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 
71 int aperf_mperf_unstable;
72 int backwards_count;
73 char *progname;
74 
75 int num_cpus;
76 cpu_set_t *cpu_mask;
77 size_t cpu_mask_size;
78 
79 struct counters {
80 	unsigned long long tsc;		/* per thread */
81 	unsigned long long aperf;	/* per thread */
82 	unsigned long long mperf;	/* per thread */
83 	unsigned long long c1;	/* per thread (calculated) */
84 	unsigned long long c3;	/* per core */
85 	unsigned long long c6;	/* per core */
86 	unsigned long long c7;	/* per core */
87 	unsigned long long pc2;	/* per package */
88 	unsigned long long pc3;	/* per package */
89 	unsigned long long pc6;	/* per package */
90 	unsigned long long pc7;	/* per package */
91 	unsigned long long extra_msr;	/* per thread */
92 	int pkg;
93 	int core;
94 	int cpu;
95 	struct counters *next;
96 };
97 
98 struct counters *cnt_even;
99 struct counters *cnt_odd;
100 struct counters *cnt_delta;
101 struct counters *cnt_average;
102 struct timeval tv_even;
103 struct timeval tv_odd;
104 struct timeval tv_delta;
105 
106 /*
107  * cpu_mask_init(ncpus)
108  *
109  * allocate and clear cpu_mask
110  * set cpu_mask_size
111  */
112 void cpu_mask_init(int ncpus)
113 {
114 	cpu_mask = CPU_ALLOC(ncpus);
115 	if (cpu_mask == NULL) {
116 		perror("CPU_ALLOC");
117 		exit(3);
118 	}
119 	cpu_mask_size = CPU_ALLOC_SIZE(ncpus);
120 	CPU_ZERO_S(cpu_mask_size, cpu_mask);
121 }
122 
123 void cpu_mask_uninit()
124 {
125 	CPU_FREE(cpu_mask);
126 	cpu_mask = NULL;
127 	cpu_mask_size = 0;
128 }
129 
130 int cpu_migrate(int cpu)
131 {
132 	CPU_ZERO_S(cpu_mask_size, cpu_mask);
133 	CPU_SET_S(cpu, cpu_mask_size, cpu_mask);
134 	if (sched_setaffinity(0, cpu_mask_size, cpu_mask) == -1)
135 		return -1;
136 	else
137 		return 0;
138 }
139 
140 int get_msr(int cpu, off_t offset, unsigned long long *msr)
141 {
142 	ssize_t retval;
143 	char pathname[32];
144 	int fd;
145 
146 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
147 	fd = open(pathname, O_RDONLY);
148 	if (fd < 0)
149 		return -1;
150 
151 	retval = pread(fd, msr, sizeof *msr, offset);
152 	close(fd);
153 
154 	if (retval != sizeof *msr)
155 		return -1;
156 
157 	return 0;
158 }
159 
160 void print_header(void)
161 {
162 	if (show_pkg)
163 		fprintf(stderr, "pk");
164 	if (show_pkg)
165 		fprintf(stderr, " ");
166 	if (show_core)
167 		fprintf(stderr, "cor");
168 	if (show_cpu)
169 		fprintf(stderr, " CPU");
170 	if (show_pkg || show_core || show_cpu)
171 		fprintf(stderr, " ");
172 	if (do_nhm_cstates)
173 		fprintf(stderr, "   %%c0");
174 	if (has_aperf)
175 		fprintf(stderr, "  GHz");
176 	fprintf(stderr, "  TSC");
177 	if (do_nhm_cstates)
178 		fprintf(stderr, "    %%c1");
179 	if (do_nhm_cstates)
180 		fprintf(stderr, "    %%c3");
181 	if (do_nhm_cstates)
182 		fprintf(stderr, "    %%c6");
183 	if (do_snb_cstates)
184 		fprintf(stderr, "    %%c7");
185 	if (do_snb_cstates)
186 		fprintf(stderr, "   %%pc2");
187 	if (do_nhm_cstates)
188 		fprintf(stderr, "   %%pc3");
189 	if (do_nhm_cstates)
190 		fprintf(stderr, "   %%pc6");
191 	if (do_snb_cstates)
192 		fprintf(stderr, "   %%pc7");
193 	if (extra_msr_offset)
194 		fprintf(stderr, "        MSR 0x%x ", extra_msr_offset);
195 
196 	putc('\n', stderr);
197 }
198 
199 void dump_cnt(struct counters *cnt)
200 {
201 	if (!cnt)
202 		return;
203 	if (cnt->pkg) fprintf(stderr, "package: %d ", cnt->pkg);
204 	if (cnt->core) fprintf(stderr, "core:: %d ", cnt->core);
205 	if (cnt->cpu) fprintf(stderr, "CPU: %d ", cnt->cpu);
206 	if (cnt->tsc) fprintf(stderr, "TSC: %016llX\n", cnt->tsc);
207 	if (cnt->c3) fprintf(stderr, "c3: %016llX\n", cnt->c3);
208 	if (cnt->c6) fprintf(stderr, "c6: %016llX\n", cnt->c6);
209 	if (cnt->c7) fprintf(stderr, "c7: %016llX\n", cnt->c7);
210 	if (cnt->aperf) fprintf(stderr, "aperf: %016llX\n", cnt->aperf);
211 	if (cnt->pc2) fprintf(stderr, "pc2: %016llX\n", cnt->pc2);
212 	if (cnt->pc3) fprintf(stderr, "pc3: %016llX\n", cnt->pc3);
213 	if (cnt->pc6) fprintf(stderr, "pc6: %016llX\n", cnt->pc6);
214 	if (cnt->pc7) fprintf(stderr, "pc7: %016llX\n", cnt->pc7);
215 	if (cnt->extra_msr) fprintf(stderr, "msr0x%x: %016llX\n", extra_msr_offset, cnt->extra_msr);
216 }
217 
218 void dump_list(struct counters *cnt)
219 {
220 	printf("dump_list 0x%p\n", cnt);
221 
222 	for (; cnt; cnt = cnt->next)
223 		dump_cnt(cnt);
224 }
225 
226 /*
227  * column formatting convention & formats
228  * package: "pk" 2 columns %2d
229  * core: "cor" 3 columns %3d
230  * CPU: "CPU" 3 columns %3d
231  * GHz: "GHz" 3 columns %3.2
232  * TSC: "TSC" 3 columns %3.2
233  * percentage " %pc3" %6.2
234  */
235 void print_cnt(struct counters *p)
236 {
237 	double interval_float;
238 
239 	interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
240 
241 	/* topology columns, print blanks on 1st (average) line */
242 	if (p == cnt_average) {
243 		if (show_pkg)
244 			fprintf(stderr, "  ");
245 		if (show_pkg && show_core)
246 			fprintf(stderr, " ");
247 		if (show_core)
248 			fprintf(stderr, "   ");
249 		if (show_cpu)
250 			fprintf(stderr, " " "   ");
251 	} else {
252 		if (show_pkg)
253 			fprintf(stderr, "%2d", p->pkg);
254 		if (show_pkg && show_core)
255 			fprintf(stderr, " ");
256 		if (show_core)
257 			fprintf(stderr, "%3d", p->core);
258 		if (show_cpu)
259 			fprintf(stderr, " %3d", p->cpu);
260 	}
261 
262 	/* %c0 */
263 	if (do_nhm_cstates) {
264 		if (show_pkg || show_core || show_cpu)
265 			fprintf(stderr, " ");
266 		if (!skip_c0)
267 			fprintf(stderr, "%6.2f", 100.0 * p->mperf/p->tsc);
268 		else
269 			fprintf(stderr, "  ****");
270 	}
271 
272 	/* GHz */
273 	if (has_aperf) {
274 		if (!aperf_mperf_unstable) {
275 			fprintf(stderr, " %3.2f",
276 				1.0 * p->tsc / units * p->aperf /
277 				p->mperf / interval_float);
278 		} else {
279 			if (p->aperf > p->tsc || p->mperf > p->tsc) {
280 				fprintf(stderr, " ***");
281 			} else {
282 				fprintf(stderr, "%3.1f*",
283 					1.0 * p->tsc /
284 					units * p->aperf /
285 					p->mperf / interval_float);
286 			}
287 		}
288 	}
289 
290 	/* TSC */
291 	fprintf(stderr, "%5.2f", 1.0 * p->tsc/units/interval_float);
292 
293 	if (do_nhm_cstates) {
294 		if (!skip_c1)
295 			fprintf(stderr, " %6.2f", 100.0 * p->c1/p->tsc);
296 		else
297 			fprintf(stderr, "  ****");
298 	}
299 	if (do_nhm_cstates)
300 		fprintf(stderr, " %6.2f", 100.0 * p->c3/p->tsc);
301 	if (do_nhm_cstates)
302 		fprintf(stderr, " %6.2f", 100.0 * p->c6/p->tsc);
303 	if (do_snb_cstates)
304 		fprintf(stderr, " %6.2f", 100.0 * p->c7/p->tsc);
305 	if (do_snb_cstates)
306 		fprintf(stderr, " %6.2f", 100.0 * p->pc2/p->tsc);
307 	if (do_nhm_cstates)
308 		fprintf(stderr, " %6.2f", 100.0 * p->pc3/p->tsc);
309 	if (do_nhm_cstates)
310 		fprintf(stderr, " %6.2f", 100.0 * p->pc6/p->tsc);
311 	if (do_snb_cstates)
312 		fprintf(stderr, " %6.2f", 100.0 * p->pc7/p->tsc);
313 	if (extra_msr_offset)
314 		fprintf(stderr, "  0x%016llx", p->extra_msr);
315 	putc('\n', stderr);
316 }
317 
318 void print_counters(struct counters *counters)
319 {
320 	struct counters *cnt;
321 	static int printed;
322 
323 
324 	if (!printed || !summary_only)
325 		print_header();
326 
327 	if (num_cpus > 1)
328 		print_cnt(cnt_average);
329 
330 	printed = 1;
331 
332 	if (summary_only)
333 		return;
334 
335 	for (cnt = counters; cnt != NULL; cnt = cnt->next)
336 		print_cnt(cnt);
337 
338 }
339 
340 #define SUBTRACT_COUNTER(after, before, delta) (delta = (after - before), (before > after))
341 
342 int compute_delta(struct counters *after,
343 	struct counters *before, struct counters *delta)
344 {
345 	int errors = 0;
346 	int perf_err = 0;
347 
348 	skip_c0 = skip_c1 = 0;
349 
350 	for ( ; after && before && delta;
351 		after = after->next, before = before->next, delta = delta->next) {
352 		if (before->cpu != after->cpu) {
353 			printf("cpu configuration changed: %d != %d\n",
354 				before->cpu, after->cpu);
355 			return -1;
356 		}
357 
358 		if (SUBTRACT_COUNTER(after->tsc, before->tsc, delta->tsc)) {
359 			fprintf(stderr, "cpu%d TSC went backwards %llX to %llX\n",
360 				before->cpu, before->tsc, after->tsc);
361 			errors++;
362 		}
363 		/* check for TSC < 1 Mcycles over interval */
364 		if (delta->tsc < (1000 * 1000)) {
365 			fprintf(stderr, "Insanely slow TSC rate,"
366 				" TSC stops in idle?\n");
367 			fprintf(stderr, "You can disable all c-states"
368 				" by booting with \"idle=poll\"\n");
369 			fprintf(stderr, "or just the deep ones with"
370 				" \"processor.max_cstate=1\"\n");
371 			exit(-3);
372 		}
373 		if (SUBTRACT_COUNTER(after->c3, before->c3, delta->c3)) {
374 			fprintf(stderr, "cpu%d c3 counter went backwards %llX to %llX\n",
375 				before->cpu, before->c3, after->c3);
376 			errors++;
377 		}
378 		if (SUBTRACT_COUNTER(after->c6, before->c6, delta->c6)) {
379 			fprintf(stderr, "cpu%d c6 counter went backwards %llX to %llX\n",
380 				before->cpu, before->c6, after->c6);
381 			errors++;
382 		}
383 		if (SUBTRACT_COUNTER(after->c7, before->c7, delta->c7)) {
384 			fprintf(stderr, "cpu%d c7 counter went backwards %llX to %llX\n",
385 				before->cpu, before->c7, after->c7);
386 			errors++;
387 		}
388 		if (SUBTRACT_COUNTER(after->pc2, before->pc2, delta->pc2)) {
389 			fprintf(stderr, "cpu%d pc2 counter went backwards %llX to %llX\n",
390 				before->cpu, before->pc2, after->pc2);
391 			errors++;
392 		}
393 		if (SUBTRACT_COUNTER(after->pc3, before->pc3, delta->pc3)) {
394 			fprintf(stderr, "cpu%d pc3 counter went backwards %llX to %llX\n",
395 				before->cpu, before->pc3, after->pc3);
396 			errors++;
397 		}
398 		if (SUBTRACT_COUNTER(after->pc6, before->pc6, delta->pc6)) {
399 			fprintf(stderr, "cpu%d pc6 counter went backwards %llX to %llX\n",
400 				before->cpu, before->pc6, after->pc6);
401 			errors++;
402 		}
403 		if (SUBTRACT_COUNTER(after->pc7, before->pc7, delta->pc7)) {
404 			fprintf(stderr, "cpu%d pc7 counter went backwards %llX to %llX\n",
405 				before->cpu, before->pc7, after->pc7);
406 			errors++;
407 		}
408 
409 		perf_err = SUBTRACT_COUNTER(after->aperf, before->aperf, delta->aperf);
410 		if (perf_err) {
411 			fprintf(stderr, "cpu%d aperf counter went backwards %llX to %llX\n",
412 				before->cpu, before->aperf, after->aperf);
413 		}
414 		perf_err |= SUBTRACT_COUNTER(after->mperf, before->mperf, delta->mperf);
415 		if (perf_err) {
416 			fprintf(stderr, "cpu%d mperf counter went backwards %llX to %llX\n",
417 				before->cpu, before->mperf, after->mperf);
418 		}
419 		if (perf_err) {
420 			if (!aperf_mperf_unstable) {
421 				fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
422 				fprintf(stderr, "* Frequency results do not cover entire interval *\n");
423 				fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
424 
425 				aperf_mperf_unstable = 1;
426 			}
427 			/*
428 			 * mperf delta is likely a huge "positive" number
429 			 * can not use it for calculating c0 time
430 			 */
431 			skip_c0 = 1;
432 			skip_c1 = 1;
433 		}
434 
435 		/*
436 		 * As mperf and tsc collection are not atomic,
437 		 * it is possible for mperf's non-halted cycles
438 		 * to exceed TSC's all cycles: show c1 = 0% in that case.
439 		 */
440 		if (delta->mperf > delta->tsc)
441 			delta->c1 = 0;
442 		else /* normal case, derive c1 */
443 			delta->c1 = delta->tsc - delta->mperf
444 				- delta->c3 - delta->c6 - delta->c7;
445 
446 		if (delta->mperf == 0)
447 			delta->mperf = 1;	/* divide by 0 protection */
448 
449 		/*
450 		 * for "extra msr", just copy the latest w/o subtracting
451 		 */
452 		delta->extra_msr = after->extra_msr;
453 		if (errors) {
454 			fprintf(stderr, "ERROR cpu%d before:\n", before->cpu);
455 			dump_cnt(before);
456 			fprintf(stderr, "ERROR cpu%d after:\n", before->cpu);
457 			dump_cnt(after);
458 			errors = 0;
459 		}
460 	}
461 	return 0;
462 }
463 
464 void compute_average(struct counters *delta, struct counters *avg)
465 {
466 	struct counters *sum;
467 
468 	sum = calloc(1, sizeof(struct counters));
469 	if (sum == NULL) {
470 		perror("calloc sum");
471 		exit(1);
472 	}
473 
474 	for (; delta; delta = delta->next) {
475 		sum->tsc += delta->tsc;
476 		sum->c1 += delta->c1;
477 		sum->c3 += delta->c3;
478 		sum->c6 += delta->c6;
479 		sum->c7 += delta->c7;
480 		sum->aperf += delta->aperf;
481 		sum->mperf += delta->mperf;
482 		sum->pc2 += delta->pc2;
483 		sum->pc3 += delta->pc3;
484 		sum->pc6 += delta->pc6;
485 		sum->pc7 += delta->pc7;
486 	}
487 	avg->tsc = sum->tsc/num_cpus;
488 	avg->c1 = sum->c1/num_cpus;
489 	avg->c3 = sum->c3/num_cpus;
490 	avg->c6 = sum->c6/num_cpus;
491 	avg->c7 = sum->c7/num_cpus;
492 	avg->aperf = sum->aperf/num_cpus;
493 	avg->mperf = sum->mperf/num_cpus;
494 	avg->pc2 = sum->pc2/num_cpus;
495 	avg->pc3 = sum->pc3/num_cpus;
496 	avg->pc6 = sum->pc6/num_cpus;
497 	avg->pc7 = sum->pc7/num_cpus;
498 
499 	free(sum);
500 }
501 
502 int get_counters(struct counters *cnt)
503 {
504 	for ( ; cnt; cnt = cnt->next) {
505 
506 		if (cpu_migrate(cnt->cpu))
507 			return -1;
508 
509 		if (get_msr(cnt->cpu, MSR_TSC, &cnt->tsc))
510 			return -1;
511 
512 		if (has_aperf) {
513 			if (get_msr(cnt->cpu, MSR_APERF, &cnt->aperf))
514 				return -1;
515 			if (get_msr(cnt->cpu, MSR_MPERF, &cnt->mperf))
516 				return -1;
517 		}
518 
519 		if (do_nhm_cstates) {
520 			if (get_msr(cnt->cpu, MSR_CORE_C3_RESIDENCY, &cnt->c3))
521 				return -1;
522 			if (get_msr(cnt->cpu, MSR_CORE_C6_RESIDENCY, &cnt->c6))
523 				return -1;
524 		}
525 
526 		if (do_snb_cstates)
527 			if (get_msr(cnt->cpu, MSR_CORE_C7_RESIDENCY, &cnt->c7))
528 				return -1;
529 
530 		if (do_nhm_cstates) {
531 			if (get_msr(cnt->cpu, MSR_PKG_C3_RESIDENCY, &cnt->pc3))
532 				return -1;
533 			if (get_msr(cnt->cpu, MSR_PKG_C6_RESIDENCY, &cnt->pc6))
534 				return -1;
535 		}
536 		if (do_snb_cstates) {
537 			if (get_msr(cnt->cpu, MSR_PKG_C2_RESIDENCY, &cnt->pc2))
538 				return -1;
539 			if (get_msr(cnt->cpu, MSR_PKG_C7_RESIDENCY, &cnt->pc7))
540 				return -1;
541 		}
542 		if (extra_msr_offset)
543 			if (get_msr(cnt->cpu, extra_msr_offset, &cnt->extra_msr))
544 				return -1;
545 	}
546 	return 0;
547 }
548 
549 void print_nehalem_info(void)
550 {
551 	unsigned long long msr;
552 	unsigned int ratio;
553 
554 	if (!do_nehalem_platform_info)
555 		return;
556 
557 	get_msr(0, MSR_NEHALEM_PLATFORM_INFO, &msr);
558 
559 	ratio = (msr >> 40) & 0xFF;
560 	fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
561 		ratio, bclk, ratio * bclk);
562 
563 	ratio = (msr >> 8) & 0xFF;
564 	fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
565 		ratio, bclk, ratio * bclk);
566 
567 	if (verbose > 1)
568 		fprintf(stderr, "MSR_NEHALEM_PLATFORM_INFO: 0x%llx\n", msr);
569 
570 	if (!do_nehalem_turbo_ratio_limit)
571 		return;
572 
573 	get_msr(0, MSR_NEHALEM_TURBO_RATIO_LIMIT, &msr);
574 
575 	ratio = (msr >> 24) & 0xFF;
576 	if (ratio)
577 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
578 			ratio, bclk, ratio * bclk);
579 
580 	ratio = (msr >> 16) & 0xFF;
581 	if (ratio)
582 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
583 			ratio, bclk, ratio * bclk);
584 
585 	ratio = (msr >> 8) & 0xFF;
586 	if (ratio)
587 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
588 			ratio, bclk, ratio * bclk);
589 
590 	ratio = (msr >> 0) & 0xFF;
591 	if (ratio)
592 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
593 			ratio, bclk, ratio * bclk);
594 
595 }
596 
597 void free_counter_list(struct counters *list)
598 {
599 	struct counters *p;
600 
601 	for (p = list; p; ) {
602 		struct counters *free_me;
603 
604 		free_me = p;
605 		p = p->next;
606 		free(free_me);
607 	}
608 }
609 
610 void free_all_counters(void)
611 {
612 	free_counter_list(cnt_even);
613 	cnt_even = NULL;
614 
615 	free_counter_list(cnt_odd);
616 	cnt_odd = NULL;
617 
618 	free_counter_list(cnt_delta);
619 	cnt_delta = NULL;
620 
621 	free_counter_list(cnt_average);
622 	cnt_average = NULL;
623 }
624 
625 void insert_counters(struct counters **list,
626 	struct counters *new)
627 {
628 	struct counters *prev;
629 
630 	/*
631 	 * list was empty
632 	 */
633 	if (*list == NULL) {
634 		new->next = *list;
635 		*list = new;
636 		return;
637 	}
638 
639 	if (!summary_only)
640 		show_cpu = 1;	/* there is more than one CPU */
641 
642 	/*
643 	 * insert on front of list.
644 	 * It is sorted by ascending package#, core#, cpu#
645 	 */
646 	if (((*list)->pkg > new->pkg) ||
647 	    (((*list)->pkg == new->pkg) && ((*list)->core > new->core)) ||
648 	    (((*list)->pkg == new->pkg) && ((*list)->core == new->core) && ((*list)->cpu > new->cpu))) {
649 		new->next = *list;
650 		*list = new;
651 		return;
652 	}
653 
654 	prev = *list;
655 
656 	while (prev->next && (prev->next->pkg < new->pkg)) {
657 		prev = prev->next;
658 		if (!summary_only)
659 			show_pkg = 1;	/* there is more than 1 package */
660 	}
661 
662 	while (prev->next && (prev->next->pkg == new->pkg)
663 		&& (prev->next->core < new->core)) {
664 		prev = prev->next;
665 		if (!summary_only)
666 			show_core = 1;	/* there is more than 1 core */
667 	}
668 
669 	while (prev->next && (prev->next->pkg == new->pkg)
670 		&& (prev->next->core == new->core)
671 		&& (prev->next->cpu < new->cpu)) {
672 		prev = prev->next;
673 	}
674 
675 	/*
676 	 * insert after "prev"
677 	 */
678 	new->next = prev->next;
679 	prev->next = new;
680 }
681 
682 void alloc_new_counters(int pkg, int core, int cpu)
683 {
684 	struct counters *new;
685 
686 	if (verbose > 1)
687 		printf("pkg%d core%d, cpu%d\n", pkg, core, cpu);
688 
689 	new = (struct counters *)calloc(1, sizeof(struct counters));
690 	if (new == NULL) {
691 		perror("calloc");
692 		exit(1);
693 	}
694 	new->pkg = pkg;
695 	new->core = core;
696 	new->cpu = cpu;
697 	insert_counters(&cnt_odd, new);
698 
699 	new = (struct counters *)calloc(1,
700 		sizeof(struct counters));
701 	if (new == NULL) {
702 		perror("calloc");
703 		exit(1);
704 	}
705 	new->pkg = pkg;
706 	new->core = core;
707 	new->cpu = cpu;
708 	insert_counters(&cnt_even, new);
709 
710 	new = (struct counters *)calloc(1, sizeof(struct counters));
711 	if (new == NULL) {
712 		perror("calloc");
713 		exit(1);
714 	}
715 	new->pkg = pkg;
716 	new->core = core;
717 	new->cpu = cpu;
718 	insert_counters(&cnt_delta, new);
719 
720 	new = (struct counters *)calloc(1, sizeof(struct counters));
721 	if (new == NULL) {
722 		perror("calloc");
723 		exit(1);
724 	}
725 	new->pkg = pkg;
726 	new->core = core;
727 	new->cpu = cpu;
728 	cnt_average = new;
729 }
730 
731 int get_physical_package_id(int cpu)
732 {
733 	char path[64];
734 	FILE *filep;
735 	int pkg;
736 
737 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
738 	filep = fopen(path, "r");
739 	if (filep == NULL) {
740 		perror(path);
741 		exit(1);
742 	}
743 	fscanf(filep, "%d", &pkg);
744 	fclose(filep);
745 	return pkg;
746 }
747 
748 int get_core_id(int cpu)
749 {
750 	char path[64];
751 	FILE *filep;
752 	int core;
753 
754 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
755 	filep = fopen(path, "r");
756 	if (filep == NULL) {
757 		perror(path);
758 		exit(1);
759 	}
760 	fscanf(filep, "%d", &core);
761 	fclose(filep);
762 	return core;
763 }
764 
765 /*
766  * run func(pkg, core, cpu) on every cpu in /proc/stat
767  */
768 
769 int for_all_cpus(void (func)(int, int, int))
770 {
771 	FILE *fp;
772 	int cpu_count;
773 	int retval;
774 
775 	fp = fopen(proc_stat, "r");
776 	if (fp == NULL) {
777 		perror(proc_stat);
778 		exit(1);
779 	}
780 
781 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
782 	if (retval != 0) {
783 		perror("/proc/stat format");
784 		exit(1);
785 	}
786 
787 	for (cpu_count = 0; ; cpu_count++) {
788 		int cpu;
789 
790 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu);
791 		if (retval != 1)
792 			break;
793 
794 		func(get_physical_package_id(cpu), get_core_id(cpu), cpu);
795 	}
796 	fclose(fp);
797 	return cpu_count;
798 }
799 
800 void re_initialize(void)
801 {
802 	free_all_counters();
803 	num_cpus = for_all_cpus(alloc_new_counters);
804 	cpu_mask_uninit();
805 	cpu_mask_init(num_cpus);
806 	printf("turbostat: re-initialized with num_cpus %d\n", num_cpus);
807 }
808 
809 void dummy(int pkg, int core, int cpu) { return; }
810 /*
811  * check to see if a cpu came on-line
812  */
813 int verify_num_cpus(void)
814 {
815 	int new_num_cpus;
816 
817 	new_num_cpus = for_all_cpus(dummy);
818 
819 	if (new_num_cpus != num_cpus) {
820 		if (verbose)
821 			printf("num_cpus was %d, is now  %d\n",
822 				num_cpus, new_num_cpus);
823 		return -1;
824 	}
825 	return 0;
826 }
827 
828 void turbostat_loop()
829 {
830 restart:
831 	get_counters(cnt_even);
832 	gettimeofday(&tv_even, (struct timezone *)NULL);
833 
834 	while (1) {
835 		if (verify_num_cpus()) {
836 			re_initialize();
837 			goto restart;
838 		}
839 		sleep(interval_sec);
840 		if (get_counters(cnt_odd)) {
841 			re_initialize();
842 			goto restart;
843 		}
844 		gettimeofday(&tv_odd, (struct timezone *)NULL);
845 		compute_delta(cnt_odd, cnt_even, cnt_delta);
846 		timersub(&tv_odd, &tv_even, &tv_delta);
847 		compute_average(cnt_delta, cnt_average);
848 		print_counters(cnt_delta);
849 		sleep(interval_sec);
850 		if (get_counters(cnt_even)) {
851 			re_initialize();
852 			goto restart;
853 		}
854 		gettimeofday(&tv_even, (struct timezone *)NULL);
855 		compute_delta(cnt_even, cnt_odd, cnt_delta);
856 		timersub(&tv_even, &tv_odd, &tv_delta);
857 		compute_average(cnt_delta, cnt_average);
858 		print_counters(cnt_delta);
859 	}
860 }
861 
862 void check_dev_msr()
863 {
864 	struct stat sb;
865 
866 	if (stat("/dev/cpu/0/msr", &sb)) {
867 		fprintf(stderr, "no /dev/cpu/0/msr\n");
868 		fprintf(stderr, "Try \"# modprobe msr\"\n");
869 		exit(-5);
870 	}
871 }
872 
873 void check_super_user()
874 {
875 	if (getuid() != 0) {
876 		fprintf(stderr, "must be root\n");
877 		exit(-6);
878 	}
879 }
880 
881 int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model)
882 {
883 	if (!genuine_intel)
884 		return 0;
885 
886 	if (family != 6)
887 		return 0;
888 
889 	switch (model) {
890 	case 0x1A:	/* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
891 	case 0x1E:	/* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
892 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
893 	case 0x25:	/* Westmere Client - Clarkdale, Arrandale */
894 	case 0x2C:	/* Westmere EP - Gulftown */
895 	case 0x2A:	/* SNB */
896 	case 0x2D:	/* SNB Xeon */
897 	case 0x3A:	/* IVB */
898 	case 0x3D:	/* IVB Xeon */
899 		return 1;
900 	case 0x2E:	/* Nehalem-EX Xeon - Beckton */
901 	case 0x2F:	/* Westmere-EX Xeon - Eagleton */
902 	default:
903 		return 0;
904 	}
905 }
906 
907 int is_snb(unsigned int family, unsigned int model)
908 {
909 	if (!genuine_intel)
910 		return 0;
911 
912 	switch (model) {
913 	case 0x2A:
914 	case 0x2D:
915 		return 1;
916 	}
917 	return 0;
918 }
919 
920 double discover_bclk(unsigned int family, unsigned int model)
921 {
922 	if (is_snb(family, model))
923 		return 100.00;
924 	else
925 		return 133.33;
926 }
927 
928 void check_cpuid()
929 {
930 	unsigned int eax, ebx, ecx, edx, max_level;
931 	unsigned int fms, family, model, stepping;
932 
933 	eax = ebx = ecx = edx = 0;
934 
935 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0));
936 
937 	if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
938 		genuine_intel = 1;
939 
940 	if (verbose)
941 		fprintf(stderr, "%.4s%.4s%.4s ",
942 			(char *)&ebx, (char *)&edx, (char *)&ecx);
943 
944 	asm("cpuid" : "=a" (fms), "=c" (ecx), "=d" (edx) : "a" (1) : "ebx");
945 	family = (fms >> 8) & 0xf;
946 	model = (fms >> 4) & 0xf;
947 	stepping = fms & 0xf;
948 	if (family == 6 || family == 0xf)
949 		model += ((fms >> 16) & 0xf) << 4;
950 
951 	if (verbose)
952 		fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
953 			max_level, family, model, stepping, family, model, stepping);
954 
955 	if (!(edx & (1 << 5))) {
956 		fprintf(stderr, "CPUID: no MSR\n");
957 		exit(1);
958 	}
959 
960 	/*
961 	 * check max extended function levels of CPUID.
962 	 * This is needed to check for invariant TSC.
963 	 * This check is valid for both Intel and AMD.
964 	 */
965 	ebx = ecx = edx = 0;
966 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000000));
967 
968 	if (max_level < 0x80000007) {
969 		fprintf(stderr, "CPUID: no invariant TSC (max_level 0x%x)\n", max_level);
970 		exit(1);
971 	}
972 
973 	/*
974 	 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
975 	 * this check is valid for both Intel and AMD
976 	 */
977 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000007));
978 	has_invariant_tsc = edx & (1 << 8);
979 
980 	if (!has_invariant_tsc) {
981 		fprintf(stderr, "No invariant TSC\n");
982 		exit(1);
983 	}
984 
985 	/*
986 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
987 	 * this check is valid for both Intel and AMD
988 	 */
989 
990 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x6));
991 	has_aperf = ecx & (1 << 0);
992 	if (!has_aperf) {
993 		fprintf(stderr, "No APERF MSR\n");
994 		exit(1);
995 	}
996 
997 	do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
998 	do_nhm_cstates = genuine_intel;	/* all Intel w/ non-stop TSC have NHM counters */
999 	do_snb_cstates = is_snb(family, model);
1000 	bclk = discover_bclk(family, model);
1001 
1002 	do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model);
1003 }
1004 
1005 
1006 void usage()
1007 {
1008 	fprintf(stderr, "%s: [-v] [-M MSR#] [-i interval_sec | command ...]\n",
1009 		progname);
1010 	exit(1);
1011 }
1012 
1013 
1014 /*
1015  * in /dev/cpu/ return success for names that are numbers
1016  * ie. filter out ".", "..", "microcode".
1017  */
1018 int dir_filter(const struct dirent *dirp)
1019 {
1020 	if (isdigit(dirp->d_name[0]))
1021 		return 1;
1022 	else
1023 		return 0;
1024 }
1025 
1026 int open_dev_cpu_msr(int dummy1)
1027 {
1028 	return 0;
1029 }
1030 
1031 void turbostat_init()
1032 {
1033 	check_cpuid();
1034 
1035 	check_dev_msr();
1036 	check_super_user();
1037 
1038 	num_cpus = for_all_cpus(alloc_new_counters);
1039 	cpu_mask_init(num_cpus);
1040 
1041 	if (verbose)
1042 		print_nehalem_info();
1043 }
1044 
1045 int fork_it(char **argv)
1046 {
1047 	int retval;
1048 	pid_t child_pid;
1049 	get_counters(cnt_even);
1050 	gettimeofday(&tv_even, (struct timezone *)NULL);
1051 
1052 	child_pid = fork();
1053 	if (!child_pid) {
1054 		/* child */
1055 		execvp(argv[0], argv);
1056 	} else {
1057 		int status;
1058 
1059 		/* parent */
1060 		if (child_pid == -1) {
1061 			perror("fork");
1062 			exit(1);
1063 		}
1064 
1065 		signal(SIGINT, SIG_IGN);
1066 		signal(SIGQUIT, SIG_IGN);
1067 		if (waitpid(child_pid, &status, 0) == -1) {
1068 			perror("wait");
1069 			exit(1);
1070 		}
1071 	}
1072 	get_counters(cnt_odd);
1073 	gettimeofday(&tv_odd, (struct timezone *)NULL);
1074 	retval = compute_delta(cnt_odd, cnt_even, cnt_delta);
1075 
1076 	timersub(&tv_odd, &tv_even, &tv_delta);
1077 	compute_average(cnt_delta, cnt_average);
1078 	if (!retval)
1079 		print_counters(cnt_delta);
1080 
1081 	fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
1082 
1083 	return 0;
1084 }
1085 
1086 void cmdline(int argc, char **argv)
1087 {
1088 	int opt;
1089 
1090 	progname = argv[0];
1091 
1092 	while ((opt = getopt(argc, argv, "+svi:M:")) != -1) {
1093 		switch (opt) {
1094 		case 's':
1095 			summary_only++;
1096 			break;
1097 		case 'v':
1098 			verbose++;
1099 			break;
1100 		case 'i':
1101 			interval_sec = atoi(optarg);
1102 			break;
1103 		case 'M':
1104 			sscanf(optarg, "%x", &extra_msr_offset);
1105 			if (verbose > 1)
1106 				fprintf(stderr, "MSR 0x%X\n", extra_msr_offset);
1107 			break;
1108 		default:
1109 			usage();
1110 		}
1111 	}
1112 }
1113 
1114 int main(int argc, char **argv)
1115 {
1116 	cmdline(argc, argv);
1117 
1118 	if (verbose > 1)
1119 		fprintf(stderr, "turbostat Dec 6, 2010"
1120 			" - Len Brown <lenb@kernel.org>\n");
1121 	if (verbose > 1)
1122 		fprintf(stderr, "http://userweb.kernel.org/~lenb/acpi/utils/pmtools/turbostat/\n");
1123 
1124 	turbostat_init();
1125 
1126 	/*
1127 	 * if any params left, it must be a command to fork
1128 	 */
1129 	if (argc - optind)
1130 		return fork_it(argv + optind);
1131 	else
1132 		turbostat_loop();
1133 
1134 	return 0;
1135 }
1136