xref: /openbmc/linux/tools/power/x86/turbostat/turbostat.c (revision 2f32edf12c1eafc8e5b1b0337360993fde1b3565)
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_NEHALEM_PLATFORM_INFO	0xCE
39 #define MSR_NEHALEM_TURBO_RATIO_LIMIT	0x1AD
40 #define MSR_IVT_TURBO_RATIO_LIMIT	0x1AE
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 do_ivt_turbo_ratio_limit;
66 unsigned int extra_msr_offset32;
67 unsigned int extra_msr_offset64;
68 double bclk;
69 unsigned int show_pkg;
70 unsigned int show_core;
71 unsigned int show_cpu;
72 unsigned int show_pkg_only;
73 unsigned int show_core_only;
74 char *output_buffer, *outp;
75 
76 int aperf_mperf_unstable;
77 int backwards_count;
78 char *progname;
79 
80 cpu_set_t *cpu_present_set, *cpu_affinity_set;
81 size_t cpu_present_setsize, cpu_affinity_setsize;
82 
83 struct thread_data {
84 	unsigned long long tsc;
85 	unsigned long long aperf;
86 	unsigned long long mperf;
87 	unsigned long long c1;	/* derived */
88 	unsigned long long extra_msr64;
89 	unsigned int extra_msr32;
90 	unsigned int cpu_id;
91 	unsigned int flags;
92 #define CPU_IS_FIRST_THREAD_IN_CORE	0x2
93 #define CPU_IS_FIRST_CORE_IN_PACKAGE	0x4
94 } *thread_even, *thread_odd;
95 
96 struct core_data {
97 	unsigned long long c3;
98 	unsigned long long c6;
99 	unsigned long long c7;
100 	unsigned int core_id;
101 } *core_even, *core_odd;
102 
103 struct pkg_data {
104 	unsigned long long pc2;
105 	unsigned long long pc3;
106 	unsigned long long pc6;
107 	unsigned long long pc7;
108 	unsigned int package_id;
109 } *package_even, *package_odd;
110 
111 #define ODD_COUNTERS thread_odd, core_odd, package_odd
112 #define EVEN_COUNTERS thread_even, core_even, package_even
113 
114 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
115 	(thread_base + (pkg_no) * topo.num_cores_per_pkg * \
116 		topo.num_threads_per_core + \
117 		(core_no) * topo.num_threads_per_core + (thread_no))
118 #define GET_CORE(core_base, core_no, pkg_no) \
119 	(core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
120 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
121 
122 struct system_summary {
123 	struct thread_data threads;
124 	struct core_data cores;
125 	struct pkg_data packages;
126 } sum, average;
127 
128 
129 struct topo_params {
130 	int num_packages;
131 	int num_cpus;
132 	int num_cores;
133 	int max_cpu_num;
134 	int num_cores_per_pkg;
135 	int num_threads_per_core;
136 } topo;
137 
138 struct timeval tv_even, tv_odd, tv_delta;
139 
140 void setup_all_buffers(void);
141 
142 int cpu_is_not_present(int cpu)
143 {
144 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
145 }
146 /*
147  * run func(thread, core, package) in topology order
148  * skip non-present cpus
149  */
150 
151 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
152 	struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
153 {
154 	int retval, pkg_no, core_no, thread_no;
155 
156 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
157 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
158 			for (thread_no = 0; thread_no <
159 				topo.num_threads_per_core; ++thread_no) {
160 				struct thread_data *t;
161 				struct core_data *c;
162 				struct pkg_data *p;
163 
164 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
165 
166 				if (cpu_is_not_present(t->cpu_id))
167 					continue;
168 
169 				c = GET_CORE(core_base, core_no, pkg_no);
170 				p = GET_PKG(pkg_base, pkg_no);
171 
172 				retval = func(t, c, p);
173 				if (retval)
174 					return retval;
175 			}
176 		}
177 	}
178 	return 0;
179 }
180 
181 int cpu_migrate(int cpu)
182 {
183 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
184 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
185 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
186 		return -1;
187 	else
188 		return 0;
189 }
190 
191 int get_msr(int cpu, off_t offset, unsigned long long *msr)
192 {
193 	ssize_t retval;
194 	char pathname[32];
195 	int fd;
196 
197 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
198 	fd = open(pathname, O_RDONLY);
199 	if (fd < 0)
200 		return -1;
201 
202 	retval = pread(fd, msr, sizeof *msr, offset);
203 	close(fd);
204 
205 	if (retval != sizeof *msr)
206 		return -1;
207 
208 	return 0;
209 }
210 
211 /*
212  * Truncate the 8 bytes we read from /dev/cpu/.../msr
213  * to the 4 bytes requested
214  */
215 
216 int get_msr32(int cpu, off_t offset, unsigned int *msr)
217 {
218 	int retval;
219 
220 	unsigned long long msr64;
221 
222 	retval = get_msr(cpu, offset, &msr64);
223 	*msr = (unsigned int) msr64;
224 
225 	return retval;
226 }
227 
228 
229 void print_header(void)
230 {
231 	if (show_pkg)
232 		outp += sprintf(outp, "pk");
233 	if (show_pkg)
234 		outp += sprintf(outp, " ");
235 	if (show_core)
236 		outp += sprintf(outp, "cor");
237 	if (show_cpu)
238 		outp += sprintf(outp, " CPU");
239 	if (show_pkg || show_core || show_cpu)
240 		outp += sprintf(outp, " ");
241 	if (do_nhm_cstates)
242 		outp += sprintf(outp, "   %%c0");
243 	if (has_aperf)
244 		outp += sprintf(outp, "  GHz");
245 	outp += sprintf(outp, "  TSC");
246 	if (extra_msr_offset32)
247 		outp += sprintf(outp, "  MSR 0x%04X", extra_msr_offset32);
248 	if (extra_msr_offset64)
249 		outp += sprintf(outp, "          MSR 0x%04X", extra_msr_offset64);
250 	if (do_nhm_cstates)
251 		outp += sprintf(outp, "    %%c1");
252 	if (do_nhm_cstates)
253 		outp += sprintf(outp, "    %%c3");
254 	if (do_nhm_cstates)
255 		outp += sprintf(outp, "    %%c6");
256 	if (do_snb_cstates)
257 		outp += sprintf(outp, "    %%c7");
258 	if (do_snb_cstates)
259 		outp += sprintf(outp, "   %%pc2");
260 	if (do_nhm_cstates)
261 		outp += sprintf(outp, "   %%pc3");
262 	if (do_nhm_cstates)
263 		outp += sprintf(outp, "   %%pc6");
264 	if (do_snb_cstates)
265 		outp += sprintf(outp, "   %%pc7");
266 
267 	outp += sprintf(outp, "\n");
268 }
269 
270 int dump_counters(struct thread_data *t, struct core_data *c,
271 	struct pkg_data *p)
272 {
273 	fprintf(stderr, "t %p, c %p, p %p\n", t, c, p);
274 
275 	if (t) {
276 		fprintf(stderr, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
277 		fprintf(stderr, "TSC: %016llX\n", t->tsc);
278 		fprintf(stderr, "aperf: %016llX\n", t->aperf);
279 		fprintf(stderr, "mperf: %016llX\n", t->mperf);
280 		fprintf(stderr, "c1: %016llX\n", t->c1);
281 		fprintf(stderr, "msr0x%x: %08X\n",
282 			extra_msr_offset32, t->extra_msr32);
283 		fprintf(stderr, "msr0x%x: %016llX\n",
284 			extra_msr_offset64, t->extra_msr64);
285 	}
286 
287 	if (c) {
288 		fprintf(stderr, "core: %d\n", c->core_id);
289 		fprintf(stderr, "c3: %016llX\n", c->c3);
290 		fprintf(stderr, "c6: %016llX\n", c->c6);
291 		fprintf(stderr, "c7: %016llX\n", c->c7);
292 	}
293 
294 	if (p) {
295 		fprintf(stderr, "package: %d\n", p->package_id);
296 		fprintf(stderr, "pc2: %016llX\n", p->pc2);
297 		fprintf(stderr, "pc3: %016llX\n", p->pc3);
298 		fprintf(stderr, "pc6: %016llX\n", p->pc6);
299 		fprintf(stderr, "pc7: %016llX\n", p->pc7);
300 	}
301 	return 0;
302 }
303 
304 /*
305  * column formatting convention & formats
306  * package: "pk" 2 columns %2d
307  * core: "cor" 3 columns %3d
308  * CPU: "CPU" 3 columns %3d
309  * GHz: "GHz" 3 columns %3.2
310  * TSC: "TSC" 3 columns %3.2
311  * percentage " %pc3" %6.2
312  */
313 int format_counters(struct thread_data *t, struct core_data *c,
314 	struct pkg_data *p)
315 {
316 	double interval_float;
317 
318 	 /* if showing only 1st thread in core and this isn't one, bail out */
319 	if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
320 		return 0;
321 
322 	 /* if showing only 1st thread in pkg and this isn't one, bail out */
323 	if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
324 		return 0;
325 
326 	interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
327 
328 	/* topo columns, print blanks on 1st (average) line */
329 	if (t == &average.threads) {
330 		if (show_pkg)
331 			outp += sprintf(outp, "  ");
332 		if (show_pkg && show_core)
333 			outp += sprintf(outp, " ");
334 		if (show_core)
335 			outp += sprintf(outp, "   ");
336 		if (show_cpu)
337 			outp += sprintf(outp, " " "   ");
338 	} else {
339 		if (show_pkg) {
340 			if (p)
341 				outp += sprintf(outp, "%2d", p->package_id);
342 			else
343 				outp += sprintf(outp, "  ");
344 		}
345 		if (show_pkg && show_core)
346 			outp += sprintf(outp, " ");
347 		if (show_core) {
348 			if (c)
349 				outp += sprintf(outp, "%3d", c->core_id);
350 			else
351 				outp += sprintf(outp, "   ");
352 		}
353 		if (show_cpu)
354 			outp += sprintf(outp, " %3d", t->cpu_id);
355 	}
356 
357 	/* %c0 */
358 	if (do_nhm_cstates) {
359 		if (show_pkg || show_core || show_cpu)
360 			outp += sprintf(outp, " ");
361 		if (!skip_c0)
362 			outp += sprintf(outp, "%6.2f", 100.0 * t->mperf/t->tsc);
363 		else
364 			outp += sprintf(outp, "  ****");
365 	}
366 
367 	/* GHz */
368 	if (has_aperf) {
369 		if (!aperf_mperf_unstable) {
370 			outp += sprintf(outp, " %3.2f",
371 				1.0 * t->tsc / units * t->aperf /
372 				t->mperf / interval_float);
373 		} else {
374 			if (t->aperf > t->tsc || t->mperf > t->tsc) {
375 				outp += sprintf(outp, " ***");
376 			} else {
377 				outp += sprintf(outp, "%3.1f*",
378 					1.0 * t->tsc /
379 					units * t->aperf /
380 					t->mperf / interval_float);
381 			}
382 		}
383 	}
384 
385 	/* TSC */
386 	outp += sprintf(outp, "%5.2f", 1.0 * t->tsc/units/interval_float);
387 
388 	/* msr */
389 	if (extra_msr_offset32)
390 		outp += sprintf(outp, "  0x%08x", t->extra_msr32);
391 
392 	/* MSR */
393 	if (extra_msr_offset64)
394 		outp += sprintf(outp, "  0x%016llx", t->extra_msr64);
395 
396 	if (do_nhm_cstates) {
397 		if (!skip_c1)
398 			outp += sprintf(outp, " %6.2f", 100.0 * t->c1/t->tsc);
399 		else
400 			outp += sprintf(outp, "  ****");
401 	}
402 
403 	/* print per-core data only for 1st thread in core */
404 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
405 		goto done;
406 
407 	if (do_nhm_cstates)
408 		outp += sprintf(outp, " %6.2f", 100.0 * c->c3/t->tsc);
409 	if (do_nhm_cstates)
410 		outp += sprintf(outp, " %6.2f", 100.0 * c->c6/t->tsc);
411 	if (do_snb_cstates)
412 		outp += sprintf(outp, " %6.2f", 100.0 * c->c7/t->tsc);
413 
414 	/* print per-package data only for 1st core in package */
415 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
416 		goto done;
417 
418 	if (do_snb_cstates)
419 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc2/t->tsc);
420 	if (do_nhm_cstates)
421 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc3/t->tsc);
422 	if (do_nhm_cstates)
423 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc6/t->tsc);
424 	if (do_snb_cstates)
425 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc7/t->tsc);
426 done:
427 	outp += sprintf(outp, "\n");
428 
429 	return 0;
430 }
431 
432 void flush_stdout()
433 {
434 	fputs(output_buffer, stdout);
435 	outp = output_buffer;
436 }
437 void flush_stderr()
438 {
439 	fputs(output_buffer, stderr);
440 	outp = output_buffer;
441 }
442 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
443 {
444 	static int printed;
445 
446 	if (!printed || !summary_only)
447 		print_header();
448 
449 	if (topo.num_cpus > 1)
450 		format_counters(&average.threads, &average.cores,
451 			&average.packages);
452 
453 	printed = 1;
454 
455 	if (summary_only)
456 		return;
457 
458 	for_all_cpus(format_counters, t, c, p);
459 }
460 
461 void
462 delta_package(struct pkg_data *new, struct pkg_data *old)
463 {
464 	old->pc2 = new->pc2 - old->pc2;
465 	old->pc3 = new->pc3 - old->pc3;
466 	old->pc6 = new->pc6 - old->pc6;
467 	old->pc7 = new->pc7 - old->pc7;
468 }
469 
470 void
471 delta_core(struct core_data *new, struct core_data *old)
472 {
473 	old->c3 = new->c3 - old->c3;
474 	old->c6 = new->c6 - old->c6;
475 	old->c7 = new->c7 - old->c7;
476 }
477 
478 /*
479  * old = new - old
480  */
481 void
482 delta_thread(struct thread_data *new, struct thread_data *old,
483 	struct core_data *core_delta)
484 {
485 	old->tsc = new->tsc - old->tsc;
486 
487 	/* check for TSC < 1 Mcycles over interval */
488 	if (old->tsc < (1000 * 1000)) {
489 		fprintf(stderr, "Insanely slow TSC rate, TSC stops in idle?\n");
490 		fprintf(stderr, "You can disable all c-states by booting with \"idle=poll\"\n");
491 		fprintf(stderr, "or just the deep ones with \"processor.max_cstate=1\"\n");
492 		exit(-3);
493 	}
494 
495 	old->c1 = new->c1 - old->c1;
496 
497 	if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
498 		old->aperf = new->aperf - old->aperf;
499 		old->mperf = new->mperf - old->mperf;
500 	} else {
501 
502 		if (!aperf_mperf_unstable) {
503 			fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
504 			fprintf(stderr, "* Frequency results do not cover entire interval *\n");
505 			fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
506 
507 			aperf_mperf_unstable = 1;
508 		}
509 		/*
510 		 * mperf delta is likely a huge "positive" number
511 		 * can not use it for calculating c0 time
512 		 */
513 		skip_c0 = 1;
514 		skip_c1 = 1;
515 	}
516 
517 
518 	/*
519 	 * As counter collection is not atomic,
520 	 * it is possible for mperf's non-halted cycles + idle states
521 	 * to exceed TSC's all cycles: show c1 = 0% in that case.
522 	 */
523 	if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
524 		old->c1 = 0;
525 	else {
526 		/* normal case, derive c1 */
527 		old->c1 = old->tsc - old->mperf - core_delta->c3
528 				- core_delta->c6 - core_delta->c7;
529 	}
530 
531 	if (old->mperf == 0) {
532 		if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
533 		old->mperf = 1;	/* divide by 0 protection */
534 	}
535 
536 	/*
537 	 * Extra MSR is a snapshot, simply copy latest w/o subtracting
538 	 */
539 	old->extra_msr32 = new->extra_msr32;
540 	old->extra_msr64 = new->extra_msr64;
541 }
542 
543 int delta_cpu(struct thread_data *t, struct core_data *c,
544 	struct pkg_data *p, struct thread_data *t2,
545 	struct core_data *c2, struct pkg_data *p2)
546 {
547 	/* calculate core delta only for 1st thread in core */
548 	if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
549 		delta_core(c, c2);
550 
551 	/* always calculate thread delta */
552 	delta_thread(t, t2, c2);	/* c2 is core delta */
553 
554 	/* calculate package delta only for 1st core in package */
555 	if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
556 		delta_package(p, p2);
557 
558 	return 0;
559 }
560 
561 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
562 {
563 	t->tsc = 0;
564 	t->aperf = 0;
565 	t->mperf = 0;
566 	t->c1 = 0;
567 
568 	/* tells format_counters to dump all fields from this set */
569 	t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
570 
571 	c->c3 = 0;
572 	c->c6 = 0;
573 	c->c7 = 0;
574 
575 	p->pc2 = 0;
576 	p->pc3 = 0;
577 	p->pc6 = 0;
578 	p->pc7 = 0;
579 }
580 int sum_counters(struct thread_data *t, struct core_data *c,
581 	struct pkg_data *p)
582 {
583 	average.threads.tsc += t->tsc;
584 	average.threads.aperf += t->aperf;
585 	average.threads.mperf += t->mperf;
586 	average.threads.c1 += t->c1;
587 
588 	/* sum per-core values only for 1st thread in core */
589 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
590 		return 0;
591 
592 	average.cores.c3 += c->c3;
593 	average.cores.c6 += c->c6;
594 	average.cores.c7 += c->c7;
595 
596 	/* sum per-pkg values only for 1st core in pkg */
597 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
598 		return 0;
599 
600 	average.packages.pc2 += p->pc2;
601 	average.packages.pc3 += p->pc3;
602 	average.packages.pc6 += p->pc6;
603 	average.packages.pc7 += p->pc7;
604 
605 	return 0;
606 }
607 /*
608  * sum the counters for all cpus in the system
609  * compute the weighted average
610  */
611 void compute_average(struct thread_data *t, struct core_data *c,
612 	struct pkg_data *p)
613 {
614 	clear_counters(&average.threads, &average.cores, &average.packages);
615 
616 	for_all_cpus(sum_counters, t, c, p);
617 
618 	average.threads.tsc /= topo.num_cpus;
619 	average.threads.aperf /= topo.num_cpus;
620 	average.threads.mperf /= topo.num_cpus;
621 	average.threads.c1 /= topo.num_cpus;
622 
623 	average.cores.c3 /= topo.num_cores;
624 	average.cores.c6 /= topo.num_cores;
625 	average.cores.c7 /= topo.num_cores;
626 
627 	average.packages.pc2 /= topo.num_packages;
628 	average.packages.pc3 /= topo.num_packages;
629 	average.packages.pc6 /= topo.num_packages;
630 	average.packages.pc7 /= topo.num_packages;
631 }
632 
633 static unsigned long long rdtsc(void)
634 {
635 	unsigned int low, high;
636 
637 	asm volatile("rdtsc" : "=a" (low), "=d" (high));
638 
639 	return low | ((unsigned long long)high) << 32;
640 }
641 
642 
643 /*
644  * get_counters(...)
645  * migrate to cpu
646  * acquire and record local counters for that cpu
647  */
648 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
649 {
650 	int cpu = t->cpu_id;
651 
652 	if (cpu_migrate(cpu))
653 		return -1;
654 
655 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
656 
657 	if (has_aperf) {
658 		if (get_msr(cpu, MSR_APERF, &t->aperf))
659 			return -3;
660 		if (get_msr(cpu, MSR_MPERF, &t->mperf))
661 			return -4;
662 	}
663 
664 	if (extra_msr_offset32)
665 		if (get_msr32(cpu, extra_msr_offset32, &t->extra_msr32))
666 			return -5;
667 
668 	if (extra_msr_offset64)
669 		if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
670 			return -5;
671 
672 	/* collect core counters only for 1st thread in core */
673 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
674 		return 0;
675 
676 	if (do_nhm_cstates) {
677 		if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
678 			return -6;
679 		if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
680 			return -7;
681 	}
682 
683 	if (do_snb_cstates)
684 		if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
685 			return -8;
686 
687 	/* collect package counters only for 1st core in package */
688 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
689 		return 0;
690 
691 	if (do_nhm_cstates) {
692 		if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
693 			return -9;
694 		if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
695 			return -10;
696 	}
697 	if (do_snb_cstates) {
698 		if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
699 			return -11;
700 		if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
701 			return -12;
702 	}
703 	return 0;
704 }
705 
706 void print_verbose_header(void)
707 {
708 	unsigned long long msr;
709 	unsigned int ratio;
710 
711 	if (!do_nehalem_platform_info)
712 		return;
713 
714 	get_msr(0, MSR_NEHALEM_PLATFORM_INFO, &msr);
715 
716 	if (verbose > 1)
717 		fprintf(stderr, "MSR_NEHALEM_PLATFORM_INFO: 0x%llx\n", msr);
718 
719 	ratio = (msr >> 40) & 0xFF;
720 	fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
721 		ratio, bclk, ratio * bclk);
722 
723 	ratio = (msr >> 8) & 0xFF;
724 	fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
725 		ratio, bclk, ratio * bclk);
726 
727 	if (!do_ivt_turbo_ratio_limit)
728 		goto print_nhm_turbo_ratio_limits;
729 
730 	get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr);
731 
732 	if (verbose > 1)
733 		fprintf(stderr, "MSR_IVT_TURBO_RATIO_LIMIT: 0x%llx\n", msr);
734 
735 	ratio = (msr >> 56) & 0xFF;
736 	if (ratio)
737 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
738 			ratio, bclk, ratio * bclk);
739 
740 	ratio = (msr >> 48) & 0xFF;
741 	if (ratio)
742 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
743 			ratio, bclk, ratio * bclk);
744 
745 	ratio = (msr >> 40) & 0xFF;
746 	if (ratio)
747 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
748 			ratio, bclk, ratio * bclk);
749 
750 	ratio = (msr >> 32) & 0xFF;
751 	if (ratio)
752 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
753 			ratio, bclk, ratio * bclk);
754 
755 	ratio = (msr >> 24) & 0xFF;
756 	if (ratio)
757 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
758 			ratio, bclk, ratio * bclk);
759 
760 	ratio = (msr >> 16) & 0xFF;
761 	if (ratio)
762 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
763 			ratio, bclk, ratio * bclk);
764 
765 	ratio = (msr >> 8) & 0xFF;
766 	if (ratio)
767 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
768 			ratio, bclk, ratio * bclk);
769 
770 	ratio = (msr >> 0) & 0xFF;
771 	if (ratio)
772 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
773 			ratio, bclk, ratio * bclk);
774 
775 print_nhm_turbo_ratio_limits:
776 
777 	if (!do_nehalem_turbo_ratio_limit)
778 		return;
779 
780 	get_msr(0, MSR_NEHALEM_TURBO_RATIO_LIMIT, &msr);
781 
782 	if (verbose > 1)
783 		fprintf(stderr, "MSR_NEHALEM_TURBO_RATIO_LIMIT: 0x%llx\n", msr);
784 
785 	ratio = (msr >> 56) & 0xFF;
786 	if (ratio)
787 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
788 			ratio, bclk, ratio * bclk);
789 
790 	ratio = (msr >> 48) & 0xFF;
791 	if (ratio)
792 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
793 			ratio, bclk, ratio * bclk);
794 
795 	ratio = (msr >> 40) & 0xFF;
796 	if (ratio)
797 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
798 			ratio, bclk, ratio * bclk);
799 
800 	ratio = (msr >> 32) & 0xFF;
801 	if (ratio)
802 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
803 			ratio, bclk, ratio * bclk);
804 
805 	ratio = (msr >> 24) & 0xFF;
806 	if (ratio)
807 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
808 			ratio, bclk, ratio * bclk);
809 
810 	ratio = (msr >> 16) & 0xFF;
811 	if (ratio)
812 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
813 			ratio, bclk, ratio * bclk);
814 
815 	ratio = (msr >> 8) & 0xFF;
816 	if (ratio)
817 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
818 			ratio, bclk, ratio * bclk);
819 
820 	ratio = (msr >> 0) & 0xFF;
821 	if (ratio)
822 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
823 			ratio, bclk, ratio * bclk);
824 }
825 
826 void free_all_buffers(void)
827 {
828 	CPU_FREE(cpu_present_set);
829 	cpu_present_set = NULL;
830 	cpu_present_set = 0;
831 
832 	CPU_FREE(cpu_affinity_set);
833 	cpu_affinity_set = NULL;
834 	cpu_affinity_setsize = 0;
835 
836 	free(thread_even);
837 	free(core_even);
838 	free(package_even);
839 
840 	thread_even = NULL;
841 	core_even = NULL;
842 	package_even = NULL;
843 
844 	free(thread_odd);
845 	free(core_odd);
846 	free(package_odd);
847 
848 	thread_odd = NULL;
849 	core_odd = NULL;
850 	package_odd = NULL;
851 
852 	free(output_buffer);
853 	output_buffer = NULL;
854 	outp = NULL;
855 }
856 
857 /*
858  * cpu_is_first_sibling_in_core(cpu)
859  * return 1 if given CPU is 1st HT sibling in the core
860  */
861 int cpu_is_first_sibling_in_core(int cpu)
862 {
863 	char path[64];
864 	FILE *filep;
865 	int first_cpu;
866 
867 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
868 	filep = fopen(path, "r");
869 	if (filep == NULL) {
870 		perror(path);
871 		exit(1);
872 	}
873 	fscanf(filep, "%d", &first_cpu);
874 	fclose(filep);
875 	return (cpu == first_cpu);
876 }
877 
878 /*
879  * cpu_is_first_core_in_package(cpu)
880  * return 1 if given CPU is 1st core in package
881  */
882 int cpu_is_first_core_in_package(int cpu)
883 {
884 	char path[64];
885 	FILE *filep;
886 	int first_cpu;
887 
888 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
889 	filep = fopen(path, "r");
890 	if (filep == NULL) {
891 		perror(path);
892 		exit(1);
893 	}
894 	fscanf(filep, "%d", &first_cpu);
895 	fclose(filep);
896 	return (cpu == first_cpu);
897 }
898 
899 int get_physical_package_id(int cpu)
900 {
901 	char path[80];
902 	FILE *filep;
903 	int pkg;
904 
905 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
906 	filep = fopen(path, "r");
907 	if (filep == NULL) {
908 		perror(path);
909 		exit(1);
910 	}
911 	fscanf(filep, "%d", &pkg);
912 	fclose(filep);
913 	return pkg;
914 }
915 
916 int get_core_id(int cpu)
917 {
918 	char path[80];
919 	FILE *filep;
920 	int core;
921 
922 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
923 	filep = fopen(path, "r");
924 	if (filep == NULL) {
925 		perror(path);
926 		exit(1);
927 	}
928 	fscanf(filep, "%d", &core);
929 	fclose(filep);
930 	return core;
931 }
932 
933 int get_num_ht_siblings(int cpu)
934 {
935 	char path[80];
936 	FILE *filep;
937 	int sib1, sib2;
938 	int matches;
939 	char character;
940 
941 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
942 	filep = fopen(path, "r");
943 	if (filep == NULL) {
944 		perror(path);
945 		exit(1);
946 	}
947 	/*
948 	 * file format:
949 	 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
950 	 * otherwinse 1 sibling (self).
951 	 */
952 	matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
953 
954 	fclose(filep);
955 
956 	if (matches == 3)
957 		return 2;
958 	else
959 		return 1;
960 }
961 
962 /*
963  * run func(thread, core, package) in topology order
964  * skip non-present cpus
965  */
966 
967 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
968 	struct pkg_data *, struct thread_data *, struct core_data *,
969 	struct pkg_data *), struct thread_data *thread_base,
970 	struct core_data *core_base, struct pkg_data *pkg_base,
971 	struct thread_data *thread_base2, struct core_data *core_base2,
972 	struct pkg_data *pkg_base2)
973 {
974 	int retval, pkg_no, core_no, thread_no;
975 
976 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
977 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
978 			for (thread_no = 0; thread_no <
979 				topo.num_threads_per_core; ++thread_no) {
980 				struct thread_data *t, *t2;
981 				struct core_data *c, *c2;
982 				struct pkg_data *p, *p2;
983 
984 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
985 
986 				if (cpu_is_not_present(t->cpu_id))
987 					continue;
988 
989 				t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
990 
991 				c = GET_CORE(core_base, core_no, pkg_no);
992 				c2 = GET_CORE(core_base2, core_no, pkg_no);
993 
994 				p = GET_PKG(pkg_base, pkg_no);
995 				p2 = GET_PKG(pkg_base2, pkg_no);
996 
997 				retval = func(t, c, p, t2, c2, p2);
998 				if (retval)
999 					return retval;
1000 			}
1001 		}
1002 	}
1003 	return 0;
1004 }
1005 
1006 /*
1007  * run func(cpu) on every cpu in /proc/stat
1008  * return max_cpu number
1009  */
1010 int for_all_proc_cpus(int (func)(int))
1011 {
1012 	FILE *fp;
1013 	int cpu_num;
1014 	int retval;
1015 
1016 	fp = fopen(proc_stat, "r");
1017 	if (fp == NULL) {
1018 		perror(proc_stat);
1019 		exit(1);
1020 	}
1021 
1022 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1023 	if (retval != 0) {
1024 		perror("/proc/stat format");
1025 		exit(1);
1026 	}
1027 
1028 	while (1) {
1029 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1030 		if (retval != 1)
1031 			break;
1032 
1033 		retval = func(cpu_num);
1034 		if (retval) {
1035 			fclose(fp);
1036 			return(retval);
1037 		}
1038 	}
1039 	fclose(fp);
1040 	return 0;
1041 }
1042 
1043 void re_initialize(void)
1044 {
1045 	free_all_buffers();
1046 	setup_all_buffers();
1047 	printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1048 }
1049 
1050 
1051 /*
1052  * count_cpus()
1053  * remember the last one seen, it will be the max
1054  */
1055 int count_cpus(int cpu)
1056 {
1057 	if (topo.max_cpu_num < cpu)
1058 		topo.max_cpu_num = cpu;
1059 
1060 	topo.num_cpus += 1;
1061 	return 0;
1062 }
1063 int mark_cpu_present(int cpu)
1064 {
1065 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1066 	return 0;
1067 }
1068 
1069 void turbostat_loop()
1070 {
1071 	int retval;
1072 
1073 restart:
1074 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1075 	if (retval) {
1076 		re_initialize();
1077 		goto restart;
1078 	}
1079 	gettimeofday(&tv_even, (struct timezone *)NULL);
1080 
1081 	while (1) {
1082 		if (for_all_proc_cpus(cpu_is_not_present)) {
1083 			re_initialize();
1084 			goto restart;
1085 		}
1086 		sleep(interval_sec);
1087 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
1088 		if (retval) {
1089 			re_initialize();
1090 			goto restart;
1091 		}
1092 		gettimeofday(&tv_odd, (struct timezone *)NULL);
1093 		timersub(&tv_odd, &tv_even, &tv_delta);
1094 		for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1095 		compute_average(EVEN_COUNTERS);
1096 		format_all_counters(EVEN_COUNTERS);
1097 		flush_stdout();
1098 		sleep(interval_sec);
1099 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1100 		if (retval) {
1101 			re_initialize();
1102 			goto restart;
1103 		}
1104 		gettimeofday(&tv_even, (struct timezone *)NULL);
1105 		timersub(&tv_even, &tv_odd, &tv_delta);
1106 		for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1107 		compute_average(ODD_COUNTERS);
1108 		format_all_counters(ODD_COUNTERS);
1109 		flush_stdout();
1110 	}
1111 }
1112 
1113 void check_dev_msr()
1114 {
1115 	struct stat sb;
1116 
1117 	if (stat("/dev/cpu/0/msr", &sb)) {
1118 		fprintf(stderr, "no /dev/cpu/0/msr\n");
1119 		fprintf(stderr, "Try \"# modprobe msr\"\n");
1120 		exit(-5);
1121 	}
1122 }
1123 
1124 void check_super_user()
1125 {
1126 	if (getuid() != 0) {
1127 		fprintf(stderr, "must be root\n");
1128 		exit(-6);
1129 	}
1130 }
1131 
1132 int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model)
1133 {
1134 	if (!genuine_intel)
1135 		return 0;
1136 
1137 	if (family != 6)
1138 		return 0;
1139 
1140 	switch (model) {
1141 	case 0x1A:	/* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1142 	case 0x1E:	/* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1143 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
1144 	case 0x25:	/* Westmere Client - Clarkdale, Arrandale */
1145 	case 0x2C:	/* Westmere EP - Gulftown */
1146 	case 0x2A:	/* SNB */
1147 	case 0x2D:	/* SNB Xeon */
1148 	case 0x3A:	/* IVB */
1149 	case 0x3E:	/* IVB Xeon */
1150 		return 1;
1151 	case 0x2E:	/* Nehalem-EX Xeon - Beckton */
1152 	case 0x2F:	/* Westmere-EX Xeon - Eagleton */
1153 	default:
1154 		return 0;
1155 	}
1156 }
1157 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1158 {
1159 	if (!genuine_intel)
1160 		return 0;
1161 
1162 	if (family != 6)
1163 		return 0;
1164 
1165 	switch (model) {
1166 	case 0x3E:	/* IVB Xeon */
1167 		return 1;
1168 	default:
1169 		return 0;
1170 	}
1171 }
1172 
1173 
1174 int is_snb(unsigned int family, unsigned int model)
1175 {
1176 	if (!genuine_intel)
1177 		return 0;
1178 
1179 	switch (model) {
1180 	case 0x2A:
1181 	case 0x2D:
1182 	case 0x3A:	/* IVB */
1183 	case 0x3E:	/* IVB Xeon */
1184 		return 1;
1185 	}
1186 	return 0;
1187 }
1188 
1189 double discover_bclk(unsigned int family, unsigned int model)
1190 {
1191 	if (is_snb(family, model))
1192 		return 100.00;
1193 	else
1194 		return 133.33;
1195 }
1196 
1197 void check_cpuid()
1198 {
1199 	unsigned int eax, ebx, ecx, edx, max_level;
1200 	unsigned int fms, family, model, stepping;
1201 
1202 	eax = ebx = ecx = edx = 0;
1203 
1204 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0));
1205 
1206 	if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1207 		genuine_intel = 1;
1208 
1209 	if (verbose)
1210 		fprintf(stderr, "%.4s%.4s%.4s ",
1211 			(char *)&ebx, (char *)&edx, (char *)&ecx);
1212 
1213 	asm("cpuid" : "=a" (fms), "=c" (ecx), "=d" (edx) : "a" (1) : "ebx");
1214 	family = (fms >> 8) & 0xf;
1215 	model = (fms >> 4) & 0xf;
1216 	stepping = fms & 0xf;
1217 	if (family == 6 || family == 0xf)
1218 		model += ((fms >> 16) & 0xf) << 4;
1219 
1220 	if (verbose)
1221 		fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
1222 			max_level, family, model, stepping, family, model, stepping);
1223 
1224 	if (!(edx & (1 << 5))) {
1225 		fprintf(stderr, "CPUID: no MSR\n");
1226 		exit(1);
1227 	}
1228 
1229 	/*
1230 	 * check max extended function levels of CPUID.
1231 	 * This is needed to check for invariant TSC.
1232 	 * This check is valid for both Intel and AMD.
1233 	 */
1234 	ebx = ecx = edx = 0;
1235 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000000));
1236 
1237 	if (max_level < 0x80000007) {
1238 		fprintf(stderr, "CPUID: no invariant TSC (max_level 0x%x)\n", max_level);
1239 		exit(1);
1240 	}
1241 
1242 	/*
1243 	 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1244 	 * this check is valid for both Intel and AMD
1245 	 */
1246 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000007));
1247 	has_invariant_tsc = edx & (1 << 8);
1248 
1249 	if (!has_invariant_tsc) {
1250 		fprintf(stderr, "No invariant TSC\n");
1251 		exit(1);
1252 	}
1253 
1254 	/*
1255 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1256 	 * this check is valid for both Intel and AMD
1257 	 */
1258 
1259 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x6));
1260 	has_aperf = ecx & (1 << 0);
1261 	if (!has_aperf) {
1262 		fprintf(stderr, "No APERF MSR\n");
1263 		exit(1);
1264 	}
1265 
1266 	do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1267 	do_nhm_cstates = genuine_intel;	/* all Intel w/ non-stop TSC have NHM counters */
1268 	do_snb_cstates = is_snb(family, model);
1269 	bclk = discover_bclk(family, model);
1270 
1271 	do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model);
1272 	do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model);
1273 }
1274 
1275 
1276 void usage()
1277 {
1278 	fprintf(stderr, "%s: [-v] [-m msr#] [-M MSR#] [-i interval_sec | command ...]\n",
1279 		progname);
1280 	exit(1);
1281 }
1282 
1283 
1284 /*
1285  * in /dev/cpu/ return success for names that are numbers
1286  * ie. filter out ".", "..", "microcode".
1287  */
1288 int dir_filter(const struct dirent *dirp)
1289 {
1290 	if (isdigit(dirp->d_name[0]))
1291 		return 1;
1292 	else
1293 		return 0;
1294 }
1295 
1296 int open_dev_cpu_msr(int dummy1)
1297 {
1298 	return 0;
1299 }
1300 
1301 void topology_probe()
1302 {
1303 	int i;
1304 	int max_core_id = 0;
1305 	int max_package_id = 0;
1306 	int max_siblings = 0;
1307 	struct cpu_topology {
1308 		int core_id;
1309 		int physical_package_id;
1310 	} *cpus;
1311 
1312 	/* Initialize num_cpus, max_cpu_num */
1313 	topo.num_cpus = 0;
1314 	topo.max_cpu_num = 0;
1315 	for_all_proc_cpus(count_cpus);
1316 	if (!summary_only && topo.num_cpus > 1)
1317 		show_cpu = 1;
1318 
1319 	if (verbose > 1)
1320 		fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1321 
1322 	cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1323 	if (cpus == NULL) {
1324 		perror("calloc cpus");
1325 		exit(1);
1326 	}
1327 
1328 	/*
1329 	 * Allocate and initialize cpu_present_set
1330 	 */
1331 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1332 	if (cpu_present_set == NULL) {
1333 		perror("CPU_ALLOC");
1334 		exit(3);
1335 	}
1336 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1337 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1338 	for_all_proc_cpus(mark_cpu_present);
1339 
1340 	/*
1341 	 * Allocate and initialize cpu_affinity_set
1342 	 */
1343 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1344 	if (cpu_affinity_set == NULL) {
1345 		perror("CPU_ALLOC");
1346 		exit(3);
1347 	}
1348 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1349 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1350 
1351 
1352 	/*
1353 	 * For online cpus
1354 	 * find max_core_id, max_package_id
1355 	 */
1356 	for (i = 0; i <= topo.max_cpu_num; ++i) {
1357 		int siblings;
1358 
1359 		if (cpu_is_not_present(i)) {
1360 			if (verbose > 1)
1361 				fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1362 			continue;
1363 		}
1364 		cpus[i].core_id = get_core_id(i);
1365 		if (cpus[i].core_id > max_core_id)
1366 			max_core_id = cpus[i].core_id;
1367 
1368 		cpus[i].physical_package_id = get_physical_package_id(i);
1369 		if (cpus[i].physical_package_id > max_package_id)
1370 			max_package_id = cpus[i].physical_package_id;
1371 
1372 		siblings = get_num_ht_siblings(i);
1373 		if (siblings > max_siblings)
1374 			max_siblings = siblings;
1375 		if (verbose > 1)
1376 			fprintf(stderr, "cpu %d pkg %d core %d\n",
1377 				i, cpus[i].physical_package_id, cpus[i].core_id);
1378 	}
1379 	topo.num_cores_per_pkg = max_core_id + 1;
1380 	if (verbose > 1)
1381 		fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
1382 			max_core_id, topo.num_cores_per_pkg);
1383 	if (!summary_only && topo.num_cores_per_pkg > 1)
1384 		show_core = 1;
1385 
1386 	topo.num_packages = max_package_id + 1;
1387 	if (verbose > 1)
1388 		fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
1389 			max_package_id, topo.num_packages);
1390 	if (!summary_only && topo.num_packages > 1)
1391 		show_pkg = 1;
1392 
1393 	topo.num_threads_per_core = max_siblings;
1394 	if (verbose > 1)
1395 		fprintf(stderr, "max_siblings %d\n", max_siblings);
1396 
1397 	free(cpus);
1398 }
1399 
1400 void
1401 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1402 {
1403 	int i;
1404 
1405 	*t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1406 		topo.num_packages, sizeof(struct thread_data));
1407 	if (*t == NULL)
1408 		goto error;
1409 
1410 	for (i = 0; i < topo.num_threads_per_core *
1411 		topo.num_cores_per_pkg * topo.num_packages; i++)
1412 		(*t)[i].cpu_id = -1;
1413 
1414 	*c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1415 		sizeof(struct core_data));
1416 	if (*c == NULL)
1417 		goto error;
1418 
1419 	for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1420 		(*c)[i].core_id = -1;
1421 
1422 	*p = calloc(topo.num_packages, sizeof(struct pkg_data));
1423 	if (*p == NULL)
1424 		goto error;
1425 
1426 	for (i = 0; i < topo.num_packages; i++)
1427 		(*p)[i].package_id = i;
1428 
1429 	return;
1430 error:
1431 	perror("calloc counters");
1432 	exit(1);
1433 }
1434 /*
1435  * init_counter()
1436  *
1437  * set cpu_id, core_num, pkg_num
1438  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1439  *
1440  * increment topo.num_cores when 1st core in pkg seen
1441  */
1442 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
1443 	struct pkg_data *pkg_base, int thread_num, int core_num,
1444 	int pkg_num, int cpu_id)
1445 {
1446 	struct thread_data *t;
1447 	struct core_data *c;
1448 	struct pkg_data *p;
1449 
1450 	t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1451 	c = GET_CORE(core_base, core_num, pkg_num);
1452 	p = GET_PKG(pkg_base, pkg_num);
1453 
1454 	t->cpu_id = cpu_id;
1455 	if (thread_num == 0) {
1456 		t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1457 		if (cpu_is_first_core_in_package(cpu_id))
1458 			t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1459 	}
1460 
1461 	c->core_id = core_num;
1462 	p->package_id = pkg_num;
1463 }
1464 
1465 
1466 int initialize_counters(int cpu_id)
1467 {
1468 	int my_thread_id, my_core_id, my_package_id;
1469 
1470 	my_package_id = get_physical_package_id(cpu_id);
1471 	my_core_id = get_core_id(cpu_id);
1472 
1473 	if (cpu_is_first_sibling_in_core(cpu_id)) {
1474 		my_thread_id = 0;
1475 		topo.num_cores++;
1476 	} else {
1477 		my_thread_id = 1;
1478 	}
1479 
1480 	init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1481 	init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1482 	return 0;
1483 }
1484 
1485 void allocate_output_buffer()
1486 {
1487 	output_buffer = calloc(1, (1 + topo.num_cpus) * 128);
1488 	outp = output_buffer;
1489 	if (outp == NULL) {
1490 		perror("calloc");
1491 		exit(-1);
1492 	}
1493 }
1494 
1495 void setup_all_buffers(void)
1496 {
1497 	topology_probe();
1498 	allocate_counters(&thread_even, &core_even, &package_even);
1499 	allocate_counters(&thread_odd, &core_odd, &package_odd);
1500 	allocate_output_buffer();
1501 	for_all_proc_cpus(initialize_counters);
1502 }
1503 void turbostat_init()
1504 {
1505 	check_cpuid();
1506 
1507 	check_dev_msr();
1508 	check_super_user();
1509 
1510 	setup_all_buffers();
1511 
1512 	if (verbose)
1513 		print_verbose_header();
1514 }
1515 
1516 int fork_it(char **argv)
1517 {
1518 	pid_t child_pid;
1519 
1520 	for_all_cpus(get_counters, EVEN_COUNTERS);
1521 	/* clear affinity side-effect of get_counters() */
1522 	sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
1523 	gettimeofday(&tv_even, (struct timezone *)NULL);
1524 
1525 	child_pid = fork();
1526 	if (!child_pid) {
1527 		/* child */
1528 		execvp(argv[0], argv);
1529 	} else {
1530 		int status;
1531 
1532 		/* parent */
1533 		if (child_pid == -1) {
1534 			perror("fork");
1535 			exit(1);
1536 		}
1537 
1538 		signal(SIGINT, SIG_IGN);
1539 		signal(SIGQUIT, SIG_IGN);
1540 		if (waitpid(child_pid, &status, 0) == -1) {
1541 			perror("wait");
1542 			exit(1);
1543 		}
1544 	}
1545 	/*
1546 	 * n.b. fork_it() does not check for errors from for_all_cpus()
1547 	 * because re-starting is problematic when forking
1548 	 */
1549 	for_all_cpus(get_counters, ODD_COUNTERS);
1550 	gettimeofday(&tv_odd, (struct timezone *)NULL);
1551 	timersub(&tv_odd, &tv_even, &tv_delta);
1552 	for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1553 	compute_average(EVEN_COUNTERS);
1554 	format_all_counters(EVEN_COUNTERS);
1555 	flush_stderr();
1556 
1557 	fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
1558 
1559 	return 0;
1560 }
1561 
1562 void cmdline(int argc, char **argv)
1563 {
1564 	int opt;
1565 
1566 	progname = argv[0];
1567 
1568 	while ((opt = getopt(argc, argv, "+cpsvi:m:M:")) != -1) {
1569 		switch (opt) {
1570 		case 'c':
1571 			show_core_only++;
1572 			break;
1573 		case 'p':
1574 			show_pkg_only++;
1575 			break;
1576 		case 's':
1577 			summary_only++;
1578 			break;
1579 		case 'v':
1580 			verbose++;
1581 			break;
1582 		case 'i':
1583 			interval_sec = atoi(optarg);
1584 			break;
1585 		case 'm':
1586 			sscanf(optarg, "%x", &extra_msr_offset32);
1587 			if (verbose > 1)
1588 				fprintf(stderr, "msr 0x%X\n", extra_msr_offset32);
1589 			break;
1590 		case 'M':
1591 			sscanf(optarg, "%x", &extra_msr_offset64);
1592 			if (verbose > 1)
1593 				fprintf(stderr, "MSR 0x%X\n", extra_msr_offset64);
1594 			break;
1595 		default:
1596 			usage();
1597 		}
1598 	}
1599 }
1600 
1601 int main(int argc, char **argv)
1602 {
1603 	cmdline(argc, argv);
1604 
1605 	if (verbose > 1)
1606 		fprintf(stderr, "turbostat v2.0 May 16, 2012"
1607 			" - Len Brown <lenb@kernel.org>\n");
1608 
1609 	turbostat_init();
1610 
1611 	/*
1612 	 * if any params left, it must be a command to fork
1613 	 */
1614 	if (argc - optind)
1615 		return fork_it(argv + optind);
1616 	else
1617 		turbostat_loop();
1618 
1619 	return 0;
1620 }
1621