1 /*
2  * turbostat -- show CPU frequency and C-state residency
3  * on modern Intel turbo-capable processors.
4  *
5  * Copyright (c) 2012 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 <stdio.h>
24 #include <unistd.h>
25 #include <sys/types.h>
26 #include <sys/wait.h>
27 #include <sys/stat.h>
28 #include <sys/resource.h>
29 #include <fcntl.h>
30 #include <signal.h>
31 #include <sys/time.h>
32 #include <stdlib.h>
33 #include <dirent.h>
34 #include <string.h>
35 #include <ctype.h>
36 #include <sched.h>
37 
38 #define MSR_TSC	0x10
39 #define MSR_NEHALEM_PLATFORM_INFO	0xCE
40 #define MSR_NEHALEM_TURBO_RATIO_LIMIT	0x1AD
41 #define MSR_APERF	0xE8
42 #define MSR_MPERF	0xE7
43 #define MSR_PKG_C2_RESIDENCY	0x60D	/* SNB only */
44 #define MSR_PKG_C3_RESIDENCY	0x3F8
45 #define MSR_PKG_C6_RESIDENCY	0x3F9
46 #define MSR_PKG_C7_RESIDENCY	0x3FA	/* SNB only */
47 #define MSR_CORE_C3_RESIDENCY	0x3FC
48 #define MSR_CORE_C6_RESIDENCY	0x3FD
49 #define MSR_CORE_C7_RESIDENCY	0x3FE	/* SNB only */
50 
51 char *proc_stat = "/proc/stat";
52 unsigned int interval_sec = 5;	/* set with -i interval_sec */
53 unsigned int verbose;		/* set with -v */
54 unsigned int summary_only;	/* set with -s */
55 unsigned int skip_c0;
56 unsigned int skip_c1;
57 unsigned int do_nhm_cstates;
58 unsigned int do_snb_cstates;
59 unsigned int has_aperf;
60 unsigned int units = 1000000000;	/* Ghz etc */
61 unsigned int genuine_intel;
62 unsigned int has_invariant_tsc;
63 unsigned int do_nehalem_platform_info;
64 unsigned int do_nehalem_turbo_ratio_limit;
65 unsigned int extra_msr_offset;
66 double bclk;
67 unsigned int show_pkg;
68 unsigned int show_core;
69 unsigned int show_cpu;
70 unsigned int show_pkg_only;
71 unsigned int show_core_only;
72 char *output_buffer, *outp;
73 
74 int aperf_mperf_unstable;
75 int backwards_count;
76 char *progname;
77 
78 cpu_set_t *cpu_present_set, *cpu_affinity_set;
79 size_t cpu_present_setsize, cpu_affinity_setsize;
80 
81 struct thread_data {
82 	unsigned long long tsc;
83 	unsigned long long aperf;
84 	unsigned long long mperf;
85 	unsigned long long c1;	/* derived */
86 	unsigned long long extra_msr;
87 	unsigned int cpu_id;
88 	unsigned int flags;
89 #define CPU_IS_FIRST_THREAD_IN_CORE	0x2
90 #define CPU_IS_FIRST_CORE_IN_PACKAGE	0x4
91 } *thread_even, *thread_odd;
92 
93 struct core_data {
94 	unsigned long long c3;
95 	unsigned long long c6;
96 	unsigned long long c7;
97 	unsigned int core_id;
98 } *core_even, *core_odd;
99 
100 struct pkg_data {
101 	unsigned long long pc2;
102 	unsigned long long pc3;
103 	unsigned long long pc6;
104 	unsigned long long pc7;
105 	unsigned int package_id;
106 } *package_even, *package_odd;
107 
108 #define ODD_COUNTERS thread_odd, core_odd, package_odd
109 #define EVEN_COUNTERS thread_even, core_even, package_even
110 
111 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
112 	(thread_base + (pkg_no) * topo.num_cores_per_pkg * \
113 		topo.num_threads_per_core + \
114 		(core_no) * topo.num_threads_per_core + (thread_no))
115 #define GET_CORE(core_base, core_no, pkg_no) \
116 	(core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
117 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
118 
119 struct system_summary {
120 	struct thread_data threads;
121 	struct core_data cores;
122 	struct pkg_data packages;
123 } sum, average;
124 
125 
126 struct topo_params {
127 	int num_packages;
128 	int num_cpus;
129 	int num_cores;
130 	int max_cpu_num;
131 	int num_cores_per_pkg;
132 	int num_threads_per_core;
133 } topo;
134 
135 struct timeval tv_even, tv_odd, tv_delta;
136 
137 void setup_all_buffers(void);
138 
139 int cpu_is_not_present(int cpu)
140 {
141 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
142 }
143 /*
144  * run func(thread, core, package) in topology order
145  * skip non-present cpus
146  */
147 
148 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
149 	struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
150 {
151 	int retval, pkg_no, core_no, thread_no;
152 
153 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
154 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
155 			for (thread_no = 0; thread_no <
156 				topo.num_threads_per_core; ++thread_no) {
157 				struct thread_data *t;
158 				struct core_data *c;
159 				struct pkg_data *p;
160 
161 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
162 
163 				if (cpu_is_not_present(t->cpu_id))
164 					continue;
165 
166 				c = GET_CORE(core_base, core_no, pkg_no);
167 				p = GET_PKG(pkg_base, pkg_no);
168 
169 				retval = func(t, c, p);
170 				if (retval)
171 					return retval;
172 			}
173 		}
174 	}
175 	return 0;
176 }
177 
178 int cpu_migrate(int cpu)
179 {
180 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
181 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
182 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
183 		return -1;
184 	else
185 		return 0;
186 }
187 
188 int get_msr(int cpu, off_t offset, unsigned long long *msr)
189 {
190 	ssize_t retval;
191 	char pathname[32];
192 	int fd;
193 
194 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
195 	fd = open(pathname, O_RDONLY);
196 	if (fd < 0)
197 		return -1;
198 
199 	retval = pread(fd, msr, sizeof *msr, offset);
200 	close(fd);
201 
202 	if (retval != sizeof *msr)
203 		return -1;
204 
205 	return 0;
206 }
207 
208 void print_header(void)
209 {
210 	if (show_pkg)
211 		outp += sprintf(outp, "pk");
212 	if (show_pkg)
213 		outp += sprintf(outp, " ");
214 	if (show_core)
215 		outp += sprintf(outp, "cor");
216 	if (show_cpu)
217 		outp += sprintf(outp, " CPU");
218 	if (show_pkg || show_core || show_cpu)
219 		outp += sprintf(outp, " ");
220 	if (do_nhm_cstates)
221 		outp += sprintf(outp, "   %%c0");
222 	if (has_aperf)
223 		outp += sprintf(outp, "  GHz");
224 	outp += sprintf(outp, "  TSC");
225 	if (do_nhm_cstates)
226 		outp += sprintf(outp, "    %%c1");
227 	if (do_nhm_cstates)
228 		outp += sprintf(outp, "    %%c3");
229 	if (do_nhm_cstates)
230 		outp += sprintf(outp, "    %%c6");
231 	if (do_snb_cstates)
232 		outp += sprintf(outp, "    %%c7");
233 	if (do_snb_cstates)
234 		outp += sprintf(outp, "   %%pc2");
235 	if (do_nhm_cstates)
236 		outp += sprintf(outp, "   %%pc3");
237 	if (do_nhm_cstates)
238 		outp += sprintf(outp, "   %%pc6");
239 	if (do_snb_cstates)
240 		outp += sprintf(outp, "   %%pc7");
241 	if (extra_msr_offset)
242 		outp += sprintf(outp, "        MSR 0x%x ", extra_msr_offset);
243 
244 	outp += sprintf(outp, "\n");
245 }
246 
247 int dump_counters(struct thread_data *t, struct core_data *c,
248 	struct pkg_data *p)
249 {
250 	fprintf(stderr, "t %p, c %p, p %p\n", t, c, p);
251 
252 	if (t) {
253 		fprintf(stderr, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
254 		fprintf(stderr, "TSC: %016llX\n", t->tsc);
255 		fprintf(stderr, "aperf: %016llX\n", t->aperf);
256 		fprintf(stderr, "mperf: %016llX\n", t->mperf);
257 		fprintf(stderr, "c1: %016llX\n", t->c1);
258 		fprintf(stderr, "msr0x%x: %016llX\n",
259 			extra_msr_offset, t->extra_msr);
260 	}
261 
262 	if (c) {
263 		fprintf(stderr, "core: %d\n", c->core_id);
264 		fprintf(stderr, "c3: %016llX\n", c->c3);
265 		fprintf(stderr, "c6: %016llX\n", c->c6);
266 		fprintf(stderr, "c7: %016llX\n", c->c7);
267 	}
268 
269 	if (p) {
270 		fprintf(stderr, "package: %d\n", p->package_id);
271 		fprintf(stderr, "pc2: %016llX\n", p->pc2);
272 		fprintf(stderr, "pc3: %016llX\n", p->pc3);
273 		fprintf(stderr, "pc6: %016llX\n", p->pc6);
274 		fprintf(stderr, "pc7: %016llX\n", p->pc7);
275 	}
276 	return 0;
277 }
278 
279 /*
280  * column formatting convention & formats
281  * package: "pk" 2 columns %2d
282  * core: "cor" 3 columns %3d
283  * CPU: "CPU" 3 columns %3d
284  * GHz: "GHz" 3 columns %3.2
285  * TSC: "TSC" 3 columns %3.2
286  * percentage " %pc3" %6.2
287  */
288 int format_counters(struct thread_data *t, struct core_data *c,
289 	struct pkg_data *p)
290 {
291 	double interval_float;
292 
293 	 /* if showing only 1st thread in core and this isn't one, bail out */
294 	if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
295 		return 0;
296 
297 	 /* if showing only 1st thread in pkg and this isn't one, bail out */
298 	if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
299 		return 0;
300 
301 	interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
302 
303 	/* topo columns, print blanks on 1st (average) line */
304 	if (t == &average.threads) {
305 		if (show_pkg)
306 			outp += sprintf(outp, "  ");
307 		if (show_pkg && show_core)
308 			outp += sprintf(outp, " ");
309 		if (show_core)
310 			outp += sprintf(outp, "   ");
311 		if (show_cpu)
312 			outp += sprintf(outp, " " "   ");
313 	} else {
314 		if (show_pkg) {
315 			if (p)
316 				outp += sprintf(outp, "%2d", p->package_id);
317 			else
318 				outp += sprintf(outp, "  ");
319 		}
320 		if (show_pkg && show_core)
321 			outp += sprintf(outp, " ");
322 		if (show_core) {
323 			if (c)
324 				outp += sprintf(outp, "%3d", c->core_id);
325 			else
326 				outp += sprintf(outp, "   ");
327 		}
328 		if (show_cpu)
329 			outp += sprintf(outp, " %3d", t->cpu_id);
330 	}
331 
332 	/* %c0 */
333 	if (do_nhm_cstates) {
334 		if (show_pkg || show_core || show_cpu)
335 			outp += sprintf(outp, " ");
336 		if (!skip_c0)
337 			outp += sprintf(outp, "%6.2f", 100.0 * t->mperf/t->tsc);
338 		else
339 			outp += sprintf(outp, "  ****");
340 	}
341 
342 	/* GHz */
343 	if (has_aperf) {
344 		if (!aperf_mperf_unstable) {
345 			outp += sprintf(outp, " %3.2f",
346 				1.0 * t->tsc / units * t->aperf /
347 				t->mperf / interval_float);
348 		} else {
349 			if (t->aperf > t->tsc || t->mperf > t->tsc) {
350 				outp += sprintf(outp, " ***");
351 			} else {
352 				outp += sprintf(outp, "%3.1f*",
353 					1.0 * t->tsc /
354 					units * t->aperf /
355 					t->mperf / interval_float);
356 			}
357 		}
358 	}
359 
360 	/* TSC */
361 	outp += sprintf(outp, "%5.2f", 1.0 * t->tsc/units/interval_float);
362 
363 	if (do_nhm_cstates) {
364 		if (!skip_c1)
365 			outp += sprintf(outp, " %6.2f", 100.0 * t->c1/t->tsc);
366 		else
367 			outp += sprintf(outp, "  ****");
368 	}
369 
370 	/* print per-core data only for 1st thread in core */
371 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
372 		goto done;
373 
374 	if (do_nhm_cstates)
375 		outp += sprintf(outp, " %6.2f", 100.0 * c->c3/t->tsc);
376 	if (do_nhm_cstates)
377 		outp += sprintf(outp, " %6.2f", 100.0 * c->c6/t->tsc);
378 	if (do_snb_cstates)
379 		outp += sprintf(outp, " %6.2f", 100.0 * c->c7/t->tsc);
380 
381 	/* print per-package data only for 1st core in package */
382 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
383 		goto done;
384 
385 	if (do_snb_cstates)
386 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc2/t->tsc);
387 	if (do_nhm_cstates)
388 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc3/t->tsc);
389 	if (do_nhm_cstates)
390 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc6/t->tsc);
391 	if (do_snb_cstates)
392 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc7/t->tsc);
393 done:
394 	if (extra_msr_offset)
395 		outp += sprintf(outp, "  0x%016llx", t->extra_msr);
396 	outp += sprintf(outp, "\n");
397 
398 	return 0;
399 }
400 
401 void flush_stdout()
402 {
403 	fputs(output_buffer, stdout);
404 	outp = output_buffer;
405 }
406 void flush_stderr()
407 {
408 	fputs(output_buffer, stderr);
409 	outp = output_buffer;
410 }
411 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
412 {
413 	static int printed;
414 
415 	if (!printed || !summary_only)
416 		print_header();
417 
418 	if (topo.num_cpus > 1)
419 		format_counters(&average.threads, &average.cores,
420 			&average.packages);
421 
422 	printed = 1;
423 
424 	if (summary_only)
425 		return;
426 
427 	for_all_cpus(format_counters, t, c, p);
428 }
429 
430 void
431 delta_package(struct pkg_data *new, struct pkg_data *old)
432 {
433 	old->pc2 = new->pc2 - old->pc2;
434 	old->pc3 = new->pc3 - old->pc3;
435 	old->pc6 = new->pc6 - old->pc6;
436 	old->pc7 = new->pc7 - old->pc7;
437 }
438 
439 void
440 delta_core(struct core_data *new, struct core_data *old)
441 {
442 	old->c3 = new->c3 - old->c3;
443 	old->c6 = new->c6 - old->c6;
444 	old->c7 = new->c7 - old->c7;
445 }
446 
447 /*
448  * old = new - old
449  */
450 void
451 delta_thread(struct thread_data *new, struct thread_data *old,
452 	struct core_data *core_delta)
453 {
454 	old->tsc = new->tsc - old->tsc;
455 
456 	/* check for TSC < 1 Mcycles over interval */
457 	if (old->tsc < (1000 * 1000)) {
458 		fprintf(stderr, "Insanely slow TSC rate, TSC stops in idle?\n");
459 		fprintf(stderr, "You can disable all c-states by booting with \"idle=poll\"\n");
460 		fprintf(stderr, "or just the deep ones with \"processor.max_cstate=1\"\n");
461 		exit(-3);
462 	}
463 
464 	old->c1 = new->c1 - old->c1;
465 
466 	if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
467 		old->aperf = new->aperf - old->aperf;
468 		old->mperf = new->mperf - old->mperf;
469 	} else {
470 
471 		if (!aperf_mperf_unstable) {
472 			fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
473 			fprintf(stderr, "* Frequency results do not cover entire interval *\n");
474 			fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
475 
476 			aperf_mperf_unstable = 1;
477 		}
478 		/*
479 		 * mperf delta is likely a huge "positive" number
480 		 * can not use it for calculating c0 time
481 		 */
482 		skip_c0 = 1;
483 		skip_c1 = 1;
484 	}
485 
486 
487 	/*
488 	 * As counter collection is not atomic,
489 	 * it is possible for mperf's non-halted cycles + idle states
490 	 * to exceed TSC's all cycles: show c1 = 0% in that case.
491 	 */
492 	if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
493 		old->c1 = 0;
494 	else {
495 		/* normal case, derive c1 */
496 		old->c1 = old->tsc - old->mperf - core_delta->c3
497 				- core_delta->c6 - core_delta->c7;
498 	}
499 
500 	if (old->mperf == 0) {
501 		if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
502 		old->mperf = 1;	/* divide by 0 protection */
503 	}
504 
505 	/*
506 	 * for "extra msr", just copy the latest w/o subtracting
507 	 */
508 	old->extra_msr = new->extra_msr;
509 }
510 
511 int delta_cpu(struct thread_data *t, struct core_data *c,
512 	struct pkg_data *p, struct thread_data *t2,
513 	struct core_data *c2, struct pkg_data *p2)
514 {
515 	/* calculate core delta only for 1st thread in core */
516 	if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
517 		delta_core(c, c2);
518 
519 	/* always calculate thread delta */
520 	delta_thread(t, t2, c2);	/* c2 is core delta */
521 
522 	/* calculate package delta only for 1st core in package */
523 	if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
524 		delta_package(p, p2);
525 
526 	return 0;
527 }
528 
529 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
530 {
531 	t->tsc = 0;
532 	t->aperf = 0;
533 	t->mperf = 0;
534 	t->c1 = 0;
535 
536 	/* tells format_counters to dump all fields from this set */
537 	t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
538 
539 	c->c3 = 0;
540 	c->c6 = 0;
541 	c->c7 = 0;
542 
543 	p->pc2 = 0;
544 	p->pc3 = 0;
545 	p->pc6 = 0;
546 	p->pc7 = 0;
547 }
548 int sum_counters(struct thread_data *t, struct core_data *c,
549 	struct pkg_data *p)
550 {
551 	average.threads.tsc += t->tsc;
552 	average.threads.aperf += t->aperf;
553 	average.threads.mperf += t->mperf;
554 	average.threads.c1 += t->c1;
555 
556 	/* sum per-core values only for 1st thread in core */
557 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
558 		return 0;
559 
560 	average.cores.c3 += c->c3;
561 	average.cores.c6 += c->c6;
562 	average.cores.c7 += c->c7;
563 
564 	/* sum per-pkg values only for 1st core in pkg */
565 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
566 		return 0;
567 
568 	average.packages.pc2 += p->pc2;
569 	average.packages.pc3 += p->pc3;
570 	average.packages.pc6 += p->pc6;
571 	average.packages.pc7 += p->pc7;
572 
573 	return 0;
574 }
575 /*
576  * sum the counters for all cpus in the system
577  * compute the weighted average
578  */
579 void compute_average(struct thread_data *t, struct core_data *c,
580 	struct pkg_data *p)
581 {
582 	clear_counters(&average.threads, &average.cores, &average.packages);
583 
584 	for_all_cpus(sum_counters, t, c, p);
585 
586 	average.threads.tsc /= topo.num_cpus;
587 	average.threads.aperf /= topo.num_cpus;
588 	average.threads.mperf /= topo.num_cpus;
589 	average.threads.c1 /= topo.num_cpus;
590 
591 	average.cores.c3 /= topo.num_cores;
592 	average.cores.c6 /= topo.num_cores;
593 	average.cores.c7 /= topo.num_cores;
594 
595 	average.packages.pc2 /= topo.num_packages;
596 	average.packages.pc3 /= topo.num_packages;
597 	average.packages.pc6 /= topo.num_packages;
598 	average.packages.pc7 /= topo.num_packages;
599 }
600 
601 static unsigned long long rdtsc(void)
602 {
603 	unsigned int low, high;
604 
605 	asm volatile("rdtsc" : "=a" (low), "=d" (high));
606 
607 	return low | ((unsigned long long)high) << 32;
608 }
609 
610 
611 /*
612  * get_counters(...)
613  * migrate to cpu
614  * acquire and record local counters for that cpu
615  */
616 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
617 {
618 	int cpu = t->cpu_id;
619 
620 	if (cpu_migrate(cpu))
621 		return -1;
622 
623 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
624 
625 	if (has_aperf) {
626 		if (get_msr(cpu, MSR_APERF, &t->aperf))
627 			return -3;
628 		if (get_msr(cpu, MSR_MPERF, &t->mperf))
629 			return -4;
630 	}
631 
632 	if (extra_msr_offset)
633 		if (get_msr(cpu, extra_msr_offset, &t->extra_msr))
634 			return -5;
635 
636 	/* collect core counters only for 1st thread in core */
637 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
638 		return 0;
639 
640 	if (do_nhm_cstates) {
641 		if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
642 			return -6;
643 		if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
644 			return -7;
645 	}
646 
647 	if (do_snb_cstates)
648 		if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
649 			return -8;
650 
651 	/* collect package counters only for 1st core in package */
652 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
653 		return 0;
654 
655 	if (do_nhm_cstates) {
656 		if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
657 			return -9;
658 		if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
659 			return -10;
660 	}
661 	if (do_snb_cstates) {
662 		if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
663 			return -11;
664 		if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
665 			return -12;
666 	}
667 	return 0;
668 }
669 
670 void print_verbose_header(void)
671 {
672 	unsigned long long msr;
673 	unsigned int ratio;
674 
675 	if (!do_nehalem_platform_info)
676 		return;
677 
678 	get_msr(0, MSR_NEHALEM_PLATFORM_INFO, &msr);
679 
680 	ratio = (msr >> 40) & 0xFF;
681 	fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
682 		ratio, bclk, ratio * bclk);
683 
684 	ratio = (msr >> 8) & 0xFF;
685 	fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
686 		ratio, bclk, ratio * bclk);
687 
688 	if (verbose > 1)
689 		fprintf(stderr, "MSR_NEHALEM_PLATFORM_INFO: 0x%llx\n", msr);
690 
691 	if (!do_nehalem_turbo_ratio_limit)
692 		return;
693 
694 	get_msr(0, MSR_NEHALEM_TURBO_RATIO_LIMIT, &msr);
695 
696 	ratio = (msr >> 24) & 0xFF;
697 	if (ratio)
698 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
699 			ratio, bclk, ratio * bclk);
700 
701 	ratio = (msr >> 16) & 0xFF;
702 	if (ratio)
703 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
704 			ratio, bclk, ratio * bclk);
705 
706 	ratio = (msr >> 8) & 0xFF;
707 	if (ratio)
708 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
709 			ratio, bclk, ratio * bclk);
710 
711 	ratio = (msr >> 0) & 0xFF;
712 	if (ratio)
713 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
714 			ratio, bclk, ratio * bclk);
715 
716 }
717 
718 void free_all_buffers(void)
719 {
720 	CPU_FREE(cpu_present_set);
721 	cpu_present_set = NULL;
722 	cpu_present_set = 0;
723 
724 	CPU_FREE(cpu_affinity_set);
725 	cpu_affinity_set = NULL;
726 	cpu_affinity_setsize = 0;
727 
728 	free(thread_even);
729 	free(core_even);
730 	free(package_even);
731 
732 	thread_even = NULL;
733 	core_even = NULL;
734 	package_even = NULL;
735 
736 	free(thread_odd);
737 	free(core_odd);
738 	free(package_odd);
739 
740 	thread_odd = NULL;
741 	core_odd = NULL;
742 	package_odd = NULL;
743 
744 	free(output_buffer);
745 	output_buffer = NULL;
746 	outp = NULL;
747 }
748 
749 /*
750  * cpu_is_first_sibling_in_core(cpu)
751  * return 1 if given CPU is 1st HT sibling in the core
752  */
753 int cpu_is_first_sibling_in_core(int cpu)
754 {
755 	char path[64];
756 	FILE *filep;
757 	int first_cpu;
758 
759 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
760 	filep = fopen(path, "r");
761 	if (filep == NULL) {
762 		perror(path);
763 		exit(1);
764 	}
765 	fscanf(filep, "%d", &first_cpu);
766 	fclose(filep);
767 	return (cpu == first_cpu);
768 }
769 
770 /*
771  * cpu_is_first_core_in_package(cpu)
772  * return 1 if given CPU is 1st core in package
773  */
774 int cpu_is_first_core_in_package(int cpu)
775 {
776 	char path[64];
777 	FILE *filep;
778 	int first_cpu;
779 
780 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
781 	filep = fopen(path, "r");
782 	if (filep == NULL) {
783 		perror(path);
784 		exit(1);
785 	}
786 	fscanf(filep, "%d", &first_cpu);
787 	fclose(filep);
788 	return (cpu == first_cpu);
789 }
790 
791 int get_physical_package_id(int cpu)
792 {
793 	char path[80];
794 	FILE *filep;
795 	int pkg;
796 
797 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
798 	filep = fopen(path, "r");
799 	if (filep == NULL) {
800 		perror(path);
801 		exit(1);
802 	}
803 	fscanf(filep, "%d", &pkg);
804 	fclose(filep);
805 	return pkg;
806 }
807 
808 int get_core_id(int cpu)
809 {
810 	char path[80];
811 	FILE *filep;
812 	int core;
813 
814 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
815 	filep = fopen(path, "r");
816 	if (filep == NULL) {
817 		perror(path);
818 		exit(1);
819 	}
820 	fscanf(filep, "%d", &core);
821 	fclose(filep);
822 	return core;
823 }
824 
825 int get_num_ht_siblings(int cpu)
826 {
827 	char path[80];
828 	FILE *filep;
829 	int sib1, sib2;
830 	int matches;
831 	char character;
832 
833 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
834 	filep = fopen(path, "r");
835 	if (filep == NULL) {
836 		perror(path);
837 		exit(1);
838 	}
839 	/*
840 	 * file format:
841 	 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
842 	 * otherwinse 1 sibling (self).
843 	 */
844 	matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
845 
846 	fclose(filep);
847 
848 	if (matches == 3)
849 		return 2;
850 	else
851 		return 1;
852 }
853 
854 /*
855  * run func(thread, core, package) in topology order
856  * skip non-present cpus
857  */
858 
859 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
860 	struct pkg_data *, struct thread_data *, struct core_data *,
861 	struct pkg_data *), struct thread_data *thread_base,
862 	struct core_data *core_base, struct pkg_data *pkg_base,
863 	struct thread_data *thread_base2, struct core_data *core_base2,
864 	struct pkg_data *pkg_base2)
865 {
866 	int retval, pkg_no, core_no, thread_no;
867 
868 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
869 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
870 			for (thread_no = 0; thread_no <
871 				topo.num_threads_per_core; ++thread_no) {
872 				struct thread_data *t, *t2;
873 				struct core_data *c, *c2;
874 				struct pkg_data *p, *p2;
875 
876 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
877 
878 				if (cpu_is_not_present(t->cpu_id))
879 					continue;
880 
881 				t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
882 
883 				c = GET_CORE(core_base, core_no, pkg_no);
884 				c2 = GET_CORE(core_base2, core_no, pkg_no);
885 
886 				p = GET_PKG(pkg_base, pkg_no);
887 				p2 = GET_PKG(pkg_base2, pkg_no);
888 
889 				retval = func(t, c, p, t2, c2, p2);
890 				if (retval)
891 					return retval;
892 			}
893 		}
894 	}
895 	return 0;
896 }
897 
898 /*
899  * run func(cpu) on every cpu in /proc/stat
900  * return max_cpu number
901  */
902 int for_all_proc_cpus(int (func)(int))
903 {
904 	FILE *fp;
905 	int cpu_num;
906 	int retval;
907 
908 	fp = fopen(proc_stat, "r");
909 	if (fp == NULL) {
910 		perror(proc_stat);
911 		exit(1);
912 	}
913 
914 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
915 	if (retval != 0) {
916 		perror("/proc/stat format");
917 		exit(1);
918 	}
919 
920 	while (1) {
921 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
922 		if (retval != 1)
923 			break;
924 
925 		retval = func(cpu_num);
926 		if (retval) {
927 			fclose(fp);
928 			return(retval);
929 		}
930 	}
931 	fclose(fp);
932 	return 0;
933 }
934 
935 void re_initialize(void)
936 {
937 	free_all_buffers();
938 	setup_all_buffers();
939 	printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
940 }
941 
942 
943 /*
944  * count_cpus()
945  * remember the last one seen, it will be the max
946  */
947 int count_cpus(int cpu)
948 {
949 	if (topo.max_cpu_num < cpu)
950 		topo.max_cpu_num = cpu;
951 
952 	topo.num_cpus += 1;
953 	return 0;
954 }
955 int mark_cpu_present(int cpu)
956 {
957 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
958 	return 0;
959 }
960 
961 void turbostat_loop()
962 {
963 	int retval;
964 
965 restart:
966 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
967 	if (retval) {
968 		re_initialize();
969 		goto restart;
970 	}
971 	gettimeofday(&tv_even, (struct timezone *)NULL);
972 
973 	while (1) {
974 		if (for_all_proc_cpus(cpu_is_not_present)) {
975 			re_initialize();
976 			goto restart;
977 		}
978 		sleep(interval_sec);
979 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
980 		if (retval) {
981 			re_initialize();
982 			goto restart;
983 		}
984 		gettimeofday(&tv_odd, (struct timezone *)NULL);
985 		timersub(&tv_odd, &tv_even, &tv_delta);
986 		for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
987 		compute_average(EVEN_COUNTERS);
988 		format_all_counters(EVEN_COUNTERS);
989 		flush_stdout();
990 		sleep(interval_sec);
991 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
992 		if (retval) {
993 			re_initialize();
994 			goto restart;
995 		}
996 		gettimeofday(&tv_even, (struct timezone *)NULL);
997 		timersub(&tv_even, &tv_odd, &tv_delta);
998 		for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
999 		compute_average(ODD_COUNTERS);
1000 		format_all_counters(ODD_COUNTERS);
1001 		flush_stdout();
1002 	}
1003 }
1004 
1005 void check_dev_msr()
1006 {
1007 	struct stat sb;
1008 
1009 	if (stat("/dev/cpu/0/msr", &sb)) {
1010 		fprintf(stderr, "no /dev/cpu/0/msr\n");
1011 		fprintf(stderr, "Try \"# modprobe msr\"\n");
1012 		exit(-5);
1013 	}
1014 }
1015 
1016 void check_super_user()
1017 {
1018 	if (getuid() != 0) {
1019 		fprintf(stderr, "must be root\n");
1020 		exit(-6);
1021 	}
1022 }
1023 
1024 int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model)
1025 {
1026 	if (!genuine_intel)
1027 		return 0;
1028 
1029 	if (family != 6)
1030 		return 0;
1031 
1032 	switch (model) {
1033 	case 0x1A:	/* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1034 	case 0x1E:	/* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1035 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
1036 	case 0x25:	/* Westmere Client - Clarkdale, Arrandale */
1037 	case 0x2C:	/* Westmere EP - Gulftown */
1038 	case 0x2A:	/* SNB */
1039 	case 0x2D:	/* SNB Xeon */
1040 	case 0x3A:	/* IVB */
1041 	case 0x3D:	/* IVB Xeon */
1042 		return 1;
1043 	case 0x2E:	/* Nehalem-EX Xeon - Beckton */
1044 	case 0x2F:	/* Westmere-EX Xeon - Eagleton */
1045 	default:
1046 		return 0;
1047 	}
1048 }
1049 
1050 int is_snb(unsigned int family, unsigned int model)
1051 {
1052 	if (!genuine_intel)
1053 		return 0;
1054 
1055 	switch (model) {
1056 	case 0x2A:
1057 	case 0x2D:
1058 	case 0x3A:	/* IVB */
1059 	case 0x3D:	/* IVB Xeon */
1060 		return 1;
1061 	}
1062 	return 0;
1063 }
1064 
1065 double discover_bclk(unsigned int family, unsigned int model)
1066 {
1067 	if (is_snb(family, model))
1068 		return 100.00;
1069 	else
1070 		return 133.33;
1071 }
1072 
1073 void check_cpuid()
1074 {
1075 	unsigned int eax, ebx, ecx, edx, max_level;
1076 	unsigned int fms, family, model, stepping;
1077 
1078 	eax = ebx = ecx = edx = 0;
1079 
1080 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0));
1081 
1082 	if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1083 		genuine_intel = 1;
1084 
1085 	if (verbose)
1086 		fprintf(stderr, "%.4s%.4s%.4s ",
1087 			(char *)&ebx, (char *)&edx, (char *)&ecx);
1088 
1089 	asm("cpuid" : "=a" (fms), "=c" (ecx), "=d" (edx) : "a" (1) : "ebx");
1090 	family = (fms >> 8) & 0xf;
1091 	model = (fms >> 4) & 0xf;
1092 	stepping = fms & 0xf;
1093 	if (family == 6 || family == 0xf)
1094 		model += ((fms >> 16) & 0xf) << 4;
1095 
1096 	if (verbose)
1097 		fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
1098 			max_level, family, model, stepping, family, model, stepping);
1099 
1100 	if (!(edx & (1 << 5))) {
1101 		fprintf(stderr, "CPUID: no MSR\n");
1102 		exit(1);
1103 	}
1104 
1105 	/*
1106 	 * check max extended function levels of CPUID.
1107 	 * This is needed to check for invariant TSC.
1108 	 * This check is valid for both Intel and AMD.
1109 	 */
1110 	ebx = ecx = edx = 0;
1111 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000000));
1112 
1113 	if (max_level < 0x80000007) {
1114 		fprintf(stderr, "CPUID: no invariant TSC (max_level 0x%x)\n", max_level);
1115 		exit(1);
1116 	}
1117 
1118 	/*
1119 	 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1120 	 * this check is valid for both Intel and AMD
1121 	 */
1122 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000007));
1123 	has_invariant_tsc = edx & (1 << 8);
1124 
1125 	if (!has_invariant_tsc) {
1126 		fprintf(stderr, "No invariant TSC\n");
1127 		exit(1);
1128 	}
1129 
1130 	/*
1131 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1132 	 * this check is valid for both Intel and AMD
1133 	 */
1134 
1135 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x6));
1136 	has_aperf = ecx & (1 << 0);
1137 	if (!has_aperf) {
1138 		fprintf(stderr, "No APERF MSR\n");
1139 		exit(1);
1140 	}
1141 
1142 	do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1143 	do_nhm_cstates = genuine_intel;	/* all Intel w/ non-stop TSC have NHM counters */
1144 	do_snb_cstates = is_snb(family, model);
1145 	bclk = discover_bclk(family, model);
1146 
1147 	do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model);
1148 }
1149 
1150 
1151 void usage()
1152 {
1153 	fprintf(stderr, "%s: [-v] [-M MSR#] [-i interval_sec | command ...]\n",
1154 		progname);
1155 	exit(1);
1156 }
1157 
1158 
1159 /*
1160  * in /dev/cpu/ return success for names that are numbers
1161  * ie. filter out ".", "..", "microcode".
1162  */
1163 int dir_filter(const struct dirent *dirp)
1164 {
1165 	if (isdigit(dirp->d_name[0]))
1166 		return 1;
1167 	else
1168 		return 0;
1169 }
1170 
1171 int open_dev_cpu_msr(int dummy1)
1172 {
1173 	return 0;
1174 }
1175 
1176 void topology_probe()
1177 {
1178 	int i;
1179 	int max_core_id = 0;
1180 	int max_package_id = 0;
1181 	int max_siblings = 0;
1182 	struct cpu_topology {
1183 		int core_id;
1184 		int physical_package_id;
1185 	} *cpus;
1186 
1187 	/* Initialize num_cpus, max_cpu_num */
1188 	topo.num_cpus = 0;
1189 	topo.max_cpu_num = 0;
1190 	for_all_proc_cpus(count_cpus);
1191 	if (!summary_only && topo.num_cpus > 1)
1192 		show_cpu = 1;
1193 
1194 	if (verbose > 1)
1195 		fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1196 
1197 	cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1198 	if (cpus == NULL) {
1199 		perror("calloc cpus");
1200 		exit(1);
1201 	}
1202 
1203 	/*
1204 	 * Allocate and initialize cpu_present_set
1205 	 */
1206 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1207 	if (cpu_present_set == NULL) {
1208 		perror("CPU_ALLOC");
1209 		exit(3);
1210 	}
1211 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1212 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1213 	for_all_proc_cpus(mark_cpu_present);
1214 
1215 	/*
1216 	 * Allocate and initialize cpu_affinity_set
1217 	 */
1218 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1219 	if (cpu_affinity_set == NULL) {
1220 		perror("CPU_ALLOC");
1221 		exit(3);
1222 	}
1223 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1224 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1225 
1226 
1227 	/*
1228 	 * For online cpus
1229 	 * find max_core_id, max_package_id
1230 	 */
1231 	for (i = 0; i <= topo.max_cpu_num; ++i) {
1232 		int siblings;
1233 
1234 		if (cpu_is_not_present(i)) {
1235 			if (verbose > 1)
1236 				fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1237 			continue;
1238 		}
1239 		cpus[i].core_id = get_core_id(i);
1240 		if (cpus[i].core_id > max_core_id)
1241 			max_core_id = cpus[i].core_id;
1242 
1243 		cpus[i].physical_package_id = get_physical_package_id(i);
1244 		if (cpus[i].physical_package_id > max_package_id)
1245 			max_package_id = cpus[i].physical_package_id;
1246 
1247 		siblings = get_num_ht_siblings(i);
1248 		if (siblings > max_siblings)
1249 			max_siblings = siblings;
1250 		if (verbose > 1)
1251 			fprintf(stderr, "cpu %d pkg %d core %d\n",
1252 				i, cpus[i].physical_package_id, cpus[i].core_id);
1253 	}
1254 	topo.num_cores_per_pkg = max_core_id + 1;
1255 	if (verbose > 1)
1256 		fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
1257 			max_core_id, topo.num_cores_per_pkg);
1258 	if (!summary_only && topo.num_cores_per_pkg > 1)
1259 		show_core = 1;
1260 
1261 	topo.num_packages = max_package_id + 1;
1262 	if (verbose > 1)
1263 		fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
1264 			max_package_id, topo.num_packages);
1265 	if (!summary_only && topo.num_packages > 1)
1266 		show_pkg = 1;
1267 
1268 	topo.num_threads_per_core = max_siblings;
1269 	if (verbose > 1)
1270 		fprintf(stderr, "max_siblings %d\n", max_siblings);
1271 
1272 	free(cpus);
1273 }
1274 
1275 void
1276 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1277 {
1278 	int i;
1279 
1280 	*t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1281 		topo.num_packages, sizeof(struct thread_data));
1282 	if (*t == NULL)
1283 		goto error;
1284 
1285 	for (i = 0; i < topo.num_threads_per_core *
1286 		topo.num_cores_per_pkg * topo.num_packages; i++)
1287 		(*t)[i].cpu_id = -1;
1288 
1289 	*c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1290 		sizeof(struct core_data));
1291 	if (*c == NULL)
1292 		goto error;
1293 
1294 	for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1295 		(*c)[i].core_id = -1;
1296 
1297 	*p = calloc(topo.num_packages, sizeof(struct pkg_data));
1298 	if (*p == NULL)
1299 		goto error;
1300 
1301 	for (i = 0; i < topo.num_packages; i++)
1302 		(*p)[i].package_id = i;
1303 
1304 	return;
1305 error:
1306 	perror("calloc counters");
1307 	exit(1);
1308 }
1309 /*
1310  * init_counter()
1311  *
1312  * set cpu_id, core_num, pkg_num
1313  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1314  *
1315  * increment topo.num_cores when 1st core in pkg seen
1316  */
1317 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
1318 	struct pkg_data *pkg_base, int thread_num, int core_num,
1319 	int pkg_num, int cpu_id)
1320 {
1321 	struct thread_data *t;
1322 	struct core_data *c;
1323 	struct pkg_data *p;
1324 
1325 	t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1326 	c = GET_CORE(core_base, core_num, pkg_num);
1327 	p = GET_PKG(pkg_base, pkg_num);
1328 
1329 	t->cpu_id = cpu_id;
1330 	if (thread_num == 0) {
1331 		t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1332 		if (cpu_is_first_core_in_package(cpu_id))
1333 			t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1334 	}
1335 
1336 	c->core_id = core_num;
1337 	p->package_id = pkg_num;
1338 }
1339 
1340 
1341 int initialize_counters(int cpu_id)
1342 {
1343 	int my_thread_id, my_core_id, my_package_id;
1344 
1345 	my_package_id = get_physical_package_id(cpu_id);
1346 	my_core_id = get_core_id(cpu_id);
1347 
1348 	if (cpu_is_first_sibling_in_core(cpu_id)) {
1349 		my_thread_id = 0;
1350 		topo.num_cores++;
1351 	} else {
1352 		my_thread_id = 1;
1353 	}
1354 
1355 	init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1356 	init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1357 	return 0;
1358 }
1359 
1360 void allocate_output_buffer()
1361 {
1362 	output_buffer = calloc(1, (1 + topo.num_cpus) * 128);
1363 	outp = output_buffer;
1364 	if (outp == NULL) {
1365 		perror("calloc");
1366 		exit(-1);
1367 	}
1368 }
1369 
1370 void setup_all_buffers(void)
1371 {
1372 	topology_probe();
1373 	allocate_counters(&thread_even, &core_even, &package_even);
1374 	allocate_counters(&thread_odd, &core_odd, &package_odd);
1375 	allocate_output_buffer();
1376 	for_all_proc_cpus(initialize_counters);
1377 }
1378 void turbostat_init()
1379 {
1380 	check_cpuid();
1381 
1382 	check_dev_msr();
1383 	check_super_user();
1384 
1385 	setup_all_buffers();
1386 
1387 	if (verbose)
1388 		print_verbose_header();
1389 }
1390 
1391 int fork_it(char **argv)
1392 {
1393 	pid_t child_pid;
1394 
1395 	for_all_cpus(get_counters, EVEN_COUNTERS);
1396 	/* clear affinity side-effect of get_counters() */
1397 	sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
1398 	gettimeofday(&tv_even, (struct timezone *)NULL);
1399 
1400 	child_pid = fork();
1401 	if (!child_pid) {
1402 		/* child */
1403 		execvp(argv[0], argv);
1404 	} else {
1405 		int status;
1406 
1407 		/* parent */
1408 		if (child_pid == -1) {
1409 			perror("fork");
1410 			exit(1);
1411 		}
1412 
1413 		signal(SIGINT, SIG_IGN);
1414 		signal(SIGQUIT, SIG_IGN);
1415 		if (waitpid(child_pid, &status, 0) == -1) {
1416 			perror("wait");
1417 			exit(1);
1418 		}
1419 	}
1420 	/*
1421 	 * n.b. fork_it() does not check for errors from for_all_cpus()
1422 	 * because re-starting is problematic when forking
1423 	 */
1424 	for_all_cpus(get_counters, ODD_COUNTERS);
1425 	gettimeofday(&tv_odd, (struct timezone *)NULL);
1426 	timersub(&tv_odd, &tv_even, &tv_delta);
1427 	for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1428 	compute_average(EVEN_COUNTERS);
1429 	format_all_counters(EVEN_COUNTERS);
1430 	flush_stderr();
1431 
1432 	fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
1433 
1434 	return 0;
1435 }
1436 
1437 void cmdline(int argc, char **argv)
1438 {
1439 	int opt;
1440 
1441 	progname = argv[0];
1442 
1443 	while ((opt = getopt(argc, argv, "+cpsvi:M:")) != -1) {
1444 		switch (opt) {
1445 		case 'c':
1446 			show_core_only++;
1447 			break;
1448 		case 'p':
1449 			show_pkg_only++;
1450 			break;
1451 		case 's':
1452 			summary_only++;
1453 			break;
1454 		case 'v':
1455 			verbose++;
1456 			break;
1457 		case 'i':
1458 			interval_sec = atoi(optarg);
1459 			break;
1460 		case 'M':
1461 			sscanf(optarg, "%x", &extra_msr_offset);
1462 			if (verbose > 1)
1463 				fprintf(stderr, "MSR 0x%X\n", extra_msr_offset);
1464 			break;
1465 		default:
1466 			usage();
1467 		}
1468 	}
1469 }
1470 
1471 int main(int argc, char **argv)
1472 {
1473 	cmdline(argc, argv);
1474 
1475 	if (verbose > 1)
1476 		fprintf(stderr, "turbostat v2.0 May 16, 2012"
1477 			" - Len Brown <lenb@kernel.org>\n");
1478 
1479 	turbostat_init();
1480 
1481 	/*
1482 	 * if any params left, it must be a command to fork
1483 	 */
1484 	if (argc - optind)
1485 		return fork_it(argv + optind);
1486 	else
1487 		turbostat_loop();
1488 
1489 	return 0;
1490 }
1491