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,
268 /*****************************
269 * MSR Manipulation helpers *
270 *****************************/
273 * Open a MSR device for reading
274 * Can change the scheduling affinity of the current process if multiple_read is 1
276 static int __attribute__((warn_unused_result))
277 open_msr(int cpu, _Bool multiple_read)
283 * If we need to do multiple read, let's migrate to the CPU
284 * Otherwise, we would lose time calling functions on another CPU
286 * If we are not yet initialized (cpu_affinity_setsize = 0),
287 * we need to skip this optimisation.
289 if (multiple_read && cpu_affinity_setsize) {
290 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
291 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
292 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
293 ERROR("Could not migrate to CPU %d", cpu);
294 return -ERR_CPU_MIGRATE;
298 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
299 fd = open(pathname, O_RDONLY);
301 return -ERR_CANT_OPEN_MSR;
306 * Read a single MSR from an open file descriptor
308 static int __attribute__((warn_unused_result))
309 read_msr(int fd, off_t offset, unsigned long long *msr)
313 retval = pread(fd, msr, sizeof *msr, offset);
315 if (retval != sizeof *msr) {
316 ERROR("MSR offset 0x%llx read failed", (unsigned long long)offset);
323 * Open a MSR device for reading, read the value asked for and close it.
324 * This call will not affect the scheduling affinity of this thread.
326 static int __attribute__((warn_unused_result))
327 get_msr(int cpu, off_t offset, unsigned long long *msr)
332 fd = open_msr(cpu, 0);
335 retval = read_msr(fd, offset, msr);
341 /********************************
342 * Raw data acquisition (1 CPU) *
343 ********************************/
346 * Read every data avalaible for a single CPU
348 * Core data is shared for all threads in one core: extracted only for the first thread
349 * Package data is shared for all core in one package: extracted only for the first thread of the first core
351 * Side effect: migrates to the targeted CPU
353 static int __attribute__((warn_unused_result))
354 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
357 unsigned long long msr;
361 msr_fd = open_msr(cpu, 1);
365 #define READ_MSR(msr, dst) \
367 if (read_msr(msr_fd, msr, dst)) { \
368 retval = -ERR_##msr; \
373 READ_MSR(MSR_IA32_TSC, &t->tsc);
375 READ_MSR(MSR_IA32_APERF, &t->aperf);
376 READ_MSR(MSR_IA32_MPERF, &t->mperf);
378 READ_MSR(MSR_SMI_COUNT, &msr);
379 t->smi_count = msr & 0xFFFFFFFF;
381 /* collect core counters only for 1st thread in core */
382 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
387 if (do_core_cstate & (1 << 3))
388 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
389 if (do_core_cstate & (1 << 6))
390 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
391 if (do_core_cstate & (1 << 7))
392 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
395 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
396 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
399 /* collect package counters only for 1st core in package */
400 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
405 if (do_pkg_cstate & (1 << 2))
406 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
407 if (do_pkg_cstate & (1 << 3))
408 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
409 if (do_pkg_cstate & (1 << 6))
410 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
411 if (do_pkg_cstate & (1 << 7))
412 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
413 if (do_pkg_cstate & (1 << 8))
414 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
415 if (do_pkg_cstate & (1 << 9))
416 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
417 if (do_pkg_cstate & (1 << 10))
418 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
420 if (do_rapl & RAPL_PKG) {
421 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
422 p->energy_pkg = msr & 0xFFFFFFFF;
424 if (do_rapl & RAPL_CORES) {
425 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
426 p->energy_cores = msr & 0xFFFFFFFF;
428 if (do_rapl & RAPL_DRAM) {
429 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
430 p->energy_dram = msr & 0xFFFFFFFF;
432 if (do_rapl & RAPL_GFX) {
433 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
434 p->energy_gfx = msr & 0xFFFFFFFF;
436 if (do_rapl & RAPL_PKG_PERF_STATUS) {
437 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
438 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
440 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
441 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
442 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
445 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
446 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
455 /**********************************
456 * Evaluating the changes (1 CPU) *
457 **********************************/
460 * Do delta = new - old on 32bits cyclique intergers
462 #define DELTA_WRAP32(delta, new, old) \
466 delta = 0x100000000 + new - old; \
470 * Extract the evolution old->new in delta at a package level
471 * (some are not new-delta, e.g. temperature)
474 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
476 delta->pc2 = new->pc2 - old->pc2;
477 delta->pc3 = new->pc3 - old->pc3;
478 delta->pc6 = new->pc6 - old->pc6;
479 delta->pc7 = new->pc7 - old->pc7;
480 delta->pc8 = new->pc8 - old->pc8;
481 delta->pc9 = new->pc9 - old->pc9;
482 delta->pc10 = new->pc10 - old->pc10;
483 delta->pkg_temp_c = new->pkg_temp_c;
485 DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
486 DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
487 DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
488 DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
489 DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
490 DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
494 * Extract the evolution old->new in delta at a core level
495 * (some are not new-delta, e.g. temperature)
498 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
500 delta->c3 = new->c3 - old->c3;
501 delta->c6 = new->c6 - old->c6;
502 delta->c7 = new->c7 - old->c7;
503 delta->core_temp_c = new->core_temp_c;
507 * Extract the evolution old->new in delta at a package level
508 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
510 static inline int __attribute__((warn_unused_result))
511 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
512 const struct core_data *core_delta)
514 delta->tsc = new->tsc - old->tsc;
516 /* check for TSC < 1 Mcycles over interval */
517 if (delta->tsc < (1000 * 1000)) {
518 WARNING("Insanely slow TSC rate, TSC stops in idle? ");
519 WARNING("You can disable all c-states by booting with \"idle=poll\" ");
520 WARNING("or just the deep ones with \"processor.max_cstate=1\"");
524 delta->c1 = new->c1 - old->c1;
526 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
527 delta->aperf = new->aperf - old->aperf;
528 delta->mperf = new->mperf - old->mperf;
530 if (!aperf_mperf_unstable) {
531 WARNING(" APERF or MPERF went backwards * ");
532 WARNING("* Frequency results do not cover entire interval *");
533 WARNING("* fix this by running Linux-2.6.30 or later *");
535 aperf_mperf_unstable = 1;
540 * As counter collection is not atomic,
541 * it is possible for mperf's non-halted cycles + idle states
542 * to exceed TSC's all cycles: show c1 = 0% in that case.
544 if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
547 /* normal case, derive c1 */
548 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
549 - core_delta->c6 - core_delta->c7;
552 if (delta->mperf == 0) {
553 WARNING("cpu%d MPERF 0!", old->cpu_id);
554 delta->mperf = 1; /* divide by 0 protection */
557 delta->smi_count = new->smi_count - old->smi_count;
562 /**********************************
563 * Submitting the results (1 CPU) *
564 **********************************/
567 * Submit one gauge value
570 turbostat_submit (const char *plugin_instance,
571 const char *type, const char *type_instance,
574 value_list_t vl = VALUE_LIST_INIT;
580 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
581 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
582 if (plugin_instance != NULL)
583 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
584 sstrncpy (vl.type, type, sizeof (vl.type));
585 if (type_instance != NULL)
586 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
588 plugin_dispatch_values (&vl);
592 * Submit every data for a single CPU
594 * Core data is shared for all threads in one core: submitted only for the first thread
595 * Package data is shared for all core in one package: submitted only for the first thread of the first core
598 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
601 double interval_float;
603 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
605 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
607 if (!aperf_mperf_unstable)
608 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
609 if (!aperf_mperf_unstable)
610 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
613 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
614 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
617 turbostat_submit(NULL, "current", name, t->smi_count);
619 /* submit per-core data only for 1st thread in core */
620 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
623 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
625 if (do_core_cstate & (1 << 3))
626 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
627 if (do_core_cstate & (1 << 6))
628 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
629 if (do_core_cstate & (1 << 7))
630 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
633 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
635 /* submit per-package data only for 1st core in package */
636 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
639 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
642 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
644 if (do_pkg_cstate & (1 << 2))
645 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
646 if (do_pkg_cstate & (1 << 3))
647 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
648 if (do_pkg_cstate & (1 << 6))
649 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
650 if (do_pkg_cstate & (1 << 7))
651 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
652 if (do_pkg_cstate & (1 << 8))
653 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
654 if (do_pkg_cstate & (1 << 9))
655 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
656 if (do_pkg_cstate & (1 << 10))
657 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
660 if (do_rapl & RAPL_PKG)
661 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
662 if (do_rapl & RAPL_CORES)
663 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
664 if (do_rapl & RAPL_GFX)
665 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
666 if (do_rapl & RAPL_DRAM)
667 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
674 /**********************************
675 * Looping function over all CPUs *
676 **********************************/
679 * Check if a given cpu id is in our compiled list of existing CPUs
682 cpu_is_not_present(int cpu)
684 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
688 * Loop on all CPUs in topological order
690 * Skip non-present cpus
691 * Return the error code at the first error or 0
693 static int __attribute__((warn_unused_result))
694 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
695 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_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;
706 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
708 if (cpu_is_not_present(t->cpu_id))
711 c = GET_CORE(core_base, core_no, pkg_no);
712 p = GET_PKG(pkg_base, pkg_no);
714 retval = func(t, c, p);
724 * Dedicated loop: Extract every data evolution for all CPU
726 * Skip non-present cpus
727 * Return the error code at the first error or 0
729 * Core data is shared for all threads in one core: extracted only for the first thread
730 * Package data is shared for all core in one package: extracted only for the first thread of the first core
732 static int __attribute__((warn_unused_result))
733 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,
734 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
736 int retval, pkg_no, core_no, thread_no;
738 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
739 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
740 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
741 struct thread_data *t_delta;
742 const struct thread_data *t_old, *t_new;
743 struct core_data *c_delta;
745 /* Get correct pointers for threads */
746 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
747 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
748 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
750 /* Skip threads that disappeared */
751 if (cpu_is_not_present(t_delta->cpu_id))
754 /* c_delta is always required for delta_thread */
755 c_delta = GET_CORE(core_delta, core_no, pkg_no);
757 /* calculate core delta only for 1st thread in core */
758 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
759 const struct core_data *c_old, *c_new;
761 c_new = GET_CORE(core_new_base, core_no, pkg_no);
762 c_old = GET_CORE(core_old_base, core_no, pkg_no);
764 delta_core(c_delta, c_new, c_old);
767 /* Always calculate thread delta */
768 retval = delta_thread(t_delta, t_new, t_old, c_delta);
772 /* calculate package delta only for 1st core in package */
773 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
774 struct pkg_data *p_delta;
775 const struct pkg_data *p_old, *p_new;
777 p_delta = GET_PKG(package_delta, pkg_no);
778 p_new = GET_PKG(pkg_new_base, pkg_no);
779 p_old = GET_PKG(pkg_old_base, pkg_no);
781 delta_package(p_delta, p_new, p_old);
791 free_all_buffers(void)
796 CPU_FREE(cpu_present_set);
797 cpu_present_set = NULL;
800 CPU_FREE(cpu_affinity_set);
801 cpu_affinity_set = NULL;
802 cpu_affinity_setsize = 0;
804 CPU_FREE(cpu_saved_affinity_set);
805 cpu_saved_affinity_set = NULL;
806 cpu_saved_affinity_setsize = 0;
830 package_delta = NULL;
839 * Read a single int from a file.
841 static int __attribute__ ((format(printf,1,2)))
842 parse_int_file(const char *fmt, ...)
850 vsnprintf(path, sizeof(path), fmt, args);
852 filep = fopen(path, "r");
854 ERROR("%s: open failed", path);
855 return -ERR_CANT_OPEN_FILE;
857 if (fscanf(filep, "%d", &value) != 1) {
858 ERROR("%s: failed to parse number from file", path);
859 return -ERR_CANT_READ_NUMBER;
866 get_threads_on_core(int cpu)
874 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
875 filep = fopen(path, "r");
877 ERROR("%s: open failed", path);
878 return -ERR_CANT_OPEN_FILE;
882 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
883 * otherwinse 1 sibling (self).
885 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
896 * run func(cpu) on every cpu in /proc/stat
897 * return max_cpu number
899 static int __attribute__((warn_unused_result))
900 for_all_proc_cpus(int (func)(int))
906 fp = fopen("/proc/stat", "r");
908 ERROR("Failed to open /proc/stat");
909 return -ERR_CANT_OPEN_FILE;
912 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
914 ERROR("Failed to parse /proc/stat");
916 return -ERR_CANT_READ_PROC_STAT;
920 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
924 retval = func(cpu_num);
935 * Update the stored topology.max_cpu_id
938 update_max_cpu_id(int cpu)
940 if (topology.max_cpu_id < cpu)
941 topology.max_cpu_id = cpu;
946 mark_cpu_present(int cpu)
948 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
953 static int setup_all_buffers(void);
956 turbostat_read(user_data_t * not_used)
961 if ((ret = setup_all_buffers()) < 0)
965 if (for_all_proc_cpus(cpu_is_not_present)) {
967 if ((ret = setup_all_buffers()) < 0)
969 if (for_all_proc_cpus(cpu_is_not_present))
970 return -ERR_CPU_NOT_PRESENT;
973 /* Saving the scheduling affinity, as it will be modified by get_counters */
974 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0)
975 return -ERR_CPU_SAVE_SCHED_AFFINITY;
978 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
980 gettimeofday(&tv_even, (struct timezone *)NULL);
988 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
990 gettimeofday(&tv_odd, (struct timezone *)NULL);
992 timersub(&tv_odd, &tv_even, &tv_delta);
993 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
995 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
998 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1000 gettimeofday(&tv_even, (struct timezone *)NULL);
1002 timersub(&tv_even, &tv_odd, &tv_delta);
1003 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1005 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1011 * Let's restore the affinity
1012 * This might fail if the number of CPU changed, but we can't do anything in that case..
1014 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1018 static int __attribute__((warn_unused_result))
1023 if (stat("/dev/cpu/0/msr", &sb)) {
1024 ERROR("no /dev/cpu/0/msr, try \"# modprobe msr\"");
1030 static int __attribute__((warn_unused_result))
1033 if (getuid() != 0) {
1034 ERROR("must be root");
1035 return -ERR_NOT_ROOT;
1041 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1042 * the Thermal Control Circuit (TCC) activates.
1043 * This is usually equal to tjMax.
1045 * Older processors do not have this MSR, so there we guess,
1046 * but also allow conficuration over-ride with "TCCActivationTemp".
1048 * Several MSR temperature values are in units of degrees-C
1049 * below this value, including the Digital Thermal Sensor (DTS),
1050 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1052 static int __attribute__((warn_unused_result))
1053 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1055 unsigned long long msr;
1056 unsigned int target_c_local;
1058 /* tcc_activation_temp is used only for dts or ptm */
1059 if (!(do_dts || do_ptm))
1062 /* this is a per-package concept */
1063 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1066 if (tcc_activation_temp != 0) {
1067 p->tcc_activation_temp = tcc_activation_temp;
1071 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
1074 target_c_local = (msr >> 16) & 0xFF;
1076 if (!target_c_local)
1079 p->tcc_activation_temp = target_c_local;
1084 p->tcc_activation_temp = TJMAX_DEFAULT;
1085 WARNING("cpu%d: Guessing tjMax %d C, Please use TCCActivationTemp to specify",
1086 t->cpu_id, p->tcc_activation_temp);
1092 * Identify the functionality of the CPU
1094 static int __attribute__((warn_unused_result))
1097 unsigned int eax, ebx, ecx, edx, max_level;
1098 unsigned int fms, family, model;
1101 * - EAX: Maximum Input Value for Basic CPUID Information
1102 * - EBX: "Genu" (0x756e6547)
1103 * - EDX: "ineI" (0x49656e69)
1104 * - ECX: "ntel" (0x6c65746e)
1106 max_level = ebx = ecx = edx = 0;
1107 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
1108 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
1109 ERROR("Unsupported CPU");
1110 return -UNSUPPORTED_CPU;
1114 * - EAX: Version Information: Type, Family, Model, and Stepping ID
1117 * + 12-13: Processor type
1118 * + 16-19: Extended Model ID
1119 * + 20-27: Extended Family ID
1120 * - EDX: Feature Information:
1121 * + 5: Support for MSR read/write operations
1123 fms = ebx = ecx = edx = 0;
1124 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
1125 family = (fms >> 8) & 0xf;
1126 model = (fms >> 4) & 0xf;
1128 family += (fms >> 20) & 0xf;
1129 if (family == 6 || family == 0xf)
1130 model += ((fms >> 16) & 0xf) << 4;
1131 if (!(edx & (1 << 5))) {
1132 ERROR("CPUID: no MSR");
1137 * CPUID(0x80000000):
1138 * - EAX: Maximum Input Value for Extended Function CPUID Information
1140 * This allows us to verify if the CPUID(0x80000007) can be called
1142 * This check is valid for both Intel and AMD.
1144 max_level = ebx = ecx = edx = 0;
1145 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
1146 if (max_level < 0x80000007) {
1147 ERROR("CPUID: no invariant TSC (max_level 0x%x)", max_level);
1148 return -ERR_NO_INVARIANT_TSC;
1152 * CPUID(0x80000007):
1154 * + 8: Invariant TSC available if set
1156 * This check is valid for both Intel and AMD
1158 eax = ebx = ecx = edx = 0;
1159 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
1160 if (!(edx & (1 << 8))) {
1161 ERROR("No invariant TSC");
1162 return -ERR_NO_INVARIANT_TSC;
1168 * + 0: Digital temperature sensor is supported if set
1169 * + 6: Package thermal management is supported if set
1171 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
1172 * + 3: The processor supports performance-energy bias preference if set.
1173 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
1175 * This check is valid for both Intel and AMD
1177 eax = ebx = ecx = edx = 0;
1178 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
1179 do_dts = eax & (1 << 0);
1180 do_ptm = eax & (1 << 6);
1181 if (!(ecx & (1 << 0))) {
1183 return -ERR_NO_APERF;
1187 * Enable or disable C states depending on the model and family
1191 /* Atom (partial) */
1194 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
1197 case 0x37: /* BYT */
1199 case 0x4D: /* AVN */
1202 do_core_cstate = (1 << 1) | (1 << 6);
1203 do_pkg_cstate = (1 << 6);
1206 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1207 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1208 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
1209 case 0x2E: /* Nehalem-EX Xeon - Beckton */
1210 do_core_cstate = (1 << 3) | (1 << 6);
1211 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1214 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
1215 case 0x2C: /* Westmere EP - Gulftown */
1216 case 0x2F: /* Westmere-EX Xeon - Eagleton */
1217 do_core_cstate = (1 << 3) | (1 << 6);
1218 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1221 case 0x2A: /* SNB */
1222 case 0x2D: /* SNB Xeon */
1223 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1224 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1227 case 0x3A: /* IVB */
1228 case 0x3E: /* IVB Xeon */
1229 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1230 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1232 /* Haswell Bridge */
1233 case 0x3C: /* HSW */
1234 case 0x3F: /* HSW */
1235 case 0x46: /* HSW */
1236 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1237 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1239 case 0x45: /* HSW */
1240 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1241 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1244 case 0x4F: /* BDW */
1245 case 0x56: /* BDX-DE */
1246 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1247 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
1249 case 0x3D: /* BDW */
1250 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
1251 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
1254 ERROR("Unsupported CPU");
1262 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
1265 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1269 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1273 do_rapl = RAPL_PKG | RAPL_CORES;
1279 ERROR("Unsupported CPU");
1280 return -UNSUPPORTED_CPU;
1284 unsigned long long msr;
1285 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1289 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1291 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1297 static int __attribute__((warn_unused_result))
1298 allocate_cpu_set(cpu_set_t * set, size_t * size) {
1299 set = CPU_ALLOC(topology.max_cpu_id + 1);
1301 ERROR("Unable to allocate CPU state");
1302 return -ERR_CPU_ALLOC;
1304 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1305 CPU_ZERO_S(*size, set);
1310 * Build a local representation of the cpu distribution
1312 static int __attribute__((warn_unused_result))
1317 int max_package_id, max_core_id, max_thread_id;
1318 max_package_id = max_core_id = max_thread_id = 0;
1320 /* Clean topology */
1321 free(topology.cpus);
1322 memset(&topology, 0, sizeof(topology));
1324 /* Can't fail (update_max_cpu_id always returns 0) */
1325 assert(for_all_proc_cpus(update_max_cpu_id));
1327 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1328 if (topology.cpus == NULL) {
1329 ERROR("Unable to allocate memory for cpu topology");
1333 ret = allocate_cpu_set(cpu_present_set, &cpu_present_setsize);
1336 ret = allocate_cpu_set(cpu_affinity_set, &cpu_affinity_setsize);
1339 ret = allocate_cpu_set(cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1343 /* Can't fail (mark_cpu_present always returns 0) */
1344 assert(for_all_proc_cpus(mark_cpu_present));
1348 * find max_core_id, max_package_id
1350 for (i = 0; i <= topology.max_cpu_id; ++i) {
1352 struct cpu_topology *cpu = &topology.cpus[i];
1354 if (cpu_is_not_present(i)) {
1355 WARNING("cpu%d NOT PRESENT", i);
1359 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1363 cpu->package_id = ret;
1364 if (cpu->package_id > max_package_id)
1365 max_package_id = cpu->package_id;
1367 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1372 if (cpu->core_id > max_core_id)
1373 max_core_id = cpu->core_id;
1374 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1377 else if (ret == cpu->core_id)
1378 cpu->first_core_in_package = 1;
1380 ret = get_threads_on_core(i);
1385 if (num_threads > max_thread_id)
1386 max_thread_id = num_threads;
1387 if (num_threads > 1) {
1388 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1391 else if (ret == num_threads)
1392 cpu->first_thread_in_core = 1;
1394 DEBUG("cpu %d pkg %d core %d\n",
1395 i, cpu->package_id, cpu->core_id);
1397 /* Num is max + 1 (need to count 0) */
1398 topology.num_packages = max_package_id + 1;
1399 topology.num_cores = max_core_id + 1;
1400 topology.num_threads = max_thread_id + 1;
1404 free(topology.cpus);
1409 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1412 int total_threads, total_cores;
1414 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1415 *threads = calloc(total_threads, sizeof(struct thread_data));
1416 if (*threads == NULL)
1419 for (i = 0; i < total_threads; ++i)
1420 (*threads)[i].cpu_id = -1;
1422 total_cores = topology.num_cores * topology.num_packages;
1423 *cores = calloc(total_cores, sizeof(struct core_data));
1425 goto err_clean_threads;
1427 for (i = 0; i < total_cores; ++i)
1428 (*cores)[i].core_id = -1;
1430 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1431 if (*packages == NULL)
1432 goto err_clean_cores;
1434 for (i = 0; i < topology.num_packages; i++)
1435 (*packages)[i].package_id = i;
1444 ERROR("calloc counters");
1451 * set cpu_id, core_id, package_id
1452 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1454 * increment topo.num_cores when 1st core in pkg seen
1457 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1458 struct pkg_data *pkg_base, int cpu_id)
1460 struct thread_data *t;
1461 struct core_data *c;
1463 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1465 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1466 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1467 p = GET_PKG(pkg_base, cpu->package_id);
1470 if (cpu->first_thread_in_core)
1471 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1472 if (cpu->first_core_in_package)
1473 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1475 c->core_id = cpu->core_id;
1476 p->package_id = cpu->package_id;
1483 initialize_counters(void)
1488 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1489 if (cpu_is_not_present(cpu_id)) {
1493 ret = init_counter(EVEN_COUNTERS, cpu_id);
1496 ret = init_counter(ODD_COUNTERS, cpu_id);
1499 ret = init_counter(DELTA_COUNTERS, cpu_id);
1506 #define DO_OR_GOTO_ERR(something) \
1508 ret = (something); \
1513 static int setup_all_buffers(void)
1517 DO_OR_GOTO_ERR(topology_probe());
1518 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1519 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1520 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1521 DO_OR_GOTO_ERR(initialize_counters());
1522 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1523 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1533 turbostat_init(void)
1537 DO_OR_GOTO_ERR(check_super_user());
1538 DO_OR_GOTO_ERR(probe_cpu());
1539 DO_OR_GOTO_ERR(check_dev_msr());
1540 DO_OR_GOTO_ERR(setup_all_buffers());
1542 plugin_register_complex_read(NULL, PLUGIN_NAME, turbostat_read, NULL, NULL);
1550 static const char *config_keys[] =
1552 "TCCActivationTemp",
1554 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
1557 turbostat_config(const char *key, const char *value)
1559 long unsigned int tmp_val;
1562 if (strcasecmp("TCCActivationTemp", key) == 0) {
1563 tmp_val = strtoul(value, &end, 0);
1564 if (*end != '\0' || tmp_val > UINT_MAX)
1566 tcc_activation_temp = (unsigned int) tmp_val;
1573 void module_register(void);
1574 void module_register(void)
1576 plugin_register_init(PLUGIN_NAME, turbostat_init);
1577 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);