Merge branches 'for-3.18/always-poll-quirk', 'for-3.18/logitech', 'for-3.18/picolcd...
[linux-drm-fsl-dcu.git] / tools / power / x86 / turbostat / turbostat.c
1 /*
2  * turbostat -- show CPU frequency and C-state residency
3  * on modern Intel turbo-capable processors.
4  *
5  * Copyright (c) 2013 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 MSRHEADER
24 #include <stdarg.h>
25 #include <stdio.h>
26 #include <err.h>
27 #include <unistd.h>
28 #include <sys/types.h>
29 #include <sys/wait.h>
30 #include <sys/stat.h>
31 #include <sys/resource.h>
32 #include <fcntl.h>
33 #include <signal.h>
34 #include <sys/time.h>
35 #include <stdlib.h>
36 #include <dirent.h>
37 #include <string.h>
38 #include <ctype.h>
39 #include <sched.h>
40 #include <cpuid.h>
41
42 char *proc_stat = "/proc/stat";
43 unsigned int interval_sec = 5;  /* set with -i interval_sec */
44 unsigned int verbose;           /* set with -v */
45 unsigned int rapl_verbose;      /* set with -R */
46 unsigned int rapl_joules;       /* set with -J */
47 unsigned int thermal_verbose;   /* set with -T */
48 unsigned int summary_only;      /* set with -S */
49 unsigned int dump_only;         /* set with -s */
50 unsigned int skip_c0;
51 unsigned int skip_c1;
52 unsigned int do_nhm_cstates;
53 unsigned int do_snb_cstates;
54 unsigned int do_c8_c9_c10;
55 unsigned int do_slm_cstates;
56 unsigned int use_c1_residency_msr;
57 unsigned int has_aperf;
58 unsigned int has_epb;
59 unsigned int units = 1000000;   /* MHz etc */
60 unsigned int genuine_intel;
61 unsigned int has_invariant_tsc;
62 unsigned int do_nehalem_platform_info;
63 unsigned int do_nehalem_turbo_ratio_limit;
64 unsigned int do_ivt_turbo_ratio_limit;
65 unsigned int extra_msr_offset32;
66 unsigned int extra_msr_offset64;
67 unsigned int extra_delta_offset32;
68 unsigned int extra_delta_offset64;
69 int do_smi;
70 double bclk;
71 unsigned int show_pkg;
72 unsigned int show_core;
73 unsigned int show_cpu;
74 unsigned int show_pkg_only;
75 unsigned int show_core_only;
76 char *output_buffer, *outp;
77 unsigned int do_rapl;
78 unsigned int do_dts;
79 unsigned int do_ptm;
80 unsigned int tcc_activation_temp;
81 unsigned int tcc_activation_temp_override;
82 double rapl_power_units, rapl_energy_units, rapl_time_units;
83 double rapl_joule_counter_range;
84
85 #define RAPL_PKG                (1 << 0)
86                                         /* 0x610 MSR_PKG_POWER_LIMIT */
87                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
88 #define RAPL_PKG_PERF_STATUS    (1 << 1)
89                                         /* 0x613 MSR_PKG_PERF_STATUS */
90 #define RAPL_PKG_POWER_INFO     (1 << 2)
91                                         /* 0x614 MSR_PKG_POWER_INFO */
92
93 #define RAPL_DRAM               (1 << 3)
94                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
95                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
96                                         /* 0x61c MSR_DRAM_POWER_INFO */
97 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
98                                         /* 0x61b MSR_DRAM_PERF_STATUS */
99
100 #define RAPL_CORES              (1 << 5)
101                                         /* 0x638 MSR_PP0_POWER_LIMIT */
102                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
103 #define RAPL_CORE_POLICY        (1 << 6)
104                                         /* 0x63a MSR_PP0_POLICY */
105
106
107 #define RAPL_GFX                (1 << 7)
108                                         /* 0x640 MSR_PP1_POWER_LIMIT */
109                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
110                                         /* 0x642 MSR_PP1_POLICY */
111 #define TJMAX_DEFAULT   100
112
113 #define MAX(a, b) ((a) > (b) ? (a) : (b))
114
115 int aperf_mperf_unstable;
116 int backwards_count;
117 char *progname;
118
119 cpu_set_t *cpu_present_set, *cpu_affinity_set;
120 size_t cpu_present_setsize, cpu_affinity_setsize;
121
122 struct thread_data {
123         unsigned long long tsc;
124         unsigned long long aperf;
125         unsigned long long mperf;
126         unsigned long long c1;
127         unsigned long long extra_msr64;
128         unsigned long long extra_delta64;
129         unsigned long long extra_msr32;
130         unsigned long long extra_delta32;
131         unsigned int smi_count;
132         unsigned int cpu_id;
133         unsigned int flags;
134 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
135 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
136 } *thread_even, *thread_odd;
137
138 struct core_data {
139         unsigned long long c3;
140         unsigned long long c6;
141         unsigned long long c7;
142         unsigned int core_temp_c;
143         unsigned int core_id;
144 } *core_even, *core_odd;
145
146 struct pkg_data {
147         unsigned long long pc2;
148         unsigned long long pc3;
149         unsigned long long pc6;
150         unsigned long long pc7;
151         unsigned long long pc8;
152         unsigned long long pc9;
153         unsigned long long pc10;
154         unsigned int package_id;
155         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
156         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
157         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
158         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
159         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
160         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
161         unsigned int pkg_temp_c;
162
163 } *package_even, *package_odd;
164
165 #define ODD_COUNTERS thread_odd, core_odd, package_odd
166 #define EVEN_COUNTERS thread_even, core_even, package_even
167
168 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
169         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
170                 topo.num_threads_per_core + \
171                 (core_no) * topo.num_threads_per_core + (thread_no))
172 #define GET_CORE(core_base, core_no, pkg_no) \
173         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
174 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
175
176 struct system_summary {
177         struct thread_data threads;
178         struct core_data cores;
179         struct pkg_data packages;
180 } sum, average;
181
182
183 struct topo_params {
184         int num_packages;
185         int num_cpus;
186         int num_cores;
187         int max_cpu_num;
188         int num_cores_per_pkg;
189         int num_threads_per_core;
190 } topo;
191
192 struct timeval tv_even, tv_odd, tv_delta;
193
194 void setup_all_buffers(void);
195
196 int cpu_is_not_present(int cpu)
197 {
198         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
199 }
200 /*
201  * run func(thread, core, package) in topology order
202  * skip non-present cpus
203  */
204
205 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
206         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
207 {
208         int retval, pkg_no, core_no, thread_no;
209
210         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
211                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
212                         for (thread_no = 0; thread_no <
213                                 topo.num_threads_per_core; ++thread_no) {
214                                 struct thread_data *t;
215                                 struct core_data *c;
216                                 struct pkg_data *p;
217
218                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
219
220                                 if (cpu_is_not_present(t->cpu_id))
221                                         continue;
222
223                                 c = GET_CORE(core_base, core_no, pkg_no);
224                                 p = GET_PKG(pkg_base, pkg_no);
225
226                                 retval = func(t, c, p);
227                                 if (retval)
228                                         return retval;
229                         }
230                 }
231         }
232         return 0;
233 }
234
235 int cpu_migrate(int cpu)
236 {
237         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
238         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
239         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
240                 return -1;
241         else
242                 return 0;
243 }
244
245 int get_msr(int cpu, off_t offset, unsigned long long *msr)
246 {
247         ssize_t retval;
248         char pathname[32];
249         int fd;
250
251         sprintf(pathname, "/dev/cpu/%d/msr", cpu);
252         fd = open(pathname, O_RDONLY);
253         if (fd < 0)
254                 return -1;
255
256         retval = pread(fd, msr, sizeof *msr, offset);
257         close(fd);
258
259         if (retval != sizeof *msr) {
260                 fprintf(stderr, "%s offset 0x%llx read failed\n", pathname, (unsigned long long)offset);
261                 return -1;
262         }
263
264         return 0;
265 }
266
267 /*
268  * Example Format w/ field column widths:
269  *
270  *  Package    Core     CPU Avg_MHz Bzy_MHz TSC_MHz     SMI   %Busy CPU_%c1 CPU_%c3 CPU_%c6 CPU_%c7 CoreTmp  PkgTmp Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 PkgWatt CorWatt GFXWatt
271  * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
272  */
273
274 void print_header(void)
275 {
276         if (show_pkg)
277                 outp += sprintf(outp, " Package");
278         if (show_core)
279                 outp += sprintf(outp, "    Core");
280         if (show_cpu)
281                 outp += sprintf(outp, "     CPU");
282         if (has_aperf)
283                 outp += sprintf(outp, " Avg_MHz");
284         if (do_nhm_cstates)
285                 outp += sprintf(outp, "   %%Busy");
286         if (has_aperf)
287                 outp += sprintf(outp, " Bzy_MHz");
288         outp += sprintf(outp, " TSC_MHz");
289         if (do_smi)
290                 outp += sprintf(outp, "     SMI");
291         if (extra_delta_offset32)
292                 outp += sprintf(outp, "  count 0x%03X", extra_delta_offset32);
293         if (extra_delta_offset64)
294                 outp += sprintf(outp, "  COUNT 0x%03X", extra_delta_offset64);
295         if (extra_msr_offset32)
296                 outp += sprintf(outp, "   MSR 0x%03X", extra_msr_offset32);
297         if (extra_msr_offset64)
298                 outp += sprintf(outp, "           MSR 0x%03X", extra_msr_offset64);
299         if (do_nhm_cstates)
300                 outp += sprintf(outp, "  CPU%%c1");
301         if (do_nhm_cstates && !do_slm_cstates)
302                 outp += sprintf(outp, "  CPU%%c3");
303         if (do_nhm_cstates)
304                 outp += sprintf(outp, "  CPU%%c6");
305         if (do_snb_cstates)
306                 outp += sprintf(outp, "  CPU%%c7");
307
308         if (do_dts)
309                 outp += sprintf(outp, " CoreTmp");
310         if (do_ptm)
311                 outp += sprintf(outp, "  PkgTmp");
312
313         if (do_snb_cstates)
314                 outp += sprintf(outp, " Pkg%%pc2");
315         if (do_nhm_cstates && !do_slm_cstates)
316                 outp += sprintf(outp, " Pkg%%pc3");
317         if (do_nhm_cstates && !do_slm_cstates)
318                 outp += sprintf(outp, " Pkg%%pc6");
319         if (do_snb_cstates)
320                 outp += sprintf(outp, " Pkg%%pc7");
321         if (do_c8_c9_c10) {
322                 outp += sprintf(outp, " Pkg%%pc8");
323                 outp += sprintf(outp, " Pkg%%pc9");
324                 outp += sprintf(outp, " Pk%%pc10");
325         }
326
327         if (do_rapl && !rapl_joules) {
328                 if (do_rapl & RAPL_PKG)
329                         outp += sprintf(outp, " PkgWatt");
330                 if (do_rapl & RAPL_CORES)
331                         outp += sprintf(outp, " CorWatt");
332                 if (do_rapl & RAPL_GFX)
333                         outp += sprintf(outp, " GFXWatt");
334                 if (do_rapl & RAPL_DRAM)
335                         outp += sprintf(outp, " RAMWatt");
336                 if (do_rapl & RAPL_PKG_PERF_STATUS)
337                         outp += sprintf(outp, "   PKG_%%");
338                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
339                         outp += sprintf(outp, "   RAM_%%");
340         } else {
341                 if (do_rapl & RAPL_PKG)
342                         outp += sprintf(outp, "   Pkg_J");
343                 if (do_rapl & RAPL_CORES)
344                         outp += sprintf(outp, "   Cor_J");
345                 if (do_rapl & RAPL_GFX)
346                         outp += sprintf(outp, "   GFX_J");
347                 if (do_rapl & RAPL_DRAM)
348                         outp += sprintf(outp, "   RAM_W");
349                 if (do_rapl & RAPL_PKG_PERF_STATUS)
350                         outp += sprintf(outp, "   PKG_%%");
351                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
352                         outp += sprintf(outp, "   RAM_%%");
353                 outp += sprintf(outp, "   time");
354
355         }
356         outp += sprintf(outp, "\n");
357 }
358
359 int dump_counters(struct thread_data *t, struct core_data *c,
360         struct pkg_data *p)
361 {
362         outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
363
364         if (t) {
365                 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
366                         t->cpu_id, t->flags);
367                 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
368                 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
369                 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
370                 outp += sprintf(outp, "c1: %016llX\n", t->c1);
371                 outp += sprintf(outp, "msr0x%x: %08llX\n",
372                         extra_delta_offset32, t->extra_delta32);
373                 outp += sprintf(outp, "msr0x%x: %016llX\n",
374                         extra_delta_offset64, t->extra_delta64);
375                 outp += sprintf(outp, "msr0x%x: %08llX\n",
376                         extra_msr_offset32, t->extra_msr32);
377                 outp += sprintf(outp, "msr0x%x: %016llX\n",
378                         extra_msr_offset64, t->extra_msr64);
379                 if (do_smi)
380                         outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
381         }
382
383         if (c) {
384                 outp += sprintf(outp, "core: %d\n", c->core_id);
385                 outp += sprintf(outp, "c3: %016llX\n", c->c3);
386                 outp += sprintf(outp, "c6: %016llX\n", c->c6);
387                 outp += sprintf(outp, "c7: %016llX\n", c->c7);
388                 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
389         }
390
391         if (p) {
392                 outp += sprintf(outp, "package: %d\n", p->package_id);
393                 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
394                 outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
395                 outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
396                 outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
397                 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
398                 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
399                 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
400                 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
401                 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
402                 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
403                 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
404                 outp += sprintf(outp, "Throttle PKG: %0X\n",
405                         p->rapl_pkg_perf_status);
406                 outp += sprintf(outp, "Throttle RAM: %0X\n",
407                         p->rapl_dram_perf_status);
408                 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
409         }
410
411         outp += sprintf(outp, "\n");
412
413         return 0;
414 }
415
416 /*
417  * column formatting convention & formats
418  */
419 int format_counters(struct thread_data *t, struct core_data *c,
420         struct pkg_data *p)
421 {
422         double interval_float;
423         char *fmt8;
424
425          /* if showing only 1st thread in core and this isn't one, bail out */
426         if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
427                 return 0;
428
429          /* if showing only 1st thread in pkg and this isn't one, bail out */
430         if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
431                 return 0;
432
433         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
434
435         /* topo columns, print blanks on 1st (average) line */
436         if (t == &average.threads) {
437                 if (show_pkg)
438                         outp += sprintf(outp, "       -");
439                 if (show_core)
440                         outp += sprintf(outp, "       -");
441                 if (show_cpu)
442                         outp += sprintf(outp, "       -");
443         } else {
444                 if (show_pkg) {
445                         if (p)
446                                 outp += sprintf(outp, "%8d", p->package_id);
447                         else
448                                 outp += sprintf(outp, "       -");
449                 }
450                 if (show_core) {
451                         if (c)
452                                 outp += sprintf(outp, "%8d", c->core_id);
453                         else
454                                 outp += sprintf(outp, "       -");
455                 }
456                 if (show_cpu)
457                         outp += sprintf(outp, "%8d", t->cpu_id);
458         }
459
460         /* AvgMHz */
461         if (has_aperf)
462                 outp += sprintf(outp, "%8.0f",
463                         1.0 / units * t->aperf / interval_float);
464
465         /* %c0 */
466         if (do_nhm_cstates) {
467                 if (!skip_c0)
468                         outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc);
469                 else
470                         outp += sprintf(outp, "********");
471         }
472
473         /* BzyMHz */
474         if (has_aperf)
475                 outp += sprintf(outp, "%8.0f",
476                         1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
477
478         /* TSC */
479         outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
480
481         /* SMI */
482         if (do_smi)
483                 outp += sprintf(outp, "%8d", t->smi_count);
484
485         /* delta */
486         if (extra_delta_offset32)
487                 outp += sprintf(outp, "  %11llu", t->extra_delta32);
488
489         /* DELTA */
490         if (extra_delta_offset64)
491                 outp += sprintf(outp, "  %11llu", t->extra_delta64);
492         /* msr */
493         if (extra_msr_offset32)
494                 outp += sprintf(outp, "  0x%08llx", t->extra_msr32);
495
496         /* MSR */
497         if (extra_msr_offset64)
498                 outp += sprintf(outp, "  0x%016llx", t->extra_msr64);
499
500         if (do_nhm_cstates) {
501                 if (!skip_c1)
502                         outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
503                 else
504                         outp += sprintf(outp, "********");
505         }
506
507         /* print per-core data only for 1st thread in core */
508         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
509                 goto done;
510
511         if (do_nhm_cstates && !do_slm_cstates)
512                 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
513         if (do_nhm_cstates)
514                 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
515         if (do_snb_cstates)
516                 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
517
518         if (do_dts)
519                 outp += sprintf(outp, "%8d", c->core_temp_c);
520
521         /* print per-package data only for 1st core in package */
522         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
523                 goto done;
524
525         if (do_ptm)
526                 outp += sprintf(outp, "%8d", p->pkg_temp_c);
527
528         if (do_snb_cstates)
529                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
530         if (do_nhm_cstates && !do_slm_cstates)
531                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
532         if (do_nhm_cstates && !do_slm_cstates)
533                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
534         if (do_snb_cstates)
535                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
536         if (do_c8_c9_c10) {
537                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
538                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
539                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
540         }
541
542         /*
543          * If measurement interval exceeds minimum RAPL Joule Counter range,
544          * indicate that results are suspect by printing "**" in fraction place.
545          */
546         if (interval_float < rapl_joule_counter_range)
547                 fmt8 = "%8.2f";
548         else
549                 fmt8 = " %6.0f**";
550
551         if (do_rapl && !rapl_joules) {
552                 if (do_rapl & RAPL_PKG)
553                         outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
554                 if (do_rapl & RAPL_CORES)
555                         outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
556                 if (do_rapl & RAPL_GFX)
557                         outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
558                 if (do_rapl & RAPL_DRAM)
559                         outp += sprintf(outp, fmt8, p->energy_dram * rapl_energy_units / interval_float);
560                 if (do_rapl & RAPL_PKG_PERF_STATUS)
561                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
562                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
563                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
564         } else {
565                 if (do_rapl & RAPL_PKG)
566                         outp += sprintf(outp, fmt8,
567                                         p->energy_pkg * rapl_energy_units);
568                 if (do_rapl & RAPL_CORES)
569                         outp += sprintf(outp, fmt8,
570                                         p->energy_cores * rapl_energy_units);
571                 if (do_rapl & RAPL_GFX)
572                         outp += sprintf(outp, fmt8,
573                                         p->energy_gfx * rapl_energy_units);
574                 if (do_rapl & RAPL_DRAM)
575                         outp += sprintf(outp, fmt8,
576                                         p->energy_dram * rapl_energy_units);
577                 if (do_rapl & RAPL_PKG_PERF_STATUS)
578                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
579                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
580                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
581         outp += sprintf(outp, fmt8, interval_float);
582
583         }
584 done:
585         outp += sprintf(outp, "\n");
586
587         return 0;
588 }
589
590 void flush_stdout()
591 {
592         fputs(output_buffer, stdout);
593         fflush(stdout);
594         outp = output_buffer;
595 }
596 void flush_stderr()
597 {
598         fputs(output_buffer, stderr);
599         outp = output_buffer;
600 }
601 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
602 {
603         static int printed;
604
605         if (!printed || !summary_only)
606                 print_header();
607
608         if (topo.num_cpus > 1)
609                 format_counters(&average.threads, &average.cores,
610                         &average.packages);
611
612         printed = 1;
613
614         if (summary_only)
615                 return;
616
617         for_all_cpus(format_counters, t, c, p);
618 }
619
620 #define DELTA_WRAP32(new, old)                  \
621         if (new > old) {                        \
622                 old = new - old;                \
623         } else {                                \
624                 old = 0x100000000 + new - old;  \
625         }
626
627 void
628 delta_package(struct pkg_data *new, struct pkg_data *old)
629 {
630         old->pc2 = new->pc2 - old->pc2;
631         old->pc3 = new->pc3 - old->pc3;
632         old->pc6 = new->pc6 - old->pc6;
633         old->pc7 = new->pc7 - old->pc7;
634         old->pc8 = new->pc8 - old->pc8;
635         old->pc9 = new->pc9 - old->pc9;
636         old->pc10 = new->pc10 - old->pc10;
637         old->pkg_temp_c = new->pkg_temp_c;
638
639         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
640         DELTA_WRAP32(new->energy_cores, old->energy_cores);
641         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
642         DELTA_WRAP32(new->energy_dram, old->energy_dram);
643         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
644         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
645 }
646
647 void
648 delta_core(struct core_data *new, struct core_data *old)
649 {
650         old->c3 = new->c3 - old->c3;
651         old->c6 = new->c6 - old->c6;
652         old->c7 = new->c7 - old->c7;
653         old->core_temp_c = new->core_temp_c;
654 }
655
656 /*
657  * old = new - old
658  */
659 void
660 delta_thread(struct thread_data *new, struct thread_data *old,
661         struct core_data *core_delta)
662 {
663         old->tsc = new->tsc - old->tsc;
664
665         /* check for TSC < 1 Mcycles over interval */
666         if (old->tsc < (1000 * 1000))
667                 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
668                      "You can disable all c-states by booting with \"idle=poll\"\n"
669                      "or just the deep ones with \"processor.max_cstate=1\"");
670
671         old->c1 = new->c1 - old->c1;
672
673         if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
674                 old->aperf = new->aperf - old->aperf;
675                 old->mperf = new->mperf - old->mperf;
676         } else {
677
678                 if (!aperf_mperf_unstable) {
679                         fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
680                         fprintf(stderr, "* Frequency results do not cover entire interval *\n");
681                         fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
682
683                         aperf_mperf_unstable = 1;
684                 }
685                 /*
686                  * mperf delta is likely a huge "positive" number
687                  * can not use it for calculating c0 time
688                  */
689                 skip_c0 = 1;
690                 skip_c1 = 1;
691         }
692
693
694         if (use_c1_residency_msr) {
695                 /*
696                  * Some models have a dedicated C1 residency MSR,
697                  * which should be more accurate than the derivation below.
698                  */
699         } else {
700                 /*
701                  * As counter collection is not atomic,
702                  * it is possible for mperf's non-halted cycles + idle states
703                  * to exceed TSC's all cycles: show c1 = 0% in that case.
704                  */
705                 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
706                         old->c1 = 0;
707                 else {
708                         /* normal case, derive c1 */
709                         old->c1 = old->tsc - old->mperf - core_delta->c3
710                                 - core_delta->c6 - core_delta->c7;
711                 }
712         }
713
714         if (old->mperf == 0) {
715                 if (verbose > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
716                 old->mperf = 1; /* divide by 0 protection */
717         }
718
719         old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
720         old->extra_delta32 &= 0xFFFFFFFF;
721
722         old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
723
724         /*
725          * Extra MSR is just a snapshot, simply copy latest w/o subtracting
726          */
727         old->extra_msr32 = new->extra_msr32;
728         old->extra_msr64 = new->extra_msr64;
729
730         if (do_smi)
731                 old->smi_count = new->smi_count - old->smi_count;
732 }
733
734 int delta_cpu(struct thread_data *t, struct core_data *c,
735         struct pkg_data *p, struct thread_data *t2,
736         struct core_data *c2, struct pkg_data *p2)
737 {
738         /* calculate core delta only for 1st thread in core */
739         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
740                 delta_core(c, c2);
741
742         /* always calculate thread delta */
743         delta_thread(t, t2, c2);        /* c2 is core delta */
744
745         /* calculate package delta only for 1st core in package */
746         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
747                 delta_package(p, p2);
748
749         return 0;
750 }
751
752 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
753 {
754         t->tsc = 0;
755         t->aperf = 0;
756         t->mperf = 0;
757         t->c1 = 0;
758
759         t->smi_count = 0;
760         t->extra_delta32 = 0;
761         t->extra_delta64 = 0;
762
763         /* tells format_counters to dump all fields from this set */
764         t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
765
766         c->c3 = 0;
767         c->c6 = 0;
768         c->c7 = 0;
769         c->core_temp_c = 0;
770
771         p->pc2 = 0;
772         p->pc3 = 0;
773         p->pc6 = 0;
774         p->pc7 = 0;
775         p->pc8 = 0;
776         p->pc9 = 0;
777         p->pc10 = 0;
778
779         p->energy_pkg = 0;
780         p->energy_dram = 0;
781         p->energy_cores = 0;
782         p->energy_gfx = 0;
783         p->rapl_pkg_perf_status = 0;
784         p->rapl_dram_perf_status = 0;
785         p->pkg_temp_c = 0;
786 }
787 int sum_counters(struct thread_data *t, struct core_data *c,
788         struct pkg_data *p)
789 {
790         average.threads.tsc += t->tsc;
791         average.threads.aperf += t->aperf;
792         average.threads.mperf += t->mperf;
793         average.threads.c1 += t->c1;
794
795         average.threads.extra_delta32 += t->extra_delta32;
796         average.threads.extra_delta64 += t->extra_delta64;
797
798         /* sum per-core values only for 1st thread in core */
799         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
800                 return 0;
801
802         average.cores.c3 += c->c3;
803         average.cores.c6 += c->c6;
804         average.cores.c7 += c->c7;
805
806         average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
807
808         /* sum per-pkg values only for 1st core in pkg */
809         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
810                 return 0;
811
812         average.packages.pc2 += p->pc2;
813         average.packages.pc3 += p->pc3;
814         average.packages.pc6 += p->pc6;
815         average.packages.pc7 += p->pc7;
816         average.packages.pc8 += p->pc8;
817         average.packages.pc9 += p->pc9;
818         average.packages.pc10 += p->pc10;
819
820         average.packages.energy_pkg += p->energy_pkg;
821         average.packages.energy_dram += p->energy_dram;
822         average.packages.energy_cores += p->energy_cores;
823         average.packages.energy_gfx += p->energy_gfx;
824
825         average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
826
827         average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
828         average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
829         return 0;
830 }
831 /*
832  * sum the counters for all cpus in the system
833  * compute the weighted average
834  */
835 void compute_average(struct thread_data *t, struct core_data *c,
836         struct pkg_data *p)
837 {
838         clear_counters(&average.threads, &average.cores, &average.packages);
839
840         for_all_cpus(sum_counters, t, c, p);
841
842         average.threads.tsc /= topo.num_cpus;
843         average.threads.aperf /= topo.num_cpus;
844         average.threads.mperf /= topo.num_cpus;
845         average.threads.c1 /= topo.num_cpus;
846
847         average.threads.extra_delta32 /= topo.num_cpus;
848         average.threads.extra_delta32 &= 0xFFFFFFFF;
849
850         average.threads.extra_delta64 /= topo.num_cpus;
851
852         average.cores.c3 /= topo.num_cores;
853         average.cores.c6 /= topo.num_cores;
854         average.cores.c7 /= topo.num_cores;
855
856         average.packages.pc2 /= topo.num_packages;
857         average.packages.pc3 /= topo.num_packages;
858         average.packages.pc6 /= topo.num_packages;
859         average.packages.pc7 /= topo.num_packages;
860
861         average.packages.pc8 /= topo.num_packages;
862         average.packages.pc9 /= topo.num_packages;
863         average.packages.pc10 /= topo.num_packages;
864 }
865
866 static unsigned long long rdtsc(void)
867 {
868         unsigned int low, high;
869
870         asm volatile("rdtsc" : "=a" (low), "=d" (high));
871
872         return low | ((unsigned long long)high) << 32;
873 }
874
875
876 /*
877  * get_counters(...)
878  * migrate to cpu
879  * acquire and record local counters for that cpu
880  */
881 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
882 {
883         int cpu = t->cpu_id;
884         unsigned long long msr;
885
886         if (cpu_migrate(cpu)) {
887                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
888                 return -1;
889         }
890
891         t->tsc = rdtsc();       /* we are running on local CPU of interest */
892
893         if (has_aperf) {
894                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
895                         return -3;
896                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
897                         return -4;
898         }
899
900         if (do_smi) {
901                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
902                         return -5;
903                 t->smi_count = msr & 0xFFFFFFFF;
904         }
905         if (extra_delta_offset32) {
906                 if (get_msr(cpu, extra_delta_offset32, &msr))
907                         return -5;
908                 t->extra_delta32 = msr & 0xFFFFFFFF;
909         }
910
911         if (extra_delta_offset64)
912                 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
913                         return -5;
914
915         if (extra_msr_offset32) {
916                 if (get_msr(cpu, extra_msr_offset32, &msr))
917                         return -5;
918                 t->extra_msr32 = msr & 0xFFFFFFFF;
919         }
920
921         if (extra_msr_offset64)
922                 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
923                         return -5;
924
925         if (use_c1_residency_msr) {
926                 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
927                         return -6;
928         }
929
930         /* collect core counters only for 1st thread in core */
931         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
932                 return 0;
933
934         if (do_nhm_cstates && !do_slm_cstates) {
935                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
936                         return -6;
937         }
938
939         if (do_nhm_cstates) {
940                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
941                         return -7;
942         }
943
944         if (do_snb_cstates)
945                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
946                         return -8;
947
948         if (do_dts) {
949                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
950                         return -9;
951                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
952         }
953
954
955         /* collect package counters only for 1st core in package */
956         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
957                 return 0;
958
959         if (do_nhm_cstates && !do_slm_cstates) {
960                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
961                         return -9;
962                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
963                         return -10;
964         }
965         if (do_snb_cstates) {
966                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
967                         return -11;
968                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
969                         return -12;
970         }
971         if (do_c8_c9_c10) {
972                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
973                         return -13;
974                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
975                         return -13;
976                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
977                         return -13;
978         }
979         if (do_rapl & RAPL_PKG) {
980                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
981                         return -13;
982                 p->energy_pkg = msr & 0xFFFFFFFF;
983         }
984         if (do_rapl & RAPL_CORES) {
985                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
986                         return -14;
987                 p->energy_cores = msr & 0xFFFFFFFF;
988         }
989         if (do_rapl & RAPL_DRAM) {
990                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
991                         return -15;
992                 p->energy_dram = msr & 0xFFFFFFFF;
993         }
994         if (do_rapl & RAPL_GFX) {
995                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
996                         return -16;
997                 p->energy_gfx = msr & 0xFFFFFFFF;
998         }
999         if (do_rapl & RAPL_PKG_PERF_STATUS) {
1000                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1001                         return -16;
1002                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1003         }
1004         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1005                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1006                         return -16;
1007                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1008         }
1009         if (do_ptm) {
1010                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1011                         return -17;
1012                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1013         }
1014         return 0;
1015 }
1016
1017 void print_verbose_header(void)
1018 {
1019         unsigned long long msr;
1020         unsigned int ratio;
1021
1022         if (!do_nehalem_platform_info)
1023                 return;
1024
1025         get_msr(0, MSR_NHM_PLATFORM_INFO, &msr);
1026
1027         fprintf(stderr, "cpu0: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", msr);
1028
1029         ratio = (msr >> 40) & 0xFF;
1030         fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency\n",
1031                 ratio, bclk, ratio * bclk);
1032
1033         ratio = (msr >> 8) & 0xFF;
1034         fprintf(stderr, "%d * %.0f = %.0f MHz TSC frequency\n",
1035                 ratio, bclk, ratio * bclk);
1036
1037         get_msr(0, MSR_IA32_POWER_CTL, &msr);
1038         fprintf(stderr, "cpu0: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1039                 msr, msr & 0x2 ? "EN" : "DIS");
1040
1041         if (!do_ivt_turbo_ratio_limit)
1042                 goto print_nhm_turbo_ratio_limits;
1043
1044         get_msr(0, MSR_IVT_TURBO_RATIO_LIMIT, &msr);
1045
1046         fprintf(stderr, "cpu0: MSR_IVT_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1047
1048         ratio = (msr >> 56) & 0xFF;
1049         if (ratio)
1050                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1051                         ratio, bclk, ratio * bclk);
1052
1053         ratio = (msr >> 48) & 0xFF;
1054         if (ratio)
1055                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1056                         ratio, bclk, ratio * bclk);
1057
1058         ratio = (msr >> 40) & 0xFF;
1059         if (ratio)
1060                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1061                         ratio, bclk, ratio * bclk);
1062
1063         ratio = (msr >> 32) & 0xFF;
1064         if (ratio)
1065                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1066                         ratio, bclk, ratio * bclk);
1067
1068         ratio = (msr >> 24) & 0xFF;
1069         if (ratio)
1070                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1071                         ratio, bclk, ratio * bclk);
1072
1073         ratio = (msr >> 16) & 0xFF;
1074         if (ratio)
1075                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1076                         ratio, bclk, ratio * bclk);
1077
1078         ratio = (msr >> 8) & 0xFF;
1079         if (ratio)
1080                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1081                         ratio, bclk, ratio * bclk);
1082
1083         ratio = (msr >> 0) & 0xFF;
1084         if (ratio)
1085                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1086                         ratio, bclk, ratio * bclk);
1087
1088 print_nhm_turbo_ratio_limits:
1089         get_msr(0, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1090
1091 #define SNB_C1_AUTO_UNDEMOTE              (1UL << 27)
1092 #define SNB_C3_AUTO_UNDEMOTE              (1UL << 28)
1093
1094         fprintf(stderr, "cpu0: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", msr);
1095
1096         fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: ",
1097                 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1098                 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1099                 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1100                 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1101                 (msr & (1 << 15)) ? "" : "UN",
1102                 (unsigned int)msr & 7);
1103
1104
1105         switch(msr & 0x7) {
1106         case 0:
1107                 fprintf(stderr, do_slm_cstates ? "no pkg states" : "pc0");
1108                 break;
1109         case 1:
1110                 fprintf(stderr, do_slm_cstates ? "no pkg states" : do_snb_cstates ? "pc2" : "pc0");
1111                 break;
1112         case 2:
1113                 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc6-noret" : "pc3");
1114                 break;
1115         case 3:
1116                 fprintf(stderr, do_slm_cstates ? "invalid" : "pc6");
1117                 break;
1118         case 4:
1119                 fprintf(stderr, do_slm_cstates ? "pc4" : "pc7");
1120                 break;
1121         case 5:
1122                 fprintf(stderr, do_slm_cstates ? "invalid" : do_snb_cstates ? "pc7s" : "invalid");
1123                 break;
1124         case 6:
1125                 fprintf(stderr, do_slm_cstates ? "pc6" : "invalid");
1126                 break;
1127         case 7:
1128                 fprintf(stderr, do_slm_cstates ? "pc7" : "unlimited");
1129                 break;
1130         default:
1131                 fprintf(stderr, "invalid");
1132         }
1133         fprintf(stderr, ")\n");
1134
1135         if (!do_nehalem_turbo_ratio_limit)
1136                 return;
1137
1138         get_msr(0, MSR_NHM_TURBO_RATIO_LIMIT, &msr);
1139
1140         fprintf(stderr, "cpu0: MSR_NHM_TURBO_RATIO_LIMIT: 0x%08llx\n", msr);
1141
1142         ratio = (msr >> 56) & 0xFF;
1143         if (ratio)
1144                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1145                         ratio, bclk, ratio * bclk);
1146
1147         ratio = (msr >> 48) & 0xFF;
1148         if (ratio)
1149                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1150                         ratio, bclk, ratio * bclk);
1151
1152         ratio = (msr >> 40) & 0xFF;
1153         if (ratio)
1154                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1155                         ratio, bclk, ratio * bclk);
1156
1157         ratio = (msr >> 32) & 0xFF;
1158         if (ratio)
1159                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1160                         ratio, bclk, ratio * bclk);
1161
1162         ratio = (msr >> 24) & 0xFF;
1163         if (ratio)
1164                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1165                         ratio, bclk, ratio * bclk);
1166
1167         ratio = (msr >> 16) & 0xFF;
1168         if (ratio)
1169                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1170                         ratio, bclk, ratio * bclk);
1171
1172         ratio = (msr >> 8) & 0xFF;
1173         if (ratio)
1174                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1175                         ratio, bclk, ratio * bclk);
1176
1177         ratio = (msr >> 0) & 0xFF;
1178         if (ratio)
1179                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1180                         ratio, bclk, ratio * bclk);
1181 }
1182
1183 void free_all_buffers(void)
1184 {
1185         CPU_FREE(cpu_present_set);
1186         cpu_present_set = NULL;
1187         cpu_present_set = 0;
1188
1189         CPU_FREE(cpu_affinity_set);
1190         cpu_affinity_set = NULL;
1191         cpu_affinity_setsize = 0;
1192
1193         free(thread_even);
1194         free(core_even);
1195         free(package_even);
1196
1197         thread_even = NULL;
1198         core_even = NULL;
1199         package_even = NULL;
1200
1201         free(thread_odd);
1202         free(core_odd);
1203         free(package_odd);
1204
1205         thread_odd = NULL;
1206         core_odd = NULL;
1207         package_odd = NULL;
1208
1209         free(output_buffer);
1210         output_buffer = NULL;
1211         outp = NULL;
1212 }
1213
1214 /*
1215  * Open a file, and exit on failure
1216  */
1217 FILE *fopen_or_die(const char *path, const char *mode)
1218 {
1219         FILE *filep = fopen(path, "r");
1220         if (!filep)
1221                 err(1, "%s: open failed", path);
1222         return filep;
1223 }
1224
1225 /*
1226  * Parse a file containing a single int.
1227  */
1228 int parse_int_file(const char *fmt, ...)
1229 {
1230         va_list args;
1231         char path[PATH_MAX];
1232         FILE *filep;
1233         int value;
1234
1235         va_start(args, fmt);
1236         vsnprintf(path, sizeof(path), fmt, args);
1237         va_end(args);
1238         filep = fopen_or_die(path, "r");
1239         if (fscanf(filep, "%d", &value) != 1)
1240                 err(1, "%s: failed to parse number from file", path);
1241         fclose(filep);
1242         return value;
1243 }
1244
1245 /*
1246  * cpu_is_first_sibling_in_core(cpu)
1247  * return 1 if given CPU is 1st HT sibling in the core
1248  */
1249 int cpu_is_first_sibling_in_core(int cpu)
1250 {
1251         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1252 }
1253
1254 /*
1255  * cpu_is_first_core_in_package(cpu)
1256  * return 1 if given CPU is 1st core in package
1257  */
1258 int cpu_is_first_core_in_package(int cpu)
1259 {
1260         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1261 }
1262
1263 int get_physical_package_id(int cpu)
1264 {
1265         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1266 }
1267
1268 int get_core_id(int cpu)
1269 {
1270         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1271 }
1272
1273 int get_num_ht_siblings(int cpu)
1274 {
1275         char path[80];
1276         FILE *filep;
1277         int sib1, sib2;
1278         int matches;
1279         char character;
1280
1281         sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1282         filep = fopen_or_die(path, "r");
1283         /*
1284          * file format:
1285          * if a pair of number with a character between: 2 siblings (eg. 1-2, or 1,4)
1286          * otherwinse 1 sibling (self).
1287          */
1288         matches = fscanf(filep, "%d%c%d\n", &sib1, &character, &sib2);
1289
1290         fclose(filep);
1291
1292         if (matches == 3)
1293                 return 2;
1294         else
1295                 return 1;
1296 }
1297
1298 /*
1299  * run func(thread, core, package) in topology order
1300  * skip non-present cpus
1301  */
1302
1303 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1304         struct pkg_data *, struct thread_data *, struct core_data *,
1305         struct pkg_data *), struct thread_data *thread_base,
1306         struct core_data *core_base, struct pkg_data *pkg_base,
1307         struct thread_data *thread_base2, struct core_data *core_base2,
1308         struct pkg_data *pkg_base2)
1309 {
1310         int retval, pkg_no, core_no, thread_no;
1311
1312         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1313                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1314                         for (thread_no = 0; thread_no <
1315                                 topo.num_threads_per_core; ++thread_no) {
1316                                 struct thread_data *t, *t2;
1317                                 struct core_data *c, *c2;
1318                                 struct pkg_data *p, *p2;
1319
1320                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1321
1322                                 if (cpu_is_not_present(t->cpu_id))
1323                                         continue;
1324
1325                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1326
1327                                 c = GET_CORE(core_base, core_no, pkg_no);
1328                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
1329
1330                                 p = GET_PKG(pkg_base, pkg_no);
1331                                 p2 = GET_PKG(pkg_base2, pkg_no);
1332
1333                                 retval = func(t, c, p, t2, c2, p2);
1334                                 if (retval)
1335                                         return retval;
1336                         }
1337                 }
1338         }
1339         return 0;
1340 }
1341
1342 /*
1343  * run func(cpu) on every cpu in /proc/stat
1344  * return max_cpu number
1345  */
1346 int for_all_proc_cpus(int (func)(int))
1347 {
1348         FILE *fp;
1349         int cpu_num;
1350         int retval;
1351
1352         fp = fopen_or_die(proc_stat, "r");
1353
1354         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1355         if (retval != 0)
1356                 err(1, "%s: failed to parse format", proc_stat);
1357
1358         while (1) {
1359                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1360                 if (retval != 1)
1361                         break;
1362
1363                 retval = func(cpu_num);
1364                 if (retval) {
1365                         fclose(fp);
1366                         return(retval);
1367                 }
1368         }
1369         fclose(fp);
1370         return 0;
1371 }
1372
1373 void re_initialize(void)
1374 {
1375         free_all_buffers();
1376         setup_all_buffers();
1377         printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1378 }
1379
1380
1381 /*
1382  * count_cpus()
1383  * remember the last one seen, it will be the max
1384  */
1385 int count_cpus(int cpu)
1386 {
1387         if (topo.max_cpu_num < cpu)
1388                 topo.max_cpu_num = cpu;
1389
1390         topo.num_cpus += 1;
1391         return 0;
1392 }
1393 int mark_cpu_present(int cpu)
1394 {
1395         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1396         return 0;
1397 }
1398
1399 void turbostat_loop()
1400 {
1401         int retval;
1402         int restarted = 0;
1403
1404 restart:
1405         restarted++;
1406
1407         retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1408         if (retval < -1) {
1409                 exit(retval);
1410         } else if (retval == -1) {
1411                 if (restarted > 1) {
1412                         exit(retval);
1413                 }
1414                 re_initialize();
1415                 goto restart;
1416         }
1417         restarted = 0;
1418         gettimeofday(&tv_even, (struct timezone *)NULL);
1419
1420         while (1) {
1421                 if (for_all_proc_cpus(cpu_is_not_present)) {
1422                         re_initialize();
1423                         goto restart;
1424                 }
1425                 sleep(interval_sec);
1426                 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1427                 if (retval < -1) {
1428                         exit(retval);
1429                 } else if (retval == -1) {
1430                         re_initialize();
1431                         goto restart;
1432                 }
1433                 gettimeofday(&tv_odd, (struct timezone *)NULL);
1434                 timersub(&tv_odd, &tv_even, &tv_delta);
1435                 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1436                 compute_average(EVEN_COUNTERS);
1437                 format_all_counters(EVEN_COUNTERS);
1438                 flush_stdout();
1439                 sleep(interval_sec);
1440                 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1441                 if (retval < -1) {
1442                         exit(retval);
1443                 } else if (retval == -1) {
1444                         re_initialize();
1445                         goto restart;
1446                 }
1447                 gettimeofday(&tv_even, (struct timezone *)NULL);
1448                 timersub(&tv_even, &tv_odd, &tv_delta);
1449                 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1450                 compute_average(ODD_COUNTERS);
1451                 format_all_counters(ODD_COUNTERS);
1452                 flush_stdout();
1453         }
1454 }
1455
1456 void check_dev_msr()
1457 {
1458         struct stat sb;
1459
1460         if (stat("/dev/cpu/0/msr", &sb))
1461                 err(-5, "no /dev/cpu/0/msr\n"
1462                     "Try \"# modprobe msr\"");
1463 }
1464
1465 void check_super_user()
1466 {
1467         if (getuid() != 0)
1468                 errx(-6, "must be root");
1469 }
1470
1471 int has_nehalem_turbo_ratio_limit(unsigned int family, unsigned int model)
1472 {
1473         if (!genuine_intel)
1474                 return 0;
1475
1476         if (family != 6)
1477                 return 0;
1478
1479         switch (model) {
1480         case 0x1A:      /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1481         case 0x1E:      /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1482         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
1483         case 0x25:      /* Westmere Client - Clarkdale, Arrandale */
1484         case 0x2C:      /* Westmere EP - Gulftown */
1485         case 0x2A:      /* SNB */
1486         case 0x2D:      /* SNB Xeon */
1487         case 0x3A:      /* IVB */
1488         case 0x3E:      /* IVB Xeon */
1489         case 0x3C:      /* HSW */
1490         case 0x3F:      /* HSX */
1491         case 0x45:      /* HSW */
1492         case 0x46:      /* HSW */
1493         case 0x37:      /* BYT */
1494         case 0x4D:      /* AVN */
1495         case 0x3D:      /* BDW */
1496         case 0x4F:      /* BDX */
1497         case 0x56:      /* BDX-DE */
1498                 return 1;
1499         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1500         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1501         default:
1502                 return 0;
1503         }
1504 }
1505 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1506 {
1507         if (!genuine_intel)
1508                 return 0;
1509
1510         if (family != 6)
1511                 return 0;
1512
1513         switch (model) {
1514         case 0x3E:      /* IVB Xeon */
1515                 return 1;
1516         default:
1517                 return 0;
1518         }
1519 }
1520
1521 /*
1522  * print_epb()
1523  * Decode the ENERGY_PERF_BIAS MSR
1524  */
1525 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1526 {
1527         unsigned long long msr;
1528         char *epb_string;
1529         int cpu;
1530
1531         if (!has_epb)
1532                 return 0;
1533
1534         cpu = t->cpu_id;
1535
1536         /* EPB is per-package */
1537         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1538                 return 0;
1539
1540         if (cpu_migrate(cpu)) {
1541                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1542                 return -1;
1543         }
1544
1545         if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
1546                 return 0;
1547
1548         switch (msr & 0x7) {
1549         case ENERGY_PERF_BIAS_PERFORMANCE:
1550                 epb_string = "performance";
1551                 break;
1552         case ENERGY_PERF_BIAS_NORMAL:
1553                 epb_string = "balanced";
1554                 break;
1555         case ENERGY_PERF_BIAS_POWERSAVE:
1556                 epb_string = "powersave";
1557                 break;
1558         default:
1559                 epb_string = "custom";
1560                 break;
1561         }
1562         fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
1563
1564         return 0;
1565 }
1566
1567 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
1568 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
1569
1570 double get_tdp(model)
1571 {
1572         unsigned long long msr;
1573
1574         if (do_rapl & RAPL_PKG_POWER_INFO)
1575                 if (!get_msr(0, MSR_PKG_POWER_INFO, &msr))
1576                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
1577
1578         switch (model) {
1579         case 0x37:
1580         case 0x4D:
1581                 return 30.0;
1582         default:
1583                 return 135.0;
1584         }
1585 }
1586
1587
1588 /*
1589  * rapl_probe()
1590  *
1591  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
1592  */
1593 void rapl_probe(unsigned int family, unsigned int model)
1594 {
1595         unsigned long long msr;
1596         unsigned int time_unit;
1597         double tdp;
1598
1599         if (!genuine_intel)
1600                 return;
1601
1602         if (family != 6)
1603                 return;
1604
1605         switch (model) {
1606         case 0x2A:
1607         case 0x3A:
1608         case 0x3C:      /* HSW */
1609         case 0x45:      /* HSW */
1610         case 0x46:      /* HSW */
1611         case 0x3D:      /* BDW */
1612                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
1613                 break;
1614         case 0x3F:      /* HSX */
1615         case 0x4F:      /* BDX */
1616         case 0x56:      /* BDX-DE */
1617                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
1618                 break;
1619         case 0x2D:
1620         case 0x3E:
1621                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
1622                 break;
1623         case 0x37:      /* BYT */
1624         case 0x4D:      /* AVN */
1625                 do_rapl = RAPL_PKG | RAPL_CORES ;
1626                 break;
1627         default:
1628                 return;
1629         }
1630
1631         /* units on package 0, verify later other packages match */
1632         if (get_msr(0, MSR_RAPL_POWER_UNIT, &msr))
1633                 return;
1634
1635         rapl_power_units = 1.0 / (1 << (msr & 0xF));
1636         if (model == 0x37)
1637                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
1638         else
1639                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
1640
1641         time_unit = msr >> 16 & 0xF;
1642         if (time_unit == 0)
1643                 time_unit = 0xA;
1644
1645         rapl_time_units = 1.0 / (1 << (time_unit));
1646
1647         tdp = get_tdp(model);
1648
1649         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
1650         if (verbose)
1651                 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
1652
1653         return;
1654 }
1655
1656 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1657 {
1658         unsigned long long msr;
1659         unsigned int dts;
1660         int cpu;
1661
1662         if (!(do_dts || do_ptm))
1663                 return 0;
1664
1665         cpu = t->cpu_id;
1666
1667         /* DTS is per-core, no need to print for each thread */
1668         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 
1669                 return 0;
1670
1671         if (cpu_migrate(cpu)) {
1672                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1673                 return -1;
1674         }
1675
1676         if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
1677                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1678                         return 0;
1679
1680                 dts = (msr >> 16) & 0x7F;
1681                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
1682                         cpu, msr, tcc_activation_temp - dts);
1683
1684 #ifdef  THERM_DEBUG
1685                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
1686                         return 0;
1687
1688                 dts = (msr >> 16) & 0x7F;
1689                 dts2 = (msr >> 8) & 0x7F;
1690                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1691                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1692 #endif
1693         }
1694
1695
1696         if (do_dts) {
1697                 unsigned int resolution;
1698
1699                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1700                         return 0;
1701
1702                 dts = (msr >> 16) & 0x7F;
1703                 resolution = (msr >> 27) & 0xF;
1704                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
1705                         cpu, msr, tcc_activation_temp - dts, resolution);
1706
1707 #ifdef THERM_DEBUG
1708                 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
1709                         return 0;
1710
1711                 dts = (msr >> 16) & 0x7F;
1712                 dts2 = (msr >> 8) & 0x7F;
1713                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
1714                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
1715 #endif
1716         }
1717
1718         return 0;
1719 }
1720         
1721 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
1722 {
1723         fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
1724                 cpu, label,
1725                 ((msr >> 15) & 1) ? "EN" : "DIS",
1726                 ((msr >> 0) & 0x7FFF) * rapl_power_units,
1727                 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
1728                 (((msr >> 16) & 1) ? "EN" : "DIS"));
1729
1730         return;
1731 }
1732
1733 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1734 {
1735         unsigned long long msr;
1736         int cpu;
1737
1738         if (!do_rapl)
1739                 return 0;
1740
1741         /* RAPL counters are per package, so print only for 1st thread/package */
1742         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1743                 return 0;
1744
1745         cpu = t->cpu_id;
1746         if (cpu_migrate(cpu)) {
1747                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1748                 return -1;
1749         }
1750
1751         if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
1752                 return -1;
1753
1754         if (verbose) {
1755                 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
1756                         "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
1757                         rapl_power_units, rapl_energy_units, rapl_time_units);
1758         }
1759         if (do_rapl & RAPL_PKG_POWER_INFO) {
1760
1761                 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
1762                         return -5;
1763
1764
1765                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1766                         cpu, msr,
1767                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1768                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1769                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1770                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1771
1772         }
1773         if (do_rapl & RAPL_PKG) {
1774
1775                 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
1776                         return -9;
1777
1778                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
1779                         cpu, msr, (msr >> 63) & 1 ? "": "UN");
1780
1781                 print_power_limit_msr(cpu, msr, "PKG Limit #1");
1782                 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
1783                         cpu,
1784                         ((msr >> 47) & 1) ? "EN" : "DIS",
1785                         ((msr >> 32) & 0x7FFF) * rapl_power_units,
1786                         (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
1787                         ((msr >> 48) & 1) ? "EN" : "DIS");
1788         }
1789
1790         if (do_rapl & RAPL_DRAM) {
1791                 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
1792                         return -6;
1793
1794
1795                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
1796                         cpu, msr,
1797                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1798                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1799                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
1800                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
1801
1802
1803                 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
1804                         return -9;
1805                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
1806                                 cpu, msr, (msr >> 31) & 1 ? "": "UN");
1807
1808                 print_power_limit_msr(cpu, msr, "DRAM Limit");
1809         }
1810         if (do_rapl & RAPL_CORE_POLICY) {
1811                 if (verbose) {
1812                         if (get_msr(cpu, MSR_PP0_POLICY, &msr))
1813                                 return -7;
1814
1815                         fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
1816                 }
1817         }
1818         if (do_rapl & RAPL_CORES) {
1819                 if (verbose) {
1820
1821                         if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
1822                                 return -9;
1823                         fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
1824                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
1825                         print_power_limit_msr(cpu, msr, "Cores Limit");
1826                 }
1827         }
1828         if (do_rapl & RAPL_GFX) {
1829                 if (verbose) {
1830                         if (get_msr(cpu, MSR_PP1_POLICY, &msr))
1831                                 return -8;
1832
1833                         fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
1834
1835                         if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
1836                                 return -9;
1837                         fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
1838                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
1839                         print_power_limit_msr(cpu, msr, "GFX Limit");
1840                 }
1841         }
1842         return 0;
1843 }
1844
1845
1846 int is_snb(unsigned int family, unsigned int model)
1847 {
1848         if (!genuine_intel)
1849                 return 0;
1850
1851         switch (model) {
1852         case 0x2A:
1853         case 0x2D:
1854         case 0x3A:      /* IVB */
1855         case 0x3E:      /* IVB Xeon */
1856         case 0x3C:      /* HSW */
1857         case 0x3F:      /* HSW */
1858         case 0x45:      /* HSW */
1859         case 0x46:      /* HSW */
1860         case 0x3D:      /* BDW */
1861         case 0x4F:      /* BDX */
1862         case 0x56:      /* BDX-DE */
1863                 return 1;
1864         }
1865         return 0;
1866 }
1867
1868 int has_c8_c9_c10(unsigned int family, unsigned int model)
1869 {
1870         if (!genuine_intel)
1871                 return 0;
1872
1873         switch (model) {
1874         case 0x45:      /* HSW */
1875         case 0x3D:      /* BDW */
1876                 return 1;
1877         }
1878         return 0;
1879 }
1880
1881
1882 int is_slm(unsigned int family, unsigned int model)
1883 {
1884         if (!genuine_intel)
1885                 return 0;
1886         switch (model) {
1887         case 0x37:      /* BYT */
1888         case 0x4D:      /* AVN */
1889                 return 1;
1890         }
1891         return 0;
1892 }
1893
1894 #define SLM_BCLK_FREQS 5
1895 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
1896
1897 double slm_bclk(void)
1898 {
1899         unsigned long long msr = 3;
1900         unsigned int i;
1901         double freq;
1902
1903         if (get_msr(0, MSR_FSB_FREQ, &msr))
1904                 fprintf(stderr, "SLM BCLK: unknown\n");
1905
1906         i = msr & 0xf;
1907         if (i >= SLM_BCLK_FREQS) {
1908                 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
1909                 msr = 3;
1910         }
1911         freq = slm_freq_table[i];
1912
1913         fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
1914
1915         return freq;
1916 }
1917
1918 double discover_bclk(unsigned int family, unsigned int model)
1919 {
1920         if (is_snb(family, model))
1921                 return 100.00;
1922         else if (is_slm(family, model))
1923                 return slm_bclk();
1924         else
1925                 return 133.33;
1926 }
1927
1928 /*
1929  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
1930  * the Thermal Control Circuit (TCC) activates.
1931  * This is usually equal to tjMax.
1932  *
1933  * Older processors do not have this MSR, so there we guess,
1934  * but also allow cmdline over-ride with -T.
1935  *
1936  * Several MSR temperature values are in units of degrees-C
1937  * below this value, including the Digital Thermal Sensor (DTS),
1938  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
1939  */
1940 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1941 {
1942         unsigned long long msr;
1943         unsigned int target_c_local;
1944         int cpu;
1945
1946         /* tcc_activation_temp is used only for dts or ptm */
1947         if (!(do_dts || do_ptm))
1948                 return 0;
1949
1950         /* this is a per-package concept */
1951         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1952                 return 0;
1953
1954         cpu = t->cpu_id;
1955         if (cpu_migrate(cpu)) {
1956                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1957                 return -1;
1958         }
1959
1960         if (tcc_activation_temp_override != 0) {
1961                 tcc_activation_temp = tcc_activation_temp_override;
1962                 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
1963                         cpu, tcc_activation_temp);
1964                 return 0;
1965         }
1966
1967         /* Temperature Target MSR is Nehalem and newer only */
1968         if (!do_nehalem_platform_info)
1969                 goto guess;
1970
1971         if (get_msr(0, MSR_IA32_TEMPERATURE_TARGET, &msr))
1972                 goto guess;
1973
1974         target_c_local = (msr >> 16) & 0xFF;
1975
1976         if (verbose)
1977                 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
1978                         cpu, msr, target_c_local);
1979
1980         if (!target_c_local)
1981                 goto guess;
1982
1983         tcc_activation_temp = target_c_local;
1984
1985         return 0;
1986
1987 guess:
1988         tcc_activation_temp = TJMAX_DEFAULT;
1989         fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
1990                 cpu, tcc_activation_temp);
1991
1992         return 0;
1993 }
1994 void check_cpuid()
1995 {
1996         unsigned int eax, ebx, ecx, edx, max_level;
1997         unsigned int fms, family, model, stepping;
1998
1999         eax = ebx = ecx = edx = 0;
2000
2001         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2002
2003         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2004                 genuine_intel = 1;
2005
2006         if (verbose)
2007                 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2008                         (char *)&ebx, (char *)&edx, (char *)&ecx);
2009
2010         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2011         family = (fms >> 8) & 0xf;
2012         model = (fms >> 4) & 0xf;
2013         stepping = fms & 0xf;
2014         if (family == 6 || family == 0xf)
2015                 model += ((fms >> 16) & 0xf) << 4;
2016
2017         if (verbose)
2018                 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2019                         max_level, family, model, stepping, family, model, stepping);
2020
2021         if (!(edx & (1 << 5)))
2022                 errx(1, "CPUID: no MSR");
2023
2024         /*
2025          * check max extended function levels of CPUID.
2026          * This is needed to check for invariant TSC.
2027          * This check is valid for both Intel and AMD.
2028          */
2029         ebx = ecx = edx = 0;
2030         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2031
2032         if (max_level < 0x80000007)
2033                 errx(1, "CPUID: no invariant TSC (max_level 0x%x)", max_level);
2034
2035         /*
2036          * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2037          * this check is valid for both Intel and AMD
2038          */
2039         __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2040         has_invariant_tsc = edx & (1 << 8);
2041
2042         if (!has_invariant_tsc)
2043                 errx(1, "No invariant TSC");
2044
2045         /*
2046          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2047          * this check is valid for both Intel and AMD
2048          */
2049
2050         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2051         has_aperf = ecx & (1 << 0);
2052         do_dts = eax & (1 << 0);
2053         do_ptm = eax & (1 << 6);
2054         has_epb = ecx & (1 << 3);
2055
2056         if (verbose)
2057                 fprintf(stderr, "CPUID(6): %s%s%s%s\n",
2058                         has_aperf ? "APERF" : "No APERF!",
2059                         do_dts ? ", DTS" : "",
2060                         do_ptm ? ", PTM": "",
2061                         has_epb ? ", EPB": "");
2062
2063         if (!has_aperf)
2064                 errx(-1, "No APERF");
2065
2066         do_nehalem_platform_info = genuine_intel && has_invariant_tsc;
2067         do_nhm_cstates = genuine_intel; /* all Intel w/ non-stop TSC have NHM counters */
2068         do_smi = do_nhm_cstates;
2069         do_snb_cstates = is_snb(family, model);
2070         do_c8_c9_c10 = has_c8_c9_c10(family, model);
2071         do_slm_cstates = is_slm(family, model);
2072         bclk = discover_bclk(family, model);
2073
2074         do_nehalem_turbo_ratio_limit = has_nehalem_turbo_ratio_limit(family, model);
2075         do_ivt_turbo_ratio_limit = has_ivt_turbo_ratio_limit(family, model);
2076         rapl_probe(family, model);
2077
2078         return;
2079 }
2080
2081
2082 void usage()
2083 {
2084         errx(1, "%s: [-v][-R][-T][-p|-P|-S][-c MSR#][-C MSR#][-m MSR#][-M MSR#][-i interval_sec | command ...]\n",
2085              progname);
2086 }
2087
2088
2089 /*
2090  * in /dev/cpu/ return success for names that are numbers
2091  * ie. filter out ".", "..", "microcode".
2092  */
2093 int dir_filter(const struct dirent *dirp)
2094 {
2095         if (isdigit(dirp->d_name[0]))
2096                 return 1;
2097         else
2098                 return 0;
2099 }
2100
2101 int open_dev_cpu_msr(int dummy1)
2102 {
2103         return 0;
2104 }
2105
2106 void topology_probe()
2107 {
2108         int i;
2109         int max_core_id = 0;
2110         int max_package_id = 0;
2111         int max_siblings = 0;
2112         struct cpu_topology {
2113                 int core_id;
2114                 int physical_package_id;
2115         } *cpus;
2116
2117         /* Initialize num_cpus, max_cpu_num */
2118         topo.num_cpus = 0;
2119         topo.max_cpu_num = 0;
2120         for_all_proc_cpus(count_cpus);
2121         if (!summary_only && topo.num_cpus > 1)
2122                 show_cpu = 1;
2123
2124         if (verbose > 1)
2125                 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2126
2127         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
2128         if (cpus == NULL)
2129                 err(1, "calloc cpus");
2130
2131         /*
2132          * Allocate and initialize cpu_present_set
2133          */
2134         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2135         if (cpu_present_set == NULL)
2136                 err(3, "CPU_ALLOC");
2137         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2138         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2139         for_all_proc_cpus(mark_cpu_present);
2140
2141         /*
2142          * Allocate and initialize cpu_affinity_set
2143          */
2144         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2145         if (cpu_affinity_set == NULL)
2146                 err(3, "CPU_ALLOC");
2147         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2148         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2149
2150
2151         /*
2152          * For online cpus
2153          * find max_core_id, max_package_id
2154          */
2155         for (i = 0; i <= topo.max_cpu_num; ++i) {
2156                 int siblings;
2157
2158                 if (cpu_is_not_present(i)) {
2159                         if (verbose > 1)
2160                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2161                         continue;
2162                 }
2163                 cpus[i].core_id = get_core_id(i);
2164                 if (cpus[i].core_id > max_core_id)
2165                         max_core_id = cpus[i].core_id;
2166
2167                 cpus[i].physical_package_id = get_physical_package_id(i);
2168                 if (cpus[i].physical_package_id > max_package_id)
2169                         max_package_id = cpus[i].physical_package_id;
2170
2171                 siblings = get_num_ht_siblings(i);
2172                 if (siblings > max_siblings)
2173                         max_siblings = siblings;
2174                 if (verbose > 1)
2175                         fprintf(stderr, "cpu %d pkg %d core %d\n",
2176                                 i, cpus[i].physical_package_id, cpus[i].core_id);
2177         }
2178         topo.num_cores_per_pkg = max_core_id + 1;
2179         if (verbose > 1)
2180                 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2181                         max_core_id, topo.num_cores_per_pkg);
2182         if (!summary_only && topo.num_cores_per_pkg > 1)
2183                 show_core = 1;
2184
2185         topo.num_packages = max_package_id + 1;
2186         if (verbose > 1)
2187                 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2188                         max_package_id, topo.num_packages);
2189         if (!summary_only && topo.num_packages > 1)
2190                 show_pkg = 1;
2191
2192         topo.num_threads_per_core = max_siblings;
2193         if (verbose > 1)
2194                 fprintf(stderr, "max_siblings %d\n", max_siblings);
2195
2196         free(cpus);
2197 }
2198
2199 void
2200 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2201 {
2202         int i;
2203
2204         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2205                 topo.num_packages, sizeof(struct thread_data));
2206         if (*t == NULL)
2207                 goto error;
2208
2209         for (i = 0; i < topo.num_threads_per_core *
2210                 topo.num_cores_per_pkg * topo.num_packages; i++)
2211                 (*t)[i].cpu_id = -1;
2212
2213         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2214                 sizeof(struct core_data));
2215         if (*c == NULL)
2216                 goto error;
2217
2218         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2219                 (*c)[i].core_id = -1;
2220
2221         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2222         if (*p == NULL)
2223                 goto error;
2224
2225         for (i = 0; i < topo.num_packages; i++)
2226                 (*p)[i].package_id = i;
2227
2228         return;
2229 error:
2230         err(1, "calloc counters");
2231 }
2232 /*
2233  * init_counter()
2234  *
2235  * set cpu_id, core_num, pkg_num
2236  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2237  *
2238  * increment topo.num_cores when 1st core in pkg seen
2239  */
2240 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2241         struct pkg_data *pkg_base, int thread_num, int core_num,
2242         int pkg_num, int cpu_id)
2243 {
2244         struct thread_data *t;
2245         struct core_data *c;
2246         struct pkg_data *p;
2247
2248         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2249         c = GET_CORE(core_base, core_num, pkg_num);
2250         p = GET_PKG(pkg_base, pkg_num);
2251
2252         t->cpu_id = cpu_id;
2253         if (thread_num == 0) {
2254                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2255                 if (cpu_is_first_core_in_package(cpu_id))
2256                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2257         }
2258
2259         c->core_id = core_num;
2260         p->package_id = pkg_num;
2261 }
2262
2263
2264 int initialize_counters(int cpu_id)
2265 {
2266         int my_thread_id, my_core_id, my_package_id;
2267
2268         my_package_id = get_physical_package_id(cpu_id);
2269         my_core_id = get_core_id(cpu_id);
2270
2271         if (cpu_is_first_sibling_in_core(cpu_id)) {
2272                 my_thread_id = 0;
2273                 topo.num_cores++;
2274         } else {
2275                 my_thread_id = 1;
2276         }
2277
2278         init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2279         init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2280         return 0;
2281 }
2282
2283 void allocate_output_buffer()
2284 {
2285         output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
2286         outp = output_buffer;
2287         if (outp == NULL)
2288                 err(-1, "calloc output buffer");
2289 }
2290
2291 void setup_all_buffers(void)
2292 {
2293         topology_probe();
2294         allocate_counters(&thread_even, &core_even, &package_even);
2295         allocate_counters(&thread_odd, &core_odd, &package_odd);
2296         allocate_output_buffer();
2297         for_all_proc_cpus(initialize_counters);
2298 }
2299
2300 void turbostat_init()
2301 {
2302         check_cpuid();
2303
2304         check_dev_msr();
2305         check_super_user();
2306
2307         setup_all_buffers();
2308
2309         if (verbose)
2310                 print_verbose_header();
2311
2312         if (verbose)
2313                 for_all_cpus(print_epb, ODD_COUNTERS);
2314
2315         if (verbose)
2316                 for_all_cpus(print_rapl, ODD_COUNTERS);
2317
2318         for_all_cpus(set_temperature_target, ODD_COUNTERS);
2319
2320         if (verbose)
2321                 for_all_cpus(print_thermal, ODD_COUNTERS);
2322 }
2323
2324 int fork_it(char **argv)
2325 {
2326         pid_t child_pid;
2327         int status;
2328
2329         status = for_all_cpus(get_counters, EVEN_COUNTERS);
2330         if (status)
2331                 exit(status);
2332         /* clear affinity side-effect of get_counters() */
2333         sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
2334         gettimeofday(&tv_even, (struct timezone *)NULL);
2335
2336         child_pid = fork();
2337         if (!child_pid) {
2338                 /* child */
2339                 execvp(argv[0], argv);
2340         } else {
2341
2342                 /* parent */
2343                 if (child_pid == -1)
2344                         err(1, "fork");
2345
2346                 signal(SIGINT, SIG_IGN);
2347                 signal(SIGQUIT, SIG_IGN);
2348                 if (waitpid(child_pid, &status, 0) == -1)
2349                         err(status, "waitpid");
2350         }
2351         /*
2352          * n.b. fork_it() does not check for errors from for_all_cpus()
2353          * because re-starting is problematic when forking
2354          */
2355         for_all_cpus(get_counters, ODD_COUNTERS);
2356         gettimeofday(&tv_odd, (struct timezone *)NULL);
2357         timersub(&tv_odd, &tv_even, &tv_delta);
2358         for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
2359         compute_average(EVEN_COUNTERS);
2360         format_all_counters(EVEN_COUNTERS);
2361         flush_stderr();
2362
2363         fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
2364
2365         return status;
2366 }
2367
2368 int get_and_dump_counters(void)
2369 {
2370         int status;
2371
2372         status = for_all_cpus(get_counters, ODD_COUNTERS);
2373         if (status)
2374                 return status;
2375
2376         status = for_all_cpus(dump_counters, ODD_COUNTERS);
2377         if (status)
2378                 return status;
2379
2380         flush_stdout();
2381
2382         return status;
2383 }
2384
2385 void cmdline(int argc, char **argv)
2386 {
2387         int opt;
2388
2389         progname = argv[0];
2390
2391         while ((opt = getopt(argc, argv, "+pPsSvi:c:C:m:M:RJT:")) != -1) {
2392                 switch (opt) {
2393                 case 'p':
2394                         show_core_only++;
2395                         break;
2396                 case 'P':
2397                         show_pkg_only++;
2398                         break;
2399                 case 's':
2400                         dump_only++;
2401                         break;
2402                 case 'S':
2403                         summary_only++;
2404                         break;
2405                 case 'v':
2406                         verbose++;
2407                         break;
2408                 case 'i':
2409                         interval_sec = atoi(optarg);
2410                         break;
2411                 case 'c':
2412                         sscanf(optarg, "%x", &extra_delta_offset32);
2413                         break;
2414                 case 'C':
2415                         sscanf(optarg, "%x", &extra_delta_offset64);
2416                         break;
2417                 case 'm':
2418                         sscanf(optarg, "%x", &extra_msr_offset32);
2419                         break;
2420                 case 'M':
2421                         sscanf(optarg, "%x", &extra_msr_offset64);
2422                         break;
2423                 case 'R':
2424                         rapl_verbose++;
2425                         break;
2426                 case 'T':
2427                         tcc_activation_temp_override = atoi(optarg);
2428                         break;
2429                 case 'J':
2430                         rapl_joules++;
2431                         break;
2432
2433                 default:
2434                         usage();
2435                 }
2436         }
2437 }
2438
2439 int main(int argc, char **argv)
2440 {
2441         cmdline(argc, argv);
2442
2443         if (verbose)
2444                 fprintf(stderr, "turbostat v3.7 Feb 6, 2014"
2445                         " - Len Brown <lenb@kernel.org>\n");
2446
2447         turbostat_init();
2448
2449         /* dump counters and exit */
2450         if (dump_only)
2451                 return get_and_dump_counters();
2452
2453         /*
2454          * if any params left, it must be a command to fork
2455          */
2456         if (argc - optind)
2457                 return fork_it(argv + optind);
2458         else
2459                 turbostat_loop();
2460
2461         return 0;
2462 }