xref: /openbmc/linux/tools/power/x86/turbostat/turbostat.c (revision 678a3bd1b3de6d2ebf604e7d708bc8150bb667e9)
1 /*
2  * turbostat -- show CPU frequency and C-state residency
3  * on modern Intel turbo-capable processors.
4  *
5  * Copyright (c) 2013 Intel Corporation.
6  * Len Brown <len.brown@intel.com>
7  *
8  * This program is free software; you can redistribute it and/or modify it
9  * under the terms and conditions of the GNU General Public License,
10  * version 2, as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope it will be useful, but WITHOUT
13  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
15  * more details.
16  *
17  * You should have received a copy of the GNU General Public License along with
18  * this program; if not, write to the Free Software Foundation, Inc.,
19  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
20  */
21 
22 #define _GNU_SOURCE
23 #include MSRHEADER
24 #include INTEL_FAMILY_HEADER
25 #include <stdarg.h>
26 #include <stdio.h>
27 #include <err.h>
28 #include <unistd.h>
29 #include <sys/types.h>
30 #include <sys/wait.h>
31 #include <sys/stat.h>
32 #include <sys/resource.h>
33 #include <fcntl.h>
34 #include <signal.h>
35 #include <sys/time.h>
36 #include <stdlib.h>
37 #include <getopt.h>
38 #include <dirent.h>
39 #include <string.h>
40 #include <ctype.h>
41 #include <sched.h>
42 #include <time.h>
43 #include <cpuid.h>
44 #include <linux/capability.h>
45 #include <errno.h>
46 
47 char *proc_stat = "/proc/stat";
48 FILE *outf;
49 int *fd_percpu;
50 struct timespec interval_ts = {5, 0};
51 unsigned int debug;
52 unsigned int rapl_joules;
53 unsigned int summary_only;
54 unsigned int dump_only;
55 unsigned int do_nhm_cstates;
56 unsigned int do_snb_cstates;
57 unsigned int do_knl_cstates;
58 unsigned int do_pc2;
59 unsigned int do_pc3;
60 unsigned int do_pc6;
61 unsigned int do_pc7;
62 unsigned int do_c8_c9_c10;
63 unsigned int do_skl_residency;
64 unsigned int do_slm_cstates;
65 unsigned int use_c1_residency_msr;
66 unsigned int has_aperf;
67 unsigned int has_epb;
68 unsigned int do_irtl_snb;
69 unsigned int do_irtl_hsw;
70 unsigned int units = 1000000;	/* MHz etc */
71 unsigned int genuine_intel;
72 unsigned int has_invariant_tsc;
73 unsigned int do_nhm_platform_info;
74 unsigned int aperf_mperf_multiplier = 1;
75 int do_irq = 1;
76 int do_smi;
77 double bclk;
78 double base_hz;
79 unsigned int has_base_hz;
80 double tsc_tweak = 1.0;
81 unsigned int show_pkg;
82 unsigned int show_core;
83 unsigned int show_cpu;
84 unsigned int show_pkg_only;
85 unsigned int show_core_only;
86 char *output_buffer, *outp;
87 unsigned int do_rapl;
88 unsigned int do_dts;
89 unsigned int do_ptm;
90 unsigned int do_gfx_rc6_ms;
91 unsigned long long  gfx_cur_rc6_ms;
92 unsigned int do_gfx_mhz;
93 unsigned int gfx_cur_mhz;
94 unsigned int tcc_activation_temp;
95 unsigned int tcc_activation_temp_override;
96 double rapl_power_units, rapl_time_units;
97 double rapl_dram_energy_units, rapl_energy_units;
98 double rapl_joule_counter_range;
99 unsigned int do_core_perf_limit_reasons;
100 unsigned int do_gfx_perf_limit_reasons;
101 unsigned int do_ring_perf_limit_reasons;
102 unsigned int crystal_hz;
103 unsigned long long tsc_hz;
104 int base_cpu;
105 double discover_bclk(unsigned int family, unsigned int model);
106 unsigned int has_hwp;	/* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
107 			/* IA32_HWP_REQUEST, IA32_HWP_STATUS */
108 unsigned int has_hwp_notify;		/* IA32_HWP_INTERRUPT */
109 unsigned int has_hwp_activity_window;	/* IA32_HWP_REQUEST[bits 41:32] */
110 unsigned int has_hwp_epp;		/* IA32_HWP_REQUEST[bits 31:24] */
111 unsigned int has_hwp_pkg;		/* IA32_HWP_REQUEST_PKG */
112 
113 #define RAPL_PKG		(1 << 0)
114 					/* 0x610 MSR_PKG_POWER_LIMIT */
115 					/* 0x611 MSR_PKG_ENERGY_STATUS */
116 #define RAPL_PKG_PERF_STATUS	(1 << 1)
117 					/* 0x613 MSR_PKG_PERF_STATUS */
118 #define RAPL_PKG_POWER_INFO	(1 << 2)
119 					/* 0x614 MSR_PKG_POWER_INFO */
120 
121 #define RAPL_DRAM		(1 << 3)
122 					/* 0x618 MSR_DRAM_POWER_LIMIT */
123 					/* 0x619 MSR_DRAM_ENERGY_STATUS */
124 #define RAPL_DRAM_PERF_STATUS	(1 << 4)
125 					/* 0x61b MSR_DRAM_PERF_STATUS */
126 #define RAPL_DRAM_POWER_INFO	(1 << 5)
127 					/* 0x61c MSR_DRAM_POWER_INFO */
128 
129 #define RAPL_CORES_POWER_LIMIT	(1 << 6)
130 					/* 0x638 MSR_PP0_POWER_LIMIT */
131 #define RAPL_CORE_POLICY	(1 << 7)
132 					/* 0x63a MSR_PP0_POLICY */
133 
134 #define RAPL_GFX		(1 << 8)
135 					/* 0x640 MSR_PP1_POWER_LIMIT */
136 					/* 0x641 MSR_PP1_ENERGY_STATUS */
137 					/* 0x642 MSR_PP1_POLICY */
138 
139 #define RAPL_CORES_ENERGY_STATUS	(1 << 9)
140 					/* 0x639 MSR_PP0_ENERGY_STATUS */
141 #define RAPL_CORES (RAPL_CORES_ENERGY_STATUS | RAPL_CORES_POWER_LIMIT)
142 #define	TJMAX_DEFAULT	100
143 
144 #define MAX(a, b) ((a) > (b) ? (a) : (b))
145 
146 /*
147  * buffer size used by sscanf() for added column names
148  * Usually truncated to 7 characters, but also handles 18 columns for raw 64-bit counters
149  */
150 #define	NAME_BYTES 20
151 
152 int backwards_count;
153 char *progname;
154 
155 cpu_set_t *cpu_present_set, *cpu_affinity_set;
156 size_t cpu_present_setsize, cpu_affinity_setsize;
157 #define MAX_ADDED_COUNTERS 16
158 
159 struct thread_data {
160 	unsigned long long tsc;
161 	unsigned long long aperf;
162 	unsigned long long mperf;
163 	unsigned long long c1;
164 	unsigned int irq_count;
165 	unsigned int smi_count;
166 	unsigned int cpu_id;
167 	unsigned int flags;
168 #define CPU_IS_FIRST_THREAD_IN_CORE	0x2
169 #define CPU_IS_FIRST_CORE_IN_PACKAGE	0x4
170 	unsigned long long counter[MAX_ADDED_COUNTERS];
171 } *thread_even, *thread_odd;
172 
173 struct core_data {
174 	unsigned long long c3;
175 	unsigned long long c6;
176 	unsigned long long c7;
177 	unsigned int core_temp_c;
178 	unsigned int core_id;
179 	unsigned long long counter[MAX_ADDED_COUNTERS];
180 } *core_even, *core_odd;
181 
182 struct pkg_data {
183 	unsigned long long pc2;
184 	unsigned long long pc3;
185 	unsigned long long pc6;
186 	unsigned long long pc7;
187 	unsigned long long pc8;
188 	unsigned long long pc9;
189 	unsigned long long pc10;
190 	unsigned long long pkg_wtd_core_c0;
191 	unsigned long long pkg_any_core_c0;
192 	unsigned long long pkg_any_gfxe_c0;
193 	unsigned long long pkg_both_core_gfxe_c0;
194 	long long gfx_rc6_ms;
195 	unsigned int gfx_mhz;
196 	unsigned int package_id;
197 	unsigned int energy_pkg;	/* MSR_PKG_ENERGY_STATUS */
198 	unsigned int energy_dram;	/* MSR_DRAM_ENERGY_STATUS */
199 	unsigned int energy_cores;	/* MSR_PP0_ENERGY_STATUS */
200 	unsigned int energy_gfx;	/* MSR_PP1_ENERGY_STATUS */
201 	unsigned int rapl_pkg_perf_status;	/* MSR_PKG_PERF_STATUS */
202 	unsigned int rapl_dram_perf_status;	/* MSR_DRAM_PERF_STATUS */
203 	unsigned int pkg_temp_c;
204 	unsigned long long counter[MAX_ADDED_COUNTERS];
205 } *package_even, *package_odd;
206 
207 #define ODD_COUNTERS thread_odd, core_odd, package_odd
208 #define EVEN_COUNTERS thread_even, core_even, package_even
209 
210 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
211 	(thread_base + (pkg_no) * topo.num_cores_per_pkg * \
212 		topo.num_threads_per_core + \
213 		(core_no) * topo.num_threads_per_core + (thread_no))
214 #define GET_CORE(core_base, core_no, pkg_no) \
215 	(core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
216 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
217 
218 enum counter_scope {SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE};
219 enum counter_type {COUNTER_CYCLES, COUNTER_SECONDS};
220 enum counter_format {FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT};
221 
222 struct msr_counter {
223 	unsigned int msr_num;
224 	char name[NAME_BYTES];
225 	unsigned int width;
226 	enum counter_type type;
227 	enum counter_format format;
228 	struct msr_counter *next;
229 };
230 
231 struct sys_counters {
232 	unsigned int added_thread_counters;
233 	unsigned int added_core_counters;
234 	unsigned int added_package_counters;
235 	struct msr_counter *tp;
236 	struct msr_counter *cp;
237 	struct msr_counter *pp;
238 } sys;
239 
240 struct system_summary {
241 	struct thread_data threads;
242 	struct core_data cores;
243 	struct pkg_data packages;
244 } average;
245 
246 
247 struct topo_params {
248 	int num_packages;
249 	int num_cpus;
250 	int num_cores;
251 	int max_cpu_num;
252 	int num_cores_per_pkg;
253 	int num_threads_per_core;
254 } topo;
255 
256 struct timeval tv_even, tv_odd, tv_delta;
257 
258 int *irq_column_2_cpu;	/* /proc/interrupts column numbers */
259 int *irqs_per_cpu;		/* indexed by cpu_num */
260 
261 void setup_all_buffers(void);
262 
263 int cpu_is_not_present(int cpu)
264 {
265 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
266 }
267 /*
268  * run func(thread, core, package) in topology order
269  * skip non-present cpus
270  */
271 
272 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
273 	struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
274 {
275 	int retval, pkg_no, core_no, thread_no;
276 
277 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
278 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
279 			for (thread_no = 0; thread_no <
280 				topo.num_threads_per_core; ++thread_no) {
281 				struct thread_data *t;
282 				struct core_data *c;
283 				struct pkg_data *p;
284 
285 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
286 
287 				if (cpu_is_not_present(t->cpu_id))
288 					continue;
289 
290 				c = GET_CORE(core_base, core_no, pkg_no);
291 				p = GET_PKG(pkg_base, pkg_no);
292 
293 				retval = func(t, c, p);
294 				if (retval)
295 					return retval;
296 			}
297 		}
298 	}
299 	return 0;
300 }
301 
302 int cpu_migrate(int cpu)
303 {
304 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
305 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
306 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
307 		return -1;
308 	else
309 		return 0;
310 }
311 int get_msr_fd(int cpu)
312 {
313 	char pathname[32];
314 	int fd;
315 
316 	fd = fd_percpu[cpu];
317 
318 	if (fd)
319 		return fd;
320 
321 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
322 	fd = open(pathname, O_RDONLY);
323 	if (fd < 0)
324 		err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
325 
326 	fd_percpu[cpu] = fd;
327 
328 	return fd;
329 }
330 
331 int get_msr(int cpu, off_t offset, unsigned long long *msr)
332 {
333 	ssize_t retval;
334 
335 	retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
336 
337 	if (retval != sizeof *msr)
338 		err(-1, "msr %d offset 0x%llx read failed", cpu, (unsigned long long)offset);
339 
340 	return 0;
341 }
342 
343 /*
344  * Example Format w/ field column widths:
345  *
346  *  Package    Core     CPU Avg_MHz Bzy_MHz TSC_MHz     IRQ   SMI   Busy% CPU_%c1 CPU_%c3 CPU_%c6 CPU_%c7 ThreadC CoreTmp  CoreCnt PkgTmp  GFXMHz Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 PkgWatt CorWatt GFXWatt PkgCnt
347  * 12345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
348  */
349 
350 void print_header(void)
351 {
352 	struct msr_counter *mp;
353 
354 	if (show_pkg)
355 		outp += sprintf(outp, "\tPackage");
356 	if (show_core)
357 		outp += sprintf(outp, "\tCore");
358 	if (show_cpu)
359 		outp += sprintf(outp, "\tCPU");
360 	if (has_aperf)
361 		outp += sprintf(outp, "\tAvg_MHz");
362 	if (has_aperf)
363 		outp += sprintf(outp, "\tBusy%%");
364 	if (has_aperf)
365 		outp += sprintf(outp, "\tBzy_MHz");
366 	outp += sprintf(outp, "\tTSC_MHz");
367 
368 	if (!debug)
369 		goto done;
370 
371 	if (do_irq)
372 		outp += sprintf(outp, "\tIRQ");
373 	if (do_smi)
374 		outp += sprintf(outp, "\tSMI");
375 
376 	if (do_nhm_cstates)
377 		outp += sprintf(outp, "\tCPU%%c1");
378 
379 	for (mp = sys.tp; mp; mp = mp->next) {
380 		if (mp->format == FORMAT_RAW) {
381 			if (mp->width == 64)
382 				outp += sprintf(outp, "\t%18.18s", mp->name);
383 			else
384 				outp += sprintf(outp, "\t%10.10s", mp->name);
385 		} else {
386 			outp += sprintf(outp, "\t%-7.7s", mp->name);
387 		}
388 	}
389 
390 	if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
391 		outp += sprintf(outp, "\tCPU%%c3");
392 	if (do_nhm_cstates)
393 		outp += sprintf(outp, "\tCPU%%c6");
394 	if (do_snb_cstates)
395 		outp += sprintf(outp, "\tCPU%%c7");
396 
397 
398 	if (do_dts)
399 		outp += sprintf(outp, "\tCoreTmp");
400 
401 	for (mp = sys.cp; mp; mp = mp->next) {
402 		if (mp->format == FORMAT_RAW) {
403 			if (mp->width == 64)
404 				outp += sprintf(outp, "\t%18.18s", mp->name);
405 			else
406 				outp += sprintf(outp, "\t%10.10s", mp->name);
407 		} else {
408 			outp += sprintf(outp, "\t%-7.7s", mp->name);
409 		}
410 	}
411 
412 	if (do_ptm)
413 		outp += sprintf(outp, "\tPkgTmp");
414 
415 	if (do_gfx_rc6_ms)
416 		outp += sprintf(outp, "\tGFX%%rc6");
417 
418 	if (do_gfx_mhz)
419 		outp += sprintf(outp, "\tGFXMHz");
420 
421 	if (do_skl_residency) {
422 		outp += sprintf(outp, "\tTotl%%C0");
423 		outp += sprintf(outp, "\tAny%%C0");
424 		outp += sprintf(outp, "\tGFX%%C0");
425 		outp += sprintf(outp, "\tCPUGFX%%");
426 	}
427 
428 	if (do_pc2)
429 		outp += sprintf(outp, "\tPkg%%pc2");
430 	if (do_pc3)
431 		outp += sprintf(outp, "\tPkg%%pc3");
432 	if (do_pc6)
433 		outp += sprintf(outp, "\tPkg%%pc6");
434 	if (do_pc7)
435 		outp += sprintf(outp, "\tPkg%%pc7");
436 	if (do_c8_c9_c10) {
437 		outp += sprintf(outp, "\tPkg%%pc8");
438 		outp += sprintf(outp, "\tPkg%%pc9");
439 		outp += sprintf(outp, "\tPk%%pc10");
440 	}
441 
442 	if (do_rapl && !rapl_joules) {
443 		if (do_rapl & RAPL_PKG)
444 			outp += sprintf(outp, "\tPkgWatt");
445 		if (do_rapl & RAPL_CORES_ENERGY_STATUS)
446 			outp += sprintf(outp, "\tCorWatt");
447 		if (do_rapl & RAPL_GFX)
448 			outp += sprintf(outp, "\tGFXWatt");
449 		if (do_rapl & RAPL_DRAM)
450 			outp += sprintf(outp, "\tRAMWatt");
451 		if (do_rapl & RAPL_PKG_PERF_STATUS)
452 			outp += sprintf(outp, "\tPKG_%%");
453 		if (do_rapl & RAPL_DRAM_PERF_STATUS)
454 			outp += sprintf(outp, "\tRAM_%%");
455 	} else if (do_rapl && rapl_joules) {
456 		if (do_rapl & RAPL_PKG)
457 			outp += sprintf(outp, "\tPkg_J");
458 		if (do_rapl & RAPL_CORES_ENERGY_STATUS)
459 			outp += sprintf(outp, "\tCor_J");
460 		if (do_rapl & RAPL_GFX)
461 			outp += sprintf(outp, "\tGFX_J");
462 		if (do_rapl & RAPL_DRAM)
463 			outp += sprintf(outp, "\tRAM_J");
464 		if (do_rapl & RAPL_PKG_PERF_STATUS)
465 			outp += sprintf(outp, "\tPKG_%%");
466 		if (do_rapl & RAPL_DRAM_PERF_STATUS)
467 			outp += sprintf(outp, "\tRAM_%%");
468 	}
469 	for (mp = sys.pp; mp; mp = mp->next) {
470 		if (mp->format == FORMAT_RAW) {
471 			if (mp->width == 64)
472 				outp += sprintf(outp, "\t%18.18s", mp->name);
473 			else
474 				outp += sprintf(outp, "\t%10.10s", mp->name);
475 		} else {
476 			outp += sprintf(outp, "\t%-7.7s", mp->name);
477 		}
478 	}
479 
480 done:
481 	outp += sprintf(outp, "\n");
482 }
483 
484 int dump_counters(struct thread_data *t, struct core_data *c,
485 	struct pkg_data *p)
486 {
487 	int i;
488 	struct msr_counter *mp;
489 
490 	outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
491 
492 	if (t) {
493 		outp += sprintf(outp, "CPU: %d flags 0x%x\n",
494 			t->cpu_id, t->flags);
495 		outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
496 		outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
497 		outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
498 		outp += sprintf(outp, "c1: %016llX\n", t->c1);
499 
500 		if (do_irq)
501 			outp += sprintf(outp, "IRQ: %08X\n", t->irq_count);
502 		if (do_smi)
503 			outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
504 
505 		for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
506 			outp += sprintf(outp, "tADDED [%d] msr0x%x: %08llX\n",
507 				i, mp->msr_num, t->counter[i]);
508 		}
509 	}
510 
511 	if (c) {
512 		outp += sprintf(outp, "core: %d\n", c->core_id);
513 		outp += sprintf(outp, "c3: %016llX\n", c->c3);
514 		outp += sprintf(outp, "c6: %016llX\n", c->c6);
515 		outp += sprintf(outp, "c7: %016llX\n", c->c7);
516 		outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
517 
518 		for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
519 			outp += sprintf(outp, "cADDED [%d] msr0x%x: %08llX\n",
520 				i, mp->msr_num, c->counter[i]);
521 		}
522 	}
523 
524 	if (p) {
525 		outp += sprintf(outp, "package: %d\n", p->package_id);
526 
527 		outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
528 		outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
529 		outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
530 		outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
531 
532 		outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
533 		if (do_pc3)
534 			outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
535 		if (do_pc6)
536 			outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
537 		if (do_pc7)
538 			outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
539 		outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
540 		outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
541 		outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
542 		outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
543 		outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
544 		outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
545 		outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
546 		outp += sprintf(outp, "Throttle PKG: %0X\n",
547 			p->rapl_pkg_perf_status);
548 		outp += sprintf(outp, "Throttle RAM: %0X\n",
549 			p->rapl_dram_perf_status);
550 		outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
551 
552 		for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
553 			outp += sprintf(outp, "pADDED [%d] msr0x%x: %08llX\n",
554 				i, mp->msr_num, p->counter[i]);
555 		}
556 	}
557 
558 	outp += sprintf(outp, "\n");
559 
560 	return 0;
561 }
562 
563 /*
564  * column formatting convention & formats
565  */
566 int format_counters(struct thread_data *t, struct core_data *c,
567 	struct pkg_data *p)
568 {
569 	double interval_float;
570 	char *fmt8;
571 	int i;
572 	struct msr_counter *mp;
573 
574 	 /* if showing only 1st thread in core and this isn't one, bail out */
575 	if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
576 		return 0;
577 
578 	 /* if showing only 1st thread in pkg and this isn't one, bail out */
579 	if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
580 		return 0;
581 
582 	interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
583 
584 	/* topo columns, print blanks on 1st (average) line */
585 	if (t == &average.threads) {
586 		if (show_pkg)
587 			outp += sprintf(outp, "\t-");
588 		if (show_core)
589 			outp += sprintf(outp, "\t-");
590 		if (show_cpu)
591 			outp += sprintf(outp, "\t-");
592 	} else {
593 		if (show_pkg) {
594 			if (p)
595 				outp += sprintf(outp, "\t%d", p->package_id);
596 			else
597 				outp += sprintf(outp, "\t-");
598 		}
599 		if (show_core) {
600 			if (c)
601 				outp += sprintf(outp, "\t%d", c->core_id);
602 			else
603 				outp += sprintf(outp, "\t-");
604 		}
605 		if (show_cpu)
606 			outp += sprintf(outp, "\t%d", t->cpu_id);
607 	}
608 
609 	/* Avg_MHz */
610 	if (has_aperf)
611 		outp += sprintf(outp, "\t%.0f",
612 			1.0 / units * t->aperf / interval_float);
613 
614 	/* Busy% */
615 	if (has_aperf)
616 		outp += sprintf(outp, "\t%.2f", 100.0 * t->mperf/t->tsc/tsc_tweak);
617 
618 	/* Bzy_MHz */
619 	if (has_aperf) {
620 		if (has_base_hz)
621 			outp += sprintf(outp, "\t%.0f", base_hz / units * t->aperf / t->mperf);
622 		else
623 			outp += sprintf(outp, "\t%.0f",
624 				1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
625 	}
626 
627 	/* TSC_MHz */
628 	outp += sprintf(outp, "\t%.0f", 1.0 * t->tsc/units/interval_float);
629 
630 	if (!debug)
631 		goto done;
632 
633 	/* IRQ */
634 	if (do_irq)
635 		outp += sprintf(outp, "\t%d", t->irq_count);
636 
637 	/* SMI */
638 	if (do_smi)
639 		outp += sprintf(outp, "\t%d", t->smi_count);
640 
641 	/* C1 */
642 	if (do_nhm_cstates)
643 		outp += sprintf(outp, "\t%.2f", 100.0 * t->c1/t->tsc);
644 
645 	/* Added counters */
646 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
647 		if (mp->format == FORMAT_RAW) {
648 			if (mp->width == 32)
649 				outp += sprintf(outp, "\t0x%08lx", (unsigned long) t->counter[i]);
650 			else
651 				outp += sprintf(outp, "\t0x%016llx", t->counter[i]);
652 		} else if (mp->format == FORMAT_DELTA) {
653 			outp += sprintf(outp, "\t%lld", t->counter[i]);
654 		} else if (mp->format == FORMAT_PERCENT) {
655 			outp += sprintf(outp, "\t%.2f", 100.0 * t->counter[i]/t->tsc);
656 		}
657 	}
658 
659 	/* print per-core data only for 1st thread in core */
660 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
661 		goto done;
662 
663 	if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
664 		outp += sprintf(outp, "\t%.2f", 100.0 * c->c3/t->tsc);
665 	if (do_nhm_cstates)
666 		outp += sprintf(outp, "\t%.2f", 100.0 * c->c6/t->tsc);
667 	if (do_snb_cstates)
668 		outp += sprintf(outp, "\t%.2f", 100.0 * c->c7/t->tsc);
669 
670 
671 	if (do_dts)
672 		outp += sprintf(outp, "\t%d", c->core_temp_c);
673 
674 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
675 		if (mp->format == FORMAT_RAW) {
676 			if (mp->width == 32)
677 				outp += sprintf(outp, "\t0x%08lx", (unsigned long) c->counter[i]);
678 			else
679 				outp += sprintf(outp, "\t0x%016llx", c->counter[i]);
680 		} else if (mp->format == FORMAT_DELTA) {
681 			outp += sprintf(outp, "\t%lld", c->counter[i]);
682 		} else if (mp->format == FORMAT_PERCENT) {
683 			outp += sprintf(outp, "\t%.2f", 100.0 * c->counter[i]/t->tsc);
684 		}
685 	}
686 
687 	/* print per-package data only for 1st core in package */
688 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
689 		goto done;
690 
691 	/* PkgTmp */
692 	if (do_ptm)
693 		outp += sprintf(outp, "\t%d", p->pkg_temp_c);
694 
695 	/* GFXrc6 */
696 	if (do_gfx_rc6_ms) {
697 		if (p->gfx_rc6_ms == -1) {	/* detect GFX counter reset */
698 			outp += sprintf(outp, "\t**.**");
699 		} else {
700 			outp += sprintf(outp, "\t%.2f",
701 				p->gfx_rc6_ms / 10.0 / interval_float);
702 		}
703 	}
704 
705 	/* GFXMHz */
706 	if (do_gfx_mhz)
707 		outp += sprintf(outp, "\t%d", p->gfx_mhz);
708 
709 	/* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
710 	if (do_skl_residency) {
711 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pkg_wtd_core_c0/t->tsc);
712 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pkg_any_core_c0/t->tsc);
713 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pkg_any_gfxe_c0/t->tsc);
714 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pkg_both_core_gfxe_c0/t->tsc);
715 	}
716 
717 	if (do_pc2)
718 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pc2/t->tsc);
719 	if (do_pc3)
720 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pc3/t->tsc);
721 	if (do_pc6)
722 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pc6/t->tsc);
723 	if (do_pc7)
724 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pc7/t->tsc);
725 	if (do_c8_c9_c10) {
726 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pc8/t->tsc);
727 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pc9/t->tsc);
728 		outp += sprintf(outp, "\t%.2f", 100.0 * p->pc10/t->tsc);
729 	}
730 
731 	/*
732  	 * If measurement interval exceeds minimum RAPL Joule Counter range,
733  	 * indicate that results are suspect by printing "**" in fraction place.
734  	 */
735 	if (interval_float < rapl_joule_counter_range)
736 		fmt8 = "\t%.2f";
737 	else
738 		fmt8 = "%6.0f**";
739 
740 	if (do_rapl && !rapl_joules) {
741 		if (do_rapl & RAPL_PKG)
742 			outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
743 		if (do_rapl & RAPL_CORES_ENERGY_STATUS)
744 			outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
745 		if (do_rapl & RAPL_GFX)
746 			outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
747 		if (do_rapl & RAPL_DRAM)
748 			outp += sprintf(outp, fmt8, p->energy_dram * rapl_dram_energy_units / interval_float);
749 		if (do_rapl & RAPL_PKG_PERF_STATUS)
750 			outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
751 		if (do_rapl & RAPL_DRAM_PERF_STATUS)
752 			outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
753 	} else if (do_rapl && rapl_joules) {
754 		if (do_rapl & RAPL_PKG)
755 			outp += sprintf(outp, fmt8,
756 					p->energy_pkg * rapl_energy_units);
757 		if (do_rapl & RAPL_CORES)
758 			outp += sprintf(outp, fmt8,
759 					p->energy_cores * rapl_energy_units);
760 		if (do_rapl & RAPL_GFX)
761 			outp += sprintf(outp, fmt8,
762 					p->energy_gfx * rapl_energy_units);
763 		if (do_rapl & RAPL_DRAM)
764 			outp += sprintf(outp, fmt8,
765 					p->energy_dram * rapl_dram_energy_units);
766 		if (do_rapl & RAPL_PKG_PERF_STATUS)
767 			outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
768 		if (do_rapl & RAPL_DRAM_PERF_STATUS)
769 			outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
770 	}
771 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
772 		if (mp->format == FORMAT_RAW) {
773 			if (mp->width == 32)
774 				outp += sprintf(outp, "\t0x%08lx", (unsigned long) p->counter[i]);
775 			else
776 				outp += sprintf(outp, "\t0x%016llx", p->counter[i]);
777 		} else if (mp->format == FORMAT_DELTA) {
778 			outp += sprintf(outp, "\t%lld", p->counter[i]);
779 		} else if (mp->format == FORMAT_PERCENT) {
780 			outp += sprintf(outp, "\t%.2f", 100.0 * p->counter[i]/t->tsc);
781 		}
782 	}
783 
784 done:
785 	outp += sprintf(outp, "\n");
786 
787 	return 0;
788 }
789 
790 void flush_output_stdout(void)
791 {
792 	FILE *filep;
793 
794 	if (outf == stderr)
795 		filep = stdout;
796 	else
797 		filep = outf;
798 
799 	fputs(output_buffer, filep);
800 	fflush(filep);
801 
802 	outp = output_buffer;
803 }
804 void flush_output_stderr(void)
805 {
806 	fputs(output_buffer, outf);
807 	fflush(outf);
808 	outp = output_buffer;
809 }
810 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
811 {
812 	static int printed;
813 
814 	if (!printed || !summary_only)
815 		print_header();
816 
817 	if (topo.num_cpus > 1)
818 		format_counters(&average.threads, &average.cores,
819 			&average.packages);
820 
821 	printed = 1;
822 
823 	if (summary_only)
824 		return;
825 
826 	for_all_cpus(format_counters, t, c, p);
827 }
828 
829 #define DELTA_WRAP32(new, old)			\
830 	if (new > old) {			\
831 		old = new - old;		\
832 	} else {				\
833 		old = 0x100000000 + new - old;	\
834 	}
835 
836 int
837 delta_package(struct pkg_data *new, struct pkg_data *old)
838 {
839 	int i;
840 	struct msr_counter *mp;
841 
842 	if (do_skl_residency) {
843 		old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
844 		old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
845 		old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
846 		old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
847 	}
848 	old->pc2 = new->pc2 - old->pc2;
849 	if (do_pc3)
850 		old->pc3 = new->pc3 - old->pc3;
851 	if (do_pc6)
852 		old->pc6 = new->pc6 - old->pc6;
853 	if (do_pc7)
854 		old->pc7 = new->pc7 - old->pc7;
855 	old->pc8 = new->pc8 - old->pc8;
856 	old->pc9 = new->pc9 - old->pc9;
857 	old->pc10 = new->pc10 - old->pc10;
858 	old->pkg_temp_c = new->pkg_temp_c;
859 
860 	/* flag an error when rc6 counter resets/wraps */
861 	if (old->gfx_rc6_ms >  new->gfx_rc6_ms)
862 		old->gfx_rc6_ms = -1;
863 	else
864 		old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
865 
866 	old->gfx_mhz = new->gfx_mhz;
867 
868 	DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
869 	DELTA_WRAP32(new->energy_cores, old->energy_cores);
870 	DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
871 	DELTA_WRAP32(new->energy_dram, old->energy_dram);
872 	DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
873 	DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
874 
875 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
876 		if (mp->format == FORMAT_RAW)
877 			old->counter[i] = new->counter[i];
878 		else
879 			old->counter[i] = new->counter[i] - old->counter[i];
880 	}
881 
882 	return 0;
883 }
884 
885 void
886 delta_core(struct core_data *new, struct core_data *old)
887 {
888 	int i;
889 	struct msr_counter *mp;
890 
891 	old->c3 = new->c3 - old->c3;
892 	old->c6 = new->c6 - old->c6;
893 	old->c7 = new->c7 - old->c7;
894 	old->core_temp_c = new->core_temp_c;
895 
896 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
897 		if (mp->format == FORMAT_RAW)
898 			old->counter[i] = new->counter[i];
899 		else
900 			old->counter[i] = new->counter[i] - old->counter[i];
901 	}
902 }
903 
904 /*
905  * old = new - old
906  */
907 int
908 delta_thread(struct thread_data *new, struct thread_data *old,
909 	struct core_data *core_delta)
910 {
911 	int i;
912 	struct msr_counter *mp;
913 
914 	old->tsc = new->tsc - old->tsc;
915 
916 	/* check for TSC < 1 Mcycles over interval */
917 	if (old->tsc < (1000 * 1000))
918 		errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
919 		     "You can disable all c-states by booting with \"idle=poll\"\n"
920 		     "or just the deep ones with \"processor.max_cstate=1\"");
921 
922 	old->c1 = new->c1 - old->c1;
923 
924 	if (has_aperf) {
925 		if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
926 			old->aperf = new->aperf - old->aperf;
927 			old->mperf = new->mperf - old->mperf;
928 		} else {
929 			return -1;
930 		}
931 	}
932 
933 
934 	if (use_c1_residency_msr) {
935 		/*
936 		 * Some models have a dedicated C1 residency MSR,
937 		 * which should be more accurate than the derivation below.
938 		 */
939 	} else {
940 		/*
941 		 * As counter collection is not atomic,
942 		 * it is possible for mperf's non-halted cycles + idle states
943 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
944 		 */
945 		if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
946 			old->c1 = 0;
947 		else {
948 			/* normal case, derive c1 */
949 			old->c1 = old->tsc - old->mperf - core_delta->c3
950 				- core_delta->c6 - core_delta->c7;
951 		}
952 	}
953 
954 	if (old->mperf == 0) {
955 		if (debug > 1)
956 			fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
957 		old->mperf = 1;	/* divide by 0 protection */
958 	}
959 
960 	if (do_irq)
961 		old->irq_count = new->irq_count - old->irq_count;
962 
963 	if (do_smi)
964 		old->smi_count = new->smi_count - old->smi_count;
965 
966 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
967 		if (mp->format == FORMAT_RAW)
968 			old->counter[i] = new->counter[i];
969 		else
970 			old->counter[i] = new->counter[i] - old->counter[i];
971 	}
972 	return 0;
973 }
974 
975 int delta_cpu(struct thread_data *t, struct core_data *c,
976 	struct pkg_data *p, struct thread_data *t2,
977 	struct core_data *c2, struct pkg_data *p2)
978 {
979 	int retval = 0;
980 
981 	/* calculate core delta only for 1st thread in core */
982 	if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
983 		delta_core(c, c2);
984 
985 	/* always calculate thread delta */
986 	retval = delta_thread(t, t2, c2);	/* c2 is core delta */
987 	if (retval)
988 		return retval;
989 
990 	/* calculate package delta only for 1st core in package */
991 	if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
992 		retval = delta_package(p, p2);
993 
994 	return retval;
995 }
996 
997 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
998 {
999 	int i;
1000 	struct msr_counter  *mp;
1001 
1002 	t->tsc = 0;
1003 	t->aperf = 0;
1004 	t->mperf = 0;
1005 	t->c1 = 0;
1006 
1007 	t->irq_count = 0;
1008 	t->smi_count = 0;
1009 
1010 	/* tells format_counters to dump all fields from this set */
1011 	t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
1012 
1013 	c->c3 = 0;
1014 	c->c6 = 0;
1015 	c->c7 = 0;
1016 	c->core_temp_c = 0;
1017 
1018 	p->pkg_wtd_core_c0 = 0;
1019 	p->pkg_any_core_c0 = 0;
1020 	p->pkg_any_gfxe_c0 = 0;
1021 	p->pkg_both_core_gfxe_c0 = 0;
1022 
1023 	p->pc2 = 0;
1024 	if (do_pc3)
1025 		p->pc3 = 0;
1026 	if (do_pc6)
1027 		p->pc6 = 0;
1028 	if (do_pc7)
1029 		p->pc7 = 0;
1030 	p->pc8 = 0;
1031 	p->pc9 = 0;
1032 	p->pc10 = 0;
1033 
1034 	p->energy_pkg = 0;
1035 	p->energy_dram = 0;
1036 	p->energy_cores = 0;
1037 	p->energy_gfx = 0;
1038 	p->rapl_pkg_perf_status = 0;
1039 	p->rapl_dram_perf_status = 0;
1040 	p->pkg_temp_c = 0;
1041 
1042 	p->gfx_rc6_ms = 0;
1043 	p->gfx_mhz = 0;
1044 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
1045 		t->counter[i] = 0;
1046 
1047 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
1048 		c->counter[i] = 0;
1049 
1050 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
1051 		p->counter[i] = 0;
1052 }
1053 int sum_counters(struct thread_data *t, struct core_data *c,
1054 	struct pkg_data *p)
1055 {
1056 	int i;
1057 	struct msr_counter *mp;
1058 
1059 	average.threads.tsc += t->tsc;
1060 	average.threads.aperf += t->aperf;
1061 	average.threads.mperf += t->mperf;
1062 	average.threads.c1 += t->c1;
1063 
1064 	average.threads.irq_count += t->irq_count;
1065 	average.threads.smi_count += t->smi_count;
1066 
1067 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1068 		if (mp->format == FORMAT_RAW)
1069 			continue;
1070 		average.threads.counter[i] += t->counter[i];
1071 	}
1072 
1073 	/* sum per-core values only for 1st thread in core */
1074 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1075 		return 0;
1076 
1077 	average.cores.c3 += c->c3;
1078 	average.cores.c6 += c->c6;
1079 	average.cores.c7 += c->c7;
1080 
1081 	average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
1082 
1083 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1084 		if (mp->format == FORMAT_RAW)
1085 			continue;
1086 		average.cores.counter[i] += c->counter[i];
1087 	}
1088 
1089 	/* sum per-pkg values only for 1st core in pkg */
1090 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1091 		return 0;
1092 
1093 	if (do_skl_residency) {
1094 		average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
1095 		average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
1096 		average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
1097 		average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
1098 	}
1099 
1100 	average.packages.pc2 += p->pc2;
1101 	if (do_pc3)
1102 		average.packages.pc3 += p->pc3;
1103 	if (do_pc6)
1104 		average.packages.pc6 += p->pc6;
1105 	if (do_pc7)
1106 		average.packages.pc7 += p->pc7;
1107 	average.packages.pc8 += p->pc8;
1108 	average.packages.pc9 += p->pc9;
1109 	average.packages.pc10 += p->pc10;
1110 
1111 	average.packages.energy_pkg += p->energy_pkg;
1112 	average.packages.energy_dram += p->energy_dram;
1113 	average.packages.energy_cores += p->energy_cores;
1114 	average.packages.energy_gfx += p->energy_gfx;
1115 
1116 	average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
1117 	average.packages.gfx_mhz = p->gfx_mhz;
1118 
1119 	average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
1120 
1121 	average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
1122 	average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
1123 
1124 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1125 		if (mp->format == FORMAT_RAW)
1126 			continue;
1127 		average.packages.counter[i] += p->counter[i];
1128 	}
1129 	return 0;
1130 }
1131 /*
1132  * sum the counters for all cpus in the system
1133  * compute the weighted average
1134  */
1135 void compute_average(struct thread_data *t, struct core_data *c,
1136 	struct pkg_data *p)
1137 {
1138 	int i;
1139 	struct msr_counter *mp;
1140 
1141 	clear_counters(&average.threads, &average.cores, &average.packages);
1142 
1143 	for_all_cpus(sum_counters, t, c, p);
1144 
1145 	average.threads.tsc /= topo.num_cpus;
1146 	average.threads.aperf /= topo.num_cpus;
1147 	average.threads.mperf /= topo.num_cpus;
1148 	average.threads.c1 /= topo.num_cpus;
1149 
1150 	average.cores.c3 /= topo.num_cores;
1151 	average.cores.c6 /= topo.num_cores;
1152 	average.cores.c7 /= topo.num_cores;
1153 
1154 	if (do_skl_residency) {
1155 		average.packages.pkg_wtd_core_c0 /= topo.num_packages;
1156 		average.packages.pkg_any_core_c0 /= topo.num_packages;
1157 		average.packages.pkg_any_gfxe_c0 /= topo.num_packages;
1158 		average.packages.pkg_both_core_gfxe_c0 /= topo.num_packages;
1159 	}
1160 
1161 	average.packages.pc2 /= topo.num_packages;
1162 	if (do_pc3)
1163 		average.packages.pc3 /= topo.num_packages;
1164 	if (do_pc6)
1165 		average.packages.pc6 /= topo.num_packages;
1166 	if (do_pc7)
1167 		average.packages.pc7 /= topo.num_packages;
1168 
1169 	average.packages.pc8 /= topo.num_packages;
1170 	average.packages.pc9 /= topo.num_packages;
1171 	average.packages.pc10 /= topo.num_packages;
1172 
1173 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1174 		if (mp->format == FORMAT_RAW)
1175 			continue;
1176 		average.threads.counter[i] /= topo.num_cpus;
1177 	}
1178 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1179 		if (mp->format == FORMAT_RAW)
1180 			continue;
1181 		average.cores.counter[i] /= topo.num_cores;
1182 	}
1183 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1184 		if (mp->format == FORMAT_RAW)
1185 			continue;
1186 		average.packages.counter[i] /= topo.num_packages;
1187 	}
1188 }
1189 
1190 static unsigned long long rdtsc(void)
1191 {
1192 	unsigned int low, high;
1193 
1194 	asm volatile("rdtsc" : "=a" (low), "=d" (high));
1195 
1196 	return low | ((unsigned long long)high) << 32;
1197 }
1198 
1199 /*
1200  * get_counters(...)
1201  * migrate to cpu
1202  * acquire and record local counters for that cpu
1203  */
1204 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1205 {
1206 	int cpu = t->cpu_id;
1207 	unsigned long long msr;
1208 	int aperf_mperf_retry_count = 0;
1209 	struct msr_counter *mp;
1210 	int i;
1211 
1212 	if (cpu_migrate(cpu)) {
1213 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
1214 		return -1;
1215 	}
1216 
1217 retry:
1218 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
1219 
1220 	if (has_aperf) {
1221 		unsigned long long tsc_before, tsc_between, tsc_after, aperf_time, mperf_time;
1222 
1223 		/*
1224 		 * The TSC, APERF and MPERF must be read together for
1225 		 * APERF/MPERF and MPERF/TSC to give accurate results.
1226 		 *
1227 		 * Unfortunately, APERF and MPERF are read by
1228 		 * individual system call, so delays may occur
1229 		 * between them.  If the time to read them
1230 		 * varies by a large amount, we re-read them.
1231 		 */
1232 
1233 		/*
1234 		 * This initial dummy APERF read has been seen to
1235 		 * reduce jitter in the subsequent reads.
1236 		 */
1237 
1238 		if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1239 			return -3;
1240 
1241 		t->tsc = rdtsc();	/* re-read close to APERF */
1242 
1243 		tsc_before = t->tsc;
1244 
1245 		if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1246 			return -3;
1247 
1248 		tsc_between = rdtsc();
1249 
1250 		if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
1251 			return -4;
1252 
1253 		tsc_after = rdtsc();
1254 
1255 		aperf_time = tsc_between - tsc_before;
1256 		mperf_time = tsc_after - tsc_between;
1257 
1258 		/*
1259 		 * If the system call latency to read APERF and MPERF
1260 		 * differ by more than 2x, then try again.
1261 		 */
1262 		if ((aperf_time > (2 * mperf_time)) || (mperf_time > (2 * aperf_time))) {
1263 			aperf_mperf_retry_count++;
1264 			if (aperf_mperf_retry_count < 5)
1265 				goto retry;
1266 			else
1267 				warnx("cpu%d jitter %lld %lld",
1268 					cpu, aperf_time, mperf_time);
1269 		}
1270 		aperf_mperf_retry_count = 0;
1271 
1272 		t->aperf = t->aperf * aperf_mperf_multiplier;
1273 		t->mperf = t->mperf * aperf_mperf_multiplier;
1274 	}
1275 
1276 	if (do_irq)
1277 		t->irq_count = irqs_per_cpu[cpu];
1278 	if (do_smi) {
1279 		if (get_msr(cpu, MSR_SMI_COUNT, &msr))
1280 			return -5;
1281 		t->smi_count = msr & 0xFFFFFFFF;
1282 	}
1283 
1284 	if (use_c1_residency_msr) {
1285 		if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
1286 			return -6;
1287 	}
1288 
1289 	for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1290 		if (get_msr(cpu, mp->msr_num, &t->counter[i]))
1291 			return -10;
1292 	}
1293 
1294 
1295 	/* collect core counters only for 1st thread in core */
1296 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1297 		return 0;
1298 
1299 	if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates) {
1300 		if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
1301 			return -6;
1302 	}
1303 
1304 	if (do_nhm_cstates && !do_knl_cstates) {
1305 		if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
1306 			return -7;
1307 	} else if (do_knl_cstates) {
1308 		if (get_msr(cpu, MSR_KNL_CORE_C6_RESIDENCY, &c->c6))
1309 			return -7;
1310 	}
1311 
1312 	if (do_snb_cstates)
1313 		if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
1314 			return -8;
1315 
1316 	if (do_dts) {
1317 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1318 			return -9;
1319 		c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1320 	}
1321 
1322 	for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1323 		if (get_msr(cpu, mp->msr_num, &c->counter[i]))
1324 			return -10;
1325 	}
1326 
1327 	/* collect package counters only for 1st core in package */
1328 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1329 		return 0;
1330 
1331 	if (do_skl_residency) {
1332 		if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
1333 			return -10;
1334 		if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
1335 			return -11;
1336 		if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
1337 			return -12;
1338 		if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
1339 			return -13;
1340 	}
1341 	if (do_pc3)
1342 		if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
1343 			return -9;
1344 	if (do_pc6)
1345 		if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
1346 			return -10;
1347 	if (do_pc2)
1348 		if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
1349 			return -11;
1350 	if (do_pc7)
1351 		if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1352 			return -12;
1353 	if (do_c8_c9_c10) {
1354 		if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1355 			return -13;
1356 		if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1357 			return -13;
1358 		if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1359 			return -13;
1360 	}
1361 	if (do_rapl & RAPL_PKG) {
1362 		if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1363 			return -13;
1364 		p->energy_pkg = msr & 0xFFFFFFFF;
1365 	}
1366 	if (do_rapl & RAPL_CORES_ENERGY_STATUS) {
1367 		if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1368 			return -14;
1369 		p->energy_cores = msr & 0xFFFFFFFF;
1370 	}
1371 	if (do_rapl & RAPL_DRAM) {
1372 		if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1373 			return -15;
1374 		p->energy_dram = msr & 0xFFFFFFFF;
1375 	}
1376 	if (do_rapl & RAPL_GFX) {
1377 		if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1378 			return -16;
1379 		p->energy_gfx = msr & 0xFFFFFFFF;
1380 	}
1381 	if (do_rapl & RAPL_PKG_PERF_STATUS) {
1382 		if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1383 			return -16;
1384 		p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1385 	}
1386 	if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1387 		if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1388 			return -16;
1389 		p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1390 	}
1391 	if (do_ptm) {
1392 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1393 			return -17;
1394 		p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1395 	}
1396 
1397 	if (do_gfx_rc6_ms)
1398 		p->gfx_rc6_ms = gfx_cur_rc6_ms;
1399 
1400 	if (do_gfx_mhz)
1401 		p->gfx_mhz = gfx_cur_mhz;
1402 
1403 	for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1404 		if (get_msr(cpu, mp->msr_num, &p->counter[i]))
1405 			return -10;
1406 	}
1407 
1408 	return 0;
1409 }
1410 
1411 /*
1412  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1413  * If you change the values, note they are used both in comparisons
1414  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1415  */
1416 
1417 #define PCLUKN 0 /* Unknown */
1418 #define PCLRSV 1 /* Reserved */
1419 #define PCL__0 2 /* PC0 */
1420 #define PCL__1 3 /* PC1 */
1421 #define PCL__2 4 /* PC2 */
1422 #define PCL__3 5 /* PC3 */
1423 #define PCL__4 6 /* PC4 */
1424 #define PCL__6 7 /* PC6 */
1425 #define PCL_6N 8 /* PC6 No Retention */
1426 #define PCL_6R 9 /* PC6 Retention */
1427 #define PCL__7 10 /* PC7 */
1428 #define PCL_7S 11 /* PC7 Shrink */
1429 #define PCL__8 12 /* PC8 */
1430 #define PCL__9 13 /* PC9 */
1431 #define PCLUNL 14 /* Unlimited */
1432 
1433 int pkg_cstate_limit = PCLUKN;
1434 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1435 	"pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "unlimited"};
1436 
1437 int nhm_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1438 int snb_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1439 int hsw_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1440 int slv_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1441 int amt_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1442 int phi_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1443 int bxt_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1444 int skx_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1445 
1446 
1447 static void
1448 calculate_tsc_tweak()
1449 {
1450 	tsc_tweak = base_hz / tsc_hz;
1451 }
1452 
1453 static void
1454 dump_nhm_platform_info(void)
1455 {
1456 	unsigned long long msr;
1457 	unsigned int ratio;
1458 
1459 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
1460 
1461 	fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
1462 
1463 	ratio = (msr >> 40) & 0xFF;
1464 	fprintf(outf, "%d * %.0f = %.0f MHz max efficiency frequency\n",
1465 		ratio, bclk, ratio * bclk);
1466 
1467 	ratio = (msr >> 8) & 0xFF;
1468 	fprintf(outf, "%d * %.0f = %.0f MHz base frequency\n",
1469 		ratio, bclk, ratio * bclk);
1470 
1471 	get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
1472 	fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1473 		base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
1474 
1475 	return;
1476 }
1477 
1478 static void
1479 dump_hsw_turbo_ratio_limits(void)
1480 {
1481 	unsigned long long msr;
1482 	unsigned int ratio;
1483 
1484 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
1485 
1486 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
1487 
1488 	ratio = (msr >> 8) & 0xFF;
1489 	if (ratio)
1490 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 18 active cores\n",
1491 			ratio, bclk, ratio * bclk);
1492 
1493 	ratio = (msr >> 0) & 0xFF;
1494 	if (ratio)
1495 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 17 active cores\n",
1496 			ratio, bclk, ratio * bclk);
1497 	return;
1498 }
1499 
1500 static void
1501 dump_ivt_turbo_ratio_limits(void)
1502 {
1503 	unsigned long long msr;
1504 	unsigned int ratio;
1505 
1506 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
1507 
1508 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
1509 
1510 	ratio = (msr >> 56) & 0xFF;
1511 	if (ratio)
1512 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1513 			ratio, bclk, ratio * bclk);
1514 
1515 	ratio = (msr >> 48) & 0xFF;
1516 	if (ratio)
1517 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1518 			ratio, bclk, ratio * bclk);
1519 
1520 	ratio = (msr >> 40) & 0xFF;
1521 	if (ratio)
1522 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1523 			ratio, bclk, ratio * bclk);
1524 
1525 	ratio = (msr >> 32) & 0xFF;
1526 	if (ratio)
1527 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1528 			ratio, bclk, ratio * bclk);
1529 
1530 	ratio = (msr >> 24) & 0xFF;
1531 	if (ratio)
1532 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1533 			ratio, bclk, ratio * bclk);
1534 
1535 	ratio = (msr >> 16) & 0xFF;
1536 	if (ratio)
1537 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1538 			ratio, bclk, ratio * bclk);
1539 
1540 	ratio = (msr >> 8) & 0xFF;
1541 	if (ratio)
1542 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1543 			ratio, bclk, ratio * bclk);
1544 
1545 	ratio = (msr >> 0) & 0xFF;
1546 	if (ratio)
1547 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1548 			ratio, bclk, ratio * bclk);
1549 	return;
1550 }
1551 
1552 static void
1553 dump_nhm_turbo_ratio_limits(void)
1554 {
1555 	unsigned long long msr;
1556 	unsigned int ratio;
1557 
1558 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
1559 
1560 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
1561 
1562 	ratio = (msr >> 56) & 0xFF;
1563 	if (ratio)
1564 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1565 			ratio, bclk, ratio * bclk);
1566 
1567 	ratio = (msr >> 48) & 0xFF;
1568 	if (ratio)
1569 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1570 			ratio, bclk, ratio * bclk);
1571 
1572 	ratio = (msr >> 40) & 0xFF;
1573 	if (ratio)
1574 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1575 			ratio, bclk, ratio * bclk);
1576 
1577 	ratio = (msr >> 32) & 0xFF;
1578 	if (ratio)
1579 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1580 			ratio, bclk, ratio * bclk);
1581 
1582 	ratio = (msr >> 24) & 0xFF;
1583 	if (ratio)
1584 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1585 			ratio, bclk, ratio * bclk);
1586 
1587 	ratio = (msr >> 16) & 0xFF;
1588 	if (ratio)
1589 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1590 			ratio, bclk, ratio * bclk);
1591 
1592 	ratio = (msr >> 8) & 0xFF;
1593 	if (ratio)
1594 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1595 			ratio, bclk, ratio * bclk);
1596 
1597 	ratio = (msr >> 0) & 0xFF;
1598 	if (ratio)
1599 		fprintf(outf, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1600 			ratio, bclk, ratio * bclk);
1601 	return;
1602 }
1603 
1604 static void
1605 dump_knl_turbo_ratio_limits(void)
1606 {
1607 	const unsigned int buckets_no = 7;
1608 
1609 	unsigned long long msr;
1610 	int delta_cores, delta_ratio;
1611 	int i, b_nr;
1612 	unsigned int cores[buckets_no];
1613 	unsigned int ratio[buckets_no];
1614 
1615 	get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
1616 
1617 	fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n",
1618 		base_cpu, msr);
1619 
1620 	/**
1621 	 * Turbo encoding in KNL is as follows:
1622 	 * [0] -- Reserved
1623 	 * [7:1] -- Base value of number of active cores of bucket 1.
1624 	 * [15:8] -- Base value of freq ratio of bucket 1.
1625 	 * [20:16] -- +ve delta of number of active cores of bucket 2.
1626 	 * i.e. active cores of bucket 2 =
1627 	 * active cores of bucket 1 + delta
1628 	 * [23:21] -- Negative delta of freq ratio of bucket 2.
1629 	 * i.e. freq ratio of bucket 2 =
1630 	 * freq ratio of bucket 1 - delta
1631 	 * [28:24]-- +ve delta of number of active cores of bucket 3.
1632 	 * [31:29]-- -ve delta of freq ratio of bucket 3.
1633 	 * [36:32]-- +ve delta of number of active cores of bucket 4.
1634 	 * [39:37]-- -ve delta of freq ratio of bucket 4.
1635 	 * [44:40]-- +ve delta of number of active cores of bucket 5.
1636 	 * [47:45]-- -ve delta of freq ratio of bucket 5.
1637 	 * [52:48]-- +ve delta of number of active cores of bucket 6.
1638 	 * [55:53]-- -ve delta of freq ratio of bucket 6.
1639 	 * [60:56]-- +ve delta of number of active cores of bucket 7.
1640 	 * [63:61]-- -ve delta of freq ratio of bucket 7.
1641 	 */
1642 
1643 	b_nr = 0;
1644 	cores[b_nr] = (msr & 0xFF) >> 1;
1645 	ratio[b_nr] = (msr >> 8) & 0xFF;
1646 
1647 	for (i = 16; i < 64; i += 8) {
1648 		delta_cores = (msr >> i) & 0x1F;
1649 		delta_ratio = (msr >> (i + 5)) & 0x7;
1650 
1651 		cores[b_nr + 1] = cores[b_nr] + delta_cores;
1652 		ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
1653 		b_nr++;
1654 	}
1655 
1656 	for (i = buckets_no - 1; i >= 0; i--)
1657 		if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
1658 			fprintf(outf,
1659 				"%d * %.0f = %.0f MHz max turbo %d active cores\n",
1660 				ratio[i], bclk, ratio[i] * bclk, cores[i]);
1661 }
1662 
1663 static void
1664 dump_nhm_cst_cfg(void)
1665 {
1666 	unsigned long long msr;
1667 
1668 	get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1669 
1670 #define SNB_C1_AUTO_UNDEMOTE              (1UL << 27)
1671 #define SNB_C3_AUTO_UNDEMOTE              (1UL << 28)
1672 
1673 	fprintf(outf, "cpu%d: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", base_cpu, msr);
1674 
1675 	fprintf(outf, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: %s)\n",
1676 		(msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1677 		(msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1678 		(msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1679 		(msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1680 		(msr & (1 << 15)) ? "" : "UN",
1681 		(unsigned int)msr & 0xF,
1682 		pkg_cstate_limit_strings[pkg_cstate_limit]);
1683 	return;
1684 }
1685 
1686 static void
1687 dump_config_tdp(void)
1688 {
1689 	unsigned long long msr;
1690 
1691 	get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
1692 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
1693 	fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
1694 
1695 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
1696 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
1697 	if (msr) {
1698 		fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
1699 		fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
1700 		fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
1701 		fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
1702 	}
1703 	fprintf(outf, ")\n");
1704 
1705 	get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
1706 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
1707 	if (msr) {
1708 		fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
1709 		fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
1710 		fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
1711 		fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
1712 	}
1713 	fprintf(outf, ")\n");
1714 
1715 	get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
1716 	fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
1717 	if ((msr) & 0x3)
1718 		fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
1719 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
1720 	fprintf(outf, ")\n");
1721 
1722 	get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
1723 	fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
1724 	fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
1725 	fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
1726 	fprintf(outf, ")\n");
1727 }
1728 
1729 unsigned int irtl_time_units[] = {1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
1730 
1731 void print_irtl(void)
1732 {
1733 	unsigned long long msr;
1734 
1735 	get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
1736 	fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
1737 	fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1738 		(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1739 
1740 	get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
1741 	fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
1742 	fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1743 		(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1744 
1745 	get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
1746 	fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
1747 	fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1748 		(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1749 
1750 	if (!do_irtl_hsw)
1751 		return;
1752 
1753 	get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
1754 	fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
1755 	fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1756 		(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1757 
1758 	get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
1759 	fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
1760 	fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1761 		(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1762 
1763 	get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
1764 	fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
1765 	fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
1766 		(msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
1767 
1768 }
1769 void free_fd_percpu(void)
1770 {
1771 	int i;
1772 
1773 	for (i = 0; i < topo.max_cpu_num + 1; ++i) {
1774 		if (fd_percpu[i] != 0)
1775 			close(fd_percpu[i]);
1776 	}
1777 
1778 	free(fd_percpu);
1779 }
1780 
1781 void free_all_buffers(void)
1782 {
1783 	CPU_FREE(cpu_present_set);
1784 	cpu_present_set = NULL;
1785 	cpu_present_setsize = 0;
1786 
1787 	CPU_FREE(cpu_affinity_set);
1788 	cpu_affinity_set = NULL;
1789 	cpu_affinity_setsize = 0;
1790 
1791 	free(thread_even);
1792 	free(core_even);
1793 	free(package_even);
1794 
1795 	thread_even = NULL;
1796 	core_even = NULL;
1797 	package_even = NULL;
1798 
1799 	free(thread_odd);
1800 	free(core_odd);
1801 	free(package_odd);
1802 
1803 	thread_odd = NULL;
1804 	core_odd = NULL;
1805 	package_odd = NULL;
1806 
1807 	free(output_buffer);
1808 	output_buffer = NULL;
1809 	outp = NULL;
1810 
1811 	free_fd_percpu();
1812 
1813 	free(irq_column_2_cpu);
1814 	free(irqs_per_cpu);
1815 }
1816 
1817 /*
1818  * Open a file, and exit on failure
1819  */
1820 FILE *fopen_or_die(const char *path, const char *mode)
1821 {
1822 	FILE *filep = fopen(path, mode);
1823 	if (!filep)
1824 		err(1, "%s: open failed", path);
1825 	return filep;
1826 }
1827 
1828 /*
1829  * Parse a file containing a single int.
1830  */
1831 int parse_int_file(const char *fmt, ...)
1832 {
1833 	va_list args;
1834 	char path[PATH_MAX];
1835 	FILE *filep;
1836 	int value;
1837 
1838 	va_start(args, fmt);
1839 	vsnprintf(path, sizeof(path), fmt, args);
1840 	va_end(args);
1841 	filep = fopen_or_die(path, "r");
1842 	if (fscanf(filep, "%d", &value) != 1)
1843 		err(1, "%s: failed to parse number from file", path);
1844 	fclose(filep);
1845 	return value;
1846 }
1847 
1848 /*
1849  * get_cpu_position_in_core(cpu)
1850  * return the position of the CPU among its HT siblings in the core
1851  * return -1 if the sibling is not in list
1852  */
1853 int get_cpu_position_in_core(int cpu)
1854 {
1855 	char path[64];
1856 	FILE *filep;
1857 	int this_cpu;
1858 	char character;
1859 	int i;
1860 
1861 	sprintf(path,
1862 		"/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list",
1863 		cpu);
1864 	filep = fopen(path, "r");
1865 	if (filep == NULL) {
1866 		perror(path);
1867 		exit(1);
1868 	}
1869 
1870 	for (i = 0; i < topo.num_threads_per_core; i++) {
1871 		fscanf(filep, "%d", &this_cpu);
1872 		if (this_cpu == cpu) {
1873 			fclose(filep);
1874 			return i;
1875 		}
1876 
1877 		/* Account for no separator after last thread*/
1878 		if (i != (topo.num_threads_per_core - 1))
1879 			fscanf(filep, "%c", &character);
1880 	}
1881 
1882 	fclose(filep);
1883 	return -1;
1884 }
1885 
1886 /*
1887  * cpu_is_first_core_in_package(cpu)
1888  * return 1 if given CPU is 1st core in package
1889  */
1890 int cpu_is_first_core_in_package(int cpu)
1891 {
1892 	return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1893 }
1894 
1895 int get_physical_package_id(int cpu)
1896 {
1897 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1898 }
1899 
1900 int get_core_id(int cpu)
1901 {
1902 	return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1903 }
1904 
1905 int get_num_ht_siblings(int cpu)
1906 {
1907 	char path[80];
1908 	FILE *filep;
1909 	int sib1;
1910 	int matches = 0;
1911 	char character;
1912 	char str[100];
1913 	char *ch;
1914 
1915 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1916 	filep = fopen_or_die(path, "r");
1917 
1918 	/*
1919 	 * file format:
1920 	 * A ',' separated or '-' separated set of numbers
1921 	 * (eg 1-2 or 1,3,4,5)
1922 	 */
1923 	fscanf(filep, "%d%c\n", &sib1, &character);
1924 	fseek(filep, 0, SEEK_SET);
1925 	fgets(str, 100, filep);
1926 	ch = strchr(str, character);
1927 	while (ch != NULL) {
1928 		matches++;
1929 		ch = strchr(ch+1, character);
1930 	}
1931 
1932 	fclose(filep);
1933 	return matches+1;
1934 }
1935 
1936 /*
1937  * run func(thread, core, package) in topology order
1938  * skip non-present cpus
1939  */
1940 
1941 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1942 	struct pkg_data *, struct thread_data *, struct core_data *,
1943 	struct pkg_data *), struct thread_data *thread_base,
1944 	struct core_data *core_base, struct pkg_data *pkg_base,
1945 	struct thread_data *thread_base2, struct core_data *core_base2,
1946 	struct pkg_data *pkg_base2)
1947 {
1948 	int retval, pkg_no, core_no, thread_no;
1949 
1950 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1951 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1952 			for (thread_no = 0; thread_no <
1953 				topo.num_threads_per_core; ++thread_no) {
1954 				struct thread_data *t, *t2;
1955 				struct core_data *c, *c2;
1956 				struct pkg_data *p, *p2;
1957 
1958 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1959 
1960 				if (cpu_is_not_present(t->cpu_id))
1961 					continue;
1962 
1963 				t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1964 
1965 				c = GET_CORE(core_base, core_no, pkg_no);
1966 				c2 = GET_CORE(core_base2, core_no, pkg_no);
1967 
1968 				p = GET_PKG(pkg_base, pkg_no);
1969 				p2 = GET_PKG(pkg_base2, pkg_no);
1970 
1971 				retval = func(t, c, p, t2, c2, p2);
1972 				if (retval)
1973 					return retval;
1974 			}
1975 		}
1976 	}
1977 	return 0;
1978 }
1979 
1980 /*
1981  * run func(cpu) on every cpu in /proc/stat
1982  * return max_cpu number
1983  */
1984 int for_all_proc_cpus(int (func)(int))
1985 {
1986 	FILE *fp;
1987 	int cpu_num;
1988 	int retval;
1989 
1990 	fp = fopen_or_die(proc_stat, "r");
1991 
1992 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1993 	if (retval != 0)
1994 		err(1, "%s: failed to parse format", proc_stat);
1995 
1996 	while (1) {
1997 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1998 		if (retval != 1)
1999 			break;
2000 
2001 		retval = func(cpu_num);
2002 		if (retval) {
2003 			fclose(fp);
2004 			return(retval);
2005 		}
2006 	}
2007 	fclose(fp);
2008 	return 0;
2009 }
2010 
2011 void re_initialize(void)
2012 {
2013 	free_all_buffers();
2014 	setup_all_buffers();
2015 	printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
2016 }
2017 
2018 
2019 /*
2020  * count_cpus()
2021  * remember the last one seen, it will be the max
2022  */
2023 int count_cpus(int cpu)
2024 {
2025 	if (topo.max_cpu_num < cpu)
2026 		topo.max_cpu_num = cpu;
2027 
2028 	topo.num_cpus += 1;
2029 	return 0;
2030 }
2031 int mark_cpu_present(int cpu)
2032 {
2033 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
2034 	return 0;
2035 }
2036 
2037 /*
2038  * snapshot_proc_interrupts()
2039  *
2040  * read and record summary of /proc/interrupts
2041  *
2042  * return 1 if config change requires a restart, else return 0
2043  */
2044 int snapshot_proc_interrupts(void)
2045 {
2046 	static FILE *fp;
2047 	int column, retval;
2048 
2049 	if (fp == NULL)
2050 		fp = fopen_or_die("/proc/interrupts", "r");
2051 	else
2052 		rewind(fp);
2053 
2054 	/* read 1st line of /proc/interrupts to get cpu* name for each column */
2055 	for (column = 0; column < topo.num_cpus; ++column) {
2056 		int cpu_number;
2057 
2058 		retval = fscanf(fp, " CPU%d", &cpu_number);
2059 		if (retval != 1)
2060 			break;
2061 
2062 		if (cpu_number > topo.max_cpu_num) {
2063 			warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
2064 			return 1;
2065 		}
2066 
2067 		irq_column_2_cpu[column] = cpu_number;
2068 		irqs_per_cpu[cpu_number] = 0;
2069 	}
2070 
2071 	/* read /proc/interrupt count lines and sum up irqs per cpu */
2072 	while (1) {
2073 		int column;
2074 		char buf[64];
2075 
2076 		retval = fscanf(fp, " %s:", buf);	/* flush irq# "N:" */
2077 		if (retval != 1)
2078 			break;
2079 
2080 		/* read the count per cpu */
2081 		for (column = 0; column < topo.num_cpus; ++column) {
2082 
2083 			int cpu_number, irq_count;
2084 
2085 			retval = fscanf(fp, " %d", &irq_count);
2086 			if (retval != 1)
2087 				break;
2088 
2089 			cpu_number = irq_column_2_cpu[column];
2090 			irqs_per_cpu[cpu_number] += irq_count;
2091 
2092 		}
2093 
2094 		while (getc(fp) != '\n')
2095 			;	/* flush interrupt description */
2096 
2097 	}
2098 	return 0;
2099 }
2100 /*
2101  * snapshot_gfx_rc6_ms()
2102  *
2103  * record snapshot of
2104  * /sys/class/drm/card0/power/rc6_residency_ms
2105  *
2106  * return 1 if config change requires a restart, else return 0
2107  */
2108 int snapshot_gfx_rc6_ms(void)
2109 {
2110 	FILE *fp;
2111 	int retval;
2112 
2113 	fp = fopen_or_die("/sys/class/drm/card0/power/rc6_residency_ms", "r");
2114 
2115 	retval = fscanf(fp, "%lld", &gfx_cur_rc6_ms);
2116 	if (retval != 1)
2117 		err(1, "GFX rc6");
2118 
2119 	fclose(fp);
2120 
2121 	return 0;
2122 }
2123 /*
2124  * snapshot_gfx_mhz()
2125  *
2126  * record snapshot of
2127  * /sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz
2128  *
2129  * return 1 if config change requires a restart, else return 0
2130  */
2131 int snapshot_gfx_mhz(void)
2132 {
2133 	static FILE *fp;
2134 	int retval;
2135 
2136 	if (fp == NULL)
2137 		fp = fopen_or_die("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", "r");
2138 	else
2139 		rewind(fp);
2140 
2141 	retval = fscanf(fp, "%d", &gfx_cur_mhz);
2142 	if (retval != 1)
2143 		err(1, "GFX MHz");
2144 
2145 	return 0;
2146 }
2147 
2148 /*
2149  * snapshot /proc and /sys files
2150  *
2151  * return 1 if configuration restart needed, else return 0
2152  */
2153 int snapshot_proc_sysfs_files(void)
2154 {
2155 	if (snapshot_proc_interrupts())
2156 		return 1;
2157 
2158 	if (do_gfx_rc6_ms)
2159 		snapshot_gfx_rc6_ms();
2160 
2161 	if (do_gfx_mhz)
2162 		snapshot_gfx_mhz();
2163 
2164 	return 0;
2165 }
2166 
2167 void turbostat_loop()
2168 {
2169 	int retval;
2170 	int restarted = 0;
2171 
2172 restart:
2173 	restarted++;
2174 
2175 	snapshot_proc_sysfs_files();
2176 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
2177 	if (retval < -1) {
2178 		exit(retval);
2179 	} else if (retval == -1) {
2180 		if (restarted > 1) {
2181 			exit(retval);
2182 		}
2183 		re_initialize();
2184 		goto restart;
2185 	}
2186 	restarted = 0;
2187 	gettimeofday(&tv_even, (struct timezone *)NULL);
2188 
2189 	while (1) {
2190 		if (for_all_proc_cpus(cpu_is_not_present)) {
2191 			re_initialize();
2192 			goto restart;
2193 		}
2194 		nanosleep(&interval_ts, NULL);
2195 		if (snapshot_proc_sysfs_files())
2196 			goto restart;
2197 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
2198 		if (retval < -1) {
2199 			exit(retval);
2200 		} else if (retval == -1) {
2201 			re_initialize();
2202 			goto restart;
2203 		}
2204 		gettimeofday(&tv_odd, (struct timezone *)NULL);
2205 		timersub(&tv_odd, &tv_even, &tv_delta);
2206 		if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
2207 			re_initialize();
2208 			goto restart;
2209 		}
2210 		compute_average(EVEN_COUNTERS);
2211 		format_all_counters(EVEN_COUNTERS);
2212 		flush_output_stdout();
2213 		nanosleep(&interval_ts, NULL);
2214 		if (snapshot_proc_sysfs_files())
2215 			goto restart;
2216 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
2217 		if (retval < -1) {
2218 			exit(retval);
2219 		} else if (retval == -1) {
2220 			re_initialize();
2221 			goto restart;
2222 		}
2223 		gettimeofday(&tv_even, (struct timezone *)NULL);
2224 		timersub(&tv_even, &tv_odd, &tv_delta);
2225 		if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
2226 			re_initialize();
2227 			goto restart;
2228 		}
2229 		compute_average(ODD_COUNTERS);
2230 		format_all_counters(ODD_COUNTERS);
2231 		flush_output_stdout();
2232 	}
2233 }
2234 
2235 void check_dev_msr()
2236 {
2237 	struct stat sb;
2238 	char pathname[32];
2239 
2240 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
2241 	if (stat(pathname, &sb))
2242  		if (system("/sbin/modprobe msr > /dev/null 2>&1"))
2243 			err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
2244 }
2245 
2246 void check_permissions()
2247 {
2248 	struct __user_cap_header_struct cap_header_data;
2249 	cap_user_header_t cap_header = &cap_header_data;
2250 	struct __user_cap_data_struct cap_data_data;
2251 	cap_user_data_t cap_data = &cap_data_data;
2252 	extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
2253 	int do_exit = 0;
2254 	char pathname[32];
2255 
2256 	/* check for CAP_SYS_RAWIO */
2257 	cap_header->pid = getpid();
2258 	cap_header->version = _LINUX_CAPABILITY_VERSION;
2259 	if (capget(cap_header, cap_data) < 0)
2260 		err(-6, "capget(2) failed");
2261 
2262 	if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
2263 		do_exit++;
2264 		warnx("capget(CAP_SYS_RAWIO) failed,"
2265 			" try \"# setcap cap_sys_rawio=ep %s\"", progname);
2266 	}
2267 
2268 	/* test file permissions */
2269 	sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
2270 	if (euidaccess(pathname, R_OK)) {
2271 		do_exit++;
2272 		warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
2273 	}
2274 
2275 	/* if all else fails, thell them to be root */
2276 	if (do_exit)
2277 		if (getuid() != 0)
2278 			warnx("... or simply run as root");
2279 
2280 	if (do_exit)
2281 		exit(-6);
2282 }
2283 
2284 /*
2285  * NHM adds support for additional MSRs:
2286  *
2287  * MSR_SMI_COUNT                   0x00000034
2288  *
2289  * MSR_PLATFORM_INFO               0x000000ce
2290  * MSR_NHM_SNB_PKG_CST_CFG_CTL     0x000000e2
2291  *
2292  * MSR_PKG_C3_RESIDENCY            0x000003f8
2293  * MSR_PKG_C6_RESIDENCY            0x000003f9
2294  * MSR_CORE_C3_RESIDENCY           0x000003fc
2295  * MSR_CORE_C6_RESIDENCY           0x000003fd
2296  *
2297  * Side effect:
2298  * sets global pkg_cstate_limit to decode MSR_NHM_SNB_PKG_CST_CFG_CTL
2299  */
2300 int probe_nhm_msrs(unsigned int family, unsigned int model)
2301 {
2302 	unsigned long long msr;
2303 	unsigned int base_ratio;
2304 	int *pkg_cstate_limits;
2305 
2306 	if (!genuine_intel)
2307 		return 0;
2308 
2309 	if (family != 6)
2310 		return 0;
2311 
2312 	bclk = discover_bclk(family, model);
2313 
2314 	switch (model) {
2315 	case INTEL_FAM6_NEHALEM_EP:	/* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
2316 	case INTEL_FAM6_NEHALEM:	/* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
2317 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
2318 	case INTEL_FAM6_WESTMERE:	/* Westmere Client - Clarkdale, Arrandale */
2319 	case INTEL_FAM6_WESTMERE_EP:	/* Westmere EP - Gulftown */
2320 	case INTEL_FAM6_NEHALEM_EX:	/* Nehalem-EX Xeon - Beckton */
2321 	case INTEL_FAM6_WESTMERE_EX:	/* Westmere-EX Xeon - Eagleton */
2322 		pkg_cstate_limits = nhm_pkg_cstate_limits;
2323 		break;
2324 	case INTEL_FAM6_SANDYBRIDGE:	/* SNB */
2325 	case INTEL_FAM6_SANDYBRIDGE_X:	/* SNB Xeon */
2326 	case INTEL_FAM6_IVYBRIDGE:	/* IVB */
2327 	case INTEL_FAM6_IVYBRIDGE_X:	/* IVB Xeon */
2328 		pkg_cstate_limits = snb_pkg_cstate_limits;
2329 		break;
2330 	case INTEL_FAM6_HASWELL_CORE:	/* HSW */
2331 	case INTEL_FAM6_HASWELL_X:	/* HSX */
2332 	case INTEL_FAM6_HASWELL_ULT:	/* HSW */
2333 	case INTEL_FAM6_HASWELL_GT3E:	/* HSW */
2334 	case INTEL_FAM6_BROADWELL_CORE:	/* BDW */
2335 	case INTEL_FAM6_BROADWELL_GT3E:	/* BDW */
2336 	case INTEL_FAM6_BROADWELL_X:	/* BDX */
2337 	case INTEL_FAM6_BROADWELL_XEON_D:	/* BDX-DE */
2338 	case INTEL_FAM6_SKYLAKE_MOBILE:	/* SKL */
2339 	case INTEL_FAM6_SKYLAKE_DESKTOP:	/* SKL */
2340 	case INTEL_FAM6_KABYLAKE_MOBILE:	/* KBL */
2341 	case INTEL_FAM6_KABYLAKE_DESKTOP:	/* KBL */
2342 		pkg_cstate_limits = hsw_pkg_cstate_limits;
2343 		break;
2344 	case INTEL_FAM6_SKYLAKE_X:	/* SKX */
2345 		pkg_cstate_limits = skx_pkg_cstate_limits;
2346 		break;
2347 	case INTEL_FAM6_ATOM_SILVERMONT1:	/* BYT */
2348 	case INTEL_FAM6_ATOM_SILVERMONT2:	/* AVN */
2349 		pkg_cstate_limits = slv_pkg_cstate_limits;
2350 		break;
2351 	case INTEL_FAM6_ATOM_AIRMONT:	/* AMT */
2352 		pkg_cstate_limits = amt_pkg_cstate_limits;
2353 		break;
2354 	case INTEL_FAM6_XEON_PHI_KNL:	/* PHI */
2355 	case INTEL_FAM6_XEON_PHI_KNM:
2356 		pkg_cstate_limits = phi_pkg_cstate_limits;
2357 		break;
2358 	case INTEL_FAM6_ATOM_GOLDMONT:	/* BXT */
2359 	case INTEL_FAM6_ATOM_DENVERTON:	/* DNV */
2360 		pkg_cstate_limits = bxt_pkg_cstate_limits;
2361 		break;
2362 	default:
2363 		return 0;
2364 	}
2365 	get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
2366 	pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
2367 
2368 	get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
2369 	base_ratio = (msr >> 8) & 0xFF;
2370 
2371 	base_hz = base_ratio * bclk * 1000000;
2372 	has_base_hz = 1;
2373 	return 1;
2374 }
2375 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
2376 {
2377 	switch (model) {
2378 	/* Nehalem compatible, but do not include turbo-ratio limit support */
2379 	case INTEL_FAM6_NEHALEM_EX:	/* Nehalem-EX Xeon - Beckton */
2380 	case INTEL_FAM6_WESTMERE_EX:	/* Westmere-EX Xeon - Eagleton */
2381 	case INTEL_FAM6_XEON_PHI_KNL:	/* PHI - Knights Landing (different MSR definition) */
2382 	case INTEL_FAM6_XEON_PHI_KNM:
2383 		return 0;
2384 	default:
2385 		return 1;
2386 	}
2387 }
2388 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
2389 {
2390 	if (!genuine_intel)
2391 		return 0;
2392 
2393 	if (family != 6)
2394 		return 0;
2395 
2396 	switch (model) {
2397 	case INTEL_FAM6_IVYBRIDGE_X:	/* IVB Xeon */
2398 	case INTEL_FAM6_HASWELL_X:	/* HSW Xeon */
2399 		return 1;
2400 	default:
2401 		return 0;
2402 	}
2403 }
2404 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
2405 {
2406 	if (!genuine_intel)
2407 		return 0;
2408 
2409 	if (family != 6)
2410 		return 0;
2411 
2412 	switch (model) {
2413 	case INTEL_FAM6_HASWELL_X:	/* HSW Xeon */
2414 		return 1;
2415 	default:
2416 		return 0;
2417 	}
2418 }
2419 
2420 int has_knl_turbo_ratio_limit(unsigned int family, unsigned int model)
2421 {
2422 	if (!genuine_intel)
2423 		return 0;
2424 
2425 	if (family != 6)
2426 		return 0;
2427 
2428 	switch (model) {
2429 	case INTEL_FAM6_XEON_PHI_KNL:	/* Knights Landing */
2430 	case INTEL_FAM6_XEON_PHI_KNM:
2431 		return 1;
2432 	default:
2433 		return 0;
2434 	}
2435 }
2436 int has_config_tdp(unsigned int family, unsigned int model)
2437 {
2438 	if (!genuine_intel)
2439 		return 0;
2440 
2441 	if (family != 6)
2442 		return 0;
2443 
2444 	switch (model) {
2445 	case INTEL_FAM6_IVYBRIDGE:	/* IVB */
2446 	case INTEL_FAM6_HASWELL_CORE:	/* HSW */
2447 	case INTEL_FAM6_HASWELL_X:	/* HSX */
2448 	case INTEL_FAM6_HASWELL_ULT:	/* HSW */
2449 	case INTEL_FAM6_HASWELL_GT3E:	/* HSW */
2450 	case INTEL_FAM6_BROADWELL_CORE:	/* BDW */
2451 	case INTEL_FAM6_BROADWELL_GT3E:	/* BDW */
2452 	case INTEL_FAM6_BROADWELL_X:	/* BDX */
2453 	case INTEL_FAM6_BROADWELL_XEON_D:	/* BDX-DE */
2454 	case INTEL_FAM6_SKYLAKE_MOBILE:	/* SKL */
2455 	case INTEL_FAM6_SKYLAKE_DESKTOP:	/* SKL */
2456 	case INTEL_FAM6_KABYLAKE_MOBILE:	/* KBL */
2457 	case INTEL_FAM6_KABYLAKE_DESKTOP:	/* KBL */
2458 	case INTEL_FAM6_SKYLAKE_X:	/* SKX */
2459 
2460 	case INTEL_FAM6_XEON_PHI_KNL:	/* Knights Landing */
2461 	case INTEL_FAM6_XEON_PHI_KNM:
2462 		return 1;
2463 	default:
2464 		return 0;
2465 	}
2466 }
2467 
2468 static void
2469 dump_cstate_pstate_config_info(unsigned int family, unsigned int model)
2470 {
2471 	if (!do_nhm_platform_info)
2472 		return;
2473 
2474 	dump_nhm_platform_info();
2475 
2476 	if (has_hsw_turbo_ratio_limit(family, model))
2477 		dump_hsw_turbo_ratio_limits();
2478 
2479 	if (has_ivt_turbo_ratio_limit(family, model))
2480 		dump_ivt_turbo_ratio_limits();
2481 
2482 	if (has_nhm_turbo_ratio_limit(family, model))
2483 		dump_nhm_turbo_ratio_limits();
2484 
2485 	if (has_knl_turbo_ratio_limit(family, model))
2486 		dump_knl_turbo_ratio_limits();
2487 
2488 	if (has_config_tdp(family, model))
2489 		dump_config_tdp();
2490 
2491 	dump_nhm_cst_cfg();
2492 }
2493 
2494 
2495 /*
2496  * print_epb()
2497  * Decode the ENERGY_PERF_BIAS MSR
2498  */
2499 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2500 {
2501 	unsigned long long msr;
2502 	char *epb_string;
2503 	int cpu;
2504 
2505 	if (!has_epb)
2506 		return 0;
2507 
2508 	cpu = t->cpu_id;
2509 
2510 	/* EPB is per-package */
2511 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2512 		return 0;
2513 
2514 	if (cpu_migrate(cpu)) {
2515 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2516 		return -1;
2517 	}
2518 
2519 	if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
2520 		return 0;
2521 
2522 	switch (msr & 0xF) {
2523 	case ENERGY_PERF_BIAS_PERFORMANCE:
2524 		epb_string = "performance";
2525 		break;
2526 	case ENERGY_PERF_BIAS_NORMAL:
2527 		epb_string = "balanced";
2528 		break;
2529 	case ENERGY_PERF_BIAS_POWERSAVE:
2530 		epb_string = "powersave";
2531 		break;
2532 	default:
2533 		epb_string = "custom";
2534 		break;
2535 	}
2536 	fprintf(outf, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
2537 
2538 	return 0;
2539 }
2540 /*
2541  * print_hwp()
2542  * Decode the MSR_HWP_CAPABILITIES
2543  */
2544 int print_hwp(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2545 {
2546 	unsigned long long msr;
2547 	int cpu;
2548 
2549 	if (!has_hwp)
2550 		return 0;
2551 
2552 	cpu = t->cpu_id;
2553 
2554 	/* MSR_HWP_CAPABILITIES is per-package */
2555 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2556 		return 0;
2557 
2558 	if (cpu_migrate(cpu)) {
2559 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2560 		return -1;
2561 	}
2562 
2563 	if (get_msr(cpu, MSR_PM_ENABLE, &msr))
2564 		return 0;
2565 
2566 	fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n",
2567 		cpu, msr, (msr & (1 << 0)) ? "" : "No-");
2568 
2569 	/* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
2570 	if ((msr & (1 << 0)) == 0)
2571 		return 0;
2572 
2573 	if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
2574 		return 0;
2575 
2576 	fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
2577 			"(high 0x%x guar 0x%x eff 0x%x low 0x%x)\n",
2578 			cpu, msr,
2579 			(unsigned int)HWP_HIGHEST_PERF(msr),
2580 			(unsigned int)HWP_GUARANTEED_PERF(msr),
2581 			(unsigned int)HWP_MOSTEFFICIENT_PERF(msr),
2582 			(unsigned int)HWP_LOWEST_PERF(msr));
2583 
2584 	if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
2585 		return 0;
2586 
2587 	fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
2588 			"(min 0x%x max 0x%x des 0x%x epp 0x%x window 0x%x pkg 0x%x)\n",
2589 			cpu, msr,
2590 			(unsigned int)(((msr) >> 0) & 0xff),
2591 			(unsigned int)(((msr) >> 8) & 0xff),
2592 			(unsigned int)(((msr) >> 16) & 0xff),
2593 			(unsigned int)(((msr) >> 24) & 0xff),
2594 			(unsigned int)(((msr) >> 32) & 0xff3),
2595 			(unsigned int)(((msr) >> 42) & 0x1));
2596 
2597 	if (has_hwp_pkg) {
2598 		if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
2599 			return 0;
2600 
2601 		fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
2602 			"(min 0x%x max 0x%x des 0x%x epp 0x%x window 0x%x)\n",
2603 			cpu, msr,
2604 			(unsigned int)(((msr) >> 0) & 0xff),
2605 			(unsigned int)(((msr) >> 8) & 0xff),
2606 			(unsigned int)(((msr) >> 16) & 0xff),
2607 			(unsigned int)(((msr) >> 24) & 0xff),
2608 			(unsigned int)(((msr) >> 32) & 0xff3));
2609 	}
2610 	if (has_hwp_notify) {
2611 		if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
2612 			return 0;
2613 
2614 		fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
2615 			"(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
2616 			cpu, msr,
2617 			((msr) & 0x1) ? "EN" : "Dis",
2618 			((msr) & 0x2) ? "EN" : "Dis");
2619 	}
2620 	if (get_msr(cpu, MSR_HWP_STATUS, &msr))
2621 		return 0;
2622 
2623 	fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
2624 			"(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
2625 			cpu, msr,
2626 			((msr) & 0x1) ? "" : "No-",
2627 			((msr) & 0x2) ? "" : "No-");
2628 
2629 	return 0;
2630 }
2631 
2632 /*
2633  * print_perf_limit()
2634  */
2635 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2636 {
2637 	unsigned long long msr;
2638 	int cpu;
2639 
2640 	cpu = t->cpu_id;
2641 
2642 	/* per-package */
2643 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2644 		return 0;
2645 
2646 	if (cpu_migrate(cpu)) {
2647 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2648 		return -1;
2649 	}
2650 
2651 	if (do_core_perf_limit_reasons) {
2652 		get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
2653 		fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2654 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
2655 			(msr & 1 << 15) ? "bit15, " : "",
2656 			(msr & 1 << 14) ? "bit14, " : "",
2657 			(msr & 1 << 13) ? "Transitions, " : "",
2658 			(msr & 1 << 12) ? "MultiCoreTurbo, " : "",
2659 			(msr & 1 << 11) ? "PkgPwrL2, " : "",
2660 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
2661 			(msr & 1 << 9) ? "CorePwr, " : "",
2662 			(msr & 1 << 8) ? "Amps, " : "",
2663 			(msr & 1 << 6) ? "VR-Therm, " : "",
2664 			(msr & 1 << 5) ? "Auto-HWP, " : "",
2665 			(msr & 1 << 4) ? "Graphics, " : "",
2666 			(msr & 1 << 2) ? "bit2, " : "",
2667 			(msr & 1 << 1) ? "ThermStatus, " : "",
2668 			(msr & 1 << 0) ? "PROCHOT, " : "");
2669 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
2670 			(msr & 1 << 31) ? "bit31, " : "",
2671 			(msr & 1 << 30) ? "bit30, " : "",
2672 			(msr & 1 << 29) ? "Transitions, " : "",
2673 			(msr & 1 << 28) ? "MultiCoreTurbo, " : "",
2674 			(msr & 1 << 27) ? "PkgPwrL2, " : "",
2675 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
2676 			(msr & 1 << 25) ? "CorePwr, " : "",
2677 			(msr & 1 << 24) ? "Amps, " : "",
2678 			(msr & 1 << 22) ? "VR-Therm, " : "",
2679 			(msr & 1 << 21) ? "Auto-HWP, " : "",
2680 			(msr & 1 << 20) ? "Graphics, " : "",
2681 			(msr & 1 << 18) ? "bit18, " : "",
2682 			(msr & 1 << 17) ? "ThermStatus, " : "",
2683 			(msr & 1 << 16) ? "PROCHOT, " : "");
2684 
2685 	}
2686 	if (do_gfx_perf_limit_reasons) {
2687 		get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
2688 		fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2689 		fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
2690 			(msr & 1 << 0) ? "PROCHOT, " : "",
2691 			(msr & 1 << 1) ? "ThermStatus, " : "",
2692 			(msr & 1 << 4) ? "Graphics, " : "",
2693 			(msr & 1 << 6) ? "VR-Therm, " : "",
2694 			(msr & 1 << 8) ? "Amps, " : "",
2695 			(msr & 1 << 9) ? "GFXPwr, " : "",
2696 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
2697 			(msr & 1 << 11) ? "PkgPwrL2, " : "");
2698 		fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
2699 			(msr & 1 << 16) ? "PROCHOT, " : "",
2700 			(msr & 1 << 17) ? "ThermStatus, " : "",
2701 			(msr & 1 << 20) ? "Graphics, " : "",
2702 			(msr & 1 << 22) ? "VR-Therm, " : "",
2703 			(msr & 1 << 24) ? "Amps, " : "",
2704 			(msr & 1 << 25) ? "GFXPwr, " : "",
2705 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
2706 			(msr & 1 << 27) ? "PkgPwrL2, " : "");
2707 	}
2708 	if (do_ring_perf_limit_reasons) {
2709 		get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
2710 		fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2711 		fprintf(outf, " (Active: %s%s%s%s%s%s)",
2712 			(msr & 1 << 0) ? "PROCHOT, " : "",
2713 			(msr & 1 << 1) ? "ThermStatus, " : "",
2714 			(msr & 1 << 6) ? "VR-Therm, " : "",
2715 			(msr & 1 << 8) ? "Amps, " : "",
2716 			(msr & 1 << 10) ? "PkgPwrL1, " : "",
2717 			(msr & 1 << 11) ? "PkgPwrL2, " : "");
2718 		fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
2719 			(msr & 1 << 16) ? "PROCHOT, " : "",
2720 			(msr & 1 << 17) ? "ThermStatus, " : "",
2721 			(msr & 1 << 22) ? "VR-Therm, " : "",
2722 			(msr & 1 << 24) ? "Amps, " : "",
2723 			(msr & 1 << 26) ? "PkgPwrL1, " : "",
2724 			(msr & 1 << 27) ? "PkgPwrL2, " : "");
2725 	}
2726 	return 0;
2727 }
2728 
2729 #define	RAPL_POWER_GRANULARITY	0x7FFF	/* 15 bit power granularity */
2730 #define	RAPL_TIME_GRANULARITY	0x3F /* 6 bit time granularity */
2731 
2732 double get_tdp(unsigned int model)
2733 {
2734 	unsigned long long msr;
2735 
2736 	if (do_rapl & RAPL_PKG_POWER_INFO)
2737 		if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
2738 			return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
2739 
2740 	switch (model) {
2741 	case INTEL_FAM6_ATOM_SILVERMONT1:
2742 	case INTEL_FAM6_ATOM_SILVERMONT2:
2743 		return 30.0;
2744 	default:
2745 		return 135.0;
2746 	}
2747 }
2748 
2749 /*
2750  * rapl_dram_energy_units_probe()
2751  * Energy units are either hard-coded, or come from RAPL Energy Unit MSR.
2752  */
2753 static double
2754 rapl_dram_energy_units_probe(int  model, double rapl_energy_units)
2755 {
2756 	/* only called for genuine_intel, family 6 */
2757 
2758 	switch (model) {
2759 	case INTEL_FAM6_HASWELL_X:	/* HSX */
2760 	case INTEL_FAM6_BROADWELL_X:	/* BDX */
2761 	case INTEL_FAM6_BROADWELL_XEON_D:	/* BDX-DE */
2762 	case INTEL_FAM6_XEON_PHI_KNL:	/* KNL */
2763 	case INTEL_FAM6_XEON_PHI_KNM:
2764 		return (rapl_dram_energy_units = 15.3 / 1000000);
2765 	default:
2766 		return (rapl_energy_units);
2767 	}
2768 }
2769 
2770 
2771 /*
2772  * rapl_probe()
2773  *
2774  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
2775  */
2776 void rapl_probe(unsigned int family, unsigned int model)
2777 {
2778 	unsigned long long msr;
2779 	unsigned int time_unit;
2780 	double tdp;
2781 
2782 	if (!genuine_intel)
2783 		return;
2784 
2785 	if (family != 6)
2786 		return;
2787 
2788 	switch (model) {
2789 	case INTEL_FAM6_SANDYBRIDGE:
2790 	case INTEL_FAM6_IVYBRIDGE:
2791 	case INTEL_FAM6_HASWELL_CORE:	/* HSW */
2792 	case INTEL_FAM6_HASWELL_ULT:	/* HSW */
2793 	case INTEL_FAM6_HASWELL_GT3E:	/* HSW */
2794 	case INTEL_FAM6_BROADWELL_CORE:	/* BDW */
2795 	case INTEL_FAM6_BROADWELL_GT3E:	/* BDW */
2796 		do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
2797 		break;
2798 	case INTEL_FAM6_ATOM_GOLDMONT:	/* BXT */
2799 		do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO;
2800 		break;
2801 	case INTEL_FAM6_SKYLAKE_MOBILE:	/* SKL */
2802 	case INTEL_FAM6_SKYLAKE_DESKTOP:	/* SKL */
2803 	case INTEL_FAM6_KABYLAKE_MOBILE:	/* KBL */
2804 	case INTEL_FAM6_KABYLAKE_DESKTOP:	/* KBL */
2805 		do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2806 		break;
2807 	case INTEL_FAM6_HASWELL_X:	/* HSX */
2808 	case INTEL_FAM6_BROADWELL_X:	/* BDX */
2809 	case INTEL_FAM6_BROADWELL_XEON_D:	/* BDX-DE */
2810 	case INTEL_FAM6_SKYLAKE_X:	/* SKX */
2811 	case INTEL_FAM6_XEON_PHI_KNL:	/* KNL */
2812 	case INTEL_FAM6_XEON_PHI_KNM:
2813 		do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2814 		break;
2815 	case INTEL_FAM6_SANDYBRIDGE_X:
2816 	case INTEL_FAM6_IVYBRIDGE_X:
2817 		do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
2818 		break;
2819 	case INTEL_FAM6_ATOM_SILVERMONT1:	/* BYT */
2820 	case INTEL_FAM6_ATOM_SILVERMONT2:	/* AVN */
2821 		do_rapl = RAPL_PKG | RAPL_CORES;
2822 		break;
2823 	case INTEL_FAM6_ATOM_DENVERTON:	/* DNV */
2824 		do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO | RAPL_CORES_ENERGY_STATUS;
2825 		break;
2826 	default:
2827 		return;
2828 	}
2829 
2830 	/* units on package 0, verify later other packages match */
2831 	if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
2832 		return;
2833 
2834 	rapl_power_units = 1.0 / (1 << (msr & 0xF));
2835 	if (model == INTEL_FAM6_ATOM_SILVERMONT1)
2836 		rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
2837 	else
2838 		rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
2839 
2840 	rapl_dram_energy_units = rapl_dram_energy_units_probe(model, rapl_energy_units);
2841 
2842 	time_unit = msr >> 16 & 0xF;
2843 	if (time_unit == 0)
2844 		time_unit = 0xA;
2845 
2846 	rapl_time_units = 1.0 / (1 << (time_unit));
2847 
2848 	tdp = get_tdp(model);
2849 
2850 	rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
2851 	if (debug)
2852 		fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
2853 
2854 	return;
2855 }
2856 
2857 void perf_limit_reasons_probe(unsigned int family, unsigned int model)
2858 {
2859 	if (!genuine_intel)
2860 		return;
2861 
2862 	if (family != 6)
2863 		return;
2864 
2865 	switch (model) {
2866 	case INTEL_FAM6_HASWELL_CORE:	/* HSW */
2867 	case INTEL_FAM6_HASWELL_ULT:	/* HSW */
2868 	case INTEL_FAM6_HASWELL_GT3E:	/* HSW */
2869 		do_gfx_perf_limit_reasons = 1;
2870 	case INTEL_FAM6_HASWELL_X:	/* HSX */
2871 		do_core_perf_limit_reasons = 1;
2872 		do_ring_perf_limit_reasons = 1;
2873 	default:
2874 		return;
2875 	}
2876 }
2877 
2878 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2879 {
2880 	unsigned long long msr;
2881 	unsigned int dts;
2882 	int cpu;
2883 
2884 	if (!(do_dts || do_ptm))
2885 		return 0;
2886 
2887 	cpu = t->cpu_id;
2888 
2889 	/* DTS is per-core, no need to print for each thread */
2890 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
2891 		return 0;
2892 
2893 	if (cpu_migrate(cpu)) {
2894 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2895 		return -1;
2896 	}
2897 
2898 	if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
2899 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
2900 			return 0;
2901 
2902 		dts = (msr >> 16) & 0x7F;
2903 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
2904 			cpu, msr, tcc_activation_temp - dts);
2905 
2906 #ifdef	THERM_DEBUG
2907 		if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
2908 			return 0;
2909 
2910 		dts = (msr >> 16) & 0x7F;
2911 		dts2 = (msr >> 8) & 0x7F;
2912 		fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2913 			cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2914 #endif
2915 	}
2916 
2917 
2918 	if (do_dts) {
2919 		unsigned int resolution;
2920 
2921 		if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
2922 			return 0;
2923 
2924 		dts = (msr >> 16) & 0x7F;
2925 		resolution = (msr >> 27) & 0xF;
2926 		fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
2927 			cpu, msr, tcc_activation_temp - dts, resolution);
2928 
2929 #ifdef THERM_DEBUG
2930 		if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
2931 			return 0;
2932 
2933 		dts = (msr >> 16) & 0x7F;
2934 		dts2 = (msr >> 8) & 0x7F;
2935 		fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2936 			cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2937 #endif
2938 	}
2939 
2940 	return 0;
2941 }
2942 
2943 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
2944 {
2945 	fprintf(outf, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
2946 		cpu, label,
2947 		((msr >> 15) & 1) ? "EN" : "DIS",
2948 		((msr >> 0) & 0x7FFF) * rapl_power_units,
2949 		(1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
2950 		(((msr >> 16) & 1) ? "EN" : "DIS"));
2951 
2952 	return;
2953 }
2954 
2955 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2956 {
2957 	unsigned long long msr;
2958 	int cpu;
2959 
2960 	if (!do_rapl)
2961 		return 0;
2962 
2963 	/* RAPL counters are per package, so print only for 1st thread/package */
2964 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2965 		return 0;
2966 
2967 	cpu = t->cpu_id;
2968 	if (cpu_migrate(cpu)) {
2969 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
2970 		return -1;
2971 	}
2972 
2973 	if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
2974 		return -1;
2975 
2976 	if (debug) {
2977 		fprintf(outf, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
2978 			"(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
2979 			rapl_power_units, rapl_energy_units, rapl_time_units);
2980 	}
2981 	if (do_rapl & RAPL_PKG_POWER_INFO) {
2982 
2983 		if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
2984                 	return -5;
2985 
2986 
2987 		fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2988 			cpu, msr,
2989 			((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2990 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2991 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2992 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2993 
2994 	}
2995 	if (do_rapl & RAPL_PKG) {
2996 
2997 		if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
2998 			return -9;
2999 
3000 		fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
3001 			cpu, msr, (msr >> 63) & 1 ? "": "UN");
3002 
3003 		print_power_limit_msr(cpu, msr, "PKG Limit #1");
3004 		fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
3005 			cpu,
3006 			((msr >> 47) & 1) ? "EN" : "DIS",
3007 			((msr >> 32) & 0x7FFF) * rapl_power_units,
3008 			(1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
3009 			((msr >> 48) & 1) ? "EN" : "DIS");
3010 	}
3011 
3012 	if (do_rapl & RAPL_DRAM_POWER_INFO) {
3013 		if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
3014                 	return -6;
3015 
3016 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
3017 			cpu, msr,
3018 			((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
3019 			((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
3020 			((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
3021 			((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
3022 	}
3023 	if (do_rapl & RAPL_DRAM) {
3024 		if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
3025 			return -9;
3026 		fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
3027 				cpu, msr, (msr >> 31) & 1 ? "": "UN");
3028 
3029 		print_power_limit_msr(cpu, msr, "DRAM Limit");
3030 	}
3031 	if (do_rapl & RAPL_CORE_POLICY) {
3032 		if (debug) {
3033 			if (get_msr(cpu, MSR_PP0_POLICY, &msr))
3034 				return -7;
3035 
3036 			fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
3037 		}
3038 	}
3039 	if (do_rapl & RAPL_CORES_POWER_LIMIT) {
3040 		if (debug) {
3041 			if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
3042 				return -9;
3043 			fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
3044 					cpu, msr, (msr >> 31) & 1 ? "": "UN");
3045 			print_power_limit_msr(cpu, msr, "Cores Limit");
3046 		}
3047 	}
3048 	if (do_rapl & RAPL_GFX) {
3049 		if (debug) {
3050 			if (get_msr(cpu, MSR_PP1_POLICY, &msr))
3051 				return -8;
3052 
3053 			fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
3054 
3055 			if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
3056 				return -9;
3057 			fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
3058 					cpu, msr, (msr >> 31) & 1 ? "": "UN");
3059 			print_power_limit_msr(cpu, msr, "GFX Limit");
3060 		}
3061 	}
3062 	return 0;
3063 }
3064 
3065 /*
3066  * SNB adds support for additional MSRs:
3067  *
3068  * MSR_PKG_C7_RESIDENCY            0x000003fa
3069  * MSR_CORE_C7_RESIDENCY           0x000003fe
3070  * MSR_PKG_C2_RESIDENCY            0x0000060d
3071  */
3072 
3073 int has_snb_msrs(unsigned int family, unsigned int model)
3074 {
3075 	if (!genuine_intel)
3076 		return 0;
3077 
3078 	switch (model) {
3079 	case INTEL_FAM6_SANDYBRIDGE:
3080 	case INTEL_FAM6_SANDYBRIDGE_X:
3081 	case INTEL_FAM6_IVYBRIDGE:	/* IVB */
3082 	case INTEL_FAM6_IVYBRIDGE_X:	/* IVB Xeon */
3083 	case INTEL_FAM6_HASWELL_CORE:	/* HSW */
3084 	case INTEL_FAM6_HASWELL_X:	/* HSW */
3085 	case INTEL_FAM6_HASWELL_ULT:	/* HSW */
3086 	case INTEL_FAM6_HASWELL_GT3E:	/* HSW */
3087 	case INTEL_FAM6_BROADWELL_CORE:	/* BDW */
3088 	case INTEL_FAM6_BROADWELL_GT3E:	/* BDW */
3089 	case INTEL_FAM6_BROADWELL_X:	/* BDX */
3090 	case INTEL_FAM6_BROADWELL_XEON_D:	/* BDX-DE */
3091 	case INTEL_FAM6_SKYLAKE_MOBILE:	/* SKL */
3092 	case INTEL_FAM6_SKYLAKE_DESKTOP:	/* SKL */
3093 	case INTEL_FAM6_KABYLAKE_MOBILE:	/* KBL */
3094 	case INTEL_FAM6_KABYLAKE_DESKTOP:	/* KBL */
3095 	case INTEL_FAM6_SKYLAKE_X:	/* SKX */
3096 	case INTEL_FAM6_ATOM_GOLDMONT:	/* BXT */
3097 	case INTEL_FAM6_ATOM_DENVERTON:	/* DNV */
3098 		return 1;
3099 	}
3100 	return 0;
3101 }
3102 
3103 /*
3104  * HSW adds support for additional MSRs:
3105  *
3106  * MSR_PKG_C8_RESIDENCY		0x00000630
3107  * MSR_PKG_C9_RESIDENCY		0x00000631
3108  * MSR_PKG_C10_RESIDENCY	0x00000632
3109  *
3110  * MSR_PKGC8_IRTL		0x00000633
3111  * MSR_PKGC9_IRTL		0x00000634
3112  * MSR_PKGC10_IRTL		0x00000635
3113  *
3114  */
3115 int has_hsw_msrs(unsigned int family, unsigned int model)
3116 {
3117 	if (!genuine_intel)
3118 		return 0;
3119 
3120 	switch (model) {
3121 	case INTEL_FAM6_HASWELL_ULT:	/* HSW */
3122 	case INTEL_FAM6_BROADWELL_CORE:	/* BDW */
3123 	case INTEL_FAM6_SKYLAKE_MOBILE:	/* SKL */
3124 	case INTEL_FAM6_SKYLAKE_DESKTOP:	/* SKL */
3125 	case INTEL_FAM6_KABYLAKE_MOBILE:	/* KBL */
3126 	case INTEL_FAM6_KABYLAKE_DESKTOP:	/* KBL */
3127 	case INTEL_FAM6_ATOM_GOLDMONT:	/* BXT */
3128 		return 1;
3129 	}
3130 	return 0;
3131 }
3132 
3133 /*
3134  * SKL adds support for additional MSRS:
3135  *
3136  * MSR_PKG_WEIGHTED_CORE_C0_RES    0x00000658
3137  * MSR_PKG_ANY_CORE_C0_RES         0x00000659
3138  * MSR_PKG_ANY_GFXE_C0_RES         0x0000065A
3139  * MSR_PKG_BOTH_CORE_GFXE_C0_RES   0x0000065B
3140  */
3141 int has_skl_msrs(unsigned int family, unsigned int model)
3142 {
3143 	if (!genuine_intel)
3144 		return 0;
3145 
3146 	switch (model) {
3147 	case INTEL_FAM6_SKYLAKE_MOBILE:	/* SKL */
3148 	case INTEL_FAM6_SKYLAKE_DESKTOP:	/* SKL */
3149 	case INTEL_FAM6_KABYLAKE_MOBILE:	/* KBL */
3150 	case INTEL_FAM6_KABYLAKE_DESKTOP:	/* KBL */
3151 		return 1;
3152 	}
3153 	return 0;
3154 }
3155 
3156 
3157 
3158 int is_slm(unsigned int family, unsigned int model)
3159 {
3160 	if (!genuine_intel)
3161 		return 0;
3162 	switch (model) {
3163 	case INTEL_FAM6_ATOM_SILVERMONT1:	/* BYT */
3164 	case INTEL_FAM6_ATOM_SILVERMONT2:	/* AVN */
3165 		return 1;
3166 	}
3167 	return 0;
3168 }
3169 
3170 int is_knl(unsigned int family, unsigned int model)
3171 {
3172 	if (!genuine_intel)
3173 		return 0;
3174 	switch (model) {
3175 	case INTEL_FAM6_XEON_PHI_KNL:	/* KNL */
3176 	case INTEL_FAM6_XEON_PHI_KNM:
3177 		return 1;
3178 	}
3179 	return 0;
3180 }
3181 
3182 unsigned int get_aperf_mperf_multiplier(unsigned int family, unsigned int model)
3183 {
3184 	if (is_knl(family, model))
3185 		return 1024;
3186 	return 1;
3187 }
3188 
3189 #define SLM_BCLK_FREQS 5
3190 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
3191 
3192 double slm_bclk(void)
3193 {
3194 	unsigned long long msr = 3;
3195 	unsigned int i;
3196 	double freq;
3197 
3198 	if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
3199 		fprintf(outf, "SLM BCLK: unknown\n");
3200 
3201 	i = msr & 0xf;
3202 	if (i >= SLM_BCLK_FREQS) {
3203 		fprintf(outf, "SLM BCLK[%d] invalid\n", i);
3204 		i = 3;
3205 	}
3206 	freq = slm_freq_table[i];
3207 
3208 	fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
3209 
3210 	return freq;
3211 }
3212 
3213 double discover_bclk(unsigned int family, unsigned int model)
3214 {
3215 	if (has_snb_msrs(family, model) || is_knl(family, model))
3216 		return 100.00;
3217 	else if (is_slm(family, model))
3218 		return slm_bclk();
3219 	else
3220 		return 133.33;
3221 }
3222 
3223 /*
3224  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
3225  * the Thermal Control Circuit (TCC) activates.
3226  * This is usually equal to tjMax.
3227  *
3228  * Older processors do not have this MSR, so there we guess,
3229  * but also allow cmdline over-ride with -T.
3230  *
3231  * Several MSR temperature values are in units of degrees-C
3232  * below this value, including the Digital Thermal Sensor (DTS),
3233  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
3234  */
3235 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3236 {
3237 	unsigned long long msr;
3238 	unsigned int target_c_local;
3239 	int cpu;
3240 
3241 	/* tcc_activation_temp is used only for dts or ptm */
3242 	if (!(do_dts || do_ptm))
3243 		return 0;
3244 
3245 	/* this is a per-package concept */
3246 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
3247 		return 0;
3248 
3249 	cpu = t->cpu_id;
3250 	if (cpu_migrate(cpu)) {
3251 		fprintf(outf, "Could not migrate to CPU %d\n", cpu);
3252 		return -1;
3253 	}
3254 
3255 	if (tcc_activation_temp_override != 0) {
3256 		tcc_activation_temp = tcc_activation_temp_override;
3257 		fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n",
3258 			cpu, tcc_activation_temp);
3259 		return 0;
3260 	}
3261 
3262 	/* Temperature Target MSR is Nehalem and newer only */
3263 	if (!do_nhm_platform_info)
3264 		goto guess;
3265 
3266 	if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
3267 		goto guess;
3268 
3269 	target_c_local = (msr >> 16) & 0xFF;
3270 
3271 	if (debug)
3272 		fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
3273 			cpu, msr, target_c_local);
3274 
3275 	if (!target_c_local)
3276 		goto guess;
3277 
3278 	tcc_activation_temp = target_c_local;
3279 
3280 	return 0;
3281 
3282 guess:
3283 	tcc_activation_temp = TJMAX_DEFAULT;
3284 	fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
3285 		cpu, tcc_activation_temp);
3286 
3287 	return 0;
3288 }
3289 
3290 void decode_feature_control_msr(void)
3291 {
3292 	unsigned long long msr;
3293 
3294 	if (!get_msr(base_cpu, MSR_IA32_FEATURE_CONTROL, &msr))
3295 		fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
3296 			base_cpu, msr,
3297 			msr & FEATURE_CONTROL_LOCKED ? "" : "UN-",
3298 			msr & (1 << 18) ? "SGX" : "");
3299 }
3300 
3301 void decode_misc_enable_msr(void)
3302 {
3303 	unsigned long long msr;
3304 
3305 	if (!get_msr(base_cpu, MSR_IA32_MISC_ENABLE, &msr))
3306 		fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%s %s %s)\n",
3307 			base_cpu, msr,
3308 			msr & (1 << 3) ? "TCC" : "",
3309 			msr & (1 << 16) ? "EIST" : "",
3310 			msr & (1 << 18) ? "MONITOR" : "");
3311 }
3312 
3313 /*
3314  * Decode MSR_MISC_PWR_MGMT
3315  *
3316  * Decode the bits according to the Nehalem documentation
3317  * bit[0] seems to continue to have same meaning going forward
3318  * bit[1] less so...
3319  */
3320 void decode_misc_pwr_mgmt_msr(void)
3321 {
3322 	unsigned long long msr;
3323 
3324 	if (!do_nhm_platform_info)
3325 		return;
3326 
3327 	if (!get_msr(base_cpu, MSR_MISC_PWR_MGMT, &msr))
3328 		fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
3329 			base_cpu, msr,
3330 			msr & (1 << 0) ? "DIS" : "EN",
3331 			msr & (1 << 1) ? "EN" : "DIS",
3332 			msr & (1 << 8) ? "EN" : "DIS");
3333 }
3334 
3335 void process_cpuid()
3336 {
3337 	unsigned int eax, ebx, ecx, edx, max_level, max_extended_level;
3338 	unsigned int fms, family, model, stepping;
3339 
3340 	eax = ebx = ecx = edx = 0;
3341 
3342 	__cpuid(0, max_level, ebx, ecx, edx);
3343 
3344 	if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
3345 		genuine_intel = 1;
3346 
3347 	if (debug)
3348 		fprintf(outf, "CPUID(0): %.4s%.4s%.4s ",
3349 			(char *)&ebx, (char *)&edx, (char *)&ecx);
3350 
3351 	__cpuid(1, fms, ebx, ecx, edx);
3352 	family = (fms >> 8) & 0xf;
3353 	model = (fms >> 4) & 0xf;
3354 	stepping = fms & 0xf;
3355 	if (family == 6 || family == 0xf)
3356 		model += ((fms >> 16) & 0xf) << 4;
3357 
3358 	if (debug) {
3359 		fprintf(outf, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
3360 			max_level, family, model, stepping, family, model, stepping);
3361 		fprintf(outf, "CPUID(1): %s %s %s %s %s %s %s %s %s\n",
3362 			ecx & (1 << 0) ? "SSE3" : "-",
3363 			ecx & (1 << 3) ? "MONITOR" : "-",
3364 			ecx & (1 << 6) ? "SMX" : "-",
3365 			ecx & (1 << 7) ? "EIST" : "-",
3366 			ecx & (1 << 8) ? "TM2" : "-",
3367 			edx & (1 << 4) ? "TSC" : "-",
3368 			edx & (1 << 5) ? "MSR" : "-",
3369 			edx & (1 << 22) ? "ACPI-TM" : "-",
3370 			edx & (1 << 29) ? "TM" : "-");
3371 	}
3372 
3373 	if (!(edx & (1 << 5)))
3374 		errx(1, "CPUID: no MSR");
3375 
3376 	/*
3377 	 * check max extended function levels of CPUID.
3378 	 * This is needed to check for invariant TSC.
3379 	 * This check is valid for both Intel and AMD.
3380 	 */
3381 	ebx = ecx = edx = 0;
3382 	__cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
3383 
3384 	if (max_extended_level >= 0x80000007) {
3385 
3386 		/*
3387 		 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
3388 		 * this check is valid for both Intel and AMD
3389 		 */
3390 		__cpuid(0x80000007, eax, ebx, ecx, edx);
3391 		has_invariant_tsc = edx & (1 << 8);
3392 	}
3393 
3394 	/*
3395 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
3396 	 * this check is valid for both Intel and AMD
3397 	 */
3398 
3399 	__cpuid(0x6, eax, ebx, ecx, edx);
3400 	has_aperf = ecx & (1 << 0);
3401 	do_dts = eax & (1 << 0);
3402 	do_ptm = eax & (1 << 6);
3403 	has_hwp = eax & (1 << 7);
3404 	has_hwp_notify = eax & (1 << 8);
3405 	has_hwp_activity_window = eax & (1 << 9);
3406 	has_hwp_epp = eax & (1 << 10);
3407 	has_hwp_pkg = eax & (1 << 11);
3408 	has_epb = ecx & (1 << 3);
3409 
3410 	if (debug)
3411 		fprintf(outf, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sHWP, "
3412 			"%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
3413 			has_aperf ? "" : "No-",
3414 			do_dts ? "" : "No-",
3415 			do_ptm ? "" : "No-",
3416 			has_hwp ? "" : "No-",
3417 			has_hwp_notify ? "" : "No-",
3418 			has_hwp_activity_window ? "" : "No-",
3419 			has_hwp_epp ? "" : "No-",
3420 			has_hwp_pkg ? "" : "No-",
3421 			has_epb ? "" : "No-");
3422 
3423 	if (debug)
3424 		decode_misc_enable_msr();
3425 
3426 	if (max_level >= 0x7 && debug) {
3427 		int has_sgx;
3428 
3429 		ecx = 0;
3430 
3431 		__cpuid_count(0x7, 0, eax, ebx, ecx, edx);
3432 
3433 		has_sgx = ebx & (1 << 2);
3434 		fprintf(outf, "CPUID(7): %sSGX\n", has_sgx ? "" : "No-");
3435 
3436 		if (has_sgx)
3437 			decode_feature_control_msr();
3438 	}
3439 
3440 	if (max_level >= 0x15) {
3441 		unsigned int eax_crystal;
3442 		unsigned int ebx_tsc;
3443 
3444 		/*
3445 		 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
3446 		 */
3447 		eax_crystal = ebx_tsc = crystal_hz = edx = 0;
3448 		__cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
3449 
3450 		if (ebx_tsc != 0) {
3451 
3452 			if (debug && (ebx != 0))
3453 				fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
3454 					eax_crystal, ebx_tsc, crystal_hz);
3455 
3456 			if (crystal_hz == 0)
3457 				switch(model) {
3458 				case INTEL_FAM6_SKYLAKE_MOBILE:	/* SKL */
3459 				case INTEL_FAM6_SKYLAKE_DESKTOP:	/* SKL */
3460 				case INTEL_FAM6_KABYLAKE_MOBILE:	/* KBL */
3461 				case INTEL_FAM6_KABYLAKE_DESKTOP:	/* KBL */
3462 					crystal_hz = 24000000;	/* 24.0 MHz */
3463 					break;
3464 				case INTEL_FAM6_SKYLAKE_X:	/* SKX */
3465 				case INTEL_FAM6_ATOM_DENVERTON:	/* DNV */
3466 					crystal_hz = 25000000;	/* 25.0 MHz */
3467 					break;
3468 				case INTEL_FAM6_ATOM_GOLDMONT:	/* BXT */
3469 					crystal_hz = 19200000;	/* 19.2 MHz */
3470 					break;
3471 				default:
3472 					crystal_hz = 0;
3473 			}
3474 
3475 			if (crystal_hz) {
3476 				tsc_hz =  (unsigned long long) crystal_hz * ebx_tsc / eax_crystal;
3477 				if (debug)
3478 					fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
3479 						tsc_hz / 1000000, crystal_hz, ebx_tsc,  eax_crystal);
3480 			}
3481 		}
3482 	}
3483 	if (max_level >= 0x16) {
3484 		unsigned int base_mhz, max_mhz, bus_mhz, edx;
3485 
3486 		/*
3487 		 * CPUID 16H Base MHz, Max MHz, Bus MHz
3488 		 */
3489 		base_mhz = max_mhz = bus_mhz = edx = 0;
3490 
3491 		__cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
3492 		if (debug)
3493 			fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
3494 				base_mhz, max_mhz, bus_mhz);
3495 	}
3496 
3497 	if (has_aperf)
3498 		aperf_mperf_multiplier = get_aperf_mperf_multiplier(family, model);
3499 
3500 	do_nhm_platform_info = do_nhm_cstates = do_smi = probe_nhm_msrs(family, model);
3501 	do_snb_cstates = has_snb_msrs(family, model);
3502 	do_irtl_snb = has_snb_msrs(family, model);
3503 	do_pc2 = do_snb_cstates && (pkg_cstate_limit >= PCL__2);
3504 	do_pc3 = (pkg_cstate_limit >= PCL__3);
3505 	do_pc6 = (pkg_cstate_limit >= PCL__6);
3506 	do_pc7 = do_snb_cstates && (pkg_cstate_limit >= PCL__7);
3507 	do_c8_c9_c10 = has_hsw_msrs(family, model);
3508 	do_irtl_hsw = has_hsw_msrs(family, model);
3509 	do_skl_residency = has_skl_msrs(family, model);
3510 	do_slm_cstates = is_slm(family, model);
3511 	do_knl_cstates  = is_knl(family, model);
3512 
3513 	if (debug)
3514 		decode_misc_pwr_mgmt_msr();
3515 
3516 	rapl_probe(family, model);
3517 	perf_limit_reasons_probe(family, model);
3518 
3519 	if (debug)
3520 		dump_cstate_pstate_config_info(family, model);
3521 
3522 	if (has_skl_msrs(family, model))
3523 		calculate_tsc_tweak();
3524 
3525 	do_gfx_rc6_ms = !access("/sys/class/drm/card0/power/rc6_residency_ms", R_OK);
3526 
3527 	do_gfx_mhz = !access("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", R_OK);
3528 
3529 	return;
3530 }
3531 
3532 void help()
3533 {
3534 	fprintf(outf,
3535 	"Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
3536 	"\n"
3537 	"Turbostat forks the specified COMMAND and prints statistics\n"
3538 	"when COMMAND completes.\n"
3539 	"If no COMMAND is specified, turbostat wakes every 5-seconds\n"
3540 	"to print statistics, until interrupted.\n"
3541 	"--add		add a counter\n"
3542 	"		eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
3543 	"--debug	run in \"debug\" mode\n"
3544 	"--interval sec	Override default 5-second measurement interval\n"
3545 	"--help		print this help message\n"
3546 	"--out file	create or truncate \"file\" for all output\n"
3547 	"--version	print version information\n"
3548 	"\n"
3549 	"For more help, run \"man turbostat\"\n");
3550 }
3551 
3552 
3553 /*
3554  * in /dev/cpu/ return success for names that are numbers
3555  * ie. filter out ".", "..", "microcode".
3556  */
3557 int dir_filter(const struct dirent *dirp)
3558 {
3559 	if (isdigit(dirp->d_name[0]))
3560 		return 1;
3561 	else
3562 		return 0;
3563 }
3564 
3565 int open_dev_cpu_msr(int dummy1)
3566 {
3567 	return 0;
3568 }
3569 
3570 void topology_probe()
3571 {
3572 	int i;
3573 	int max_core_id = 0;
3574 	int max_package_id = 0;
3575 	int max_siblings = 0;
3576 	struct cpu_topology {
3577 		int core_id;
3578 		int physical_package_id;
3579 	} *cpus;
3580 
3581 	/* Initialize num_cpus, max_cpu_num */
3582 	topo.num_cpus = 0;
3583 	topo.max_cpu_num = 0;
3584 	for_all_proc_cpus(count_cpus);
3585 	if (!summary_only && topo.num_cpus > 1)
3586 		show_cpu = 1;
3587 
3588 	if (debug > 1)
3589 		fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
3590 
3591 	cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
3592 	if (cpus == NULL)
3593 		err(1, "calloc cpus");
3594 
3595 	/*
3596 	 * Allocate and initialize cpu_present_set
3597 	 */
3598 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
3599 	if (cpu_present_set == NULL)
3600 		err(3, "CPU_ALLOC");
3601 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
3602 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
3603 	for_all_proc_cpus(mark_cpu_present);
3604 
3605 	/*
3606 	 * Allocate and initialize cpu_affinity_set
3607 	 */
3608 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
3609 	if (cpu_affinity_set == NULL)
3610 		err(3, "CPU_ALLOC");
3611 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
3612 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
3613 
3614 
3615 	/*
3616 	 * For online cpus
3617 	 * find max_core_id, max_package_id
3618 	 */
3619 	for (i = 0; i <= topo.max_cpu_num; ++i) {
3620 		int siblings;
3621 
3622 		if (cpu_is_not_present(i)) {
3623 			if (debug > 1)
3624 				fprintf(outf, "cpu%d NOT PRESENT\n", i);
3625 			continue;
3626 		}
3627 		cpus[i].core_id = get_core_id(i);
3628 		if (cpus[i].core_id > max_core_id)
3629 			max_core_id = cpus[i].core_id;
3630 
3631 		cpus[i].physical_package_id = get_physical_package_id(i);
3632 		if (cpus[i].physical_package_id > max_package_id)
3633 			max_package_id = cpus[i].physical_package_id;
3634 
3635 		siblings = get_num_ht_siblings(i);
3636 		if (siblings > max_siblings)
3637 			max_siblings = siblings;
3638 		if (debug > 1)
3639 			fprintf(outf, "cpu %d pkg %d core %d\n",
3640 				i, cpus[i].physical_package_id, cpus[i].core_id);
3641 	}
3642 	topo.num_cores_per_pkg = max_core_id + 1;
3643 	if (debug > 1)
3644 		fprintf(outf, "max_core_id %d, sizing for %d cores per package\n",
3645 			max_core_id, topo.num_cores_per_pkg);
3646 	if (debug && !summary_only && topo.num_cores_per_pkg > 1)
3647 		show_core = 1;
3648 
3649 	topo.num_packages = max_package_id + 1;
3650 	if (debug > 1)
3651 		fprintf(outf, "max_package_id %d, sizing for %d packages\n",
3652 			max_package_id, topo.num_packages);
3653 	if (debug && !summary_only && topo.num_packages > 1)
3654 		show_pkg = 1;
3655 
3656 	topo.num_threads_per_core = max_siblings;
3657 	if (debug > 1)
3658 		fprintf(outf, "max_siblings %d\n", max_siblings);
3659 
3660 	free(cpus);
3661 }
3662 
3663 void
3664 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
3665 {
3666 	int i;
3667 
3668 	*t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
3669 		topo.num_packages, sizeof(struct thread_data));
3670 	if (*t == NULL)
3671 		goto error;
3672 
3673 	for (i = 0; i < topo.num_threads_per_core *
3674 		topo.num_cores_per_pkg * topo.num_packages; i++)
3675 		(*t)[i].cpu_id = -1;
3676 
3677 	*c = calloc(topo.num_cores_per_pkg * topo.num_packages,
3678 		sizeof(struct core_data));
3679 	if (*c == NULL)
3680 		goto error;
3681 
3682 	for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
3683 		(*c)[i].core_id = -1;
3684 
3685 	*p = calloc(topo.num_packages, sizeof(struct pkg_data));
3686 	if (*p == NULL)
3687 		goto error;
3688 
3689 	for (i = 0; i < topo.num_packages; i++)
3690 		(*p)[i].package_id = i;
3691 
3692 	return;
3693 error:
3694 	err(1, "calloc counters");
3695 }
3696 /*
3697  * init_counter()
3698  *
3699  * set cpu_id, core_num, pkg_num
3700  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
3701  *
3702  * increment topo.num_cores when 1st core in pkg seen
3703  */
3704 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
3705 	struct pkg_data *pkg_base, int thread_num, int core_num,
3706 	int pkg_num, int cpu_id)
3707 {
3708 	struct thread_data *t;
3709 	struct core_data *c;
3710 	struct pkg_data *p;
3711 
3712 	t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
3713 	c = GET_CORE(core_base, core_num, pkg_num);
3714 	p = GET_PKG(pkg_base, pkg_num);
3715 
3716 	t->cpu_id = cpu_id;
3717 	if (thread_num == 0) {
3718 		t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
3719 		if (cpu_is_first_core_in_package(cpu_id))
3720 			t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
3721 	}
3722 
3723 	c->core_id = core_num;
3724 	p->package_id = pkg_num;
3725 }
3726 
3727 
3728 int initialize_counters(int cpu_id)
3729 {
3730 	int my_thread_id, my_core_id, my_package_id;
3731 
3732 	my_package_id = get_physical_package_id(cpu_id);
3733 	my_core_id = get_core_id(cpu_id);
3734 	my_thread_id = get_cpu_position_in_core(cpu_id);
3735 	if (!my_thread_id)
3736 		topo.num_cores++;
3737 
3738 	init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
3739 	init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
3740 	return 0;
3741 }
3742 
3743 void allocate_output_buffer()
3744 {
3745 	output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
3746 	outp = output_buffer;
3747 	if (outp == NULL)
3748 		err(-1, "calloc output buffer");
3749 }
3750 void allocate_fd_percpu(void)
3751 {
3752 	fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
3753 	if (fd_percpu == NULL)
3754 		err(-1, "calloc fd_percpu");
3755 }
3756 void allocate_irq_buffers(void)
3757 {
3758 	irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
3759 	if (irq_column_2_cpu == NULL)
3760 		err(-1, "calloc %d", topo.num_cpus);
3761 
3762 	irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
3763 	if (irqs_per_cpu == NULL)
3764 		err(-1, "calloc %d", topo.max_cpu_num + 1);
3765 }
3766 void setup_all_buffers(void)
3767 {
3768 	topology_probe();
3769 	allocate_irq_buffers();
3770 	allocate_fd_percpu();
3771 	allocate_counters(&thread_even, &core_even, &package_even);
3772 	allocate_counters(&thread_odd, &core_odd, &package_odd);
3773 	allocate_output_buffer();
3774 	for_all_proc_cpus(initialize_counters);
3775 }
3776 
3777 void set_base_cpu(void)
3778 {
3779 	base_cpu = sched_getcpu();
3780 	if (base_cpu < 0)
3781 		err(-ENODEV, "No valid cpus found");
3782 
3783 	if (debug > 1)
3784 		fprintf(outf, "base_cpu = %d\n", base_cpu);
3785 }
3786 
3787 void turbostat_init()
3788 {
3789 	setup_all_buffers();
3790 	set_base_cpu();
3791 	check_dev_msr();
3792 	check_permissions();
3793 	process_cpuid();
3794 
3795 
3796 	if (debug)
3797 		for_all_cpus(print_hwp, ODD_COUNTERS);
3798 
3799 	if (debug)
3800 		for_all_cpus(print_epb, ODD_COUNTERS);
3801 
3802 	if (debug)
3803 		for_all_cpus(print_perf_limit, ODD_COUNTERS);
3804 
3805 	if (debug)
3806 		for_all_cpus(print_rapl, ODD_COUNTERS);
3807 
3808 	for_all_cpus(set_temperature_target, ODD_COUNTERS);
3809 
3810 	if (debug)
3811 		for_all_cpus(print_thermal, ODD_COUNTERS);
3812 
3813 	if (debug && do_irtl_snb)
3814 		print_irtl();
3815 }
3816 
3817 int fork_it(char **argv)
3818 {
3819 	pid_t child_pid;
3820 	int status;
3821 
3822 	status = for_all_cpus(get_counters, EVEN_COUNTERS);
3823 	if (status)
3824 		exit(status);
3825 	/* clear affinity side-effect of get_counters() */
3826 	sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
3827 	gettimeofday(&tv_even, (struct timezone *)NULL);
3828 
3829 	child_pid = fork();
3830 	if (!child_pid) {
3831 		/* child */
3832 		execvp(argv[0], argv);
3833 	} else {
3834 
3835 		/* parent */
3836 		if (child_pid == -1)
3837 			err(1, "fork");
3838 
3839 		signal(SIGINT, SIG_IGN);
3840 		signal(SIGQUIT, SIG_IGN);
3841 		if (waitpid(child_pid, &status, 0) == -1)
3842 			err(status, "waitpid");
3843 	}
3844 	/*
3845 	 * n.b. fork_it() does not check for errors from for_all_cpus()
3846 	 * because re-starting is problematic when forking
3847 	 */
3848 	for_all_cpus(get_counters, ODD_COUNTERS);
3849 	gettimeofday(&tv_odd, (struct timezone *)NULL);
3850 	timersub(&tv_odd, &tv_even, &tv_delta);
3851 	if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
3852 		fprintf(outf, "%s: Counter reset detected\n", progname);
3853 	else {
3854 		compute_average(EVEN_COUNTERS);
3855 		format_all_counters(EVEN_COUNTERS);
3856 	}
3857 
3858 	fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
3859 
3860 	flush_output_stderr();
3861 
3862 	return status;
3863 }
3864 
3865 int get_and_dump_counters(void)
3866 {
3867 	int status;
3868 
3869 	status = for_all_cpus(get_counters, ODD_COUNTERS);
3870 	if (status)
3871 		return status;
3872 
3873 	status = for_all_cpus(dump_counters, ODD_COUNTERS);
3874 	if (status)
3875 		return status;
3876 
3877 	flush_output_stdout();
3878 
3879 	return status;
3880 }
3881 
3882 void print_version() {
3883 	fprintf(outf, "turbostat version 4.16 24 Dec 2016"
3884 		" - Len Brown <lenb@kernel.org>\n");
3885 }
3886 
3887 int add_counter(unsigned int msr_num, char *name, unsigned int width,
3888 	enum counter_scope scope, enum counter_type type,
3889 	enum counter_format format)
3890 {
3891 	struct msr_counter *msrp;
3892 
3893 	msrp = calloc(1, sizeof(struct msr_counter));
3894 	if (msrp == NULL) {
3895 		perror("calloc");
3896 		exit(1);
3897 	}
3898 
3899 	msrp->msr_num = msr_num;
3900 	strncpy(msrp->name, name, NAME_BYTES);
3901 	msrp->width = width;
3902 	msrp->type = type;
3903 	msrp->format = format;
3904 
3905 	switch (scope) {
3906 
3907 	case SCOPE_CPU:
3908 		msrp->next = sys.tp;
3909 		sys.tp = msrp;
3910 		sys.added_thread_counters++;
3911 		if (sys.added_thread_counters > MAX_ADDED_COUNTERS) {
3912 			fprintf(stderr, "exceeded max %d added thread counters\n",
3913 				MAX_ADDED_COUNTERS);
3914 			exit(-1);
3915 		}
3916 		break;
3917 
3918 	case SCOPE_CORE:
3919 		msrp->next = sys.cp;
3920 		sys.cp = msrp;
3921 		sys.added_core_counters++;
3922 		if (sys.added_core_counters > MAX_ADDED_COUNTERS) {
3923 			fprintf(stderr, "exceeded max %d added core counters\n",
3924 				MAX_ADDED_COUNTERS);
3925 			exit(-1);
3926 		}
3927 		break;
3928 
3929 	case SCOPE_PACKAGE:
3930 		msrp->next = sys.pp;
3931 		sys.pp = msrp;
3932 		sys.added_package_counters++;
3933 		if (sys.added_package_counters > MAX_ADDED_COUNTERS) {
3934 			fprintf(stderr, "exceeded max %d added package counters\n",
3935 				MAX_ADDED_COUNTERS);
3936 			exit(-1);
3937 		}
3938 		break;
3939 	}
3940 
3941 	return 0;
3942 }
3943 
3944 void parse_add_command(char *add_command)
3945 {
3946 	int msr_num = 0;
3947 	char name_buffer[NAME_BYTES];
3948 	int width = 64;
3949 	int fail = 0;
3950 	enum counter_scope scope = SCOPE_CPU;
3951 	enum counter_type type = COUNTER_CYCLES;
3952 	enum counter_format format = FORMAT_DELTA;
3953 
3954 	while (add_command) {
3955 
3956 		if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
3957 			goto next;
3958 
3959 		if (sscanf(add_command, "msr%d", &msr_num) == 1)
3960 			goto next;
3961 
3962 		if (sscanf(add_command, "u%d", &width) == 1) {
3963 			if ((width == 32) || (width == 64))
3964 				goto next;
3965 			width = 64;
3966 		}
3967 		if (!strncmp(add_command, "cpu", strlen("cpu"))) {
3968 			scope = SCOPE_CPU;
3969 			goto next;
3970 		}
3971 		if (!strncmp(add_command, "core", strlen("core"))) {
3972 			scope = SCOPE_CORE;
3973 			goto next;
3974 		}
3975 		if (!strncmp(add_command, "package", strlen("package"))) {
3976 			scope = SCOPE_PACKAGE;
3977 			goto next;
3978 		}
3979 		if (!strncmp(add_command, "cycles", strlen("cycles"))) {
3980 			type = COUNTER_CYCLES;
3981 			goto next;
3982 		}
3983 		if (!strncmp(add_command, "seconds", strlen("seconds"))) {
3984 			type = COUNTER_SECONDS;
3985 			goto next;
3986 		}
3987 		if (!strncmp(add_command, "raw", strlen("raw"))) {
3988 			format = FORMAT_RAW;
3989 			goto next;
3990 		}
3991 		if (!strncmp(add_command, "delta", strlen("delta"))) {
3992 			format = FORMAT_DELTA;
3993 			goto next;
3994 		}
3995 		if (!strncmp(add_command, "percent", strlen("percent"))) {
3996 			format = FORMAT_PERCENT;
3997 			goto next;
3998 		}
3999 
4000 		if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) {	/* 18 < NAME_BYTES */
4001 			char *eos;
4002 
4003 			eos = strchr(name_buffer, ',');
4004 			if (eos)
4005 				*eos = '\0';
4006 			goto next;
4007 		}
4008 
4009 next:
4010 		add_command = strchr(add_command, ',');
4011 		if (add_command)
4012 			add_command++;
4013 
4014 	}
4015 	if (msr_num == 0) {
4016 		fprintf(stderr, "--add: (msrDDD | msr0xXXX) required\n");
4017 		fail++;
4018 	}
4019 
4020 	/* generate default column header */
4021 	if (*name_buffer == '\0') {
4022 		if (format == FORMAT_RAW) {
4023 			if (width == 32)
4024 				sprintf(name_buffer, "msr%d", msr_num);
4025 			else
4026 				sprintf(name_buffer, "MSR%d", msr_num);
4027 		} else if (format == FORMAT_DELTA) {
4028 			if (width == 32)
4029 				sprintf(name_buffer, "cnt%d", msr_num);
4030 			else
4031 				sprintf(name_buffer, "CNT%d", msr_num);
4032 		} else if (format == FORMAT_PERCENT) {
4033 			if (width == 32)
4034 				sprintf(name_buffer, "msr%d%%", msr_num);
4035 			else
4036 				sprintf(name_buffer, "MSR%d%%", msr_num);
4037 		}
4038 	}
4039 
4040 	if (add_counter(msr_num, name_buffer, width, scope, type, format))
4041 		fail++;
4042 
4043 	if (fail) {
4044 		help();
4045 		exit(1);
4046 	}
4047 }
4048 void cmdline(int argc, char **argv)
4049 {
4050 	int opt;
4051 	int option_index = 0;
4052 	static struct option long_options[] = {
4053 		{"add",		required_argument,	0, 'a'},
4054 		{"Dump",	no_argument,		0, 'D'},
4055 		{"debug",	no_argument,		0, 'd'},
4056 		{"interval",	required_argument,	0, 'i'},
4057 		{"help",	no_argument,		0, 'h'},
4058 		{"Joules",	no_argument,		0, 'J'},
4059 		{"out",		required_argument,	0, 'o'},
4060 		{"Package",	no_argument,		0, 'p'},
4061 		{"processor",	no_argument,		0, 'p'},
4062 		{"Summary",	no_argument,		0, 'S'},
4063 		{"TCC",		required_argument,	0, 'T'},
4064 		{"version",	no_argument,		0, 'v' },
4065 		{0,		0,			0,  0 }
4066 	};
4067 
4068 	progname = argv[0];
4069 
4070 	while ((opt = getopt_long_only(argc, argv, "+C:c:Ddhi:JM:m:o:PpST:v",
4071 				long_options, &option_index)) != -1) {
4072 		switch (opt) {
4073 		case 'a':
4074 			parse_add_command(optarg);
4075 			break;
4076 		case 'D':
4077 			dump_only++;
4078 			break;
4079 		case 'd':
4080 			debug++;
4081 			break;
4082 		case 'h':
4083 		default:
4084 			help();
4085 			exit(1);
4086 		case 'i':
4087 			{
4088 				double interval = strtod(optarg, NULL);
4089 
4090 				if (interval < 0.001) {
4091 					fprintf(outf, "interval %f seconds is too small\n",
4092 						interval);
4093 					exit(2);
4094 				}
4095 
4096 				interval_ts.tv_sec = interval;
4097 				interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
4098 			}
4099 			break;
4100 		case 'J':
4101 			rapl_joules++;
4102 			break;
4103 		case 'o':
4104 			outf = fopen_or_die(optarg, "w");
4105 			break;
4106 		case 'P':
4107 			show_pkg_only++;
4108 			break;
4109 		case 'p':
4110 			show_core_only++;
4111 			break;
4112 		case 'S':
4113 			summary_only++;
4114 			break;
4115 		case 'T':
4116 			tcc_activation_temp_override = atoi(optarg);
4117 			break;
4118 		case 'v':
4119 			print_version();
4120 			exit(0);
4121 			break;
4122 		}
4123 	}
4124 }
4125 
4126 int main(int argc, char **argv)
4127 {
4128 	outf = stderr;
4129 
4130 	cmdline(argc, argv);
4131 
4132 	if (debug)
4133 		print_version();
4134 
4135 	turbostat_init();
4136 
4137 	/* dump counters and exit */
4138 	if (dump_only)
4139 		return get_and_dump_counters();
4140 
4141 	/*
4142 	 * if any params left, it must be a command to fork
4143 	 */
4144 	if (argc - optind)
4145 		return fork_it(argv + optind);
4146 	else
4147 		turbostat_loop();
4148 
4149 	return 0;
4150 }
4151