2 * turbostat -- Log CPU frequency and C-state residency
3 * on modern Intel turbo-capable processors for collectd.
5 * Based on the 'turbostat' tool of the Linux kernel, found at
6 * linux/tools/power/x86/turbostat/turbostat.c:
8 * Copyright (c) 2013 Intel Corporation.
9 * Len Brown <len.brown@intel.com>
11 * This program is free software; you can redistribute it and/or modify it
12 * under the terms and conditions of the GNU General Public License,
13 * version 2, as published by the Free Software Foundation.
15 * This program is distributed in the hope it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
20 * You should have received a copy of the GNU General Public License along with
21 * this program; if not, write to the Free Software Foundation, Inc.,
22 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
24 * Ported to collectd by Vincent Brillault <git@lerya.net>
28 * _GNU_SOURCE is required because of the following functions:
38 #include <asm/msr-index.h>
43 #include <sys/types.h>
46 #include <sys/resource.h>
61 #define PLUGIN_NAME "turbostat"
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 struct timeval tv_even, tv_odd, tv_delta;
226 /*****************************
227 * MSR Manipulation helpers *
228 *****************************/
231 * Open a MSR device for reading
232 * Can change the scheduling affinity of the current process if multiple_read is 1
234 static int __attribute__((warn_unused_result))
235 open_msr(int cpu, _Bool multiple_read)
241 * If we need to do multiple read, let's migrate to the CPU
242 * Otherwise, we would lose time calling functions on another CPU
244 * If we are not yet initialized (cpu_affinity_setsize = 0),
245 * we need to skip this optimisation.
247 if (multiple_read && cpu_affinity_setsize) {
248 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
249 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
250 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
251 ERROR("Turbostat plugin: Could not migrate to CPU %d", cpu);
256 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
257 fd = open(pathname, O_RDONLY);
259 ERROR("Turbostat plugin: failed to open %s", pathname);
266 * Read a single MSR from an open file descriptor
268 static int __attribute__((warn_unused_result))
269 read_msr(int fd, off_t offset, unsigned long long *msr)
273 retval = pread(fd, msr, sizeof *msr, offset);
275 if (retval != sizeof *msr) {
276 ERROR("Turbostat plugin: MSR offset 0x%llx read failed",
277 (unsigned long long)offset);
284 * Open a MSR device for reading, read the value asked for and close it.
285 * This call will not affect the scheduling affinity of this thread.
287 static int __attribute__((warn_unused_result))
288 get_msr(int cpu, off_t offset, unsigned long long *msr)
293 fd = open_msr(cpu, 0);
296 retval = read_msr(fd, offset, msr);
302 /********************************
303 * Raw data acquisition (1 CPU) *
304 ********************************/
307 * Read every data avalaible for a single CPU
309 * Core data is shared for all threads in one core: extracted only for the first thread
310 * Package data is shared for all core in one package: extracted only for the first thread of the first core
312 * Side effect: migrates to the targeted CPU
314 static int __attribute__((warn_unused_result))
315 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
318 unsigned long long msr;
322 msr_fd = open_msr(cpu, 1);
326 #define READ_MSR(msr, dst) \
328 if (read_msr(msr_fd, msr, dst)) { \
329 ERROR("Turbostat plugin: Unable to read " #msr); \
335 READ_MSR(MSR_IA32_TSC, &t->tsc);
337 READ_MSR(MSR_IA32_APERF, &t->aperf);
338 READ_MSR(MSR_IA32_MPERF, &t->mperf);
340 READ_MSR(MSR_SMI_COUNT, &msr);
341 t->smi_count = msr & 0xFFFFFFFF;
343 /* collect core counters only for 1st thread in core */
344 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
349 if (do_core_cstate & (1 << 3))
350 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
351 if (do_core_cstate & (1 << 6))
352 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
353 if (do_core_cstate & (1 << 7))
354 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
357 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
358 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
361 /* collect package counters only for 1st core in package */
362 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
367 if (do_pkg_cstate & (1 << 2))
368 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
369 if (do_pkg_cstate & (1 << 3))
370 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
371 if (do_pkg_cstate & (1 << 6))
372 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
373 if (do_pkg_cstate & (1 << 7))
374 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
375 if (do_pkg_cstate & (1 << 8))
376 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
377 if (do_pkg_cstate & (1 << 9))
378 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
379 if (do_pkg_cstate & (1 << 10))
380 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
382 if (do_rapl & RAPL_PKG) {
383 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
384 p->energy_pkg = msr & 0xFFFFFFFF;
386 if (do_rapl & RAPL_CORES) {
387 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
388 p->energy_cores = msr & 0xFFFFFFFF;
390 if (do_rapl & RAPL_DRAM) {
391 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
392 p->energy_dram = msr & 0xFFFFFFFF;
394 if (do_rapl & RAPL_GFX) {
395 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
396 p->energy_gfx = msr & 0xFFFFFFFF;
398 if (do_rapl & RAPL_PKG_PERF_STATUS) {
399 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
400 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
402 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
403 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
404 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
407 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
408 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
417 /**********************************
418 * Evaluating the changes (1 CPU) *
419 **********************************/
422 * Do delta = new - old on 32bits cyclique intergers
424 #define DELTA_WRAP32(delta, new, old) \
428 delta = 0x100000000 + new - old; \
432 * Extract the evolution old->new in delta at a package level
433 * (some are not new-delta, e.g. temperature)
436 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
438 delta->pc2 = new->pc2 - old->pc2;
439 delta->pc3 = new->pc3 - old->pc3;
440 delta->pc6 = new->pc6 - old->pc6;
441 delta->pc7 = new->pc7 - old->pc7;
442 delta->pc8 = new->pc8 - old->pc8;
443 delta->pc9 = new->pc9 - old->pc9;
444 delta->pc10 = new->pc10 - old->pc10;
445 delta->pkg_temp_c = new->pkg_temp_c;
447 DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
448 DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
449 DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
450 DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
451 DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
452 DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
456 * Extract the evolution old->new in delta at a core level
457 * (some are not new-delta, e.g. temperature)
460 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
462 delta->c3 = new->c3 - old->c3;
463 delta->c6 = new->c6 - old->c6;
464 delta->c7 = new->c7 - old->c7;
465 delta->core_temp_c = new->core_temp_c;
469 * Extract the evolution old->new in delta at a package level
470 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
472 static inline int __attribute__((warn_unused_result))
473 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
474 const struct core_data *core_delta)
476 delta->tsc = new->tsc - old->tsc;
478 /* check for TSC < 1 Mcycles over interval */
479 if (delta->tsc < (1000 * 1000)) {
480 WARNING("Turbostat plugin: Insanely slow TSC rate, TSC stops "
481 "in idle? You can disable all c-states by booting with"
482 " 'idle=poll' or just the deep ones with"
483 " 'processor.max_cstate=1'");
487 delta->c1 = new->c1 - old->c1;
489 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
490 delta->aperf = new->aperf - old->aperf;
491 delta->mperf = new->mperf - old->mperf;
493 if (!aperf_mperf_unstable) {
494 WARNING("Turbostat plugin: APERF or MPERF went "
495 "backwards. Frequency results do not cover "
496 "the entire interval. Fix this by running "
497 "Linux-2.6.30 or later.");
499 aperf_mperf_unstable = 1;
504 * As counter collection is not atomic,
505 * it is possible for mperf's non-halted cycles + idle states
506 * to exceed TSC's all cycles: show c1 = 0% in that case.
508 if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
511 /* normal case, derive c1 */
512 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
513 - core_delta->c6 - core_delta->c7;
516 if (delta->mperf == 0) {
517 WARNING("Turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
518 delta->mperf = 1; /* divide by 0 protection */
521 delta->smi_count = new->smi_count - old->smi_count;
526 /**********************************
527 * Submitting the results (1 CPU) *
528 **********************************/
531 * Submit one gauge value
534 turbostat_submit (const char *plugin_instance,
535 const char *type, const char *type_instance,
538 value_list_t vl = VALUE_LIST_INIT;
544 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
545 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
546 if (plugin_instance != NULL)
547 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
548 sstrncpy (vl.type, type, sizeof (vl.type));
549 if (type_instance != NULL)
550 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
552 plugin_dispatch_values (&vl);
556 * Submit every data for a single CPU
558 * Core data is shared for all threads in one core: submitted only for the first thread
559 * Package data is shared for all core in one package: submitted only for the first thread of the first core
562 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
564 char name[DATA_MAX_NAME_LEN];
565 double interval_float;
567 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
569 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
571 if (!aperf_mperf_unstable)
572 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
573 if (!aperf_mperf_unstable)
574 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
577 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
578 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
581 turbostat_submit(NULL, "current", name, t->smi_count);
583 /* submit per-core data only for 1st thread in core */
584 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
587 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
589 if (do_core_cstate & (1 << 3))
590 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
591 if (do_core_cstate & (1 << 6))
592 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
593 if (do_core_cstate & (1 << 7))
594 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
597 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
599 /* submit per-package data only for 1st core in package */
600 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
603 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
606 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
608 if (do_pkg_cstate & (1 << 2))
609 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
610 if (do_pkg_cstate & (1 << 3))
611 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
612 if (do_pkg_cstate & (1 << 6))
613 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
614 if (do_pkg_cstate & (1 << 7))
615 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
616 if (do_pkg_cstate & (1 << 8))
617 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
618 if (do_pkg_cstate & (1 << 9))
619 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
620 if (do_pkg_cstate & (1 << 10))
621 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
624 if (do_rapl & RAPL_PKG)
625 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
626 if (do_rapl & RAPL_CORES)
627 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
628 if (do_rapl & RAPL_GFX)
629 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
630 if (do_rapl & RAPL_DRAM)
631 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
638 /**********************************
639 * Looping function over all CPUs *
640 **********************************/
643 * Check if a given cpu id is in our compiled list of existing CPUs
646 cpu_is_not_present(int cpu)
648 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
652 * Loop on all CPUs in topological order
654 * Skip non-present cpus
655 * Return the error code at the first error or 0
657 static int __attribute__((warn_unused_result))
658 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
659 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
661 int retval, pkg_no, core_no, thread_no;
663 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
664 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
665 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
666 struct thread_data *t;
670 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
672 if (cpu_is_not_present(t->cpu_id))
675 c = GET_CORE(core_base, core_no, pkg_no);
676 p = GET_PKG(pkg_base, pkg_no);
678 retval = func(t, c, p);
688 * Dedicated loop: Extract every data evolution for all CPU
690 * Skip non-present cpus
691 * Return the error code at the first error or 0
693 * Core data is shared for all threads in one core: extracted only for the first thread
694 * Package data is shared for all core in one package: extracted only for the first thread of the first core
696 static int __attribute__((warn_unused_result))
697 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,
698 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
700 int retval, pkg_no, core_no, thread_no;
702 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
703 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
704 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
705 struct thread_data *t_delta;
706 const struct thread_data *t_old, *t_new;
707 struct core_data *c_delta;
709 /* Get correct pointers for threads */
710 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
711 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
712 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
714 /* Skip threads that disappeared */
715 if (cpu_is_not_present(t_delta->cpu_id))
718 /* c_delta is always required for delta_thread */
719 c_delta = GET_CORE(core_delta, core_no, pkg_no);
721 /* calculate core delta only for 1st thread in core */
722 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
723 const struct core_data *c_old, *c_new;
725 c_new = GET_CORE(core_new_base, core_no, pkg_no);
726 c_old = GET_CORE(core_old_base, core_no, pkg_no);
728 delta_core(c_delta, c_new, c_old);
731 /* Always calculate thread delta */
732 retval = delta_thread(t_delta, t_new, t_old, c_delta);
736 /* calculate package delta only for 1st core in package */
737 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
738 struct pkg_data *p_delta;
739 const struct pkg_data *p_old, *p_new;
741 p_delta = GET_PKG(package_delta, pkg_no);
742 p_new = GET_PKG(pkg_new_base, pkg_no);
743 p_old = GET_PKG(pkg_old_base, pkg_no);
745 delta_package(p_delta, p_new, p_old);
759 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
760 * the Thermal Control Circuit (TCC) activates.
761 * This is usually equal to tjMax.
763 * Older processors do not have this MSR, so there we guess,
764 * but also allow conficuration over-ride with "TCCActivationTemp".
766 * Several MSR temperature values are in units of degrees-C
767 * below this value, including the Digital Thermal Sensor (DTS),
768 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
770 static int __attribute__((warn_unused_result))
771 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
773 unsigned long long msr;
774 unsigned int target_c_local;
776 /* tcc_activation_temp is used only for dts or ptm */
777 if (!(do_dts || do_ptm))
780 /* this is a per-package concept */
781 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
784 if (tcc_activation_temp != 0) {
785 p->tcc_activation_temp = tcc_activation_temp;
789 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
792 target_c_local = (msr >> 16) & 0xFF;
797 p->tcc_activation_temp = target_c_local;
802 p->tcc_activation_temp = TJMAX_DEFAULT;
803 WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
804 " Please use TCCActivationTemp to specify it.",
805 t->cpu_id, p->tcc_activation_temp);
811 * Identify the functionality of the CPU
813 static int __attribute__((warn_unused_result))
816 unsigned int eax, ebx, ecx, edx, max_level;
817 unsigned int fms, family, model;
820 * - EAX: Maximum Input Value for Basic CPUID Information
821 * - EBX: "Genu" (0x756e6547)
822 * - EDX: "ineI" (0x49656e69)
823 * - ECX: "ntel" (0x6c65746e)
825 max_level = ebx = ecx = edx = 0;
826 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
827 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
828 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
833 * - EAX: Version Information: Type, Family, Model, and Stepping ID
836 * + 12-13: Processor type
837 * + 16-19: Extended Model ID
838 * + 20-27: Extended Family ID
839 * - EDX: Feature Information:
840 * + 5: Support for MSR read/write operations
842 fms = ebx = ecx = edx = 0;
843 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
844 family = (fms >> 8) & 0xf;
845 model = (fms >> 4) & 0xf;
847 family += (fms >> 20) & 0xf;
848 if (family == 6 || family == 0xf)
849 model += ((fms >> 16) & 0xf) << 4;
850 if (!(edx & (1 << 5))) {
851 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
857 * - EAX: Maximum Input Value for Extended Function CPUID Information
859 * This allows us to verify if the CPUID(0x80000007) can be called
861 * This check is valid for both Intel and AMD.
863 max_level = ebx = ecx = edx = 0;
864 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
865 if (max_level < 0x80000007) {
866 ERROR("Turbostat plugin: Unsupported CPU (no invariant TSC, "
867 " Maximum Extended Function: 0x%x)", max_level);
874 * + 8: Invariant TSC available if set
876 * This check is valid for both Intel and AMD
878 eax = ebx = ecx = edx = 0;
879 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
880 if (!(edx & (1 << 8))) {
881 ERROR("Turbostat plugin: Unsupported CPU (No invariant TSC)");
888 * + 0: Digital temperature sensor is supported if set
889 * + 6: Package thermal management is supported if set
891 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
892 * + 3: The processor supports performance-energy bias preference if set.
893 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
895 * This check is valid for both Intel and AMD
897 eax = ebx = ecx = edx = 0;
898 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
899 do_dts = eax & (1 << 0);
900 do_ptm = eax & (1 << 6);
901 if (!(ecx & (1 << 0))) {
902 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
907 * Enable or disable C states depending on the model and family
914 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
922 do_core_cstate = (1 << 1) | (1 << 6);
923 do_pkg_cstate = (1 << 6);
926 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
927 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
928 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
929 case 0x2E: /* Nehalem-EX Xeon - Beckton */
930 do_core_cstate = (1 << 3) | (1 << 6);
931 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
934 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
935 case 0x2C: /* Westmere EP - Gulftown */
936 case 0x2F: /* Westmere-EX Xeon - Eagleton */
937 do_core_cstate = (1 << 3) | (1 << 6);
938 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
942 case 0x2D: /* SNB Xeon */
943 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
944 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
948 case 0x3E: /* IVB Xeon */
949 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
950 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
956 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
957 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
960 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
961 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
965 case 0x56: /* BDX-DE */
966 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
967 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
970 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
971 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
974 ERROR("Turbostat plugin: Unsupported CPU (family: %#x,"
975 " model: %#x)", family, model);
983 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
986 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
990 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
994 do_rapl = RAPL_PKG | RAPL_CORES;
1000 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
1001 "model: %#x)", family, model);
1006 unsigned long long msr;
1007 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1011 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1013 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1020 /********************
1021 * Topology Probing *
1022 ********************/
1025 * Read a single int from a file.
1027 static int __attribute__ ((format(printf,1,2)))
1028 parse_int_file(const char *fmt, ...)
1031 char path[PATH_MAX];
1035 va_start(args, fmt);
1036 len = vsnprintf(path, sizeof(path), fmt, args);
1038 if (len < 0 || len >= PATH_MAX) {
1039 ERROR("Turbostat plugin: path truncated: '%s'", path);
1043 filep = fopen(path, "r");
1045 ERROR("Turbostat plugin: Failed to open '%s'", path);
1048 if (fscanf(filep, "%d", &value) != 1) {
1049 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1057 get_threads_on_core(int cpu)
1065 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1066 filep = fopen(path, "r");
1068 ERROR("Turbostat plugin: Failed to open '%s'", path);
1073 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1074 * otherwinse 1 sibling (self).
1076 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1087 * run func(cpu) on every cpu in /proc/stat
1088 * return max_cpu number
1090 static int __attribute__((warn_unused_result))
1091 for_all_proc_cpus(int (func)(int))
1097 fp = fopen("/proc/stat", "r");
1099 ERROR("Turbostat plugin: Failed to open /proc/stat");
1103 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1105 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1111 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1115 retval = func(cpu_num);
1126 * Update the stored topology.max_cpu_id
1129 update_max_cpu_id(int cpu)
1131 if (topology.max_cpu_id < cpu)
1132 topology.max_cpu_id = cpu;
1137 mark_cpu_present(int cpu)
1139 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1143 static int __attribute__((warn_unused_result))
1144 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1145 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1147 ERROR("Turbostat plugin: Unable to allocate CPU state");
1150 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1151 CPU_ZERO_S(*size, *set);
1156 * Build a local representation of the cpu distribution
1158 static int __attribute__((warn_unused_result))
1163 int max_package_id, max_core_id, max_thread_id;
1164 max_package_id = max_core_id = max_thread_id = 0;
1166 /* Clean topology */
1167 free(topology.cpus);
1168 memset(&topology, 0, sizeof(topology));
1170 ret = for_all_proc_cpus(update_max_cpu_id);
1174 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1175 if (topology.cpus == NULL) {
1176 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1180 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1183 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1186 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1190 ret = for_all_proc_cpus(mark_cpu_present);
1196 * find max_core_id, max_package_id
1198 for (i = 0; i <= topology.max_cpu_id; ++i) {
1200 struct cpu_topology *cpu = &topology.cpus[i];
1202 if (cpu_is_not_present(i)) {
1203 WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1207 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1211 cpu->package_id = ret;
1212 if (cpu->package_id > max_package_id)
1213 max_package_id = cpu->package_id;
1215 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1220 if (cpu->core_id > max_core_id)
1221 max_core_id = cpu->core_id;
1222 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1226 cpu->first_core_in_package = 1;
1228 ret = get_threads_on_core(i);
1233 if (num_threads > max_thread_id)
1234 max_thread_id = num_threads;
1235 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1239 cpu->first_thread_in_core = 1;
1241 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1242 i, cpu->package_id, cpu->core_id);
1244 /* Num is max + 1 (need to count 0) */
1245 topology.num_packages = max_package_id + 1;
1246 topology.num_cores = max_core_id + 1;
1247 topology.num_threads = max_thread_id + 1;
1251 free(topology.cpus);
1256 /************************
1257 * Main alloc/init/free *
1258 ************************/
1261 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1264 int total_threads, total_cores;
1266 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1267 *threads = calloc(total_threads, sizeof(struct thread_data));
1268 if (*threads == NULL)
1271 for (i = 0; i < total_threads; ++i)
1272 (*threads)[i].cpu_id = -1;
1274 total_cores = topology.num_cores * topology.num_packages;
1275 *cores = calloc(total_cores, sizeof(struct core_data));
1277 goto err_clean_threads;
1279 for (i = 0; i < total_cores; ++i)
1280 (*cores)[i].core_id = -1;
1282 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1283 if (*packages == NULL)
1284 goto err_clean_cores;
1286 for (i = 0; i < topology.num_packages; i++)
1287 (*packages)[i].package_id = i;
1296 ERROR("Turbostat plugin: Failled to allocate memory for counters");
1301 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1302 struct pkg_data *pkg_base, int cpu_id)
1304 struct thread_data *t;
1305 struct core_data *c;
1307 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1309 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1310 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1311 p = GET_PKG(pkg_base, cpu->package_id);
1314 if (cpu->first_thread_in_core)
1315 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1316 if (cpu->first_core_in_package)
1317 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1319 c->core_id = cpu->core_id;
1320 p->package_id = cpu->package_id;
1324 initialize_counters(void)
1328 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1329 if (cpu_is_not_present(cpu_id))
1331 init_counter(EVEN_COUNTERS, cpu_id);
1332 init_counter(ODD_COUNTERS, cpu_id);
1333 init_counter(DELTA_COUNTERS, cpu_id);
1340 free_all_buffers(void)
1345 CPU_FREE(cpu_present_set);
1346 cpu_present_set = NULL;
1347 cpu_present_set = 0;
1349 CPU_FREE(cpu_affinity_set);
1350 cpu_affinity_set = NULL;
1351 cpu_affinity_setsize = 0;
1353 CPU_FREE(cpu_saved_affinity_set);
1354 cpu_saved_affinity_set = NULL;
1355 cpu_saved_affinity_setsize = 0;
1363 package_even = NULL;
1375 free(package_delta);
1377 thread_delta = NULL;
1379 package_delta = NULL;
1383 /**********************
1384 * Collectd functions *
1385 **********************/
1387 #define DO_OR_GOTO_ERR(something) \
1389 ret = (something); \
1394 static int setup_all_buffers(void)
1398 DO_OR_GOTO_ERR(topology_probe());
1399 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1400 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1401 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1402 initialize_counters();
1403 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1404 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1414 turbostat_read(void)
1419 if ((ret = setup_all_buffers()) < 0)
1423 if (for_all_proc_cpus(cpu_is_not_present)) {
1425 if ((ret = setup_all_buffers()) < 0)
1427 if (for_all_proc_cpus(cpu_is_not_present)) {
1428 ERROR("Turbostat plugin: CPU appeared just after "
1434 /* Saving the scheduling affinity, as it will be modified by get_counters */
1435 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1436 ERROR("Turbostat plugin: Unable to save the CPU affinity");
1441 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1443 gettimeofday(&tv_even, (struct timezone *)NULL);
1451 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1453 gettimeofday(&tv_odd, (struct timezone *)NULL);
1455 timersub(&tv_odd, &tv_even, &tv_delta);
1456 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1458 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1461 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1463 gettimeofday(&tv_even, (struct timezone *)NULL);
1465 timersub(&tv_even, &tv_odd, &tv_delta);
1466 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1468 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1474 * Let's restore the affinity
1475 * This might fail if the number of CPU changed, but we can't do anything in that case..
1477 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1482 turbostat_init(void)
1487 if (getuid() != 0) {
1488 ERROR("Turbostat plugin: Initialization failed: this plugin "
1489 "requires collectd to run as root in order to read "
1490 "special CPU registers");
1494 DO_OR_GOTO_ERR(probe_cpu());
1496 if (stat("/dev/cpu/0/msr", &sb)) {
1497 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1498 " does not exist while the CPU supports MSR. You may be "
1499 "missing the corresponding kernel module, please try '# "
1504 DO_OR_GOTO_ERR(setup_all_buffers());
1506 plugin_register_read(PLUGIN_NAME, turbostat_read);
1514 static const char *config_keys[] =
1516 "TCCActivationTemp",
1518 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
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);