cd2cf434bc7d5594d2fabaf431b5014095628562
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"
39 #include "common.h"
40 #include "plugin.h"
41 #include "utils_time.h"
43 #include <asm/msr-index.h>
44 #include <stdarg.h>
45 #include <stdio.h>
46 #include <err.h>
47 #include <unistd.h>
48 #include <sys/types.h>
49 #include <sys/wait.h>
50 #include <sys/stat.h>
51 #include <sys/resource.h>
52 #include <fcntl.h>
53 #include <signal.h>
54 #include <stdlib.h>
55 #include <dirent.h>
56 #include <string.h>
57 #include <ctype.h>
58 #include <sched.h>
59 #include <cpuid.h>
60 #include <sys/capability.h>
62 #define PLUGIN_NAME "turbostat"
64 /*
65 * This tool uses the Model-Specific Registers (MSRs) present on Intel processors.
66 * The general description each of these registers, depending on the architecture,
67 * can be found in the IntelĀ® 64 and IA-32 Architectures Software Developer Manual,
68 * Volume 3 Chapter 35.
69 */
71 /*
72 * If set, aperf_mperf_unstable disables a/mperf based stats.
73 * This includes: C0 & C1 states, frequency
74 *
75 * This value is automatically set if mperf or aperf go backward
76 */
77 static _Bool aperf_mperf_unstable;
79 /*
80 * Bitmask of the list of core C states supported by the processor.
81 * Currently supported C-states (by this plugin): 3, 6, 7
82 */
83 static unsigned int do_core_cstate;
85 /*
86 * Bitmask of the list of pacages C states supported by the processor.
87 * Currently supported C-states (by this plugin): 2, 3, 6, 7, 8, 9, 10
88 */
89 static unsigned int do_pkg_cstate;
91 /*
92 * Boolean indicating if the processor supports 'Digital temperature sensor'
93 * This feature enables the monitoring of the temperature of each core
94 *
95 * This feature has two limitations:
96 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
97 * - Temperatures above the tcc_activation_temp are not recorded
98 */
99 static _Bool do_dts;
101 /*
102 * Boolean indicating if the processor supports 'Package thermal management'
103 * This feature allows the monitoring of the temperature of each package
104 *
105 * This feature has two limitations:
106 * - if MSR_IA32_TEMPERATURE_TARGET is not supported, the absolute temperature might be wrong
107 * - Temperatures above the tcc_activation_temp are not recorded
108 */
109 static _Bool do_ptm;
111 /*
112 * Thermal Control Circuit Activation Temperature as configured by the user.
113 * This override the automated detection via MSR_IA32_TEMPERATURE_TARGET
114 * and should only be used if the automated detection fails.
115 */
116 static unsigned int tcc_activation_temp;
118 static unsigned int do_rapl;
119 static double rapl_energy_units;
121 #define RAPL_PKG (1 << 0)
122 /* 0x610 MSR_PKG_POWER_LIMIT */
123 /* 0x611 MSR_PKG_ENERGY_STATUS */
124 #define RAPL_PKG_PERF_STATUS (1 << 1)
125 /* 0x613 MSR_PKG_PERF_STATUS */
126 #define RAPL_PKG_POWER_INFO (1 << 2)
127 /* 0x614 MSR_PKG_POWER_INFO */
129 #define RAPL_DRAM (1 << 3)
130 /* 0x618 MSR_DRAM_POWER_LIMIT */
131 /* 0x619 MSR_DRAM_ENERGY_STATUS */
132 /* 0x61c MSR_DRAM_POWER_INFO */
133 #define RAPL_DRAM_PERF_STATUS (1 << 4)
134 /* 0x61b MSR_DRAM_PERF_STATUS */
136 #define RAPL_CORES (1 << 5)
137 /* 0x638 MSR_PP0_POWER_LIMIT */
138 /* 0x639 MSR_PP0_ENERGY_STATUS */
139 #define RAPL_CORE_POLICY (1 << 6)
140 /* 0x63a MSR_PP0_POLICY */
143 #define RAPL_GFX (1 << 7)
144 /* 0x640 MSR_PP1_POWER_LIMIT */
145 /* 0x641 MSR_PP1_ENERGY_STATUS */
146 /* 0x642 MSR_PP1_POLICY */
147 #define TJMAX_DEFAULT 100
149 static cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_saved_affinity_set;
150 static size_t cpu_present_setsize, cpu_affinity_setsize, cpu_saved_affinity_setsize;
152 static struct thread_data {
153 unsigned long long tsc;
154 unsigned long long aperf;
155 unsigned long long mperf;
156 unsigned long long c1;
157 unsigned int smi_count;
158 unsigned int cpu_id;
159 unsigned int flags;
160 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
161 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
162 } *thread_delta, *thread_even, *thread_odd;
164 static struct core_data {
165 unsigned long long c3;
166 unsigned long long c6;
167 unsigned long long c7;
168 unsigned int core_temp_c;
169 unsigned int core_id;
170 } *core_delta, *core_even, *core_odd;
172 static struct pkg_data {
173 unsigned long long pc2;
174 unsigned long long pc3;
175 unsigned long long pc6;
176 unsigned long long pc7;
177 unsigned long long pc8;
178 unsigned long long pc9;
179 unsigned long long pc10;
180 unsigned int package_id;
181 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
182 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
183 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
184 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
185 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
186 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
187 unsigned int tcc_activation_temp;
188 unsigned int pkg_temp_c;
189 } *package_delta, *package_even, *package_odd;
191 #define DELTA_COUNTERS thread_delta, core_delta, package_delta
192 #define ODD_COUNTERS thread_odd, core_odd, package_odd
193 #define EVEN_COUNTERS thread_even, core_even, package_even
194 static _Bool is_even = 1;
196 static _Bool allocated = 0;
197 static _Bool initialized = 0;
199 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
200 (thread_base + \
201 (pkg_no) * topology.num_cores * topology.num_threads + \
202 (core_no) * topology.num_threads + \
203 (thread_no))
204 #define GET_CORE(core_base, core_no, pkg_no) \
205 (core_base + \
206 (pkg_no) * topology.num_cores + \
207 (core_no))
208 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
210 struct cpu_topology {
211 int package_id;
212 int core_id;
213 _Bool first_core_in_package;
214 _Bool first_thread_in_core;
215 };
217 static struct topology {
218 int max_cpu_id;
219 int num_packages;
220 int num_cores;
221 int num_threads;
222 struct cpu_topology *cpus;
223 } topology;
225 static cdtime_t time_even, time_odd, time_delta;
227 static const char *config_keys[] =
228 {
229 "TCCActivationTemp",
230 };
231 static const int config_keys_num = STATIC_ARRAY_SIZE (config_keys);
233 /*****************************
234 * MSR Manipulation helpers *
235 *****************************/
237 /*
238 * Open a MSR device for reading
239 * Can change the scheduling affinity of the current process if multiple_read is 1
240 */
241 static int __attribute__((warn_unused_result))
242 open_msr(int cpu, _Bool multiple_read)
243 {
244 char pathname[32];
245 int fd;
247 /*
248 * If we need to do multiple read, let's migrate to the CPU
249 * Otherwise, we would lose time calling functions on another CPU
250 *
251 * If we are not yet initialized (cpu_affinity_setsize = 0),
252 * we need to skip this optimisation.
253 */
254 if (multiple_read && cpu_affinity_setsize) {
255 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
256 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
257 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1) {
258 ERROR("Turbostat plugin: Could not migrate to CPU %d", cpu);
259 return -1;
260 }
261 }
263 ssnprintf(pathname, sizeof(pathname), "/dev/cpu/%d/msr", cpu);
264 fd = open(pathname, O_RDONLY);
265 if (fd < 0) {
266 ERROR("Turbostat plugin: failed to open %s", pathname);
267 return -1;
268 }
269 return fd;
270 }
272 /*
273 * Read a single MSR from an open file descriptor
274 */
275 static int __attribute__((warn_unused_result))
276 read_msr(int fd, off_t offset, unsigned long long *msr)
277 {
278 ssize_t retval;
280 retval = pread(fd, msr, sizeof *msr, offset);
282 if (retval != sizeof *msr) {
283 ERROR("Turbostat plugin: MSR offset 0x%llx read failed",
284 (unsigned long long)offset);
285 return -1;
286 }
287 return 0;
288 }
290 /*
291 * Open a MSR device for reading, read the value asked for and close it.
292 * This call will not affect the scheduling affinity of this thread.
293 */
294 static int __attribute__((warn_unused_result))
295 get_msr(int cpu, off_t offset, unsigned long long *msr)
296 {
297 ssize_t retval;
298 int fd;
300 fd = open_msr(cpu, 0);
301 if (fd < 0)
302 return fd;
303 retval = read_msr(fd, offset, msr);
304 close(fd);
305 return retval;
306 }
309 /********************************
310 * Raw data acquisition (1 CPU) *
311 ********************************/
313 /*
314 * Read every data avalaible for a single CPU
315 *
316 * Core data is shared for all threads in one core: extracted only for the first thread
317 * Package data is shared for all core in one package: extracted only for the first thread of the first core
318 *
319 * Side effect: migrates to the targeted CPU
320 */
321 static int __attribute__((warn_unused_result))
322 get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
323 {
324 int cpu = t->cpu_id;
325 unsigned long long msr;
326 int msr_fd;
327 int retval = 0;
329 msr_fd = open_msr(cpu, 1);
330 if (msr_fd < 0)
331 return msr_fd;
333 #define READ_MSR(msr, dst) \
334 do { \
335 if (read_msr(msr_fd, msr, dst)) { \
336 ERROR("Turbostat plugin: Unable to read " #msr); \
337 retval = -1; \
338 goto out; \
339 } \
340 } while (0)
342 READ_MSR(MSR_IA32_TSC, &t->tsc);
344 READ_MSR(MSR_IA32_APERF, &t->aperf);
345 READ_MSR(MSR_IA32_MPERF, &t->mperf);
347 READ_MSR(MSR_SMI_COUNT, &msr);
348 t->smi_count = msr & 0xFFFFFFFF;
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_rapl & RAPL_PKG_PERF_STATUS) {
406 READ_MSR(MSR_PKG_PERF_STATUS, &msr);
407 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
408 }
409 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
410 READ_MSR(MSR_DRAM_PERF_STATUS, &msr);
411 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
412 }
413 if (do_ptm) {
414 READ_MSR(MSR_IA32_PACKAGE_THERM_STATUS, &msr);
415 p->pkg_temp_c = p->tcc_activation_temp - ((msr >> 16) & 0x7F);
416 }
418 out:
419 close(msr_fd);
420 return retval;
421 }
424 /**********************************
425 * Evaluating the changes (1 CPU) *
426 **********************************/
428 /*
429 * Do delta = new - old on 32bits cyclique intergers
430 */
431 #define DELTA_WRAP32(delta, new, old) \
432 if (new > old) { \
433 delta = new - old; \
434 } else { \
435 delta = 0x100000000 + new - old; \
436 }
438 /*
439 * Extract the evolution old->new in delta at a package level
440 * (some are not new-delta, e.g. temperature)
441 */
442 static inline void
443 delta_package(struct pkg_data *delta, const struct pkg_data *new, const struct pkg_data *old)
444 {
445 delta->pc2 = new->pc2 - old->pc2;
446 delta->pc3 = new->pc3 - old->pc3;
447 delta->pc6 = new->pc6 - old->pc6;
448 delta->pc7 = new->pc7 - old->pc7;
449 delta->pc8 = new->pc8 - old->pc8;
450 delta->pc9 = new->pc9 - old->pc9;
451 delta->pc10 = new->pc10 - old->pc10;
452 delta->pkg_temp_c = new->pkg_temp_c;
454 DELTA_WRAP32(delta->energy_pkg, new->energy_pkg, old->energy_pkg);
455 DELTA_WRAP32(delta->energy_cores, new->energy_cores, old->energy_cores);
456 DELTA_WRAP32(delta->energy_gfx, new->energy_gfx, old->energy_gfx);
457 DELTA_WRAP32(delta->energy_dram, new->energy_dram, old->energy_dram);
458 DELTA_WRAP32(delta->rapl_pkg_perf_status, new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
459 DELTA_WRAP32(delta->rapl_dram_perf_status, new->rapl_dram_perf_status, old->rapl_dram_perf_status);
460 }
462 /*
463 * Extract the evolution old->new in delta at a core level
464 * (some are not new-delta, e.g. temperature)
465 */
466 static inline void
467 delta_core(struct core_data *delta, const struct core_data *new, const struct core_data *old)
468 {
469 delta->c3 = new->c3 - old->c3;
470 delta->c6 = new->c6 - old->c6;
471 delta->c7 = new->c7 - old->c7;
472 delta->core_temp_c = new->core_temp_c;
473 }
475 /*
476 * Extract the evolution old->new in delta at a package level
477 * core_delta is required for c1 estimation (tsc - c0 - all core cstates)
478 */
479 static inline int __attribute__((warn_unused_result))
480 delta_thread(struct thread_data *delta, const struct thread_data *new, const struct thread_data *old,
481 const struct core_data *core_delta)
482 {
483 delta->tsc = new->tsc - old->tsc;
485 /* check for TSC < 1 Mcycles over interval */
486 if (delta->tsc < (1000 * 1000)) {
487 WARNING("Turbostat plugin: Insanely slow TSC rate, TSC stops "
488 "in idle? You can disable all c-states by booting with"
489 " 'idle=poll' or just the deep ones with"
490 " 'processor.max_cstate=1'");
491 return -1;
492 }
494 delta->c1 = new->c1 - old->c1;
496 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
497 delta->aperf = new->aperf - old->aperf;
498 delta->mperf = new->mperf - old->mperf;
499 } else {
500 if (!aperf_mperf_unstable) {
501 WARNING("Turbostat plugin: APERF or MPERF went "
502 "backwards. Frequency results do not cover "
503 "the entire interval. Fix this by running "
504 "Linux-2.6.30 or later.");
506 aperf_mperf_unstable = 1;
507 }
508 }
510 /*
511 * As counter collection is not atomic,
512 * it is possible for mperf's non-halted cycles + idle states
513 * to exceed TSC's all cycles: show c1 = 0% in that case.
514 */
515 if ((delta->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > delta->tsc)
516 delta->c1 = 0;
517 else {
518 /* normal case, derive c1 */
519 delta->c1 = delta->tsc - delta->mperf - core_delta->c3
520 - core_delta->c6 - core_delta->c7;
521 }
523 if (delta->mperf == 0) {
524 WARNING("Turbostat plugin: cpu%d MPERF 0!", old->cpu_id);
525 delta->mperf = 1; /* divide by 0 protection */
526 }
528 delta->smi_count = new->smi_count - old->smi_count;
530 return 0;
531 }
533 /**********************************
534 * Submitting the results (1 CPU) *
535 **********************************/
537 /*
538 * Submit one gauge value
539 */
540 static void
541 turbostat_submit (const char *plugin_instance,
542 const char *type, const char *type_instance,
543 gauge_t value)
544 {
545 value_list_t vl = VALUE_LIST_INIT;
546 value_t v;
548 v.gauge = value;
549 vl.values = &v;
550 vl.values_len = 1;
551 sstrncpy (vl.host, hostname_g, sizeof (vl.host));
552 sstrncpy (vl.plugin, PLUGIN_NAME, sizeof (vl.plugin));
553 if (plugin_instance != NULL)
554 sstrncpy (vl.plugin_instance, plugin_instance, sizeof (vl.plugin_instance));
555 sstrncpy (vl.type, type, sizeof (vl.type));
556 if (type_instance != NULL)
557 sstrncpy (vl.type_instance, type_instance, sizeof (vl.type_instance));
559 plugin_dispatch_values (&vl);
560 }
562 /*
563 * Submit every data for a single CPU
564 *
565 * Core data is shared for all threads in one core: submitted only for the first thread
566 * Package data is shared for all core in one package: submitted only for the first thread of the first core
567 */
568 static int
569 submit_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
570 {
571 char name[DATA_MAX_NAME_LEN];
572 double interval_float;
574 interval_float = CDTIME_T_TO_DOUBLE(time_delta);
576 ssnprintf(name, sizeof(name), "cpu%02d", t->cpu_id);
578 if (!aperf_mperf_unstable)
579 turbostat_submit(name, "percent", "c0", 100.0 * t->mperf/t->tsc);
580 if (!aperf_mperf_unstable)
581 turbostat_submit(name, "percent", "c1", 100.0 * t->c1/t->tsc);
583 /* GHz */
584 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
585 turbostat_submit(NULL, "frequency", name, 1.0 * t->tsc / 1000000000 * t->aperf / t->mperf / interval_float);
587 /* SMI */
588 turbostat_submit(NULL, "current", name, t->smi_count);
590 /* submit per-core data only for 1st thread in core */
591 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
592 goto done;
594 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
596 if (do_core_cstate & (1 << 3))
597 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
598 if (do_core_cstate & (1 << 6))
599 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
600 if (do_core_cstate & (1 << 7))
601 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
603 if (do_dts)
604 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
606 /* submit per-package data only for 1st core in package */
607 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
608 goto done;
610 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
612 if (do_ptm)
613 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
615 if (do_pkg_cstate & (1 << 2))
616 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
617 if (do_pkg_cstate & (1 << 3))
618 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
619 if (do_pkg_cstate & (1 << 6))
620 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
621 if (do_pkg_cstate & (1 << 7))
622 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
623 if (do_pkg_cstate & (1 << 8))
624 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
625 if (do_pkg_cstate & (1 << 9))
626 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
627 if (do_pkg_cstate & (1 << 10))
628 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
630 if (do_rapl) {
631 if (do_rapl & RAPL_PKG)
632 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
633 if (do_rapl & RAPL_CORES)
634 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
635 if (do_rapl & RAPL_GFX)
636 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
637 if (do_rapl & RAPL_DRAM)
638 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
639 }
640 done:
641 return 0;
642 }
645 /**********************************
646 * Looping function over all CPUs *
647 **********************************/
649 /*
650 * Check if a given cpu id is in our compiled list of existing CPUs
651 */
652 static int
653 cpu_is_not_present(int cpu)
654 {
655 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
656 }
658 /*
659 * Loop on all CPUs in topological order
660 *
661 * Skip non-present cpus
662 * Return the error code at the first error or 0
663 */
664 static int __attribute__((warn_unused_result))
665 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
666 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
667 {
668 int retval, pkg_no, core_no, thread_no;
670 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
671 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
672 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
673 struct thread_data *t;
674 struct core_data *c;
675 struct pkg_data *p;
677 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
679 if (cpu_is_not_present(t->cpu_id))
680 continue;
682 c = GET_CORE(core_base, core_no, pkg_no);
683 p = GET_PKG(pkg_base, pkg_no);
685 retval = func(t, c, p);
686 if (retval)
687 return retval;
688 }
689 }
690 }
691 return 0;
692 }
694 /*
695 * Dedicated loop: Extract every data evolution for all CPU
696 *
697 * Skip non-present cpus
698 * Return the error code at the first error or 0
699 *
700 * Core data is shared for all threads in one core: extracted only for the first thread
701 * Package data is shared for all core in one package: extracted only for the first thread of the first core
702 */
703 static int __attribute__((warn_unused_result))
704 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,
705 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
706 {
707 int retval, pkg_no, core_no, thread_no;
709 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
710 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
711 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
712 struct thread_data *t_delta;
713 const struct thread_data *t_old, *t_new;
714 struct core_data *c_delta;
716 /* Get correct pointers for threads */
717 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
718 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
719 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
721 /* Skip threads that disappeared */
722 if (cpu_is_not_present(t_delta->cpu_id))
723 continue;
725 /* c_delta is always required for delta_thread */
726 c_delta = GET_CORE(core_delta, core_no, pkg_no);
728 /* calculate core delta only for 1st thread in core */
729 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
730 const struct core_data *c_old, *c_new;
732 c_new = GET_CORE(core_new_base, core_no, pkg_no);
733 c_old = GET_CORE(core_old_base, core_no, pkg_no);
735 delta_core(c_delta, c_new, c_old);
736 }
738 /* Always calculate thread delta */
739 retval = delta_thread(t_delta, t_new, t_old, c_delta);
740 if (retval)
741 return retval;
743 /* calculate package delta only for 1st core in package */
744 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
745 struct pkg_data *p_delta;
746 const struct pkg_data *p_old, *p_new;
748 p_delta = GET_PKG(package_delta, pkg_no);
749 p_new = GET_PKG(pkg_new_base, pkg_no);
750 p_old = GET_PKG(pkg_old_base, pkg_no);
752 delta_package(p_delta, p_new, p_old);
753 }
754 }
755 }
756 }
757 return 0;
758 }
761 /***************
762 * CPU Probing *
763 ***************/
765 /*
766 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
767 * the Thermal Control Circuit (TCC) activates.
768 * This is usually equal to tjMax.
769 *
770 * Older processors do not have this MSR, so there we guess,
771 * but also allow conficuration over-ride with "TCCActivationTemp".
772 *
773 * Several MSR temperature values are in units of degrees-C
774 * below this value, including the Digital Thermal Sensor (DTS),
775 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
776 */
777 static int __attribute__((warn_unused_result))
778 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
779 {
780 unsigned long long msr;
781 unsigned int target_c_local;
783 /* tcc_activation_temp is used only for dts or ptm */
784 if (!(do_dts || do_ptm))
785 return 0;
787 /* this is a per-package concept */
788 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
789 return 0;
791 if (tcc_activation_temp != 0) {
792 p->tcc_activation_temp = tcc_activation_temp;
793 return 0;
794 }
796 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
797 goto guess;
799 target_c_local = (msr >> 16) & 0xFF;
801 if (!target_c_local)
802 goto guess;
804 p->tcc_activation_temp = target_c_local;
806 return 0;
808 guess:
809 p->tcc_activation_temp = TJMAX_DEFAULT;
810 WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
811 " Please use TCCActivationTemp to specify it.",
812 t->cpu_id, p->tcc_activation_temp);
814 return 0;
815 }
817 /*
818 * Identify the functionality of the CPU
819 */
820 static int __attribute__((warn_unused_result))
821 probe_cpu()
822 {
823 unsigned int eax, ebx, ecx, edx, max_level;
824 unsigned int fms, family, model;
826 /* CPUID(0):
827 * - EAX: Maximum Input Value for Basic CPUID Information
828 * - EBX: "Genu" (0x756e6547)
829 * - EDX: "ineI" (0x49656e69)
830 * - ECX: "ntel" (0x6c65746e)
831 */
832 max_level = ebx = ecx = edx = 0;
833 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
834 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
835 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
836 return -1;
837 }
839 /* CPUID(1):
840 * - EAX: Version Information: Type, Family, Model, and Stepping ID
841 * + 4-7: Model ID
842 * + 8-11: Family ID
843 * + 12-13: Processor type
844 * + 16-19: Extended Model ID
845 * + 20-27: Extended Family ID
846 * - EDX: Feature Information:
847 * + 5: Support for MSR read/write operations
848 */
849 fms = ebx = ecx = edx = 0;
850 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
851 family = (fms >> 8) & 0xf;
852 model = (fms >> 4) & 0xf;
853 if (family == 0xf)
854 family += (fms >> 20) & 0xf;
855 if (family == 6 || family == 0xf)
856 model += ((fms >> 16) & 0xf) << 4;
857 if (!(edx & (1 << 5))) {
858 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
859 return -1;
860 }
862 /*
863 * CPUID(0x80000000):
864 * - EAX: Maximum Input Value for Extended Function CPUID Information
865 *
866 * This allows us to verify if the CPUID(0x80000007) can be called
867 *
868 * This check is valid for both Intel and AMD.
869 */
870 max_level = ebx = ecx = edx = 0;
871 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
872 if (max_level < 0x80000007) {
873 ERROR("Turbostat plugin: Unsupported CPU (no invariant TSC, "
874 " Maximum Extended Function: 0x%x)", max_level);
875 return -1;
876 }
878 /*
879 * CPUID(0x80000007):
880 * - EDX:
881 * + 8: Invariant TSC available if set
882 *
883 * This check is valid for both Intel and AMD
884 */
885 eax = ebx = ecx = edx = 0;
886 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
887 if (!(edx & (1 << 8))) {
888 ERROR("Turbostat plugin: Unsupported CPU (No invariant TSC)");
889 return -1;
890 }
892 /*
893 * CPUID(6):
894 * - EAX:
895 * + 0: Digital temperature sensor is supported if set
896 * + 6: Package thermal management is supported if set
897 * - ECX:
898 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
899 * + 3: The processor supports performance-energy bias preference if set.
900 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
901 *
902 * This check is valid for both Intel and AMD
903 */
904 eax = ebx = ecx = edx = 0;
905 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
906 do_dts = eax & (1 << 0);
907 do_ptm = eax & (1 << 6);
908 if (!(ecx & (1 << 0))) {
909 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
910 return -1;
911 }
913 /*
914 * Enable or disable C states depending on the model and family
915 */
916 if (family == 6) {
917 switch (model) {
918 /* Atom (partial) */
919 case 0x27:
920 do_core_cstate = 0;
921 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
922 break;
923 /* Silvermont */
924 case 0x37: /* BYT */
925 case 0x4A:
926 case 0x4D: /* AVN */
927 case 0x5A:
928 case 0x5D:
929 do_core_cstate = (1 << 1) | (1 << 6);
930 do_pkg_cstate = (1 << 6);
931 break;
932 /* Nehalem */
933 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
934 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
935 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
936 case 0x2E: /* Nehalem-EX Xeon - Beckton */
937 do_core_cstate = (1 << 3) | (1 << 6);
938 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
939 break;
940 /* Westmere */
941 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
942 case 0x2C: /* Westmere EP - Gulftown */
943 case 0x2F: /* Westmere-EX Xeon - Eagleton */
944 do_core_cstate = (1 << 3) | (1 << 6);
945 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
946 break;
947 /* Sandy Bridge */
948 case 0x2A: /* SNB */
949 case 0x2D: /* SNB Xeon */
950 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
951 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
952 break;
953 /* Ivy Bridge */
954 case 0x3A: /* IVB */
955 case 0x3E: /* IVB Xeon */
956 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
957 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
958 break;
959 /* Haswell Bridge */
960 case 0x3C: /* HSW */
961 case 0x3F: /* HSW */
962 case 0x46: /* HSW */
963 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
964 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
965 break;
966 case 0x45: /* HSW */
967 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
968 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
969 break;
970 /* Broadwel */
971 case 0x4F: /* BDW */
972 case 0x56: /* BDX-DE */
973 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
974 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
975 break;
976 case 0x3D: /* BDW */
977 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
978 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
979 break;
980 default:
981 ERROR("Turbostat plugin: Unsupported CPU (family: %#x,"
982 " model: %#x)", family, model);
983 }
984 switch (model) {
985 case 0x2A:
986 case 0x3A:
987 case 0x3C:
988 case 0x45:
989 case 0x46:
990 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
991 break;
992 case 0x3F:
993 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
994 break;
995 case 0x2D:
996 case 0x3E:
997 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
998 break;
999 case 0x37:
1000 case 0x4D:
1001 do_rapl = RAPL_PKG | RAPL_CORES;
1002 break;
1003 default:
1004 do_rapl = 0;
1005 }
1006 } else {
1007 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
1008 "model: %#x)", family, model);
1009 return -1;
1010 }
1012 if (do_rapl) {
1013 unsigned long long msr;
1014 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1015 return 0;
1017 if (model == 0x37)
1018 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1019 else
1020 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1021 }
1023 return 0;
1024 }
1027 /********************
1028 * Topology Probing *
1029 ********************/
1031 /*
1032 * Read a single int from a file.
1033 */
1034 static int __attribute__ ((format(printf,1,2)))
1035 parse_int_file(const char *fmt, ...)
1036 {
1037 va_list args;
1038 char path[PATH_MAX];
1039 FILE *filep;
1040 int len, value;
1042 va_start(args, fmt);
1043 len = vsnprintf(path, sizeof(path), fmt, args);
1044 va_end(args);
1045 if (len < 0 || len >= PATH_MAX) {
1046 ERROR("Turbostat plugin: path truncated: '%s'", path);
1047 return -1;
1048 }
1050 filep = fopen(path, "r");
1051 if (!filep) {
1052 ERROR("Turbostat plugin: Failed to open '%s'", path);
1053 return -1;
1054 }
1055 if (fscanf(filep, "%d", &value) != 1) {
1056 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1057 return -1;
1058 }
1059 fclose(filep);
1060 return value;
1061 }
1063 static int
1064 get_threads_on_core(int cpu)
1065 {
1066 char path[80];
1067 FILE *filep;
1068 int sib1, sib2;
1069 int matches;
1070 char character;
1072 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1073 filep = fopen(path, "r");
1074 if (!filep) {
1075 ERROR("Turbostat plugin: Failed to open '%s'", path);
1076 return -1;
1077 }
1078 /*
1079 * file format:
1080 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1081 * otherwinse 1 sibling (self).
1082 */
1083 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1085 fclose(filep);
1087 if (matches == 3)
1088 return 2;
1089 else
1090 return 1;
1091 }
1093 /*
1094 * run func(cpu) on every cpu in /proc/stat
1095 * return max_cpu number
1096 */
1097 static int __attribute__((warn_unused_result))
1098 for_all_proc_cpus(int (func)(int))
1099 {
1100 FILE *fp;
1101 int cpu_num;
1102 int retval;
1104 fp = fopen("/proc/stat", "r");
1105 if (!fp) {
1106 ERROR("Turbostat plugin: Failed to open /proc/stat");
1107 return -1;
1108 }
1110 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1111 if (retval != 0) {
1112 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1113 fclose(fp);
1114 return -1;
1115 }
1117 while (1) {
1118 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1119 if (retval != 1)
1120 break;
1122 retval = func(cpu_num);
1123 if (retval) {
1124 fclose(fp);
1125 return(retval);
1126 }
1127 }
1128 fclose(fp);
1129 return 0;
1130 }
1132 /*
1133 * Update the stored topology.max_cpu_id
1134 */
1135 static int
1136 update_max_cpu_id(int cpu)
1137 {
1138 if (topology.max_cpu_id < cpu)
1139 topology.max_cpu_id = cpu;
1140 return 0;
1141 }
1143 static int
1144 mark_cpu_present(int cpu)
1145 {
1146 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1147 return 0;
1148 }
1150 static int __attribute__((warn_unused_result))
1151 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1152 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1153 if (*set == NULL) {
1154 ERROR("Turbostat plugin: Unable to allocate CPU state");
1155 return -1;
1156 }
1157 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1158 CPU_ZERO_S(*size, *set);
1159 return 0;
1160 }
1162 /*
1163 * Build a local representation of the cpu distribution
1164 */
1165 static int __attribute__((warn_unused_result))
1166 topology_probe()
1167 {
1168 int i;
1169 int ret;
1170 int max_package_id, max_core_id, max_thread_id;
1171 max_package_id = max_core_id = max_thread_id = 0;
1173 /* Clean topology */
1174 free(topology.cpus);
1175 memset(&topology, 0, sizeof(topology));
1177 ret = for_all_proc_cpus(update_max_cpu_id);
1178 if (ret != 0)
1179 goto err;
1181 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1182 if (topology.cpus == NULL) {
1183 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1184 return -1;
1185 }
1187 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1188 if (ret != 0)
1189 goto err;
1190 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1191 if (ret != 0)
1192 goto err;
1193 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1194 if (ret != 0)
1195 goto err;
1197 ret = for_all_proc_cpus(mark_cpu_present);
1198 if (ret != 0)
1199 goto err;
1201 /*
1202 * For online cpus
1203 * find max_core_id, max_package_id
1204 */
1205 for (i = 0; i <= topology.max_cpu_id; ++i) {
1206 int num_threads;
1207 struct cpu_topology *cpu = &topology.cpus[i];
1209 if (cpu_is_not_present(i)) {
1210 WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1211 continue;
1212 }
1214 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1215 if (ret < 0)
1216 goto err;
1217 else
1218 cpu->package_id = ret;
1219 if (cpu->package_id > max_package_id)
1220 max_package_id = cpu->package_id;
1222 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1223 if (ret < 0)
1224 goto err;
1225 else
1226 cpu->core_id = ret;
1227 if (cpu->core_id > max_core_id)
1228 max_core_id = cpu->core_id;
1229 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1230 if (ret < 0)
1231 goto err;
1232 else if (ret == i)
1233 cpu->first_core_in_package = 1;
1235 ret = get_threads_on_core(i);
1236 if (ret < 0)
1237 goto err;
1238 else
1239 num_threads = ret;
1240 if (num_threads > max_thread_id)
1241 max_thread_id = num_threads;
1242 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1243 if (ret < 0)
1244 goto err;
1245 else if (ret == i)
1246 cpu->first_thread_in_core = 1;
1248 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1249 i, cpu->package_id, cpu->core_id);
1250 }
1251 /* Num is max + 1 (need to count 0) */
1252 topology.num_packages = max_package_id + 1;
1253 topology.num_cores = max_core_id + 1;
1254 topology.num_threads = max_thread_id + 1;
1256 return 0;
1257 err:
1258 free(topology.cpus);
1259 return ret;
1260 }
1263 /************************
1264 * Main alloc/init/free *
1265 ************************/
1267 static int
1268 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1269 {
1270 int i;
1271 int total_threads, total_cores;
1273 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1274 *threads = calloc(total_threads, sizeof(struct thread_data));
1275 if (*threads == NULL)
1276 goto err;
1278 for (i = 0; i < total_threads; ++i)
1279 (*threads)[i].cpu_id = -1;
1281 total_cores = topology.num_cores * topology.num_packages;
1282 *cores = calloc(total_cores, sizeof(struct core_data));
1283 if (*cores == NULL)
1284 goto err_clean_threads;
1286 for (i = 0; i < total_cores; ++i)
1287 (*cores)[i].core_id = -1;
1289 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1290 if (*packages == NULL)
1291 goto err_clean_cores;
1293 for (i = 0; i < topology.num_packages; i++)
1294 (*packages)[i].package_id = i;
1296 return 0;
1298 err_clean_cores:
1299 free(*cores);
1300 err_clean_threads:
1301 free(*threads);
1302 err:
1303 ERROR("Turbostat plugin: Failled to allocate memory for counters");
1304 return -1;
1305 }
1307 static void
1308 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1309 struct pkg_data *pkg_base, int cpu_id)
1310 {
1311 struct thread_data *t;
1312 struct core_data *c;
1313 struct pkg_data *p;
1314 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1316 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1317 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1318 p = GET_PKG(pkg_base, cpu->package_id);
1320 t->cpu_id = cpu_id;
1321 if (cpu->first_thread_in_core)
1322 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1323 if (cpu->first_core_in_package)
1324 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1326 c->core_id = cpu->core_id;
1327 p->package_id = cpu->package_id;
1328 }
1330 static void
1331 initialize_counters(void)
1332 {
1333 int cpu_id;
1335 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1336 if (cpu_is_not_present(cpu_id))
1337 continue;
1338 init_counter(EVEN_COUNTERS, cpu_id);
1339 init_counter(ODD_COUNTERS, cpu_id);
1340 init_counter(DELTA_COUNTERS, cpu_id);
1341 }
1342 }
1346 static void
1347 free_all_buffers(void)
1348 {
1349 allocated = 0;
1350 initialized = 0;
1352 CPU_FREE(cpu_present_set);
1353 cpu_present_set = NULL;
1354 cpu_present_set = 0;
1356 CPU_FREE(cpu_affinity_set);
1357 cpu_affinity_set = NULL;
1358 cpu_affinity_setsize = 0;
1360 CPU_FREE(cpu_saved_affinity_set);
1361 cpu_saved_affinity_set = NULL;
1362 cpu_saved_affinity_setsize = 0;
1364 free(thread_even);
1365 free(core_even);
1366 free(package_even);
1368 thread_even = NULL;
1369 core_even = NULL;
1370 package_even = NULL;
1372 free(thread_odd);
1373 free(core_odd);
1374 free(package_odd);
1376 thread_odd = NULL;
1377 core_odd = NULL;
1378 package_odd = NULL;
1380 free(thread_delta);
1381 free(core_delta);
1382 free(package_delta);
1384 thread_delta = NULL;
1385 core_delta = NULL;
1386 package_delta = NULL;
1387 }
1390 /**********************
1391 * Collectd functions *
1392 **********************/
1394 #define DO_OR_GOTO_ERR(something) \
1395 do { \
1396 ret = (something); \
1397 if (ret < 0) \
1398 goto err; \
1399 } while (0)
1401 static int setup_all_buffers(void)
1402 {
1403 int ret;
1405 DO_OR_GOTO_ERR(topology_probe());
1406 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1407 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1408 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1409 initialize_counters();
1410 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1411 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1413 allocated = 1;
1414 return 0;
1415 err:
1416 free_all_buffers();
1417 return ret;
1418 }
1420 static int
1421 turbostat_read(void)
1422 {
1423 int ret;
1425 if (!allocated) {
1426 if ((ret = setup_all_buffers()) < 0)
1427 return ret;
1428 }
1430 if (for_all_proc_cpus(cpu_is_not_present)) {
1431 free_all_buffers();
1432 if ((ret = setup_all_buffers()) < 0)
1433 return ret;
1434 if (for_all_proc_cpus(cpu_is_not_present)) {
1435 ERROR("Turbostat plugin: CPU appeared just after "
1436 "initialization");
1437 return -1;
1438 }
1439 }
1441 /* Saving the scheduling affinity, as it will be modified by get_counters */
1442 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1443 ERROR("Turbostat plugin: Unable to save the CPU affinity");
1444 return -1;
1445 }
1447 if (!initialized) {
1448 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1449 goto out;
1450 time_even = cdtime();
1451 is_even = 1;
1452 initialized = 1;
1453 ret = 0;
1454 goto out;
1455 }
1457 if (is_even) {
1458 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1459 goto out;
1460 time_odd = cdtime();
1461 is_even = 0;
1462 time_delta = time_odd - time_even;
1463 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1464 goto out;
1465 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1466 goto out;
1467 } else {
1468 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1469 goto out;
1470 time_even = cdtime();
1471 is_even = 1;
1472 time_delta = time_even - time_odd;
1473 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1474 goto out;
1475 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1476 goto out;
1477 }
1478 ret = 0;
1479 out:
1480 /*
1481 * Let's restore the affinity
1482 * This might fail if the number of CPU changed, but we can't do anything in that case..
1483 */
1484 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1485 return ret;
1486 }
1488 static int
1489 check_permissions(void)
1490 {
1491 struct __user_cap_header_struct cap_header_data;
1492 cap_user_header_t cap_header = &cap_header_data;
1493 struct __user_cap_data_struct cap_data_data;
1494 cap_user_data_t cap_data = &cap_data_data;
1495 int ret = 0;
1497 if (getuid() == 0) {
1498 /* We have everything we need */
1499 return 0;
1500 }
1502 /* check for CAP_SYS_RAWIO */
1503 cap_header->pid = getpid();
1504 cap_header->version = _LINUX_CAPABILITY_VERSION;
1505 if (capget(cap_header, cap_data) < 0) {
1506 ERROR("Turbostat plugin: capget failed");
1507 return -1;
1508 }
1510 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1511 WARNING("Turbostat plugin: Collectd doesn't have the "
1512 "CAP_SYS_RAWIO capability. If you don't want to run "
1513 "collectd as root, try running \"setcap "
1514 "cap_sys_rawio=ep\" on collectd binary");
1515 ret = -1;
1516 }
1518 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1519 WARNING("Turbostat plugin: Collectd cannot open"
1520 "/dev/cpu/0/msr. If you don't want to run collectd as "
1521 "root, you need to change the ownership (chown) and "
1522 "permissions on /dev/cpu/*/msr to allow such access");
1523 ret = -1;
1524 }
1526 if (ret != 0)
1527 ERROR("Turbostat plugin: Initialization failed: this plugin "
1528 "requires collectd to either to run as root or give "
1529 "collectd a special capability (CAP_SYS_RAWIO) and read "
1530 "access to /dev/cpu/*/msr (see previous warnings)");
1531 return ret;
1532 }
1534 static int
1535 turbostat_init(void)
1536 {
1537 struct stat sb;
1538 int ret;
1540 if (stat("/dev/cpu/0/msr", &sb)) {
1541 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1542 " does not exist while the CPU supports MSR. You may be "
1543 "missing the corresponding kernel module, please try '# "
1544 "modprobe msr'");
1545 return -1;
1546 }
1548 DO_OR_GOTO_ERR(check_permissions());
1550 DO_OR_GOTO_ERR(probe_cpu());
1552 DO_OR_GOTO_ERR(setup_all_buffers());
1554 plugin_register_read(PLUGIN_NAME, turbostat_read);
1556 return 0;
1557 err:
1558 free_all_buffers();
1559 return ret;
1560 }
1562 static int
1563 turbostat_config(const char *key, const char *value)
1564 {
1565 long unsigned int tmp_val;
1566 char *end;
1568 if (strcasecmp("TCCActivationTemp", key) == 0) {
1569 tmp_val = strtoul(value, &end, 0);
1570 if (*end != '\0' || tmp_val > UINT_MAX) {
1571 ERROR("Turbostat plugin: Invalid TCCActivationTemp '%s'",
1572 value);
1573 return -1;
1574 }
1575 tcc_activation_temp = (unsigned int) tmp_val;
1576 } else {
1577 ERROR("Turbostat plugin: Invalid configuration option '%s'",
1578 key);
1579 return -1;
1580 }
1581 return 0;
1582 }
1584 void module_register(void)
1585 {
1586 plugin_register_init(PLUGIN_NAME, turbostat_init);
1587 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1588 }