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 'I/O System-Management Interrupt counter'
97 * Boolean indicating if the processor supports 'Digital temperature sensor'
98 * This feature enables the monitoring of the temperature of each core
100 * This feature has two limitations:
101 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
102 * - Temperatures above the tcc_activation_temp are not recorded
107 * Boolean indicating if the processor supports 'Package thermal management'
108 * This feature allows the monitoring of the temperature of each package
110 * This feature has two limitations:
111 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
112 * - Temperatures above the tcc_activation_temp are not recorded
117 * Thermal Control Circuit Activation Temperature as configured by the user.
118 * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
119 * and should only be used if the automated detection fails.
121 static unsigned int tcc_activation_temp;
123 static unsigned int do_rapl;
124 static double rapl_energy_units;
126 #define RAPL_PKG (1 << 0)
127 /* 0x610 MSR_PKG_POWER_LIMIT */
128 /* 0x611 MSR_PKG_ENERGY_STATUS */
129 #define RAPL_DRAM (1 << 1)
130 /* 0x618 MSR_DRAM_POWER_LIMIT */
131 /* 0x619 MSR_DRAM_ENERGY_STATUS */
132 /* 0x61c MSR_DRAM_POWER_INFO */
133 #define RAPL_CORES (1 << 2)
134 /* 0x638 MSR_PP0_POWER_LIMIT */
135 /* 0x639 MSR_PP0_ENERGY_STATUS */
137 #define RAPL_GFX (1 << 3)
138 /* 0x640 MSR_PP1_POWER_LIMIT */
139 /* 0x641 MSR_PP1_ENERGY_STATUS */
140 /* 0x642 MSR_PP1_POLICY */
141 #define TJMAX_DEFAULT 100
143 static cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
144 static size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
146 static struct thread_data {
147 unsigned long long tsc;
148 unsigned long long aperf;
149 unsigned long long mperf;
150 unsigned long long c1;
151 unsigned int smi_count;
154 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
155 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
156 } *thread_delta, *thread_even, *thread_odd;
158 static struct core_data {
159 unsigned long long c3;
160 unsigned long long c6;
161 unsigned long long c7;
162 unsigned int core_temp_c;
163 unsigned int core_id;
164 } *core_delta, *core_even, *core_odd;
166 static struct pkg_data {
167 unsigned long long pc2;
168 unsigned long long pc3;
169 unsigned long long pc6;
170 unsigned long long pc7;
171 unsigned long long pc8;
172 unsigned long long pc9;
173 unsigned long long pc10;
174 unsigned int package_id;
175 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
176 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
177 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
178 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
179 unsigned int tcc_activation_temp;
180 unsigned int pkg_temp_c;
181 } *package_delta, *package_even, *package_odd;
183 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
184 #define ODD_COUNTERS thread_odd, core_odd, package_odd
185 #define EVEN_COUNTERS thread_even, core_even, package_even
186 static _Bool is_even = 1;
188 static _Bool allocated = 0;
189 static _Bool initialized = 0;
191 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
193 (pkg_no) * topology.num_cores * topology.num_threads + \
194 (core_no) * topology.num_threads + \
196 #define GET_CORE(core_base, core_no, pkg_no) \
198 (pkg_no) * topology.num_cores + \
200 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
202 struct cpu_topology {
205 _Bool first_core_in_package;
206 _Bool first_thread_in_core;
209 static struct topology {
214 struct cpu_topology *cpus;
217 static cdtime_t time_even, time_odd, time_delta;
219 static const char *config_keys[] =
223 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
225 /*****************************
226 * MSR Manipulation helpers *
227 *****************************/
230 * Open a MSR device for reading
231 * Can change the scheduling affinity of the current process if multiple_read is 1
233 static int __attribute__((warn_unused_result))
234 open_msr(int cpu, _Bool multiple_read)
240 * If we need to do multiple read, let's migrate to the CPU
241 * Otherwise, we would lose time calling functions on another CPU
243 * If we are not yet initialized (cpu_affinity_setsize = 0),
244 * we need to skip this optimisation.
246 if (multiple_read && cpu_affinity_setsize) {
247 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
248 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
249 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
250 ERROR("Turbostat plugin: Could not migrate to CPU %d", cpu);
255 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
256 fd = open(pathname, O_RDONLY);
258 ERROR("Turbostat plugin: failed to open %s", pathname);
265 * Read a single MSR from an open file descriptor
267 static int __attribute__((warn_unused_result))
268 read_msr(int fd, off_t offset, unsigned long long *msr)
272 retval = pread(fd, msr, sizeof *msr, offset);
274 if (retval != sizeof *msr) {
275 ERROR("Turbostat plugin: MSR offset 0x%llx read failed",
276 (unsigned long long)offset);
283 * Open a MSR device for reading, read the value asked for and close it.
284 * This call will not affect the scheduling affinity of this thread.
286 static int __attribute__((warn_unused_result))
287 get_msr(int cpu, off_t offset, unsigned long long *msr)
292 fd = open_msr(cpu, 0);
295 retval = read_msr(fd, offset, msr);
301 /********************************
302 * Raw data acquisition (1 CPU) *
303 ********************************/
306 * Read every data avalaible for a single CPU
308 * Core data is shared for all threads in one core: extracted only for the first thread
309 * Package data is shared for all core in one package: extracted only for the first thread of the first core
311 * Side effect: migrates to the targeted CPU
313 static int __attribute__((warn_unused_result))
314 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
317 unsigned long long msr;
321 msr_fd = open_msr(cpu, 1);
325 #define READ_MSR(msr, dst) \
327 if (read_msr(msr_fd, msr, dst)) { \
328 ERROR("Turbostat plugin: Unable to read " #msr); \
334 READ_MSR(MSR_IA32_TSC, &t->tsc);
336 READ_MSR(MSR_IA32_APERF, &t->aperf);
337 READ_MSR(MSR_IA32_MPERF, &t->mperf);
340 READ_MSR(MSR_SMI_COUNT, &msr);
341 t->smi_count = msr & 0xFFFFFFFF;
344 /* collect core counters only for 1st thread in core */
345 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
350 if (do_core_cstate & (1 << 3))
351 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
352 if (do_core_cstate & (1 << 6))
353 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
354 if (do_core_cstate & (1 << 7))
355 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
358 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
359 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
362 /* collect package counters only for 1st core in package */
363 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
368 if (do_pkg_cstate & (1 << 2))
369 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
370 if (do_pkg_cstate & (1 << 3))
371 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
372 if (do_pkg_cstate & (1 << 6))
373 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
374 if (do_pkg_cstate & (1 << 7))
375 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
376 if (do_pkg_cstate & (1 << 8))
377 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
378 if (do_pkg_cstate & (1 << 9))
379 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
380 if (do_pkg_cstate & (1 << 10))
381 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
383 if (do_rapl & RAPL_PKG) {
384 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
385 p->energy_pkg = msr & 0xFFFFFFFF;
387 if (do_rapl & RAPL_CORES) {
388 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
389 p->energy_cores = msr & 0xFFFFFFFF;
391 if (do_rapl & RAPL_DRAM) {
392 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
393 p->energy_dram = msr & 0xFFFFFFFF;
395 if (do_rapl & RAPL_GFX) {
396 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
397 p->energy_gfx = msr & 0xFFFFFFFF;
400 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
401 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
410 /**********************************
411 * Evaluating the changes (1 CPU) *
412 **********************************/
415 * Do delta = new - old on 32bits cyclique intergers
417 #define DELTA_WRAP32(delta, new, old) \
421 delta = 0x100000000 + new - old; \
425 * Extract the evolution old->new in delta at a package level
426 * (some are not new-delta, e.g. temperature)
429 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
431 delta->pc2 = new->pc2 - old->pc2;
432 delta->pc3 = new->pc3 - old->pc3;
433 delta->pc6 = new->pc6 - old->pc6;
434 delta->pc7 = new->pc7 - old->pc7;
435 delta->pc8 = new->pc8 - old->pc8;
436 delta->pc9 = new->pc9 - old->pc9;
437 delta->pc10 = new->pc10 - old->pc10;
438 delta->pkg_temp_c = new->pkg_temp_c;
440 DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
441 DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
442 DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
443 DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
447 * Extract the evolution old->new in delta at a core level
448 * (some are not new-delta, e.g. temperature)
451 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
453 delta->c3 = new->c3 - old->c3;
454 delta->c6 = new->c6 - old->c6;
455 delta->c7 = new->c7 - old->c7;
456 delta->core_temp_c = new->core_temp_c;
460 * Extract the evolution old->new in delta at a package level
461 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
463 static inline int __attribute__((warn_unused_result))
464 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
465 const struct core_data *core_delta)
467 delta->tsc = new->tsc - old->tsc;
469 /* check for TSC < 1 Mcycles over interval */
470 if (delta->tsc < (1000 * 1000)) {
471 WARNING("Turbostat plugin: Insanely slow TSC rate, TSC stops "
472 "in idle? You can disable all c-states by booting with"
473 " 'idle=poll' or just the deep ones with"
474 " 'processor.max_cstate=1'");
478 delta->c1 = new->c1 - old->c1;
480 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
481 delta->aperf = new->aperf - old->aperf;
482 delta->mperf = new->mperf - old->mperf;
484 if (!aperf_mperf_unstable) {
485 WARNING("Turbostat plugin: APERF or MPERF went "
486 "backwards. Frequency results do not cover "
487 "the entire interval. Fix this by running "
488 "Linux-2.6.30 or later.");
490 aperf_mperf_unstable = 1;
495 * As counter collection is not atomic,
496 * it is possible for mperf's non-halted cycles + idle states
497 * to exceed TSC's all cycles: show c1 = 0% in that case.
499 if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
502 /* normal case, derive c1 */
503 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
504 - core_delta->c6 - core_delta->c7;
507 if (delta->mperf == 0) {
508 WARNING("Turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
509 delta->mperf = 1; /* divide by 0 protection */
513 delta->smi_count = new->smi_count - old->smi_count;
518 /**********************************
519 * Submitting the results (1 CPU) *
520 **********************************/
523 * Submit one gauge value
526 turbostat_submit (const char *plugin_instance,
527 const char *type, const char *type_instance,
530 value_list_t vl = VALUE_LIST_INIT;
536 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
537 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
538 if (plugin_instance != NULL)
539 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
540 sstrncpy (vl.type, type, sizeof (vl.type));
541 if (type_instance != NULL)
542 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
544 plugin_dispatch_values (&vl);
548 * Submit every data for a single CPU
550 * Core data is shared for all threads in one core: submitted only for the first thread
551 * Package data is shared for all core in one package: submitted only for the first thread of the first core
554 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
556 char name[DATA_MAX_NAME_LEN];
557 double interval_float;
559 interval_float = CDTIME_T_TO_DOUBLE(time_delta);
561 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
563 if (!aperf_mperf_unstable)
564 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
565 if (!aperf_mperf_unstable)
566 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
568 turbostat_submit("Average", "frequency", name, 1.0 / 1000000 * t->aperf / interval_float);
570 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
571 turbostat_submit("Buzy", "frequency", name, 1.0 * t->tsc / 1000000 * t->aperf / t->mperf / interval_float);
573 /* Sanity check (should stay stable) */
574 turbostat_submit("TSC", "gauge", name, 1.0 * t->tsc / 1000000 / interval_float);
578 turbostat_submit(NULL, "current", name, t->smi_count);
580 /* submit per-core data only for 1st thread in core */
581 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
584 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
586 if (do_core_cstate & (1 << 3))
587 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
588 if (do_core_cstate & (1 << 6))
589 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
590 if (do_core_cstate & (1 << 7))
591 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
594 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
596 /* submit per-package data only for 1st core in package */
597 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
600 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
603 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
605 if (do_pkg_cstate & (1 << 2))
606 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
607 if (do_pkg_cstate & (1 << 3))
608 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
609 if (do_pkg_cstate & (1 << 6))
610 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
611 if (do_pkg_cstate & (1 << 7))
612 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
613 if (do_pkg_cstate & (1 << 8))
614 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
615 if (do_pkg_cstate & (1 << 9))
616 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
617 if (do_pkg_cstate & (1 << 10))
618 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
621 if (do_rapl & RAPL_PKG)
622 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
623 if (do_rapl & RAPL_CORES)
624 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
625 if (do_rapl & RAPL_GFX)
626 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
627 if (do_rapl & RAPL_DRAM)
628 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
635 /**********************************
636 * Looping function over all CPUs *
637 **********************************/
640 * Check if a given cpu id is in our compiled list of existing CPUs
643 cpu_is_not_present(int cpu)
645 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
649 * Loop on all CPUs in topological order
651 * Skip non-present cpus
652 * Return the error code at the first error or 0
654 static int __attribute__((warn_unused_result))
655 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
656 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
658 int retval, pkg_no, core_no, thread_no;
660 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
661 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
662 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
663 struct thread_data *t;
667 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
669 if (cpu_is_not_present(t->cpu_id))
672 c = GET_CORE(core_base, core_no, pkg_no);
673 p = GET_PKG(pkg_base, pkg_no);
675 retval = func(t, c, p);
685 * Dedicated loop: Extract every data evolution for all CPU
687 * Skip non-present cpus
688 * Return the error code at the first error or 0
690 * Core data is shared for all threads in one core: extracted only for the first thread
691 * Package data is shared for all core in one package: extracted only for the first thread of the first core
693 static int __attribute__((warn_unused_result))
694 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,
695 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
697 int retval, pkg_no, core_no, thread_no;
699 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
700 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
701 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
702 struct thread_data *t_delta;
703 const struct thread_data *t_old, *t_new;
704 struct core_data *c_delta;
706 /* Get correct pointers for threads */
707 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
708 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
709 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
711 /* Skip threads that disappeared */
712 if (cpu_is_not_present(t_delta->cpu_id))
715 /* c_delta is always required for delta_thread */
716 c_delta = GET_CORE(core_delta, core_no, pkg_no);
718 /* calculate core delta only for 1st thread in core */
719 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
720 const struct core_data *c_old, *c_new;
722 c_new = GET_CORE(core_new_base, core_no, pkg_no);
723 c_old = GET_CORE(core_old_base, core_no, pkg_no);
725 delta_core(c_delta, c_new, c_old);
728 /* Always calculate thread delta */
729 retval = delta_thread(t_delta, t_new, t_old, c_delta);
733 /* calculate package delta only for 1st core in package */
734 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
735 struct pkg_data *p_delta;
736 const struct pkg_data *p_old, *p_new;
738 p_delta = GET_PKG(package_delta, pkg_no);
739 p_new = GET_PKG(pkg_new_base, pkg_no);
740 p_old = GET_PKG(pkg_old_base, pkg_no);
742 delta_package(p_delta, p_new, p_old);
756 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
757 * the Thermal Control Circuit (TCC) activates.
758 * This is usually equal to tjMax.
760 * Older processors do not have this MSR, so there we guess,
761 * but also allow conficuration over-ride with "TCCActivationTemp".
763 * Several MSR temperature values are in units of degrees-C
764 * below this value, including the Digital Thermal Sensor (DTS),
765 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
767 static int __attribute__((warn_unused_result))
768 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
770 unsigned long long msr;
771 unsigned int target_c_local;
773 /* tcc_activation_temp is used only for dts or ptm */
774 if (!(do_dts || do_ptm))
777 /* this is a per-package concept */
778 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
781 if (tcc_activation_temp != 0) {
782 p->tcc_activation_temp = tcc_activation_temp;
786 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
789 target_c_local = (msr >> 16) & 0xFF;
794 p->tcc_activation_temp = target_c_local;
799 p->tcc_activation_temp = TJMAX_DEFAULT;
800 WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
801 " Please use TCCActivationTemp to specify it.",
802 t->cpu_id, p->tcc_activation_temp);
808 * Identify the functionality of the CPU
810 static int __attribute__((warn_unused_result))
813 unsigned int eax, ebx, ecx, edx, max_level;
814 unsigned int fms, family, model;
817 * - EAX: Maximum Input Value for Basic CPUID Information
818 * - EBX: "Genu" (0x756e6547)
819 * - EDX: "ineI" (0x49656e69)
820 * - ECX: "ntel" (0x6c65746e)
822 max_level = ebx = ecx = edx = 0;
823 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
824 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
825 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
830 * - EAX: Version Information: Type, Family, Model, and Stepping ID
833 * + 12-13: Processor type
834 * + 16-19: Extended Model ID
835 * + 20-27: Extended Family ID
836 * - EDX: Feature Information:
837 * + 5: Support for MSR read/write operations
839 fms = ebx = ecx = edx = 0;
840 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
841 family = (fms >> 8) & 0xf;
842 model = (fms >> 4) & 0xf;
844 family += (fms >> 20) & 0xf;
845 if (family == 6 || family == 0xf)
846 model += ((fms >> 16) & 0xf) << 4;
847 if (!(edx & (1 << 5))) {
848 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
855 * + 0: Digital temperature sensor is supported if set
856 * + 6: Package thermal management is supported if set
858 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
859 * + 3: The processor supports performance-energy bias preference if set.
860 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
862 * This check is valid for both Intel and AMD
864 eax = ebx = ecx = edx = 0;
865 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
866 do_dts = eax & (1 << 0);
867 do_ptm = eax & (1 << 6);
868 if (!(ecx & (1 << 0))) {
869 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
874 * Enable or disable C states depending on the model and family
882 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
888 do_core_cstate = (1 << 1) | (1 << 6);
889 do_pkg_cstate = (1 << 6);
892 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
893 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
894 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
895 case 0x2E: /* Nehalem-EX Xeon - Beckton */
897 do_core_cstate = (1 << 3) | (1 << 6);
898 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
901 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
902 case 0x2C: /* Westmere EP - Gulftown */
903 case 0x2F: /* Westmere-EX Xeon - Eagleton */
905 do_core_cstate = (1 << 3) | (1 << 6);
906 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
910 case 0x2D: /* SNB Xeon */
912 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
913 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
917 case 0x3E: /* IVB Xeon */
919 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
920 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
927 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
928 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
932 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
933 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
937 case 0x56: /* BDX-DE */
939 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
940 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
944 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
945 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
960 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_GFX;
964 case 0x56: /* BDX-DE */
965 do_rapl = RAPL_PKG | RAPL_DRAM ;
967 case 0x2D: /* SNB Xeon */
968 case 0x3E: /* IVB Xeon */
969 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_DRAM;
973 do_rapl = RAPL_PKG | RAPL_CORES;
979 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
980 "model: %#x)", family, model);
985 unsigned long long msr;
986 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
990 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
992 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
999 /********************
1000 * Topology Probing *
1001 ********************/
1004 * Read a single int from a file.
1006 static int __attribute__ ((format(printf,1,2)))
1007 parse_int_file(const char *fmt, ...)
1010 char path[PATH_MAX];
1014 va_start(args, fmt);
1015 len = vsnprintf(path, sizeof(path), fmt, args);
1017 if (len < 0 || len >= PATH_MAX) {
1018 ERROR("Turbostat plugin: path truncated: '%s'", path);
1022 filep = fopen(path, "r");
1024 ERROR("Turbostat plugin: Failed to open '%s'", path);
1027 if (fscanf(filep, "%d", &value) != 1) {
1028 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1036 get_threads_on_core(int cpu)
1044 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1045 filep = fopen(path, "r");
1047 ERROR("Turbostat plugin: Failed to open '%s'", path);
1052 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1053 * otherwinse 1 sibling (self).
1055 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1066 * run func(cpu) on every cpu in /proc/stat
1067 * return max_cpu number
1069 static int __attribute__((warn_unused_result))
1070 for_all_proc_cpus(int (func)(int))
1076 fp = fopen("/proc/stat", "r");
1078 ERROR("Turbostat plugin: Failed to open /proc/stat");
1082 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1084 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1090 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1094 retval = func(cpu_num);
1105 * Update the stored topology.max_cpu_id
1108 update_max_cpu_id(int cpu)
1110 if (topology.max_cpu_id < cpu)
1111 topology.max_cpu_id = cpu;
1116 mark_cpu_present(int cpu)
1118 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1122 static int __attribute__((warn_unused_result))
1123 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1124 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1126 ERROR("Turbostat plugin: Unable to allocate CPU state");
1129 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1130 CPU_ZERO_S(*size, *set);
1135 * Build a local representation of the cpu distribution
1137 static int __attribute__((warn_unused_result))
1142 int max_package_id, max_core_id, max_thread_id;
1143 max_package_id = max_core_id = max_thread_id = 0;
1145 /* Clean topology */
1146 free(topology.cpus);
1147 memset(&topology, 0, sizeof(topology));
1149 ret = for_all_proc_cpus(update_max_cpu_id);
1153 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1154 if (topology.cpus == NULL) {
1155 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1159 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1162 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1165 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1169 ret = for_all_proc_cpus(mark_cpu_present);
1175 * find max_core_id, max_package_id
1177 for (i = 0; i <= topology.max_cpu_id; ++i) {
1179 struct cpu_topology *cpu = &topology.cpus[i];
1181 if (cpu_is_not_present(i)) {
1182 WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1186 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1190 cpu->package_id = ret;
1191 if (cpu->package_id > max_package_id)
1192 max_package_id = cpu->package_id;
1194 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1199 if (cpu->core_id > max_core_id)
1200 max_core_id = cpu->core_id;
1201 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1205 cpu->first_core_in_package = 1;
1207 ret = get_threads_on_core(i);
1212 if (num_threads > max_thread_id)
1213 max_thread_id = num_threads;
1214 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1218 cpu->first_thread_in_core = 1;
1220 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1221 i, cpu->package_id, cpu->core_id);
1223 /* Num is max + 1 (need to count 0) */
1224 topology.num_packages = max_package_id + 1;
1225 topology.num_cores = max_core_id + 1;
1226 topology.num_threads = max_thread_id + 1;
1230 free(topology.cpus);
1235 /************************
1236 * Main alloc/init/free *
1237 ************************/
1240 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1243 int total_threads, total_cores;
1245 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1246 *threads = calloc(total_threads, sizeof(struct thread_data));
1247 if (*threads == NULL)
1250 for (i = 0; i < total_threads; ++i)
1251 (*threads)[i].cpu_id = -1;
1253 total_cores = topology.num_cores * topology.num_packages;
1254 *cores = calloc(total_cores, sizeof(struct core_data));
1256 goto err_clean_threads;
1258 for (i = 0; i < total_cores; ++i)
1259 (*cores)[i].core_id = -1;
1261 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1262 if (*packages == NULL)
1263 goto err_clean_cores;
1265 for (i = 0; i < topology.num_packages; i++)
1266 (*packages)[i].package_id = i;
1275 ERROR("Turbostat plugin: Failled to allocate memory for counters");
1280 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1281 struct pkg_data *pkg_base, int cpu_id)
1283 struct thread_data *t;
1284 struct core_data *c;
1286 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1288 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1289 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1290 p = GET_PKG(pkg_base, cpu->package_id);
1293 if (cpu->first_thread_in_core)
1294 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1295 if (cpu->first_core_in_package)
1296 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1298 c->core_id = cpu->core_id;
1299 p->package_id = cpu->package_id;
1303 initialize_counters(void)
1307 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1308 if (cpu_is_not_present(cpu_id))
1310 init_counter(EVEN_COUNTERS, cpu_id);
1311 init_counter(ODD_COUNTERS, cpu_id);
1312 init_counter(DELTA_COUNTERS, cpu_id);
1319 free_all_buffers(void)
1324 CPU_FREE(cpu_present_set);
1325 cpu_present_set = NULL;
1326 cpu_present_set = 0;
1328 CPU_FREE(cpu_affinity_set);
1329 cpu_affinity_set = NULL;
1330 cpu_affinity_setsize = 0;
1332 CPU_FREE(cpu_saved_affinity_set);
1333 cpu_saved_affinity_set = NULL;
1334 cpu_saved_affinity_setsize = 0;
1342 package_even = NULL;
1354 free(package_delta);
1356 thread_delta = NULL;
1358 package_delta = NULL;
1362 /**********************
1363 * Collectd functions *
1364 **********************/
1366 #define DO_OR_GOTO_ERR(something) \
1368 ret = (something); \
1373 static int setup_all_buffers(void)
1377 DO_OR_GOTO_ERR(topology_probe());
1378 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1379 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1380 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1381 initialize_counters();
1382 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1383 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1393 turbostat_read(void)
1398 if ((ret = setup_all_buffers()) < 0)
1402 if (for_all_proc_cpus(cpu_is_not_present)) {
1404 if ((ret = setup_all_buffers()) < 0)
1406 if (for_all_proc_cpus(cpu_is_not_present)) {
1407 ERROR("Turbostat plugin: CPU appeared just after "
1413 /* Saving the scheduling affinity, as it will be modified by get_counters */
1414 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1415 ERROR("Turbostat plugin: Unable to save the CPU affinity");
1420 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1422 time_even = cdtime();
1430 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1432 time_odd = cdtime();
1434 time_delta = time_odd - time_even;
1435 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1437 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1440 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1442 time_even = cdtime();
1444 time_delta = time_even - time_odd;
1445 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1447 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1453 * Let's restore the affinity
1454 * This might fail if the number of CPU changed, but we can't do anything in that case..
1456 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1461 check_permissions(void)
1463 struct __user_cap_header_struct cap_header_data;
1464 cap_user_header_t cap_header = &cap_header_data;
1465 struct __user_cap_data_struct cap_data_data;
1466 cap_user_data_t cap_data = &cap_data_data;
1469 if (getuid() == 0) {
1470 /* We have everything we need */
1474 /* check for CAP_SYS_RAWIO */
1475 cap_header->pid = getpid();
1476 cap_header->version = _LINUX_CAPABILITY_VERSION;
1477 if (capget(cap_header, cap_data) < 0) {
1478 ERROR("Turbostat plugin: capget failed");
1482 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1483 WARNING("Turbostat plugin: Collectd doesn't have the "
1484 "CAP_SYS_RAWIO capability. If you don't want to run "
1485 "collectd as root, try running \"setcap "
1486 "cap_sys_rawio=ep\" on collectd binary");
1490 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1491 WARNING("Turbostat plugin: Collectd cannot open"
1492 "/dev/cpu/0/msr. If you don't want to run collectd as "
1493 "root, you need to change the ownership (chown) and "
1494 "permissions on /dev/cpu/*/msr to allow such access");
1499 ERROR("Turbostat plugin: Initialization failed: this plugin "
1500 "requires collectd to either to run as root or give "
1501 "collectd a special capability (CAP_SYS_RAWIO) and read "
1502 "access to /dev/cpu/*/msr (see previous warnings)");
1507 turbostat_init(void)
1512 if (stat("/dev/cpu/0/msr", &sb)) {
1513 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1514 " does not exist while the CPU supports MSR. You may be "
1515 "missing the corresponding kernel module, please try '# "
1520 DO_OR_GOTO_ERR(check_permissions());
1522 DO_OR_GOTO_ERR(probe_cpu());
1524 DO_OR_GOTO_ERR(setup_all_buffers());
1526 plugin_register_read(PLUGIN_NAME, turbostat_read);
1535 turbostat_config(const char *key, const char *value)
1537 long unsigned int tmp_val;
1540 if (strcasecmp("TCCActivationTemp", key) == 0) {
1541 tmp_val = strtoul(value, &end, 0);
1542 if (*end != '\0' || tmp_val > UINT_MAX) {
1543 ERROR("Turbostat plugin: Invalid TCCActivationTemp '%s'",
1547 tcc_activation_temp = (unsigned int) tmp_val;
1549 ERROR("Turbostat plugin: Invalid configuration option '%s'",
1556 void module_register(void)
1558 plugin_register_init(PLUGIN_NAME, turbostat_init);
1559 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);