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 turbostat_submit("Average", "frequency", name, 1.0 / 1000000 * t->aperf / interval_float);
585 if ((!aperf_mperf_unstable) || (!(t->aperf > t->tsc || t->mperf > t->tsc)))
586 turbostat_submit("Buzy", "frequency", name, 1.0 * t->tsc / 1000000 * t->aperf / t->mperf / interval_float);
588 /* Sanity check (should stay stable) */
589 turbostat_submit("TSC", "gauge", name, 1.0 * t->tsc / 1000000 / interval_float);
591 /* SMI */
592 turbostat_submit(NULL, "current", name, t->smi_count);
594 /* submit per-core data only for 1st thread in core */
595 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
596 goto done;
598 ssnprintf(name, sizeof(name), "core%02d", c->core_id);
600 if (do_core_cstate & (1 << 3))
601 turbostat_submit(name, "percent", "c3", 100.0 * c->c3/t->tsc);
602 if (do_core_cstate & (1 << 6))
603 turbostat_submit(name, "percent", "c6", 100.0 * c->c6/t->tsc);
604 if (do_core_cstate & (1 << 7))
605 turbostat_submit(name, "percent", "c7", 100.0 * c->c7/t->tsc);
607 if (do_dts)
608 turbostat_submit(NULL, "temperature", name, c->core_temp_c);
610 /* submit per-package data only for 1st core in package */
611 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
612 goto done;
614 ssnprintf(name, sizeof(name), "pkg%02d", p->package_id);
616 if (do_ptm)
617 turbostat_submit(NULL, "temperature", name, p->pkg_temp_c);
619 if (do_pkg_cstate & (1 << 2))
620 turbostat_submit(name, "percent", "pc2", 100.0 * p->pc2/t->tsc);
621 if (do_pkg_cstate & (1 << 3))
622 turbostat_submit(name, "percent", "pc3", 100.0 * p->pc3/t->tsc);
623 if (do_pkg_cstate & (1 << 6))
624 turbostat_submit(name, "percent", "pc6", 100.0 * p->pc6/t->tsc);
625 if (do_pkg_cstate & (1 << 7))
626 turbostat_submit(name, "percent", "pc7", 100.0 * p->pc7/t->tsc);
627 if (do_pkg_cstate & (1 << 8))
628 turbostat_submit(name, "percent", "pc8", 100.0 * p->pc8/t->tsc);
629 if (do_pkg_cstate & (1 << 9))
630 turbostat_submit(name, "percent", "pc9", 100.0 * p->pc9/t->tsc);
631 if (do_pkg_cstate & (1 << 10))
632 turbostat_submit(name, "percent", "pc10", 100.0 * p->pc10/t->tsc);
634 if (do_rapl) {
635 if (do_rapl & RAPL_PKG)
636 turbostat_submit(name, "power", "Pkg_W", p->energy_pkg * rapl_energy_units / interval_float);
637 if (do_rapl & RAPL_CORES)
638 turbostat_submit(name, "power", "Cor_W", p->energy_cores * rapl_energy_units / interval_float);
639 if (do_rapl & RAPL_GFX)
640 turbostat_submit(name, "power", "GFX_W", p->energy_gfx * rapl_energy_units / interval_float);
641 if (do_rapl & RAPL_DRAM)
642 turbostat_submit(name, "power", "RAM_W", p->energy_dram * rapl_energy_units / interval_float);
643 }
644 done:
645 return 0;
646 }
649 /**********************************
650 * Looping function over all CPUs *
651 **********************************/
653 /*
654 * Check if a given cpu id is in our compiled list of existing CPUs
655 */
656 static int
657 cpu_is_not_present(int cpu)
658 {
659 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
660 }
662 /*
663 * Loop on all CPUs in topological order
664 *
665 * Skip non-present cpus
666 * Return the error code at the first error or 0
667 */
668 static int __attribute__((warn_unused_result))
669 for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
670 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
671 {
672 int retval, pkg_no, core_no, thread_no;
674 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
675 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
676 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
677 struct thread_data *t;
678 struct core_data *c;
679 struct pkg_data *p;
681 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
683 if (cpu_is_not_present(t->cpu_id))
684 continue;
686 c = GET_CORE(core_base, core_no, pkg_no);
687 p = GET_PKG(pkg_base, pkg_no);
689 retval = func(t, c, p);
690 if (retval)
691 return retval;
692 }
693 }
694 }
695 return 0;
696 }
698 /*
699 * Dedicated loop: Extract every data evolution for all CPU
700 *
701 * Skip non-present cpus
702 * Return the error code at the first error or 0
703 *
704 * Core data is shared for all threads in one core: extracted only for the first thread
705 * Package data is shared for all core in one package: extracted only for the first thread of the first core
706 */
707 static int __attribute__((warn_unused_result))
708 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,
709 const struct thread_data *thread_old_base, const struct core_data *core_old_base, const struct pkg_data *pkg_old_base)
710 {
711 int retval, pkg_no, core_no, thread_no;
713 for (pkg_no = 0; pkg_no < topology.num_packages; ++pkg_no) {
714 for (core_no = 0; core_no < topology.num_cores; ++core_no) {
715 for (thread_no = 0; thread_no < topology.num_threads; ++thread_no) {
716 struct thread_data *t_delta;
717 const struct thread_data *t_old, *t_new;
718 struct core_data *c_delta;
720 /* Get correct pointers for threads */
721 t_delta = GET_THREAD(thread_delta, thread_no, core_no, pkg_no);
722 t_new = GET_THREAD(thread_new_base, thread_no, core_no, pkg_no);
723 t_old = GET_THREAD(thread_old_base, thread_no, core_no, pkg_no);
725 /* Skip threads that disappeared */
726 if (cpu_is_not_present(t_delta->cpu_id))
727 continue;
729 /* c_delta is always required for delta_thread */
730 c_delta = GET_CORE(core_delta, core_no, pkg_no);
732 /* calculate core delta only for 1st thread in core */
733 if (t_new->flags & CPU_IS_FIRST_THREAD_IN_CORE) {
734 const struct core_data *c_old, *c_new;
736 c_new = GET_CORE(core_new_base, core_no, pkg_no);
737 c_old = GET_CORE(core_old_base, core_no, pkg_no);
739 delta_core(c_delta, c_new, c_old);
740 }
742 /* Always calculate thread delta */
743 retval = delta_thread(t_delta, t_new, t_old, c_delta);
744 if (retval)
745 return retval;
747 /* calculate package delta only for 1st core in package */
748 if (t_new->flags & CPU_IS_FIRST_CORE_IN_PACKAGE) {
749 struct pkg_data *p_delta;
750 const struct pkg_data *p_old, *p_new;
752 p_delta = GET_PKG(package_delta, pkg_no);
753 p_new = GET_PKG(pkg_new_base, pkg_no);
754 p_old = GET_PKG(pkg_old_base, pkg_no);
756 delta_package(p_delta, p_new, p_old);
757 }
758 }
759 }
760 }
761 return 0;
762 }
765 /***************
766 * CPU Probing *
767 ***************/
769 /*
770 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
771 * the Thermal Control Circuit (TCC) activates.
772 * This is usually equal to tjMax.
773 *
774 * Older processors do not have this MSR, so there we guess,
775 * but also allow conficuration over-ride with "TCCActivationTemp".
776 *
777 * Several MSR temperature values are in units of degrees-C
778 * below this value, including the Digital Thermal Sensor (DTS),
779 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
780 */
781 static int __attribute__((warn_unused_result))
782 set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
783 {
784 unsigned long long msr;
785 unsigned int target_c_local;
787 /* tcc_activation_temp is used only for dts or ptm */
788 if (!(do_dts || do_ptm))
789 return 0;
791 /* this is a per-package concept */
792 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
793 return 0;
795 if (tcc_activation_temp != 0) {
796 p->tcc_activation_temp = tcc_activation_temp;
797 return 0;
798 }
800 if (get_msr(t->cpu_id, MSR_IA32_TEMPERATURE_TARGET, &msr))
801 goto guess;
803 target_c_local = (msr >> 16) & 0xFF;
805 if (!target_c_local)
806 goto guess;
808 p->tcc_activation_temp = target_c_local;
810 return 0;
812 guess:
813 p->tcc_activation_temp = TJMAX_DEFAULT;
814 WARNING("Turbostat plugin: cpu%d: Guessing tjMax %d C,"
815 " Please use TCCActivationTemp to specify it.",
816 t->cpu_id, p->tcc_activation_temp);
818 return 0;
819 }
821 /*
822 * Identify the functionality of the CPU
823 */
824 static int __attribute__((warn_unused_result))
825 probe_cpu()
826 {
827 unsigned int eax, ebx, ecx, edx, max_level;
828 unsigned int fms, family, model;
830 /* CPUID(0):
831 * - EAX: Maximum Input Value for Basic CPUID Information
832 * - EBX: "Genu" (0x756e6547)
833 * - EDX: "ineI" (0x49656e69)
834 * - ECX: "ntel" (0x6c65746e)
835 */
836 max_level = ebx = ecx = edx = 0;
837 __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
838 if (ebx != 0x756e6547 && edx != 0x49656e69 && ecx != 0x6c65746e) {
839 ERROR("Turbostat plugin: Unsupported CPU (not Intel)");
840 return -1;
841 }
843 /* CPUID(1):
844 * - EAX: Version Information: Type, Family, Model, and Stepping ID
845 * + 4-7: Model ID
846 * + 8-11: Family ID
847 * + 12-13: Processor type
848 * + 16-19: Extended Model ID
849 * + 20-27: Extended Family ID
850 * - EDX: Feature Information:
851 * + 5: Support for MSR read/write operations
852 */
853 fms = ebx = ecx = edx = 0;
854 __get_cpuid(1, &fms, &ebx, &ecx, &edx);
855 family = (fms >> 8) & 0xf;
856 model = (fms >> 4) & 0xf;
857 if (family == 0xf)
858 family += (fms >> 20) & 0xf;
859 if (family == 6 || family == 0xf)
860 model += ((fms >> 16) & 0xf) << 4;
861 if (!(edx & (1 << 5))) {
862 ERROR("Turbostat plugin: Unsupported CPU (no MSR support)");
863 return -1;
864 }
866 /*
867 * CPUID(0x80000000):
868 * - EAX: Maximum Input Value for Extended Function CPUID Information
869 *
870 * This allows us to verify if the CPUID(0x80000007) can be called
871 *
872 * This check is valid for both Intel and AMD.
873 */
874 max_level = ebx = ecx = edx = 0;
875 __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
876 if (max_level < 0x80000007) {
877 ERROR("Turbostat plugin: Unsupported CPU (no invariant TSC, "
878 " Maximum Extended Function: 0x%x)", max_level);
879 return -1;
880 }
882 /*
883 * CPUID(0x80000007):
884 * - EDX:
885 * + 8: Invariant TSC available if set
886 *
887 * This check is valid for both Intel and AMD
888 */
889 eax = ebx = ecx = edx = 0;
890 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
891 if (!(edx & (1 << 8))) {
892 ERROR("Turbostat plugin: Unsupported CPU (No invariant TSC)");
893 return -1;
894 }
896 /*
897 * CPUID(6):
898 * - EAX:
899 * + 0: Digital temperature sensor is supported if set
900 * + 6: Package thermal management is supported if set
901 * - ECX:
902 * + 0: Hardware Coordination Feedback Capability (Presence of IA32_MPERF and IA32_APERF).
903 * + 3: The processor supports performance-energy bias preference if set.
904 * It also implies the presence of a new architectural MSR called IA32_ENERGY_PERF_BIAS
905 *
906 * This check is valid for both Intel and AMD
907 */
908 eax = ebx = ecx = edx = 0;
909 __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
910 do_dts = eax & (1 << 0);
911 do_ptm = eax & (1 << 6);
912 if (!(ecx & (1 << 0))) {
913 ERROR("Turbostat plugin: Unsupported CPU (No APERF)");
914 return -1;
915 }
917 /*
918 * Enable or disable C states depending on the model and family
919 */
920 if (family == 6) {
921 switch (model) {
922 /* Atom (partial) */
923 case 0x27:
924 do_core_cstate = 0;
925 do_pkg_cstate = (1 << 2) | (1 << 4) | (1 << 6);
926 break;
927 /* Silvermont */
928 case 0x37: /* BYT */
929 case 0x4D: /* AVN */
930 do_core_cstate = (1 << 1) | (1 << 6);
931 do_pkg_cstate = (1 << 6);
932 break;
933 /* Nehalem */
934 case 0x1A: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
935 case 0x1E: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
936 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
937 case 0x2E: /* Nehalem-EX Xeon - Beckton */
938 do_core_cstate = (1 << 3) | (1 << 6);
939 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
940 break;
941 /* Westmere */
942 case 0x25: /* Westmere Client - Clarkdale, Arrandale */
943 case 0x2C: /* Westmere EP - Gulftown */
944 case 0x2F: /* Westmere-EX Xeon - Eagleton */
945 do_core_cstate = (1 << 3) | (1 << 6);
946 do_pkg_cstate = (1 << 3) | (1 << 6) | (1 << 7);
947 break;
948 /* Sandy Bridge */
949 case 0x2A: /* SNB */
950 case 0x2D: /* SNB Xeon */
951 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
952 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
953 break;
954 /* Ivy Bridge */
955 case 0x3A: /* IVB */
956 case 0x3E: /* IVB Xeon */
957 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
958 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
959 break;
960 /* Haswell Bridge */
961 case 0x3C: /* HSW */
962 case 0x3F: /* HSW */
963 case 0x46: /* HSW */
964 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
965 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
966 break;
967 case 0x45: /* HSW */
968 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
969 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
970 break;
971 /* Broadwel */
972 case 0x4F: /* BDW */
973 case 0x56: /* BDX-DE */
974 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
975 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7);
976 break;
977 case 0x3D: /* BDW */
978 do_core_cstate = (1 << 3) | (1 << 6) | (1 << 7);
979 do_pkg_cstate = (1 << 2) | (1 << 3) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 9) | (1 << 10);
980 break;
981 default:
982 ERROR("Turbostat plugin: Unsupported CPU (family: %#x,"
983 " model: %#x)", family, model);
984 }
985 switch (model) {
986 case 0x2A: /* SNB */
987 case 0x3A: /* IVB */
988 case 0x3C: /* HSW */
989 case 0x45: /* HSW */
990 case 0x46: /* HSW */
991 case 0x3D: /* BDW */
992 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_GFX;
993 break;
994 case 0x3F: /* HSX */
995 case 0x4F: /* BDX */
996 case 0x56: /* BDX-DE */
997 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
998 break;
999 case 0x2D: /* SNB Xeon */
1000 case 0x3E: /* IVB Xeon */
1001 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_PKG_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM | RAPL_DRAM_PERF_STATUS;
1002 break;
1003 case 0x37: /* BYT */
1004 case 0x4D: /* AVN */
1005 do_rapl = RAPL_PKG | RAPL_CORES;
1006 break;
1007 default:
1008 do_rapl = 0;
1009 }
1010 } else {
1011 ERROR("Turbostat plugin: Unsupported CPU (family: %#x, "
1012 "model: %#x)", family, model);
1013 return -1;
1014 }
1016 if (do_rapl) {
1017 unsigned long long msr;
1018 if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1019 return 0;
1021 if (model == 0x37)
1022 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1023 else
1024 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1025 }
1027 return 0;
1028 }
1031 /********************
1032 * Topology Probing *
1033 ********************/
1035 /*
1036 * Read a single int from a file.
1037 */
1038 static int __attribute__ ((format(printf,1,2)))
1039 parse_int_file(const char *fmt, ...)
1040 {
1041 va_list args;
1042 char path[PATH_MAX];
1043 FILE *filep;
1044 int len, value;
1046 va_start(args, fmt);
1047 len = vsnprintf(path, sizeof(path), fmt, args);
1048 va_end(args);
1049 if (len < 0 || len >= PATH_MAX) {
1050 ERROR("Turbostat plugin: path truncated: '%s'", path);
1051 return -1;
1052 }
1054 filep = fopen(path, "r");
1055 if (!filep) {
1056 ERROR("Turbostat plugin: Failed to open '%s'", path);
1057 return -1;
1058 }
1059 if (fscanf(filep, "%d", &value) != 1) {
1060 ERROR("Turbostat plugin: Failed to parse number from '%s'", path);
1061 return -1;
1062 }
1063 fclose(filep);
1064 return value;
1065 }
1067 static int
1068 get_threads_on_core(int cpu)
1069 {
1070 char path[80];
1071 FILE *filep;
1072 int sib1, sib2;
1073 int matches;
1074 char character;
1076 ssnprintf(path, sizeof(path), "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1077 filep = fopen(path, "r");
1078 if (!filep) {
1079 ERROR("Turbostat plugin: Failed to open '%s'", path);
1080 return -1;
1081 }
1082 /*
1083 * file format:
1084 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1085 * otherwinse 1 sibling (self).
1086 */
1087 matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1089 fclose(filep);
1091 if (matches == 3)
1092 return 2;
1093 else
1094 return 1;
1095 }
1097 /*
1098 * run func(cpu) on every cpu in /proc/stat
1099 * return max_cpu number
1100 */
1101 static int __attribute__((warn_unused_result))
1102 for_all_proc_cpus(int (func)(int))
1103 {
1104 FILE *fp;
1105 int cpu_num;
1106 int retval;
1108 fp = fopen("/proc/stat", "r");
1109 if (!fp) {
1110 ERROR("Turbostat plugin: Failed to open /proc/stat");
1111 return -1;
1112 }
1114 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1115 if (retval != 0) {
1116 ERROR("Turbostat plugin: Failed to parse /proc/stat");
1117 fclose(fp);
1118 return -1;
1119 }
1121 while (1) {
1122 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1123 if (retval != 1)
1124 break;
1126 retval = func(cpu_num);
1127 if (retval) {
1128 fclose(fp);
1129 return(retval);
1130 }
1131 }
1132 fclose(fp);
1133 return 0;
1134 }
1136 /*
1137 * Update the stored topology.max_cpu_id
1138 */
1139 static int
1140 update_max_cpu_id(int cpu)
1141 {
1142 if (topology.max_cpu_id < cpu)
1143 topology.max_cpu_id = cpu;
1144 return 0;
1145 }
1147 static int
1148 mark_cpu_present(int cpu)
1149 {
1150 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1151 return 0;
1152 }
1154 static int __attribute__((warn_unused_result))
1155 allocate_cpu_set(cpu_set_t ** set, size_t * size) {
1156 *set = CPU_ALLOC(topology.max_cpu_id + 1);
1157 if (*set == NULL) {
1158 ERROR("Turbostat plugin: Unable to allocate CPU state");
1159 return -1;
1160 }
1161 *size = CPU_ALLOC_SIZE(topology.max_cpu_id + 1);
1162 CPU_ZERO_S(*size, *set);
1163 return 0;
1164 }
1166 /*
1167 * Build a local representation of the cpu distribution
1168 */
1169 static int __attribute__((warn_unused_result))
1170 topology_probe()
1171 {
1172 int i;
1173 int ret;
1174 int max_package_id, max_core_id, max_thread_id;
1175 max_package_id = max_core_id = max_thread_id = 0;
1177 /* Clean topology */
1178 free(topology.cpus);
1179 memset(&topology, 0, sizeof(topology));
1181 ret = for_all_proc_cpus(update_max_cpu_id);
1182 if (ret != 0)
1183 goto err;
1185 topology.cpus = calloc(1, (topology.max_cpu_id + 1) * sizeof(struct cpu_topology));
1186 if (topology.cpus == NULL) {
1187 ERROR("Turbostat plugin: Unable to allocate memory for CPU topology");
1188 return -1;
1189 }
1191 ret = allocate_cpu_set(&cpu_present_set, &cpu_present_setsize);
1192 if (ret != 0)
1193 goto err;
1194 ret = allocate_cpu_set(&cpu_affinity_set, &cpu_affinity_setsize);
1195 if (ret != 0)
1196 goto err;
1197 ret = allocate_cpu_set(&cpu_saved_affinity_set, &cpu_saved_affinity_setsize);
1198 if (ret != 0)
1199 goto err;
1201 ret = for_all_proc_cpus(mark_cpu_present);
1202 if (ret != 0)
1203 goto err;
1205 /*
1206 * For online cpus
1207 * find max_core_id, max_package_id
1208 */
1209 for (i = 0; i <= topology.max_cpu_id; ++i) {
1210 int num_threads;
1211 struct cpu_topology *cpu = &topology.cpus[i];
1213 if (cpu_is_not_present(i)) {
1214 WARNING("Turbostat plugin: cpu%d NOT PRESENT", i);
1215 continue;
1216 }
1218 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", i);
1219 if (ret < 0)
1220 goto err;
1221 else
1222 cpu->package_id = ret;
1223 if (cpu->package_id > max_package_id)
1224 max_package_id = cpu->package_id;
1226 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", i);
1227 if (ret < 0)
1228 goto err;
1229 else
1230 cpu->core_id = ret;
1231 if (cpu->core_id > max_core_id)
1232 max_core_id = cpu->core_id;
1233 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", i);
1234 if (ret < 0)
1235 goto err;
1236 else if (ret == i)
1237 cpu->first_core_in_package = 1;
1239 ret = get_threads_on_core(i);
1240 if (ret < 0)
1241 goto err;
1242 else
1243 num_threads = ret;
1244 if (num_threads > max_thread_id)
1245 max_thread_id = num_threads;
1246 ret = parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", i);
1247 if (ret < 0)
1248 goto err;
1249 else if (ret == i)
1250 cpu->first_thread_in_core = 1;
1252 DEBUG("Turbostat plugin: cpu %d pkg %d core %d\n",
1253 i, cpu->package_id, cpu->core_id);
1254 }
1255 /* Num is max + 1 (need to count 0) */
1256 topology.num_packages = max_package_id + 1;
1257 topology.num_cores = max_core_id + 1;
1258 topology.num_threads = max_thread_id + 1;
1260 return 0;
1261 err:
1262 free(topology.cpus);
1263 return ret;
1264 }
1267 /************************
1268 * Main alloc/init/free *
1269 ************************/
1271 static int
1272 allocate_counters(struct thread_data **threads, struct core_data **cores, struct pkg_data **packages)
1273 {
1274 int i;
1275 int total_threads, total_cores;
1277 total_threads = topology.num_threads * topology.num_cores * topology.num_packages;
1278 *threads = calloc(total_threads, sizeof(struct thread_data));
1279 if (*threads == NULL)
1280 goto err;
1282 for (i = 0; i < total_threads; ++i)
1283 (*threads)[i].cpu_id = -1;
1285 total_cores = topology.num_cores * topology.num_packages;
1286 *cores = calloc(total_cores, sizeof(struct core_data));
1287 if (*cores == NULL)
1288 goto err_clean_threads;
1290 for (i = 0; i < total_cores; ++i)
1291 (*cores)[i].core_id = -1;
1293 *packages = calloc(topology.num_packages, sizeof(struct pkg_data));
1294 if (*packages == NULL)
1295 goto err_clean_cores;
1297 for (i = 0; i < topology.num_packages; i++)
1298 (*packages)[i].package_id = i;
1300 return 0;
1302 err_clean_cores:
1303 free(*cores);
1304 err_clean_threads:
1305 free(*threads);
1306 err:
1307 ERROR("Turbostat plugin: Failled to allocate memory for counters");
1308 return -1;
1309 }
1311 static void
1312 init_counter(struct thread_data *thread_base, struct core_data *core_base,
1313 struct pkg_data *pkg_base, int cpu_id)
1314 {
1315 struct thread_data *t;
1316 struct core_data *c;
1317 struct pkg_data *p;
1318 struct cpu_topology *cpu = &topology.cpus[cpu_id];
1320 t = GET_THREAD(thread_base, !(cpu->first_thread_in_core), cpu->core_id, cpu->package_id);
1321 c = GET_CORE(core_base, cpu->core_id, cpu->package_id);
1322 p = GET_PKG(pkg_base, cpu->package_id);
1324 t->cpu_id = cpu_id;
1325 if (cpu->first_thread_in_core)
1326 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1327 if (cpu->first_core_in_package)
1328 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1330 c->core_id = cpu->core_id;
1331 p->package_id = cpu->package_id;
1332 }
1334 static void
1335 initialize_counters(void)
1336 {
1337 int cpu_id;
1339 for (cpu_id = 0; cpu_id <= topology.max_cpu_id; ++cpu_id) {
1340 if (cpu_is_not_present(cpu_id))
1341 continue;
1342 init_counter(EVEN_COUNTERS, cpu_id);
1343 init_counter(ODD_COUNTERS, cpu_id);
1344 init_counter(DELTA_COUNTERS, cpu_id);
1345 }
1346 }
1350 static void
1351 free_all_buffers(void)
1352 {
1353 allocated = 0;
1354 initialized = 0;
1356 CPU_FREE(cpu_present_set);
1357 cpu_present_set = NULL;
1358 cpu_present_set = 0;
1360 CPU_FREE(cpu_affinity_set);
1361 cpu_affinity_set = NULL;
1362 cpu_affinity_setsize = 0;
1364 CPU_FREE(cpu_saved_affinity_set);
1365 cpu_saved_affinity_set = NULL;
1366 cpu_saved_affinity_setsize = 0;
1368 free(thread_even);
1369 free(core_even);
1370 free(package_even);
1372 thread_even = NULL;
1373 core_even = NULL;
1374 package_even = NULL;
1376 free(thread_odd);
1377 free(core_odd);
1378 free(package_odd);
1380 thread_odd = NULL;
1381 core_odd = NULL;
1382 package_odd = NULL;
1384 free(thread_delta);
1385 free(core_delta);
1386 free(package_delta);
1388 thread_delta = NULL;
1389 core_delta = NULL;
1390 package_delta = NULL;
1391 }
1394 /**********************
1395 * Collectd functions *
1396 **********************/
1398 #define DO_OR_GOTO_ERR(something) \
1399 do { \
1400 ret = (something); \
1401 if (ret < 0) \
1402 goto err; \
1403 } while (0)
1405 static int setup_all_buffers(void)
1406 {
1407 int ret;
1409 DO_OR_GOTO_ERR(topology_probe());
1410 DO_OR_GOTO_ERR(allocate_counters(&thread_even, &core_even, &package_even));
1411 DO_OR_GOTO_ERR(allocate_counters(&thread_odd, &core_odd, &package_odd));
1412 DO_OR_GOTO_ERR(allocate_counters(&thread_delta, &core_delta, &package_delta));
1413 initialize_counters();
1414 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, EVEN_COUNTERS));
1415 DO_OR_GOTO_ERR(for_all_cpus(set_temperature_target, ODD_COUNTERS));
1417 allocated = 1;
1418 return 0;
1419 err:
1420 free_all_buffers();
1421 return ret;
1422 }
1424 static int
1425 turbostat_read(void)
1426 {
1427 int ret;
1429 if (!allocated) {
1430 if ((ret = setup_all_buffers()) < 0)
1431 return ret;
1432 }
1434 if (for_all_proc_cpus(cpu_is_not_present)) {
1435 free_all_buffers();
1436 if ((ret = setup_all_buffers()) < 0)
1437 return ret;
1438 if (for_all_proc_cpus(cpu_is_not_present)) {
1439 ERROR("Turbostat plugin: CPU appeared just after "
1440 "initialization");
1441 return -1;
1442 }
1443 }
1445 /* Saving the scheduling affinity, as it will be modified by get_counters */
1446 if (sched_getaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set) != 0) {
1447 ERROR("Turbostat plugin: Unable to save the CPU affinity");
1448 return -1;
1449 }
1451 if (!initialized) {
1452 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1453 goto out;
1454 time_even = cdtime();
1455 is_even = 1;
1456 initialized = 1;
1457 ret = 0;
1458 goto out;
1459 }
1461 if (is_even) {
1462 if ((ret = for_all_cpus(get_counters, ODD_COUNTERS)) < 0)
1463 goto out;
1464 time_odd = cdtime();
1465 is_even = 0;
1466 time_delta = time_odd - time_even;
1467 if ((ret = for_all_cpus_delta(ODD_COUNTERS, EVEN_COUNTERS)) < 0)
1468 goto out;
1469 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1470 goto out;
1471 } else {
1472 if ((ret = for_all_cpus(get_counters, EVEN_COUNTERS)) < 0)
1473 goto out;
1474 time_even = cdtime();
1475 is_even = 1;
1476 time_delta = time_even - time_odd;
1477 if ((ret = for_all_cpus_delta(EVEN_COUNTERS, ODD_COUNTERS)) < 0)
1478 goto out;
1479 if ((ret = for_all_cpus(submit_counters, DELTA_COUNTERS)) < 0)
1480 goto out;
1481 }
1482 ret = 0;
1483 out:
1484 /*
1485 * Let's restore the affinity
1486 * This might fail if the number of CPU changed, but we can't do anything in that case..
1487 */
1488 (void)sched_setaffinity(0, cpu_saved_affinity_setsize, cpu_saved_affinity_set);
1489 return ret;
1490 }
1492 static int
1493 check_permissions(void)
1494 {
1495 struct __user_cap_header_struct cap_header_data;
1496 cap_user_header_t cap_header = &cap_header_data;
1497 struct __user_cap_data_struct cap_data_data;
1498 cap_user_data_t cap_data = &cap_data_data;
1499 int ret = 0;
1501 if (getuid() == 0) {
1502 /* We have everything we need */
1503 return 0;
1504 }
1506 /* check for CAP_SYS_RAWIO */
1507 cap_header->pid = getpid();
1508 cap_header->version = _LINUX_CAPABILITY_VERSION;
1509 if (capget(cap_header, cap_data) < 0) {
1510 ERROR("Turbostat plugin: capget failed");
1511 return -1;
1512 }
1514 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1515 WARNING("Turbostat plugin: Collectd doesn't have the "
1516 "CAP_SYS_RAWIO capability. If you don't want to run "
1517 "collectd as root, try running \"setcap "
1518 "cap_sys_rawio=ep\" on collectd binary");
1519 ret = -1;
1520 }
1522 if (euidaccess("/dev/cpu/0/msr", R_OK)) {
1523 WARNING("Turbostat plugin: Collectd cannot open"
1524 "/dev/cpu/0/msr. If you don't want to run collectd as "
1525 "root, you need to change the ownership (chown) and "
1526 "permissions on /dev/cpu/*/msr to allow such access");
1527 ret = -1;
1528 }
1530 if (ret != 0)
1531 ERROR("Turbostat plugin: Initialization failed: this plugin "
1532 "requires collectd to either to run as root or give "
1533 "collectd a special capability (CAP_SYS_RAWIO) and read "
1534 "access to /dev/cpu/*/msr (see previous warnings)");
1535 return ret;
1536 }
1538 static int
1539 turbostat_init(void)
1540 {
1541 struct stat sb;
1542 int ret;
1544 if (stat("/dev/cpu/0/msr", &sb)) {
1545 ERROR("Turbostat plugin: Initialization failed: /dev/cpu/0/msr"
1546 " does not exist while the CPU supports MSR. You may be "
1547 "missing the corresponding kernel module, please try '# "
1548 "modprobe msr'");
1549 return -1;
1550 }
1552 DO_OR_GOTO_ERR(check_permissions());
1554 DO_OR_GOTO_ERR(probe_cpu());
1556 DO_OR_GOTO_ERR(setup_all_buffers());
1558 plugin_register_read(PLUGIN_NAME, turbostat_read);
1560 return 0;
1561 err:
1562 free_all_buffers();
1563 return ret;
1564 }
1566 static int
1567 turbostat_config(const char *key, const char *value)
1568 {
1569 long unsigned int tmp_val;
1570 char *end;
1572 if (strcasecmp("TCCActivationTemp", key) == 0) {
1573 tmp_val = strtoul(value, &end, 0);
1574 if (*end != '\0' || tmp_val > UINT_MAX) {
1575 ERROR("Turbostat plugin: Invalid TCCActivationTemp '%s'",
1576 value);
1577 return -1;
1578 }
1579 tcc_activation_temp = (unsigned int) tmp_val;
1580 } else {
1581 ERROR("Turbostat plugin: Invalid configuration option '%s'",
1582 key);
1583 return -1;
1584 }
1585 return 0;
1586 }
1588 void module_register(void)
1589 {
1590 plugin_register_init(PLUGIN_NAME, turbostat_init);
1591 plugin_register_config(PLUGIN_NAME, turbostat_config, config_keys, config_keys_num);
1592 }