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;
229 ERR_CPU_SAVE_SCHED_AFFINITY,
233 ERR_MSR_CORE_C3_RESIDENCY,
234 ERR_MSR_CORE_C6_RESIDENCY,
235 ERR_MSR_CORE_C7_RESIDENCY,
236 ERR_MSR_IA32_THERM_STATUS,
237 ERR_MSR_PKG_C3_RESIDENCY,
238 ERR_MSR_PKG_C6_RESIDENCY,
239 ERR_MSR_PKG_C2_RESIDENCY,
240 ERR_MSR_PKG_C7_RESIDENCY,
241 ERR_MSR_PKG_C8_RESIDENCY,
242 ERR_MSR_PKG_C9_RESIDENCY,
243 ERR_MSR_PKG_C10_RESIDENCY,
244 ERR_MSR_PKG_ENERGY_STATUS,
245 ERR_MSR_PKG_POWER_INFO,
246 ERR_MSR_PP0_ENERGY_STATUS,
247 ERR_MSR_DRAM_ENERGY_STATUS,
248 ERR_MSR_PP1_ENERGY_STATUS,
249 ERR_MSR_PKG_PERF_STATUS,
250 ERR_MSR_DRAM_PERF_STATUS,
251 ERR_MSR_IA32_PACKAGE_THERM_STATUS,
257 ERR_CANT_READ_NUMBER,
258 ERR_CANT_READ_PROC_STAT,
259 ERR_NO_INVARIANT_TSC,
269 /*****************************
270 * MSR Manipulation helpers *
271 *****************************/
274 * Open a MSR device for reading
275 * Can change the scheduling affinity of the current process if multiple_read is 1
277 static int __attribute__((warn_unused_result))
278 open_msr(int cpu, _Bool multiple_read)
284 * If we need to do multiple read, let's migrate to the CPU
285 * Otherwise, we would lose time calling functions on another CPU
287 * If we are not yet initialized (cpu_affinity_setsize = 0),
288 * we need to skip this optimisation.
290 if (multiple_read && cpu_affinity_setsize) {
291 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
292 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
293 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
294 ERROR("Turbostat plugin: Could not migrate to CPU %d", cpu);
295 return -ERR_CPU_MIGRATE;
299 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
300 fd = open(pathname, O_RDONLY);
302 return -ERR_CANT_OPEN_MSR;
307 * Read a single MSR from an open file descriptor
309 static int __attribute__((warn_unused_result))
310 read_msr(int fd, off_t offset, unsigned long long *msr)
314 retval = pread(fd, msr, sizeof *msr, offset);
316 if (retval != sizeof *msr) {
317 ERROR("Turbostat plugin: MSR offset 0x%llx read failed",
318 (unsigned long long)offset);
325 * Open a MSR device for reading, read the value asked for and close it.
326 * This call will not affect the scheduling affinity of this thread.
328 static int __attribute__((warn_unused_result))
329 get_msr(int cpu, off_t offset, unsigned long long *msr)
334 fd = open_msr(cpu, 0);
337 retval = read_msr(fd, offset, msr);
343 /********************************
344 * Raw data acquisition (1 CPU) *
345 ********************************/
348 * Read every data avalaible for a single CPU
350 * Core data is shared for all threads in one core: extracted only for the first thread
351 * Package data is shared for all core in one package: extracted only for the first thread of the first core
353 * Side effect: migrates to the targeted CPU
355 static int __attribute__((warn_unused_result))
356 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
359 unsigned long long msr;
363 msr_fd = open_msr(cpu, 1);
367 #define READ_MSR(msr, dst) \
369 if (read_msr(msr_fd, msr, dst)) { \
370 retval = -ERR_##msr; \
375 READ_MSR(MSR_IA32_TSC, &t->tsc);
377 READ_MSR(MSR_IA32_APERF, &t->aperf);
378 READ_MSR(MSR_IA32_MPERF, &t->mperf);
380 READ_MSR(MSR_SMI_COUNT, &msr);
381 t->smi_count = msr & 0xFFFFFFFF;
383 /* collect core counters only for 1st thread in core */
384 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
389 if (do_core_cstate & (1 << 3))
390 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
391 if (do_core_cstate & (1 << 6))
392 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
393 if (do_core_cstate & (1 << 7))
394 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
397 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
398 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
401 /* collect package counters only for 1st core in package */
402 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
407 if (do_pkg_cstate & (1 << 2))
408 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
409 if (do_pkg_cstate & (1 << 3))
410 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
411 if (do_pkg_cstate & (1 << 6))
412 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
413 if (do_pkg_cstate & (1 << 7))
414 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
415 if (do_pkg_cstate & (1 << 8))
416 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
417 if (do_pkg_cstate & (1 << 9))
418 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
419 if (do_pkg_cstate & (1 << 10))
420 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
422 if (do_rapl & RAPL_PKG) {
423 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
424 p->energy_pkg = msr & 0xFFFFFFFF;
426 if (do_rapl & RAPL_CORES) {
427 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
428 p->energy_cores = msr & 0xFFFFFFFF;
430 if (do_rapl & RAPL_DRAM) {
431 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
432 p->energy_dram = msr & 0xFFFFFFFF;
434 if (do_rapl & RAPL_GFX) {
435 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
436 p->energy_gfx = msr & 0xFFFFFFFF;
438 if (do_rapl & RAPL_PKG_PERF_STATUS) {
439 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
440 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
442 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
443 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
444 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
447 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
448 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
457 /**********************************
458 * Evaluating the changes (1 CPU) *
459 **********************************/
462 * Do delta = new - old on 32bits cyclique intergers
464 #define DELTA_WRAP32(delta, new, old) \
468 delta = 0x100000000 + new - old; \
472 * Extract the evolution old->new in delta at a package level
473 * (some are not new-delta, e.g. temperature)
476 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
478 delta->pc2 = new->pc2 - old->pc2;
479 delta->pc3 = new->pc3 - old->pc3;
480 delta->pc6 = new->pc6 - old->pc6;
481 delta->pc7 = new->pc7 - old->pc7;
482 delta->pc8 = new->pc8 - old->pc8;
483 delta->pc9 = new->pc9 - old->pc9;
484 delta->pc10 = new->pc10 - old->pc10;
485 delta->pkg_temp_c = new->pkg_temp_c;
487 DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
488 DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
489 DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
490 DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
491 DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
492 DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
496 * Extract the evolution old->new in delta at a core level
497 * (some are not new-delta, e.g. temperature)
500 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
502 delta->c3 = new->c3 - old->c3;
503 delta->c6 = new->c6 - old->c6;
504 delta->c7 = new->c7 - old->c7;
505 delta->core_temp_c = new->core_temp_c;
509 * Extract the evolution old->new in delta at a package level
510 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
512 static inline int __attribute__((warn_unused_result))
513 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
514 const struct core_data *core_delta)
516 delta->tsc = new->tsc - old->tsc;
518 /* check for TSC < 1 Mcycles over interval */
519 if (delta->tsc < (1000 * 1000)) {
520 WARNING("Turbostat plugin: Insanely slow TSC rate, TSC stops "
521 "in idle? You can disable all c-states by booting with"
522 " 'idle=poll' or just the deep ones with"
523 " 'processor.max_cstate=1'");
527 delta->c1 = new->c1 - old->c1;
529 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
530 delta->aperf = new->aperf - old->aperf;
531 delta->mperf = new->mperf - old->mperf;
533 if (!aperf_mperf_unstable) {
534 WARNING("Turbostat plugin: APERF or MPERF went "
535 "backwards. Frequency results do not cover "
536 "the entire interval. Fix this by running "
537 "Linux-2.6.30 or later.");
539 aperf_mperf_unstable = 1;
544 * As counter collection is not atomic,
545 * it is possible for mperf's non-halted cycles + idle states
546 * to exceed TSC's all cycles: show c1 = 0% in that case.
548 if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
551 /* normal case, derive c1 */
552 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
553 - core_delta->c6 - core_delta->c7;
556 if (delta->mperf == 0) {
557 WARNING("Turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
558 delta->mperf = 1; /* divide by 0 protection */
561 delta->smi_count = new->smi_count - old->smi_count;
566 /**********************************
567 * Submitting the results (1 CPU) *
568 **********************************/
571 * Submit one gauge value
574 turbostat_submit (const char *plugin_instance,
575 const char *type, const char *type_instance,
578 value_list_t vl = VALUE_LIST_INIT;
584 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
585 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
586 if (plugin_instance != NULL)
587 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
588 sstrncpy (vl.type, type, sizeof (vl.type));
589 if (type_instance != NULL)
590 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
592 plugin_dispatch_values (&vl);
596 * Submit every data for a single CPU
598 * Core data is shared for all threads in one core: submitted only for the first thread
599 * Package data is shared for all core in one package: submitted only for the first thread of the first core
602 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
605 double interval_float;
607 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
609 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
611 if (!aperf_mperf_unstable)
612 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
613 if (!aperf_mperf_unstable)
614 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
617 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
618 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
621 turbostat_submit(NULL, "current", name, t->smi_count);
623 /* submit per-core data only for 1st thread in core */
624 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
627 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
629 if (do_core_cstate & (1 << 3))
630 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
631 if (do_core_cstate & (1 << 6))
632 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
633 if (do_core_cstate & (1 << 7))
634 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
637 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
639 /* submit per-package data only for 1st core in package */
640 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
643 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
646 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
648 if (do_pkg_cstate & (1 << 2))
649 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
650 if (do_pkg_cstate & (1 << 3))
651 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
652 if (do_pkg_cstate & (1 << 6))
653 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
654 if (do_pkg_cstate & (1 << 7))
655 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
656 if (do_pkg_cstate & (1 << 8))
657 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
658 if (do_pkg_cstate & (1 << 9))
659 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
660 if (do_pkg_cstate & (1 << 10))
661 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
664 if (do_rapl & RAPL_PKG)
665 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
666 if (do_rapl & RAPL_CORES)
667 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
668 if (do_rapl & RAPL_GFX)
669 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
670 if (do_rapl & RAPL_DRAM)
671 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
678 /**********************************
679 * Looping function over all CPUs *
680 **********************************/
683 * Check if a given cpu id is in our compiled list of existing CPUs
686 cpu_is_not_present(int cpu)
688 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
692 * Loop on all CPUs in topological order
694 * Skip non-present cpus
695 * Return the error code at the first error or 0
697 static int __attribute__((warn_unused_result))
698 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
699 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
701 int retval, pkg_no, core_no, thread_no;
703 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
704 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
705 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
706 struct thread_data *t;
710 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
712 if (cpu_is_not_present(t->cpu_id))
715 c = GET_CORE(core_base, core_no, pkg_no);
716 p = GET_PKG(pkg_base, pkg_no);
718 retval = func(t, c, p);
728 * Dedicated loop: Extract every data evolution for all CPU
730 * Skip non-present cpus
731 * Return the error code at the first error or 0
733 * Core data is shared for all threads in one core: extracted only for the first thread
734 * Package data is shared for all core in one package: extracted only for the first thread of the first core
736 static int __attribute__((warn_unused_result))
737 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,
738 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
740 int retval, pkg_no, core_no, thread_no;
742 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
743 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
744 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
745 struct thread_data *t_delta;
746 const struct thread_data *t_old, *t_new;
747 struct core_data *c_delta;
749 /* Get correct pointers for threads */
750 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
751 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
752 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
754 /* Skip threads that disappeared */
755 if (cpu_is_not_present(t_delta->cpu_id))
758 /* c_delta is always required for delta_thread */
759 c_delta = GET_CORE(core_delta, core_no, pkg_no);
761 /* calculate core delta only for 1st thread in core */
762 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
763 const struct core_data *c_old, *c_new;
765 c_new = GET_CORE(core_new_base, core_no, pkg_no);
766 c_old = GET_CORE(core_old_base, core_no, pkg_no);
768 delta_core(c_delta, c_new, c_old);
771 /* Always calculate thread delta */
772 retval = delta_thread(t_delta, t_new, t_old, c_delta);
776 /* calculate package delta only for 1st core in package */
777 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
778 struct pkg_data *p_delta;
779 const struct pkg_data *p_old, *p_new;
781 p_delta = GET_PKG(package_delta, pkg_no);
782 p_new = GET_PKG(pkg_new_base, pkg_no);
783 p_old = GET_PKG(pkg_old_base, pkg_no);
785 delta_package(p_delta, p_new, p_old);
799 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
800 * the Thermal Control Circuit (TCC) activates.
801 * This is usually equal to tjMax.
803 * Older processors do not have this MSR, so there we guess,
804 * but also allow conficuration over-ride with "TCCActivationTemp".
806 * Several MSR temperature values are in units of degrees-C
807 * below this value, including the Digital Thermal Sensor (DTS),
808 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
810 static int __attribute__((warn_unused_result))
811 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
813 unsigned long long msr;
814 unsigned int target_c_local;
816 /* tcc_activation_temp is used only for dts or ptm */
817 if (!(do_dts || do_ptm))
820 /* this is a per-package concept */
821 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
824 if (tcc_activation_temp != 0) {
825 p->tcc_activation_temp = tcc_activation_temp;
829 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
832 target_c_local = (msr >> 16) & 0xFF;
837 p->tcc_activation_temp = target_c_local;
842 p->tcc_activation_temp = TJMAX_DEFAULT;
843 WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
844 " Please use TCCActivationTemp to specify it.",
845 t->cpu_id, p->tcc_activation_temp);
851 * Identify the functionality of the CPU
853 static int __attribute__((warn_unused_result))
856 unsigned int eax, ebx, ecx, edx, max_level;
857 unsigned int fms, family, model;
860 * - EAX: Maximum Input Value for Basic CPUID Information
861 * - EBX: "Genu" (0x756e6547)
862 * - EDX: "ineI" (0x49656e69)
863 * - ECX: "ntel" (0x6c65746e)
865 max_level = ebx = ecx = edx = 0;
866 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
867 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
868 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
869 return -UNSUPPORTED_CPU;
873 * - EAX: Version Information: Type, Family, Model, and Stepping ID
876 * + 12-13: Processor type
877 * + 16-19: Extended Model ID
878 * + 20-27: Extended Family ID
879 * - EDX: Feature Information:
880 * + 5: Support for MSR read/write operations
882 fms = ebx = ecx = edx = 0;
883 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
884 family = (fms >> 8) & 0xf;
885 model = (fms >> 4) & 0xf;
887 family += (fms >> 20) & 0xf;
888 if (family == 6 || family == 0xf)
889 model += ((fms >> 16) & 0xf) << 4;
890 if (!(edx & (1 << 5))) {
891 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
897 * - EAX: Maximum Input Value for Extended Function CPUID Information
899 * This allows us to verify if the CPUID(0x80000007) can be called
901 * This check is valid for both Intel and AMD.
903 max_level = ebx = ecx = edx = 0;
904 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
905 if (max_level < 0x80000007) {
906 ERROR("Turbostat plugin: Unsupported CPU (no invariant TSC, "
907 " Maximum Extended Function: 0x%x)", max_level);
908 return -ERR_NO_INVARIANT_TSC;
914 * + 8: Invariant TSC available if set
916 * This check is valid for both Intel and AMD
918 eax = ebx = ecx = edx = 0;
919 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
920 if (!(edx & (1 << 8))) {
921 ERROR("Turbostat plugin: Unsupported CPU (No invariant TSC)");
922 return -ERR_NO_INVARIANT_TSC;
928 * + 0: Digital temperature sensor is supported if set
929 * + 6: Package thermal management is supported if set
931 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
932 * + 3: The processor supports performance-energy bias preference if set.
933 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
935 * This check is valid for both Intel and AMD
937 eax = ebx = ecx = edx = 0;
938 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
939 do_dts = eax & (1 << 0);
940 do_ptm = eax & (1 << 6);
941 if (!(ecx & (1 << 0))) {
942 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
943 return -ERR_NO_APERF;
947 * Enable or disable C states depending on the model and family
954 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
962 do_core_cstate = (1 << 1) | (1 << 6);
963 do_pkg_cstate = (1 << 6);
966 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
967 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
968 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
969 case 0x2E: /* Nehalem-EX Xeon - Beckton */
970 do_core_cstate = (1 << 3) | (1 << 6);
971 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
974 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
975 case 0x2C: /* Westmere EP - Gulftown */
976 case 0x2F: /* Westmere-EX Xeon - Eagleton */
977 do_core_cstate = (1 << 3) | (1 << 6);
978 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
982 case 0x2D: /* SNB Xeon */
983 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
984 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
988 case 0x3E: /* IVB Xeon */
989 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
990 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
996 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
997 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1000 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1001 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1004 case 0x4F: /* BDW */
1005 case 0x56: /* BDX-DE */
1006 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1007 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1009 case 0x3D: /* BDW */
1010 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1011 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1014 ERROR("Turbostat plugin: Unsupported CPU (family: %#x,"
1015 " model: %#x)", family, model);
1023 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1026 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1030 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1034 do_rapl = RAPL_PKG | RAPL_CORES;
1040 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
1041 "model: %#x)", family, model);
1042 return -UNSUPPORTED_CPU;
1046 unsigned long long msr;
1047 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1051 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1053 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1060 /********************
1061 * Topology Probing *
1062 ********************/
1065 * Read a single int from a file.
1067 static int __attribute__ ((format(printf,1,2)))
1068 parse_int_file(const char *fmt, ...)
1071 char path[PATH_MAX];
1075 va_start(args, fmt);
1076 len = vsnprintf(path, sizeof(path), fmt, args);
1078 if (len < 0 || len >= PATH_MAX) {
1079 ERROR("Turbostat plugin: path truncated: '%s'", path);
1080 return -ERR_PATH_TOO_LONG;
1083 filep = fopen(path, "r");
1085 ERROR("Turbostat plugin: Failed to open '%s'", path);
1086 return -ERR_CANT_OPEN_FILE;
1088 if (fscanf(filep, "%d", &value) != 1) {
1089 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1090 return -ERR_CANT_READ_NUMBER;
1097 get_threads_on_core(int cpu)
1105 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1106 filep = fopen(path, "r");
1108 ERROR("Turbostat plugin: Failed to open '%s'", path);
1109 return -ERR_CANT_OPEN_FILE;
1113 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1114 * otherwinse 1 sibling (self).
1116 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1127 * run func(cpu) on every cpu in /proc/stat
1128 * return max_cpu number
1130 static int __attribute__((warn_unused_result))
1131 for_all_proc_cpus(int (func)(int))
1137 fp = fopen("/proc/stat", "r");
1139 ERROR("Turbostat plugin: Failed to open /proc/stat");
1140 return -ERR_CANT_OPEN_FILE;
1143 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1145 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1147 return -ERR_CANT_READ_PROC_STAT;
1151 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1155 retval = func(cpu_num);
1166 * Update the stored topology.max_cpu_id
1169 update_max_cpu_id(int cpu)
1171 if (topology.max_cpu_id < cpu)
1172 topology.max_cpu_id = cpu;
1177 mark_cpu_present(int cpu)
1179 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1183 static int __attribute__((warn_unused_result))
1184 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1185 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1187 ERROR("Turbostat plugin: Unable to allocate CPU state");
1188 return -ERR_CPU_ALLOC;
1190 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1191 CPU_ZERO_S(*size, *set);
1196 * Build a local representation of the cpu distribution
1198 static int __attribute__((warn_unused_result))
1203 int max_package_id, max_core_id, max_thread_id;
1204 max_package_id = max_core_id = max_thread_id = 0;
1206 /* Clean topology */
1207 free(topology.cpus);
1208 memset(&topology, 0, sizeof(topology));
1210 ret = for_all_proc_cpus(update_max_cpu_id);
1214 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1215 if (topology.cpus == NULL) {
1216 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1220 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1223 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1226 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1230 ret = for_all_proc_cpus(mark_cpu_present);
1236 * find max_core_id, max_package_id
1238 for (i = 0; i <= topology.max_cpu_id; ++i) {
1240 struct cpu_topology *cpu = &topology.cpus[i];
1242 if (cpu_is_not_present(i)) {
1243 WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1247 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1251 cpu->package_id = ret;
1252 if (cpu->package_id > max_package_id)
1253 max_package_id = cpu->package_id;
1255 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1260 if (cpu->core_id > max_core_id)
1261 max_core_id = cpu->core_id;
1262 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1266 cpu->first_core_in_package = 1;
1268 ret = get_threads_on_core(i);
1273 if (num_threads > max_thread_id)
1274 max_thread_id = num_threads;
1275 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1279 cpu->first_thread_in_core = 1;
1281 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1282 i, cpu->package_id, cpu->core_id);
1284 /* Num is max + 1 (need to count 0) */
1285 topology.num_packages = max_package_id + 1;
1286 topology.num_cores = max_core_id + 1;
1287 topology.num_threads = max_thread_id + 1;
1291 free(topology.cpus);
1296 /************************
1297 * Main alloc/init/free *
1298 ************************/
1301 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1304 int total_threads, total_cores;
1306 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1307 *threads = calloc(total_threads, sizeof(struct thread_data));
1308 if (*threads == NULL)
1311 for (i = 0; i < total_threads; ++i)
1312 (*threads)[i].cpu_id = -1;
1314 total_cores = topology.num_cores * topology.num_packages;
1315 *cores = calloc(total_cores, sizeof(struct core_data));
1317 goto err_clean_threads;
1319 for (i = 0; i < total_cores; ++i)
1320 (*cores)[i].core_id = -1;
1322 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1323 if (*packages == NULL)
1324 goto err_clean_cores;
1326 for (i = 0; i < topology.num_packages; i++)
1327 (*packages)[i].package_id = i;
1336 ERROR("Turbostat plugin: Failled to allocate memory for counters");
1341 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1342 struct pkg_data *pkg_base, int cpu_id)
1344 struct thread_data *t;
1345 struct core_data *c;
1347 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1349 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1350 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1351 p = GET_PKG(pkg_base, cpu->package_id);
1354 if (cpu->first_thread_in_core)
1355 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1356 if (cpu->first_core_in_package)
1357 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1359 c->core_id = cpu->core_id;
1360 p->package_id = cpu->package_id;
1366 initialize_counters(void)
1371 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1372 if (cpu_is_not_present(cpu_id)) {
1376 ret = init_counter(EVEN_COUNTERS, cpu_id);
1379 ret = init_counter(ODD_COUNTERS, cpu_id);
1382 ret = init_counter(DELTA_COUNTERS, cpu_id);
1392 free_all_buffers(void)
1397 CPU_FREE(cpu_present_set);
1398 cpu_present_set = NULL;
1399 cpu_present_set = 0;
1401 CPU_FREE(cpu_affinity_set);
1402 cpu_affinity_set = NULL;
1403 cpu_affinity_setsize = 0;
1405 CPU_FREE(cpu_saved_affinity_set);
1406 cpu_saved_affinity_set = NULL;
1407 cpu_saved_affinity_setsize = 0;
1415 package_even = NULL;
1427 free(package_delta);
1429 thread_delta = NULL;
1431 package_delta = NULL;
1435 /**********************
1436 * Collectd functions *
1437 **********************/
1439 #define DO_OR_GOTO_ERR(something) \
1441 ret = (something); \
1446 static int setup_all_buffers(void)
1450 DO_OR_GOTO_ERR(topology_probe());
1451 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1452 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1453 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1454 DO_OR_GOTO_ERR(initialize_counters());
1455 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1456 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1466 turbostat_read(void)
1471 if ((ret = setup_all_buffers()) < 0)
1475 if (for_all_proc_cpus(cpu_is_not_present)) {
1477 if ((ret = setup_all_buffers()) < 0)
1479 if (for_all_proc_cpus(cpu_is_not_present))
1480 return -ERR_CPU_NOT_PRESENT;
1483 /* Saving the scheduling affinity, as it will be modified by get_counters */
1484 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0)
1485 return -ERR_CPU_SAVE_SCHED_AFFINITY;
1488 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1490 gettimeofday(&tv_even, (struct timezone *)NULL);
1498 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1500 gettimeofday(&tv_odd, (struct timezone *)NULL);
1502 timersub(&tv_odd, &tv_even, &tv_delta);
1503 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1505 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1508 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1510 gettimeofday(&tv_even, (struct timezone *)NULL);
1512 timersub(&tv_even, &tv_odd, &tv_delta);
1513 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1515 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1521 * Let's restore the affinity
1522 * This might fail if the number of CPU changed, but we can't do anything in that case..
1524 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1529 turbostat_init(void)
1534 if (getuid() != 0) {
1535 ERROR("Turbostat plugin: Initialization failed: this plugin "
1536 "requires collectd to run as root in order to read "
1537 "special CPU registers");
1538 return -ERR_NOT_ROOT;
1541 DO_OR_GOTO_ERR(probe_cpu());
1543 if (stat("/dev/cpu/0/msr", &sb)) {
1544 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1545 " does not exist while the CPU supports MSR. You may be "
1546 "missing the corresponding kernel module, please try '# "
1551 DO_OR_GOTO_ERR(setup_all_buffers());
1553 plugin_register_read(PLUGIN_NAME, turbostat_read);
1561 static const char *config_keys[] =
1563 "TCCActivationTemp",
1565 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1568 turbostat_config(const char *key, const char *value)
1570 long unsigned int tmp_val;
1573 if (strcasecmp("TCCActivationTemp", key) == 0) {
1574 tmp_val = strtoul(value, &end, 0);
1575 if (*end != '\0' || tmp_val > UINT_MAX)
1577 tcc_activation_temp = (unsigned int) tmp_val;
1584 void module_register(void)
1586 plugin_register_init(PLUGIN_NAME, turbostat_init);
1587 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);