1 /*
2  * turbostat -- show CPU frequency and C-state residency
3  * on modern Intel turbo-capable processors.
4  *
5  * Copyright (c) 2010, 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 #include <stdio.h>
23 #include <unistd.h>
24 #include <sys/types.h>
25 #include <sys/wait.h>
26 #include <sys/stat.h>
27 #include <sys/resource.h>
28 #include <fcntl.h>
29 #include <signal.h>
30 #include <sys/time.h>
31 #include <stdlib.h>
32 #include <dirent.h>
33 #include <string.h>
34 #include <ctype.h>
35 
36 #define MSR_TSC	0x10
37 #define MSR_NEHALEM_PLATFORM_INFO	0xCE
38 #define MSR_NEHALEM_TURBO_RATIO_LIMIT	0x1AD
39 #define MSR_APERF	0xE8
40 #define MSR_MPERF	0xE7
41 #define MSR_PKG_C2_RESIDENCY	0x60D	/* SNB only */
42 #define MSR_PKG_C3_RESIDENCY	0x3F8
43 #define MSR_PKG_C6_RESIDENCY	0x3F9
44 #define MSR_PKG_C7_RESIDENCY	0x3FA	/* SNB only */
45 #define MSR_CORE_C3_RESIDENCY	0x3FC
46 #define MSR_CORE_C6_RESIDENCY	0x3FD
47 #define MSR_CORE_C7_RESIDENCY	0x3FE	/* SNB only */
48 
49 char *proc_stat = "/proc/stat";
50 unsigned int interval_sec = 5;	/* set with -i interval_sec */
51 unsigned int verbose;		/* set with -v */
52 unsigned int skip_c0;
53 unsigned int skip_c1;
54 unsigned int do_nhm_cstates;
55 unsigned int do_snb_cstates;
56 unsigned int has_aperf;
57 unsigned int units = 1000000000;	/* Ghz etc */
58 unsigned int genuine_intel;
59 unsigned int has_invariant_tsc;
60 unsigned int do_nehalem_platform_info;
61 unsigned int do_nehalem_turbo_ratio_limit;
62 unsigned int extra_msr_offset;
63 double bclk;
64 unsigned int show_pkg;
65 unsigned int show_core;
66 unsigned int show_cpu;
67 
68 int aperf_mperf_unstable;
69 int backwards_count;
70 char *progname;
71 int need_reinitialize;
72 
73 int num_cpus;
74 
75 struct counters {
76 	unsigned long long tsc;		/* per thread */
77 	unsigned long long aperf;	/* per thread */
78 	unsigned long long mperf;	/* per thread */
79 	unsigned long long c1;	/* per thread (calculated) */
80 	unsigned long long c3;	/* per core */
81 	unsigned long long c6;	/* per core */
82 	unsigned long long c7;	/* per core */
83 	unsigned long long pc2;	/* per package */
84 	unsigned long long pc3;	/* per package */
85 	unsigned long long pc6;	/* per package */
86 	unsigned long long pc7;	/* per package */
87 	unsigned long long extra_msr;	/* per thread */
88 	int pkg;
89 	int core;
90 	int cpu;
91 	struct counters *next;
92 };
93 
94 struct counters *cnt_even;
95 struct counters *cnt_odd;
96 struct counters *cnt_delta;
97 struct counters *cnt_average;
98 struct timeval tv_even;
99 struct timeval tv_odd;
100 struct timeval tv_delta;
101 
102 unsigned long long get_msr(int cpu, off_t offset)
103 {
104 	ssize_t retval;
105 	unsigned long long msr;
106 	char pathname[32];
107 	int fd;
108 
109 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
110 	fd = open(pathname, O_RDONLY);
111 	if (fd < 0) {
112 		perror(pathname);
113 		need_reinitialize = 1;
114 		return 0;
115 	}
116 
117 	retval = pread(fd, &msr, sizeof msr, offset);
118 	if (retval != sizeof msr) {
119 		fprintf(stderr, "cpu%d pread(..., 0x%zx) = %jd\n",
120 			cpu, offset, retval);
121 		exit(-2);
122 	}
123 
124 	close(fd);
125 	return msr;
126 }
127 
128 void print_header(void)
129 {
130 	if (show_pkg)
131 		fprintf(stderr, "pk");
132 	if (show_core)
133 		fprintf(stderr, " cr");
134 	if (show_cpu)
135 		fprintf(stderr, " CPU");
136 	if (do_nhm_cstates)
137 		fprintf(stderr, "    %%c0 ");
138 	if (has_aperf)
139 		fprintf(stderr, " GHz");
140 	fprintf(stderr, "  TSC");
141 	if (do_nhm_cstates)
142 		fprintf(stderr, "    %%c1");
143 	if (do_nhm_cstates)
144 		fprintf(stderr, "    %%c3");
145 	if (do_nhm_cstates)
146 		fprintf(stderr, "    %%c6");
147 	if (do_snb_cstates)
148 		fprintf(stderr, "    %%c7");
149 	if (do_snb_cstates)
150 		fprintf(stderr, "  %%pc2");
151 	if (do_nhm_cstates)
152 		fprintf(stderr, "  %%pc3");
153 	if (do_nhm_cstates)
154 		fprintf(stderr, "  %%pc6");
155 	if (do_snb_cstates)
156 		fprintf(stderr, "  %%pc7");
157 	if (extra_msr_offset)
158 		fprintf(stderr, "        MSR 0x%x ", extra_msr_offset);
159 
160 	putc('\n', stderr);
161 }
162 
163 void dump_cnt(struct counters *cnt)
164 {
165 	if (!cnt)
166 		return;
167 	if (cnt->pkg) fprintf(stderr, "package: %d ", cnt->pkg);
168 	if (cnt->core) fprintf(stderr, "core:: %d ", cnt->core);
169 	if (cnt->cpu) fprintf(stderr, "CPU: %d ", cnt->cpu);
170 	if (cnt->tsc) fprintf(stderr, "TSC: %016llX\n", cnt->tsc);
171 	if (cnt->c3) fprintf(stderr, "c3: %016llX\n", cnt->c3);
172 	if (cnt->c6) fprintf(stderr, "c6: %016llX\n", cnt->c6);
173 	if (cnt->c7) fprintf(stderr, "c7: %016llX\n", cnt->c7);
174 	if (cnt->aperf) fprintf(stderr, "aperf: %016llX\n", cnt->aperf);
175 	if (cnt->pc2) fprintf(stderr, "pc2: %016llX\n", cnt->pc2);
176 	if (cnt->pc3) fprintf(stderr, "pc3: %016llX\n", cnt->pc3);
177 	if (cnt->pc6) fprintf(stderr, "pc6: %016llX\n", cnt->pc6);
178 	if (cnt->pc7) fprintf(stderr, "pc7: %016llX\n", cnt->pc7);
179 	if (cnt->extra_msr) fprintf(stderr, "msr0x%x: %016llX\n", extra_msr_offset, cnt->extra_msr);
180 }
181 
182 void dump_list(struct counters *cnt)
183 {
184 	printf("dump_list 0x%p\n", cnt);
185 
186 	for (; cnt; cnt = cnt->next)
187 		dump_cnt(cnt);
188 }
189 
190 void print_cnt(struct counters *p)
191 {
192 	double interval_float;
193 
194 	interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
195 
196 	/* topology columns, print blanks on 1st (average) line */
197 	if (p == cnt_average) {
198 		if (show_pkg)
199 			fprintf(stderr, " ");
200 		if (show_core)
201 			fprintf(stderr, "    ");
202 		if (show_cpu)
203 			fprintf(stderr, "    ");
204 	} else {
205 		if (show_pkg)
206 			fprintf(stderr, "%d", p->pkg);
207 		if (show_core)
208 			fprintf(stderr, "%4d", p->core);
209 		if (show_cpu)
210 			fprintf(stderr, "%4d", p->cpu);
211 	}
212 
213 	/* %c0 */
214 	if (do_nhm_cstates) {
215 		if (!skip_c0)
216 			fprintf(stderr, "%7.2f", 100.0 * p->mperf/p->tsc);
217 		else
218 			fprintf(stderr, "   ****");
219 	}
220 
221 	/* GHz */
222 	if (has_aperf) {
223 		if (!aperf_mperf_unstable) {
224 			fprintf(stderr, "%5.2f",
225 				1.0 * p->tsc / units * p->aperf /
226 				p->mperf / interval_float);
227 		} else {
228 			if (p->aperf > p->tsc || p->mperf > p->tsc) {
229 				fprintf(stderr, " ****");
230 			} else {
231 				fprintf(stderr, "%4.1f*",
232 					1.0 * p->tsc /
233 					units * p->aperf /
234 					p->mperf / interval_float);
235 			}
236 		}
237 	}
238 
239 	/* TSC */
240 	fprintf(stderr, "%5.2f", 1.0 * p->tsc/units/interval_float);
241 
242 	if (do_nhm_cstates) {
243 		if (!skip_c1)
244 			fprintf(stderr, "%7.2f", 100.0 * p->c1/p->tsc);
245 		else
246 			fprintf(stderr, "  ****");
247 	}
248 	if (do_nhm_cstates)
249 		fprintf(stderr, " %6.2f", 100.0 * p->c3/p->tsc);
250 	if (do_nhm_cstates)
251 		fprintf(stderr, " %6.2f", 100.0 * p->c6/p->tsc);
252 	if (do_snb_cstates)
253 		fprintf(stderr, " %6.2f", 100.0 * p->c7/p->tsc);
254 	if (do_snb_cstates)
255 		fprintf(stderr, " %5.2f", 100.0 * p->pc2/p->tsc);
256 	if (do_nhm_cstates)
257 		fprintf(stderr, " %5.2f", 100.0 * p->pc3/p->tsc);
258 	if (do_nhm_cstates)
259 		fprintf(stderr, " %5.2f", 100.0 * p->pc6/p->tsc);
260 	if (do_snb_cstates)
261 		fprintf(stderr, " %5.2f", 100.0 * p->pc7/p->tsc);
262 	if (extra_msr_offset)
263 		fprintf(stderr, "  0x%016llx", p->extra_msr);
264 	putc('\n', stderr);
265 }
266 
267 void print_counters(struct counters *counters)
268 {
269 	struct counters *cnt;
270 
271 	print_header();
272 
273 	if (num_cpus > 1)
274 		print_cnt(cnt_average);
275 
276 	for (cnt = counters; cnt != NULL; cnt = cnt->next)
277 		print_cnt(cnt);
278 
279 }
280 
281 #define SUBTRACT_COUNTER(after, before, delta) (delta = (after - before), (before > after))
282 
283 int compute_delta(struct counters *after,
284 	struct counters *before, struct counters *delta)
285 {
286 	int errors = 0;
287 	int perf_err = 0;
288 
289 	skip_c0 = skip_c1 = 0;
290 
291 	for ( ; after && before && delta;
292 		after = after->next, before = before->next, delta = delta->next) {
293 		if (before->cpu != after->cpu) {
294 			printf("cpu configuration changed: %d != %d\n",
295 				before->cpu, after->cpu);
296 			return -1;
297 		}
298 
299 		if (SUBTRACT_COUNTER(after->tsc, before->tsc, delta->tsc)) {
300 			fprintf(stderr, "cpu%d TSC went backwards %llX to %llX\n",
301 				before->cpu, before->tsc, after->tsc);
302 			errors++;
303 		}
304 		/* check for TSC < 1 Mcycles over interval */
305 		if (delta->tsc < (1000 * 1000)) {
306 			fprintf(stderr, "Insanely slow TSC rate,"
307 				" TSC stops in idle?\n");
308 			fprintf(stderr, "You can disable all c-states"
309 				" by booting with \"idle=poll\"\n");
310 			fprintf(stderr, "or just the deep ones with"
311 				" \"processor.max_cstate=1\"\n");
312 			exit(-3);
313 		}
314 		if (SUBTRACT_COUNTER(after->c3, before->c3, delta->c3)) {
315 			fprintf(stderr, "cpu%d c3 counter went backwards %llX to %llX\n",
316 				before->cpu, before->c3, after->c3);
317 			errors++;
318 		}
319 		if (SUBTRACT_COUNTER(after->c6, before->c6, delta->c6)) {
320 			fprintf(stderr, "cpu%d c6 counter went backwards %llX to %llX\n",
321 				before->cpu, before->c6, after->c6);
322 			errors++;
323 		}
324 		if (SUBTRACT_COUNTER(after->c7, before->c7, delta->c7)) {
325 			fprintf(stderr, "cpu%d c7 counter went backwards %llX to %llX\n",
326 				before->cpu, before->c7, after->c7);
327 			errors++;
328 		}
329 		if (SUBTRACT_COUNTER(after->pc2, before->pc2, delta->pc2)) {
330 			fprintf(stderr, "cpu%d pc2 counter went backwards %llX to %llX\n",
331 				before->cpu, before->pc2, after->pc2);
332 			errors++;
333 		}
334 		if (SUBTRACT_COUNTER(after->pc3, before->pc3, delta->pc3)) {
335 			fprintf(stderr, "cpu%d pc3 counter went backwards %llX to %llX\n",
336 				before->cpu, before->pc3, after->pc3);
337 			errors++;
338 		}
339 		if (SUBTRACT_COUNTER(after->pc6, before->pc6, delta->pc6)) {
340 			fprintf(stderr, "cpu%d pc6 counter went backwards %llX to %llX\n",
341 				before->cpu, before->pc6, after->pc6);
342 			errors++;
343 		}
344 		if (SUBTRACT_COUNTER(after->pc7, before->pc7, delta->pc7)) {
345 			fprintf(stderr, "cpu%d pc7 counter went backwards %llX to %llX\n",
346 				before->cpu, before->pc7, after->pc7);
347 			errors++;
348 		}
349 
350 		perf_err = SUBTRACT_COUNTER(after->aperf, before->aperf, delta->aperf);
351 		if (perf_err) {
352 			fprintf(stderr, "cpu%d aperf counter went backwards %llX to %llX\n",
353 				before->cpu, before->aperf, after->aperf);
354 		}
355 		perf_err |= SUBTRACT_COUNTER(after->mperf, before->mperf, delta->mperf);
356 		if (perf_err) {
357 			fprintf(stderr, "cpu%d mperf counter went backwards %llX to %llX\n",
358 				before->cpu, before->mperf, after->mperf);
359 		}
360 		if (perf_err) {
361 			if (!aperf_mperf_unstable) {
362 				fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
363 				fprintf(stderr, "* Frequency results do not cover entire interval *\n");
364 				fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
365 
366 				aperf_mperf_unstable = 1;
367 			}
368 			/*
369 			 * mperf delta is likely a huge "positive" number
370 			 * can not use it for calculating c0 time
371 			 */
372 			skip_c0 = 1;
373 			skip_c1 = 1;
374 		}
375 
376 		/*
377 		 * As mperf and tsc collection are not atomic,
378 		 * it is possible for mperf's non-halted cycles
379 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
380 		 */
381 		if (delta->mperf > delta->tsc)
382 			delta->c1 = 0;
383 		else /* normal case, derive c1 */
384 			delta->c1 = delta->tsc - delta->mperf
385 				- delta->c3 - delta->c6 - delta->c7;
386 
387 		if (delta->mperf == 0)
388 			delta->mperf = 1;	/* divide by 0 protection */
389 
390 		/*
391 		 * for "extra msr", just copy the latest w/o subtracting
392 		 */
393 		delta->extra_msr = after->extra_msr;
394 		if (errors) {
395 			fprintf(stderr, "ERROR cpu%d before:\n", before->cpu);
396 			dump_cnt(before);
397 			fprintf(stderr, "ERROR cpu%d after:\n", before->cpu);
398 			dump_cnt(after);
399 			errors = 0;
400 		}
401 	}
402 	return 0;
403 }
404 
405 void compute_average(struct counters *delta, struct counters *avg)
406 {
407 	struct counters *sum;
408 
409 	sum = calloc(1, sizeof(struct counters));
410 	if (sum == NULL) {
411 		perror("calloc sum");
412 		exit(1);
413 	}
414 
415 	for (; delta; delta = delta->next) {
416 		sum->tsc += delta->tsc;
417 		sum->c1 += delta->c1;
418 		sum->c3 += delta->c3;
419 		sum->c6 += delta->c6;
420 		sum->c7 += delta->c7;
421 		sum->aperf += delta->aperf;
422 		sum->mperf += delta->mperf;
423 		sum->pc2 += delta->pc2;
424 		sum->pc3 += delta->pc3;
425 		sum->pc6 += delta->pc6;
426 		sum->pc7 += delta->pc7;
427 	}
428 	avg->tsc = sum->tsc/num_cpus;
429 	avg->c1 = sum->c1/num_cpus;
430 	avg->c3 = sum->c3/num_cpus;
431 	avg->c6 = sum->c6/num_cpus;
432 	avg->c7 = sum->c7/num_cpus;
433 	avg->aperf = sum->aperf/num_cpus;
434 	avg->mperf = sum->mperf/num_cpus;
435 	avg->pc2 = sum->pc2/num_cpus;
436 	avg->pc3 = sum->pc3/num_cpus;
437 	avg->pc6 = sum->pc6/num_cpus;
438 	avg->pc7 = sum->pc7/num_cpus;
439 
440 	free(sum);
441 }
442 
443 void get_counters(struct counters *cnt)
444 {
445 	for ( ; cnt; cnt = cnt->next) {
446 		cnt->tsc = get_msr(cnt->cpu, MSR_TSC);
447 		if (do_nhm_cstates)
448 			cnt->c3 = get_msr(cnt->cpu, MSR_CORE_C3_RESIDENCY);
449 		if (do_nhm_cstates)
450 			cnt->c6 = get_msr(cnt->cpu, MSR_CORE_C6_RESIDENCY);
451 		if (do_snb_cstates)
452 			cnt->c7 = get_msr(cnt->cpu, MSR_CORE_C7_RESIDENCY);
453 		if (has_aperf)
454 			cnt->aperf = get_msr(cnt->cpu, MSR_APERF);
455 		if (has_aperf)
456 			cnt->mperf = get_msr(cnt->cpu, MSR_MPERF);
457 		if (do_snb_cstates)
458 			cnt->pc2 = get_msr(cnt->cpu, MSR_PKG_C2_RESIDENCY);
459 		if (do_nhm_cstates)
460 			cnt->pc3 = get_msr(cnt->cpu, MSR_PKG_C3_RESIDENCY);
461 		if (do_nhm_cstates)
462 			cnt->pc6 = get_msr(cnt->cpu, MSR_PKG_C6_RESIDENCY);
463 		if (do_snb_cstates)
464 			cnt->pc7 = get_msr(cnt->cpu, MSR_PKG_C7_RESIDENCY);
465 		if (extra_msr_offset)
466 			cnt->extra_msr = get_msr(cnt->cpu, extra_msr_offset);
467 	}
468 }
469 
470 void print_nehalem_info(void)
471 {
472 	unsigned long long msr;
473 	unsigned int ratio;
474 
475 	if (!do_nehalem_platform_info)
476 		return;
477 
478 	msr = get_msr(0, MSR_NEHALEM_PLATFORM_INFO);
479 
480 	ratio = (msr >> 40) & 0xFF;
481 	fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
482 		ratio, bclk, ratio * bclk);
483 
484 	ratio = (msr >> 8) & 0xFF;
485 	fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
486 		ratio, bclk, ratio * bclk);
487 
488 	if (verbose > 1)
489 		fprintf(stderr, "MSR_NEHALEM_PLATFORM_INFO: 0x%llx\n", msr);
490 
491 	if (!do_nehalem_turbo_ratio_limit)
492 		return;
493 
494 	msr = get_msr(0, MSR_NEHALEM_TURBO_RATIO_LIMIT);
495 
496 	ratio = (msr >> 24) & 0xFF;
497 	if (ratio)
498 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
499 			ratio, bclk, ratio * bclk);
500 
501 	ratio = (msr >> 16) & 0xFF;
502 	if (ratio)
503 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
504 			ratio, bclk, ratio * bclk);
505 
506 	ratio = (msr >> 8) & 0xFF;
507 	if (ratio)
508 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
509 			ratio, bclk, ratio * bclk);
510 
511 	ratio = (msr >> 0) & 0xFF;
512 	if (ratio)
513 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
514 			ratio, bclk, ratio * bclk);
515 
516 }
517 
518 void free_counter_list(struct counters *list)
519 {
520 	struct counters *p;
521 
522 	for (p = list; p; ) {
523 		struct counters *free_me;
524 
525 		free_me = p;
526 		p = p->next;
527 		free(free_me);
528 	}
529 }
530 
531 void free_all_counters(void)
532 {
533 	free_counter_list(cnt_even);
534 	cnt_even = NULL;
535 
536 	free_counter_list(cnt_odd);
537 	cnt_odd = NULL;
538 
539 	free_counter_list(cnt_delta);
540 	cnt_delta = NULL;
541 
542 	free_counter_list(cnt_average);
543 	cnt_average = NULL;
544 }
545 
546 void insert_counters(struct counters **list,
547 	struct counters *new)
548 {
549 	struct counters *prev;
550 
551 	/*
552 	 * list was empty
553 	 */
554 	if (*list == NULL) {
555 		new->next = *list;
556 		*list = new;
557 		return;
558 	}
559 
560 	show_cpu = 1;	/* there is more than one CPU */
561 
562 	/*
563 	 * insert on front of list.
564 	 * It is sorted by ascending package#, core#, cpu#
565 	 */
566 	if (((*list)->pkg > new->pkg) ||
567 	    (((*list)->pkg == new->pkg) && ((*list)->core > new->core)) ||
568 	    (((*list)->pkg == new->pkg) && ((*list)->core == new->core) && ((*list)->cpu > new->cpu))) {
569 		new->next = *list;
570 		*list = new;
571 		return;
572 	}
573 
574 	prev = *list;
575 
576 	while (prev->next && (prev->next->pkg < new->pkg)) {
577 		prev = prev->next;
578 		show_pkg = 1;	/* there is more than 1 package */
579 	}
580 
581 	while (prev->next && (prev->next->pkg == new->pkg)
582 		&& (prev->next->core < new->core)) {
583 		prev = prev->next;
584 		show_core = 1;	/* there is more than 1 core */
585 	}
586 
587 	while (prev->next && (prev->next->pkg == new->pkg)
588 		&& (prev->next->core == new->core)
589 		&& (prev->next->cpu < new->cpu)) {
590 		prev = prev->next;
591 	}
592 
593 	/*
594 	 * insert after "prev"
595 	 */
596 	new->next = prev->next;
597 	prev->next = new;
598 }
599 
600 void alloc_new_counters(int pkg, int core, int cpu)
601 {
602 	struct counters *new;
603 
604 	if (verbose > 1)
605 		printf("pkg%d core%d, cpu%d\n", pkg, core, cpu);
606 
607 	new = (struct counters *)calloc(1, sizeof(struct counters));
608 	if (new == NULL) {
609 		perror("calloc");
610 		exit(1);
611 	}
612 	new->pkg = pkg;
613 	new->core = core;
614 	new->cpu = cpu;
615 	insert_counters(&cnt_odd, new);
616 
617 	new = (struct counters *)calloc(1,
618 		sizeof(struct counters));
619 	if (new == NULL) {
620 		perror("calloc");
621 		exit(1);
622 	}
623 	new->pkg = pkg;
624 	new->core = core;
625 	new->cpu = cpu;
626 	insert_counters(&cnt_even, new);
627 
628 	new = (struct counters *)calloc(1, sizeof(struct counters));
629 	if (new == NULL) {
630 		perror("calloc");
631 		exit(1);
632 	}
633 	new->pkg = pkg;
634 	new->core = core;
635 	new->cpu = cpu;
636 	insert_counters(&cnt_delta, new);
637 
638 	new = (struct counters *)calloc(1, sizeof(struct counters));
639 	if (new == NULL) {
640 		perror("calloc");
641 		exit(1);
642 	}
643 	new->pkg = pkg;
644 	new->core = core;
645 	new->cpu = cpu;
646 	cnt_average = new;
647 }
648 
649 int get_physical_package_id(int cpu)
650 {
651 	char path[64];
652 	FILE *filep;
653 	int pkg;
654 
655 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
656 	filep = fopen(path, "r");
657 	if (filep == NULL) {
658 		perror(path);
659 		exit(1);
660 	}
661 	fscanf(filep, "%d", &pkg);
662 	fclose(filep);
663 	return pkg;
664 }
665 
666 int get_core_id(int cpu)
667 {
668 	char path[64];
669 	FILE *filep;
670 	int core;
671 
672 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
673 	filep = fopen(path, "r");
674 	if (filep == NULL) {
675 		perror(path);
676 		exit(1);
677 	}
678 	fscanf(filep, "%d", &core);
679 	fclose(filep);
680 	return core;
681 }
682 
683 /*
684  * run func(index, cpu) on every cpu in /proc/stat
685  */
686 
687 int for_all_cpus(void (func)(int, int, int))
688 {
689 	FILE *fp;
690 	int cpu_count;
691 	int retval;
692 
693 	fp = fopen(proc_stat, "r");
694 	if (fp == NULL) {
695 		perror(proc_stat);
696 		exit(1);
697 	}
698 
699 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
700 	if (retval != 0) {
701 		perror("/proc/stat format");
702 		exit(1);
703 	}
704 
705 	for (cpu_count = 0; ; cpu_count++) {
706 		int cpu;
707 
708 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu);
709 		if (retval != 1)
710 			break;
711 
712 		func(get_physical_package_id(cpu), get_core_id(cpu), cpu);
713 	}
714 	fclose(fp);
715 	return cpu_count;
716 }
717 
718 void re_initialize(void)
719 {
720 	printf("turbostat: topology changed, re-initializing.\n");
721 	free_all_counters();
722 	num_cpus = for_all_cpus(alloc_new_counters);
723 	need_reinitialize = 0;
724 	printf("num_cpus is now %d\n", num_cpus);
725 }
726 
727 void dummy(int pkg, int core, int cpu) { return; }
728 /*
729  * check to see if a cpu came on-line
730  */
731 void verify_num_cpus(void)
732 {
733 	int new_num_cpus;
734 
735 	new_num_cpus = for_all_cpus(dummy);
736 
737 	if (new_num_cpus != num_cpus) {
738 		if (verbose)
739 			printf("num_cpus was %d, is now  %d\n",
740 				num_cpus, new_num_cpus);
741 		need_reinitialize = 1;
742 	}
743 }
744 
745 void turbostat_loop()
746 {
747 restart:
748 	get_counters(cnt_even);
749 	gettimeofday(&tv_even, (struct timezone *)NULL);
750 
751 	while (1) {
752 		verify_num_cpus();
753 		if (need_reinitialize) {
754 			re_initialize();
755 			goto restart;
756 		}
757 		sleep(interval_sec);
758 		get_counters(cnt_odd);
759 		gettimeofday(&tv_odd, (struct timezone *)NULL);
760 
761 		compute_delta(cnt_odd, cnt_even, cnt_delta);
762 		timersub(&tv_odd, &tv_even, &tv_delta);
763 		compute_average(cnt_delta, cnt_average);
764 		print_counters(cnt_delta);
765 		if (need_reinitialize) {
766 			re_initialize();
767 			goto restart;
768 		}
769 		sleep(interval_sec);
770 		get_counters(cnt_even);
771 		gettimeofday(&tv_even, (struct timezone *)NULL);
772 		compute_delta(cnt_even, cnt_odd, cnt_delta);
773 		timersub(&tv_even, &tv_odd, &tv_delta);
774 		compute_average(cnt_delta, cnt_average);
775 		print_counters(cnt_delta);
776 	}
777 }
778 
779 void check_dev_msr()
780 {
781 	struct stat sb;
782 
783 	if (stat("/dev/cpu/0/msr", &sb)) {
784 		fprintf(stderr, "no /dev/cpu/0/msr\n");
785 		fprintf(stderr, "Try \"# modprobe msr\"\n");
786 		exit(-5);
787 	}
788 }
789 
790 void check_super_user()
791 {
792 	if (getuid() != 0) {
793 		fprintf(stderr, "must be root\n");
794 		exit(-6);
795 	}
796 }
797 
798 int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model)
799 {
800 	if (!genuine_intel)
801 		return 0;
802 
803 	if (family != 6)
804 		return 0;
805 
806 	switch (model) {
807 	case 0x1A:	/* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
808 	case 0x1E:	/* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
809 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
810 	case 0x25:	/* Westmere Client - Clarkdale, Arrandale */
811 	case 0x2C:	/* Westmere EP - Gulftown */
812 	case 0x2A:	/* SNB */
813 	case 0x2D:	/* SNB Xeon */
814 		return 1;
815 	case 0x2E:	/* Nehalem-EX Xeon - Beckton */
816 	case 0x2F:	/* Westmere-EX Xeon - Eagleton */
817 	default:
818 		return 0;
819 	}
820 }
821 
822 int is_snb(unsigned int family, unsigned int model)
823 {
824 	if (!genuine_intel)
825 		return 0;
826 
827 	switch (model) {
828 	case 0x2A:
829 	case 0x2D:
830 		return 1;
831 	}
832 	return 0;
833 }
834 
835 double discover_bclk(unsigned int family, unsigned int model)
836 {
837 	if (is_snb(family, model))
838 		return 100.00;
839 	else
840 		return 133.33;
841 }
842 
843 void check_cpuid()
844 {
845 	unsigned int eax, ebx, ecx, edx, max_level;
846 	unsigned int fms, family, model, stepping;
847 
848 	eax = ebx = ecx = edx = 0;
849 
850 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0));
851 
852 	if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
853 		genuine_intel = 1;
854 
855 	if (verbose)
856 		fprintf(stderr, "%.4s%.4s%.4s ",
857 			(char *)&ebx, (char *)&edx, (char *)&ecx);
858 
859 	asm("cpuid" : "=a" (fms), "=c" (ecx), "=d" (edx) : "a" (1) : "ebx");
860 	family = (fms >> 8) & 0xf;
861 	model = (fms >> 4) & 0xf;
862 	stepping = fms & 0xf;
863 	if (family == 6 || family == 0xf)
864 		model += ((fms >> 16) & 0xf) << 4;
865 
866 	if (verbose)
867 		fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
868 			max_level, family, model, stepping, family, model, stepping);
869 
870 	if (!(edx & (1 << 5))) {
871 		fprintf(stderr, "CPUID: no MSR\n");
872 		exit(1);
873 	}
874 
875 	/*
876 	 * check max extended function levels of CPUID.
877 	 * This is needed to check for invariant TSC.
878 	 * This check is valid for both Intel and AMD.
879 	 */
880 	ebx = ecx = edx = 0;
881 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000000));
882 
883 	if (max_level < 0x80000007) {
884 		fprintf(stderr, "CPUID: no invariant TSC (max_level 0x%x)\n", max_level);
885 		exit(1);
886 	}
887 
888 	/*
889 	 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
890 	 * this check is valid for both Intel and AMD
891 	 */
892 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000007));
893 	has_invariant_tsc = edx & (1 << 8);
894 
895 	if (!has_invariant_tsc) {
896 		fprintf(stderr, "No invariant TSC\n");
897 		exit(1);
898 	}
899 
900 	/*
901 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
902 	 * this check is valid for both Intel and AMD
903 	 */
904 
905 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x6));
906 	has_aperf = ecx & (1 << 0);
907 	if (!has_aperf) {
908 		fprintf(stderr, "No APERF MSR\n");
909 		exit(1);
910 	}
911 
912 	do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
913 	do_nhm_cstates = genuine_intel;	/* all Intel w/ non-stop TSC have NHM counters */
914 	do_snb_cstates = is_snb(family, model);
915 	bclk = discover_bclk(family, model);
916 
917 	do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model);
918 }
919 
920 
921 void usage()
922 {
923 	fprintf(stderr, "%s: [-v] [-M MSR#] [-i interval_sec | command ...]\n",
924 		progname);
925 	exit(1);
926 }
927 
928 
929 /*
930  * in /dev/cpu/ return success for names that are numbers
931  * ie. filter out ".", "..", "microcode".
932  */
933 int dir_filter(const struct dirent *dirp)
934 {
935 	if (isdigit(dirp->d_name[0]))
936 		return 1;
937 	else
938 		return 0;
939 }
940 
941 int open_dev_cpu_msr(int dummy1)
942 {
943 	return 0;
944 }
945 
946 void turbostat_init()
947 {
948 	check_cpuid();
949 
950 	check_dev_msr();
951 	check_super_user();
952 
953 	num_cpus = for_all_cpus(alloc_new_counters);
954 
955 	if (verbose)
956 		print_nehalem_info();
957 }
958 
959 int fork_it(char **argv)
960 {
961 	int retval;
962 	pid_t child_pid;
963 	get_counters(cnt_even);
964 	gettimeofday(&tv_even, (struct timezone *)NULL);
965 
966 	child_pid = fork();
967 	if (!child_pid) {
968 		/* child */
969 		execvp(argv[0], argv);
970 	} else {
971 		int status;
972 
973 		/* parent */
974 		if (child_pid == -1) {
975 			perror("fork");
976 			exit(1);
977 		}
978 
979 		signal(SIGINT, SIG_IGN);
980 		signal(SIGQUIT, SIG_IGN);
981 		if (waitpid(child_pid, &status, 0) == -1) {
982 			perror("wait");
983 			exit(1);
984 		}
985 	}
986 	get_counters(cnt_odd);
987 	gettimeofday(&tv_odd, (struct timezone *)NULL);
988 	retval = compute_delta(cnt_odd, cnt_even, cnt_delta);
989 
990 	timersub(&tv_odd, &tv_even, &tv_delta);
991 	compute_average(cnt_delta, cnt_average);
992 	if (!retval)
993 		print_counters(cnt_delta);
994 
995 	fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
996 
997 	return 0;
998 }
999 
1000 void cmdline(int argc, char **argv)
1001 {
1002 	int opt;
1003 
1004 	progname = argv[0];
1005 
1006 	while ((opt = getopt(argc, argv, "+vi:M:")) != -1) {
1007 		switch (opt) {
1008 		case 'v':
1009 			verbose++;
1010 			break;
1011 		case 'i':
1012 			interval_sec = atoi(optarg);
1013 			break;
1014 		case 'M':
1015 			sscanf(optarg, "%x", &extra_msr_offset);
1016 			if (verbose > 1)
1017 				fprintf(stderr, "MSR 0x%X\n", extra_msr_offset);
1018 			break;
1019 		default:
1020 			usage();
1021 		}
1022 	}
1023 }
1024 
1025 int main(int argc, char **argv)
1026 {
1027 	cmdline(argc, argv);
1028 
1029 	if (verbose > 1)
1030 		fprintf(stderr, "turbostat Dec 6, 2010"
1031 			" - Len Brown <lenb@kernel.org>\n");
1032 	if (verbose > 1)
1033 		fprintf(stderr, "http://userweb.kernel.org/~lenb/acpi/utils/pmtools/turbostat/\n");
1034 
1035 	turbostat_init();
1036 
1037 	/*
1038 	 * if any params left, it must be a command to fork
1039 	 */
1040 	if (argc - optind)
1041 		return fork_it(argv + optind);
1042 	else
1043 		turbostat_loop();
1044 
1045 	return 0;
1046 }
1047