 * Read a single MSR from an open file descriptor
274 static int __attribute__((warn_unused_result))
275 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 int __attribute__((warn_unused_result))
294 get_msr(int cpu, off_t offset, unsigned long long *msr)
299 fd = open_msr(cpu, 0);
302 retval = read_msr(fd, offset, msr);
308 /********************************
309 * Raw data acquisition (1 CPU) *
310 ********************************/
313 * Read every data avalaible for a single CPU
315 * Core data is shared for all threads in one core: extracted only for the first thread
316 * Package data is shared for all core in one package: extracted only for the 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)
324 unsigned long long msr;
328 msr_fd = open_msr(cpu, 1);
332 #define READ_MSR(msr, dst) \
334 if (read_msr(msr_fd, msr, dst)) { \
335 ERROR("Turbostat plugin: Unable to read " #msr); \
341 READ_MSR(MSR_IA32_TSC, &t->tsc);
343 READ_MSR(MSR_IA32_APERF, &t->aperf);
344 READ_MSR(MSR_IA32_MPERF, &t->mperf);
346 READ_MSR(MSR_SMI_COUNT, &msr);
347 t->smi_count = msr & 0xFFFFFFFF;
349 /* collect core counters only for 1st thread in core */
350 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
355 if (do_core_cstate & (1 << 3))
356 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
357 if (do_core_cstate & (1 << 6))
358 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
359 if (do_core_cstate & (1 << 7))
360 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
363 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
364 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
367 /* collect package counters only for 1st core in package */
368 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
373 if (do_pkg_cstate & (1 << 2))
374 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
375 if (do_pkg_cstate & (1 << 3))
376 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
377 if (do_pkg_cstate & (1 << 6))
378 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
379 if (do_pkg_cstate & (1 << 7))
380 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
381 if (do_pkg_cstate & (1 << 8))
382 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
383 if (do_pkg_cstate & (1 << 9))
384 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
385 if (do_pkg_cstate & (1 << 10))
386 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
388 if (do_rapl & RAPL_PKG) {
389 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
390 p->energy_pkg = msr & 0xFFFFFFFF;
392 if (do_rapl & RAPL_CORES) {
393 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
394 p->energy_cores = msr & 0xFFFFFFFF;
396 if (do_rapl & RAPL_DRAM) {
397 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
398 p->energy_dram = msr & 0xFFFFFFFF;
400 if (do_rapl & RAPL_GFX) {
401 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
402 p->energy_gfx = msr & 0xFFFFFFFF;
404 if (do_rapl & RAPL_PKG_PERF_STATUS) {
405 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
406 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
408 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
409 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
410 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
413 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
414 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
423 /**********************************
424 * Evaluating the changes (1 CPU) *
425 **********************************/
428 * Do delta = new - old on 32bits cyclique intergers
430 #define DELTA_WRAP32(delta, new, old) \
434 delta = 0x100000000 + new - old; \
438 * Extract the evolution old->new in delta at a package level
439 * (some are not new-delta, e.g. temperature)
442 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
444 delta->pc2 = new->pc2 - old->pc2;
445 delta->pc3 = new->pc3 - old->pc3;
446 delta->pc6 = new->pc6 - old->pc6;
447 delta->pc7 = new->pc7 - old->pc7;
448 delta->pc8 = new->pc8 - old->pc8;
449 delta->pc9 = new->pc9 - old->pc9;
450 delta->pc10 = new->pc10 - old->pc10;
451 delta->pkg_temp_c = new->pkg_temp_c;
453 DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
454 DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
455 DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
456 DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
457 DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
458 DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
462 * Extract the evolution old->new in delta at a core level
463 * (some are not new-delta, e.g. temperature)
466 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
468 delta->c3 = new->c3 - old->c3;
469 delta->c6 = new->c6 - old->c6;
470 delta->c7 = new->c7 - old->c7;
471 delta->core_temp_c = new->core_temp_c;
475 * Extract the evolution old->new in delta at a package level
476 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
478 static inline int __attribute__((warn_unused_result))
479 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
480 const struct core_data *core_delta)
482 delta->tsc = new->tsc - old->tsc;
484 /* check for TSC < 1 Mcycles over interval */
485 if (delta->tsc < (1000 * 1000)) {
486 WARNING("Turbostat plugin: Insanely slow TSC rate, TSC stops "
487 "in idle? You can disable all c-states by booting with"
488 " 'idle=poll' or just the deep ones with"
489 " 'processor.max_cstate=1'");
493 delta->c1 = new->c1 - old->c1;
495 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
496 delta->aperf = new->aperf - old->aperf;
497 delta->mperf = new->mperf - old->mperf;
499 if (!aperf_mperf_unstable) {
500 WARNING("Turbostat plugin: APERF or MPERF went "
501 "backwards. Frequency results do not cover "
502 "the entire interval. Fix this by running "
503 "Linux-2.6.30 or later.");
505 aperf_mperf_unstable = 1;
510 * As counter collection is not atomic,
511 * it is possible for mperf's non-halted cycles + idle states
512 * to exceed TSC's all cycles: show c1 = 0% in that case.
514 if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
517 /* normal case, derive c1 */
518 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
519 - core_delta->c6 - core_delta->c7;
522 if (delta->mperf == 0) {
523 WARNING("Turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
524 delta->mperf = 1; /* divide by 0 protection */
527 delta->smi_count = new->smi_count - old->smi_count;
532 /**********************************
533 * Submitting the results (1 CPU) *
534 **********************************/
537 * Submit one gauge value
540 turbostat_submit (const char *plugin_instance,
541 const char *type, const char *type_instance,
544 value_list_t vl = VALUE_LIST_INIT;
550 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
551 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
552 if (plugin_instance != NULL)
553 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
554 sstrncpy (vl.type, type, sizeof (vl.type));
555 if (type_instance != NULL)
556 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
558 plugin_dispatch_values (&vl);
562 * Submit every data for a single CPU
564 * Core data is shared for all threads in one core: submitted only for the first thread
565 * Package data is shared for all core in one package: submitted only for the first thread of the first core
568 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
570 char name[DATA_MAX_NAME_LEN];
571 double interval_float;
573 interval_float = CDTIME_T_TO_DOUBLE(time_delta);
575 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
577 if (!aperf_mperf_unstable)
578 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
579 if (!aperf_mperf_unstable)
580 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
583 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
584 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
587 turbostat_submit(NULL, "current", name, t->smi_count);
589 /* submit per-core data only for 1st thread in core */
590 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
593 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
595 if (do_core_cstate & (1 << 3))
596 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
597 if (do_core_cstate & (1 << 6))
598 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
599 if (do_core_cstate & (1 << 7))
600 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
603 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
605 /* submit per-package data only for 1st core in package */
606 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
609 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
612 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
614 if (do_pkg_cstate & (1 << 2))
615 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
616 if (do_pkg_cstate & (1 << 3))
617 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
618 if (do_pkg_cstate & (1 << 6))
619 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
620 if (do_pkg_cstate & (1 << 7))
621 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
622 if (do_pkg_cstate & (1 << 8))
623 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
624 if (do_pkg_cstate & (1 << 9))
625 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
626 if (do_pkg_cstate & (1 << 10))
627 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
630 if (do_rapl & RAPL_PKG)
631 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
632 if (do_rapl & RAPL_CORES)
633 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
634 if (do_rapl & RAPL_GFX)
635 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
636 if (do_rapl & RAPL_DRAM)
637 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
644 /**********************************
645 * Looping function over all CPUs *
646 **********************************/
649 * Check if a given cpu id is in our compiled list of existing CPUs
652 cpu_is_not_present(int cpu)
654 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
658 * Loop on all CPUs in topological order
660 * Skip non-present cpus
661 * Return the error code at the first error or 0
663 static int __attribute__((warn_unused_result))
664 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
665 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
667 int retval, pkg_no, core_no, thread_no;
669 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
670 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
671 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
672 struct thread_data *t;
676 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
678 if (cpu_is_not_present(t->cpu_id))
681 c = GET_CORE(core_base, core_no, pkg_no);
682 p = GET_PKG(pkg_base, pkg_no);
684 retval = func(t, c, p);
694 * Dedicated loop: Extract every data evolution for all CPU
696 * Skip non-present cpus
697 * Return the error code at the first error or 0
699 * Core data is shared for all threads in one core: extracted only for the first thread
700 * Package data is shared for all core in one package: extracted only for the first thread of the first core
702 static int __attribute__((warn_unused_result))
703 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,
704 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
706 int retval, pkg_no, core_no, thread_no;
708 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
709 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
710 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
711 struct thread_data *t_delta;
712 const struct thread_data *t_old, *t_new;
713 struct core_data *c_delta;
715 /* Get correct pointers for threads */
716 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
717 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
718 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
720 /* Skip threads that disappeared */
721 if (cpu_is_not_present(t_delta->cpu_id))
724 /* c_delta is always required for delta_thread */
725 c_delta = GET_CORE(core_delta, core_no, pkg_no);
727 /* calculate core delta only for 1st thread in core */
728 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
729 const struct core_data *c_old, *c_new;
731 c_new = GET_CORE(core_new_base, core_no, pkg_no);
732 c_old = GET_CORE(core_old_base, core_no, pkg_no);
734 delta_core(c_delta, c_new, c_old);
737 /* Always calculate thread delta */
738 retval = delta_thread(t_delta, t_new, t_old, c_delta);
742 /* calculate package delta only for 1st core in package */
743 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
744 struct pkg_data *p_delta;
745 const struct pkg_data *p_old, *p_new;
747 p_delta = GET_PKG(package_delta, pkg_no);
748 p_new = GET_PKG(pkg_new_base, pkg_no);
749 p_old = GET_PKG(pkg_old_base, pkg_no);
751 delta_package(p_delta, p_new, p_old);
765 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
766 * the Thermal Control Circuit (TCC) activates.
767 * This is usually equal to tjMax.
769 * Older processors do not have this MSR, so there we guess,
770 * but also allow conficuration over-ride with "TCCActivationTemp".
772 * Several MSR temperature values are in units of degrees-C
773 * below this value, including the Digital Thermal Sensor (DTS),
774 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
776 static int __attribute__((warn_unused_result))
777 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
779 unsigned long long msr;
780 unsigned int target_c_local;
782 /* tcc_activation_temp is used only for dts or ptm */
783 if (!(do_dts || do_ptm))
786 /* this is a per-package concept */
787 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
790 if (tcc_activation_temp != 0) {
791 p->tcc_activation_temp = tcc_activation_temp;
795 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
798 target_c_local = (msr >> 16) & 0xFF;
803 p->tcc_activation_temp = target_c_local;
808 p->tcc_activation_temp = TJMAX_DEFAULT;
809 WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
810 " Please use TCCActivationTemp to specify it.",
811 t->cpu_id, p->tcc_activation_temp);
817 * Identify the functionality of the CPU
819 static int __attribute__((warn_unused_result))
822 unsigned int eax, ebx, ecx, edx, max_level;
823 unsigned int fms, family, model;
826 * - EAX: Maximum Input Value for Basic CPUID Information
827 * - EBX: "Genu" (0x756e6547)
828 * - EDX: "ineI" (0x49656e69)
829 * - ECX: "ntel" (0x6c65746e)
831 max_level = ebx = ecx = edx = 0;
832 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
833 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
834 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
839 * - EAX: Version Information: Type, Family, Model, and Stepping ID
842 * + 12-13: Processor type
843 * + 16-19: Extended Model ID
844 * + 20-27: Extended Family ID
845 * - EDX: Feature Information:
846 * + 5: Support for MSR read/write operations
848 fms = ebx = ecx = edx = 0;
849 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
850 family = (fms >> 8) & 0xf;
851 model = (fms >> 4) & 0xf;
853 family += (fms >> 20) & 0xf;
854 if (family == 6 || family == 0xf)
855 model += ((fms >> 16) & 0xf) << 4;
856 if (!(edx & (1 << 5))) {
857 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
863 * - EAX: Maximum Input Value for Extended Function CPUID Information
865 * This allows us to verify if the CPUID(0x80000007) can be called
867 * This check is valid for both Intel and AMD.
869 max_level = ebx = ecx = edx = 0;
870 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
871 if (max_level < 0x80000007) {
872 ERROR("Turbostat plugin: Unsupported CPU (no invariant TSC, "
873 " Maximum Extended Function: 0x%x)", max_level);
880 * + 8: Invariant TSC available if set
882 * This check is valid for both Intel and AMD
884 eax = ebx = ecx = edx = 0;
885 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
886 if (!(edx & (1 << 8))) {
887 ERROR("Turbostat plugin: Unsupported CPU (No invariant TSC)");
894 * + 0: Digital temperature sensor is supported if set
895 * + 6: Package thermal management is supported if set
897 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
898 * + 3: The processor supports performance-energy bias preference if set.
899 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
901 * This check is valid for both Intel and AMD
903 eax = ebx = ecx = edx = 0;
904 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
905 do_dts = eax & (1 << 0);
906 do_ptm = eax & (1 << 6);
907 if (!(ecx & (1 << 0))) {
908 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
913 * Enable or disable C states depending on the model and family
920 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
928 do_core_cstate = (1 << 1) | (1 << 6);
929 do_pkg_cstate = (1 << 6);
932 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
933 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
934 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
935 case 0x2E: /* Nehalem-EX Xeon - Beckton */
936 do_core_cstate = (1 << 3) | (1 << 6);
937 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
940 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
941 case 0x2C: /* Westmere EP - Gulftown */
942 case 0x2F: /* Westmere-EX Xeon - Eagleton */
943 do_core_cstate = (1 << 3) | (1 << 6);
944 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
948 case 0x2D: /* SNB Xeon */
949 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
950 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
954 case 0x3E: /* IVB Xeon */
955 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
956 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
962 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
963 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
966 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
967 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
971 case 0x56: /* BDX-DE */
972 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
973 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
976 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
977 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
980 ERROR("Turbostat plugin: Unsupported CPU (family: %#x,"
981 " model: %#x)", family, model);
989 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
992 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
996 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1000 do_rapl = RAPL_PKG | RAPL_CORES;
1006 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
1007 "model: %#x)", family, model);
1012 unsigned long long msr;
1013 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1017 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1019 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1026 /********************
1027 * Topology Probing *
1028 ********************/
1031 * Read a single int from a file.
1033 static int __attribute__ ((format(printf,1,2)))
1034 parse_int_file(const char *fmt, ...)
1037 char path[PATH_MAX];
1041 va_start(args, fmt);
1042 len = vsnprintf(path, sizeof(path), fmt, args);
1044 if (len < 0 || len >= PATH_MAX) {
1045 ERROR("Turbostat plugin: path truncated: '%s'", path);
1049 filep = fopen(path, "r");
1051 ERROR("Turbostat plugin: Failed to open '%s'", path);
1054 if (fscanf(filep, "%d", &value) != 1) {
1055 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1063 get_threads_on_core(int cpu)
1071 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1072 filep = fopen(path, "r");
1074 ERROR("Turbostat plugin: Failed to open '%s'", path);
1079 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1080 * otherwinse 1 sibling (self).
1082 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1093 * run func(cpu) on every cpu in /proc/stat
1094 * return max_cpu number
1096 static int __attribute__((warn_unused_result))
1097 for_all_proc_cpus(int (func)(int))
1103 fp = fopen("/proc/stat", "r");
1105 ERROR("Turbostat plugin: Failed to open /proc/stat");
1109 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1111 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1117 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1121 retval = func(cpu_num);
1132 * Update the stored topology.max_cpu_id
1135 update_max_cpu_id(int cpu)
1137 if (topology.max_cpu_id < cpu)
1138 topology.max_cpu_id = cpu;
1143 mark_cpu_present(int cpu)
1145 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1149 static int __attribute__((warn_unused_result))
1150 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1151 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1153 ERROR("Turbostat plugin: Unable to allocate CPU state");
1156 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1157 CPU_ZERO_S(*size, *set);
1162 * Build a local representation of the cpu distribution
1164 static int __attribute__((warn_unused_result))
1169 int max_package_id, max_core_id, max_thread_id;
1170 max_package_id = max_core_id = max_thread_id = 0;
1172 /* Clean topology */
1173 free(topology.cpus);
1174 memset(&topology, 0, sizeof(topology));
1176 ret = for_all_proc_cpus(update_max_cpu_id);
1180 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1181 if (topology.cpus == NULL) {
1182 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1186 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1189 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1192 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1196 ret = for_all_proc_cpus(mark_cpu_present);
1202 * find max_core_id, max_package_id
1204 for (i = 0; i <= topology.max_cpu_id; ++i) {
1206 struct cpu_topology *cpu = &topology.cpus[i];
1208 if (cpu_is_not_present(i)) {
1209 WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1213 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1217 cpu->package_id = ret;
1218 if (cpu->package_id > max_package_id)
1219 max_package_id = cpu->package_id;
1221 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1226 if (cpu->core_id > max_core_id)
1227 max_core_id = cpu->core_id;
1228 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1232 cpu->first_core_in_package = 1;
1234 ret = get_threads_on_core(i);
1239 if (num_threads > max_thread_id)
1240 max_thread_id = num_threads;
1241 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1245 cpu->first_thread_in_core = 1;
1247 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1248 i, cpu->package_id, cpu->core_id);
1250 /* Num is max + 1 (need to count 0) */
1251 topology.num_packages = max_package_id + 1;
1252 topology.num_cores = max_core_id + 1;
1253 topology.num_threads = max_thread_id + 1;
1257 free(topology.cpus);
1262 /************************
1263 * Main alloc/init/free *
1264 ************************/
1267 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1270 int total_threads, total_cores;
1272 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1273 *threads = calloc(total_threads, sizeof(struct thread_data));
1274 if (*threads == NULL)
1277 for (i = 0; i < total_threads; ++i)
1278 (*threads)[i].cpu_id = -1;
1280 total_cores = topology.num_cores * topology.num_packages;
1281 *cores = calloc(total_cores, sizeof(struct core_data));
1283 goto err_clean_threads;
1285 for (i = 0; i < total_cores; ++i)
1286 (*cores)[i].core_id = -1;
1288 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1289 if (*packages == NULL)
1290 goto err_clean_cores;
1292 for (i = 0; i < topology.num_packages; i++)
1293 (*packages)[i].package_id = i;
1302 ERROR("Turbostat plugin: Failled to allocate memory for counters");
1307 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1308 struct pkg_data *pkg_base, int cpu_id)
1310 struct thread_data *t;
1311 struct core_data *c;
1313 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1315 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1316 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1317 p = GET_PKG(pkg_base, cpu->package_id);
1320 if (cpu->first_thread_in_core)
1321 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1322 if (cpu->first_core_in_package)
1323 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1325 c->core_id = cpu->core_id;
1326 p->package_id = cpu->package_id;
1330 initialize_counters(void)
1334 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1335 if (cpu_is_not_present(cpu_id))
1337 init_counter(EVEN_COUNTERS, cpu_id);
1338 init_counter(ODD_COUNTERS, cpu_id);
1339 init_counter(DELTA_COUNTERS, cpu_id);
1346 free_all_buffers(void)
1351 CPU_FREE(cpu_present_set);
1352 cpu_present_set = NULL;
1353 cpu_present_set = 0;
1355 CPU_FREE(cpu_affinity_set);
1356 cpu_affinity_set = NULL;
1357 cpu_affinity_setsize = 0;
1359 CPU_FREE(cpu_saved_affinity_set);
1360 cpu_saved_affinity_set = NULL;
1361 cpu_saved_affinity_setsize = 0;
1369 package_even = NULL;
1381 free(package_delta);
1383 thread_delta = NULL;
1385 package_delta = NULL;
1389 /**********************
1390 * Collectd functions *
1391 **********************/
1393 #define DO_OR_GOTO_ERR(something) \
1395 ret = (something); \
1400 static int setup_all_buffers(void)
1404 DO_OR_GOTO_ERR(topology_probe());
1405 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1406 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1407 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1408 initialize_counters();
1409 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1410 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1420 turbostat_read(void)
1425 if ((ret = setup_all_buffers()) < 0)
1429 if (for_all_proc_cpus(cpu_is_not_present)) {
1431 if ((ret = setup_all_buffers()) < 0)
1433 if (for_all_proc_cpus(cpu_is_not_present)) {
1434 ERROR("Turbostat plugin: CPU appeared just after "
1440 /* Saving the scheduling affinity, as it will be modified by get_counters */
1441 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1442 ERROR("Turbostat plugin: Unable to save the CPU affinity");
1447 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1449 time_even = cdtime();
1457 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1459 time_odd = cdtime();
1461 time_delta = time_odd - time_even;
1462 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1464 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1467 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1469 time_even = cdtime();
1471 time_delta = time_even - time_odd;
1472 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1474 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1480 * Let's restore the affinity
1481 * This might fail if the number of CPU changed, but we can't do anything in that case..
1483 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1488 turbostat_init(void)
1493 if (getuid() != 0) {
1494 ERROR("Turbostat plugin: Initialization failed: this plugin "
1495 "requires collectd to run as root in order to read "
1496 "special CPU registers");
1500 DO_OR_GOTO_ERR(probe_cpu());
1502 if (stat("/dev/cpu/0/msr", &sb)) {
1503 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1504 " does not exist while the CPU supports MSR. You may be "
1505 "missing the corresponding kernel module, please try '# "
1510 DO_OR_GOTO_ERR(setup_all_buffers());
1512 plugin_register_read(PLUGIN_NAME, turbostat_read);
1521 turbostat_config(const char *key, const char *value)
1523 long unsigned int tmp_val;
1526 if (strcasecmp("TCCActivationTemp", key) == 0) {
1527 tmp_val = strtoul(value, &end, 0);
1528 if (*end != '\0' || tmp_val > UINT_MAX) {
1529 ERROR("Turbostat plugin: Invalid TCCActivationTemp '%s'",
1533 tcc_activation_temp = (unsigned int) tmp_val;
1535 ERROR("Turbostat plugin: Invalid configuration option '%s'",
1542 void module_register(void)
1544 plugin_register_init(PLUGIN_NAME, turbostat_init);
1545 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);