1103a8feaSLen Brown /*
2103a8feaSLen Brown  * turbostat -- show CPU frequency and C-state residency
3103a8feaSLen Brown  * on modern Intel turbo-capable processors.
4103a8feaSLen Brown  *
5e23da037SLen Brown  * Copyright (c) 2012 Intel Corporation.
6103a8feaSLen Brown  * Len Brown <len.brown@intel.com>
7103a8feaSLen Brown  *
8103a8feaSLen Brown  * This program is free software; you can redistribute it and/or modify it
9103a8feaSLen Brown  * under the terms and conditions of the GNU General Public License,
10103a8feaSLen Brown  * version 2, as published by the Free Software Foundation.
11103a8feaSLen Brown  *
12103a8feaSLen Brown  * This program is distributed in the hope it will be useful, but WITHOUT
13103a8feaSLen Brown  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14103a8feaSLen Brown  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
15103a8feaSLen Brown  * more details.
16103a8feaSLen Brown  *
17103a8feaSLen Brown  * You should have received a copy of the GNU General Public License along with
18103a8feaSLen Brown  * this program; if not, write to the Free Software Foundation, Inc.,
19103a8feaSLen Brown  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
20103a8feaSLen Brown  */
21103a8feaSLen Brown 
2288c3281fSLen Brown #define _GNU_SOURCE
23103a8feaSLen Brown #include <stdio.h>
24103a8feaSLen Brown #include <unistd.h>
25103a8feaSLen Brown #include <sys/types.h>
26103a8feaSLen Brown #include <sys/wait.h>
27103a8feaSLen Brown #include <sys/stat.h>
28103a8feaSLen Brown #include <sys/resource.h>
29103a8feaSLen Brown #include <fcntl.h>
30103a8feaSLen Brown #include <signal.h>
31103a8feaSLen Brown #include <sys/time.h>
32103a8feaSLen Brown #include <stdlib.h>
33103a8feaSLen Brown #include <dirent.h>
34103a8feaSLen Brown #include <string.h>
35103a8feaSLen Brown #include <ctype.h>
3688c3281fSLen Brown #include <sched.h>
37103a8feaSLen Brown 
38103a8feaSLen Brown #define MSR_NEHALEM_PLATFORM_INFO	0xCE
39103a8feaSLen Brown #define MSR_NEHALEM_TURBO_RATIO_LIMIT	0x1AD
406574a5d5SLen Brown #define MSR_IVT_TURBO_RATIO_LIMIT	0x1AE
41103a8feaSLen Brown #define MSR_APERF	0xE8
42103a8feaSLen Brown #define MSR_MPERF	0xE7
43103a8feaSLen Brown #define MSR_PKG_C2_RESIDENCY	0x60D	/* SNB only */
44103a8feaSLen Brown #define MSR_PKG_C3_RESIDENCY	0x3F8
45103a8feaSLen Brown #define MSR_PKG_C6_RESIDENCY	0x3F9
46103a8feaSLen Brown #define MSR_PKG_C7_RESIDENCY	0x3FA	/* SNB only */
47103a8feaSLen Brown #define MSR_CORE_C3_RESIDENCY	0x3FC
48103a8feaSLen Brown #define MSR_CORE_C6_RESIDENCY	0x3FD
49103a8feaSLen Brown #define MSR_CORE_C7_RESIDENCY	0x3FE	/* SNB only */
50103a8feaSLen Brown 
51103a8feaSLen Brown char *proc_stat = "/proc/stat";
52103a8feaSLen Brown unsigned int interval_sec = 5;	/* set with -i interval_sec */
53103a8feaSLen Brown unsigned int verbose;		/* set with -v */
54e23da037SLen Brown unsigned int summary_only;	/* set with -s */
55103a8feaSLen Brown unsigned int skip_c0;
56103a8feaSLen Brown unsigned int skip_c1;
57103a8feaSLen Brown unsigned int do_nhm_cstates;
58103a8feaSLen Brown unsigned int do_snb_cstates;
59103a8feaSLen Brown unsigned int has_aperf;
60103a8feaSLen Brown unsigned int units = 1000000000;	/* Ghz etc */
61103a8feaSLen Brown unsigned int genuine_intel;
62103a8feaSLen Brown unsigned int has_invariant_tsc;
63103a8feaSLen Brown unsigned int do_nehalem_platform_info;
64103a8feaSLen Brown unsigned int do_nehalem_turbo_ratio_limit;
656574a5d5SLen Brown unsigned int do_ivt_turbo_ratio_limit;
662f32edf1SLen Brown unsigned int extra_msr_offset32;
672f32edf1SLen Brown unsigned int extra_msr_offset64;
688e180f3cSLen Brown unsigned int extra_delta_offset32;
698e180f3cSLen Brown unsigned int extra_delta_offset64;
70103a8feaSLen Brown double bclk;
71103a8feaSLen Brown unsigned int show_pkg;
72103a8feaSLen Brown unsigned int show_core;
73103a8feaSLen Brown unsigned int show_cpu;
74c98d5d94SLen Brown unsigned int show_pkg_only;
75c98d5d94SLen Brown unsigned int show_core_only;
76c98d5d94SLen Brown char *output_buffer, *outp;
77103a8feaSLen Brown 
78103a8feaSLen Brown int aperf_mperf_unstable;
79103a8feaSLen Brown int backwards_count;
80103a8feaSLen Brown char *progname;
81103a8feaSLen Brown 
82c98d5d94SLen Brown cpu_set_t *cpu_present_set, *cpu_affinity_set;
83c98d5d94SLen Brown size_t cpu_present_setsize, cpu_affinity_setsize;
84c98d5d94SLen Brown 
85c98d5d94SLen Brown struct thread_data {
86c98d5d94SLen Brown 	unsigned long long tsc;
87c98d5d94SLen Brown 	unsigned long long aperf;
88c98d5d94SLen Brown 	unsigned long long mperf;
89c98d5d94SLen Brown 	unsigned long long c1;	/* derived */
902f32edf1SLen Brown 	unsigned long long extra_msr64;
918e180f3cSLen Brown 	unsigned long long extra_delta64;
928e180f3cSLen Brown 	unsigned long long extra_msr32;
938e180f3cSLen Brown 	unsigned long long extra_delta32;
94c98d5d94SLen Brown 	unsigned int cpu_id;
95c98d5d94SLen Brown 	unsigned int flags;
96c98d5d94SLen Brown #define CPU_IS_FIRST_THREAD_IN_CORE	0x2
97c98d5d94SLen Brown #define CPU_IS_FIRST_CORE_IN_PACKAGE	0x4
98c98d5d94SLen Brown } *thread_even, *thread_odd;
99c98d5d94SLen Brown 
100c98d5d94SLen Brown struct core_data {
101c98d5d94SLen Brown 	unsigned long long c3;
102c98d5d94SLen Brown 	unsigned long long c6;
103c98d5d94SLen Brown 	unsigned long long c7;
104c98d5d94SLen Brown 	unsigned int core_id;
105c98d5d94SLen Brown } *core_even, *core_odd;
106c98d5d94SLen Brown 
107c98d5d94SLen Brown struct pkg_data {
108c98d5d94SLen Brown 	unsigned long long pc2;
109c98d5d94SLen Brown 	unsigned long long pc3;
110c98d5d94SLen Brown 	unsigned long long pc6;
111c98d5d94SLen Brown 	unsigned long long pc7;
112c98d5d94SLen Brown 	unsigned int package_id;
113c98d5d94SLen Brown } *package_even, *package_odd;
114c98d5d94SLen Brown 
115c98d5d94SLen Brown #define ODD_COUNTERS thread_odd, core_odd, package_odd
116c98d5d94SLen Brown #define EVEN_COUNTERS thread_even, core_even, package_even
117c98d5d94SLen Brown 
118c98d5d94SLen Brown #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
119c98d5d94SLen Brown 	(thread_base + (pkg_no) * topo.num_cores_per_pkg * \
120c98d5d94SLen Brown 		topo.num_threads_per_core + \
121c98d5d94SLen Brown 		(core_no) * topo.num_threads_per_core + (thread_no))
122c98d5d94SLen Brown #define GET_CORE(core_base, core_no, pkg_no) \
123c98d5d94SLen Brown 	(core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
124c98d5d94SLen Brown #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
125c98d5d94SLen Brown 
126c98d5d94SLen Brown struct system_summary {
127c98d5d94SLen Brown 	struct thread_data threads;
128c98d5d94SLen Brown 	struct core_data cores;
129c98d5d94SLen Brown 	struct pkg_data packages;
130c98d5d94SLen Brown } sum, average;
131c98d5d94SLen Brown 
132c98d5d94SLen Brown 
133c98d5d94SLen Brown struct topo_params {
134c98d5d94SLen Brown 	int num_packages;
135103a8feaSLen Brown 	int num_cpus;
136c98d5d94SLen Brown 	int num_cores;
137c98d5d94SLen Brown 	int max_cpu_num;
138c98d5d94SLen Brown 	int num_cores_per_pkg;
139c98d5d94SLen Brown 	int num_threads_per_core;
140c98d5d94SLen Brown } topo;
141103a8feaSLen Brown 
142c98d5d94SLen Brown struct timeval tv_even, tv_odd, tv_delta;
143103a8feaSLen Brown 
144c98d5d94SLen Brown void setup_all_buffers(void);
145103a8feaSLen Brown 
146c98d5d94SLen Brown int cpu_is_not_present(int cpu)
147d15cf7c1SLen Brown {
148c98d5d94SLen Brown 	return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
149c98d5d94SLen Brown }
150c98d5d94SLen Brown /*
151c98d5d94SLen Brown  * run func(thread, core, package) in topology order
152c98d5d94SLen Brown  * skip non-present cpus
153c98d5d94SLen Brown  */
154c98d5d94SLen Brown 
155c98d5d94SLen Brown int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
156c98d5d94SLen Brown 	struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
157c98d5d94SLen Brown {
158c98d5d94SLen Brown 	int retval, pkg_no, core_no, thread_no;
159c98d5d94SLen Brown 
160c98d5d94SLen Brown 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
161c98d5d94SLen Brown 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
162c98d5d94SLen Brown 			for (thread_no = 0; thread_no <
163c98d5d94SLen Brown 				topo.num_threads_per_core; ++thread_no) {
164c98d5d94SLen Brown 				struct thread_data *t;
165c98d5d94SLen Brown 				struct core_data *c;
166c98d5d94SLen Brown 				struct pkg_data *p;
167c98d5d94SLen Brown 
168c98d5d94SLen Brown 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
169c98d5d94SLen Brown 
170c98d5d94SLen Brown 				if (cpu_is_not_present(t->cpu_id))
171c98d5d94SLen Brown 					continue;
172c98d5d94SLen Brown 
173c98d5d94SLen Brown 				c = GET_CORE(core_base, core_no, pkg_no);
174c98d5d94SLen Brown 				p = GET_PKG(pkg_base, pkg_no);
175c98d5d94SLen Brown 
176c98d5d94SLen Brown 				retval = func(t, c, p);
177c98d5d94SLen Brown 				if (retval)
178c98d5d94SLen Brown 					return retval;
179c98d5d94SLen Brown 			}
180c98d5d94SLen Brown 		}
181c98d5d94SLen Brown 	}
182d15cf7c1SLen Brown 	return 0;
183d15cf7c1SLen Brown }
184d15cf7c1SLen Brown 
18588c3281fSLen Brown int cpu_migrate(int cpu)
18688c3281fSLen Brown {
187c98d5d94SLen Brown 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
188c98d5d94SLen Brown 	CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
189c98d5d94SLen Brown 	if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
19088c3281fSLen Brown 		return -1;
19188c3281fSLen Brown 	else
19288c3281fSLen Brown 		return 0;
19388c3281fSLen Brown }
19488c3281fSLen Brown 
19515aaa346SLen Brown int get_msr(int cpu, off_t offset, unsigned long long *msr)
196103a8feaSLen Brown {
197103a8feaSLen Brown 	ssize_t retval;
198103a8feaSLen Brown 	char pathname[32];
199103a8feaSLen Brown 	int fd;
200103a8feaSLen Brown 
201103a8feaSLen Brown 	sprintf(pathname, "/dev/cpu/%d/msr", cpu);
202103a8feaSLen Brown 	fd = open(pathname, O_RDONLY);
20315aaa346SLen Brown 	if (fd < 0)
20415aaa346SLen Brown 		return -1;
205103a8feaSLen Brown 
20615aaa346SLen Brown 	retval = pread(fd, msr, sizeof *msr, offset);
207103a8feaSLen Brown 	close(fd);
20815aaa346SLen Brown 
209d91bb17cSLen Brown 	if (retval != sizeof *msr) {
210d91bb17cSLen Brown 		fprintf(stderr, "%s offset 0x%zx read failed\n", pathname, offset);
21115aaa346SLen Brown 		return -1;
212d91bb17cSLen Brown 	}
21315aaa346SLen Brown 
21415aaa346SLen Brown 	return 0;
215103a8feaSLen Brown }
216103a8feaSLen Brown 
217a829eb4dSLen Brown void print_header(void)
218103a8feaSLen Brown {
219103a8feaSLen Brown 	if (show_pkg)
220c98d5d94SLen Brown 		outp += sprintf(outp, "pk");
221e23da037SLen Brown 	if (show_pkg)
222c98d5d94SLen Brown 		outp += sprintf(outp, " ");
223103a8feaSLen Brown 	if (show_core)
224c98d5d94SLen Brown 		outp += sprintf(outp, "cor");
225103a8feaSLen Brown 	if (show_cpu)
226c98d5d94SLen Brown 		outp += sprintf(outp, " CPU");
227e23da037SLen Brown 	if (show_pkg || show_core || show_cpu)
228c98d5d94SLen Brown 		outp += sprintf(outp, " ");
229103a8feaSLen Brown 	if (do_nhm_cstates)
230c98d5d94SLen Brown 		outp += sprintf(outp, "   %%c0");
231103a8feaSLen Brown 	if (has_aperf)
232c98d5d94SLen Brown 		outp += sprintf(outp, "  GHz");
233c98d5d94SLen Brown 	outp += sprintf(outp, "  TSC");
2348e180f3cSLen Brown 	if (extra_delta_offset32)
235f9240813SLen Brown 		outp += sprintf(outp, "  count 0x%03X", extra_delta_offset32);
2368e180f3cSLen Brown 	if (extra_delta_offset64)
237f9240813SLen Brown 		outp += sprintf(outp, "  COUNT 0x%03X", extra_delta_offset64);
2382f32edf1SLen Brown 	if (extra_msr_offset32)
2398e180f3cSLen Brown 		outp += sprintf(outp, "   MSR 0x%03X", extra_msr_offset32);
2402f32edf1SLen Brown 	if (extra_msr_offset64)
2418e180f3cSLen Brown 		outp += sprintf(outp, "           MSR 0x%03X", extra_msr_offset64);
242103a8feaSLen Brown 	if (do_nhm_cstates)
243c98d5d94SLen Brown 		outp += sprintf(outp, "    %%c1");
244103a8feaSLen Brown 	if (do_nhm_cstates)
245c98d5d94SLen Brown 		outp += sprintf(outp, "    %%c3");
246103a8feaSLen Brown 	if (do_nhm_cstates)
247c98d5d94SLen Brown 		outp += sprintf(outp, "    %%c6");
248103a8feaSLen Brown 	if (do_snb_cstates)
249c98d5d94SLen Brown 		outp += sprintf(outp, "    %%c7");
250103a8feaSLen Brown 	if (do_snb_cstates)
251c98d5d94SLen Brown 		outp += sprintf(outp, "   %%pc2");
252103a8feaSLen Brown 	if (do_nhm_cstates)
253c98d5d94SLen Brown 		outp += sprintf(outp, "   %%pc3");
254103a8feaSLen Brown 	if (do_nhm_cstates)
255c98d5d94SLen Brown 		outp += sprintf(outp, "   %%pc6");
256103a8feaSLen Brown 	if (do_snb_cstates)
257c98d5d94SLen Brown 		outp += sprintf(outp, "   %%pc7");
258103a8feaSLen Brown 
259c98d5d94SLen Brown 	outp += sprintf(outp, "\n");
260103a8feaSLen Brown }
261103a8feaSLen Brown 
262c98d5d94SLen Brown int dump_counters(struct thread_data *t, struct core_data *c,
263c98d5d94SLen Brown 	struct pkg_data *p)
264103a8feaSLen Brown {
265c98d5d94SLen Brown 	fprintf(stderr, "t %p, c %p, p %p\n", t, c, p);
266c98d5d94SLen Brown 
267c98d5d94SLen Brown 	if (t) {
268c98d5d94SLen Brown 		fprintf(stderr, "CPU: %d flags 0x%x\n", t->cpu_id, t->flags);
269c98d5d94SLen Brown 		fprintf(stderr, "TSC: %016llX\n", t->tsc);
270c98d5d94SLen Brown 		fprintf(stderr, "aperf: %016llX\n", t->aperf);
271c98d5d94SLen Brown 		fprintf(stderr, "mperf: %016llX\n", t->mperf);
272c98d5d94SLen Brown 		fprintf(stderr, "c1: %016llX\n", t->c1);
2738e180f3cSLen Brown 		fprintf(stderr, "msr0x%x: %08llX\n",
2748e180f3cSLen Brown 			extra_delta_offset32, t->extra_delta32);
2758e180f3cSLen Brown 		fprintf(stderr, "msr0x%x: %016llX\n",
2768e180f3cSLen Brown 			extra_delta_offset64, t->extra_delta64);
2778e180f3cSLen Brown 		fprintf(stderr, "msr0x%x: %08llX\n",
2782f32edf1SLen Brown 			extra_msr_offset32, t->extra_msr32);
279c98d5d94SLen Brown 		fprintf(stderr, "msr0x%x: %016llX\n",
2802f32edf1SLen Brown 			extra_msr_offset64, t->extra_msr64);
281103a8feaSLen Brown 	}
282103a8feaSLen Brown 
283c98d5d94SLen Brown 	if (c) {
284c98d5d94SLen Brown 		fprintf(stderr, "core: %d\n", c->core_id);
285c98d5d94SLen Brown 		fprintf(stderr, "c3: %016llX\n", c->c3);
286c98d5d94SLen Brown 		fprintf(stderr, "c6: %016llX\n", c->c6);
287c98d5d94SLen Brown 		fprintf(stderr, "c7: %016llX\n", c->c7);
288c98d5d94SLen Brown 	}
289103a8feaSLen Brown 
290c98d5d94SLen Brown 	if (p) {
291c98d5d94SLen Brown 		fprintf(stderr, "package: %d\n", p->package_id);
292c98d5d94SLen Brown 		fprintf(stderr, "pc2: %016llX\n", p->pc2);
293c98d5d94SLen Brown 		fprintf(stderr, "pc3: %016llX\n", p->pc3);
294c98d5d94SLen Brown 		fprintf(stderr, "pc6: %016llX\n", p->pc6);
295c98d5d94SLen Brown 		fprintf(stderr, "pc7: %016llX\n", p->pc7);
296c98d5d94SLen Brown 	}
297c98d5d94SLen Brown 	return 0;
298103a8feaSLen Brown }
299103a8feaSLen Brown 
300e23da037SLen Brown /*
301e23da037SLen Brown  * column formatting convention & formats
302e23da037SLen Brown  * package: "pk" 2 columns %2d
303e23da037SLen Brown  * core: "cor" 3 columns %3d
304e23da037SLen Brown  * CPU: "CPU" 3 columns %3d
305e23da037SLen Brown  * GHz: "GHz" 3 columns %3.2
306e23da037SLen Brown  * TSC: "TSC" 3 columns %3.2
307e23da037SLen Brown  * percentage " %pc3" %6.2
308e23da037SLen Brown  */
309c98d5d94SLen Brown int format_counters(struct thread_data *t, struct core_data *c,
310c98d5d94SLen Brown 	struct pkg_data *p)
311103a8feaSLen Brown {
312103a8feaSLen Brown 	double interval_float;
313103a8feaSLen Brown 
314c98d5d94SLen Brown 	 /* if showing only 1st thread in core and this isn't one, bail out */
315c98d5d94SLen Brown 	if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
316c98d5d94SLen Brown 		return 0;
317c98d5d94SLen Brown 
318c98d5d94SLen Brown 	 /* if showing only 1st thread in pkg and this isn't one, bail out */
319c98d5d94SLen Brown 	if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
320c98d5d94SLen Brown 		return 0;
321c98d5d94SLen Brown 
322103a8feaSLen Brown 	interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
323103a8feaSLen Brown 
324c98d5d94SLen Brown 	/* topo columns, print blanks on 1st (average) line */
325c98d5d94SLen Brown 	if (t == &average.threads) {
326103a8feaSLen Brown 		if (show_pkg)
327c98d5d94SLen Brown 			outp += sprintf(outp, "  ");
328e23da037SLen Brown 		if (show_pkg && show_core)
329c98d5d94SLen Brown 			outp += sprintf(outp, " ");
330103a8feaSLen Brown 		if (show_core)
331c98d5d94SLen Brown 			outp += sprintf(outp, "   ");
332103a8feaSLen Brown 		if (show_cpu)
333c98d5d94SLen Brown 			outp += sprintf(outp, " " "   ");
334103a8feaSLen Brown 	} else {
335c98d5d94SLen Brown 		if (show_pkg) {
336c98d5d94SLen Brown 			if (p)
337c98d5d94SLen Brown 				outp += sprintf(outp, "%2d", p->package_id);
338c98d5d94SLen Brown 			else
339c98d5d94SLen Brown 				outp += sprintf(outp, "  ");
340c98d5d94SLen Brown 		}
341e23da037SLen Brown 		if (show_pkg && show_core)
342c98d5d94SLen Brown 			outp += sprintf(outp, " ");
343c98d5d94SLen Brown 		if (show_core) {
344c98d5d94SLen Brown 			if (c)
345c98d5d94SLen Brown 				outp += sprintf(outp, "%3d", c->core_id);
346c98d5d94SLen Brown 			else
347c98d5d94SLen Brown 				outp += sprintf(outp, "   ");
348c98d5d94SLen Brown 		}
349103a8feaSLen Brown 		if (show_cpu)
350c98d5d94SLen Brown 			outp += sprintf(outp, " %3d", t->cpu_id);
351103a8feaSLen Brown 	}
352103a8feaSLen Brown 
353103a8feaSLen Brown 	/* %c0 */
354103a8feaSLen Brown 	if (do_nhm_cstates) {
355e23da037SLen Brown 		if (show_pkg || show_core || show_cpu)
356c98d5d94SLen Brown 			outp += sprintf(outp, " ");
357103a8feaSLen Brown 		if (!skip_c0)
358c98d5d94SLen Brown 			outp += sprintf(outp, "%6.2f", 100.0 * t->mperf/t->tsc);
359103a8feaSLen Brown 		else
360c98d5d94SLen Brown 			outp += sprintf(outp, "  ****");
361103a8feaSLen Brown 	}
362103a8feaSLen Brown 
363103a8feaSLen Brown 	/* GHz */
364103a8feaSLen Brown 	if (has_aperf) {
365103a8feaSLen Brown 		if (!aperf_mperf_unstable) {
366c98d5d94SLen Brown 			outp += sprintf(outp, " %3.2f",
367c98d5d94SLen Brown 				1.0 * t->tsc / units * t->aperf /
368c98d5d94SLen Brown 				t->mperf / interval_float);
369103a8feaSLen Brown 		} else {
370c98d5d94SLen Brown 			if (t->aperf > t->tsc || t->mperf > t->tsc) {
371c98d5d94SLen Brown 				outp += sprintf(outp, " ***");
372103a8feaSLen Brown 			} else {
373c98d5d94SLen Brown 				outp += sprintf(outp, "%3.1f*",
374c98d5d94SLen Brown 					1.0 * t->tsc /
375c98d5d94SLen Brown 					units * t->aperf /
376c98d5d94SLen Brown 					t->mperf / interval_float);
377103a8feaSLen Brown 			}
378103a8feaSLen Brown 		}
379103a8feaSLen Brown 	}
380103a8feaSLen Brown 
381103a8feaSLen Brown 	/* TSC */
382c98d5d94SLen Brown 	outp += sprintf(outp, "%5.2f", 1.0 * t->tsc/units/interval_float);
383103a8feaSLen Brown 
3848e180f3cSLen Brown 	/* delta */
3858e180f3cSLen Brown 	if (extra_delta_offset32)
3868e180f3cSLen Brown 		outp += sprintf(outp, "  %11llu", t->extra_delta32);
3878e180f3cSLen Brown 
3888e180f3cSLen Brown 	/* DELTA */
3898e180f3cSLen Brown 	if (extra_delta_offset64)
3908e180f3cSLen Brown 		outp += sprintf(outp, "  %11llu", t->extra_delta64);
3912f32edf1SLen Brown 	/* msr */
3922f32edf1SLen Brown 	if (extra_msr_offset32)
3938e180f3cSLen Brown 		outp += sprintf(outp, "  0x%08llx", t->extra_msr32);
3942f32edf1SLen Brown 
395130ff304SLen Brown 	/* MSR */
3962f32edf1SLen Brown 	if (extra_msr_offset64)
3972f32edf1SLen Brown 		outp += sprintf(outp, "  0x%016llx", t->extra_msr64);
398130ff304SLen Brown 
399103a8feaSLen Brown 	if (do_nhm_cstates) {
400103a8feaSLen Brown 		if (!skip_c1)
401c98d5d94SLen Brown 			outp += sprintf(outp, " %6.2f", 100.0 * t->c1/t->tsc);
402103a8feaSLen Brown 		else
403c98d5d94SLen Brown 			outp += sprintf(outp, "  ****");
404103a8feaSLen Brown 	}
405c98d5d94SLen Brown 
406c98d5d94SLen Brown 	/* print per-core data only for 1st thread in core */
407c98d5d94SLen Brown 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
408c98d5d94SLen Brown 		goto done;
409c98d5d94SLen Brown 
410103a8feaSLen Brown 	if (do_nhm_cstates)
411c98d5d94SLen Brown 		outp += sprintf(outp, " %6.2f", 100.0 * c->c3/t->tsc);
412103a8feaSLen Brown 	if (do_nhm_cstates)
413c98d5d94SLen Brown 		outp += sprintf(outp, " %6.2f", 100.0 * c->c6/t->tsc);
414103a8feaSLen Brown 	if (do_snb_cstates)
415c98d5d94SLen Brown 		outp += sprintf(outp, " %6.2f", 100.0 * c->c7/t->tsc);
416c98d5d94SLen Brown 
417c98d5d94SLen Brown 	/* print per-package data only for 1st core in package */
418c98d5d94SLen Brown 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
419c98d5d94SLen Brown 		goto done;
420c98d5d94SLen Brown 
421103a8feaSLen Brown 	if (do_snb_cstates)
422c98d5d94SLen Brown 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc2/t->tsc);
423103a8feaSLen Brown 	if (do_nhm_cstates)
424c98d5d94SLen Brown 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc3/t->tsc);
425103a8feaSLen Brown 	if (do_nhm_cstates)
426c98d5d94SLen Brown 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc6/t->tsc);
427103a8feaSLen Brown 	if (do_snb_cstates)
428c98d5d94SLen Brown 		outp += sprintf(outp, " %6.2f", 100.0 * p->pc7/t->tsc);
429c98d5d94SLen Brown done:
430c98d5d94SLen Brown 	outp += sprintf(outp, "\n");
431c98d5d94SLen Brown 
432c98d5d94SLen Brown 	return 0;
433103a8feaSLen Brown }
434103a8feaSLen Brown 
435c98d5d94SLen Brown void flush_stdout()
436103a8feaSLen Brown {
437c98d5d94SLen Brown 	fputs(output_buffer, stdout);
438c98d5d94SLen Brown 	outp = output_buffer;
439c98d5d94SLen Brown }
440c98d5d94SLen Brown void flush_stderr()
441c98d5d94SLen Brown {
442c98d5d94SLen Brown 	fputs(output_buffer, stderr);
443c98d5d94SLen Brown 	outp = output_buffer;
444c98d5d94SLen Brown }
445c98d5d94SLen Brown void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
446c98d5d94SLen Brown {
447e23da037SLen Brown 	static int printed;
448103a8feaSLen Brown 
449e23da037SLen Brown 	if (!printed || !summary_only)
450103a8feaSLen Brown 		print_header();
451103a8feaSLen Brown 
452c98d5d94SLen Brown 	if (topo.num_cpus > 1)
453c98d5d94SLen Brown 		format_counters(&average.threads, &average.cores,
454c98d5d94SLen Brown 			&average.packages);
455103a8feaSLen Brown 
456e23da037SLen Brown 	printed = 1;
457e23da037SLen Brown 
458e23da037SLen Brown 	if (summary_only)
459e23da037SLen Brown 		return;
460e23da037SLen Brown 
461c98d5d94SLen Brown 	for_all_cpus(format_counters, t, c, p);
462103a8feaSLen Brown }
463103a8feaSLen Brown 
464c98d5d94SLen Brown void
465c98d5d94SLen Brown delta_package(struct pkg_data *new, struct pkg_data *old)
466103a8feaSLen Brown {
467c98d5d94SLen Brown 	old->pc2 = new->pc2 - old->pc2;
468c98d5d94SLen Brown 	old->pc3 = new->pc3 - old->pc3;
469c98d5d94SLen Brown 	old->pc6 = new->pc6 - old->pc6;
470c98d5d94SLen Brown 	old->pc7 = new->pc7 - old->pc7;
471103a8feaSLen Brown }
472103a8feaSLen Brown 
473c98d5d94SLen Brown void
474c98d5d94SLen Brown delta_core(struct core_data *new, struct core_data *old)
475c98d5d94SLen Brown {
476c98d5d94SLen Brown 	old->c3 = new->c3 - old->c3;
477c98d5d94SLen Brown 	old->c6 = new->c6 - old->c6;
478c98d5d94SLen Brown 	old->c7 = new->c7 - old->c7;
479103a8feaSLen Brown }
480c98d5d94SLen Brown 
481c3ae331dSLen Brown /*
482c3ae331dSLen Brown  * old = new - old
483c3ae331dSLen Brown  */
484c98d5d94SLen Brown void
485c98d5d94SLen Brown delta_thread(struct thread_data *new, struct thread_data *old,
486c98d5d94SLen Brown 	struct core_data *core_delta)
487c98d5d94SLen Brown {
488c98d5d94SLen Brown 	old->tsc = new->tsc - old->tsc;
489c98d5d94SLen Brown 
490103a8feaSLen Brown 	/* check for TSC < 1 Mcycles over interval */
491c98d5d94SLen Brown 	if (old->tsc < (1000 * 1000)) {
492c98d5d94SLen Brown 		fprintf(stderr, "Insanely slow TSC rate, TSC stops in idle?\n");
493c98d5d94SLen Brown 		fprintf(stderr, "You can disable all c-states by booting with \"idle=poll\"\n");
494c98d5d94SLen Brown 		fprintf(stderr, "or just the deep ones with \"processor.max_cstate=1\"\n");
495103a8feaSLen Brown 		exit(-3);
496103a8feaSLen Brown 	}
497103a8feaSLen Brown 
498c98d5d94SLen Brown 	old->c1 = new->c1 - old->c1;
499c98d5d94SLen Brown 
500c98d5d94SLen Brown 	if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
501c98d5d94SLen Brown 		old->aperf = new->aperf - old->aperf;
502c98d5d94SLen Brown 		old->mperf = new->mperf - old->mperf;
503c98d5d94SLen Brown 	} else {
504c98d5d94SLen Brown 
505103a8feaSLen Brown 		if (!aperf_mperf_unstable) {
506103a8feaSLen Brown 			fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
507103a8feaSLen Brown 			fprintf(stderr, "* Frequency results do not cover entire interval *\n");
508103a8feaSLen Brown 			fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
509103a8feaSLen Brown 
510103a8feaSLen Brown 			aperf_mperf_unstable = 1;
511103a8feaSLen Brown 		}
512103a8feaSLen Brown 		/*
513103a8feaSLen Brown 		 * mperf delta is likely a huge "positive" number
514103a8feaSLen Brown 		 * can not use it for calculating c0 time
515103a8feaSLen Brown 		 */
516103a8feaSLen Brown 		skip_c0 = 1;
517103a8feaSLen Brown 		skip_c1 = 1;
518103a8feaSLen Brown 	}
519103a8feaSLen Brown 
520c98d5d94SLen Brown 
521103a8feaSLen Brown 	/*
522c3ae331dSLen Brown 	 * As counter collection is not atomic,
523c3ae331dSLen Brown 	 * it is possible for mperf's non-halted cycles + idle states
524103a8feaSLen Brown 	 * to exceed TSC's all cycles: show c1 = 0% in that case.
525103a8feaSLen Brown 	 */
526c3ae331dSLen Brown 	if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
527c98d5d94SLen Brown 		old->c1 = 0;
528c98d5d94SLen Brown 	else {
529c98d5d94SLen Brown 		/* normal case, derive c1 */
530c98d5d94SLen Brown 		old->c1 = old->tsc - old->mperf - core_delta->c3
531c98d5d94SLen Brown 				- core_delta->c6 - core_delta->c7;
532c98d5d94SLen Brown 	}
533c3ae331dSLen Brown 
534c98d5d94SLen Brown 	if (old->mperf == 0) {
535c3ae331dSLen Brown 		if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
536c98d5d94SLen Brown 		old->mperf = 1;	/* divide by 0 protection */
537c98d5d94SLen Brown 	}
538103a8feaSLen Brown 
5398e180f3cSLen Brown 	old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
5408e180f3cSLen Brown 	old->extra_delta32 &= 0xFFFFFFFF;
5418e180f3cSLen Brown 
5428e180f3cSLen Brown 	old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
5438e180f3cSLen Brown 
544103a8feaSLen Brown 	/*
5458e180f3cSLen Brown 	 * Extra MSR is just a snapshot, simply copy latest w/o subtracting
546103a8feaSLen Brown 	 */
5472f32edf1SLen Brown 	old->extra_msr32 = new->extra_msr32;
5482f32edf1SLen Brown 	old->extra_msr64 = new->extra_msr64;
549103a8feaSLen Brown }
550c98d5d94SLen Brown 
551c98d5d94SLen Brown int delta_cpu(struct thread_data *t, struct core_data *c,
552c98d5d94SLen Brown 	struct pkg_data *p, struct thread_data *t2,
553c98d5d94SLen Brown 	struct core_data *c2, struct pkg_data *p2)
554c98d5d94SLen Brown {
555c98d5d94SLen Brown 	/* calculate core delta only for 1st thread in core */
556c98d5d94SLen Brown 	if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
557c98d5d94SLen Brown 		delta_core(c, c2);
558c98d5d94SLen Brown 
559c98d5d94SLen Brown 	/* always calculate thread delta */
560c98d5d94SLen Brown 	delta_thread(t, t2, c2);	/* c2 is core delta */
561c98d5d94SLen Brown 
562c98d5d94SLen Brown 	/* calculate package delta only for 1st core in package */
563c98d5d94SLen Brown 	if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
564c98d5d94SLen Brown 		delta_package(p, p2);
565c98d5d94SLen Brown 
566103a8feaSLen Brown 	return 0;
567103a8feaSLen Brown }
568103a8feaSLen Brown 
569c98d5d94SLen Brown void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
570103a8feaSLen Brown {
571c98d5d94SLen Brown 	t->tsc = 0;
572c98d5d94SLen Brown 	t->aperf = 0;
573c98d5d94SLen Brown 	t->mperf = 0;
574c98d5d94SLen Brown 	t->c1 = 0;
575103a8feaSLen Brown 
5768e180f3cSLen Brown 	t->extra_delta32 = 0;
5778e180f3cSLen Brown 	t->extra_delta64 = 0;
5788e180f3cSLen Brown 
579c98d5d94SLen Brown 	/* tells format_counters to dump all fields from this set */
580c98d5d94SLen Brown 	t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
581c98d5d94SLen Brown 
582c98d5d94SLen Brown 	c->c3 = 0;
583c98d5d94SLen Brown 	c->c6 = 0;
584c98d5d94SLen Brown 	c->c7 = 0;
585c98d5d94SLen Brown 
586c98d5d94SLen Brown 	p->pc2 = 0;
587c98d5d94SLen Brown 	p->pc3 = 0;
588c98d5d94SLen Brown 	p->pc6 = 0;
589c98d5d94SLen Brown 	p->pc7 = 0;
590103a8feaSLen Brown }
591c98d5d94SLen Brown int sum_counters(struct thread_data *t, struct core_data *c,
592c98d5d94SLen Brown 	struct pkg_data *p)
593103a8feaSLen Brown {
594c98d5d94SLen Brown 	average.threads.tsc += t->tsc;
595c98d5d94SLen Brown 	average.threads.aperf += t->aperf;
596c98d5d94SLen Brown 	average.threads.mperf += t->mperf;
597c98d5d94SLen Brown 	average.threads.c1 += t->c1;
59815aaa346SLen Brown 
5998e180f3cSLen Brown 	average.threads.extra_delta32 += t->extra_delta32;
6008e180f3cSLen Brown 	average.threads.extra_delta64 += t->extra_delta64;
6018e180f3cSLen Brown 
602c98d5d94SLen Brown 	/* sum per-core values only for 1st thread in core */
603c98d5d94SLen Brown 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
604c98d5d94SLen Brown 		return 0;
605c98d5d94SLen Brown 
606c98d5d94SLen Brown 	average.cores.c3 += c->c3;
607c98d5d94SLen Brown 	average.cores.c6 += c->c6;
608c98d5d94SLen Brown 	average.cores.c7 += c->c7;
609c98d5d94SLen Brown 
610c98d5d94SLen Brown 	/* sum per-pkg values only for 1st core in pkg */
611c98d5d94SLen Brown 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
612c98d5d94SLen Brown 		return 0;
613c98d5d94SLen Brown 
614c98d5d94SLen Brown 	average.packages.pc2 += p->pc2;
615c98d5d94SLen Brown 	average.packages.pc3 += p->pc3;
616c98d5d94SLen Brown 	average.packages.pc6 += p->pc6;
617c98d5d94SLen Brown 	average.packages.pc7 += p->pc7;
618c98d5d94SLen Brown 
619c98d5d94SLen Brown 	return 0;
620c98d5d94SLen Brown }
621c98d5d94SLen Brown /*
622c98d5d94SLen Brown  * sum the counters for all cpus in the system
623c98d5d94SLen Brown  * compute the weighted average
624c98d5d94SLen Brown  */
625c98d5d94SLen Brown void compute_average(struct thread_data *t, struct core_data *c,
626c98d5d94SLen Brown 	struct pkg_data *p)
627c98d5d94SLen Brown {
628c98d5d94SLen Brown 	clear_counters(&average.threads, &average.cores, &average.packages);
629c98d5d94SLen Brown 
630c98d5d94SLen Brown 	for_all_cpus(sum_counters, t, c, p);
631c98d5d94SLen Brown 
632c98d5d94SLen Brown 	average.threads.tsc /= topo.num_cpus;
633c98d5d94SLen Brown 	average.threads.aperf /= topo.num_cpus;
634c98d5d94SLen Brown 	average.threads.mperf /= topo.num_cpus;
635c98d5d94SLen Brown 	average.threads.c1 /= topo.num_cpus;
636c98d5d94SLen Brown 
6378e180f3cSLen Brown 	average.threads.extra_delta32 /= topo.num_cpus;
6388e180f3cSLen Brown 	average.threads.extra_delta32 &= 0xFFFFFFFF;
6398e180f3cSLen Brown 
6408e180f3cSLen Brown 	average.threads.extra_delta64 /= topo.num_cpus;
6418e180f3cSLen Brown 
642c98d5d94SLen Brown 	average.cores.c3 /= topo.num_cores;
643c98d5d94SLen Brown 	average.cores.c6 /= topo.num_cores;
644c98d5d94SLen Brown 	average.cores.c7 /= topo.num_cores;
645c98d5d94SLen Brown 
646c98d5d94SLen Brown 	average.packages.pc2 /= topo.num_packages;
647c98d5d94SLen Brown 	average.packages.pc3 /= topo.num_packages;
648c98d5d94SLen Brown 	average.packages.pc6 /= topo.num_packages;
649c98d5d94SLen Brown 	average.packages.pc7 /= topo.num_packages;
650c98d5d94SLen Brown }
651c98d5d94SLen Brown 
652c98d5d94SLen Brown static unsigned long long rdtsc(void)
653c98d5d94SLen Brown {
654c98d5d94SLen Brown 	unsigned int low, high;
655c98d5d94SLen Brown 
656c98d5d94SLen Brown 	asm volatile("rdtsc" : "=a" (low), "=d" (high));
657c98d5d94SLen Brown 
658c98d5d94SLen Brown 	return low | ((unsigned long long)high) << 32;
659c98d5d94SLen Brown }
660c98d5d94SLen Brown 
661c98d5d94SLen Brown 
662c98d5d94SLen Brown /*
663c98d5d94SLen Brown  * get_counters(...)
664c98d5d94SLen Brown  * migrate to cpu
665c98d5d94SLen Brown  * acquire and record local counters for that cpu
666c98d5d94SLen Brown  */
667c98d5d94SLen Brown int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
668c98d5d94SLen Brown {
669c98d5d94SLen Brown 	int cpu = t->cpu_id;
670c98d5d94SLen Brown 
671c98d5d94SLen Brown 	if (cpu_migrate(cpu))
67215aaa346SLen Brown 		return -1;
67315aaa346SLen Brown 
674c98d5d94SLen Brown 	t->tsc = rdtsc();	/* we are running on local CPU of interest */
67515aaa346SLen Brown 
67615aaa346SLen Brown 	if (has_aperf) {
677c98d5d94SLen Brown 		if (get_msr(cpu, MSR_APERF, &t->aperf))
678c98d5d94SLen Brown 			return -3;
679c98d5d94SLen Brown 		if (get_msr(cpu, MSR_MPERF, &t->mperf))
680c98d5d94SLen Brown 			return -4;
68188c3281fSLen Brown 	}
68288c3281fSLen Brown 
6838e180f3cSLen Brown 	if (extra_delta_offset32) {
6848e180f3cSLen Brown 		if (get_msr(cpu, extra_delta_offset32, &t->extra_delta32))
6852f32edf1SLen Brown 			return -5;
6868e180f3cSLen Brown 		t->extra_delta32 &= 0xFFFFFFFF;
6878e180f3cSLen Brown 	}
6888e180f3cSLen Brown 
6898e180f3cSLen Brown 	if (extra_delta_offset64)
6908e180f3cSLen Brown 		if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
6918e180f3cSLen Brown 			return -5;
6928e180f3cSLen Brown 
6938e180f3cSLen Brown 	if (extra_msr_offset32) {
6948e180f3cSLen Brown 		if (get_msr(cpu, extra_msr_offset32, &t->extra_msr32))
6958e180f3cSLen Brown 			return -5;
6968e180f3cSLen Brown 		t->extra_msr32 &= 0xFFFFFFFF;
6978e180f3cSLen Brown 	}
6982f32edf1SLen Brown 
6992f32edf1SLen Brown 	if (extra_msr_offset64)
7002f32edf1SLen Brown 		if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
701c98d5d94SLen Brown 			return -5;
702c98d5d94SLen Brown 
703c98d5d94SLen Brown 	/* collect core counters only for 1st thread in core */
704c98d5d94SLen Brown 	if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
705c98d5d94SLen Brown 		return 0;
706c98d5d94SLen Brown 
70715aaa346SLen Brown 	if (do_nhm_cstates) {
708c98d5d94SLen Brown 		if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
709c98d5d94SLen Brown 			return -6;
710c98d5d94SLen Brown 		if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
711c98d5d94SLen Brown 			return -7;
712103a8feaSLen Brown 	}
71315aaa346SLen Brown 
71415aaa346SLen Brown 	if (do_snb_cstates)
715c98d5d94SLen Brown 		if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
716c98d5d94SLen Brown 			return -8;
717c98d5d94SLen Brown 
718c98d5d94SLen Brown 	/* collect package counters only for 1st core in package */
719c98d5d94SLen Brown 	if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
720c98d5d94SLen Brown 		return 0;
72115aaa346SLen Brown 
72215aaa346SLen Brown 	if (do_nhm_cstates) {
723c98d5d94SLen Brown 		if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
724c98d5d94SLen Brown 			return -9;
725c98d5d94SLen Brown 		if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
726c98d5d94SLen Brown 			return -10;
72715aaa346SLen Brown 	}
72815aaa346SLen Brown 	if (do_snb_cstates) {
729c98d5d94SLen Brown 		if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
730c98d5d94SLen Brown 			return -11;
731c98d5d94SLen Brown 		if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
732c98d5d94SLen Brown 			return -12;
73315aaa346SLen Brown 	}
73415aaa346SLen Brown 	return 0;
735103a8feaSLen Brown }
736103a8feaSLen Brown 
737c98d5d94SLen Brown void print_verbose_header(void)
738103a8feaSLen Brown {
739103a8feaSLen Brown 	unsigned long long msr;
740103a8feaSLen Brown 	unsigned int ratio;
741103a8feaSLen Brown 
742103a8feaSLen Brown 	if (!do_nehalem_platform_info)
743103a8feaSLen Brown 		return;
744103a8feaSLen Brown 
74515aaa346SLen Brown 	get_msr(0, MSR_NEHALEM_PLATFORM_INFO, &msr);
746103a8feaSLen Brown 
7476574a5d5SLen Brown 	if (verbose > 1)
7486574a5d5SLen Brown 		fprintf(stderr, "MSR_NEHALEM_PLATFORM_INFO: 0x%llx\n", msr);
7496574a5d5SLen Brown 
750103a8feaSLen Brown 	ratio = (msr >> 40) & 0xFF;
751103a8feaSLen Brown 	fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
752103a8feaSLen Brown 		ratio, bclk, ratio * bclk);
753103a8feaSLen Brown 
754103a8feaSLen Brown 	ratio = (msr >> 8) & 0xFF;
755103a8feaSLen Brown 	fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
756103a8feaSLen Brown 		ratio, bclk, ratio * bclk);
757103a8feaSLen Brown 
7586574a5d5SLen Brown 	if (!do_ivt_turbo_ratio_limit)
7596574a5d5SLen Brown 		goto print_nhm_turbo_ratio_limits;
7606574a5d5SLen Brown 
7616574a5d5SLen Brown 	get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr);
7626574a5d5SLen Brown 
763103a8feaSLen Brown 	if (verbose > 1)
7646574a5d5SLen Brown 		fprintf(stderr, "MSR_IVT_TURBO_RATIO_LIMIT: 0x%llx\n", msr);
7656574a5d5SLen Brown 
7666574a5d5SLen Brown 	ratio = (msr >> 56) & 0xFF;
7676574a5d5SLen Brown 	if (ratio)
7686574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
7696574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
7706574a5d5SLen Brown 
7716574a5d5SLen Brown 	ratio = (msr >> 48) & 0xFF;
7726574a5d5SLen Brown 	if (ratio)
7736574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
7746574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
7756574a5d5SLen Brown 
7766574a5d5SLen Brown 	ratio = (msr >> 40) & 0xFF;
7776574a5d5SLen Brown 	if (ratio)
7786574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
7796574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
7806574a5d5SLen Brown 
7816574a5d5SLen Brown 	ratio = (msr >> 32) & 0xFF;
7826574a5d5SLen Brown 	if (ratio)
7836574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
7846574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
7856574a5d5SLen Brown 
7866574a5d5SLen Brown 	ratio = (msr >> 24) & 0xFF;
7876574a5d5SLen Brown 	if (ratio)
7886574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
7896574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
7906574a5d5SLen Brown 
7916574a5d5SLen Brown 	ratio = (msr >> 16) & 0xFF;
7926574a5d5SLen Brown 	if (ratio)
7936574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
7946574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
7956574a5d5SLen Brown 
7966574a5d5SLen Brown 	ratio = (msr >> 8) & 0xFF;
7976574a5d5SLen Brown 	if (ratio)
7986574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
7996574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
8006574a5d5SLen Brown 
8016574a5d5SLen Brown 	ratio = (msr >> 0) & 0xFF;
8026574a5d5SLen Brown 	if (ratio)
8036574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
8046574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
8056574a5d5SLen Brown 
8066574a5d5SLen Brown print_nhm_turbo_ratio_limits:
807103a8feaSLen Brown 
808103a8feaSLen Brown 	if (!do_nehalem_turbo_ratio_limit)
809103a8feaSLen Brown 		return;
810103a8feaSLen Brown 
81115aaa346SLen Brown 	get_msr(0, MSR_NEHALEM_TURBO_RATIO_LIMIT, &msr);
812103a8feaSLen Brown 
8136574a5d5SLen Brown 	if (verbose > 1)
8146574a5d5SLen Brown 		fprintf(stderr, "MSR_NEHALEM_TURBO_RATIO_LIMIT: 0x%llx\n", msr);
8156574a5d5SLen Brown 
8166574a5d5SLen Brown 	ratio = (msr >> 56) & 0xFF;
8176574a5d5SLen Brown 	if (ratio)
8186574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
8196574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
8206574a5d5SLen Brown 
8216574a5d5SLen Brown 	ratio = (msr >> 48) & 0xFF;
8226574a5d5SLen Brown 	if (ratio)
8236574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
8246574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
8256574a5d5SLen Brown 
8266574a5d5SLen Brown 	ratio = (msr >> 40) & 0xFF;
8276574a5d5SLen Brown 	if (ratio)
8286574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
8296574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
8306574a5d5SLen Brown 
8316574a5d5SLen Brown 	ratio = (msr >> 32) & 0xFF;
8326574a5d5SLen Brown 	if (ratio)
8336574a5d5SLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
8346574a5d5SLen Brown 			ratio, bclk, ratio * bclk);
8356574a5d5SLen Brown 
836103a8feaSLen Brown 	ratio = (msr >> 24) & 0xFF;
837103a8feaSLen Brown 	if (ratio)
838103a8feaSLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
839103a8feaSLen Brown 			ratio, bclk, ratio * bclk);
840103a8feaSLen Brown 
841103a8feaSLen Brown 	ratio = (msr >> 16) & 0xFF;
842103a8feaSLen Brown 	if (ratio)
843103a8feaSLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
844103a8feaSLen Brown 			ratio, bclk, ratio * bclk);
845103a8feaSLen Brown 
846103a8feaSLen Brown 	ratio = (msr >> 8) & 0xFF;
847103a8feaSLen Brown 	if (ratio)
848103a8feaSLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
849103a8feaSLen Brown 			ratio, bclk, ratio * bclk);
850103a8feaSLen Brown 
851103a8feaSLen Brown 	ratio = (msr >> 0) & 0xFF;
852103a8feaSLen Brown 	if (ratio)
853103a8feaSLen Brown 		fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
854103a8feaSLen Brown 			ratio, bclk, ratio * bclk);
855103a8feaSLen Brown }
856103a8feaSLen Brown 
857c98d5d94SLen Brown void free_all_buffers(void)
858103a8feaSLen Brown {
859c98d5d94SLen Brown 	CPU_FREE(cpu_present_set);
860c98d5d94SLen Brown 	cpu_present_set = NULL;
861c98d5d94SLen Brown 	cpu_present_set = 0;
862103a8feaSLen Brown 
863c98d5d94SLen Brown 	CPU_FREE(cpu_affinity_set);
864c98d5d94SLen Brown 	cpu_affinity_set = NULL;
865c98d5d94SLen Brown 	cpu_affinity_setsize = 0;
866103a8feaSLen Brown 
867c98d5d94SLen Brown 	free(thread_even);
868c98d5d94SLen Brown 	free(core_even);
869c98d5d94SLen Brown 	free(package_even);
870103a8feaSLen Brown 
871c98d5d94SLen Brown 	thread_even = NULL;
872c98d5d94SLen Brown 	core_even = NULL;
873c98d5d94SLen Brown 	package_even = NULL;
874103a8feaSLen Brown 
875c98d5d94SLen Brown 	free(thread_odd);
876c98d5d94SLen Brown 	free(core_odd);
877c98d5d94SLen Brown 	free(package_odd);
878103a8feaSLen Brown 
879c98d5d94SLen Brown 	thread_odd = NULL;
880c98d5d94SLen Brown 	core_odd = NULL;
881c98d5d94SLen Brown 	package_odd = NULL;
882103a8feaSLen Brown 
883c98d5d94SLen Brown 	free(output_buffer);
884c98d5d94SLen Brown 	output_buffer = NULL;
885c98d5d94SLen Brown 	outp = NULL;
886103a8feaSLen Brown }
887103a8feaSLen Brown 
888103a8feaSLen Brown /*
889c98d5d94SLen Brown  * cpu_is_first_sibling_in_core(cpu)
890c98d5d94SLen Brown  * return 1 if given CPU is 1st HT sibling in the core
891103a8feaSLen Brown  */
892c98d5d94SLen Brown int cpu_is_first_sibling_in_core(int cpu)
893103a8feaSLen Brown {
894c98d5d94SLen Brown 	char path[64];
895c98d5d94SLen Brown 	FILE *filep;
896c98d5d94SLen Brown 	int first_cpu;
897103a8feaSLen Brown 
898c98d5d94SLen Brown 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
899c98d5d94SLen Brown 	filep = fopen(path, "r");
900c98d5d94SLen Brown 	if (filep == NULL) {
901c98d5d94SLen Brown 		perror(path);
902103a8feaSLen Brown 		exit(1);
903103a8feaSLen Brown 	}
904c98d5d94SLen Brown 	fscanf(filep, "%d", &first_cpu);
905c98d5d94SLen Brown 	fclose(filep);
906c98d5d94SLen Brown 	return (cpu == first_cpu);
907c98d5d94SLen Brown }
908103a8feaSLen Brown 
909c98d5d94SLen Brown /*
910c98d5d94SLen Brown  * cpu_is_first_core_in_package(cpu)
911c98d5d94SLen Brown  * return 1 if given CPU is 1st core in package
912c98d5d94SLen Brown  */
913c98d5d94SLen Brown int cpu_is_first_core_in_package(int cpu)
914c98d5d94SLen Brown {
915c98d5d94SLen Brown 	char path[64];
916c98d5d94SLen Brown 	FILE *filep;
917c98d5d94SLen Brown 	int first_cpu;
918c98d5d94SLen Brown 
919c98d5d94SLen Brown 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
920c98d5d94SLen Brown 	filep = fopen(path, "r");
921c98d5d94SLen Brown 	if (filep == NULL) {
922c98d5d94SLen Brown 		perror(path);
923103a8feaSLen Brown 		exit(1);
924103a8feaSLen Brown 	}
925c98d5d94SLen Brown 	fscanf(filep, "%d", &first_cpu);
926c98d5d94SLen Brown 	fclose(filep);
927c98d5d94SLen Brown 	return (cpu == first_cpu);
928103a8feaSLen Brown }
929103a8feaSLen Brown 
930103a8feaSLen Brown int get_physical_package_id(int cpu)
931103a8feaSLen Brown {
932c98d5d94SLen Brown 	char path[80];
933103a8feaSLen Brown 	FILE *filep;
934103a8feaSLen Brown 	int pkg;
935103a8feaSLen Brown 
936103a8feaSLen Brown 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
937103a8feaSLen Brown 	filep = fopen(path, "r");
938103a8feaSLen Brown 	if (filep == NULL) {
939103a8feaSLen Brown 		perror(path);
940103a8feaSLen Brown 		exit(1);
941103a8feaSLen Brown 	}
942103a8feaSLen Brown 	fscanf(filep, "%d", &pkg);
943103a8feaSLen Brown 	fclose(filep);
944103a8feaSLen Brown 	return pkg;
945103a8feaSLen Brown }
946103a8feaSLen Brown 
947103a8feaSLen Brown int get_core_id(int cpu)
948103a8feaSLen Brown {
949c98d5d94SLen Brown 	char path[80];
950103a8feaSLen Brown 	FILE *filep;
951103a8feaSLen Brown 	int core;
952103a8feaSLen Brown 
953103a8feaSLen Brown 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
954103a8feaSLen Brown 	filep = fopen(path, "r");
955103a8feaSLen Brown 	if (filep == NULL) {
956103a8feaSLen Brown 		perror(path);
957103a8feaSLen Brown 		exit(1);
958103a8feaSLen Brown 	}
959103a8feaSLen Brown 	fscanf(filep, "%d", &core);
960103a8feaSLen Brown 	fclose(filep);
961103a8feaSLen Brown 	return core;
962103a8feaSLen Brown }
963103a8feaSLen Brown 
964c98d5d94SLen Brown int get_num_ht_siblings(int cpu)
965c98d5d94SLen Brown {
966c98d5d94SLen Brown 	char path[80];
967c98d5d94SLen Brown 	FILE *filep;
968c98d5d94SLen Brown 	int sib1, sib2;
969c98d5d94SLen Brown 	int matches;
970c98d5d94SLen Brown 	char character;
971c98d5d94SLen Brown 
972c98d5d94SLen Brown 	sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
973c98d5d94SLen Brown 	filep = fopen(path, "r");
974c98d5d94SLen Brown 	if (filep == NULL) {
975c98d5d94SLen Brown 		perror(path);
976c98d5d94SLen Brown 		exit(1);
977c98d5d94SLen Brown 	}
978103a8feaSLen Brown 	/*
979c98d5d94SLen Brown 	 * file format:
980c98d5d94SLen Brown 	 * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
981c98d5d94SLen Brown 	 * otherwinse 1 sibling (self).
982c98d5d94SLen Brown 	 */
983c98d5d94SLen Brown 	matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
984c98d5d94SLen Brown 
985c98d5d94SLen Brown 	fclose(filep);
986c98d5d94SLen Brown 
987c98d5d94SLen Brown 	if (matches == 3)
988c98d5d94SLen Brown 		return 2;
989c98d5d94SLen Brown 	else
990c98d5d94SLen Brown 		return 1;
991c98d5d94SLen Brown }
992c98d5d94SLen Brown 
993c98d5d94SLen Brown /*
994c98d5d94SLen Brown  * run func(thread, core, package) in topology order
995c98d5d94SLen Brown  * skip non-present cpus
996103a8feaSLen Brown  */
997103a8feaSLen Brown 
998c98d5d94SLen Brown int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
999c98d5d94SLen Brown 	struct pkg_data *, struct thread_data *, struct core_data *,
1000c98d5d94SLen Brown 	struct pkg_data *), struct thread_data *thread_base,
1001c98d5d94SLen Brown 	struct core_data *core_base, struct pkg_data *pkg_base,
1002c98d5d94SLen Brown 	struct thread_data *thread_base2, struct core_data *core_base2,
1003c98d5d94SLen Brown 	struct pkg_data *pkg_base2)
1004c98d5d94SLen Brown {
1005c98d5d94SLen Brown 	int retval, pkg_no, core_no, thread_no;
1006c98d5d94SLen Brown 
1007c98d5d94SLen Brown 	for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1008c98d5d94SLen Brown 		for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1009c98d5d94SLen Brown 			for (thread_no = 0; thread_no <
1010c98d5d94SLen Brown 				topo.num_threads_per_core; ++thread_no) {
1011c98d5d94SLen Brown 				struct thread_data *t, *t2;
1012c98d5d94SLen Brown 				struct core_data *c, *c2;
1013c98d5d94SLen Brown 				struct pkg_data *p, *p2;
1014c98d5d94SLen Brown 
1015c98d5d94SLen Brown 				t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1016c98d5d94SLen Brown 
1017c98d5d94SLen Brown 				if (cpu_is_not_present(t->cpu_id))
1018c98d5d94SLen Brown 					continue;
1019c98d5d94SLen Brown 
1020c98d5d94SLen Brown 				t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1021c98d5d94SLen Brown 
1022c98d5d94SLen Brown 				c = GET_CORE(core_base, core_no, pkg_no);
1023c98d5d94SLen Brown 				c2 = GET_CORE(core_base2, core_no, pkg_no);
1024c98d5d94SLen Brown 
1025c98d5d94SLen Brown 				p = GET_PKG(pkg_base, pkg_no);
1026c98d5d94SLen Brown 				p2 = GET_PKG(pkg_base2, pkg_no);
1027c98d5d94SLen Brown 
1028c98d5d94SLen Brown 				retval = func(t, c, p, t2, c2, p2);
1029c98d5d94SLen Brown 				if (retval)
1030c98d5d94SLen Brown 					return retval;
1031c98d5d94SLen Brown 			}
1032c98d5d94SLen Brown 		}
1033c98d5d94SLen Brown 	}
1034c98d5d94SLen Brown 	return 0;
1035c98d5d94SLen Brown }
1036c98d5d94SLen Brown 
1037c98d5d94SLen Brown /*
1038c98d5d94SLen Brown  * run func(cpu) on every cpu in /proc/stat
1039c98d5d94SLen Brown  * return max_cpu number
1040c98d5d94SLen Brown  */
1041c98d5d94SLen Brown int for_all_proc_cpus(int (func)(int))
1042103a8feaSLen Brown {
1043103a8feaSLen Brown 	FILE *fp;
1044c98d5d94SLen Brown 	int cpu_num;
1045103a8feaSLen Brown 	int retval;
1046103a8feaSLen Brown 
1047103a8feaSLen Brown 	fp = fopen(proc_stat, "r");
1048103a8feaSLen Brown 	if (fp == NULL) {
1049103a8feaSLen Brown 		perror(proc_stat);
1050103a8feaSLen Brown 		exit(1);
1051103a8feaSLen Brown 	}
1052103a8feaSLen Brown 
1053103a8feaSLen Brown 	retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1054103a8feaSLen Brown 	if (retval != 0) {
1055103a8feaSLen Brown 		perror("/proc/stat format");
1056103a8feaSLen Brown 		exit(1);
1057103a8feaSLen Brown 	}
1058103a8feaSLen Brown 
1059c98d5d94SLen Brown 	while (1) {
1060c98d5d94SLen Brown 		retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1061103a8feaSLen Brown 		if (retval != 1)
1062103a8feaSLen Brown 			break;
1063103a8feaSLen Brown 
1064c98d5d94SLen Brown 		retval = func(cpu_num);
1065c98d5d94SLen Brown 		if (retval) {
1066c98d5d94SLen Brown 			fclose(fp);
1067c98d5d94SLen Brown 			return(retval);
1068c98d5d94SLen Brown 		}
1069103a8feaSLen Brown 	}
1070103a8feaSLen Brown 	fclose(fp);
1071c98d5d94SLen Brown 	return 0;
1072103a8feaSLen Brown }
1073103a8feaSLen Brown 
1074103a8feaSLen Brown void re_initialize(void)
1075103a8feaSLen Brown {
1076c98d5d94SLen Brown 	free_all_buffers();
1077c98d5d94SLen Brown 	setup_all_buffers();
1078c98d5d94SLen Brown 	printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1079103a8feaSLen Brown }
1080103a8feaSLen Brown 
1081c98d5d94SLen Brown 
1082103a8feaSLen Brown /*
1083c98d5d94SLen Brown  * count_cpus()
1084c98d5d94SLen Brown  * remember the last one seen, it will be the max
1085103a8feaSLen Brown  */
1086c98d5d94SLen Brown int count_cpus(int cpu)
1087103a8feaSLen Brown {
1088c98d5d94SLen Brown 	if (topo.max_cpu_num < cpu)
1089c98d5d94SLen Brown 		topo.max_cpu_num = cpu;
1090103a8feaSLen Brown 
1091c98d5d94SLen Brown 	topo.num_cpus += 1;
1092c98d5d94SLen Brown 	return 0;
1093103a8feaSLen Brown }
1094c98d5d94SLen Brown int mark_cpu_present(int cpu)
1095c98d5d94SLen Brown {
1096c98d5d94SLen Brown 	CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
109715aaa346SLen Brown 	return 0;
1098103a8feaSLen Brown }
1099103a8feaSLen Brown 
1100103a8feaSLen Brown void turbostat_loop()
1101103a8feaSLen Brown {
1102c98d5d94SLen Brown 	int retval;
1103c98d5d94SLen Brown 
1104103a8feaSLen Brown restart:
1105c98d5d94SLen Brown 	retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1106d91bb17cSLen Brown 	if (retval < -1) {
1107d91bb17cSLen Brown 		exit(retval);
1108d91bb17cSLen Brown 	} else if (retval == -1) {
1109c98d5d94SLen Brown 		re_initialize();
1110c98d5d94SLen Brown 		goto restart;
1111c98d5d94SLen Brown 	}
1112103a8feaSLen Brown 	gettimeofday(&tv_even, (struct timezone *)NULL);
1113103a8feaSLen Brown 
1114103a8feaSLen Brown 	while (1) {
1115c98d5d94SLen Brown 		if (for_all_proc_cpus(cpu_is_not_present)) {
1116103a8feaSLen Brown 			re_initialize();
1117103a8feaSLen Brown 			goto restart;
1118103a8feaSLen Brown 		}
1119103a8feaSLen Brown 		sleep(interval_sec);
1120c98d5d94SLen Brown 		retval = for_all_cpus(get_counters, ODD_COUNTERS);
1121d91bb17cSLen Brown 		if (retval < -1) {
1122d91bb17cSLen Brown 			exit(retval);
1123d91bb17cSLen Brown 		} else if (retval == -1) {
112415aaa346SLen Brown 			re_initialize();
112515aaa346SLen Brown 			goto restart;
112615aaa346SLen Brown 		}
1127103a8feaSLen Brown 		gettimeofday(&tv_odd, (struct timezone *)NULL);
1128103a8feaSLen Brown 		timersub(&tv_odd, &tv_even, &tv_delta);
1129c98d5d94SLen Brown 		for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1130c98d5d94SLen Brown 		compute_average(EVEN_COUNTERS);
1131c98d5d94SLen Brown 		format_all_counters(EVEN_COUNTERS);
1132c98d5d94SLen Brown 		flush_stdout();
113315aaa346SLen Brown 		sleep(interval_sec);
1134c98d5d94SLen Brown 		retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1135d91bb17cSLen Brown 		if (retval < -1) {
1136d91bb17cSLen Brown 			exit(retval);
1137d91bb17cSLen Brown 		} else if (retval == -1) {
1138103a8feaSLen Brown 			re_initialize();
1139103a8feaSLen Brown 			goto restart;
1140103a8feaSLen Brown 		}
1141103a8feaSLen Brown 		gettimeofday(&tv_even, (struct timezone *)NULL);
1142103a8feaSLen Brown 		timersub(&tv_even, &tv_odd, &tv_delta);
1143c98d5d94SLen Brown 		for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1144c98d5d94SLen Brown 		compute_average(ODD_COUNTERS);
1145c98d5d94SLen Brown 		format_all_counters(ODD_COUNTERS);
1146c98d5d94SLen Brown 		flush_stdout();
1147103a8feaSLen Brown 	}
1148103a8feaSLen Brown }
1149103a8feaSLen Brown 
1150103a8feaSLen Brown void check_dev_msr()
1151103a8feaSLen Brown {
1152103a8feaSLen Brown 	struct stat sb;
1153103a8feaSLen Brown 
1154103a8feaSLen Brown 	if (stat("/dev/cpu/0/msr", &sb)) {
1155103a8feaSLen Brown 		fprintf(stderr, "no /dev/cpu/0/msr\n");
1156103a8feaSLen Brown 		fprintf(stderr, "Try \"# modprobe msr\"\n");
1157103a8feaSLen Brown 		exit(-5);
1158103a8feaSLen Brown 	}
1159103a8feaSLen Brown }
1160103a8feaSLen Brown 
1161103a8feaSLen Brown void check_super_user()
1162103a8feaSLen Brown {
1163103a8feaSLen Brown 	if (getuid() != 0) {
1164103a8feaSLen Brown 		fprintf(stderr, "must be root\n");
1165103a8feaSLen Brown 		exit(-6);
1166103a8feaSLen Brown 	}
1167103a8feaSLen Brown }
1168103a8feaSLen Brown 
1169103a8feaSLen Brown int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model)
1170103a8feaSLen Brown {
1171103a8feaSLen Brown 	if (!genuine_intel)
1172103a8feaSLen Brown 		return 0;
1173103a8feaSLen Brown 
1174103a8feaSLen Brown 	if (family != 6)
1175103a8feaSLen Brown 		return 0;
1176103a8feaSLen Brown 
1177103a8feaSLen Brown 	switch (model) {
1178103a8feaSLen Brown 	case 0x1A:	/* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1179103a8feaSLen Brown 	case 0x1E:	/* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1180103a8feaSLen Brown 	case 0x1F:	/* Core i7 and i5 Processor - Nehalem */
1181103a8feaSLen Brown 	case 0x25:	/* Westmere Client - Clarkdale, Arrandale */
1182103a8feaSLen Brown 	case 0x2C:	/* Westmere EP - Gulftown */
1183103a8feaSLen Brown 	case 0x2A:	/* SNB */
1184103a8feaSLen Brown 	case 0x2D:	/* SNB Xeon */
1185553575f1SLen Brown 	case 0x3A:	/* IVB */
11861300651bSLen Brown 	case 0x3E:	/* IVB Xeon */
1187103a8feaSLen Brown 		return 1;
1188103a8feaSLen Brown 	case 0x2E:	/* Nehalem-EX Xeon - Beckton */
1189103a8feaSLen Brown 	case 0x2F:	/* Westmere-EX Xeon - Eagleton */
1190103a8feaSLen Brown 	default:
1191103a8feaSLen Brown 		return 0;
1192103a8feaSLen Brown 	}
1193103a8feaSLen Brown }
11946574a5d5SLen Brown int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
11956574a5d5SLen Brown {
11966574a5d5SLen Brown 	if (!genuine_intel)
11976574a5d5SLen Brown 		return 0;
11986574a5d5SLen Brown 
11996574a5d5SLen Brown 	if (family != 6)
12006574a5d5SLen Brown 		return 0;
12016574a5d5SLen Brown 
12026574a5d5SLen Brown 	switch (model) {
12036574a5d5SLen Brown 	case 0x3E:	/* IVB Xeon */
12046574a5d5SLen Brown 		return 1;
12056574a5d5SLen Brown 	default:
12066574a5d5SLen Brown 		return 0;
12076574a5d5SLen Brown 	}
12086574a5d5SLen Brown }
12096574a5d5SLen Brown 
1210103a8feaSLen Brown 
1211103a8feaSLen Brown int is_snb(unsigned int family, unsigned int model)
1212103a8feaSLen Brown {
1213103a8feaSLen Brown 	if (!genuine_intel)
1214103a8feaSLen Brown 		return 0;
1215103a8feaSLen Brown 
1216103a8feaSLen Brown 	switch (model) {
1217103a8feaSLen Brown 	case 0x2A:
1218103a8feaSLen Brown 	case 0x2D:
1219650a37f3SLen Brown 	case 0x3A:	/* IVB */
12201300651bSLen Brown 	case 0x3E:	/* IVB Xeon */
1221103a8feaSLen Brown 		return 1;
1222103a8feaSLen Brown 	}
1223103a8feaSLen Brown 	return 0;
1224103a8feaSLen Brown }
1225103a8feaSLen Brown 
1226103a8feaSLen Brown double discover_bclk(unsigned int family, unsigned int model)
1227103a8feaSLen Brown {
1228103a8feaSLen Brown 	if (is_snb(family, model))
1229103a8feaSLen Brown 		return 100.00;
1230103a8feaSLen Brown 	else
1231103a8feaSLen Brown 		return 133.33;
1232103a8feaSLen Brown }
1233103a8feaSLen Brown 
1234103a8feaSLen Brown void check_cpuid()
1235103a8feaSLen Brown {
1236103a8feaSLen Brown 	unsigned int eax, ebx, ecx, edx, max_level;
1237103a8feaSLen Brown 	unsigned int fms, family, model, stepping;
1238103a8feaSLen Brown 
1239103a8feaSLen Brown 	eax = ebx = ecx = edx = 0;
1240103a8feaSLen Brown 
1241103a8feaSLen Brown 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0));
1242103a8feaSLen Brown 
1243103a8feaSLen Brown 	if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
1244103a8feaSLen Brown 		genuine_intel = 1;
1245103a8feaSLen Brown 
1246103a8feaSLen Brown 	if (verbose)
1247103a8feaSLen Brown 		fprintf(stderr, "%.4s%.4s%.4s ",
1248103a8feaSLen Brown 			(char *)&ebx, (char *)&edx, (char *)&ecx);
1249103a8feaSLen Brown 
1250103a8feaSLen Brown 	asm("cpuid" : "=a" (fms), "=c" (ecx), "=d" (edx) : "a" (1) : "ebx");
1251103a8feaSLen Brown 	family = (fms >> 8) & 0xf;
1252103a8feaSLen Brown 	model = (fms >> 4) & 0xf;
1253103a8feaSLen Brown 	stepping = fms & 0xf;
1254103a8feaSLen Brown 	if (family == 6 || family == 0xf)
1255103a8feaSLen Brown 		model += ((fms >> 16) & 0xf) << 4;
1256103a8feaSLen Brown 
1257103a8feaSLen Brown 	if (verbose)
1258103a8feaSLen Brown 		fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
1259103a8feaSLen Brown 			max_level, family, model, stepping, family, model, stepping);
1260103a8feaSLen Brown 
1261103a8feaSLen Brown 	if (!(edx & (1 << 5))) {
1262103a8feaSLen Brown 		fprintf(stderr, "CPUID: no MSR\n");
1263103a8feaSLen Brown 		exit(1);
1264103a8feaSLen Brown 	}
1265103a8feaSLen Brown 
1266103a8feaSLen Brown 	/*
1267103a8feaSLen Brown 	 * check max extended function levels of CPUID.
1268103a8feaSLen Brown 	 * This is needed to check for invariant TSC.
1269103a8feaSLen Brown 	 * This check is valid for both Intel and AMD.
1270103a8feaSLen Brown 	 */
1271103a8feaSLen Brown 	ebx = ecx = edx = 0;
1272103a8feaSLen Brown 	asm("cpuid" : "=a" (max_level), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000000));
1273103a8feaSLen Brown 
1274103a8feaSLen Brown 	if (max_level < 0x80000007) {
1275103a8feaSLen Brown 		fprintf(stderr, "CPUID: no invariant TSC (max_level 0x%x)\n", max_level);
1276103a8feaSLen Brown 		exit(1);
1277103a8feaSLen Brown 	}
1278103a8feaSLen Brown 
1279103a8feaSLen Brown 	/*
1280103a8feaSLen Brown 	 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
1281103a8feaSLen Brown 	 * this check is valid for both Intel and AMD
1282103a8feaSLen Brown 	 */
1283103a8feaSLen Brown 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x80000007));
12848209e054SThomas Renninger 	has_invariant_tsc = edx & (1 << 8);
1285103a8feaSLen Brown 
1286103a8feaSLen Brown 	if (!has_invariant_tsc) {
1287103a8feaSLen Brown 		fprintf(stderr, "No invariant TSC\n");
1288103a8feaSLen Brown 		exit(1);
1289103a8feaSLen Brown 	}
1290103a8feaSLen Brown 
1291103a8feaSLen Brown 	/*
1292103a8feaSLen Brown 	 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
1293103a8feaSLen Brown 	 * this check is valid for both Intel and AMD
1294103a8feaSLen Brown 	 */
1295103a8feaSLen Brown 
1296103a8feaSLen Brown 	asm("cpuid" : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx) : "a" (0x6));
12978209e054SThomas Renninger 	has_aperf = ecx & (1 << 0);
1298103a8feaSLen Brown 	if (!has_aperf) {
1299103a8feaSLen Brown 		fprintf(stderr, "No APERF MSR\n");
1300103a8feaSLen Brown 		exit(1);
1301103a8feaSLen Brown 	}
1302103a8feaSLen Brown 
1303103a8feaSLen Brown 	do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
1304103a8feaSLen Brown 	do_nhm_cstates = genuine_intel;	/* all Intel w/ non-stop TSC have NHM counters */
1305103a8feaSLen Brown 	do_snb_cstates = is_snb(family, model);
1306103a8feaSLen Brown 	bclk = discover_bclk(family, model);
1307103a8feaSLen Brown 
1308103a8feaSLen Brown 	do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model);
13096574a5d5SLen Brown 	do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model);
1310103a8feaSLen Brown }
1311103a8feaSLen Brown 
1312103a8feaSLen Brown 
1313103a8feaSLen Brown void usage()
1314103a8feaSLen Brown {
1315f9240813SLen Brown 	fprintf(stderr, "%s: [-v][-p|-P|-S][-c MSR# | -s]][-C MSR#][-m MSR#][-M MSR#][-i interval_sec | command ...]\n",
1316103a8feaSLen Brown 		progname);
1317103a8feaSLen Brown 	exit(1);
1318103a8feaSLen Brown }
1319103a8feaSLen Brown 
1320103a8feaSLen Brown 
1321103a8feaSLen Brown /*
1322103a8feaSLen Brown  * in /dev/cpu/ return success for names that are numbers
1323103a8feaSLen Brown  * ie. filter out ".", "..", "microcode".
1324103a8feaSLen Brown  */
1325103a8feaSLen Brown int dir_filter(const struct dirent *dirp)
1326103a8feaSLen Brown {
1327103a8feaSLen Brown 	if (isdigit(dirp->d_name[0]))
1328103a8feaSLen Brown 		return 1;
1329103a8feaSLen Brown 	else
1330103a8feaSLen Brown 		return 0;
1331103a8feaSLen Brown }
1332103a8feaSLen Brown 
1333103a8feaSLen Brown int open_dev_cpu_msr(int dummy1)
1334103a8feaSLen Brown {
1335103a8feaSLen Brown 	return 0;
1336103a8feaSLen Brown }
1337103a8feaSLen Brown 
1338c98d5d94SLen Brown void topology_probe()
1339c98d5d94SLen Brown {
1340c98d5d94SLen Brown 	int i;
1341c98d5d94SLen Brown 	int max_core_id = 0;
1342c98d5d94SLen Brown 	int max_package_id = 0;
1343c98d5d94SLen Brown 	int max_siblings = 0;
1344c98d5d94SLen Brown 	struct cpu_topology {
1345c98d5d94SLen Brown 		int core_id;
1346c98d5d94SLen Brown 		int physical_package_id;
1347c98d5d94SLen Brown 	} *cpus;
1348c98d5d94SLen Brown 
1349c98d5d94SLen Brown 	/* Initialize num_cpus, max_cpu_num */
1350c98d5d94SLen Brown 	topo.num_cpus = 0;
1351c98d5d94SLen Brown 	topo.max_cpu_num = 0;
1352c98d5d94SLen Brown 	for_all_proc_cpus(count_cpus);
1353c98d5d94SLen Brown 	if (!summary_only && topo.num_cpus > 1)
1354c98d5d94SLen Brown 		show_cpu = 1;
1355c98d5d94SLen Brown 
1356c98d5d94SLen Brown 	if (verbose > 1)
1357c98d5d94SLen Brown 		fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
1358c98d5d94SLen Brown 
1359c98d5d94SLen Brown 	cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
1360c98d5d94SLen Brown 	if (cpus == NULL) {
1361c98d5d94SLen Brown 		perror("calloc cpus");
1362c98d5d94SLen Brown 		exit(1);
1363c98d5d94SLen Brown 	}
1364c98d5d94SLen Brown 
1365c98d5d94SLen Brown 	/*
1366c98d5d94SLen Brown 	 * Allocate and initialize cpu_present_set
1367c98d5d94SLen Brown 	 */
1368c98d5d94SLen Brown 	cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
1369c98d5d94SLen Brown 	if (cpu_present_set == NULL) {
1370c98d5d94SLen Brown 		perror("CPU_ALLOC");
1371c98d5d94SLen Brown 		exit(3);
1372c98d5d94SLen Brown 	}
1373c98d5d94SLen Brown 	cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1374c98d5d94SLen Brown 	CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
1375c98d5d94SLen Brown 	for_all_proc_cpus(mark_cpu_present);
1376c98d5d94SLen Brown 
1377c98d5d94SLen Brown 	/*
1378c98d5d94SLen Brown 	 * Allocate and initialize cpu_affinity_set
1379c98d5d94SLen Brown 	 */
1380c98d5d94SLen Brown 	cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
1381c98d5d94SLen Brown 	if (cpu_affinity_set == NULL) {
1382c98d5d94SLen Brown 		perror("CPU_ALLOC");
1383c98d5d94SLen Brown 		exit(3);
1384c98d5d94SLen Brown 	}
1385c98d5d94SLen Brown 	cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
1386c98d5d94SLen Brown 	CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
1387c98d5d94SLen Brown 
1388c98d5d94SLen Brown 
1389c98d5d94SLen Brown 	/*
1390c98d5d94SLen Brown 	 * For online cpus
1391c98d5d94SLen Brown 	 * find max_core_id, max_package_id
1392c98d5d94SLen Brown 	 */
1393c98d5d94SLen Brown 	for (i = 0; i <= topo.max_cpu_num; ++i) {
1394c98d5d94SLen Brown 		int siblings;
1395c98d5d94SLen Brown 
1396c98d5d94SLen Brown 		if (cpu_is_not_present(i)) {
1397c98d5d94SLen Brown 			if (verbose > 1)
1398c98d5d94SLen Brown 				fprintf(stderr, "cpu%d NOT PRESENT\n", i);
1399c98d5d94SLen Brown 			continue;
1400c98d5d94SLen Brown 		}
1401c98d5d94SLen Brown 		cpus[i].core_id = get_core_id(i);
1402c98d5d94SLen Brown 		if (cpus[i].core_id > max_core_id)
1403c98d5d94SLen Brown 			max_core_id = cpus[i].core_id;
1404c98d5d94SLen Brown 
1405c98d5d94SLen Brown 		cpus[i].physical_package_id = get_physical_package_id(i);
1406c98d5d94SLen Brown 		if (cpus[i].physical_package_id > max_package_id)
1407c98d5d94SLen Brown 			max_package_id = cpus[i].physical_package_id;
1408c98d5d94SLen Brown 
1409c98d5d94SLen Brown 		siblings = get_num_ht_siblings(i);
1410c98d5d94SLen Brown 		if (siblings > max_siblings)
1411c98d5d94SLen Brown 			max_siblings = siblings;
1412c98d5d94SLen Brown 		if (verbose > 1)
1413c98d5d94SLen Brown 			fprintf(stderr, "cpu %d pkg %d core %d\n",
1414c98d5d94SLen Brown 				i, cpus[i].physical_package_id, cpus[i].core_id);
1415c98d5d94SLen Brown 	}
1416c98d5d94SLen Brown 	topo.num_cores_per_pkg = max_core_id + 1;
1417c98d5d94SLen Brown 	if (verbose > 1)
1418c98d5d94SLen Brown 		fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
1419c98d5d94SLen Brown 			max_core_id, topo.num_cores_per_pkg);
1420c98d5d94SLen Brown 	if (!summary_only && topo.num_cores_per_pkg > 1)
1421c98d5d94SLen Brown 		show_core = 1;
1422c98d5d94SLen Brown 
1423c98d5d94SLen Brown 	topo.num_packages = max_package_id + 1;
1424c98d5d94SLen Brown 	if (verbose > 1)
1425c98d5d94SLen Brown 		fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
1426c98d5d94SLen Brown 			max_package_id, topo.num_packages);
1427c98d5d94SLen Brown 	if (!summary_only && topo.num_packages > 1)
1428c98d5d94SLen Brown 		show_pkg = 1;
1429c98d5d94SLen Brown 
1430c98d5d94SLen Brown 	topo.num_threads_per_core = max_siblings;
1431c98d5d94SLen Brown 	if (verbose > 1)
1432c98d5d94SLen Brown 		fprintf(stderr, "max_siblings %d\n", max_siblings);
1433c98d5d94SLen Brown 
1434c98d5d94SLen Brown 	free(cpus);
1435c98d5d94SLen Brown }
1436c98d5d94SLen Brown 
1437c98d5d94SLen Brown void
1438c98d5d94SLen Brown allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
1439c98d5d94SLen Brown {
1440c98d5d94SLen Brown 	int i;
1441c98d5d94SLen Brown 
1442c98d5d94SLen Brown 	*t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
1443c98d5d94SLen Brown 		topo.num_packages, sizeof(struct thread_data));
1444c98d5d94SLen Brown 	if (*t == NULL)
1445c98d5d94SLen Brown 		goto error;
1446c98d5d94SLen Brown 
1447c98d5d94SLen Brown 	for (i = 0; i < topo.num_threads_per_core *
1448c98d5d94SLen Brown 		topo.num_cores_per_pkg * topo.num_packages; i++)
1449c98d5d94SLen Brown 		(*t)[i].cpu_id = -1;
1450c98d5d94SLen Brown 
1451c98d5d94SLen Brown 	*c = calloc(topo.num_cores_per_pkg * topo.num_packages,
1452c98d5d94SLen Brown 		sizeof(struct core_data));
1453c98d5d94SLen Brown 	if (*c == NULL)
1454c98d5d94SLen Brown 		goto error;
1455c98d5d94SLen Brown 
1456c98d5d94SLen Brown 	for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
1457c98d5d94SLen Brown 		(*c)[i].core_id = -1;
1458c98d5d94SLen Brown 
1459c98d5d94SLen Brown 	*p = calloc(topo.num_packages, sizeof(struct pkg_data));
1460c98d5d94SLen Brown 	if (*p == NULL)
1461c98d5d94SLen Brown 		goto error;
1462c98d5d94SLen Brown 
1463c98d5d94SLen Brown 	for (i = 0; i < topo.num_packages; i++)
1464c98d5d94SLen Brown 		(*p)[i].package_id = i;
1465c98d5d94SLen Brown 
1466c98d5d94SLen Brown 	return;
1467c98d5d94SLen Brown error:
1468c98d5d94SLen Brown 	perror("calloc counters");
1469c98d5d94SLen Brown 	exit(1);
1470c98d5d94SLen Brown }
1471c98d5d94SLen Brown /*
1472c98d5d94SLen Brown  * init_counter()
1473c98d5d94SLen Brown  *
1474c98d5d94SLen Brown  * set cpu_id, core_num, pkg_num
1475c98d5d94SLen Brown  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
1476c98d5d94SLen Brown  *
1477c98d5d94SLen Brown  * increment topo.num_cores when 1st core in pkg seen
1478c98d5d94SLen Brown  */
1479c98d5d94SLen Brown void init_counter(struct thread_data *thread_base, struct core_data *core_base,
1480c98d5d94SLen Brown 	struct pkg_data *pkg_base, int thread_num, int core_num,
1481c98d5d94SLen Brown 	int pkg_num, int cpu_id)
1482c98d5d94SLen Brown {
1483c98d5d94SLen Brown 	struct thread_data *t;
1484c98d5d94SLen Brown 	struct core_data *c;
1485c98d5d94SLen Brown 	struct pkg_data *p;
1486c98d5d94SLen Brown 
1487c98d5d94SLen Brown 	t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
1488c98d5d94SLen Brown 	c = GET_CORE(core_base, core_num, pkg_num);
1489c98d5d94SLen Brown 	p = GET_PKG(pkg_base, pkg_num);
1490c98d5d94SLen Brown 
1491c98d5d94SLen Brown 	t->cpu_id = cpu_id;
1492c98d5d94SLen Brown 	if (thread_num == 0) {
1493c98d5d94SLen Brown 		t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
1494c98d5d94SLen Brown 		if (cpu_is_first_core_in_package(cpu_id))
1495c98d5d94SLen Brown 			t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
1496c98d5d94SLen Brown 	}
1497c98d5d94SLen Brown 
1498c98d5d94SLen Brown 	c->core_id = core_num;
1499c98d5d94SLen Brown 	p->package_id = pkg_num;
1500c98d5d94SLen Brown }
1501c98d5d94SLen Brown 
1502c98d5d94SLen Brown 
1503c98d5d94SLen Brown int initialize_counters(int cpu_id)
1504c98d5d94SLen Brown {
1505c98d5d94SLen Brown 	int my_thread_id, my_core_id, my_package_id;
1506c98d5d94SLen Brown 
1507c98d5d94SLen Brown 	my_package_id = get_physical_package_id(cpu_id);
1508c98d5d94SLen Brown 	my_core_id = get_core_id(cpu_id);
1509c98d5d94SLen Brown 
1510c98d5d94SLen Brown 	if (cpu_is_first_sibling_in_core(cpu_id)) {
1511c98d5d94SLen Brown 		my_thread_id = 0;
1512c98d5d94SLen Brown 		topo.num_cores++;
1513c98d5d94SLen Brown 	} else {
1514c98d5d94SLen Brown 		my_thread_id = 1;
1515c98d5d94SLen Brown 	}
1516c98d5d94SLen Brown 
1517c98d5d94SLen Brown 	init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1518c98d5d94SLen Brown 	init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
1519c98d5d94SLen Brown 	return 0;
1520c98d5d94SLen Brown }
1521c98d5d94SLen Brown 
1522c98d5d94SLen Brown void allocate_output_buffer()
1523c98d5d94SLen Brown {
1524c98d5d94SLen Brown 	output_buffer = calloc(1, (1 + topo.num_cpus) * 128);
1525c98d5d94SLen Brown 	outp = output_buffer;
1526c98d5d94SLen Brown 	if (outp == NULL) {
1527c98d5d94SLen Brown 		perror("calloc");
1528c98d5d94SLen Brown 		exit(-1);
1529c98d5d94SLen Brown 	}
1530c98d5d94SLen Brown }
1531c98d5d94SLen Brown 
1532c98d5d94SLen Brown void setup_all_buffers(void)
1533c98d5d94SLen Brown {
1534c98d5d94SLen Brown 	topology_probe();
1535c98d5d94SLen Brown 	allocate_counters(&thread_even, &core_even, &package_even);
1536c98d5d94SLen Brown 	allocate_counters(&thread_odd, &core_odd, &package_odd);
1537c98d5d94SLen Brown 	allocate_output_buffer();
1538c98d5d94SLen Brown 	for_all_proc_cpus(initialize_counters);
1539c98d5d94SLen Brown }
1540103a8feaSLen Brown void turbostat_init()
1541103a8feaSLen Brown {
1542103a8feaSLen Brown 	check_cpuid();
1543103a8feaSLen Brown 
1544103a8feaSLen Brown 	check_dev_msr();
1545103a8feaSLen Brown 	check_super_user();
1546103a8feaSLen Brown 
1547c98d5d94SLen Brown 	setup_all_buffers();
1548103a8feaSLen Brown 
1549103a8feaSLen Brown 	if (verbose)
1550c98d5d94SLen Brown 		print_verbose_header();
1551103a8feaSLen Brown }
1552103a8feaSLen Brown 
1553103a8feaSLen Brown int fork_it(char **argv)
1554103a8feaSLen Brown {
1555103a8feaSLen Brown 	pid_t child_pid;
1556d91bb17cSLen Brown 	int status;
1557d15cf7c1SLen Brown 
1558d91bb17cSLen Brown 	status = for_all_cpus(get_counters, EVEN_COUNTERS);
1559d91bb17cSLen Brown 	if (status)
1560d91bb17cSLen Brown 		exit(status);
1561d15cf7c1SLen Brown 	/* clear affinity side-effect of get_counters() */
1562d15cf7c1SLen Brown 	sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
1563103a8feaSLen Brown 	gettimeofday(&tv_even, (struct timezone *)NULL);
1564103a8feaSLen Brown 
1565103a8feaSLen Brown 	child_pid = fork();
1566103a8feaSLen Brown 	if (!child_pid) {
1567103a8feaSLen Brown 		/* child */
1568103a8feaSLen Brown 		execvp(argv[0], argv);
1569103a8feaSLen Brown 	} else {
1570103a8feaSLen Brown 
1571103a8feaSLen Brown 		/* parent */
1572103a8feaSLen Brown 		if (child_pid == -1) {
1573103a8feaSLen Brown 			perror("fork");
1574103a8feaSLen Brown 			exit(1);
1575103a8feaSLen Brown 		}
1576103a8feaSLen Brown 
1577103a8feaSLen Brown 		signal(SIGINT, SIG_IGN);
1578103a8feaSLen Brown 		signal(SIGQUIT, SIG_IGN);
1579103a8feaSLen Brown 		if (waitpid(child_pid, &status, 0) == -1) {
1580103a8feaSLen Brown 			perror("wait");
1581d91bb17cSLen Brown 			exit(status);
1582103a8feaSLen Brown 		}
1583103a8feaSLen Brown 	}
1584c98d5d94SLen Brown 	/*
1585c98d5d94SLen Brown 	 * n.b. fork_it() does not check for errors from for_all_cpus()
1586c98d5d94SLen Brown 	 * because re-starting is problematic when forking
1587c98d5d94SLen Brown 	 */
1588c98d5d94SLen Brown 	for_all_cpus(get_counters, ODD_COUNTERS);
1589103a8feaSLen Brown 	gettimeofday(&tv_odd, (struct timezone *)NULL);
1590103a8feaSLen Brown 	timersub(&tv_odd, &tv_even, &tv_delta);
1591c98d5d94SLen Brown 	for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1592c98d5d94SLen Brown 	compute_average(EVEN_COUNTERS);
1593c98d5d94SLen Brown 	format_all_counters(EVEN_COUNTERS);
1594c98d5d94SLen Brown 	flush_stderr();
1595103a8feaSLen Brown 
15966eab04a8SJustin P. Mattock 	fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
1597103a8feaSLen Brown 
1598d91bb17cSLen Brown 	return status;
1599103a8feaSLen Brown }
1600103a8feaSLen Brown 
1601103a8feaSLen Brown void cmdline(int argc, char **argv)
1602103a8feaSLen Brown {
1603103a8feaSLen Brown 	int opt;
1604103a8feaSLen Brown 
1605103a8feaSLen Brown 	progname = argv[0];
1606103a8feaSLen Brown 
160739300ffbSLen Brown 	while ((opt = getopt(argc, argv, "+pPSvi:sc:sC:m:M:")) != -1) {
1608103a8feaSLen Brown 		switch (opt) {
1609f9240813SLen Brown 		case 'p':
1610c98d5d94SLen Brown 			show_core_only++;
1611c98d5d94SLen Brown 			break;
1612f9240813SLen Brown 		case 'P':
1613c98d5d94SLen Brown 			show_pkg_only++;
1614c98d5d94SLen Brown 			break;
1615f9240813SLen Brown 		case 'S':
1616e23da037SLen Brown 			summary_only++;
1617e23da037SLen Brown 			break;
1618103a8feaSLen Brown 		case 'v':
1619103a8feaSLen Brown 			verbose++;
1620103a8feaSLen Brown 			break;
1621103a8feaSLen Brown 		case 'i':
1622103a8feaSLen Brown 			interval_sec = atoi(optarg);
1623103a8feaSLen Brown 			break;
1624f9240813SLen Brown 		case 'c':
16258e180f3cSLen Brown 			sscanf(optarg, "%x", &extra_delta_offset32);
16268e180f3cSLen Brown 			break;
1627f9240813SLen Brown 		case 's':
1628f9240813SLen Brown 			extra_delta_offset32 = 0x34;	/* SMI counter */
1629f9240813SLen Brown 			break;
1630f9240813SLen Brown 		case 'C':
16318e180f3cSLen Brown 			sscanf(optarg, "%x", &extra_delta_offset64);
16328e180f3cSLen Brown 			break;
16332f32edf1SLen Brown 		case 'm':
16342f32edf1SLen Brown 			sscanf(optarg, "%x", &extra_msr_offset32);
16352f32edf1SLen Brown 			break;
16362f32edf1SLen Brown 		case 'M':
16372f32edf1SLen Brown 			sscanf(optarg, "%x", &extra_msr_offset64);
1638103a8feaSLen Brown 			break;
1639103a8feaSLen Brown 		default:
1640103a8feaSLen Brown 			usage();
1641103a8feaSLen Brown 		}
1642103a8feaSLen Brown 	}
1643103a8feaSLen Brown }
1644103a8feaSLen Brown 
1645103a8feaSLen Brown int main(int argc, char **argv)
1646103a8feaSLen Brown {
1647103a8feaSLen Brown 	cmdline(argc, argv);
1648103a8feaSLen Brown 
1649103a8feaSLen Brown 	if (verbose > 1)
1650f9240813SLen Brown 		fprintf(stderr, "turbostat v2.1 October 6, 2012"
1651103a8feaSLen Brown 			" - Len Brown <lenb@kernel.org>\n");
1652103a8feaSLen Brown 
1653103a8feaSLen Brown 	turbostat_init();
1654103a8feaSLen Brown 
1655103a8feaSLen Brown 	/*
1656103a8feaSLen Brown 	 * if any params left, it must be a command to fork
1657103a8feaSLen Brown 	 */
1658103a8feaSLen Brown 	if (argc - optind)
1659103a8feaSLen Brown 		return fork_it(argv + optind);
1660103a8feaSLen Brown 	else
1661103a8feaSLen Brown 		turbostat_loop();
1662103a8feaSLen Brown 
1663103a8feaSLen Brown 	return 0;
1664103a8feaSLen Brown }
1665