1 /*
2 * turbostat -- Log CPU frequency and C-state residency
3 * on modern Intel turbo-capable processors for collectd.
4 *
5 * Based on the 'turbostat' tool of the Linux kernel, found at
6 * linux/tools/power/x86/turbostat/turbostat.c:
7 * ----
8 * Copyright (c) 2013 Intel Corporation.
9 * Len Brown <len.brown@intel.com>
10 *
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.
14 *
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
18 * more details.
19 *
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.
23 * ----
24 * Ported to collectd by Vincent Brillault <git@lerya.net>
25 */
27 /*
28 * _GNU_SOURCE is required because of the following functions:
29 * - CPU_ISSET_S
30 * - CPU_ZERO_S
31 * - CPU_SET_S
32 * - CPU_FREE
33 * - CPU_ALLOC
34 * - CPU_ALLOC_SIZE
35 */
36 #define _GNU_SOURCE
38 #include "collectd.h"
40 #include "common.h"
41 #include "plugin.h"
42 #include "utils_time.h"
44 #include <asm/msr-index.h>
45 #include <cpuid.h>
46 #ifdef HAVE_SYS_CAPABILITY_H
47 #include <sys/capability.h>
48 #endif /* HAVE_SYS_CAPABILITY_H */
50 #define PLUGIN_NAME "turbostat"
52 /*
53 * This tool uses the Model-Specific Registers (MSRs) present on Intel processors.
54 * The general description each of these registers, depending on the architecture,
55 * can be found in the IntelĀ® 64 and IA-32 Architectures Software Developer Manual,
56 * Volume 3 Chapter 35.
57 */
59 /*
60 * If set, aperf_mperf_unstable disables a/mperf based stats.
61 * This includes: C0 & C1 states, frequency
62 *
63 * This value is automatically set if mperf or aperf go backward
64 */
65 static _Bool aperf_mperf_unstable;
67 /*
68 * If set, use kernel logical core numbering for all "per core" metrics.
69 */
70 static _Bool config_lcn;
72 /*
73 * Bitmask of the list of core C states supported by the processor.
74 * Currently supported C-states (by this plugin): 3, 6, 7
75 */
76 static unsigned int do_core_cstate;
77 static unsigned int config_core_cstate;
78 static _Bool apply_config_core_cstate;
80 /*
81 * Bitmask of the list of pacages C states supported by the processor.
82 * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
83 */
84 static unsigned int do_pkg_cstate;
85 static unsigned int config_pkg_cstate;
86 static _Bool apply_config_pkg_cstate;
88 /*
89 * Boolean indicating if the processor supports 'I/O System-Management Interrupt counter'
90 */
91 static _Bool do_smi;
92 static _Bool config_smi;
93 static _Bool apply_config_smi;
95 /*
96 * Boolean indicating if the processor supports 'Digital temperature sensor'
97 * This feature enables the monitoring of the temperature of each core
98 *
99 * This feature has two limitations:
100 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
101 * - Temperatures above the tcc_activation_temp are not recorded
102 */
103 static _Bool do_dts;
104 static _Bool config_dts;
105 static _Bool apply_config_dts;
107 /*
108 * Boolean indicating if the processor supports 'Package thermal management'
109 * This feature allows the monitoring of the temperature of each package
110 *
111 * This feature has two limitations:
112 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
113 * - Temperatures above the tcc_activation_temp are not recorded
114 */
115 static _Bool do_ptm;
116 static _Bool config_ptm;
117 static _Bool apply_config_ptm;
119 /*
120 * Thermal Control Circuit Activation Temperature as configured by the user.
121 * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
122 * and should only be used if the automated detection fails.
123 */
124 static unsigned int tcc_activation_temp;
126 static unsigned int do_rapl;
127 static unsigned int config_rapl;
128 static _Bool apply_config_rapl;
129 static double rapl_energy_units;
131 #define RAPL_PKG (1 << 0)
132 /* 0x610 MSR_PKG_POWER_LIMIT */
133 /* 0x611 MSR_PKG_ENERGY_STATUS */
134 #define RAPL_DRAM (1 << 1)
135 /* 0x618 MSR_DRAM_POWER_LIMIT */
136 /* 0x619 MSR_DRAM_ENERGY_STATUS */
137 /* 0x61c MSR_DRAM_POWER_INFO */
138 #define RAPL_CORES (1 << 2)
139 /* 0x638 MSR_PP0_POWER_LIMIT */
140 /* 0x639 MSR_PP0_ENERGY_STATUS */
142 #define RAPL_GFX (1 << 3)
143 /* 0x640 MSR_PP1_POWER_LIMIT */
144 /* 0x641 MSR_PP1_ENERGY_STATUS */
145 /* 0x642 MSR_PP1_POLICY */
146 #define TJMAX_DEFAULT 100
148 static cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
149 static size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
151 static struct thread_data {
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;
157 unsigned int cpu_id;
158 unsigned int flags;
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;
163 static struct core_data {
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;
171 static struct pkg_data {
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 uint32_t energy_pkg; /* MSR_PKG_ENERGY_STATUS */
181 uint32_t energy_dram; /* MSR_DRAM_ENERGY_STATUS */
182 uint32_t energy_cores; /* MSR_PP0_ENERGY_STATUS */
183 uint32_t energy_gfx; /* MSR_PP1_ENERGY_STATUS */
184 unsigned int tcc_activation_temp;
185 unsigned int pkg_temp_c;
186 } *package_delta, *package_even, *package_odd;
188 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
189 #define ODD_COUNTERS thread_odd, core_odd, package_odd
190 #define EVEN_COUNTERS thread_even, core_even, package_even
191 static _Bool is_even = 1;
193 static _Bool allocated = 0;
194 static _Bool initialized = 0;
196 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
197 (thread_base + \
198 (pkg_no) * topology.num_cores * topology.num_threads + \
199 (core_no) * topology.num_threads + \
200 (thread_no))
201 #define GET_CORE(core_base, core_no, pkg_no) \
202 (core_base + \
203 (pkg_no) * topology.num_cores + \
204 (core_no))
205 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
207 struct cpu_topology {
208 unsigned int package_id;
209 unsigned int core_id;
210 _Bool first_core_in_package;
211 _Bool first_thread_in_core;
212 };
214 static struct topology {
215 unsigned int max_cpu_id;
216 unsigned int num_packages;
217 unsigned int num_cores;
218 unsigned int num_threads;
219 struct cpu_topology *cpus;
220 } topology;
222 static cdtime_t time_even, time_odd, time_delta;
224 static const char *config_keys[] =
225 {
226 "CoreCstates",
227 "PackageCstates",
228 "SystemManagementInterrupt",
229 "DigitalTemperatureSensor",
230 "PackageThermalManagement",
231 "TCCActivationTemp",
232 "RunningAveragePowerLimit",
233 "LogicalCoreNames",
234 };
235 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
237 /*****************************
238 * MSR Manipulation helpers *
239 *****************************/
241 /*
242 * Open a MSR device for reading
243 * Can change the scheduling affinity of the current process if multiple_read is 1
244 */
245 static int __attribute__((warn_unused_result))
246 open_msr(unsigned int cpu, _Bool multiple_read)
247 {
248 char pathname[32];
249 int fd;
251 /*
252 * If we need to do multiple read, let's migrate to the CPU
253 * Otherwise, we would lose time calling functions on another CPU
254 *
255 * If we are not yet initialized (cpu_affinity_setsize = 0),
256 * we need to skip this optimisation.
257 */
258 if (multiple_read && cpu_affinity_setsize) {
259 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
260 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
261 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
262 ERROR("turbostat plugin: Could not migrate to CPU %d", cpu);
263 return -1;
264 }
265 }
267 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
268 fd = open(pathname, O_RDONLY);
269 if (fd < 0) {
270 ERROR("turbostat plugin: failed to open %s", pathname);
271 return -1;
272 }
273 return fd;
274 }
276 /*
277 * Read a single MSR from an open file descriptor
278 */
279 static int __attribute__((warn_unused_result))
280 read_msr(int fd, off_t offset, unsigned long long *msr)
281 {
282 ssize_t retval;
284 retval = pread(fd, msr, sizeof *msr, offset);
286 if (retval != sizeof *msr) {
287 ERROR("turbostat plugin: MSR offset 0x%llx read failed",
288 (unsigned long long)offset);
289 return -1;
290 }
291 return 0;
292 }
294 /*
295 * Open a MSR device for reading, read the value asked for and close it.
296 * This call will not affect the scheduling affinity of this thread.
297 */
298 static ssize_t __attribute__((warn_unused_result))
299 get_msr(unsigned int cpu, off_t offset, unsigned long long *msr)
300 {
301 ssize_t retval;
302 int fd;
304 fd = open_msr(cpu, 0);
305 if (fd < 0)
306 return fd;
307 retval = read_msr(fd, offset, msr);
308 close(fd);
309 return retval;
310 }
313 /********************************
314 * Raw data acquisition (1 CPU) *
315 ********************************/
317 /*
318 * Read every data avalaible for a single CPU
319 *
320 * Core data is shared for all threads in one core: extracted only for the first thread
321 * Package data is shared for all core in one package: extracted only for the first thread of the first core
322 *
323 * Side effect: migrates to the targeted CPU
324 */
325 static int __attribute__((warn_unused_result))
326 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
327 {
328 unsigned int cpu = t->cpu_id;
329 unsigned long long msr;
330 int msr_fd;
331 int retval = 0;
333 msr_fd = open_msr(cpu, 1);
334 if (msr_fd < 0)
335 return msr_fd;
337 #define READ_MSR(msr, dst) \
338 do { \
339 if (read_msr(msr_fd, msr, dst)) { \
340 ERROR("turbostat plugin: Unable to read " #msr); \
341 retval = -1; \
342 goto out; \
343 } \
344 } while (0)
346 READ_MSR(MSR_IA32_TSC, &t->tsc);
348 READ_MSR(MSR_IA32_APERF, &t->aperf);
349 READ_MSR(MSR_IA32_MPERF, &t->mperf);
351 if (do_smi) {
352 READ_MSR(MSR_SMI_COUNT, &msr);
353 t->smi_count = msr & 0xFFFFFFFF;
354 }
356 /* collect core counters only for 1st thread in core */
357 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
358 retval = 0;
359 goto out;
360 }
362 if (do_core_cstate & (1 << 3))
363 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
364 if (do_core_cstate & (1 << 6))
365 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
366 if (do_core_cstate & (1 << 7))
367 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
369 if (do_dts) {
370 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
371 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
372 }
374 /* collect package counters only for 1st core in package */
375 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
376 retval = 0;
377 goto out;
378 }
380 if (do_pkg_cstate & (1 << 2))
381 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
382 if (do_pkg_cstate & (1 << 3))
383 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
384 if (do_pkg_cstate & (1 << 6))
385 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
386 if (do_pkg_cstate & (1 << 7))
387 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
388 if (do_pkg_cstate & (1 << 8))
389 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
390 if (do_pkg_cstate & (1 << 9))
391 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
392 if (do_pkg_cstate & (1 << 10))
393 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
395 if (do_rapl & RAPL_PKG) {
396 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
397 p->energy_pkg = msr & 0xFFFFFFFF;
398 }
399 if (do_rapl & RAPL_CORES) {
400 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
401 p->energy_cores = msr & 0xFFFFFFFF;
402 }
403 if (do_rapl & RAPL_DRAM) {
404 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
405 p->energy_dram = msr & 0xFFFFFFFF;
406 }
407 if (do_rapl & RAPL_GFX) {
408 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
409 p->energy_gfx = msr & 0xFFFFFFFF;
410 }
411 if (do_ptm) {
412 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
413 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
414 }
416 out:
417 close(msr_fd);
418 return retval;
419 }
422 /**********************************
423 * Evaluating the changes (1 CPU) *
424 **********************************/
426 /*
427 * Extract the evolution old->new in delta at a package level
428 * (some are not new-delta, e.g. temperature)
429 */
430 static inline void
431 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
432 {
433 delta->pc2 = new->pc2 - old->pc2;
434 delta->pc3 = new->pc3 - old->pc3;
435 delta->pc6 = new->pc6 - old->pc6;
436 delta->pc7 = new->pc7 - old->pc7;
437 delta->pc8 = new->pc8 - old->pc8;
438 delta->pc9 = new->pc9 - old->pc9;
439 delta->pc10 = new->pc10 - old->pc10;
440 delta->pkg_temp_c = new->pkg_temp_c;
442 delta->energy_pkg = new->energy_pkg - old->energy_pkg;
443 delta->energy_cores = new->energy_cores - old->energy_cores;
444 delta->energy_gfx = new->energy_gfx - old->energy_gfx;
445 delta->energy_dram = new->energy_dram - old->energy_dram;
446 }
448 /*
449 * Extract the evolution old->new in delta at a core level
450 * (some are not new-delta, e.g. temperature)
451 */
452 static inline void
453 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
454 {
455 delta->c3 = new->c3 - old->c3;
456 delta->c6 = new->c6 - old->c6;
457 delta->c7 = new->c7 - old->c7;
458 delta->core_temp_c = new->core_temp_c;
459 }
461 /*
462 * Extract the evolution old->new in delta at a package level
463 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
464 */
465 static inline int __attribute__((warn_unused_result))
466 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
467 const struct core_data *cdelta)
468 {
469 delta->tsc = new->tsc - old->tsc;
471 /* check for TSC < 1 Mcycles over interval */
472 if (delta->tsc < (1000 * 1000)) {
473 WARNING("turbostat plugin: Insanely slow TSC rate, TSC stops "
474 "in idle? You can disable all c-states by booting with"
475 " 'idle=poll' or just the deep ones with"
476 " 'processor.max_cstate=1'");
477 return -1;
478 }
480 delta->c1 = new->c1 - old->c1;
482 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
483 delta->aperf = new->aperf - old->aperf;
484 delta->mperf = new->mperf - old->mperf;
485 } else {
486 if (!aperf_mperf_unstable) {
487 WARNING("turbostat plugin: APERF or MPERF went "
488 "backwards. Frequency results do not cover "
489 "the entire interval. Fix this by running "
490 "Linux-2.6.30 or later.");
492 aperf_mperf_unstable = 1;
493 }
494 }
496 /*
497 * As counter collection is not atomic,
498 * it is possible for mperf's non-halted cycles + idle states
499 * to exceed TSC's all cycles: show c1 = 0% in that case.
500 */
501 if ((delta->mperf + cdelta->c3 + cdelta->c6 + cdelta->c7) > delta->tsc)
502 delta->c1 = 0;
503 else {
504 /* normal case, derive c1 */
505 delta->c1 = delta->tsc - delta->mperf - cdelta->c3
506 - cdelta->c6 - cdelta->c7;
507 }
509 if (delta->mperf == 0) {
510 WARNING("turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
511 delta->mperf = 1; /* divide by 0 protection */
512 }
514 if (do_smi)
515 delta->smi_count = new->smi_count - old->smi_count;
517 return 0;
518 }
520 /**********************************
521 * Submitting the results (1 CPU) *
522 **********************************/
524 /*
525 * Submit one gauge value
526 */
527 static void
528 turbostat_submit (const char *plugin_instance,
529 const char *type, const char *type_instance,
530 gauge_t value)
531 {
532 value_list_t vl = VALUE_LIST_INIT;
534 vl.values = &(value_t) { .gauge = value };
535 vl.values_len = 1;
536 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
537 if (plugin_instance != NULL)
538 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
539 sstrncpy (vl.type, type, sizeof (vl.type));
540 if (type_instance != NULL)
541 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
543 plugin_dispatch_values (&vl);
544 }
546 /*
547 * Submit every data for a single CPU
548 *
549 * Core data is shared for all threads in one core: submitted only for the first thread
550 * Package data is shared for all core in one package: submitted only for the first thread of the first core
551 */
552 static int
553 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
554 {
555 char name[DATA_MAX_NAME_LEN];
556 double interval_float;
558 interval_float = CDTIME_T_TO_DOUBLE(time_delta);
560 DEBUG("turbostat plugin: submit stats for cpu: %d, core: %d, pkg: %d",
561 t->cpu_id, c->core_id, p->package_id);
563 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
565 if (!aperf_mperf_unstable)
566 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
567 if (!aperf_mperf_unstable)
568 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
570 turbostat_submit(name, "frequency", "average", 1.0 / 1000000 * t->aperf / interval_float);
572 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
573 turbostat_submit(name, "frequency", "busy", 1.0 * t->tsc / 1000000 * t->aperf / t->mperf / interval_float);
575 /* Sanity check (should stay stable) */
576 turbostat_submit(name, "gauge", "TSC", 1.0 * t->tsc / 1000000 / interval_float);
578 /* SMI */
579 if (do_smi)
580 turbostat_submit(name, "count", NULL, t->smi_count);
582 /* submit per-core data only for 1st thread in core */
583 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
584 goto done;
586 /* If not using logical core numbering, set core id */
587 if (!config_lcn) {
588 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
589 }
591 if (do_core_cstate & (1 << 3))
592 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
593 if (do_core_cstate & (1 << 6))
594 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
595 if (do_core_cstate & (1 << 7))
596 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
598 if (do_dts)
599 turbostat_submit(name, "temperature", NULL, c->core_temp_c);
601 /* submit per-package data only for 1st core in package */
602 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
603 goto done;
605 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
607 if (do_ptm)
608 turbostat_submit(name, "temperature", NULL, p->pkg_temp_c);
610 if (do_pkg_cstate & (1 << 2))
611 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
612 if (do_pkg_cstate & (1 << 3))
613 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
614 if (do_pkg_cstate & (1 << 6))
615 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
616 if (do_pkg_cstate & (1 << 7))
617 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
618 if (do_pkg_cstate & (1 << 8))
619 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
620 if (do_pkg_cstate & (1 << 9))
621 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
622 if (do_pkg_cstate & (1 << 10))
623 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
625 if (do_rapl) {
626 if (do_rapl & RAPL_PKG)
627 turbostat_submit(name, "power", "pkg", p->energy_pkg * rapl_energy_units / interval_float);
628 if (do_rapl & RAPL_CORES)
629 turbostat_submit(name, "power", "cores", p->energy_cores * rapl_energy_units / interval_float);
630 if (do_rapl & RAPL_GFX)
631 turbostat_submit(name, "power", "GFX", p->energy_gfx * rapl_energy_units / interval_float);
632 if (do_rapl & RAPL_DRAM)
633 turbostat_submit(name, "power", "DRAM", p->energy_dram * rapl_energy_units / interval_float);
634 }
635 done:
636 return 0;
637 }
640 /**********************************
641 * Looping function over all CPUs *
642 **********************************/
644 /*
645 * Check if a given cpu id is in our compiled list of existing CPUs
646 */
647 static int
648 cpu_is_not_present(unsigned int cpu)
649 {
650 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
651 }
653 /*
654 * Loop on all CPUs in topological order
655 *
656 * Skip non-present cpus
657 * Return the error code at the first error or 0
658 */
659 static int __attribute__((warn_unused_result))
660 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
661 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
662 {
663 int retval;
665 for (unsigned int pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
666 for (unsigned int core_no = 0; core_no < topology.num_cores; ++core_no) {
667 for (unsigned int thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
668 struct thread_data *t;
669 struct core_data *c;
670 struct pkg_data *p;
672 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
674 if (cpu_is_not_present(t->cpu_id))
675 continue;
677 c = GET_CORE(core_base, core_no, pkg_no);
678 p = GET_PKG(pkg_base, pkg_no);
680 retval = func(t, c, p);
681 if (retval)
682 return retval;
683 }
684 }
685 }
686 return 0;
687 }
689 /*
690 * Dedicated loop: Extract every data evolution for all CPU
691 *
692 * Skip non-present cpus
693 * Return the error code at the first error or 0
694 *
695 * Core data is shared for all threads in one core: extracted only for the first thread
696 * Package data is shared for all core in one package: extracted only for the first thread of the first core
697 */
698 static int __attribute__((warn_unused_result))
699 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,
700 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
701 {
702 int retval;
704 for (unsigned int pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
705 for (unsigned int core_no = 0; core_no < topology.num_cores; ++core_no) {
706 for (unsigned int thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
707 struct thread_data *t_delta;
708 const struct thread_data *t_old, *t_new;
709 struct core_data *c_delta;
711 /* Get correct pointers for threads */
712 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
713 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
714 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
716 /* Skip threads that disappeared */
717 if (cpu_is_not_present(t_delta->cpu_id))
718 continue;
720 /* c_delta is always required for delta_thread */
721 c_delta = GET_CORE(core_delta, core_no, pkg_no);
723 /* calculate core delta only for 1st thread in core */
724 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
725 const struct core_data *c_old, *c_new;
727 c_new = GET_CORE(core_new_base, core_no, pkg_no);
728 c_old = GET_CORE(core_old_base, core_no, pkg_no);
730 delta_core(c_delta, c_new, c_old);
731 }
733 /* Always calculate thread delta */
734 retval = delta_thread(t_delta, t_new, t_old, c_delta);
735 if (retval)
736 return retval;
738 /* calculate package delta only for 1st core in package */
739 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
740 struct pkg_data *p_delta;
741 const struct pkg_data *p_old, *p_new;
743 p_delta = GET_PKG(package_delta, pkg_no);
744 p_new = GET_PKG(pkg_new_base, pkg_no);
745 p_old = GET_PKG(pkg_old_base, pkg_no);
747 delta_package(p_delta, p_new, p_old);
748 }
749 }
750 }
751 }
752 return 0;
753 }
756 /***************
757 * CPU Probing *
758 ***************/
760 /*
761 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
762 * the Thermal Control Circuit (TCC) activates.
763 * This is usually equal to tjMax.
764 *
765 * Older processors do not have this MSR, so there we guess,
766 * but also allow conficuration over-ride with "TCCActivationTemp".
767 *
768 * Several MSR temperature values are in units of degrees-C
769 * below this value, including the Digital Thermal Sensor (DTS),
770 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
771 */
772 static int __attribute__((warn_unused_result))
773 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
774 {
775 unsigned long long msr;
776 unsigned int target_c_local;
778 /* tcc_activation_temp is used only for dts or ptm */
779 if (!(do_dts || do_ptm))
780 return 0;
782 /* this is a per-package concept */
783 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
784 return 0;
786 if (tcc_activation_temp != 0) {
787 p->tcc_activation_temp = tcc_activation_temp;
788 return 0;
789 }
791 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
792 goto guess;
794 target_c_local = (msr >> 16) & 0xFF;
796 if (!target_c_local)
797 goto guess;
799 p->tcc_activation_temp = target_c_local;
801 return 0;
803 guess:
804 p->tcc_activation_temp = TJMAX_DEFAULT;
805 WARNING("turbostat plugin: cpu%d: Guessing tjMax %d C,"
806 " Please use TCCActivationTemp to specify it.",
807 t->cpu_id, p->tcc_activation_temp);
809 return 0;
810 }
812 /*
813 * Identify the functionality of the CPU
814 */
815 static int __attribute__((warn_unused_result))
816 probe_cpu(void)
817 {
818 unsigned int eax, ebx, ecx, edx, max_level;
819 unsigned int fms, family, model;
821 /* CPUID(0):
822 * - EAX: Maximum Input Value for Basic CPUID Information
823 * - EBX: "Genu" (0x756e6547)
824 * - EDX: "ineI" (0x49656e69)
825 * - ECX: "ntel" (0x6c65746e)
826 */
827 max_level = ebx = ecx = edx = 0;
828 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
829 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
830 ERROR("turbostat plugin: Unsupported CPU (not Intel)");
831 return -1;
832 }
834 /* CPUID(1):
835 * - EAX: Version Information: Type, Family, Model, and Stepping ID
836 * + 4-7: Model ID
837 * + 8-11: Family ID
838 * + 12-13: Processor type
839 * + 16-19: Extended Model ID
840 * + 20-27: Extended Family ID
841 * - EDX: Feature Information:
842 * + 5: Support for MSR read/write operations
843 */
844 fms = ebx = ecx = edx = 0;
845 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
846 family = (fms >> 8) & 0xf;
847 model = (fms >> 4) & 0xf;
848 if (family == 0xf)
849 family += (fms >> 20) & 0xf;
850 if (family == 6 || family == 0xf)
851 model += ((fms >> 16) & 0xf) << 4;
852 if (!(edx & (1 << 5))) {
853 ERROR("turbostat plugin: Unsupported CPU (no MSR support)");
854 return -1;
855 }
857 /*
858 * CPUID(6):
859 * - EAX:
860 * + 0: Digital temperature sensor is supported if set
861 * + 6: Package thermal management is supported if set
862 * - ECX:
863 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
864 * + 3: The processor supports performance-energy bias preference if set.
865 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
866 *
867 * This check is valid for both Intel and AMD
868 */
869 eax = ebx = ecx = edx = 0;
870 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
871 do_dts = eax & (1 << 0);
872 do_ptm = eax & (1 << 6);
873 if (!(ecx & (1 << 0))) {
874 ERROR("turbostat plugin: Unsupported CPU (No APERF)");
875 return -1;
876 }
878 /*
879 * Enable or disable C states depending on the model and family
880 */
881 if (family == 6) {
882 switch (model) {
883 /* Atom (partial) */
884 case 0x27:
885 do_smi = 0;
886 do_core_cstate = 0;
887 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
888 break;
889 /* Silvermont */
890 case 0x37: /* BYT */
891 case 0x4D: /* AVN */
892 do_smi = 1;
893 do_core_cstate = (1 << 1) | (1 << 6);
894 do_pkg_cstate = (1 << 6);
895 break;
896 /* Nehalem */
897 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
898 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
899 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
900 case 0x2E: /* Nehalem-EX Xeon - Beckton */
901 do_smi = 1;
902 do_core_cstate = (1 << 3) | (1 << 6);
903 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
904 break;
905 /* Westmere */
906 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
907 case 0x2C: /* Westmere EP - Gulftown */
908 case 0x2F: /* Westmere-EX Xeon - Eagleton */
909 do_smi = 1;
910 do_core_cstate = (1 << 3) | (1 << 6);
911 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
912 break;
913 /* Sandy Bridge */
914 case 0x2A: /* SNB */
915 case 0x2D: /* SNB Xeon */
916 do_smi = 1;
917 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
918 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
919 break;
920 /* Ivy Bridge */
921 case 0x3A: /* IVB */
922 case 0x3E: /* IVB Xeon */
923 do_smi = 1;
924 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
925 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
926 break;
927 /* Haswell Bridge */
928 case 0x3C: /* HSW */
929 case 0x3F: /* HSW */
930 case 0x46: /* HSW */
931 do_smi = 1;
932 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
933 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
934 break;
935 case 0x45: /* HSW */
936 do_smi = 1;
937 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
938 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
939 break;
940 /* Broadwel */
941 case 0x4F: /* BDW */
942 case 0x56: /* BDX-DE */
943 do_smi = 1;
944 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
945 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
946 break;
947 case 0x3D: /* BDW */
948 do_smi = 1;
949 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
950 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
951 break;
952 default:
953 do_smi = 0;
954 do_core_cstate = 0;
955 do_pkg_cstate = 0;
956 break;
957 }
958 switch (model) {
959 case 0x2A: /* SNB */
960 case 0x3A: /* IVB */
961 case 0x3C: /* HSW */
962 case 0x45: /* HSW */
963 case 0x46: /* HSW */
964 case 0x3D: /* BDW */
965 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_GFX;
966 break;
967 case 0x3F: /* HSX */
968 case 0x4F: /* BDX */
969 case 0x56: /* BDX-DE */
970 do_rapl = RAPL_PKG | RAPL_DRAM ;
971 break;
972 case 0x2D: /* SNB Xeon */
973 case 0x3E: /* IVB Xeon */
974 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_DRAM;
975 break;
976 case 0x37: /* BYT */
977 case 0x4D: /* AVN */
978 do_rapl = RAPL_PKG | RAPL_CORES;
979 break;
980 default:
981 do_rapl = 0;
982 }
983 } else {
984 ERROR("turbostat plugin: Unsupported CPU (family: %#x, "
985 "model: %#x)", family, model);
986 return -1;
987 }
989 /* Override detected values with configuration */
990 if (apply_config_core_cstate)
991 do_core_cstate = config_core_cstate;
992 if (apply_config_pkg_cstate)
993 do_pkg_cstate = config_pkg_cstate;
994 if (apply_config_smi)
995 do_smi = config_smi;
996 if (apply_config_dts)
997 do_dts = config_dts;
998 if (apply_config_ptm)
999 do_ptm = config_ptm;
1000 if (apply_config_rapl)
1001 do_rapl = config_rapl;
1003 if (do_rapl) {
1004 unsigned long long msr;
1005 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1006 return 0;
1008 if (model == 0x37)
1009 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1010 else
1011 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1012 }
1014 return 0;
1015 }
1018 /********************
1019 * Topology Probing *
1020 ********************/
1022 /*
1023 * Read a single int from a file.
1024 */
1025 static int __attribute__ ((format(printf,1,2)))
1026 parse_int_file(const char *fmt, ...)
1027 {
1028 va_list args;
1029 char path[PATH_MAX];
1030 int len;
1032 va_start(args, fmt);
1033 len = vsnprintf(path, sizeof(path), fmt, args);
1034 va_end(args);
1035 if (len < 0 || len >= PATH_MAX) {
1036 ERROR("turbostat plugin: path truncated: '%s'", path);
1037 return -1;
1038 }
1040 value_t v;
1041 if (parse_value_file (path, &v, DS_TYPE_DERIVE) != 0) {
1042 ERROR ("turbostat plugin: Parsing \"%s\" failed.", path);
1043 return -1;
1044 }
1046 return (int) v.derive;
1047 }
1049 static int
1050 get_threads_on_core(unsigned int cpu)
1051 {
1052 char path[80];
1053 FILE *filep;
1054 int sib1, sib2;
1055 int matches;
1056 char character;
1058 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1059 filep = fopen(path, "r");
1060 if (!filep) {
1061 ERROR("turbostat plugin: Failed to open '%s'", path);
1062 return -1;
1063 }
1064 /*
1065 * file format:
1066 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1067 * otherwinse 1 sibling (self).
1068 */
1069 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1071 fclose(filep);
1073 if (matches == 3)
1074 return 2;
1075 else
1076 return 1;
1077 }
1079 /*
1080 * run func(cpu) on every cpu in /proc/stat
1081 * return max_cpu number
1082 */
1083 static int __attribute__((warn_unused_result))
1084 for_all_proc_cpus(int (func)(unsigned int))
1085 {
1086 FILE *fp;
1087 unsigned int cpu_num;
1088 int retval;
1090 fp = fopen("/proc/stat", "r");
1091 if (!fp) {
1092 ERROR("turbostat plugin: Failed to open /proc/stat");
1093 return -1;
1094 }
1096 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1097 if (retval != 0) {
1098 ERROR("turbostat plugin: Failed to parse /proc/stat");
1099 fclose(fp);
1100 return -1;
1101 }
1103 while (1) {
1104 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1105 if (retval != 1)
1106 break;
1108 retval = func(cpu_num);
1109 if (retval) {
1110 fclose(fp);
1111 return(retval);
1112 }
1113 }
1114 fclose(fp);
1115 return 0;
1116 }
1118 /*
1119 * Update the stored topology.max_cpu_id
1120 */
1121 static int
1122 update_max_cpu_id(unsigned int cpu)
1123 {
1124 if (topology.max_cpu_id < cpu)
1125 topology.max_cpu_id = cpu;
1126 return 0;
1127 }
1129 static int
1130 mark_cpu_present(unsigned int cpu)
1131 {
1132 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1133 return 0;
1134 }
1136 static int __attribute__((warn_unused_result))
1137 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1138 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1139 if (*set == NULL) {
1140 ERROR("turbostat plugin: Unable to allocate CPU state");
1141 return -1;
1142 }
1143 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1144 CPU_ZERO_S(*size, *set);
1145 return 0;
1146 }
1148 /*
1149 * Build a local representation of the cpu distribution
1150 */
1151 static int __attribute__((warn_unused_result))
1152 topology_probe(void)
1153 {
1154 int ret;
1155 unsigned int max_package_id, max_core_id, max_threads;
1156 max_package_id = max_core_id = max_threads = 0;
1158 /* Clean topology */
1159 free(topology.cpus);
1160 memset(&topology, 0, sizeof(topology));
1162 ret = for_all_proc_cpus(update_max_cpu_id);
1163 if (ret != 0)
1164 goto err;
1166 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1167 if (topology.cpus == NULL) {
1168 ERROR("turbostat plugin: Unable to allocate memory for CPU topology");
1169 return -1;
1170 }
1172 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1173 if (ret != 0)
1174 goto err;
1175 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1176 if (ret != 0)
1177 goto err;
1178 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1179 if (ret != 0)
1180 goto err;
1182 ret = for_all_proc_cpus(mark_cpu_present);
1183 if (ret != 0)
1184 goto err;
1186 /*
1187 * For online cpus
1188 * find max_core_id, max_package_id
1189 */
1190 for (unsigned int i = 0; i <= topology.max_cpu_id; ++i) {
1191 unsigned int num_threads;
1192 struct cpu_topology *cpu = &topology.cpus[i];
1194 if (cpu_is_not_present(i)) {
1195 WARNING("turbostat plugin: cpu%d NOT PRESENT", i);
1196 continue;
1197 }
1199 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1200 if (ret < 0)
1201 goto err;
1202 else
1203 cpu->package_id = (unsigned int) ret;
1204 if (cpu->package_id > max_package_id)
1205 max_package_id = cpu->package_id;
1207 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1208 if (ret < 0)
1209 goto err;
1210 else
1211 cpu->core_id = (unsigned int) ret;
1212 if (cpu->core_id > max_core_id)
1213 max_core_id = cpu->core_id;
1214 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1215 if (ret < 0)
1216 goto err;
1217 else if ((unsigned int) ret == i)
1218 cpu->first_core_in_package = 1;
1220 ret = get_threads_on_core(i);
1221 if (ret < 0)
1222 goto err;
1223 else
1224 num_threads = (unsigned int) ret;
1225 if (num_threads > max_threads)
1226 max_threads = num_threads;
1227 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1228 if (ret < 0)
1229 goto err;
1230 else if ((unsigned int) ret == i)
1231 cpu->first_thread_in_core = 1;
1233 DEBUG("turbostat plugin: cpu %d pkg %d core %d\n",
1234 i, cpu->package_id, cpu->core_id);
1235 }
1236 /* Num is max + 1 (need to count 0) */
1237 topology.num_packages = max_package_id + 1;
1238 topology.num_cores = max_core_id + 1;
1239 topology.num_threads = max_threads;
1241 return 0;
1242 err:
1243 free(topology.cpus);
1244 return ret;
1245 }
1248 /************************
1249 * Main alloc/init/free *
1250 ************************/
1252 static int
1253 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1254 {
1255 unsigned int total_threads, total_cores;
1257 if ((topology.num_threads == 0)
1258 || (topology.num_cores == 0)
1259 || (topology.num_packages == 0))
1260 {
1261 ERROR ("turbostat plugin: Invalid topology: %u threads, %u cores, %u packages",
1262 topology.num_threads, topology.num_cores, topology.num_packages);
1263 return -1;
1264 }
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)
1269 {
1270 ERROR ("turbostat plugin: calloc failed");
1271 return -1;
1272 }
1274 for (unsigned int i = 0; i < total_threads; ++i)
1275 (*threads)[i].cpu_id = topology.max_cpu_id + 1;
1277 total_cores = topology.num_cores * topology.num_packages;
1278 *cores = calloc(total_cores, sizeof(struct core_data));
1279 if (*cores == NULL)
1280 {
1281 ERROR ("turbostat plugin: calloc failed");
1282 sfree (threads);
1283 return -1;
1284 }
1286 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1287 if (*packages == NULL)
1288 {
1289 ERROR ("turbostat plugin: calloc failed");
1290 sfree (cores);
1291 sfree (threads);
1292 return -1;
1293 }
1295 return 0;
1296 }
1298 static void
1299 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1300 struct pkg_data *pkg_base, unsigned int cpu_id)
1301 {
1302 struct thread_data *t;
1303 struct core_data *c;
1304 struct pkg_data *p;
1305 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1307 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1308 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1309 p = GET_PKG(pkg_base, cpu->package_id);
1311 t->cpu_id = cpu_id;
1312 if (cpu->first_thread_in_core)
1313 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1314 if (cpu->first_core_in_package)
1315 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1317 c->core_id = cpu->core_id;
1318 p->package_id = cpu->package_id;
1319 }
1321 static void
1322 initialize_counters(void)
1323 {
1324 for (unsigned int cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1325 if (cpu_is_not_present(cpu_id))
1326 continue;
1327 init_counter(EVEN_COUNTERS, cpu_id);
1328 init_counter(ODD_COUNTERS, cpu_id);
1329 init_counter(DELTA_COUNTERS, cpu_id);
1330 }
1331 }
1335 static void
1336 free_all_buffers(void)
1337 {
1338 allocated = 0;
1339 initialized = 0;
1341 CPU_FREE(cpu_present_set);
1342 cpu_present_set = NULL;
1343 cpu_present_setsize = 0;
1345 CPU_FREE(cpu_affinity_set);
1346 cpu_affinity_set = NULL;
1347 cpu_affinity_setsize = 0;
1349 CPU_FREE(cpu_saved_affinity_set);
1350 cpu_saved_affinity_set = NULL;
1351 cpu_saved_affinity_setsize = 0;
1353 free(thread_even);
1354 free(core_even);
1355 free(package_even);
1357 thread_even = NULL;
1358 core_even = NULL;
1359 package_even = NULL;
1361 free(thread_odd);
1362 free(core_odd);
1363 free(package_odd);
1365 thread_odd = NULL;
1366 core_odd = NULL;
1367 package_odd = NULL;
1369 free(thread_delta);
1370 free(core_delta);
1371 free(package_delta);
1373 thread_delta = NULL;
1374 core_delta = NULL;
1375 package_delta = NULL;
1376 }
1379 /**********************
1380 * Collectd functions *
1381 **********************/
1383 #define DO_OR_GOTO_ERR(something) \
1384 do { \
1385 ret = (something); \
1386 if (ret < 0) \
1387 goto err; \
1388 } while (0)
1390 static int setup_all_buffers(void)
1391 {
1392 int ret;
1394 DO_OR_GOTO_ERR(topology_probe());
1395 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1396 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1397 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1398 initialize_counters();
1399 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1400 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1402 allocated = 1;
1403 return 0;
1404 err:
1405 free_all_buffers();
1406 return ret;
1407 }
1409 static int
1410 turbostat_read(void)
1411 {
1412 int ret;
1414 if (!allocated) {
1415 if ((ret = setup_all_buffers()) < 0)
1416 return ret;
1417 }
1419 if (for_all_proc_cpus(cpu_is_not_present)) {
1420 free_all_buffers();
1421 if ((ret = setup_all_buffers()) < 0)
1422 return ret;
1423 if (for_all_proc_cpus(cpu_is_not_present)) {
1424 ERROR("turbostat plugin: CPU appeared just after "
1425 "initialization");
1426 return -1;
1427 }
1428 }
1430 /* Saving the scheduling affinity, as it will be modified by get_counters */
1431 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1432 ERROR("turbostat plugin: Unable to save the CPU affinity");
1433 return -1;
1434 }
1436 if (!initialized) {
1437 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1438 goto out;
1439 time_even = cdtime();
1440 is_even = 1;
1441 initialized = 1;
1442 ret = 0;
1443 goto out;
1444 }
1446 if (is_even) {
1447 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1448 goto out;
1449 time_odd = cdtime();
1450 is_even = 0;
1451 time_delta = time_odd - time_even;
1452 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1453 goto out;
1454 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1455 goto out;
1456 } else {
1457 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1458 goto out;
1459 time_even = cdtime();
1460 is_even = 1;
1461 time_delta = time_even - time_odd;
1462 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1463 goto out;
1464 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1465 goto out;
1466 }
1467 ret = 0;
1468 out:
1469 /*
1470 * Let's restore the affinity
1471 * This might fail if the number of CPU changed, but we can't do anything in that case..
1472 */
1473 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1474 return ret;
1475 }
1477 static int
1478 check_permissions(void)
1479 {
1481 if (getuid() == 0) {
1482 /* We have everything we need */
1483 return 0;
1484 #if !defined(HAVE_SYS_CAPABILITY_H) && !defined(CAP_SYS_RAWIO)
1485 } else {
1486 ERROR("turbostat plugin: Initialization failed: this plugin "
1487 "requires collectd to run as root");
1488 return -1;
1489 }
1490 #else /* HAVE_SYS_CAPABILITY_H && CAP_SYS_RAWIO */
1491 }
1493 int ret = 0;
1495 if (check_capability(CAP_SYS_RAWIO) != 0) {
1496 WARNING("turbostat plugin: Collectd doesn't have the "
1497 "CAP_SYS_RAWIO capability. If you don't want to run "
1498 "collectd as root, try running \"setcap "
1499 "cap_sys_rawio=ep\" on collectd binary");
1500 ret = -1;
1501 }
1503 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1504 WARNING("turbostat plugin: Collectd cannot open "
1505 "/dev/cpu/0/msr. If you don't want to run collectd as "
1506 "root, you need to change the ownership (chown) and "
1507 "permissions on /dev/cpu/*/msr to allow such access");
1508 ret = -1;
1509 }
1511 if (ret != 0)
1512 ERROR("turbostat plugin: Initialization failed: this plugin "
1513 "requires collectd to either to run as root or give "
1514 "collectd a special capability (CAP_SYS_RAWIO) and read "
1515 "access to /dev/cpu/*/msr (see previous warnings)");
1516 return ret;
1517 #endif /* HAVE_SYS_CAPABILITY_H && CAP_SYS_RAWIO */
1518 }
1520 static int
1521 turbostat_init(void)
1522 {
1523 struct stat sb;
1524 int ret;
1526 if (stat("/dev/cpu/0/msr", &sb)) {
1527 ERROR("turbostat plugin: Initialization failed: /dev/cpu/0/msr "
1528 "does not exist while the CPU supports MSR. You may be "
1529 "missing the corresponding kernel module, please try '# "
1530 "modprobe msr'");
1531 return -1;
1532 }
1534 DO_OR_GOTO_ERR(check_permissions());
1536 DO_OR_GOTO_ERR(probe_cpu());
1538 DO_OR_GOTO_ERR(setup_all_buffers());
1540 plugin_register_read(PLUGIN_NAME, turbostat_read);
1542 return 0;
1543 err:
1544 free_all_buffers();
1545 return ret;
1546 }
1548 static int
1549 turbostat_config(const char *key, const char *value)
1550 {
1551 long unsigned int tmp_val;
1552 char *end;
1554 if (strcasecmp("CoreCstates", key) == 0) {
1555 tmp_val = strtoul(value, &end, 0);
1556 if (*end != '\0' || tmp_val > UINT_MAX) {
1557 ERROR("turbostat plugin: Invalid CoreCstates '%s'",
1558 value);
1559 return -1;
1560 }
1561 config_core_cstate = (unsigned int) tmp_val;
1562 apply_config_core_cstate = 1;
1563 } else if (strcasecmp("PackageCstates", key) == 0) {
1564 tmp_val = strtoul(value, &end, 0);
1565 if (*end != '\0' || tmp_val > UINT_MAX) {
1566 ERROR("turbostat plugin: Invalid PackageCstates '%s'",
1567 value);
1568 return -1;
1569 }
1570 config_pkg_cstate = (unsigned int) tmp_val;
1571 apply_config_pkg_cstate = 1;
1572 } else if (strcasecmp("SystemManagementInterrupt", key) == 0) {
1573 config_smi = IS_TRUE(value);
1574 apply_config_smi = 1;
1575 } else if (strcasecmp("DigitalTemperatureSensor", key) == 0) {
1576 config_dts = IS_TRUE(value);
1577 apply_config_dts = 1;
1578 } else if (strcasecmp("PackageThermalManagement", key) == 0) {
1579 config_ptm = IS_TRUE(value);
1580 apply_config_ptm = 1;
1581 } else if (strcasecmp("LogicalCoreNames", key) == 0) {
1582 config_lcn = 1;
1583 } else if (strcasecmp("RunningAveragePowerLimit", key) == 0) {
1584 tmp_val = strtoul(value, &end, 0);
1585 if (*end != '\0' || tmp_val > UINT_MAX) {
1586 ERROR("turbostat plugin: Invalid RunningAveragePowerLimit '%s'",
1587 value);
1588 return -1;
1589 }
1590 config_rapl = (unsigned int) tmp_val;
1591 apply_config_rapl = 1;
1592 } else if (strcasecmp("TCCActivationTemp", key) == 0) {
1593 tmp_val = strtoul(value, &end, 0);
1594 if (*end != '\0' || tmp_val > UINT_MAX) {
1595 ERROR("turbostat plugin: Invalid TCCActivationTemp '%s'",
1596 value);
1597 return -1;
1598 }
1599 tcc_activation_temp = (unsigned int) tmp_val;
1600 } else {
1601 ERROR("turbostat plugin: Invalid configuration option '%s'",
1602 key);
1603 return -1;
1604 }
1605 return 0;
1606 }
1608 void module_register(void)
1609 {
1610 plugin_register_init(PLUGIN_NAME, turbostat_init);
1611 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1612 }