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 * Bitmask of the list of core C states supported by the processor.
69 * Currently supported C-states (by this plugin): 3, 6, 7
70 */
71 static unsigned int do_core_cstate;
72 static unsigned int config_core_cstate;
73 static _Bool apply_config_core_cstate;
75 /*
76 * Bitmask of the list of pacages C states supported by the processor.
77 * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
78 */
79 static unsigned int do_pkg_cstate;
80 static unsigned int config_pkg_cstate;
81 static _Bool apply_config_pkg_cstate;
83 /*
84 * Boolean indicating if the processor supports 'I/O System-Management Interrupt counter'
85 */
86 static _Bool do_smi;
87 static _Bool config_smi;
88 static _Bool apply_config_smi;
90 /*
91 * Boolean indicating if the processor supports 'Digital temperature sensor'
92 * This feature enables the monitoring of the temperature of each core
93 *
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
97 */
98 static _Bool do_dts;
99 static _Bool config_dts;
100 static _Bool apply_config_dts;
102 /*
103 * Boolean indicating if the processor supports 'Package thermal management'
104 * This feature allows the monitoring of the temperature of each package
105 *
106 * This feature has two limitations:
107 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
108 * - Temperatures above the tcc_activation_temp are not recorded
109 */
110 static _Bool do_ptm;
111 static _Bool config_ptm;
112 static _Bool apply_config_ptm;
114 /*
115 * Thermal Control Circuit Activation Temperature as configured by the user.
116 * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
117 * and should only be used if the automated detection fails.
118 */
119 static unsigned int tcc_activation_temp;
121 static unsigned int do_rapl;
122 static unsigned int config_rapl;
123 static _Bool apply_config_rapl;
124 static double rapl_energy_units;
126 #define RAPL_PKG (1 << 0)
127 /* 0x610 MSR_PKG_POWER_LIMIT */
128 /* 0x611 MSR_PKG_ENERGY_STATUS */
129 #define RAPL_DRAM (1 << 1)
130 /* 0x618 MSR_DRAM_POWER_LIMIT */
131 /* 0x619 MSR_DRAM_ENERGY_STATUS */
132 /* 0x61c MSR_DRAM_POWER_INFO */
133 #define RAPL_CORES (1 << 2)
134 /* 0x638 MSR_PP0_POWER_LIMIT */
135 /* 0x639 MSR_PP0_ENERGY_STATUS */
137 #define RAPL_GFX (1 << 3)
138 /* 0x640 MSR_PP1_POWER_LIMIT */
139 /* 0x641 MSR_PP1_ENERGY_STATUS */
140 /* 0x642 MSR_PP1_POLICY */
141 #define TJMAX_DEFAULT 100
143 static cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
144 static size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
146 static struct thread_data {
147 unsigned long long tsc;
148 unsigned long long aperf;
149 unsigned long long mperf;
150 unsigned long long c1;
151 unsigned int smi_count;
152 unsigned int cpu_id;
153 unsigned int flags;
154 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
155 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
156 } *thread_delta, *thread_even, *thread_odd;
158 static struct core_data {
159 unsigned long long c3;
160 unsigned long long c6;
161 unsigned long long c7;
162 unsigned int core_temp_c;
163 unsigned int core_id;
164 } *core_delta, *core_even, *core_odd;
166 static struct pkg_data {
167 unsigned long long pc2;
168 unsigned long long pc3;
169 unsigned long long pc6;
170 unsigned long long pc7;
171 unsigned long long pc8;
172 unsigned long long pc9;
173 unsigned long long pc10;
174 unsigned int package_id;
175 uint32_t energy_pkg; /* MSR_PKG_ENERGY_STATUS */
176 uint32_t energy_dram; /* MSR_DRAM_ENERGY_STATUS */
177 uint32_t energy_cores; /* MSR_PP0_ENERGY_STATUS */
178 uint32_t energy_gfx; /* MSR_PP1_ENERGY_STATUS */
179 unsigned int tcc_activation_temp;
180 unsigned int pkg_temp_c;
181 } *package_delta, *package_even, *package_odd;
183 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
184 #define ODD_COUNTERS thread_odd, core_odd, package_odd
185 #define EVEN_COUNTERS thread_even, core_even, package_even
186 static _Bool is_even = 1;
188 static _Bool allocated = 0;
189 static _Bool initialized = 0;
191 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
192 (thread_base + \
193 (pkg_no) * topology.num_cores * topology.num_threads + \
194 (core_no) * topology.num_threads + \
195 (thread_no))
196 #define GET_CORE(core_base, core_no, pkg_no) \
197 (core_base + \
198 (pkg_no) * topology.num_cores + \
199 (core_no))
200 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
202 struct cpu_topology {
203 unsigned int package_id;
204 unsigned int core_id;
205 _Bool first_core_in_package;
206 _Bool first_thread_in_core;
207 };
209 static struct topology {
210 unsigned int max_cpu_id;
211 unsigned int num_packages;
212 unsigned int num_cores;
213 unsigned int num_threads;
214 struct cpu_topology *cpus;
215 } topology;
217 static cdtime_t time_even, time_odd, time_delta;
219 static const char *config_keys[] =
220 {
221 "CoreCstates",
222 "PackageCstates",
223 "SystemManagementInterrupt",
224 "DigitalTemperatureSensor",
225 "PackageThermalManagement",
226 "TCCActivationTemp",
227 "RunningAveragePowerLimit",
228 };
229 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
231 /*****************************
232 * MSR Manipulation helpers *
233 *****************************/
235 /*
236 * Open a MSR device for reading
237 * Can change the scheduling affinity of the current process if multiple_read is 1
238 */
239 static int __attribute__((warn_unused_result))
240 open_msr(unsigned int cpu, _Bool multiple_read)
241 {
242 char pathname[32];
243 int fd;
245 /*
246 * If we need to do multiple read, let's migrate to the CPU
247 * Otherwise, we would lose time calling functions on another CPU
248 *
249 * If we are not yet initialized (cpu_affinity_setsize = 0),
250 * we need to skip this optimisation.
251 */
252 if (multiple_read && cpu_affinity_setsize) {
253 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
254 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
255 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
256 ERROR("turbostat plugin: Could not migrate to CPU %d", cpu);
257 return -1;
258 }
259 }
261 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
262 fd = open(pathname, O_RDONLY);
263 if (fd < 0) {
264 ERROR("turbostat plugin: failed to open %s", pathname);
265 return -1;
266 }
267 return fd;
268 }
270 /*
271 * Read a single MSR from an open file descriptor
272 */
273 static int __attribute__((warn_unused_result))
274 read_msr(int fd, off_t offset, unsigned long long *msr)
275 {
276 ssize_t retval;
278 retval = pread(fd, msr, sizeof *msr, offset);
280 if (retval != sizeof *msr) {
281 ERROR("turbostat plugin: MSR offset 0x%llx read failed",
282 (unsigned long long)offset);
283 return -1;
284 }
285 return 0;
286 }
288 /*
289 * Open a MSR device for reading, read the value asked for and close it.
290 * This call will not affect the scheduling affinity of this thread.
291 */
292 static ssize_t __attribute__((warn_unused_result))
293 get_msr(unsigned int cpu, off_t offset, unsigned long long *msr)
294 {
295 ssize_t retval;
296 int fd;
298 fd = open_msr(cpu, 0);
299 if (fd < 0)
300 return fd;
301 retval = read_msr(fd, offset, msr);
302 close(fd);
303 return retval;
304 }
307 /********************************
308 * Raw data acquisition (1 CPU) *
309 ********************************/
311 /*
312 * Read every data avalaible for a single CPU
313 *
314 * Core data is shared for all threads in one core: extracted only for the first thread
315 * Package data is shared for all core in one package: extracted only for the first thread of the first core
316 *
317 * Side effect: migrates to the targeted CPU
318 */
319 static int __attribute__((warn_unused_result))
320 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
321 {
322 unsigned int cpu = t->cpu_id;
323 unsigned long long msr;
324 int msr_fd;
325 int retval = 0;
327 msr_fd = open_msr(cpu, 1);
328 if (msr_fd < 0)
329 return msr_fd;
331 #define READ_MSR(msr, dst) \
332 do { \
333 if (read_msr(msr_fd, msr, dst)) { \
334 ERROR("turbostat plugin: Unable to read " #msr); \
335 retval = -1; \
336 goto out; \
337 } \
338 } while (0)
340 READ_MSR(MSR_IA32_TSC, &t->tsc);
342 READ_MSR(MSR_IA32_APERF, &t->aperf);
343 READ_MSR(MSR_IA32_MPERF, &t->mperf);
345 if (do_smi) {
346 READ_MSR(MSR_SMI_COUNT, &msr);
347 t->smi_count = msr & 0xFFFFFFFF;
348 }
350 /* collect core counters only for 1st thread in core */
351 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) {
352 retval = 0;
353 goto out;
354 }
356 if (do_core_cstate & (1 << 3))
357 READ_MSR(MSR_CORE_C3_RESIDENCY, &c->c3);
358 if (do_core_cstate & (1 << 6))
359 READ_MSR(MSR_CORE_C6_RESIDENCY, &c->c6);
360 if (do_core_cstate & (1 << 7))
361 READ_MSR(MSR_CORE_C7_RESIDENCY, &c->c7);
363 if (do_dts) {
364 READ_MSR(MSR_IA32_THERM_STATUS, &msr);
365 c->core_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
366 }
368 /* collect package counters only for 1st core in package */
369 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
370 retval = 0;
371 goto out;
372 }
374 if (do_pkg_cstate & (1 << 2))
375 READ_MSR(MSR_PKG_C2_RESIDENCY, &p->pc2);
376 if (do_pkg_cstate & (1 << 3))
377 READ_MSR(MSR_PKG_C3_RESIDENCY, &p->pc3);
378 if (do_pkg_cstate & (1 << 6))
379 READ_MSR(MSR_PKG_C6_RESIDENCY, &p->pc6);
380 if (do_pkg_cstate & (1 << 7))
381 READ_MSR(MSR_PKG_C7_RESIDENCY, &p->pc7);
382 if (do_pkg_cstate & (1 << 8))
383 READ_MSR(MSR_PKG_C8_RESIDENCY, &p->pc8);
384 if (do_pkg_cstate & (1 << 9))
385 READ_MSR(MSR_PKG_C9_RESIDENCY, &p->pc9);
386 if (do_pkg_cstate & (1 << 10))
387 READ_MSR(MSR_PKG_C10_RESIDENCY, &p->pc10);
389 if (do_rapl & RAPL_PKG) {
390 READ_MSR(MSR_PKG_ENERGY_STATUS, &msr);
391 p->energy_pkg = msr & 0xFFFFFFFF;
392 }
393 if (do_rapl & RAPL_CORES) {
394 READ_MSR(MSR_PP0_ENERGY_STATUS, &msr);
395 p->energy_cores = msr & 0xFFFFFFFF;
396 }
397 if (do_rapl & RAPL_DRAM) {
398 READ_MSR(MSR_DRAM_ENERGY_STATUS, &msr);
399 p->energy_dram = msr & 0xFFFFFFFF;
400 }
401 if (do_rapl & RAPL_GFX) {
402 READ_MSR(MSR_PP1_ENERGY_STATUS, &msr);
403 p->energy_gfx = msr & 0xFFFFFFFF;
404 }
405 if (do_ptm) {
406 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
407 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
408 }
410 out:
411 close(msr_fd);
412 return retval;
413 }
416 /**********************************
417 * Evaluating the changes (1 CPU) *
418 **********************************/
420 /*
421 * Extract the evolution old->new in delta at a package level
422 * (some are not new-delta, e.g. temperature)
423 */
424 static inline void
425 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
426 {
427 delta->pc2 = new->pc2 - old->pc2;
428 delta->pc3 = new->pc3 - old->pc3;
429 delta->pc6 = new->pc6 - old->pc6;
430 delta->pc7 = new->pc7 - old->pc7;
431 delta->pc8 = new->pc8 - old->pc8;
432 delta->pc9 = new->pc9 - old->pc9;
433 delta->pc10 = new->pc10 - old->pc10;
434 delta->pkg_temp_c = new->pkg_temp_c;
436 delta->energy_pkg = new->energy_pkg - old->energy_pkg;
437 delta->energy_cores = new->energy_cores - old->energy_cores;
438 delta->energy_gfx = new->energy_gfx - old->energy_gfx;
439 delta->energy_dram = new->energy_dram - old->energy_dram;
440 }
442 /*
443 * Extract the evolution old->new in delta at a core level
444 * (some are not new-delta, e.g. temperature)
445 */
446 static inline void
447 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
448 {
449 delta->c3 = new->c3 - old->c3;
450 delta->c6 = new->c6 - old->c6;
451 delta->c7 = new->c7 - old->c7;
452 delta->core_temp_c = new->core_temp_c;
453 }
455 /*
456 * Extract the evolution old->new in delta at a package level
457 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
458 */
459 static inline int __attribute__((warn_unused_result))
460 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
461 const struct core_data *cdelta)
462 {
463 delta->tsc = new->tsc - old->tsc;
465 /* check for TSC < 1 Mcycles over interval */
466 if (delta->tsc < (1000 * 1000)) {
467 WARNING("turbostat plugin: Insanely slow TSC rate, TSC stops "
468 "in idle? You can disable all c-states by booting with"
469 " 'idle=poll' or just the deep ones with"
470 " 'processor.max_cstate=1'");
471 return -1;
472 }
474 delta->c1 = new->c1 - old->c1;
476 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
477 delta->aperf = new->aperf - old->aperf;
478 delta->mperf = new->mperf - old->mperf;
479 } else {
480 if (!aperf_mperf_unstable) {
481 WARNING("turbostat plugin: APERF or MPERF went "
482 "backwards. Frequency results do not cover "
483 "the entire interval. Fix this by running "
484 "Linux-2.6.30 or later.");
486 aperf_mperf_unstable = 1;
487 }
488 }
490 /*
491 * As counter collection is not atomic,
492 * it is possible for mperf's non-halted cycles + idle states
493 * to exceed TSC's all cycles: show c1 = 0% in that case.
494 */
495 if ((delta->mperf + cdelta->c3 + cdelta->c6 + cdelta->c7) > delta->tsc)
496 delta->c1 = 0;
497 else {
498 /* normal case, derive c1 */
499 delta->c1 = delta->tsc - delta->mperf - cdelta->c3
500 - cdelta->c6 - cdelta->c7;
501 }
503 if (delta->mperf == 0) {
504 WARNING("turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
505 delta->mperf = 1; /* divide by 0 protection */
506 }
508 if (do_smi)
509 delta->smi_count = new->smi_count - old->smi_count;
511 return 0;
512 }
514 /**********************************
515 * Submitting the results (1 CPU) *
516 **********************************/
518 /*
519 * Submit one gauge value
520 */
521 static void
522 turbostat_submit (const char *plugin_instance,
523 const char *type, const char *type_instance,
524 gauge_t value)
525 {
526 value_list_t vl = VALUE_LIST_INIT;
527 value_t v;
529 v.gauge = value;
530 vl.values = &v;
531 vl.values_len = 1;
532 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
533 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
534 if (plugin_instance != NULL)
535 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
536 sstrncpy (vl.type, type, sizeof (vl.type));
537 if (type_instance != NULL)
538 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
540 plugin_dispatch_values (&vl);
541 }
543 /*
544 * Submit every data for a single CPU
545 *
546 * Core data is shared for all threads in one core: submitted only for the first thread
547 * Package data is shared for all core in one package: submitted only for the first thread of the first core
548 */
549 static int
550 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
551 {
552 char name[DATA_MAX_NAME_LEN];
553 double interval_float;
555 interval_float = CDTIME_T_TO_DOUBLE(time_delta);
557 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
559 if (!aperf_mperf_unstable)
560 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
561 if (!aperf_mperf_unstable)
562 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
564 turbostat_submit(name, "frequency", "average", 1.0 / 1000000 * t->aperf / interval_float);
566 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
567 turbostat_submit(name, "frequency", "busy", 1.0 * t->tsc / 1000000 * t->aperf / t->mperf / interval_float);
569 /* Sanity check (should stay stable) */
570 turbostat_submit(name, "gauge", "TSC", 1.0 * t->tsc / 1000000 / interval_float);
572 /* SMI */
573 if (do_smi)
574 turbostat_submit(name, "count", NULL, t->smi_count);
576 /* submit per-core data only for 1st thread in core */
577 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
578 goto done;
580 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
582 if (do_core_cstate & (1 << 3))
583 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
584 if (do_core_cstate & (1 << 6))
585 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
586 if (do_core_cstate & (1 << 7))
587 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
589 if (do_dts)
590 turbostat_submit(name, "temperature", NULL, c->core_temp_c);
592 /* submit per-package data only for 1st core in package */
593 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
594 goto done;
596 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
598 if (do_ptm)
599 turbostat_submit(name, "temperature", NULL, p->pkg_temp_c);
601 if (do_pkg_cstate & (1 << 2))
602 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
603 if (do_pkg_cstate & (1 << 3))
604 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
605 if (do_pkg_cstate & (1 << 6))
606 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
607 if (do_pkg_cstate & (1 << 7))
608 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
609 if (do_pkg_cstate & (1 << 8))
610 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
611 if (do_pkg_cstate & (1 << 9))
612 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
613 if (do_pkg_cstate & (1 << 10))
614 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
616 if (do_rapl) {
617 if (do_rapl & RAPL_PKG)
618 turbostat_submit(name, "power", "pkg", p->energy_pkg * rapl_energy_units / interval_float);
619 if (do_rapl & RAPL_CORES)
620 turbostat_submit(name, "power", "cores", p->energy_cores * rapl_energy_units / interval_float);
621 if (do_rapl & RAPL_GFX)
622 turbostat_submit(name, "power", "GFX", p->energy_gfx * rapl_energy_units / interval_float);
623 if (do_rapl & RAPL_DRAM)
624 turbostat_submit(name, "power", "DRAM", p->energy_dram * rapl_energy_units / interval_float);
625 }
626 done:
627 return 0;
628 }
631 /**********************************
632 * Looping function over all CPUs *
633 **********************************/
635 /*
636 * Check if a given cpu id is in our compiled list of existing CPUs
637 */
638 static int
639 cpu_is_not_present(unsigned int cpu)
640 {
641 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
642 }
644 /*
645 * Loop on all CPUs in topological order
646 *
647 * Skip non-present cpus
648 * Return the error code at the first error or 0
649 */
650 static int __attribute__((warn_unused_result))
651 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
652 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
653 {
654 int retval;
656 for (unsigned int pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
657 for (unsigned int core_no = 0; core_no < topology.num_cores; ++core_no) {
658 for (unsigned int thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
659 struct thread_data *t;
660 struct core_data *c;
661 struct pkg_data *p;
663 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
665 if (cpu_is_not_present(t->cpu_id))
666 continue;
668 c = GET_CORE(core_base, core_no, pkg_no);
669 p = GET_PKG(pkg_base, pkg_no);
671 retval = func(t, c, p);
672 if (retval)
673 return retval;
674 }
675 }
676 }
677 return 0;
678 }
680 /*
681 * Dedicated loop: Extract every data evolution for all CPU
682 *
683 * Skip non-present cpus
684 * Return the error code at the first error or 0
685 *
686 * Core data is shared for all threads in one core: extracted only for the first thread
687 * Package data is shared for all core in one package: extracted only for the first thread of the first core
688 */
689 static int __attribute__((warn_unused_result))
690 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,
691 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
692 {
693 int retval;
695 for (unsigned int pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
696 for (unsigned int core_no = 0; core_no < topology.num_cores; ++core_no) {
697 for (unsigned int thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
698 struct thread_data *t_delta;
699 const struct thread_data *t_old, *t_new;
700 struct core_data *c_delta;
702 /* Get correct pointers for threads */
703 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
704 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
705 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
707 /* Skip threads that disappeared */
708 if (cpu_is_not_present(t_delta->cpu_id))
709 continue;
711 /* c_delta is always required for delta_thread */
712 c_delta = GET_CORE(core_delta, core_no, pkg_no);
714 /* calculate core delta only for 1st thread in core */
715 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
716 const struct core_data *c_old, *c_new;
718 c_new = GET_CORE(core_new_base, core_no, pkg_no);
719 c_old = GET_CORE(core_old_base, core_no, pkg_no);
721 delta_core(c_delta, c_new, c_old);
722 }
724 /* Always calculate thread delta */
725 retval = delta_thread(t_delta, t_new, t_old, c_delta);
726 if (retval)
727 return retval;
729 /* calculate package delta only for 1st core in package */
730 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
731 struct pkg_data *p_delta;
732 const struct pkg_data *p_old, *p_new;
734 p_delta = GET_PKG(package_delta, pkg_no);
735 p_new = GET_PKG(pkg_new_base, pkg_no);
736 p_old = GET_PKG(pkg_old_base, pkg_no);
738 delta_package(p_delta, p_new, p_old);
739 }
740 }
741 }
742 }
743 return 0;
744 }
747 /***************
748 * CPU Probing *
749 ***************/
751 /*
752 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
753 * the Thermal Control Circuit (TCC) activates.
754 * This is usually equal to tjMax.
755 *
756 * Older processors do not have this MSR, so there we guess,
757 * but also allow conficuration over-ride with "TCCActivationTemp".
758 *
759 * Several MSR temperature values are in units of degrees-C
760 * below this value, including the Digital Thermal Sensor (DTS),
761 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
762 */
763 static int __attribute__((warn_unused_result))
764 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
765 {
766 unsigned long long msr;
767 unsigned int target_c_local;
769 /* tcc_activation_temp is used only for dts or ptm */
770 if (!(do_dts || do_ptm))
771 return 0;
773 /* this is a per-package concept */
774 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
775 return 0;
777 if (tcc_activation_temp != 0) {
778 p->tcc_activation_temp = tcc_activation_temp;
779 return 0;
780 }
782 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
783 goto guess;
785 target_c_local = (msr >> 16) & 0xFF;
787 if (!target_c_local)
788 goto guess;
790 p->tcc_activation_temp = target_c_local;
792 return 0;
794 guess:
795 p->tcc_activation_temp = TJMAX_DEFAULT;
796 WARNING("turbostat plugin: cpu%d: Guessing tjMax %d C,"
797 " Please use TCCActivationTemp to specify it.",
798 t->cpu_id, p->tcc_activation_temp);
800 return 0;
801 }
803 /*
804 * Identify the functionality of the CPU
805 */
806 static int __attribute__((warn_unused_result))
807 probe_cpu(void)
808 {
809 unsigned int eax, ebx, ecx, edx, max_level;
810 unsigned int fms, family, model;
812 /* CPUID(0):
813 * - EAX: Maximum Input Value for Basic CPUID Information
814 * - EBX: "Genu" (0x756e6547)
815 * - EDX: "ineI" (0x49656e69)
816 * - ECX: "ntel" (0x6c65746e)
817 */
818 max_level = ebx = ecx = edx = 0;
819 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
820 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
821 ERROR("turbostat plugin: Unsupported CPU (not Intel)");
822 return -1;
823 }
825 /* CPUID(1):
826 * - EAX: Version Information: Type, Family, Model, and Stepping ID
827 * + 4-7: Model ID
828 * + 8-11: Family ID
829 * + 12-13: Processor type
830 * + 16-19: Extended Model ID
831 * + 20-27: Extended Family ID
832 * - EDX: Feature Information:
833 * + 5: Support for MSR read/write operations
834 */
835 fms = ebx = ecx = edx = 0;
836 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
837 family = (fms >> 8) & 0xf;
838 model = (fms >> 4) & 0xf;
839 if (family == 0xf)
840 family += (fms >> 20) & 0xf;
841 if (family == 6 || family == 0xf)
842 model += ((fms >> 16) & 0xf) << 4;
843 if (!(edx & (1 << 5))) {
844 ERROR("turbostat plugin: Unsupported CPU (no MSR support)");
845 return -1;
846 }
848 /*
849 * CPUID(6):
850 * - EAX:
851 * + 0: Digital temperature sensor is supported if set
852 * + 6: Package thermal management is supported if set
853 * - ECX:
854 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
855 * + 3: The processor supports performance-energy bias preference if set.
856 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
857 *
858 * This check is valid for both Intel and AMD
859 */
860 eax = ebx = ecx = edx = 0;
861 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
862 do_dts = eax & (1 << 0);
863 do_ptm = eax & (1 << 6);
864 if (!(ecx & (1 << 0))) {
865 ERROR("turbostat plugin: Unsupported CPU (No APERF)");
866 return -1;
867 }
869 /*
870 * Enable or disable C states depending on the model and family
871 */
872 if (family == 6) {
873 switch (model) {
874 /* Atom (partial) */
875 case 0x27:
876 do_smi = 0;
877 do_core_cstate = 0;
878 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
879 break;
880 /* Silvermont */
881 case 0x37: /* BYT */
882 case 0x4D: /* AVN */
883 do_smi = 1;
884 do_core_cstate = (1 << 1) | (1 << 6);
885 do_pkg_cstate = (1 << 6);
886 break;
887 /* Nehalem */
888 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
889 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
890 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
891 case 0x2E: /* Nehalem-EX Xeon - Beckton */
892 do_smi = 1;
893 do_core_cstate = (1 << 3) | (1 << 6);
894 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
895 break;
896 /* Westmere */
897 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
898 case 0x2C: /* Westmere EP - Gulftown */
899 case 0x2F: /* Westmere-EX Xeon - Eagleton */
900 do_smi = 1;
901 do_core_cstate = (1 << 3) | (1 << 6);
902 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
903 break;
904 /* Sandy Bridge */
905 case 0x2A: /* SNB */
906 case 0x2D: /* SNB Xeon */
907 do_smi = 1;
908 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
909 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
910 break;
911 /* Ivy Bridge */
912 case 0x3A: /* IVB */
913 case 0x3E: /* IVB Xeon */
914 do_smi = 1;
915 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
916 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
917 break;
918 /* Haswell Bridge */
919 case 0x3C: /* HSW */
920 case 0x3F: /* HSW */
921 case 0x46: /* HSW */
922 do_smi = 1;
923 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
924 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
925 break;
926 case 0x45: /* HSW */
927 do_smi = 1;
928 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
929 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
930 break;
931 /* Broadwel */
932 case 0x4F: /* BDW */
933 case 0x56: /* BDX-DE */
934 do_smi = 1;
935 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
936 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
937 break;
938 case 0x3D: /* BDW */
939 do_smi = 1;
940 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
941 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
942 break;
943 default:
944 do_smi = 0;
945 do_core_cstate = 0;
946 do_pkg_cstate = 0;
947 break;
948 }
949 switch (model) {
950 case 0x2A: /* SNB */
951 case 0x3A: /* IVB */
952 case 0x3C: /* HSW */
953 case 0x45: /* HSW */
954 case 0x46: /* HSW */
955 case 0x3D: /* BDW */
956 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_GFX;
957 break;
958 case 0x3F: /* HSX */
959 case 0x4F: /* BDX */
960 case 0x56: /* BDX-DE */
961 do_rapl = RAPL_PKG | RAPL_DRAM ;
962 break;
963 case 0x2D: /* SNB Xeon */
964 case 0x3E: /* IVB Xeon */
965 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_DRAM;
966 break;
967 case 0x37: /* BYT */
968 case 0x4D: /* AVN */
969 do_rapl = RAPL_PKG | RAPL_CORES;
970 break;
971 default:
972 do_rapl = 0;
973 }
974 } else {
975 ERROR("turbostat plugin: Unsupported CPU (family: %#x, "
976 "model: %#x)", family, model);
977 return -1;
978 }
980 /* Override detected values with configuration */
981 if (apply_config_core_cstate)
982 do_core_cstate = config_core_cstate;
983 if (apply_config_pkg_cstate)
984 do_pkg_cstate = config_pkg_cstate;
985 if (apply_config_smi)
986 do_smi = config_smi;
987 if (apply_config_dts)
988 do_dts = config_dts;
989 if (apply_config_ptm)
990 do_ptm = config_ptm;
991 if (apply_config_rapl)
992 do_rapl = config_rapl;
994 if (do_rapl) {
995 unsigned long long msr;
996 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
997 return 0;
999 if (model == 0x37)
1000 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1001 else
1002 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1003 }
1005 return 0;
1006 }
1009 /********************
1010 * Topology Probing *
1011 ********************/
1013 /*
1014 * Read a single int from a file.
1015 */
1016 static int __attribute__ ((format(printf,1,2)))
1017 parse_int_file(const char *fmt, ...)
1018 {
1019 va_list args;
1020 char path[PATH_MAX];
1021 int len;
1023 va_start(args, fmt);
1024 len = vsnprintf(path, sizeof(path), fmt, args);
1025 va_end(args);
1026 if (len < 0 || len >= PATH_MAX) {
1027 ERROR("turbostat plugin: path truncated: '%s'", path);
1028 return -1;
1029 }
1031 value_t v;
1032 if (parse_value_file (path, &v, DS_TYPE_DERIVE) != 0) {
1033 ERROR ("turbostat plugin: Parsing \"%s\" failed.", path);
1034 return -1;
1035 }
1037 return (int) v.derive;
1038 }
1040 static int
1041 get_threads_on_core(unsigned int cpu)
1042 {
1043 char path[80];
1044 FILE *filep;
1045 int sib1, sib2;
1046 int matches;
1047 char character;
1049 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1050 filep = fopen(path, "r");
1051 if (!filep) {
1052 ERROR("turbostat plugin: Failed to open '%s'", path);
1053 return -1;
1054 }
1055 /*
1056 * file format:
1057 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1058 * otherwinse 1 sibling (self).
1059 */
1060 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1062 fclose(filep);
1064 if (matches == 3)
1065 return 2;
1066 else
1067 return 1;
1068 }
1070 /*
1071 * run func(cpu) on every cpu in /proc/stat
1072 * return max_cpu number
1073 */
1074 static int __attribute__((warn_unused_result))
1075 for_all_proc_cpus(int (func)(unsigned int))
1076 {
1077 FILE *fp;
1078 unsigned int cpu_num;
1079 int retval;
1081 fp = fopen("/proc/stat", "r");
1082 if (!fp) {
1083 ERROR("turbostat plugin: Failed to open /proc/stat");
1084 return -1;
1085 }
1087 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1088 if (retval != 0) {
1089 ERROR("turbostat plugin: Failed to parse /proc/stat");
1090 fclose(fp);
1091 return -1;
1092 }
1094 while (1) {
1095 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1096 if (retval != 1)
1097 break;
1099 retval = func(cpu_num);
1100 if (retval) {
1101 fclose(fp);
1102 return(retval);
1103 }
1104 }
1105 fclose(fp);
1106 return 0;
1107 }
1109 /*
1110 * Update the stored topology.max_cpu_id
1111 */
1112 static int
1113 update_max_cpu_id(unsigned int cpu)
1114 {
1115 if (topology.max_cpu_id < cpu)
1116 topology.max_cpu_id = cpu;
1117 return 0;
1118 }
1120 static int
1121 mark_cpu_present(unsigned int cpu)
1122 {
1123 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1124 return 0;
1125 }
1127 static int __attribute__((warn_unused_result))
1128 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1129 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1130 if (*set == NULL) {
1131 ERROR("turbostat plugin: Unable to allocate CPU state");
1132 return -1;
1133 }
1134 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1135 CPU_ZERO_S(*size, *set);
1136 return 0;
1137 }
1139 /*
1140 * Build a local representation of the cpu distribution
1141 */
1142 static int __attribute__((warn_unused_result))
1143 topology_probe(void)
1144 {
1145 int ret;
1146 unsigned int max_package_id, max_core_id, max_threads;
1147 max_package_id = max_core_id = max_threads = 0;
1149 /* Clean topology */
1150 free(topology.cpus);
1151 memset(&topology, 0, sizeof(topology));
1153 ret = for_all_proc_cpus(update_max_cpu_id);
1154 if (ret != 0)
1155 goto err;
1157 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1158 if (topology.cpus == NULL) {
1159 ERROR("turbostat plugin: Unable to allocate memory for CPU topology");
1160 return -1;
1161 }
1163 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1164 if (ret != 0)
1165 goto err;
1166 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1167 if (ret != 0)
1168 goto err;
1169 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1170 if (ret != 0)
1171 goto err;
1173 ret = for_all_proc_cpus(mark_cpu_present);
1174 if (ret != 0)
1175 goto err;
1177 /*
1178 * For online cpus
1179 * find max_core_id, max_package_id
1180 */
1181 for (unsigned int i = 0; i <= topology.max_cpu_id; ++i) {
1182 unsigned int num_threads;
1183 struct cpu_topology *cpu = &topology.cpus[i];
1185 if (cpu_is_not_present(i)) {
1186 WARNING("turbostat plugin: cpu%d NOT PRESENT", i);
1187 continue;
1188 }
1190 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1191 if (ret < 0)
1192 goto err;
1193 else
1194 cpu->package_id = (unsigned int) ret;
1195 if (cpu->package_id > max_package_id)
1196 max_package_id = cpu->package_id;
1198 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1199 if (ret < 0)
1200 goto err;
1201 else
1202 cpu->core_id = (unsigned int) ret;
1203 if (cpu->core_id > max_core_id)
1204 max_core_id = cpu->core_id;
1205 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1206 if (ret < 0)
1207 goto err;
1208 else if ((unsigned int) ret == i)
1209 cpu->first_core_in_package = 1;
1211 ret = get_threads_on_core(i);
1212 if (ret < 0)
1213 goto err;
1214 else
1215 num_threads = (unsigned int) ret;
1216 if (num_threads > max_threads)
1217 max_threads = num_threads;
1218 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1219 if (ret < 0)
1220 goto err;
1221 else if ((unsigned int) ret == i)
1222 cpu->first_thread_in_core = 1;
1224 DEBUG("turbostat plugin: cpu %d pkg %d core %d\n",
1225 i, cpu->package_id, cpu->core_id);
1226 }
1227 /* Num is max + 1 (need to count 0) */
1228 topology.num_packages = max_package_id + 1;
1229 topology.num_cores = max_core_id + 1;
1230 topology.num_threads = max_threads;
1232 return 0;
1233 err:
1234 free(topology.cpus);
1235 return ret;
1236 }
1239 /************************
1240 * Main alloc/init/free *
1241 ************************/
1243 static int
1244 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1245 {
1246 unsigned int total_threads, total_cores;
1248 if ((topology.num_threads == 0)
1249 || (topology.num_cores == 0)
1250 || (topology.num_packages == 0))
1251 {
1252 ERROR ("turbostat plugin: Invalid topology: %u threads, %u cores, %u packages",
1253 topology.num_threads, topology.num_cores, topology.num_packages);
1254 return -1;
1255 }
1257 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1258 *threads = calloc(total_threads, sizeof(struct thread_data));
1259 if (*threads == NULL)
1260 {
1261 ERROR ("turbostat plugin: calloc failed");
1262 return -1;
1263 }
1265 for (unsigned int i = 0; i < total_threads; ++i)
1266 (*threads)[i].cpu_id = topology.max_cpu_id + 1;
1268 total_cores = topology.num_cores * topology.num_packages;
1269 *cores = calloc(total_cores, sizeof(struct core_data));
1270 if (*cores == NULL)
1271 {
1272 ERROR ("turbostat plugin: calloc failed");
1273 sfree (threads);
1274 return -1;
1275 }
1277 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1278 if (*packages == NULL)
1279 {
1280 ERROR ("turbostat plugin: calloc failed");
1281 sfree (cores);
1282 sfree (threads);
1283 return -1;
1284 }
1286 return 0;
1287 }
1289 static void
1290 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1291 struct pkg_data *pkg_base, unsigned int cpu_id)
1292 {
1293 struct thread_data *t;
1294 struct core_data *c;
1295 struct pkg_data *p;
1296 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1298 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1299 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1300 p = GET_PKG(pkg_base, cpu->package_id);
1302 t->cpu_id = cpu_id;
1303 if (cpu->first_thread_in_core)
1304 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1305 if (cpu->first_core_in_package)
1306 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1308 c->core_id = cpu->core_id;
1309 p->package_id = cpu->package_id;
1310 }
1312 static void
1313 initialize_counters(void)
1314 {
1315 for (unsigned int cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1316 if (cpu_is_not_present(cpu_id))
1317 continue;
1318 init_counter(EVEN_COUNTERS, cpu_id);
1319 init_counter(ODD_COUNTERS, cpu_id);
1320 init_counter(DELTA_COUNTERS, cpu_id);
1321 }
1322 }
1326 static void
1327 free_all_buffers(void)
1328 {
1329 allocated = 0;
1330 initialized = 0;
1332 CPU_FREE(cpu_present_set);
1333 cpu_present_set = NULL;
1334 cpu_present_setsize = 0;
1336 CPU_FREE(cpu_affinity_set);
1337 cpu_affinity_set = NULL;
1338 cpu_affinity_setsize = 0;
1340 CPU_FREE(cpu_saved_affinity_set);
1341 cpu_saved_affinity_set = NULL;
1342 cpu_saved_affinity_setsize = 0;
1344 free(thread_even);
1345 free(core_even);
1346 free(package_even);
1348 thread_even = NULL;
1349 core_even = NULL;
1350 package_even = NULL;
1352 free(thread_odd);
1353 free(core_odd);
1354 free(package_odd);
1356 thread_odd = NULL;
1357 core_odd = NULL;
1358 package_odd = NULL;
1360 free(thread_delta);
1361 free(core_delta);
1362 free(package_delta);
1364 thread_delta = NULL;
1365 core_delta = NULL;
1366 package_delta = NULL;
1367 }
1370 /**********************
1371 * Collectd functions *
1372 **********************/
1374 #define DO_OR_GOTO_ERR(something) \
1375 do { \
1376 ret = (something); \
1377 if (ret < 0) \
1378 goto err; \
1379 } while (0)
1381 static int setup_all_buffers(void)
1382 {
1383 int ret;
1385 DO_OR_GOTO_ERR(topology_probe());
1386 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1387 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1388 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1389 initialize_counters();
1390 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1391 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1393 allocated = 1;
1394 return 0;
1395 err:
1396 free_all_buffers();
1397 return ret;
1398 }
1400 static int
1401 turbostat_read(void)
1402 {
1403 int ret;
1405 if (!allocated) {
1406 if ((ret = setup_all_buffers()) < 0)
1407 return ret;
1408 }
1410 if (for_all_proc_cpus(cpu_is_not_present)) {
1411 free_all_buffers();
1412 if ((ret = setup_all_buffers()) < 0)
1413 return ret;
1414 if (for_all_proc_cpus(cpu_is_not_present)) {
1415 ERROR("turbostat plugin: CPU appeared just after "
1416 "initialization");
1417 return -1;
1418 }
1419 }
1421 /* Saving the scheduling affinity, as it will be modified by get_counters */
1422 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1423 ERROR("turbostat plugin: Unable to save the CPU affinity");
1424 return -1;
1425 }
1427 if (!initialized) {
1428 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1429 goto out;
1430 time_even = cdtime();
1431 is_even = 1;
1432 initialized = 1;
1433 ret = 0;
1434 goto out;
1435 }
1437 if (is_even) {
1438 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1439 goto out;
1440 time_odd = cdtime();
1441 is_even = 0;
1442 time_delta = time_odd - time_even;
1443 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1444 goto out;
1445 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1446 goto out;
1447 } else {
1448 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1449 goto out;
1450 time_even = cdtime();
1451 is_even = 1;
1452 time_delta = time_even - time_odd;
1453 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1454 goto out;
1455 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1456 goto out;
1457 }
1458 ret = 0;
1459 out:
1460 /*
1461 * Let's restore the affinity
1462 * This might fail if the number of CPU changed, but we can't do anything in that case..
1463 */
1464 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1465 return ret;
1466 }
1468 static int
1469 check_permissions(void)
1470 {
1472 if (getuid() == 0) {
1473 /* We have everything we need */
1474 return 0;
1475 #if !defined(HAVE_SYS_CAPABILITY_H) && !defined(CAP_SYS_RAWIO)
1476 } else {
1477 ERROR("turbostat plugin: Initialization failed: this plugin "
1478 "requires collectd to run as root");
1479 return -1;
1480 }
1481 #else /* HAVE_SYS_CAPABILITY_H && CAP_SYS_RAWIO */
1482 }
1484 int ret = 0;
1486 if (check_capability(CAP_SYS_RAWIO) != 0) {
1487 WARNING("turbostat plugin: Collectd doesn't have the "
1488 "CAP_SYS_RAWIO capability. If you don't want to run "
1489 "collectd as root, try running \"setcap "
1490 "cap_sys_rawio=ep\" on collectd binary");
1491 ret = -1;
1492 }
1494 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1495 WARNING("turbostat plugin: Collectd cannot open "
1496 "/dev/cpu/0/msr. If you don't want to run collectd as "
1497 "root, you need to change the ownership (chown) and "
1498 "permissions on /dev/cpu/*/msr to allow such access");
1499 ret = -1;
1500 }
1502 if (ret != 0)
1503 ERROR("turbostat plugin: Initialization failed: this plugin "
1504 "requires collectd to either to run as root or give "
1505 "collectd a special capability (CAP_SYS_RAWIO) and read "
1506 "access to /dev/cpu/*/msr (see previous warnings)");
1507 return ret;
1508 #endif /* HAVE_SYS_CAPABILITY_H && CAP_SYS_RAWIO */
1509 }
1511 static int
1512 turbostat_init(void)
1513 {
1514 struct stat sb;
1515 int ret;
1517 if (stat("/dev/cpu/0/msr", &sb)) {
1518 ERROR("turbostat plugin: Initialization failed: /dev/cpu/0/msr "
1519 "does not exist while the CPU supports MSR. You may be "
1520 "missing the corresponding kernel module, please try '# "
1521 "modprobe msr'");
1522 return -1;
1523 }
1525 DO_OR_GOTO_ERR(check_permissions());
1527 DO_OR_GOTO_ERR(probe_cpu());
1529 DO_OR_GOTO_ERR(setup_all_buffers());
1531 plugin_register_read(PLUGIN_NAME, turbostat_read);
1533 return 0;
1534 err:
1535 free_all_buffers();
1536 return ret;
1537 }
1539 static int
1540 turbostat_config(const char *key, const char *value)
1541 {
1542 long unsigned int tmp_val;
1543 char *end;
1545 if (strcasecmp("CoreCstates", key) == 0) {
1546 tmp_val = strtoul(value, &end, 0);
1547 if (*end != '\0' || tmp_val > UINT_MAX) {
1548 ERROR("turbostat plugin: Invalid CoreCstates '%s'",
1549 value);
1550 return -1;
1551 }
1552 config_core_cstate = (unsigned int) tmp_val;
1553 apply_config_core_cstate = 1;
1554 } else if (strcasecmp("PackageCstates", key) == 0) {
1555 tmp_val = strtoul(value, &end, 0);
1556 if (*end != '\0' || tmp_val > UINT_MAX) {
1557 ERROR("turbostat plugin: Invalid PackageCstates '%s'",
1558 value);
1559 return -1;
1560 }
1561 config_pkg_cstate = (unsigned int) tmp_val;
1562 apply_config_pkg_cstate = 1;
1563 } else if (strcasecmp("SystemManagementInterrupt", key) == 0) {
1564 config_smi = IS_TRUE(value);
1565 apply_config_smi = 1;
1566 } else if (strcasecmp("DigitalTemperatureSensor", key) == 0) {
1567 config_dts = IS_TRUE(value);
1568 apply_config_dts = 1;
1569 } else if (strcasecmp("PackageThermalManagement", key) == 0) {
1570 config_ptm = IS_TRUE(value);
1571 apply_config_ptm = 1;
1572 } else if (strcasecmp("RunningAveragePowerLimit", key) == 0) {
1573 tmp_val = strtoul(value, &end, 0);
1574 if (*end != '\0' || tmp_val > UINT_MAX) {
1575 ERROR("turbostat plugin: Invalid RunningAveragePowerLimit '%s'",
1576 value);
1577 return -1;
1578 }
1579 config_rapl = (unsigned int) tmp_val;
1580 apply_config_rapl = 1;
1581 } else if (strcasecmp("TCCActivationTemp", key) == 0) {
1582 tmp_val = strtoul(value, &end, 0);
1583 if (*end != '\0' || tmp_val > UINT_MAX) {
1584 ERROR("turbostat plugin: Invalid TCCActivationTemp '%s'",
1585 value);
1586 return -1;
1587 }
1588 tcc_activation_temp = (unsigned int) tmp_val;
1589 } else {
1590 ERROR("turbostat plugin: Invalid configuration option '%s'",
1591 key);
1592 return -1;
1593 }
1594 return 0;
1595 }
1597 void module_register(void)
1598 {
1599 plugin_register_init(PLUGIN_NAME, turbostat_init);
1600 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1601 }