tools/power turbostat: KNL workaround for %Busy and Avg_MHz
[muen/linux.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 <getopt.h>
37 #include <dirent.h>
38 #include <string.h>
39 #include <ctype.h>
40 #include <sched.h>
41 #include <cpuid.h>
42 #include <linux/capability.h>
43 #include <errno.h>
44
45 char *proc_stat = "/proc/stat";
46 unsigned int interval_sec = 5;
47 unsigned int debug;
48 unsigned int rapl_joules;
49 unsigned int summary_only;
50 unsigned int dump_only;
51 unsigned int skip_c0;
52 unsigned int skip_c1;
53 unsigned int do_nhm_cstates;
54 unsigned int do_snb_cstates;
55 unsigned int do_knl_cstates;
56 unsigned int do_pc2;
57 unsigned int do_pc3;
58 unsigned int do_pc6;
59 unsigned int do_pc7;
60 unsigned int do_c8_c9_c10;
61 unsigned int do_skl_residency;
62 unsigned int do_slm_cstates;
63 unsigned int use_c1_residency_msr;
64 unsigned int has_aperf;
65 unsigned int has_epb;
66 unsigned int units = 1000000;   /* MHz etc */
67 unsigned int genuine_intel;
68 unsigned int has_invariant_tsc;
69 unsigned int do_nhm_platform_info;
70 unsigned int extra_msr_offset32;
71 unsigned int extra_msr_offset64;
72 unsigned int extra_delta_offset32;
73 unsigned int extra_delta_offset64;
74 unsigned int aperf_mperf_multiplier = 1;
75 int do_smi;
76 double bclk;
77 unsigned int show_pkg;
78 unsigned int show_core;
79 unsigned int show_cpu;
80 unsigned int show_pkg_only;
81 unsigned int show_core_only;
82 char *output_buffer, *outp;
83 unsigned int do_rapl;
84 unsigned int do_dts;
85 unsigned int do_ptm;
86 unsigned int tcc_activation_temp;
87 unsigned int tcc_activation_temp_override;
88 double rapl_power_units, rapl_time_units;
89 double rapl_dram_energy_units, rapl_energy_units;
90 double rapl_joule_counter_range;
91 unsigned int do_core_perf_limit_reasons;
92 unsigned int do_gfx_perf_limit_reasons;
93 unsigned int do_ring_perf_limit_reasons;
94 unsigned int crystal_hz;
95 unsigned long long tsc_hz;
96 int base_cpu;
97
98 #define RAPL_PKG                (1 << 0)
99                                         /* 0x610 MSR_PKG_POWER_LIMIT */
100                                         /* 0x611 MSR_PKG_ENERGY_STATUS */
101 #define RAPL_PKG_PERF_STATUS    (1 << 1)
102                                         /* 0x613 MSR_PKG_PERF_STATUS */
103 #define RAPL_PKG_POWER_INFO     (1 << 2)
104                                         /* 0x614 MSR_PKG_POWER_INFO */
105
106 #define RAPL_DRAM               (1 << 3)
107                                         /* 0x618 MSR_DRAM_POWER_LIMIT */
108                                         /* 0x619 MSR_DRAM_ENERGY_STATUS */
109 #define RAPL_DRAM_PERF_STATUS   (1 << 4)
110                                         /* 0x61b MSR_DRAM_PERF_STATUS */
111 #define RAPL_DRAM_POWER_INFO    (1 << 5)
112                                         /* 0x61c MSR_DRAM_POWER_INFO */
113
114 #define RAPL_CORES              (1 << 6)
115                                         /* 0x638 MSR_PP0_POWER_LIMIT */
116                                         /* 0x639 MSR_PP0_ENERGY_STATUS */
117 #define RAPL_CORE_POLICY        (1 << 7)
118                                         /* 0x63a MSR_PP0_POLICY */
119
120 #define RAPL_GFX                (1 << 8)
121                                         /* 0x640 MSR_PP1_POWER_LIMIT */
122                                         /* 0x641 MSR_PP1_ENERGY_STATUS */
123                                         /* 0x642 MSR_PP1_POLICY */
124 #define TJMAX_DEFAULT   100
125
126 #define MAX(a, b) ((a) > (b) ? (a) : (b))
127
128 int aperf_mperf_unstable;
129 int backwards_count;
130 char *progname;
131
132 cpu_set_t *cpu_present_set, *cpu_affinity_set;
133 size_t cpu_present_setsize, cpu_affinity_setsize;
134
135 struct thread_data {
136         unsigned long long tsc;
137         unsigned long long aperf;
138         unsigned long long mperf;
139         unsigned long long c1;
140         unsigned long long extra_msr64;
141         unsigned long long extra_delta64;
142         unsigned long long extra_msr32;
143         unsigned long long extra_delta32;
144         unsigned int smi_count;
145         unsigned int cpu_id;
146         unsigned int flags;
147 #define CPU_IS_FIRST_THREAD_IN_CORE     0x2
148 #define CPU_IS_FIRST_CORE_IN_PACKAGE    0x4
149 } *thread_even, *thread_odd;
150
151 struct core_data {
152         unsigned long long c3;
153         unsigned long long c6;
154         unsigned long long c7;
155         unsigned int core_temp_c;
156         unsigned int core_id;
157 } *core_even, *core_odd;
158
159 struct pkg_data {
160         unsigned long long pc2;
161         unsigned long long pc3;
162         unsigned long long pc6;
163         unsigned long long pc7;
164         unsigned long long pc8;
165         unsigned long long pc9;
166         unsigned long long pc10;
167         unsigned long long pkg_wtd_core_c0;
168         unsigned long long pkg_any_core_c0;
169         unsigned long long pkg_any_gfxe_c0;
170         unsigned long long pkg_both_core_gfxe_c0;
171         unsigned int package_id;
172         unsigned int energy_pkg;        /* MSR_PKG_ENERGY_STATUS */
173         unsigned int energy_dram;       /* MSR_DRAM_ENERGY_STATUS */
174         unsigned int energy_cores;      /* MSR_PP0_ENERGY_STATUS */
175         unsigned int energy_gfx;        /* MSR_PP1_ENERGY_STATUS */
176         unsigned int rapl_pkg_perf_status;      /* MSR_PKG_PERF_STATUS */
177         unsigned int rapl_dram_perf_status;     /* MSR_DRAM_PERF_STATUS */
178         unsigned int pkg_temp_c;
179
180 } *package_even, *package_odd;
181
182 #define ODD_COUNTERS thread_odd, core_odd, package_odd
183 #define EVEN_COUNTERS thread_even, core_even, package_even
184
185 #define GET_THREAD(thread_base, thread_no, core_no, pkg_no) \
186         (thread_base + (pkg_no) * topo.num_cores_per_pkg * \
187                 topo.num_threads_per_core + \
188                 (core_no) * topo.num_threads_per_core + (thread_no))
189 #define GET_CORE(core_base, core_no, pkg_no) \
190         (core_base + (pkg_no) * topo.num_cores_per_pkg + (core_no))
191 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
192
193 struct system_summary {
194         struct thread_data threads;
195         struct core_data cores;
196         struct pkg_data packages;
197 } sum, average;
198
199
200 struct topo_params {
201         int num_packages;
202         int num_cpus;
203         int num_cores;
204         int max_cpu_num;
205         int num_cores_per_pkg;
206         int num_threads_per_core;
207 } topo;
208
209 struct timeval tv_even, tv_odd, tv_delta;
210
211 void setup_all_buffers(void);
212
213 int cpu_is_not_present(int cpu)
214 {
215         return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
216 }
217 /*
218  * run func(thread, core, package) in topology order
219  * skip non-present cpus
220  */
221
222 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
223         struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
224 {
225         int retval, pkg_no, core_no, thread_no;
226
227         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
228                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
229                         for (thread_no = 0; thread_no <
230                                 topo.num_threads_per_core; ++thread_no) {
231                                 struct thread_data *t;
232                                 struct core_data *c;
233                                 struct pkg_data *p;
234
235                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
236
237                                 if (cpu_is_not_present(t->cpu_id))
238                                         continue;
239
240                                 c = GET_CORE(core_base, core_no, pkg_no);
241                                 p = GET_PKG(pkg_base, pkg_no);
242
243                                 retval = func(t, c, p);
244                                 if (retval)
245                                         return retval;
246                         }
247                 }
248         }
249         return 0;
250 }
251
252 int cpu_migrate(int cpu)
253 {
254         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
255         CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
256         if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
257                 return -1;
258         else
259                 return 0;
260 }
261
262 int get_msr(int cpu, off_t offset, unsigned long long *msr)
263 {
264         ssize_t retval;
265         char pathname[32];
266         int fd;
267
268         sprintf(pathname, "/dev/cpu/%d/msr", cpu);
269         fd = open(pathname, O_RDONLY);
270         if (fd < 0)
271                 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
272
273         retval = pread(fd, msr, sizeof *msr, offset);
274         close(fd);
275
276         if (retval != sizeof *msr)
277                 err(-1, "%s offset 0x%llx read failed", pathname, (unsigned long long)offset);
278
279         return 0;
280 }
281
282 /*
283  * Example Format w/ field column widths:
284  *
285  *  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
286  * 123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678123456781234567812345678
287  */
288
289 void print_header(void)
290 {
291         if (show_pkg)
292                 outp += sprintf(outp, " Package");
293         if (show_core)
294                 outp += sprintf(outp, "    Core");
295         if (show_cpu)
296                 outp += sprintf(outp, "     CPU");
297         if (has_aperf)
298                 outp += sprintf(outp, " Avg_MHz");
299         if (has_aperf)
300                 outp += sprintf(outp, "   %%Busy");
301         if (has_aperf)
302                 outp += sprintf(outp, " Bzy_MHz");
303         outp += sprintf(outp, " TSC_MHz");
304
305         if (extra_delta_offset32)
306                 outp += sprintf(outp, "  count 0x%03X", extra_delta_offset32);
307         if (extra_delta_offset64)
308                 outp += sprintf(outp, "  COUNT 0x%03X", extra_delta_offset64);
309         if (extra_msr_offset32)
310                 outp += sprintf(outp, "   MSR 0x%03X", extra_msr_offset32);
311         if (extra_msr_offset64)
312                 outp += sprintf(outp, "           MSR 0x%03X", extra_msr_offset64);
313
314         if (!debug)
315                 goto done;
316
317         if (do_smi)
318                 outp += sprintf(outp, "     SMI");
319
320         if (do_nhm_cstates)
321                 outp += sprintf(outp, "  CPU%%c1");
322         if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
323                 outp += sprintf(outp, "  CPU%%c3");
324         if (do_nhm_cstates)
325                 outp += sprintf(outp, "  CPU%%c6");
326         if (do_snb_cstates)
327                 outp += sprintf(outp, "  CPU%%c7");
328
329         if (do_dts)
330                 outp += sprintf(outp, " CoreTmp");
331         if (do_ptm)
332                 outp += sprintf(outp, "  PkgTmp");
333
334         if (do_skl_residency) {
335                 outp += sprintf(outp, " Totl%%C0");
336                 outp += sprintf(outp, "  Any%%C0");
337                 outp += sprintf(outp, "  GFX%%C0");
338                 outp += sprintf(outp, " CPUGFX%%");
339         }
340
341         if (do_pc2)
342                 outp += sprintf(outp, " Pkg%%pc2");
343         if (do_pc3)
344                 outp += sprintf(outp, " Pkg%%pc3");
345         if (do_pc6)
346                 outp += sprintf(outp, " Pkg%%pc6");
347         if (do_pc7)
348                 outp += sprintf(outp, " Pkg%%pc7");
349         if (do_c8_c9_c10) {
350                 outp += sprintf(outp, " Pkg%%pc8");
351                 outp += sprintf(outp, " Pkg%%pc9");
352                 outp += sprintf(outp, " Pk%%pc10");
353         }
354
355         if (do_rapl && !rapl_joules) {
356                 if (do_rapl & RAPL_PKG)
357                         outp += sprintf(outp, " PkgWatt");
358                 if (do_rapl & RAPL_CORES)
359                         outp += sprintf(outp, " CorWatt");
360                 if (do_rapl & RAPL_GFX)
361                         outp += sprintf(outp, " GFXWatt");
362                 if (do_rapl & RAPL_DRAM)
363                         outp += sprintf(outp, " RAMWatt");
364                 if (do_rapl & RAPL_PKG_PERF_STATUS)
365                         outp += sprintf(outp, "   PKG_%%");
366                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
367                         outp += sprintf(outp, "   RAM_%%");
368         } else if (do_rapl && rapl_joules) {
369                 if (do_rapl & RAPL_PKG)
370                         outp += sprintf(outp, "   Pkg_J");
371                 if (do_rapl & RAPL_CORES)
372                         outp += sprintf(outp, "   Cor_J");
373                 if (do_rapl & RAPL_GFX)
374                         outp += sprintf(outp, "   GFX_J");
375                 if (do_rapl & RAPL_DRAM)
376                         outp += sprintf(outp, "   RAM_J");
377                 if (do_rapl & RAPL_PKG_PERF_STATUS)
378                         outp += sprintf(outp, "   PKG_%%");
379                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
380                         outp += sprintf(outp, "   RAM_%%");
381                 outp += sprintf(outp, "   time");
382
383         }
384     done:
385         outp += sprintf(outp, "\n");
386 }
387
388 int dump_counters(struct thread_data *t, struct core_data *c,
389         struct pkg_data *p)
390 {
391         outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
392
393         if (t) {
394                 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
395                         t->cpu_id, t->flags);
396                 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
397                 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
398                 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
399                 outp += sprintf(outp, "c1: %016llX\n", t->c1);
400                 outp += sprintf(outp, "msr0x%x: %08llX\n",
401                         extra_delta_offset32, t->extra_delta32);
402                 outp += sprintf(outp, "msr0x%x: %016llX\n",
403                         extra_delta_offset64, t->extra_delta64);
404                 outp += sprintf(outp, "msr0x%x: %08llX\n",
405                         extra_msr_offset32, t->extra_msr32);
406                 outp += sprintf(outp, "msr0x%x: %016llX\n",
407                         extra_msr_offset64, t->extra_msr64);
408                 if (do_smi)
409                         outp += sprintf(outp, "SMI: %08X\n", t->smi_count);
410         }
411
412         if (c) {
413                 outp += sprintf(outp, "core: %d\n", c->core_id);
414                 outp += sprintf(outp, "c3: %016llX\n", c->c3);
415                 outp += sprintf(outp, "c6: %016llX\n", c->c6);
416                 outp += sprintf(outp, "c7: %016llX\n", c->c7);
417                 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
418         }
419
420         if (p) {
421                 outp += sprintf(outp, "package: %d\n", p->package_id);
422
423                 outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
424                 outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
425                 outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
426                 outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
427
428                 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
429                 if (do_pc3)
430                         outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
431                 if (do_pc6)
432                         outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
433                 if (do_pc7)
434                         outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
435                 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
436                 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
437                 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
438                 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
439                 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
440                 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
441                 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
442                 outp += sprintf(outp, "Throttle PKG: %0X\n",
443                         p->rapl_pkg_perf_status);
444                 outp += sprintf(outp, "Throttle RAM: %0X\n",
445                         p->rapl_dram_perf_status);
446                 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
447         }
448
449         outp += sprintf(outp, "\n");
450
451         return 0;
452 }
453
454 /*
455  * column formatting convention & formats
456  */
457 int format_counters(struct thread_data *t, struct core_data *c,
458         struct pkg_data *p)
459 {
460         double interval_float;
461         char *fmt8;
462
463          /* if showing only 1st thread in core and this isn't one, bail out */
464         if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
465                 return 0;
466
467          /* if showing only 1st thread in pkg and this isn't one, bail out */
468         if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
469                 return 0;
470
471         interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
472
473         /* topo columns, print blanks on 1st (average) line */
474         if (t == &average.threads) {
475                 if (show_pkg)
476                         outp += sprintf(outp, "       -");
477                 if (show_core)
478                         outp += sprintf(outp, "       -");
479                 if (show_cpu)
480                         outp += sprintf(outp, "       -");
481         } else {
482                 if (show_pkg) {
483                         if (p)
484                                 outp += sprintf(outp, "%8d", p->package_id);
485                         else
486                                 outp += sprintf(outp, "       -");
487                 }
488                 if (show_core) {
489                         if (c)
490                                 outp += sprintf(outp, "%8d", c->core_id);
491                         else
492                                 outp += sprintf(outp, "       -");
493                 }
494                 if (show_cpu)
495                         outp += sprintf(outp, "%8d", t->cpu_id);
496         }
497
498         /* Avg_MHz */
499         if (has_aperf)
500                 outp += sprintf(outp, "%8.0f",
501                         1.0 / units * t->aperf / interval_float);
502
503         /* %Busy */
504         if (has_aperf) {
505                 if (!skip_c0)
506                         outp += sprintf(outp, "%8.2f", 100.0 * t->mperf/t->tsc);
507                 else
508                         outp += sprintf(outp, "********");
509         }
510
511         /* Bzy_MHz */
512         if (has_aperf)
513                 outp += sprintf(outp, "%8.0f",
514                         1.0 * t->tsc / units * t->aperf / t->mperf / interval_float);
515
516         /* TSC_MHz */
517         outp += sprintf(outp, "%8.0f", 1.0 * t->tsc/units/interval_float);
518
519         /* delta */
520         if (extra_delta_offset32)
521                 outp += sprintf(outp, "  %11llu", t->extra_delta32);
522
523         /* DELTA */
524         if (extra_delta_offset64)
525                 outp += sprintf(outp, "  %11llu", t->extra_delta64);
526         /* msr */
527         if (extra_msr_offset32)
528                 outp += sprintf(outp, "  0x%08llx", t->extra_msr32);
529
530         /* MSR */
531         if (extra_msr_offset64)
532                 outp += sprintf(outp, "  0x%016llx", t->extra_msr64);
533
534         if (!debug)
535                 goto done;
536
537         /* SMI */
538         if (do_smi)
539                 outp += sprintf(outp, "%8d", t->smi_count);
540
541         if (do_nhm_cstates) {
542                 if (!skip_c1)
543                         outp += sprintf(outp, "%8.2f", 100.0 * t->c1/t->tsc);
544                 else
545                         outp += sprintf(outp, "********");
546         }
547
548         /* print per-core data only for 1st thread in core */
549         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
550                 goto done;
551
552         if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates)
553                 outp += sprintf(outp, "%8.2f", 100.0 * c->c3/t->tsc);
554         if (do_nhm_cstates)
555                 outp += sprintf(outp, "%8.2f", 100.0 * c->c6/t->tsc);
556         if (do_snb_cstates)
557                 outp += sprintf(outp, "%8.2f", 100.0 * c->c7/t->tsc);
558
559         if (do_dts)
560                 outp += sprintf(outp, "%8d", c->core_temp_c);
561
562         /* print per-package data only for 1st core in package */
563         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
564                 goto done;
565
566         /* PkgTmp */
567         if (do_ptm)
568                 outp += sprintf(outp, "%8d", p->pkg_temp_c);
569
570         /* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
571         if (do_skl_residency) {
572                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_wtd_core_c0/t->tsc);
573                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_any_core_c0/t->tsc);
574                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_any_gfxe_c0/t->tsc);
575                 outp += sprintf(outp, "%8.2f", 100.0 * p->pkg_both_core_gfxe_c0/t->tsc);
576         }
577
578         if (do_pc2)
579                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc2/t->tsc);
580         if (do_pc3)
581                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc3/t->tsc);
582         if (do_pc6)
583                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc6/t->tsc);
584         if (do_pc7)
585                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc7/t->tsc);
586         if (do_c8_c9_c10) {
587                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc8/t->tsc);
588                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc9/t->tsc);
589                 outp += sprintf(outp, "%8.2f", 100.0 * p->pc10/t->tsc);
590         }
591
592         /*
593          * If measurement interval exceeds minimum RAPL Joule Counter range,
594          * indicate that results are suspect by printing "**" in fraction place.
595          */
596         if (interval_float < rapl_joule_counter_range)
597                 fmt8 = "%8.2f";
598         else
599                 fmt8 = " %6.0f**";
600
601         if (do_rapl && !rapl_joules) {
602                 if (do_rapl & RAPL_PKG)
603                         outp += sprintf(outp, fmt8, p->energy_pkg * rapl_energy_units / interval_float);
604                 if (do_rapl & RAPL_CORES)
605                         outp += sprintf(outp, fmt8, p->energy_cores * rapl_energy_units / interval_float);
606                 if (do_rapl & RAPL_GFX)
607                         outp += sprintf(outp, fmt8, p->energy_gfx * rapl_energy_units / interval_float);
608                 if (do_rapl & RAPL_DRAM)
609                         outp += sprintf(outp, fmt8, p->energy_dram * rapl_dram_energy_units / interval_float);
610                 if (do_rapl & RAPL_PKG_PERF_STATUS)
611                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
612                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
613                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
614         } else if (do_rapl && rapl_joules) {
615                 if (do_rapl & RAPL_PKG)
616                         outp += sprintf(outp, fmt8,
617                                         p->energy_pkg * rapl_energy_units);
618                 if (do_rapl & RAPL_CORES)
619                         outp += sprintf(outp, fmt8,
620                                         p->energy_cores * rapl_energy_units);
621                 if (do_rapl & RAPL_GFX)
622                         outp += sprintf(outp, fmt8,
623                                         p->energy_gfx * rapl_energy_units);
624                 if (do_rapl & RAPL_DRAM)
625                         outp += sprintf(outp, fmt8,
626                                         p->energy_dram * rapl_dram_energy_units);
627                 if (do_rapl & RAPL_PKG_PERF_STATUS)
628                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
629                 if (do_rapl & RAPL_DRAM_PERF_STATUS)
630                         outp += sprintf(outp, fmt8, 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
631
632                 outp += sprintf(outp, fmt8, interval_float);
633         }
634 done:
635         outp += sprintf(outp, "\n");
636
637         return 0;
638 }
639
640 void flush_stdout()
641 {
642         fputs(output_buffer, stdout);
643         fflush(stdout);
644         outp = output_buffer;
645 }
646 void flush_stderr()
647 {
648         fputs(output_buffer, stderr);
649         outp = output_buffer;
650 }
651 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
652 {
653         static int printed;
654
655         if (!printed || !summary_only)
656                 print_header();
657
658         if (topo.num_cpus > 1)
659                 format_counters(&average.threads, &average.cores,
660                         &average.packages);
661
662         printed = 1;
663
664         if (summary_only)
665                 return;
666
667         for_all_cpus(format_counters, t, c, p);
668 }
669
670 #define DELTA_WRAP32(new, old)                  \
671         if (new > old) {                        \
672                 old = new - old;                \
673         } else {                                \
674                 old = 0x100000000 + new - old;  \
675         }
676
677 void
678 delta_package(struct pkg_data *new, struct pkg_data *old)
679 {
680
681         if (do_skl_residency) {
682                 old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
683                 old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
684                 old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
685                 old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
686         }
687         old->pc2 = new->pc2 - old->pc2;
688         if (do_pc3)
689                 old->pc3 = new->pc3 - old->pc3;
690         if (do_pc6)
691                 old->pc6 = new->pc6 - old->pc6;
692         if (do_pc7)
693                 old->pc7 = new->pc7 - old->pc7;
694         old->pc8 = new->pc8 - old->pc8;
695         old->pc9 = new->pc9 - old->pc9;
696         old->pc10 = new->pc10 - old->pc10;
697         old->pkg_temp_c = new->pkg_temp_c;
698
699         DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
700         DELTA_WRAP32(new->energy_cores, old->energy_cores);
701         DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
702         DELTA_WRAP32(new->energy_dram, old->energy_dram);
703         DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
704         DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
705 }
706
707 void
708 delta_core(struct core_data *new, struct core_data *old)
709 {
710         old->c3 = new->c3 - old->c3;
711         old->c6 = new->c6 - old->c6;
712         old->c7 = new->c7 - old->c7;
713         old->core_temp_c = new->core_temp_c;
714 }
715
716 /*
717  * old = new - old
718  */
719 void
720 delta_thread(struct thread_data *new, struct thread_data *old,
721         struct core_data *core_delta)
722 {
723         old->tsc = new->tsc - old->tsc;
724
725         /* check for TSC < 1 Mcycles over interval */
726         if (old->tsc < (1000 * 1000))
727                 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
728                      "You can disable all c-states by booting with \"idle=poll\"\n"
729                      "or just the deep ones with \"processor.max_cstate=1\"");
730
731         old->c1 = new->c1 - old->c1;
732
733         if (has_aperf) {
734                 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
735                         old->aperf = new->aperf - old->aperf;
736                         old->mperf = new->mperf - old->mperf;
737                 } else {
738
739                         if (!aperf_mperf_unstable) {
740                                 fprintf(stderr, "%s: APERF or MPERF went backwards *\n", progname);
741                                 fprintf(stderr, "* Frequency results do not cover entire interval *\n");
742                                 fprintf(stderr, "* fix this by running Linux-2.6.30 or later *\n");
743
744                                 aperf_mperf_unstable = 1;
745                         }
746                         /*
747                          * mperf delta is likely a huge "positive" number
748                          * can not use it for calculating c0 time
749                          */
750                         skip_c0 = 1;
751                         skip_c1 = 1;
752                 }
753         }
754
755
756         if (use_c1_residency_msr) {
757                 /*
758                  * Some models have a dedicated C1 residency MSR,
759                  * which should be more accurate than the derivation below.
760                  */
761         } else {
762                 /*
763                  * As counter collection is not atomic,
764                  * it is possible for mperf's non-halted cycles + idle states
765                  * to exceed TSC's all cycles: show c1 = 0% in that case.
766                  */
767                 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > old->tsc)
768                         old->c1 = 0;
769                 else {
770                         /* normal case, derive c1 */
771                         old->c1 = old->tsc - old->mperf - core_delta->c3
772                                 - core_delta->c6 - core_delta->c7;
773                 }
774         }
775
776         if (old->mperf == 0) {
777                 if (debug > 1) fprintf(stderr, "cpu%d MPERF 0!\n", old->cpu_id);
778                 old->mperf = 1; /* divide by 0 protection */
779         }
780
781         old->extra_delta32 = new->extra_delta32 - old->extra_delta32;
782         old->extra_delta32 &= 0xFFFFFFFF;
783
784         old->extra_delta64 = new->extra_delta64 - old->extra_delta64;
785
786         /*
787          * Extra MSR is just a snapshot, simply copy latest w/o subtracting
788          */
789         old->extra_msr32 = new->extra_msr32;
790         old->extra_msr64 = new->extra_msr64;
791
792         if (do_smi)
793                 old->smi_count = new->smi_count - old->smi_count;
794 }
795
796 int delta_cpu(struct thread_data *t, struct core_data *c,
797         struct pkg_data *p, struct thread_data *t2,
798         struct core_data *c2, struct pkg_data *p2)
799 {
800         /* calculate core delta only for 1st thread in core */
801         if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
802                 delta_core(c, c2);
803
804         /* always calculate thread delta */
805         delta_thread(t, t2, c2);        /* c2 is core delta */
806
807         /* calculate package delta only for 1st core in package */
808         if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
809                 delta_package(p, p2);
810
811         return 0;
812 }
813
814 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
815 {
816         t->tsc = 0;
817         t->aperf = 0;
818         t->mperf = 0;
819         t->c1 = 0;
820
821         t->smi_count = 0;
822         t->extra_delta32 = 0;
823         t->extra_delta64 = 0;
824
825         /* tells format_counters to dump all fields from this set */
826         t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
827
828         c->c3 = 0;
829         c->c6 = 0;
830         c->c7 = 0;
831         c->core_temp_c = 0;
832
833         p->pkg_wtd_core_c0 = 0;
834         p->pkg_any_core_c0 = 0;
835         p->pkg_any_gfxe_c0 = 0;
836         p->pkg_both_core_gfxe_c0 = 0;
837
838         p->pc2 = 0;
839         if (do_pc3)
840                 p->pc3 = 0;
841         if (do_pc6)
842                 p->pc6 = 0;
843         if (do_pc7)
844                 p->pc7 = 0;
845         p->pc8 = 0;
846         p->pc9 = 0;
847         p->pc10 = 0;
848
849         p->energy_pkg = 0;
850         p->energy_dram = 0;
851         p->energy_cores = 0;
852         p->energy_gfx = 0;
853         p->rapl_pkg_perf_status = 0;
854         p->rapl_dram_perf_status = 0;
855         p->pkg_temp_c = 0;
856 }
857 int sum_counters(struct thread_data *t, struct core_data *c,
858         struct pkg_data *p)
859 {
860         average.threads.tsc += t->tsc;
861         average.threads.aperf += t->aperf;
862         average.threads.mperf += t->mperf;
863         average.threads.c1 += t->c1;
864
865         average.threads.extra_delta32 += t->extra_delta32;
866         average.threads.extra_delta64 += t->extra_delta64;
867
868         /* sum per-core values only for 1st thread in core */
869         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
870                 return 0;
871
872         average.cores.c3 += c->c3;
873         average.cores.c6 += c->c6;
874         average.cores.c7 += c->c7;
875
876         average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
877
878         /* sum per-pkg values only for 1st core in pkg */
879         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
880                 return 0;
881
882         if (do_skl_residency) {
883                 average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
884                 average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
885                 average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
886                 average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
887         }
888
889         average.packages.pc2 += p->pc2;
890         if (do_pc3)
891                 average.packages.pc3 += p->pc3;
892         if (do_pc6)
893                 average.packages.pc6 += p->pc6;
894         if (do_pc7)
895                 average.packages.pc7 += p->pc7;
896         average.packages.pc8 += p->pc8;
897         average.packages.pc9 += p->pc9;
898         average.packages.pc10 += p->pc10;
899
900         average.packages.energy_pkg += p->energy_pkg;
901         average.packages.energy_dram += p->energy_dram;
902         average.packages.energy_cores += p->energy_cores;
903         average.packages.energy_gfx += p->energy_gfx;
904
905         average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
906
907         average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
908         average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
909         return 0;
910 }
911 /*
912  * sum the counters for all cpus in the system
913  * compute the weighted average
914  */
915 void compute_average(struct thread_data *t, struct core_data *c,
916         struct pkg_data *p)
917 {
918         clear_counters(&average.threads, &average.cores, &average.packages);
919
920         for_all_cpus(sum_counters, t, c, p);
921
922         average.threads.tsc /= topo.num_cpus;
923         average.threads.aperf /= topo.num_cpus;
924         average.threads.mperf /= topo.num_cpus;
925         average.threads.c1 /= topo.num_cpus;
926
927         average.threads.extra_delta32 /= topo.num_cpus;
928         average.threads.extra_delta32 &= 0xFFFFFFFF;
929
930         average.threads.extra_delta64 /= topo.num_cpus;
931
932         average.cores.c3 /= topo.num_cores;
933         average.cores.c6 /= topo.num_cores;
934         average.cores.c7 /= topo.num_cores;
935
936         if (do_skl_residency) {
937                 average.packages.pkg_wtd_core_c0 /= topo.num_packages;
938                 average.packages.pkg_any_core_c0 /= topo.num_packages;
939                 average.packages.pkg_any_gfxe_c0 /= topo.num_packages;
940                 average.packages.pkg_both_core_gfxe_c0 /= topo.num_packages;
941         }
942
943         average.packages.pc2 /= topo.num_packages;
944         if (do_pc3)
945                 average.packages.pc3 /= topo.num_packages;
946         if (do_pc6)
947                 average.packages.pc6 /= topo.num_packages;
948         if (do_pc7)
949                 average.packages.pc7 /= topo.num_packages;
950
951         average.packages.pc8 /= topo.num_packages;
952         average.packages.pc9 /= topo.num_packages;
953         average.packages.pc10 /= topo.num_packages;
954 }
955
956 static unsigned long long rdtsc(void)
957 {
958         unsigned int low, high;
959
960         asm volatile("rdtsc" : "=a" (low), "=d" (high));
961
962         return low | ((unsigned long long)high) << 32;
963 }
964
965
966 /*
967  * get_counters(...)
968  * migrate to cpu
969  * acquire and record local counters for that cpu
970  */
971 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
972 {
973         int cpu = t->cpu_id;
974         unsigned long long msr;
975
976         if (cpu_migrate(cpu)) {
977                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
978                 return -1;
979         }
980
981         t->tsc = rdtsc();       /* we are running on local CPU of interest */
982
983         if (has_aperf) {
984                 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
985                         return -3;
986                 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
987                         return -4;
988                 t->aperf = t->aperf * aperf_mperf_multiplier;
989                 t->mperf = t->mperf * aperf_mperf_multiplier;
990         }
991
992         if (do_smi) {
993                 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
994                         return -5;
995                 t->smi_count = msr & 0xFFFFFFFF;
996         }
997         if (extra_delta_offset32) {
998                 if (get_msr(cpu, extra_delta_offset32, &msr))
999                         return -5;
1000                 t->extra_delta32 = msr & 0xFFFFFFFF;
1001         }
1002
1003         if (extra_delta_offset64)
1004                 if (get_msr(cpu, extra_delta_offset64, &t->extra_delta64))
1005                         return -5;
1006
1007         if (extra_msr_offset32) {
1008                 if (get_msr(cpu, extra_msr_offset32, &msr))
1009                         return -5;
1010                 t->extra_msr32 = msr & 0xFFFFFFFF;
1011         }
1012
1013         if (extra_msr_offset64)
1014                 if (get_msr(cpu, extra_msr_offset64, &t->extra_msr64))
1015                         return -5;
1016
1017         if (use_c1_residency_msr) {
1018                 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
1019                         return -6;
1020         }
1021
1022         /* collect core counters only for 1st thread in core */
1023         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1024                 return 0;
1025
1026         if (do_nhm_cstates && !do_slm_cstates && !do_knl_cstates) {
1027                 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
1028                         return -6;
1029         }
1030
1031         if (do_nhm_cstates && !do_knl_cstates) {
1032                 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
1033                         return -7;
1034         } else if (do_knl_cstates) {
1035                 if (get_msr(cpu, MSR_KNL_CORE_C6_RESIDENCY, &c->c6))
1036                         return -7;
1037         }
1038
1039         if (do_snb_cstates)
1040                 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
1041                         return -8;
1042
1043         if (do_dts) {
1044                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1045                         return -9;
1046                 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1047         }
1048
1049
1050         /* collect package counters only for 1st core in package */
1051         if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1052                 return 0;
1053
1054         if (do_skl_residency) {
1055                 if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
1056                         return -10;
1057                 if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
1058                         return -11;
1059                 if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
1060                         return -12;
1061                 if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
1062                         return -13;
1063         }
1064         if (do_pc3)
1065                 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
1066                         return -9;
1067         if (do_pc6)
1068                 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
1069                         return -10;
1070         if (do_pc2)
1071                 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
1072                         return -11;
1073         if (do_pc7)
1074                 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1075                         return -12;
1076         if (do_c8_c9_c10) {
1077                 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1078                         return -13;
1079                 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1080                         return -13;
1081                 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1082                         return -13;
1083         }
1084         if (do_rapl & RAPL_PKG) {
1085                 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1086                         return -13;
1087                 p->energy_pkg = msr & 0xFFFFFFFF;
1088         }
1089         if (do_rapl & RAPL_CORES) {
1090                 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1091                         return -14;
1092                 p->energy_cores = msr & 0xFFFFFFFF;
1093         }
1094         if (do_rapl & RAPL_DRAM) {
1095                 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1096                         return -15;
1097                 p->energy_dram = msr & 0xFFFFFFFF;
1098         }
1099         if (do_rapl & RAPL_GFX) {
1100                 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1101                         return -16;
1102                 p->energy_gfx = msr & 0xFFFFFFFF;
1103         }
1104         if (do_rapl & RAPL_PKG_PERF_STATUS) {
1105                 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1106                         return -16;
1107                 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1108         }
1109         if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1110                 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1111                         return -16;
1112                 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1113         }
1114         if (do_ptm) {
1115                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1116                         return -17;
1117                 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1118         }
1119         return 0;
1120 }
1121
1122 /*
1123  * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1124  * If you change the values, note they are used both in comparisons
1125  * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1126  */
1127
1128 #define PCLUKN 0 /* Unknown */
1129 #define PCLRSV 1 /* Reserved */
1130 #define PCL__0 2 /* PC0 */
1131 #define PCL__1 3 /* PC1 */
1132 #define PCL__2 4 /* PC2 */
1133 #define PCL__3 5 /* PC3 */
1134 #define PCL__4 6 /* PC4 */
1135 #define PCL__6 7 /* PC6 */
1136 #define PCL_6N 8 /* PC6 No Retention */
1137 #define PCL_6R 9 /* PC6 Retention */
1138 #define PCL__7 10 /* PC7 */
1139 #define PCL_7S 11 /* PC7 Shrink */
1140 #define PCL__8 12 /* PC8 */
1141 #define PCL__9 13 /* PC9 */
1142 #define PCLUNL 14 /* Unlimited */
1143
1144 int pkg_cstate_limit = PCLUKN;
1145 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1146         "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "unlimited"};
1147
1148 int nhm_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1149 int snb_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1150 int hsw_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1151 int slv_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1152 int amt_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1153 int phi_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1154
1155 static void
1156 dump_nhm_platform_info(void)
1157 {
1158         unsigned long long msr;
1159         unsigned int ratio;
1160
1161         get_msr(base_cpu, MSR_NHM_PLATFORM_INFO, &msr);
1162
1163         fprintf(stderr, "cpu%d: MSR_NHM_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
1164
1165         ratio = (msr >> 40) & 0xFF;
1166         fprintf(stderr, "%d * %.0f = %.0f MHz max efficiency frequency\n",
1167                 ratio, bclk, ratio * bclk);
1168
1169         ratio = (msr >> 8) & 0xFF;
1170         fprintf(stderr, "%d * %.0f = %.0f MHz base frequency\n",
1171                 ratio, bclk, ratio * bclk);
1172
1173         get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
1174         fprintf(stderr, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1175                 base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
1176
1177         return;
1178 }
1179
1180 static void
1181 dump_hsw_turbo_ratio_limits(void)
1182 {
1183         unsigned long long msr;
1184         unsigned int ratio;
1185
1186         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
1187
1188         fprintf(stderr, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
1189
1190         ratio = (msr >> 8) & 0xFF;
1191         if (ratio)
1192                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 18 active cores\n",
1193                         ratio, bclk, ratio * bclk);
1194
1195         ratio = (msr >> 0) & 0xFF;
1196         if (ratio)
1197                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 17 active cores\n",
1198                         ratio, bclk, ratio * bclk);
1199         return;
1200 }
1201
1202 static void
1203 dump_ivt_turbo_ratio_limits(void)
1204 {
1205         unsigned long long msr;
1206         unsigned int ratio;
1207
1208         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
1209
1210         fprintf(stderr, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
1211
1212         ratio = (msr >> 56) & 0xFF;
1213         if (ratio)
1214                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 16 active cores\n",
1215                         ratio, bclk, ratio * bclk);
1216
1217         ratio = (msr >> 48) & 0xFF;
1218         if (ratio)
1219                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 15 active cores\n",
1220                         ratio, bclk, ratio * bclk);
1221
1222         ratio = (msr >> 40) & 0xFF;
1223         if (ratio)
1224                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 14 active cores\n",
1225                         ratio, bclk, ratio * bclk);
1226
1227         ratio = (msr >> 32) & 0xFF;
1228         if (ratio)
1229                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 13 active cores\n",
1230                         ratio, bclk, ratio * bclk);
1231
1232         ratio = (msr >> 24) & 0xFF;
1233         if (ratio)
1234                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 12 active cores\n",
1235                         ratio, bclk, ratio * bclk);
1236
1237         ratio = (msr >> 16) & 0xFF;
1238         if (ratio)
1239                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 11 active cores\n",
1240                         ratio, bclk, ratio * bclk);
1241
1242         ratio = (msr >> 8) & 0xFF;
1243         if (ratio)
1244                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 10 active cores\n",
1245                         ratio, bclk, ratio * bclk);
1246
1247         ratio = (msr >> 0) & 0xFF;
1248         if (ratio)
1249                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 9 active cores\n",
1250                         ratio, bclk, ratio * bclk);
1251         return;
1252 }
1253
1254 static void
1255 dump_nhm_turbo_ratio_limits(void)
1256 {
1257         unsigned long long msr;
1258         unsigned int ratio;
1259
1260         get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
1261
1262         fprintf(stderr, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
1263
1264         ratio = (msr >> 56) & 0xFF;
1265         if (ratio)
1266                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 8 active cores\n",
1267                         ratio, bclk, ratio * bclk);
1268
1269         ratio = (msr >> 48) & 0xFF;
1270         if (ratio)
1271                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 7 active cores\n",
1272                         ratio, bclk, ratio * bclk);
1273
1274         ratio = (msr >> 40) & 0xFF;
1275         if (ratio)
1276                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 6 active cores\n",
1277                         ratio, bclk, ratio * bclk);
1278
1279         ratio = (msr >> 32) & 0xFF;
1280         if (ratio)
1281                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 5 active cores\n",
1282                         ratio, bclk, ratio * bclk);
1283
1284         ratio = (msr >> 24) & 0xFF;
1285         if (ratio)
1286                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 4 active cores\n",
1287                         ratio, bclk, ratio * bclk);
1288
1289         ratio = (msr >> 16) & 0xFF;
1290         if (ratio)
1291                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 3 active cores\n",
1292                         ratio, bclk, ratio * bclk);
1293
1294         ratio = (msr >> 8) & 0xFF;
1295         if (ratio)
1296                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 2 active cores\n",
1297                         ratio, bclk, ratio * bclk);
1298
1299         ratio = (msr >> 0) & 0xFF;
1300         if (ratio)
1301                 fprintf(stderr, "%d * %.0f = %.0f MHz max turbo 1 active cores\n",
1302                         ratio, bclk, ratio * bclk);
1303         return;
1304 }
1305
1306 static void
1307 dump_knl_turbo_ratio_limits(void)
1308 {
1309         int cores;
1310         unsigned int ratio;
1311         unsigned long long msr;
1312         int delta_cores;
1313         int delta_ratio;
1314         int i;
1315
1316         get_msr(base_cpu, MSR_NHM_TURBO_RATIO_LIMIT, &msr);
1317
1318         fprintf(stderr, "cpu%d: MSR_NHM_TURBO_RATIO_LIMIT: 0x%08llx\n",
1319                 base_cpu, msr);
1320
1321         /**
1322          * Turbo encoding in KNL is as follows:
1323          * [7:0] -- Base value of number of active cores of bucket 1.
1324          * [15:8] -- Base value of freq ratio of bucket 1.
1325          * [20:16] -- +ve delta of number of active cores of bucket 2.
1326          * i.e. active cores of bucket 2 =
1327          * active cores of bucket 1 + delta
1328          * [23:21] -- Negative delta of freq ratio of bucket 2.
1329          * i.e. freq ratio of bucket 2 =
1330          * freq ratio of bucket 1 - delta
1331          * [28:24]-- +ve delta of number of active cores of bucket 3.
1332          * [31:29]-- -ve delta of freq ratio of bucket 3.
1333          * [36:32]-- +ve delta of number of active cores of bucket 4.
1334          * [39:37]-- -ve delta of freq ratio of bucket 4.
1335          * [44:40]-- +ve delta of number of active cores of bucket 5.
1336          * [47:45]-- -ve delta of freq ratio of bucket 5.
1337          * [52:48]-- +ve delta of number of active cores of bucket 6.
1338          * [55:53]-- -ve delta of freq ratio of bucket 6.
1339          * [60:56]-- +ve delta of number of active cores of bucket 7.
1340          * [63:61]-- -ve delta of freq ratio of bucket 7.
1341          */
1342         cores = msr & 0xFF;
1343         ratio = (msr >> 8) && 0xFF;
1344         if (ratio > 0)
1345                 fprintf(stderr,
1346                         "%d * %.0f = %.0f MHz max turbo %d active cores\n",
1347                         ratio, bclk, ratio * bclk, cores);
1348
1349         for (i = 16; i < 64; i = i + 8) {
1350                 delta_cores = (msr >> i) & 0x1F;
1351                 delta_ratio = (msr >> (i + 5)) && 0x7;
1352                 if (!delta_cores || !delta_ratio)
1353                         return;
1354                 cores = cores + delta_cores;
1355                 ratio = ratio - delta_ratio;
1356
1357                 /** -ve ratios will make successive ratio calculations
1358                  * negative. Hence return instead of carrying on.
1359                  */
1360                 if (ratio > 0)
1361                         fprintf(stderr,
1362                                 "%d * %.0f = %.0f MHz max turbo %d active cores\n",
1363                                 ratio, bclk, ratio * bclk, cores);
1364         }
1365 }
1366
1367 static void
1368 dump_nhm_cst_cfg(void)
1369 {
1370         unsigned long long msr;
1371
1372         get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1373
1374 #define SNB_C1_AUTO_UNDEMOTE              (1UL << 27)
1375 #define SNB_C3_AUTO_UNDEMOTE              (1UL << 28)
1376
1377         fprintf(stderr, "cpu%d: MSR_NHM_SNB_PKG_CST_CFG_CTL: 0x%08llx", base_cpu, msr);
1378
1379         fprintf(stderr, " (%s%s%s%s%slocked: pkg-cstate-limit=%d: %s)\n",
1380                 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
1381                 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
1382                 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
1383                 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
1384                 (msr & (1 << 15)) ? "" : "UN",
1385                 (unsigned int)msr & 7,
1386                 pkg_cstate_limit_strings[pkg_cstate_limit]);
1387         return;
1388 }
1389
1390 static void
1391 dump_config_tdp(void)
1392 {
1393         unsigned long long msr;
1394
1395         get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
1396         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
1397         fprintf(stderr, " (base_ratio=%d)\n", (unsigned int)msr & 0xEF);
1398
1399         get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
1400         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
1401         if (msr) {
1402                 fprintf(stderr, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0xEFFF);
1403                 fprintf(stderr, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0xEFFF);
1404                 fprintf(stderr, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xEF);
1405                 fprintf(stderr, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0xEFFF);
1406         }
1407         fprintf(stderr, ")\n");
1408
1409         get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
1410         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
1411         if (msr) {
1412                 fprintf(stderr, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0xEFFF);
1413                 fprintf(stderr, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0xEFFF);
1414                 fprintf(stderr, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xEF);
1415                 fprintf(stderr, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0xEFFF);
1416         }
1417         fprintf(stderr, ")\n");
1418
1419         get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
1420         fprintf(stderr, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
1421         if ((msr) & 0x3)
1422                 fprintf(stderr, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
1423         fprintf(stderr, " lock=%d", (unsigned int)(msr >> 31) & 1);
1424         fprintf(stderr, ")\n");
1425         
1426         get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
1427         fprintf(stderr, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
1428         fprintf(stderr, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xEF);
1429         fprintf(stderr, " lock=%d", (unsigned int)(msr >> 31) & 1);
1430         fprintf(stderr, ")\n");
1431 }
1432
1433 void free_all_buffers(void)
1434 {
1435         CPU_FREE(cpu_present_set);
1436         cpu_present_set = NULL;
1437         cpu_present_set = 0;
1438
1439         CPU_FREE(cpu_affinity_set);
1440         cpu_affinity_set = NULL;
1441         cpu_affinity_setsize = 0;
1442
1443         free(thread_even);
1444         free(core_even);
1445         free(package_even);
1446
1447         thread_even = NULL;
1448         core_even = NULL;
1449         package_even = NULL;
1450
1451         free(thread_odd);
1452         free(core_odd);
1453         free(package_odd);
1454
1455         thread_odd = NULL;
1456         core_odd = NULL;
1457         package_odd = NULL;
1458
1459         free(output_buffer);
1460         output_buffer = NULL;
1461         outp = NULL;
1462 }
1463
1464 /*
1465  * Open a file, and exit on failure
1466  */
1467 FILE *fopen_or_die(const char *path, const char *mode)
1468 {
1469         FILE *filep = fopen(path, "r");
1470         if (!filep)
1471                 err(1, "%s: open failed", path);
1472         return filep;
1473 }
1474
1475 /*
1476  * Parse a file containing a single int.
1477  */
1478 int parse_int_file(const char *fmt, ...)
1479 {
1480         va_list args;
1481         char path[PATH_MAX];
1482         FILE *filep;
1483         int value;
1484
1485         va_start(args, fmt);
1486         vsnprintf(path, sizeof(path), fmt, args);
1487         va_end(args);
1488         filep = fopen_or_die(path, "r");
1489         if (fscanf(filep, "%d", &value) != 1)
1490                 err(1, "%s: failed to parse number from file", path);
1491         fclose(filep);
1492         return value;
1493 }
1494
1495 /*
1496  * get_cpu_position_in_core(cpu)
1497  * return the position of the CPU among its HT siblings in the core
1498  * return -1 if the sibling is not in list
1499  */
1500 int get_cpu_position_in_core(int cpu)
1501 {
1502         char path[64];
1503         FILE *filep;
1504         int this_cpu;
1505         char character;
1506         int i;
1507
1508         sprintf(path,
1509                 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list",
1510                 cpu);
1511         filep = fopen(path, "r");
1512         if (filep == NULL) {
1513                 perror(path);
1514                 exit(1);
1515         }
1516
1517         for (i = 0; i < topo.num_threads_per_core; i++) {
1518                 fscanf(filep, "%d", &this_cpu);
1519                 if (this_cpu == cpu) {
1520                         fclose(filep);
1521                         return i;
1522                 }
1523
1524                 /* Account for no separator after last thread*/
1525                 if (i != (topo.num_threads_per_core - 1))
1526                         fscanf(filep, "%c", &character);
1527         }
1528
1529         fclose(filep);
1530         return -1;
1531 }
1532
1533 /*
1534  * cpu_is_first_core_in_package(cpu)
1535  * return 1 if given CPU is 1st core in package
1536  */
1537 int cpu_is_first_core_in_package(int cpu)
1538 {
1539         return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
1540 }
1541
1542 int get_physical_package_id(int cpu)
1543 {
1544         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
1545 }
1546
1547 int get_core_id(int cpu)
1548 {
1549         return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
1550 }
1551
1552 int get_num_ht_siblings(int cpu)
1553 {
1554         char path[80];
1555         FILE *filep;
1556         int sib1;
1557         int matches = 0;
1558         char character;
1559         char str[100];
1560         char *ch;
1561
1562         sprintf(path, "/sys/devices/system/cpu/cpu%d/topology/thread_siblings_list", cpu);
1563         filep = fopen_or_die(path, "r");
1564
1565         /*
1566          * file format:
1567          * A ',' separated or '-' separated set of numbers
1568          * (eg 1-2 or 1,3,4,5)
1569          */
1570         fscanf(filep, "%d%c\n", &sib1, &character);
1571         fseek(filep, 0, SEEK_SET);
1572         fgets(str, 100, filep);
1573         ch = strchr(str, character);
1574         while (ch != NULL) {
1575                 matches++;
1576                 ch = strchr(ch+1, character);
1577         }
1578
1579         fclose(filep);
1580         return matches+1;
1581 }
1582
1583 /*
1584  * run func(thread, core, package) in topology order
1585  * skip non-present cpus
1586  */
1587
1588 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
1589         struct pkg_data *, struct thread_data *, struct core_data *,
1590         struct pkg_data *), struct thread_data *thread_base,
1591         struct core_data *core_base, struct pkg_data *pkg_base,
1592         struct thread_data *thread_base2, struct core_data *core_base2,
1593         struct pkg_data *pkg_base2)
1594 {
1595         int retval, pkg_no, core_no, thread_no;
1596
1597         for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
1598                 for (core_no = 0; core_no < topo.num_cores_per_pkg; ++core_no) {
1599                         for (thread_no = 0; thread_no <
1600                                 topo.num_threads_per_core; ++thread_no) {
1601                                 struct thread_data *t, *t2;
1602                                 struct core_data *c, *c2;
1603                                 struct pkg_data *p, *p2;
1604
1605                                 t = GET_THREAD(thread_base, thread_no, core_no, pkg_no);
1606
1607                                 if (cpu_is_not_present(t->cpu_id))
1608                                         continue;
1609
1610                                 t2 = GET_THREAD(thread_base2, thread_no, core_no, pkg_no);
1611
1612                                 c = GET_CORE(core_base, core_no, pkg_no);
1613                                 c2 = GET_CORE(core_base2, core_no, pkg_no);
1614
1615                                 p = GET_PKG(pkg_base, pkg_no);
1616                                 p2 = GET_PKG(pkg_base2, pkg_no);
1617
1618                                 retval = func(t, c, p, t2, c2, p2);
1619                                 if (retval)
1620                                         return retval;
1621                         }
1622                 }
1623         }
1624         return 0;
1625 }
1626
1627 /*
1628  * run func(cpu) on every cpu in /proc/stat
1629  * return max_cpu number
1630  */
1631 int for_all_proc_cpus(int (func)(int))
1632 {
1633         FILE *fp;
1634         int cpu_num;
1635         int retval;
1636
1637         fp = fopen_or_die(proc_stat, "r");
1638
1639         retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
1640         if (retval != 0)
1641                 err(1, "%s: failed to parse format", proc_stat);
1642
1643         while (1) {
1644                 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
1645                 if (retval != 1)
1646                         break;
1647
1648                 retval = func(cpu_num);
1649                 if (retval) {
1650                         fclose(fp);
1651                         return(retval);
1652                 }
1653         }
1654         fclose(fp);
1655         return 0;
1656 }
1657
1658 void re_initialize(void)
1659 {
1660         free_all_buffers();
1661         setup_all_buffers();
1662         printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
1663 }
1664
1665
1666 /*
1667  * count_cpus()
1668  * remember the last one seen, it will be the max
1669  */
1670 int count_cpus(int cpu)
1671 {
1672         if (topo.max_cpu_num < cpu)
1673                 topo.max_cpu_num = cpu;
1674
1675         topo.num_cpus += 1;
1676         return 0;
1677 }
1678 int mark_cpu_present(int cpu)
1679 {
1680         CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
1681         return 0;
1682 }
1683
1684 void turbostat_loop()
1685 {
1686         int retval;
1687         int restarted = 0;
1688
1689 restart:
1690         restarted++;
1691
1692         retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1693         if (retval < -1) {
1694                 exit(retval);
1695         } else if (retval == -1) {
1696                 if (restarted > 1) {
1697                         exit(retval);
1698                 }
1699                 re_initialize();
1700                 goto restart;
1701         }
1702         restarted = 0;
1703         gettimeofday(&tv_even, (struct timezone *)NULL);
1704
1705         while (1) {
1706                 if (for_all_proc_cpus(cpu_is_not_present)) {
1707                         re_initialize();
1708                         goto restart;
1709                 }
1710                 sleep(interval_sec);
1711                 retval = for_all_cpus(get_counters, ODD_COUNTERS);
1712                 if (retval < -1) {
1713                         exit(retval);
1714                 } else if (retval == -1) {
1715                         re_initialize();
1716                         goto restart;
1717                 }
1718                 gettimeofday(&tv_odd, (struct timezone *)NULL);
1719                 timersub(&tv_odd, &tv_even, &tv_delta);
1720                 for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
1721                 compute_average(EVEN_COUNTERS);
1722                 format_all_counters(EVEN_COUNTERS);
1723                 flush_stdout();
1724                 sleep(interval_sec);
1725                 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
1726                 if (retval < -1) {
1727                         exit(retval);
1728                 } else if (retval == -1) {
1729                         re_initialize();
1730                         goto restart;
1731                 }
1732                 gettimeofday(&tv_even, (struct timezone *)NULL);
1733                 timersub(&tv_even, &tv_odd, &tv_delta);
1734                 for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS);
1735                 compute_average(ODD_COUNTERS);
1736                 format_all_counters(ODD_COUNTERS);
1737                 flush_stdout();
1738         }
1739 }
1740
1741 void check_dev_msr()
1742 {
1743         struct stat sb;
1744         char pathname[32];
1745
1746         sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
1747         if (stat(pathname, &sb))
1748                 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
1749                         err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
1750 }
1751
1752 void check_permissions()
1753 {
1754         struct __user_cap_header_struct cap_header_data;
1755         cap_user_header_t cap_header = &cap_header_data;
1756         struct __user_cap_data_struct cap_data_data;
1757         cap_user_data_t cap_data = &cap_data_data;
1758         extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
1759         int do_exit = 0;
1760         char pathname[32];
1761
1762         /* check for CAP_SYS_RAWIO */
1763         cap_header->pid = getpid();
1764         cap_header->version = _LINUX_CAPABILITY_VERSION;
1765         if (capget(cap_header, cap_data) < 0)
1766                 err(-6, "capget(2) failed");
1767
1768         if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
1769                 do_exit++;
1770                 warnx("capget(CAP_SYS_RAWIO) failed,"
1771                         " try \"# setcap cap_sys_rawio=ep %s\"", progname);
1772         }
1773
1774         /* test file permissions */
1775         sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
1776         if (euidaccess(pathname, R_OK)) {
1777                 do_exit++;
1778                 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
1779         }
1780
1781         /* if all else fails, thell them to be root */
1782         if (do_exit)
1783                 if (getuid() != 0)
1784                         warnx("... or simply run as root");
1785
1786         if (do_exit)
1787                 exit(-6);
1788 }
1789
1790 /*
1791  * NHM adds support for additional MSRs:
1792  *
1793  * MSR_SMI_COUNT                   0x00000034
1794  *
1795  * MSR_NHM_PLATFORM_INFO           0x000000ce
1796  * MSR_NHM_SNB_PKG_CST_CFG_CTL     0x000000e2
1797  *
1798  * MSR_PKG_C3_RESIDENCY            0x000003f8
1799  * MSR_PKG_C6_RESIDENCY            0x000003f9
1800  * MSR_CORE_C3_RESIDENCY           0x000003fc
1801  * MSR_CORE_C6_RESIDENCY           0x000003fd
1802  *
1803  * Side effect:
1804  * sets global pkg_cstate_limit to decode MSR_NHM_SNB_PKG_CST_CFG_CTL
1805  */
1806 int probe_nhm_msrs(unsigned int family, unsigned int model)
1807 {
1808         unsigned long long msr;
1809         int *pkg_cstate_limits;
1810
1811         if (!genuine_intel)
1812                 return 0;
1813
1814         if (family != 6)
1815                 return 0;
1816
1817         switch (model) {
1818         case 0x1A:      /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
1819         case 0x1E:      /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
1820         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
1821         case 0x25:      /* Westmere Client - Clarkdale, Arrandale */
1822         case 0x2C:      /* Westmere EP - Gulftown */
1823         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1824         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1825                 pkg_cstate_limits = nhm_pkg_cstate_limits;
1826                 break;
1827         case 0x2A:      /* SNB */
1828         case 0x2D:      /* SNB Xeon */
1829         case 0x3A:      /* IVB */
1830         case 0x3E:      /* IVB Xeon */
1831                 pkg_cstate_limits = snb_pkg_cstate_limits;
1832                 break;
1833         case 0x3C:      /* HSW */
1834         case 0x3F:      /* HSX */
1835         case 0x45:      /* HSW */
1836         case 0x46:      /* HSW */
1837         case 0x3D:      /* BDW */
1838         case 0x47:      /* BDW */
1839         case 0x4F:      /* BDX */
1840         case 0x56:      /* BDX-DE */
1841         case 0x4E:      /* SKL */
1842         case 0x5E:      /* SKL */
1843                 pkg_cstate_limits = hsw_pkg_cstate_limits;
1844                 break;
1845         case 0x37:      /* BYT */
1846         case 0x4D:      /* AVN */
1847                 pkg_cstate_limits = slv_pkg_cstate_limits;
1848                 break;
1849         case 0x4C:      /* AMT */
1850                 pkg_cstate_limits = amt_pkg_cstate_limits;
1851                 break;
1852         case 0x57:      /* PHI */
1853                 pkg_cstate_limits = phi_pkg_cstate_limits;
1854                 break;
1855         default:
1856                 return 0;
1857         }
1858         get_msr(base_cpu, MSR_NHM_SNB_PKG_CST_CFG_CTL, &msr);
1859
1860         pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
1861
1862         return 1;
1863 }
1864 int has_nhm_turbo_ratio_limit(unsigned int family, unsigned int model)
1865 {
1866         switch (model) {
1867         /* Nehalem compatible, but do not include turbo-ratio limit support */
1868         case 0x2E:      /* Nehalem-EX Xeon - Beckton */
1869         case 0x2F:      /* Westmere-EX Xeon - Eagleton */
1870                 return 0;
1871         default:
1872                 return 1;
1873         }
1874 }
1875 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
1876 {
1877         if (!genuine_intel)
1878                 return 0;
1879
1880         if (family != 6)
1881                 return 0;
1882
1883         switch (model) {
1884         case 0x3E:      /* IVB Xeon */
1885         case 0x3F:      /* HSW Xeon */
1886                 return 1;
1887         default:
1888                 return 0;
1889         }
1890 }
1891 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
1892 {
1893         if (!genuine_intel)
1894                 return 0;
1895
1896         if (family != 6)
1897                 return 0;
1898
1899         switch (model) {
1900         case 0x3F:      /* HSW Xeon */
1901                 return 1;
1902         default:
1903                 return 0;
1904         }
1905 }
1906
1907 int has_knl_turbo_ratio_limit(unsigned int family, unsigned int model)
1908 {
1909         if (!genuine_intel)
1910                 return 0;
1911
1912         if (family != 6)
1913                 return 0;
1914
1915         switch (model) {
1916         case 0x57:      /* Knights Landing */
1917                 return 1;
1918         default:
1919                 return 0;
1920         }
1921 }
1922 int has_config_tdp(unsigned int family, unsigned int model)
1923 {
1924         if (!genuine_intel)
1925                 return 0;
1926
1927         if (family != 6)
1928                 return 0;
1929
1930         switch (model) {
1931         case 0x3A:      /* IVB */
1932         case 0x3C:      /* HSW */
1933         case 0x3F:      /* HSX */
1934         case 0x45:      /* HSW */
1935         case 0x46:      /* HSW */
1936         case 0x3D:      /* BDW */
1937         case 0x47:      /* BDW */
1938         case 0x4F:      /* BDX */
1939         case 0x56:      /* BDX-DE */
1940         case 0x4E:      /* SKL */
1941         case 0x5E:      /* SKL */
1942
1943         case 0x57:      /* Knights Landing */
1944                 return 1;
1945         default:
1946                 return 0;
1947         }
1948 }
1949
1950 static void
1951 dump_cstate_pstate_config_info(family, model)
1952 {
1953         if (!do_nhm_platform_info)
1954                 return;
1955
1956         dump_nhm_platform_info();
1957
1958         if (has_hsw_turbo_ratio_limit(family, model))
1959                 dump_hsw_turbo_ratio_limits();
1960
1961         if (has_ivt_turbo_ratio_limit(family, model))
1962                 dump_ivt_turbo_ratio_limits();
1963
1964         if (has_nhm_turbo_ratio_limit(family, model))
1965                 dump_nhm_turbo_ratio_limits();
1966
1967         if (has_knl_turbo_ratio_limit(family, model))
1968                 dump_knl_turbo_ratio_limits();
1969
1970         if (has_config_tdp(family, model))
1971                 dump_config_tdp();
1972
1973         dump_nhm_cst_cfg();
1974 }
1975
1976
1977 /*
1978  * print_epb()
1979  * Decode the ENERGY_PERF_BIAS MSR
1980  */
1981 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1982 {
1983         unsigned long long msr;
1984         char *epb_string;
1985         int cpu;
1986
1987         if (!has_epb)
1988                 return 0;
1989
1990         cpu = t->cpu_id;
1991
1992         /* EPB is per-package */
1993         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1994                 return 0;
1995
1996         if (cpu_migrate(cpu)) {
1997                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
1998                 return -1;
1999         }
2000
2001         if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
2002                 return 0;
2003
2004         switch (msr & 0xF) {
2005         case ENERGY_PERF_BIAS_PERFORMANCE:
2006                 epb_string = "performance";
2007                 break;
2008         case ENERGY_PERF_BIAS_NORMAL:
2009                 epb_string = "balanced";
2010                 break;
2011         case ENERGY_PERF_BIAS_POWERSAVE:
2012                 epb_string = "powersave";
2013                 break;
2014         default:
2015                 epb_string = "custom";
2016                 break;
2017         }
2018         fprintf(stderr, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
2019
2020         return 0;
2021 }
2022
2023 /*
2024  * print_perf_limit()
2025  */
2026 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2027 {
2028         unsigned long long msr;
2029         int cpu;
2030
2031         cpu = t->cpu_id;
2032
2033         /* per-package */
2034         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2035                 return 0;
2036
2037         if (cpu_migrate(cpu)) {
2038                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2039                 return -1;
2040         }
2041
2042         if (do_core_perf_limit_reasons) {
2043                 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
2044                 fprintf(stderr, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2045                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
2046                         (msr & 1 << 15) ? "bit15, " : "",
2047                         (msr & 1 << 14) ? "bit14, " : "",
2048                         (msr & 1 << 13) ? "Transitions, " : "",
2049                         (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
2050                         (msr & 1 << 11) ? "PkgPwrL2, " : "",
2051                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
2052                         (msr & 1 << 9) ? "CorePwr, " : "",
2053                         (msr & 1 << 8) ? "Amps, " : "",
2054                         (msr & 1 << 6) ? "VR-Therm, " : "",
2055                         (msr & 1 << 5) ? "Auto-HWP, " : "",
2056                         (msr & 1 << 4) ? "Graphics, " : "",
2057                         (msr & 1 << 2) ? "bit2, " : "",
2058                         (msr & 1 << 1) ? "ThermStatus, " : "",
2059                         (msr & 1 << 0) ? "PROCHOT, " : "");
2060                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
2061                         (msr & 1 << 31) ? "bit31, " : "",
2062                         (msr & 1 << 30) ? "bit30, " : "",
2063                         (msr & 1 << 29) ? "Transitions, " : "",
2064                         (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
2065                         (msr & 1 << 27) ? "PkgPwrL2, " : "",
2066                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
2067                         (msr & 1 << 25) ? "CorePwr, " : "",
2068                         (msr & 1 << 24) ? "Amps, " : "",
2069                         (msr & 1 << 22) ? "VR-Therm, " : "",
2070                         (msr & 1 << 21) ? "Auto-HWP, " : "",
2071                         (msr & 1 << 20) ? "Graphics, " : "",
2072                         (msr & 1 << 18) ? "bit18, " : "",
2073                         (msr & 1 << 17) ? "ThermStatus, " : "",
2074                         (msr & 1 << 16) ? "PROCHOT, " : "");
2075
2076         }
2077         if (do_gfx_perf_limit_reasons) {
2078                 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
2079                 fprintf(stderr, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2080                 fprintf(stderr, " (Active: %s%s%s%s%s%s%s%s)",
2081                         (msr & 1 << 0) ? "PROCHOT, " : "",
2082                         (msr & 1 << 1) ? "ThermStatus, " : "",
2083                         (msr & 1 << 4) ? "Graphics, " : "",
2084                         (msr & 1 << 6) ? "VR-Therm, " : "",
2085                         (msr & 1 << 8) ? "Amps, " : "",
2086                         (msr & 1 << 9) ? "GFXPwr, " : "",
2087                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
2088                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
2089                 fprintf(stderr, " (Logged: %s%s%s%s%s%s%s%s)\n",
2090                         (msr & 1 << 16) ? "PROCHOT, " : "",
2091                         (msr & 1 << 17) ? "ThermStatus, " : "",
2092                         (msr & 1 << 20) ? "Graphics, " : "",
2093                         (msr & 1 << 22) ? "VR-Therm, " : "",
2094                         (msr & 1 << 24) ? "Amps, " : "",
2095                         (msr & 1 << 25) ? "GFXPwr, " : "",
2096                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
2097                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
2098         }
2099         if (do_ring_perf_limit_reasons) {
2100                 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
2101                 fprintf(stderr, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
2102                 fprintf(stderr, " (Active: %s%s%s%s%s%s)",
2103                         (msr & 1 << 0) ? "PROCHOT, " : "",
2104                         (msr & 1 << 1) ? "ThermStatus, " : "",
2105                         (msr & 1 << 6) ? "VR-Therm, " : "",
2106                         (msr & 1 << 8) ? "Amps, " : "",
2107                         (msr & 1 << 10) ? "PkgPwrL1, " : "",
2108                         (msr & 1 << 11) ? "PkgPwrL2, " : "");
2109                 fprintf(stderr, " (Logged: %s%s%s%s%s%s)\n",
2110                         (msr & 1 << 16) ? "PROCHOT, " : "",
2111                         (msr & 1 << 17) ? "ThermStatus, " : "",
2112                         (msr & 1 << 22) ? "VR-Therm, " : "",
2113                         (msr & 1 << 24) ? "Amps, " : "",
2114                         (msr & 1 << 26) ? "PkgPwrL1, " : "",
2115                         (msr & 1 << 27) ? "PkgPwrL2, " : "");
2116         }
2117         return 0;
2118 }
2119
2120 #define RAPL_POWER_GRANULARITY  0x7FFF  /* 15 bit power granularity */
2121 #define RAPL_TIME_GRANULARITY   0x3F /* 6 bit time granularity */
2122
2123 double get_tdp(model)
2124 {
2125         unsigned long long msr;
2126
2127         if (do_rapl & RAPL_PKG_POWER_INFO)
2128                 if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
2129                         return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
2130
2131         switch (model) {
2132         case 0x37:
2133         case 0x4D:
2134                 return 30.0;
2135         default:
2136                 return 135.0;
2137         }
2138 }
2139
2140 /*
2141  * rapl_dram_energy_units_probe()
2142  * Energy units are either hard-coded, or come from RAPL Energy Unit MSR.
2143  */
2144 static double
2145 rapl_dram_energy_units_probe(int  model, double rapl_energy_units)
2146 {
2147         /* only called for genuine_intel, family 6 */
2148
2149         switch (model) {
2150         case 0x3F:      /* HSX */
2151         case 0x4F:      /* BDX */
2152         case 0x56:      /* BDX-DE */
2153         case 0x57:      /* KNL */
2154                 return (rapl_dram_energy_units = 15.3 / 1000000);
2155         default:
2156                 return (rapl_energy_units);
2157         }
2158 }
2159
2160
2161 /*
2162  * rapl_probe()
2163  *
2164  * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
2165  */
2166 void rapl_probe(unsigned int family, unsigned int model)
2167 {
2168         unsigned long long msr;
2169         unsigned int time_unit;
2170         double tdp;
2171
2172         if (!genuine_intel)
2173                 return;
2174
2175         if (family != 6)
2176                 return;
2177
2178         switch (model) {
2179         case 0x2A:
2180         case 0x3A:
2181         case 0x3C:      /* HSW */
2182         case 0x45:      /* HSW */
2183         case 0x46:      /* HSW */
2184         case 0x3D:      /* BDW */
2185         case 0x47:      /* BDW */
2186                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
2187                 break;
2188         case 0x4E:      /* SKL */
2189         case 0x5E:      /* SKL */
2190                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2191                 break;
2192         case 0x3F:      /* HSX */
2193         case 0x4F:      /* BDX */
2194         case 0x56:      /* BDX-DE */
2195         case 0x57:      /* KNL */
2196                 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
2197                 break;
2198         case 0x2D:
2199         case 0x3E:
2200                 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
2201                 break;
2202         case 0x37:      /* BYT */
2203         case 0x4D:      /* AVN */
2204                 do_rapl = RAPL_PKG | RAPL_CORES ;
2205                 break;
2206         default:
2207                 return;
2208         }
2209
2210         /* units on package 0, verify later other packages match */
2211         if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
2212                 return;
2213
2214         rapl_power_units = 1.0 / (1 << (msr & 0xF));
2215         if (model == 0x37)
2216                 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
2217         else
2218                 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
2219
2220         rapl_dram_energy_units = rapl_dram_energy_units_probe(model, rapl_energy_units);
2221
2222         time_unit = msr >> 16 & 0xF;
2223         if (time_unit == 0)
2224                 time_unit = 0xA;
2225
2226         rapl_time_units = 1.0 / (1 << (time_unit));
2227
2228         tdp = get_tdp(model);
2229
2230         rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
2231         if (debug)
2232                 fprintf(stderr, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
2233
2234         return;
2235 }
2236
2237 void perf_limit_reasons_probe(family, model)
2238 {
2239         if (!genuine_intel)
2240                 return;
2241
2242         if (family != 6)
2243                 return;
2244
2245         switch (model) {
2246         case 0x3C:      /* HSW */
2247         case 0x45:      /* HSW */
2248         case 0x46:      /* HSW */
2249                 do_gfx_perf_limit_reasons = 1;
2250         case 0x3F:      /* HSX */
2251                 do_core_perf_limit_reasons = 1;
2252                 do_ring_perf_limit_reasons = 1;
2253         default:
2254                 return;
2255         }
2256 }
2257
2258 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2259 {
2260         unsigned long long msr;
2261         unsigned int dts;
2262         int cpu;
2263
2264         if (!(do_dts || do_ptm))
2265                 return 0;
2266
2267         cpu = t->cpu_id;
2268
2269         /* DTS is per-core, no need to print for each thread */
2270         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE)) 
2271                 return 0;
2272
2273         if (cpu_migrate(cpu)) {
2274                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2275                 return -1;
2276         }
2277
2278         if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
2279                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
2280                         return 0;
2281
2282                 dts = (msr >> 16) & 0x7F;
2283                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
2284                         cpu, msr, tcc_activation_temp - dts);
2285
2286 #ifdef  THERM_DEBUG
2287                 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
2288                         return 0;
2289
2290                 dts = (msr >> 16) & 0x7F;
2291                 dts2 = (msr >> 8) & 0x7F;
2292                 fprintf(stderr, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2293                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2294 #endif
2295         }
2296
2297
2298         if (do_dts) {
2299                 unsigned int resolution;
2300
2301                 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
2302                         return 0;
2303
2304                 dts = (msr >> 16) & 0x7F;
2305                 resolution = (msr >> 27) & 0xF;
2306                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
2307                         cpu, msr, tcc_activation_temp - dts, resolution);
2308
2309 #ifdef THERM_DEBUG
2310                 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
2311                         return 0;
2312
2313                 dts = (msr >> 16) & 0x7F;
2314                 dts2 = (msr >> 8) & 0x7F;
2315                 fprintf(stderr, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
2316                         cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
2317 #endif
2318         }
2319
2320         return 0;
2321 }
2322         
2323 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
2324 {
2325         fprintf(stderr, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
2326                 cpu, label,
2327                 ((msr >> 15) & 1) ? "EN" : "DIS",
2328                 ((msr >> 0) & 0x7FFF) * rapl_power_units,
2329                 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
2330                 (((msr >> 16) & 1) ? "EN" : "DIS"));
2331
2332         return;
2333 }
2334
2335 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2336 {
2337         unsigned long long msr;
2338         int cpu;
2339
2340         if (!do_rapl)
2341                 return 0;
2342
2343         /* RAPL counters are per package, so print only for 1st thread/package */
2344         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2345                 return 0;
2346
2347         cpu = t->cpu_id;
2348         if (cpu_migrate(cpu)) {
2349                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2350                 return -1;
2351         }
2352
2353         if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
2354                 return -1;
2355
2356         if (debug) {
2357                 fprintf(stderr, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx "
2358                         "(%f Watts, %f Joules, %f sec.)\n", cpu, msr,
2359                         rapl_power_units, rapl_energy_units, rapl_time_units);
2360         }
2361         if (do_rapl & RAPL_PKG_POWER_INFO) {
2362
2363                 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
2364                         return -5;
2365
2366
2367                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2368                         cpu, msr,
2369                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2370                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2371                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2372                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2373
2374         }
2375         if (do_rapl & RAPL_PKG) {
2376
2377                 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
2378                         return -9;
2379
2380                 fprintf(stderr, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
2381                         cpu, msr, (msr >> 63) & 1 ? "": "UN");
2382
2383                 print_power_limit_msr(cpu, msr, "PKG Limit #1");
2384                 fprintf(stderr, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
2385                         cpu,
2386                         ((msr >> 47) & 1) ? "EN" : "DIS",
2387                         ((msr >> 32) & 0x7FFF) * rapl_power_units,
2388                         (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
2389                         ((msr >> 48) & 1) ? "EN" : "DIS");
2390         }
2391
2392         if (do_rapl & RAPL_DRAM_POWER_INFO) {
2393                 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
2394                         return -6;
2395
2396                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
2397                         cpu, msr,
2398                         ((msr >>  0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2399                         ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2400                         ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
2401                         ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
2402         }
2403         if (do_rapl & RAPL_DRAM) {
2404                 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
2405                         return -9;
2406                 fprintf(stderr, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
2407                                 cpu, msr, (msr >> 31) & 1 ? "": "UN");
2408
2409                 print_power_limit_msr(cpu, msr, "DRAM Limit");
2410         }
2411         if (do_rapl & RAPL_CORE_POLICY) {
2412                 if (debug) {
2413                         if (get_msr(cpu, MSR_PP0_POLICY, &msr))
2414                                 return -7;
2415
2416                         fprintf(stderr, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
2417                 }
2418         }
2419         if (do_rapl & RAPL_CORES) {
2420                 if (debug) {
2421
2422                         if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
2423                                 return -9;
2424                         fprintf(stderr, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
2425                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2426                         print_power_limit_msr(cpu, msr, "Cores Limit");
2427                 }
2428         }
2429         if (do_rapl & RAPL_GFX) {
2430                 if (debug) {
2431                         if (get_msr(cpu, MSR_PP1_POLICY, &msr))
2432                                 return -8;
2433
2434                         fprintf(stderr, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
2435
2436                         if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
2437                                 return -9;
2438                         fprintf(stderr, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
2439                                         cpu, msr, (msr >> 31) & 1 ? "": "UN");
2440                         print_power_limit_msr(cpu, msr, "GFX Limit");
2441                 }
2442         }
2443         return 0;
2444 }
2445
2446 /*
2447  * SNB adds support for additional MSRs:
2448  *
2449  * MSR_PKG_C7_RESIDENCY            0x000003fa
2450  * MSR_CORE_C7_RESIDENCY           0x000003fe
2451  * MSR_PKG_C2_RESIDENCY            0x0000060d
2452  */
2453
2454 int has_snb_msrs(unsigned int family, unsigned int model)
2455 {
2456         if (!genuine_intel)
2457                 return 0;
2458
2459         switch (model) {
2460         case 0x2A:
2461         case 0x2D:
2462         case 0x3A:      /* IVB */
2463         case 0x3E:      /* IVB Xeon */
2464         case 0x3C:      /* HSW */
2465         case 0x3F:      /* HSW */
2466         case 0x45:      /* HSW */
2467         case 0x46:      /* HSW */
2468         case 0x3D:      /* BDW */
2469         case 0x47:      /* BDW */
2470         case 0x4F:      /* BDX */
2471         case 0x56:      /* BDX-DE */
2472         case 0x4E:      /* SKL */
2473         case 0x5E:      /* SKL */
2474                 return 1;
2475         }
2476         return 0;
2477 }
2478
2479 /*
2480  * HSW adds support for additional MSRs:
2481  *
2482  * MSR_PKG_C8_RESIDENCY            0x00000630
2483  * MSR_PKG_C9_RESIDENCY            0x00000631
2484  * MSR_PKG_C10_RESIDENCY           0x00000632
2485  */
2486 int has_hsw_msrs(unsigned int family, unsigned int model)
2487 {
2488         if (!genuine_intel)
2489                 return 0;
2490
2491         switch (model) {
2492         case 0x45:      /* HSW */
2493         case 0x3D:      /* BDW */
2494         case 0x4E:      /* SKL */
2495         case 0x5E:      /* SKL */
2496                 return 1;
2497         }
2498         return 0;
2499 }
2500
2501 /*
2502  * SKL adds support for additional MSRS:
2503  *
2504  * MSR_PKG_WEIGHTED_CORE_C0_RES    0x00000658
2505  * MSR_PKG_ANY_CORE_C0_RES         0x00000659
2506  * MSR_PKG_ANY_GFXE_C0_RES         0x0000065A
2507  * MSR_PKG_BOTH_CORE_GFXE_C0_RES   0x0000065B
2508  */
2509 int has_skl_msrs(unsigned int family, unsigned int model)
2510 {
2511         if (!genuine_intel)
2512                 return 0;
2513
2514         switch (model) {
2515         case 0x4E:      /* SKL */
2516         case 0x5E:      /* SKL */
2517                 return 1;
2518         }
2519         return 0;
2520 }
2521
2522
2523
2524 int is_slm(unsigned int family, unsigned int model)
2525 {
2526         if (!genuine_intel)
2527                 return 0;
2528         switch (model) {
2529         case 0x37:      /* BYT */
2530         case 0x4D:      /* AVN */
2531                 return 1;
2532         }
2533         return 0;
2534 }
2535
2536 int is_knl(unsigned int family, unsigned int model)
2537 {
2538         if (!genuine_intel)
2539                 return 0;
2540         switch (model) {
2541         case 0x57:      /* KNL */
2542                 return 1;
2543         }
2544         return 0;
2545 }
2546
2547 unsigned int get_aperf_mperf_multiplier(unsigned int family, unsigned int model)
2548 {
2549         if (is_knl(family, model))
2550                 return 1024;
2551         return 1;
2552 }
2553
2554 #define SLM_BCLK_FREQS 5
2555 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
2556
2557 double slm_bclk(void)
2558 {
2559         unsigned long long msr = 3;
2560         unsigned int i;
2561         double freq;
2562
2563         if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
2564                 fprintf(stderr, "SLM BCLK: unknown\n");
2565
2566         i = msr & 0xf;
2567         if (i >= SLM_BCLK_FREQS) {
2568                 fprintf(stderr, "SLM BCLK[%d] invalid\n", i);
2569                 msr = 3;
2570         }
2571         freq = slm_freq_table[i];
2572
2573         fprintf(stderr, "SLM BCLK: %.1f Mhz\n", freq);
2574
2575         return freq;
2576 }
2577
2578 double discover_bclk(unsigned int family, unsigned int model)
2579 {
2580         if (has_snb_msrs(family, model))
2581                 return 100.00;
2582         else if (is_slm(family, model))
2583                 return slm_bclk();
2584         else
2585                 return 133.33;
2586 }
2587
2588 /*
2589  * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
2590  * the Thermal Control Circuit (TCC) activates.
2591  * This is usually equal to tjMax.
2592  *
2593  * Older processors do not have this MSR, so there we guess,
2594  * but also allow cmdline over-ride with -T.
2595  *
2596  * Several MSR temperature values are in units of degrees-C
2597  * below this value, including the Digital Thermal Sensor (DTS),
2598  * Package Thermal Management Sensor (PTM), and thermal event thresholds.
2599  */
2600 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
2601 {
2602         unsigned long long msr;
2603         unsigned int target_c_local;
2604         int cpu;
2605
2606         /* tcc_activation_temp is used only for dts or ptm */
2607         if (!(do_dts || do_ptm))
2608                 return 0;
2609
2610         /* this is a per-package concept */
2611         if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2612                 return 0;
2613
2614         cpu = t->cpu_id;
2615         if (cpu_migrate(cpu)) {
2616                 fprintf(stderr, "Could not migrate to CPU %d\n", cpu);
2617                 return -1;
2618         }
2619
2620         if (tcc_activation_temp_override != 0) {
2621                 tcc_activation_temp = tcc_activation_temp_override;
2622                 fprintf(stderr, "cpu%d: Using cmdline TCC Target (%d C)\n",
2623                         cpu, tcc_activation_temp);
2624                 return 0;
2625         }
2626
2627         /* Temperature Target MSR is Nehalem and newer only */
2628         if (!do_nhm_platform_info)
2629                 goto guess;
2630
2631         if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
2632                 goto guess;
2633
2634         target_c_local = (msr >> 16) & 0xFF;
2635
2636         if (debug)
2637                 fprintf(stderr, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
2638                         cpu, msr, target_c_local);
2639
2640         if (!target_c_local)
2641                 goto guess;
2642
2643         tcc_activation_temp = target_c_local;
2644
2645         return 0;
2646
2647 guess:
2648         tcc_activation_temp = TJMAX_DEFAULT;
2649         fprintf(stderr, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
2650                 cpu, tcc_activation_temp);
2651
2652         return 0;
2653 }
2654 void process_cpuid()
2655 {
2656         unsigned int eax, ebx, ecx, edx, max_level;
2657         unsigned int fms, family, model, stepping;
2658
2659         eax = ebx = ecx = edx = 0;
2660
2661         __get_cpuid(0, &max_level, &ebx, &ecx, &edx);
2662
2663         if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
2664                 genuine_intel = 1;
2665
2666         if (debug)
2667                 fprintf(stderr, "CPUID(0): %.4s%.4s%.4s ",
2668                         (char *)&ebx, (char *)&edx, (char *)&ecx);
2669
2670         __get_cpuid(1, &fms, &ebx, &ecx, &edx);
2671         family = (fms >> 8) & 0xf;
2672         model = (fms >> 4) & 0xf;
2673         stepping = fms & 0xf;
2674         if (family == 6 || family == 0xf)
2675                 model += ((fms >> 16) & 0xf) << 4;
2676
2677         if (debug)
2678                 fprintf(stderr, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
2679                         max_level, family, model, stepping, family, model, stepping);
2680
2681         if (!(edx & (1 << 5)))
2682                 errx(1, "CPUID: no MSR");
2683
2684         /*
2685          * check max extended function levels of CPUID.
2686          * This is needed to check for invariant TSC.
2687          * This check is valid for both Intel and AMD.
2688          */
2689         ebx = ecx = edx = 0;
2690         __get_cpuid(0x80000000, &max_level, &ebx, &ecx, &edx);
2691
2692         if (max_level >= 0x80000007) {
2693
2694                 /*
2695                  * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
2696                  * this check is valid for both Intel and AMD
2697                  */
2698                 __get_cpuid(0x80000007, &eax, &ebx, &ecx, &edx);
2699                 has_invariant_tsc = edx & (1 << 8);
2700         }
2701
2702         /*
2703          * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
2704          * this check is valid for both Intel and AMD
2705          */
2706
2707         __get_cpuid(0x6, &eax, &ebx, &ecx, &edx);
2708         has_aperf = ecx & (1 << 0);
2709         do_dts = eax & (1 << 0);
2710         do_ptm = eax & (1 << 6);
2711         has_epb = ecx & (1 << 3);
2712
2713         if (debug)
2714                 fprintf(stderr, "CPUID(6): %sAPERF, %sDTS, %sPTM, %sEPB\n",
2715                         has_aperf ? "" : "No ",
2716                         do_dts ? "" : "No ",
2717                         do_ptm ? "" : "No ",
2718                         has_epb ? "" : "No ");
2719
2720         if (max_level > 0x15) {
2721                 unsigned int eax_crystal;
2722                 unsigned int ebx_tsc;
2723
2724                 /*
2725                  * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
2726                  */
2727                 eax_crystal = ebx_tsc = crystal_hz = edx = 0;
2728                 __get_cpuid(0x15, &eax_crystal, &ebx_tsc, &crystal_hz, &edx);
2729
2730                 if (ebx_tsc != 0) {
2731
2732                         if (debug && (ebx != 0))
2733                                 fprintf(stderr, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
2734                                         eax_crystal, ebx_tsc, crystal_hz);
2735
2736                         if (crystal_hz == 0)
2737                                 switch(model) {
2738                                 case 0x4E:      /* SKL */
2739                                 case 0x5E:      /* SKL */
2740                                         crystal_hz = 24000000;  /* 24 MHz */
2741                                         break;
2742                                 default:
2743                                         crystal_hz = 0;
2744                         }
2745
2746                         if (crystal_hz) {
2747                                 tsc_hz =  (unsigned long long) crystal_hz * ebx_tsc / eax_crystal;
2748                                 if (debug)
2749                                         fprintf(stderr, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
2750                                                 tsc_hz / 1000000, crystal_hz, ebx_tsc,  eax_crystal);
2751                         }
2752                 }
2753         }
2754
2755         if (has_aperf)
2756                 aperf_mperf_multiplier = get_aperf_mperf_multiplier(family, model);
2757
2758         do_nhm_platform_info = do_nhm_cstates = do_smi = probe_nhm_msrs(family, model);
2759         do_snb_cstates = has_snb_msrs(family, model);
2760         do_pc2 = do_snb_cstates && (pkg_cstate_limit >= PCL__2);
2761         do_pc3 = (pkg_cstate_limit >= PCL__3);
2762         do_pc6 = (pkg_cstate_limit >= PCL__6);
2763         do_pc7 = do_snb_cstates && (pkg_cstate_limit >= PCL__7);
2764         do_c8_c9_c10 = has_hsw_msrs(family, model);
2765         do_skl_residency = has_skl_msrs(family, model);
2766         do_slm_cstates = is_slm(family, model);
2767         do_knl_cstates  = is_knl(family, model);
2768         bclk = discover_bclk(family, model);
2769
2770         rapl_probe(family, model);
2771         perf_limit_reasons_probe(family, model);
2772
2773         if (debug)
2774                 dump_cstate_pstate_config_info();
2775
2776         return;
2777 }
2778
2779 void help()
2780 {
2781         fprintf(stderr,
2782         "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
2783         "\n"
2784         "Turbostat forks the specified COMMAND and prints statistics\n"
2785         "when COMMAND completes.\n"
2786         "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
2787         "to print statistics, until interrupted.\n"
2788         "--debug        run in \"debug\" mode\n"
2789         "--interval sec Override default 5-second measurement interval\n"
2790         "--help         print this help message\n"
2791         "--counter msr  print 32-bit counter at address \"msr\"\n"
2792         "--Counter msr  print 64-bit Counter at address \"msr\"\n"
2793         "--msr msr      print 32-bit value at address \"msr\"\n"
2794         "--MSR msr      print 64-bit Value at address \"msr\"\n"
2795         "--version      print version information\n"
2796         "\n"
2797         "For more help, run \"man turbostat\"\n");
2798 }
2799
2800
2801 /*
2802  * in /dev/cpu/ return success for names that are numbers
2803  * ie. filter out ".", "..", "microcode".
2804  */
2805 int dir_filter(const struct dirent *dirp)
2806 {
2807         if (isdigit(dirp->d_name[0]))
2808                 return 1;
2809         else
2810                 return 0;
2811 }
2812
2813 int open_dev_cpu_msr(int dummy1)
2814 {
2815         return 0;
2816 }
2817
2818 void topology_probe()
2819 {
2820         int i;
2821         int max_core_id = 0;
2822         int max_package_id = 0;
2823         int max_siblings = 0;
2824         struct cpu_topology {
2825                 int core_id;
2826                 int physical_package_id;
2827         } *cpus;
2828
2829         /* Initialize num_cpus, max_cpu_num */
2830         topo.num_cpus = 0;
2831         topo.max_cpu_num = 0;
2832         for_all_proc_cpus(count_cpus);
2833         if (!summary_only && topo.num_cpus > 1)
2834                 show_cpu = 1;
2835
2836         if (debug > 1)
2837                 fprintf(stderr, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
2838
2839         cpus = calloc(1, (topo.max_cpu_num  + 1) * sizeof(struct cpu_topology));
2840         if (cpus == NULL)
2841                 err(1, "calloc cpus");
2842
2843         /*
2844          * Allocate and initialize cpu_present_set
2845          */
2846         cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
2847         if (cpu_present_set == NULL)
2848                 err(3, "CPU_ALLOC");
2849         cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2850         CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
2851         for_all_proc_cpus(mark_cpu_present);
2852
2853         /*
2854          * Allocate and initialize cpu_affinity_set
2855          */
2856         cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
2857         if (cpu_affinity_set == NULL)
2858                 err(3, "CPU_ALLOC");
2859         cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2860         CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
2861
2862
2863         /*
2864          * For online cpus
2865          * find max_core_id, max_package_id
2866          */
2867         for (i = 0; i <= topo.max_cpu_num; ++i) {
2868                 int siblings;
2869
2870                 if (cpu_is_not_present(i)) {
2871                         if (debug > 1)
2872                                 fprintf(stderr, "cpu%d NOT PRESENT\n", i);
2873                         continue;
2874                 }
2875                 cpus[i].core_id = get_core_id(i);
2876                 if (cpus[i].core_id > max_core_id)
2877                         max_core_id = cpus[i].core_id;
2878
2879                 cpus[i].physical_package_id = get_physical_package_id(i);
2880                 if (cpus[i].physical_package_id > max_package_id)
2881                         max_package_id = cpus[i].physical_package_id;
2882
2883                 siblings = get_num_ht_siblings(i);
2884                 if (siblings > max_siblings)
2885                         max_siblings = siblings;
2886                 if (debug > 1)
2887                         fprintf(stderr, "cpu %d pkg %d core %d\n",
2888                                 i, cpus[i].physical_package_id, cpus[i].core_id);
2889         }
2890         topo.num_cores_per_pkg = max_core_id + 1;
2891         if (debug > 1)
2892                 fprintf(stderr, "max_core_id %d, sizing for %d cores per package\n",
2893                         max_core_id, topo.num_cores_per_pkg);
2894         if (debug && !summary_only && topo.num_cores_per_pkg > 1)
2895                 show_core = 1;
2896
2897         topo.num_packages = max_package_id + 1;
2898         if (debug > 1)
2899                 fprintf(stderr, "max_package_id %d, sizing for %d packages\n",
2900                         max_package_id, topo.num_packages);
2901         if (debug && !summary_only && topo.num_packages > 1)
2902                 show_pkg = 1;
2903
2904         topo.num_threads_per_core = max_siblings;
2905         if (debug > 1)
2906                 fprintf(stderr, "max_siblings %d\n", max_siblings);
2907
2908         free(cpus);
2909 }
2910
2911 void
2912 allocate_counters(struct thread_data **t, struct core_data **c, struct pkg_data **p)
2913 {
2914         int i;
2915
2916         *t = calloc(topo.num_threads_per_core * topo.num_cores_per_pkg *
2917                 topo.num_packages, sizeof(struct thread_data));
2918         if (*t == NULL)
2919                 goto error;
2920
2921         for (i = 0; i < topo.num_threads_per_core *
2922                 topo.num_cores_per_pkg * topo.num_packages; i++)
2923                 (*t)[i].cpu_id = -1;
2924
2925         *c = calloc(topo.num_cores_per_pkg * topo.num_packages,
2926                 sizeof(struct core_data));
2927         if (*c == NULL)
2928                 goto error;
2929
2930         for (i = 0; i < topo.num_cores_per_pkg * topo.num_packages; i++)
2931                 (*c)[i].core_id = -1;
2932
2933         *p = calloc(topo.num_packages, sizeof(struct pkg_data));
2934         if (*p == NULL)
2935                 goto error;
2936
2937         for (i = 0; i < topo.num_packages; i++)
2938                 (*p)[i].package_id = i;
2939
2940         return;
2941 error:
2942         err(1, "calloc counters");
2943 }
2944 /*
2945  * init_counter()
2946  *
2947  * set cpu_id, core_num, pkg_num
2948  * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
2949  *
2950  * increment topo.num_cores when 1st core in pkg seen
2951  */
2952 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
2953         struct pkg_data *pkg_base, int thread_num, int core_num,
2954         int pkg_num, int cpu_id)
2955 {
2956         struct thread_data *t;
2957         struct core_data *c;
2958         struct pkg_data *p;
2959
2960         t = GET_THREAD(thread_base, thread_num, core_num, pkg_num);
2961         c = GET_CORE(core_base, core_num, pkg_num);
2962         p = GET_PKG(pkg_base, pkg_num);
2963
2964         t->cpu_id = cpu_id;
2965         if (thread_num == 0) {
2966                 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
2967                 if (cpu_is_first_core_in_package(cpu_id))
2968                         t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
2969         }
2970
2971         c->core_id = core_num;
2972         p->package_id = pkg_num;
2973 }
2974
2975
2976 int initialize_counters(int cpu_id)
2977 {
2978         int my_thread_id, my_core_id, my_package_id;
2979
2980         my_package_id = get_physical_package_id(cpu_id);
2981         my_core_id = get_core_id(cpu_id);
2982         my_thread_id = get_cpu_position_in_core(cpu_id);
2983         if (!my_thread_id)
2984                 topo.num_cores++;
2985
2986         init_counter(EVEN_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2987         init_counter(ODD_COUNTERS, my_thread_id, my_core_id, my_package_id, cpu_id);
2988         return 0;
2989 }
2990
2991 void allocate_output_buffer()
2992 {
2993         output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
2994         outp = output_buffer;
2995         if (outp == NULL)
2996                 err(-1, "calloc output buffer");
2997 }
2998
2999 void setup_all_buffers(void)
3000 {
3001         topology_probe();
3002         allocate_counters(&thread_even, &core_even, &package_even);
3003         allocate_counters(&thread_odd, &core_odd, &package_odd);
3004         allocate_output_buffer();
3005         for_all_proc_cpus(initialize_counters);
3006 }
3007
3008 void set_base_cpu(void)
3009 {
3010         base_cpu = sched_getcpu();
3011         if (base_cpu < 0)
3012                 err(-ENODEV, "No valid cpus found");
3013
3014         if (debug > 1)
3015                 fprintf(stderr, "base_cpu = %d\n", base_cpu);
3016 }
3017
3018 void turbostat_init()
3019 {
3020         setup_all_buffers();
3021         set_base_cpu();
3022         check_dev_msr();
3023         check_permissions();
3024         process_cpuid();
3025
3026
3027         if (debug)
3028                 for_all_cpus(print_epb, ODD_COUNTERS);
3029
3030         if (debug)
3031                 for_all_cpus(print_perf_limit, ODD_COUNTERS);
3032
3033         if (debug)
3034                 for_all_cpus(print_rapl, ODD_COUNTERS);
3035
3036         for_all_cpus(set_temperature_target, ODD_COUNTERS);
3037
3038         if (debug)
3039                 for_all_cpus(print_thermal, ODD_COUNTERS);
3040 }
3041
3042 int fork_it(char **argv)
3043 {
3044         pid_t child_pid;
3045         int status;
3046
3047         status = for_all_cpus(get_counters, EVEN_COUNTERS);
3048         if (status)
3049                 exit(status);
3050         /* clear affinity side-effect of get_counters() */
3051         sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
3052         gettimeofday(&tv_even, (struct timezone *)NULL);
3053
3054         child_pid = fork();
3055         if (!child_pid) {
3056                 /* child */
3057                 execvp(argv[0], argv);
3058         } else {
3059
3060                 /* parent */
3061                 if (child_pid == -1)
3062                         err(1, "fork");
3063
3064                 signal(SIGINT, SIG_IGN);
3065                 signal(SIGQUIT, SIG_IGN);
3066                 if (waitpid(child_pid, &status, 0) == -1)
3067                         err(status, "waitpid");
3068         }
3069         /*
3070          * n.b. fork_it() does not check for errors from for_all_cpus()
3071          * because re-starting is problematic when forking
3072          */
3073         for_all_cpus(get_counters, ODD_COUNTERS);
3074         gettimeofday(&tv_odd, (struct timezone *)NULL);
3075         timersub(&tv_odd, &tv_even, &tv_delta);
3076         for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS);
3077         compute_average(EVEN_COUNTERS);
3078         format_all_counters(EVEN_COUNTERS);
3079         flush_stderr();
3080
3081         fprintf(stderr, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
3082
3083         return status;
3084 }
3085
3086 int get_and_dump_counters(void)
3087 {
3088         int status;
3089
3090         status = for_all_cpus(get_counters, ODD_COUNTERS);
3091         if (status)
3092                 return status;
3093
3094         status = for_all_cpus(dump_counters, ODD_COUNTERS);
3095         if (status)
3096                 return status;
3097
3098         flush_stdout();
3099
3100         return status;
3101 }
3102
3103 void print_version() {
3104         fprintf(stderr, "turbostat version 4.7 17-June, 2015"
3105                 " - Len Brown <lenb@kernel.org>\n");
3106 }
3107
3108 void cmdline(int argc, char **argv)
3109 {
3110         int opt;
3111         int option_index = 0;
3112         static struct option long_options[] = {
3113                 {"Counter",     required_argument,      0, 'C'},
3114                 {"counter",     required_argument,      0, 'c'},
3115                 {"Dump",        no_argument,            0, 'D'},
3116                 {"debug",       no_argument,            0, 'd'},
3117                 {"interval",    required_argument,      0, 'i'},
3118                 {"help",        no_argument,            0, 'h'},
3119                 {"Joules",      no_argument,            0, 'J'},
3120                 {"MSR",         required_argument,      0, 'M'},
3121                 {"msr",         required_argument,      0, 'm'},
3122                 {"Package",     no_argument,            0, 'p'},
3123                 {"processor",   no_argument,            0, 'p'},
3124                 {"Summary",     no_argument,            0, 'S'},
3125                 {"TCC",         required_argument,      0, 'T'},
3126                 {"version",     no_argument,            0, 'v' },
3127                 {0,             0,                      0,  0 }
3128         };
3129
3130         progname = argv[0];
3131
3132         while ((opt = getopt_long_only(argc, argv, "+C:c:Ddhi:JM:m:PpST:v",
3133                                 long_options, &option_index)) != -1) {
3134                 switch (opt) {
3135                 case 'C':
3136                         sscanf(optarg, "%x", &extra_delta_offset64);
3137                         break;
3138                 case 'c':
3139                         sscanf(optarg, "%x", &extra_delta_offset32);
3140                         break;
3141                 case 'D':
3142                         dump_only++;
3143                         break;
3144                 case 'd':
3145                         debug++;
3146                         break;
3147                 case 'h':
3148                 default:
3149                         help();
3150                         exit(1);
3151                 case 'i':
3152                         interval_sec = atoi(optarg);
3153                         break;
3154                 case 'J':
3155                         rapl_joules++;
3156                         break;
3157                 case 'M':
3158                         sscanf(optarg, "%x", &extra_msr_offset64);
3159                         break;
3160                 case 'm':
3161                         sscanf(optarg, "%x", &extra_msr_offset32);
3162                         break;
3163                 case 'P':
3164                         show_pkg_only++;
3165                         break;
3166                 case 'p':
3167                         show_core_only++;
3168                         break;
3169                 case 'S':
3170                         summary_only++;
3171                         break;
3172                 case 'T':
3173                         tcc_activation_temp_override = atoi(optarg);
3174                         break;
3175                 case 'v':
3176                         print_version();
3177                         exit(0);
3178                         break;
3179                 }
3180         }
3181 }
3182
3183 int main(int argc, char **argv)
3184 {
3185         cmdline(argc, argv);
3186
3187         if (debug)
3188                 print_version();
3189
3190         turbostat_init();
3191
3192         /* dump counters and exit */
3193         if (dump_only)
3194                 return get_and_dump_counters();
3195
3196         /*
3197          * if any params left, it must be a command to fork
3198          */
3199         if (argc - optind)
3200                 return fork_it(argv + optind);
3201         else
3202                 turbostat_loop();
3203
3204         return 0;
3205 }